[llvm-branch-commits] [llvm] f6b9afa - [AMDGPU] Extend and reorganize memory legalizer tests

Thu Dec 3 11:42:00 PST 2020

Author: Scott Linder
Date: 2020-12-03T19:36:33Z
New Revision: f6b9afae00d685fe869c64a27502cb23ab5aed5e

URL: https://github.com/llvm/llvm-project/commit/f6b9afae00d685fe869c64a27502cb23ab5aed5e
DIFF: https://github.com/llvm/llvm-project/commit/f6b9afae00d685fe869c64a27502cb23ab5aed5e.diff

LOG: [AMDGPU] Extend and reorganize memory legalizer tests

* Rename some tests to try to make a convention (where all components
  are optional) of:

    <addrspace>_<syncscope>_<memory-orders>_<operation>

* Split up at a level of granularity appropriate for the different RUN
  lines (i.e. split on addrspace so GFX6 can avoid FLAT) and that makes
  running a specific test reasonable in terms of wall time taken. This
  also means when run as part of the test suite the testing is not one
  serial bottleneck.

* Auto-generate check lines with `update_llc_test_checks.py` to make
  future maintenance more tractable.

Reviewed By: rampitec, t-tye

Differential Revision: https://reviews.llvm.org/D91545

Added: 
    llvm/test/CodeGen/AMDGPU/memory-legalizer-fence.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll

Modified: 
    

Removed: 
    llvm/test/CodeGen/AMDGPU/memory-legalizer-amdpal.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-cmpxchg.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-fence.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-rmw.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-load.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-mesa3d.ll
    llvm/test/CodeGen/AMDGPU/memory-legalizer-store.ll


################################################################################
diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-amdpal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-amdpal.ll
deleted file mode 100644
index b414c83374b8..000000000000

--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-amdpal.ll
+++ /dev/null
@@ -1,526 +0,0 @@
-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GCN9,CACHE_INV %s
-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GCN9,CACHE_INV %s
-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GCN9,CACHE_INV %s
-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GCN9,CACHE_INV %s
-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GCN10,CACHE_INV10 %s
-
-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GCN9,SKIP_CACHE_INV %s
-; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1010 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GCN10,SKIP_CACHE_INV %s
-
-
-; FUNC-LABEL: {{^}}system_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_acquire() {
-entry:
-  fence acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_release() {
-entry:
-  fence release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_acq_rel() {
-entry:
-  fence acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_seq_cst() {
-entry:
-  fence seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_acquire() {
-entry:
-  fence syncscope("one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_release() {
-entry:
-  fence syncscope("one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_acq_rel() {
-entry:
-  fence syncscope("one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:    buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_seq_cst() {
-entry:
-  fence syncscope("one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_acquire() {
-entry:
-  fence syncscope("singlethread") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_release() {
-entry:
-  fence syncscope("singlethread") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_acq_rel() {
-entry:
-  fence syncscope("singlethread") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_seq_cst() {
-entry:
-  fence syncscope("singlethread") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_acquire() {
-entry:
-  fence syncscope("singlethread-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_release() {
-entry:
-  fence syncscope("singlethread-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_acq_rel() {
-entry:
-  fence syncscope("singlethread-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_seq_cst() {
-entry:
-  fence syncscope("singlethread-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_acquire() {
-entry:
-  fence syncscope("agent") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_release() {
-entry:
-  fence syncscope("agent") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_acq_rel() {
-entry:
-  fence syncscope("agent") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_seq_cst() {
-entry:
-  fence syncscope("agent") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_acquire() {
-entry:
-  fence syncscope("agent-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_release() {
-entry:
-  fence syncscope("agent-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_acq_rel() {
-entry:
-  fence syncscope("agent-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; CACHE_INV:  buffer_wbinvl1{{$}}
-; CACHE_INV10: buffer_gl0_inv
-; CACHE_INV10: buffer_gl1_inv
-; SKIP_CACHE_INV-NOT: buffer_wbinvl1{{$}}
-; SKIP_CACHE_INV-NOT: buffer_gl
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_seq_cst() {
-entry:
-  fence syncscope("agent-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_acquire:
-; GCN:        %bb.0
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_acquire() {
-entry:
-  fence syncscope("workgroup") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_release:
-; GCN:        %bb.0
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_release() {
-entry:
-  fence syncscope("workgroup") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_acq_rel:
-; GCN:        %bb.0
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_acq_rel() {
-entry:
-  fence syncscope("workgroup") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_seq_cst() {
-entry:
-  fence syncscope("workgroup") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_acquire:
-; GCN:        %bb.0
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0)
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_acquire() {
-entry:
-  fence syncscope("workgroup-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_release:
-; GCN:        %bb.0
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0)
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_release() {
-entry:
-  fence syncscope("workgroup-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0)
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_acq_rel() {
-entry:
-  fence syncscope("workgroup-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN9-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN10:      s_waitcnt vmcnt(0)
-; GCN10:      s_waitcnt_vscnt null, 0x0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_seq_cst() {
-entry:
-  fence syncscope("workgroup-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_acquire() {
-entry:
-  fence syncscope("wavefront") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_release() {
-entry:
-  fence syncscope("wavefront") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_acq_rel() {
-entry:
-  fence syncscope("wavefront") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_seq_cst() {
-entry:
-  fence syncscope("wavefront") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_acquire() {
-entry:
-  fence syncscope("wavefront-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_release() {
-entry:
-  fence syncscope("wavefront-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_acq_rel() {
-entry:
-  fence syncscope("wavefront-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_seq_cst() {
-entry:
-  fence syncscope("wavefront-one-as") seq_cst
-  ret void
-}

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-cmpxchg.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-cmpxchg.ll
deleted file mode 100644
index f840a1537f8f..000000000000
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-cmpxchg.ll
+++ /dev/null
@@ -1,3292 +0,0 @@
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX8 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX8 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX10,GFX10WGP %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX10,GFX10CU %s
-
-; GCN-LABEL: {{^}}system_one_as_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel system_one_as_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acquire_monotonic:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_release_monotonic:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:   buffer_wbinvl1_vol
-; GFX10-NOT:  buffer_gl._inv
-; GFX10:         .amdhsa_kernel system_one_as_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acq_rel_monotonic:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst_monotonic:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acquire_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_release_acquire:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acq_rel_acquire:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst_acquire:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_acquire_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_release_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; Gfx8-NOT:  buffer_wbinvl1_vol
-; GCN-NOT:   buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_acq_rel_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_seq_cst_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_acquire_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_release_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_acq_rel_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_seq_cst_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_seq_cst_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel agent_one_as_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acquire_monotonic:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_release_monotonic:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:    buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel agent_one_as_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acq_rel_monotonic:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst_monotonic:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acquire_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_release_acquire:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acq_rel_acquire:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst_acquire:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_one_as_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acquire_monotonic:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:     s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_release_monotonic:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:       buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel workgroup_one_as_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acq_rel_monotonic:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst_monotonic:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acquire_acquire:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_release_acquire:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acq_rel_acquire:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst_acquire:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst_seq_cst:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_acquire_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_release_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_acq_rel_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_seq_cst_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_acquire_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_release_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_acq_rel_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_seq_cst_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_seq_cst_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acquire_monotonic_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acquire_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acq_rel_monotonic_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acq_rel_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acq_rel_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst_monotonic_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acquire_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acquire_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_release_acquire_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_release_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_release_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acq_rel_acquire_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acq_rel_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acq_rel_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst_acquire_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst_seq_cst_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst_seq_cst_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acquire_monotonic_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acquire_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acq_rel_monotonic_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acq_rel_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acq_rel_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst_monotonic_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acquire_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acquire_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_release_acquire_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_release_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_release_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acq_rel_acquire_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acq_rel_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acq_rel_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst_acquire_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst_seq_cst_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst_seq_cst_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acquire_monotonic_ret:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acq_rel_monotonic_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acq_rel_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst_monotonic_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acquire_acquire_ret:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_release_acquire_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_release_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_release_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acq_rel_acquire_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acq_rel_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst_acquire_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst_seq_cst_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst_seq_cst_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel system_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acquire_monotonic:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_release_monotonic:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:   buffer_wbinvl1_vol
-; GFX10-NOT:  buffer_gl._inv
-; GFX10:         .amdhsa_kernel system_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acq_rel_monotonic:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst_monotonic:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acquire_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_release_acquire:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acq_rel_acquire:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst_acquire:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst_seq_cst:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel singlethread_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_acquire_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel singlethread_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_release_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; Gfx8-NOT:  buffer_wbinvl1_vol
-; GCN-NOT:   buffer_gl{{[01]}}_inv
-; GFX10:         .amdhsa_kernel singlethread_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_acq_rel_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_seq_cst_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_acquire_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_release_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_acq_rel_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_seq_cst_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_seq_cst_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel singlethread_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel agent_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acquire_monotonic:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_release_monotonic:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:    buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel agent_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acq_rel_monotonic:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst_monotonic:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acquire_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_release_acquire:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acq_rel_acquire:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst_acquire:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst_seq_cst:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel workgroup_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acquire_monotonic:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:     s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_release_monotonic:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:       buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel workgroup_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") release monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acq_rel_monotonic:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst_monotonic:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acquire_acquire:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_release_acquire:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acq_rel_acquire:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst_acquire:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst_seq_cst:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_monotonic_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_monotonic_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_monotonic_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") monotonic monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_acquire_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_acquire_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acquire_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acquire monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_release_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_release_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_release_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") release monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_acq_rel_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_acq_rel_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acq_rel_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_seq_cst_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_seq_cst_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_seq_cst_monotonic(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_acquire_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_acquire_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acquire_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acquire acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_release_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_release_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_release_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") release acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_acq_rel_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_acq_rel_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acq_rel_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_seq_cst_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_seq_cst_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_seq_cst_acquire(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_seq_cst_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_cmpswap v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX8-NOT:  buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}._inv
-; GFX10:         .amdhsa_kernel wavefront_seq_cst_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_seq_cst_seq_cst(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acquire_monotonic_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acquire_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acq_rel_monotonic_ret:
-; GCN:        s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acq_rel_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acq_rel_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst_monotonic_ret:
-; GCN:        s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acquire_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acquire_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_release_acquire_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_release_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_release_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in release acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acq_rel_acquire_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acq_rel_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acq_rel_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst_acquire_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst_seq_cst_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst_seq_cst_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst seq_cst
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acquire_monotonic_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acquire_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acq_rel_monotonic_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acq_rel_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acq_rel_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst_monotonic_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acquire_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acquire_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_release_acquire_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_release_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_release_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") release acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acq_rel_acquire_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acq_rel_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acq_rel_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst_acquire_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst_seq_cst_ret:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst_seq_cst_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acquire_monotonic_ret:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8:          s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_acquire_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acq_rel_monotonic_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8:          s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acq_rel_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acq_rel_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst_monotonic_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8:          s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst_monotonic_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst_monotonic_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acquire_acquire_ret:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8:          s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10:         .amdhsa_kernel workgroup_acquire_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_release_acquire_ret:
-; GFX8:          s_waitcnt lgkmcnt(0){{$}}
-; GFX8:          flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10:         flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX8:          s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_release_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_release_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acq_rel_acquire_ret:
-; GFX8:          s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acq_rel_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acq_rel_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst_acquire_ret:
-; GFX8:          s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst_acquire_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst_seq_cst_ret:
-; GFX8:          s_waitcnt lgkmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_cmpswap v{{[0-9]+}}, v[{{[0-9]+\:[0-9]+}}], v[{{[0-9]+}}:{{[0-9]+}}]{{( offset:[0-9]+)*}} glc{{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP:      buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst_seq_cst_ret(
-    i32* %out, i32 %in, i32 %old) {
-entry:
-  %gep = getelementptr i32, i32* %out, i32 4
-  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
-  %val0 = extractvalue { i32, i1 } %val, 0
-  store i32 %val0, i32* %out, align 4
-  ret void
-}

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-fence.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-fence.ll
deleted file mode 100644
index cb6c019a3d7c..000000000000
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-fence.ll
+++ /dev/null
@@ -1,719 +0,0 @@
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX6,GFX68 %s
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX8,GFX68 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10WGP %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN,GFX10,GFX10CU %s
-
-; FUNC-LABEL: {{^}}system_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0){{$}}
-; GFX6-NEXT:  buffer_wbinvl1{{$}}
-; GFX8:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
-; GFX10:      s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire() {
-entry:
-  fence syncscope("one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_release() {
-entry:
-  fence syncscope("one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acq_rel() {
-entry:
-  fence syncscope("one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst() {
-entry:
-  fence syncscope("one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acquire() {
-entry:
-  fence syncscope("singlethread-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_release() {
-entry:
-  fence syncscope("singlethread-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acq_rel() {
-entry:
-  fence syncscope("singlethread-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_seq_cst() {
-entry:
-  fence syncscope("singlethread-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0){{$}}
-; GFX6-NEXT:  buffer_wbinvl1{{$}}
-; GFX8:       s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
-; GFX10:      s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire() {
-entry:
-  fence syncscope("agent-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_release() {
-entry:
-  fence syncscope("agent-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acq_rel() {
-entry:
-  fence syncscope("agent-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst() {
-entry:
-  fence syncscope("agent-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_one_as_acquire:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv{{$}}
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire() {
-entry:
-  fence syncscope("workgroup-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_one_as_release:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10-NOT:     buffer_gl0_inv
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_release() {
-entry:
-  fence syncscope("workgroup-one-as") release
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_one_as_acq_rel:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv{{$}}
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acq_rel() {
-entry:
-  fence syncscope("workgroup-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_one_as_seq_cst:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv{{$}}
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst() {
-entry:
-  fence syncscope("workgroup-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acquire() {
-entry:
-  fence syncscope("wavefront-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_release() {
-entry:
-  fence syncscope("wavefront-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acq_rel() {
-entry:
-  fence syncscope("wavefront-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_seq_cst() {
-entry:
-  fence syncscope("wavefront-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX6-NEXT:  buffer_wbinvl1{{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire() {
-entry:
-  fence acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_release() {
-entry:
-  fence release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acq_rel() {
-entry:
-  fence acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel system_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst() {
-entry:
-  fence seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acquire() {
-entry:
-  fence syncscope("singlethread") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_release() {
-entry:
-  fence syncscope("singlethread") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acq_rel() {
-entry:
-  fence syncscope("singlethread") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel singlethread_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_seq_cst() {
-entry:
-  fence syncscope("singlethread") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX6-NEXT:  buffer_wbinvl1{{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol{{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire() {
-entry:
-  fence syncscope("agent") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_release() {
-entry:
-  fence syncscope("agent") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acq_rel() {
-entry:
-  fence syncscope("agent") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GFX6:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX6:       buffer_wbinvl1{{$}}
-; GFX8:       buffer_wbinvl1_vol{{$}}
-; GFX10-NEXT: buffer_gl0_inv{{$}}
-; GFX10-NEXT: buffer_gl1_inv{{$}}
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel agent_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst() {
-entry:
-  fence syncscope("agent") seq_cst
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_acquire:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv{{$}}
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire() {
-entry:
-  fence syncscope("workgroup") acquire
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_release:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10-NOT:     buffer_gl0_inv
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_release() {
-entry:
-  fence syncscope("workgroup") release
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_acq_rel:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv{{$}}
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acq_rel() {
-entry:
-  fence syncscope("workgroup") acq_rel
-  ret void
-}
-
-; FUNC-LABEL:    {{^}}workgroup_seq_cst:
-; GCN:           %bb.0
-; GFX68-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10WGP-NEXT: buffer_gl0_inv{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   buffer_gl0_inv{{$}}
-; GCN-NOT:       ATOMIC_FENCE
-; GCN:           s_endpgm
-; GFX10:         .amdhsa_kernel workgroup_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst() {
-entry:
-  fence syncscope("workgroup") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acquire() {
-entry:
-  fence syncscope("wavefront") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_release() {
-entry:
-  fence syncscope("wavefront") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acq_rel() {
-entry:
-  fence syncscope("wavefront") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-; GFX10:         .amdhsa_kernel wavefront_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_seq_cst() {
-entry:
-  fence syncscope("wavefront") seq_cst
-  ret void
-}

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-rmw.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-rmw.ll
deleted file mode 100644
index ba49cdcde655..000000000000
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-atomic-rmw.ll
+++ /dev/null
@@ -1,1370 +0,0 @@
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX8 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX8 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX10,GFX10WGP %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX10,GFX10CU %s
-
-; GCN-LABEL: {{^}}system_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel system_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_release:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:    buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel system_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") release
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acq_rel:
-; GCN:         s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT:  s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:    flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT:  s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") release
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_acq_rel:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel agent_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_release:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:    buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel agent_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") release
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acq_rel:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel workgroup_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acquire:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_release:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:       buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel workgroup_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") release
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acq_rel:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acq_rel
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") release
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_acq_rel:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acquire
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acq_rel_ret:
-; GCN:         s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT:  s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:    flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NEXT:    s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_acq_rel_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acq_rel_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acq_rel
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") seq_cst
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acquire
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acq_rel_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_acq_rel_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acq_rel_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acq_rel
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst_ret:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") seq_cst
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acquire_ret:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acquire
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acq_rel_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_acq_rel_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acq_rel_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acq_rel
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") seq_cst
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel system_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_release:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:    buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel system_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in release
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acq_rel:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:    flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") release
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_acq_rel:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel singlethread_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel agent_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_release:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:    buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel agent_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") release
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acq_rel:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst:
-; GFX8:       s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt lgkmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel workgroup_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") monotonic
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acquire:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acquire
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_release:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:       buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel workgroup_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") release
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acq_rel:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acq_rel
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_monotonic(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") monotonic
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acquire(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") acquire
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_release(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") release
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_acq_rel:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_acq_rel
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acq_rel(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") acq_rel
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_atomic_swap v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}}{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:   s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN-NOT:   buffer_{{wbinvl1_vol|gl._inv}}
-; GFX10:         .amdhsa_kernel wavefront_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_seq_cst(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") seq_cst
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:  buffer_wbinvl1_vol
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in acquire
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acq_rel_ret:
-; GFX8:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:    flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NEXT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_acq_rel_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acq_rel_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in acq_rel
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst_ret:
-; GFX8:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NEXT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel system_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in seq_cst
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acquire_ret:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:        flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NEXT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acquire
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acq_rel_ret:
-; GFX8:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NEXT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_acq_rel_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acq_rel_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acq_rel
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst_ret:
-; GFX8:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt lgkmcnt(0){{$}}
-; GFX10:       s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NEXT:   s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX8-NEXT:   buffer_wbinvl1_vol
-; GFX10-NEXT:  s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT:  buffer_gl0_inv
-; GFX10-NEXT:  buffer_gl1_inv
-; GFX10:         .amdhsa_kernel agent_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") seq_cst
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acquire_ret:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:       s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:           flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acquire_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acquire
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acq_rel_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_acq_rel_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acq_rel_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acq_rel
-  store i32 %val, i32* %out, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst_ret:
-; GFX8-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_atomic_swap v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc{{$}}
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX8-NOT:      buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GFX10:         .amdhsa_kernel workgroup_seq_cst_ret
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst_ret(
-    i32* %out, i32 %in) {
-entry:
-  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") seq_cst
-  store i32 %val, i32* %out, align 4
-  ret void
-}

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-fence.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-fence.ll
new file mode 100644
index 000000000000..e7ca50248d48
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-fence.ll
@@ -0,0 +1,1229 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @singlethread_acquire_fence() {
+; GFX6-LABEL: singlethread_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread") acquire
+  ret void
+}
+
+define amdgpu_kernel void @singlethread_release_fence() {
+; GFX6-LABEL: singlethread_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread") release
+  ret void
+}
+
+define amdgpu_kernel void @singlethread_acq_rel_fence() {
+; GFX6-LABEL: singlethread_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @singlethread_seq_cst_fence() {
+; GFX6-LABEL: singlethread_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @singlethread_one_as_acquire_fence() {
+; GFX6-LABEL: singlethread_one_as_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_one_as_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_one_as_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_one_as_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_one_as_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @singlethread_one_as_release_fence() {
+; GFX6-LABEL: singlethread_one_as_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_one_as_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_one_as_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_one_as_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_one_as_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @singlethread_one_as_acq_rel_fence() {
+; GFX6-LABEL: singlethread_one_as_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_one_as_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_one_as_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_one_as_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_one_as_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @singlethread_one_as_seq_cst_fence() {
+; GFX6-LABEL: singlethread_one_as_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: singlethread_one_as_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: singlethread_one_as_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: singlethread_one_as_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: singlethread_one_as_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("singlethread-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_acquire_fence() {
+; GFX6-LABEL: wavefront_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront") acquire
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_release_fence() {
+; GFX6-LABEL: wavefront_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront") release
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_acq_rel_fence() {
+; GFX6-LABEL: wavefront_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_seq_cst_fence() {
+; GFX6-LABEL: wavefront_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_one_as_acquire_fence() {
+; GFX6-LABEL: wavefront_one_as_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_one_as_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_one_as_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_one_as_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_one_as_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_one_as_release_fence() {
+; GFX6-LABEL: wavefront_one_as_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_one_as_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_one_as_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_one_as_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_one_as_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_one_as_acq_rel_fence() {
+; GFX6-LABEL: wavefront_one_as_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_one_as_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_one_as_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_one_as_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_one_as_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @wavefront_one_as_seq_cst_fence() {
+; GFX6-LABEL: wavefront_one_as_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: wavefront_one_as_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: wavefront_one_as_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: wavefront_one_as_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: wavefront_one_as_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("wavefront-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_acquire_fence() {
+; GFX6-LABEL: workgroup_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup") acquire
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_release_fence() {
+; GFX6-LABEL: workgroup_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup") release
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_acq_rel_fence() {
+; GFX6-LABEL: workgroup_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_seq_cst_fence() {
+; GFX6-LABEL: workgroup_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_one_as_acquire_fence() {
+; GFX6-LABEL: workgroup_one_as_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_one_as_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_one_as_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_one_as_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_one_as_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_one_as_release_fence() {
+; GFX6-LABEL: workgroup_one_as_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_one_as_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_one_as_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_one_as_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_one_as_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_one_as_acq_rel_fence() {
+; GFX6-LABEL: workgroup_one_as_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_one_as_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_one_as_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_one_as_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_one_as_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @workgroup_one_as_seq_cst_fence() {
+; GFX6-LABEL: workgroup_one_as_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: workgroup_one_as_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: workgroup_one_as_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: workgroup_one_as_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: workgroup_one_as_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("workgroup-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @agent_acquire_fence() {
+; GFX6-LABEL: agent_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent") acquire
+  ret void
+}
+
+define amdgpu_kernel void @agent_release_fence() {
+; GFX6-LABEL: agent_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent") release
+  ret void
+}
+
+define amdgpu_kernel void @agent_acq_rel_fence() {
+; GFX6-LABEL: agent_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @agent_seq_cst_fence() {
+; GFX6-LABEL: agent_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @agent_one_as_acquire_fence() {
+; GFX6-LABEL: agent_one_as_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_one_as_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_one_as_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_one_as_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_one_as_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @agent_one_as_release_fence() {
+; GFX6-LABEL: agent_one_as_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_one_as_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_one_as_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_one_as_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_one_as_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @agent_one_as_acq_rel_fence() {
+; GFX6-LABEL: agent_one_as_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_one_as_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_one_as_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_one_as_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_one_as_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @agent_one_as_seq_cst_fence() {
+; GFX6-LABEL: agent_one_as_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: agent_one_as_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: agent_one_as_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: agent_one_as_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: agent_one_as_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("agent-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @system_acquire_fence() {
+; GFX6-LABEL: system_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence acquire
+  ret void
+}
+
+define amdgpu_kernel void @system_release_fence() {
+; GFX6-LABEL: system_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence release
+  ret void
+}
+
+define amdgpu_kernel void @system_acq_rel_fence() {
+; GFX6-LABEL: system_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @system_seq_cst_fence() {
+; GFX6-LABEL: system_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @system_one_as_acquire_fence() {
+; GFX6-LABEL: system_one_as_acquire_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_one_as_acquire_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_one_as_acquire_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_one_as_acquire_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_one_as_acquire_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @system_one_as_release_fence() {
+; GFX6-LABEL: system_one_as_release_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_one_as_release_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_one_as_release_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_one_as_release_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_one_as_release_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @system_one_as_acq_rel_fence() {
+; GFX6-LABEL: system_one_as_acq_rel_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_one_as_acq_rel_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_one_as_acq_rel_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_one_as_acq_rel_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_one_as_acq_rel_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @system_one_as_seq_cst_fence() {
+; GFX6-LABEL: system_one_as_seq_cst_fence:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: system_one_as_seq_cst_fence:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: system_one_as_seq_cst_fence:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: system_one_as_seq_cst_fence:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: system_one_as_seq_cst_fence:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+entry:
+  fence syncscope("one-as") seq_cst
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll
new file mode 100644
index 000000000000..c43e579a94b1
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll
@@ -0,0 +1,5098 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @flat_agent_unordered_load(
+; GFX7-LABEL: flat_agent_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_monotonic_load(
+; GFX7-LABEL: flat_agent_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acquire_load(
+; GFX7-LABEL: flat_agent_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_load(
+; GFX7-LABEL: flat_agent_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_unordered_store(
+; GFX7-LABEL: flat_agent_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_monotonic_store(
+; GFX7-LABEL: flat_agent_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_release_store(
+; GFX7-LABEL: flat_agent_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_store(
+; GFX7-LABEL: flat_agent_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_monotonic_atomicrmw(
+; GFX7-LABEL: flat_agent_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acquire_atomicrmw(
+; GFX7-LABEL: flat_agent_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_release_atomicrmw(
+; GFX7-LABEL: flat_agent_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_agent_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_agent_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_agent_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_agent_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_agent_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_agent_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_unordered_load(
+; GFX7-LABEL: flat_agent_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent-one-as") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_monotonic_load(
+; GFX7-LABEL: flat_agent_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent-one-as") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acquire_load(
+; GFX7-LABEL: flat_agent_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent-one-as") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_load(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("agent-one-as") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_unordered_store(
+; GFX7-LABEL: flat_agent_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_monotonic_store(
+; GFX7-LABEL: flat_agent_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_release_store(
+; GFX7-LABEL: flat_agent_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_store(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("agent-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_monotonic_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acquire_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_release_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("agent-one-as") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_agent_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll
new file mode 100644
index 000000000000..162b5a105614
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll
@@ -0,0 +1,260 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @flat_nontemporal_load_0(
+; GFX7-LABEL: flat_nontemporal_load_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc slc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_nontemporal_load_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] slc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_nontemporal_load_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] slc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_nontemporal_load_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc slc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load i32, i32* %in, align 4, !nontemporal !0
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_nontemporal_load_1(
+; GFX7-LABEL: flat_nontemporal_load_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v3, s1
+; GFX7-NEXT:    v_add_i32_e32 v2, vcc, s0, v2
+; GFX7-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GFX7-NEXT:    flat_load_dword v2, v[2:3] glc slc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_nontemporal_load_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_add_co_u32_e64 v0, s0, s0, v0
+; GFX10-WGP-NEXT:    v_add_co_ci_u32_e64 v1, s0, s1, 0, s0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] slc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_nontemporal_load_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_add_co_u32_e64 v0, s0, s0, v0
+; GFX10-CU-NEXT:    v_add_co_ci_u32_e64 v1, s0, s1, 0, s0
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] slc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_nontemporal_load_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_add_i32_e32 v2, vcc, s0, v2
+; SKIP-CACHE-INV-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v2, v[2:3] glc slc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val.gep = getelementptr inbounds i32, i32* %in, i32 %tid
+  %val = load i32, i32* %val.gep, align 4, !nontemporal !0
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_nontemporal_store_0(
+; GFX7-LABEL: flat_nontemporal_store_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0 glc slc
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_nontemporal_store_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_nontemporal_store_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_nontemporal_store_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0 glc slc
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load i32, i32* %in, align 4
+  store i32 %val, i32* %out, !nontemporal !0
+  ret void
+}
+
+define amdgpu_kernel void @flat_nontemporal_store_1(
+; GFX7-LABEL: flat_nontemporal_store_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    flat_load_dword v2, v[1:2]
+; GFX7-NEXT:    v_mov_b32_e32 v1, s3
+; GFX7-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
+; GFX7-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2 glc slc
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_nontemporal_store_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    v_add_co_u32_e64 v0, s0, s2, v0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[1:2]
+; GFX10-WGP-NEXT:    v_add_co_ci_u32_e64 v1, s0, s3, 0, s0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_nontemporal_store_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    v_add_co_u32_e64 v0, s0, s2, v0
+; GFX10-CU-NEXT:    flat_load_dword v2, v[1:2]
+; GFX10-CU-NEXT:    v_add_co_ci_u32_e64 v1, s0, s3, 0, s0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_nontemporal_store_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v2, v[1:2]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
+; SKIP-CACHE-INV-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2 glc slc
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val = load i32, i32* %in, align 4
+  %out.gep = getelementptr inbounds i32, i32* %out, i32 %tid
+  store i32 %val, i32* %out.gep, !nontemporal !0
+  ret void
+}
+
+!0 = !{i32 1}
+declare i32 @llvm.amdgcn.workitem.id.x()

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll
new file mode 100644
index 000000000000..8e6758739af6
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll
@@ -0,0 +1,4408 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @flat_singlethread_unordered_load(
+; GFX7-LABEL: flat_singlethread_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_monotonic_load(
+; GFX7-LABEL: flat_singlethread_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acquire_load(
+; GFX7-LABEL: flat_singlethread_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_load(
+; GFX7-LABEL: flat_singlethread_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_unordered_store(
+; GFX7-LABEL: flat_singlethread_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_monotonic_store(
+; GFX7-LABEL: flat_singlethread_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_release_store(
+; GFX7-LABEL: flat_singlethread_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_store(
+; GFX7-LABEL: flat_singlethread_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_monotonic_atomicrmw(
+; GFX7-LABEL: flat_singlethread_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acquire_atomicrmw(
+; GFX7-LABEL: flat_singlethread_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_release_atomicrmw(
+; GFX7-LABEL: flat_singlethread_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_singlethread_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_singlethread_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_singlethread_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_singlethread_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_singlethread_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_unordered_load(
+; GFX7-LABEL: flat_singlethread_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_monotonic_load(
+; GFX7-LABEL: flat_singlethread_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acquire_load(
+; GFX7-LABEL: flat_singlethread_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_load(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_unordered_store(
+; GFX7-LABEL: flat_singlethread_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_monotonic_store(
+; GFX7-LABEL: flat_singlethread_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_release_store(
+; GFX7-LABEL: flat_singlethread_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_store(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_monotonic_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acquire_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_release_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("singlethread-one-as") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll
new file mode 100644
index 000000000000..e80b2c9a4ced
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll
@@ -0,0 +1,5098 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @flat_system_unordered_load(
+; GFX7-LABEL: flat_system_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_monotonic_load(
+; GFX7-LABEL: flat_system_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acquire_load(
+; GFX7-LABEL: flat_system_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_load(
+; GFX7-LABEL: flat_system_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_unordered_store(
+; GFX7-LABEL: flat_system_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_monotonic_store(
+; GFX7-LABEL: flat_system_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_release_store(
+; GFX7-LABEL: flat_system_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_store(
+; GFX7-LABEL: flat_system_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_monotonic_atomicrmw(
+; GFX7-LABEL: flat_system_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acquire_atomicrmw(
+; GFX7-LABEL: flat_system_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_release_atomicrmw(
+; GFX7-LABEL: flat_system_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in release
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_system_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_system_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_system_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_system_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_system_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_system_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_system_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_unordered_load(
+; GFX7-LABEL: flat_system_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("one-as") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_monotonic_load(
+; GFX7-LABEL: flat_system_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("one-as") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acquire_load(
+; GFX7-LABEL: flat_system_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("one-as") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_load(
+; GFX7-LABEL: flat_system_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("one-as") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_unordered_store(
+; GFX7-LABEL: flat_system_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_monotonic_store(
+; GFX7-LABEL: flat_system_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_release_store(
+; GFX7-LABEL: flat_system_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_store(
+; GFX7-LABEL: flat_system_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_monotonic_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acquire_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_release_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_system_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("one-as") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_system_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll
new file mode 100644
index 000000000000..fc43c931043e
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll
@@ -0,0 +1,4408 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @flat_wavefront_unordered_load(
+; GFX7-LABEL: flat_wavefront_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_monotonic_load(
+; GFX7-LABEL: flat_wavefront_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acquire_load(
+; GFX7-LABEL: flat_wavefront_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_load(
+; GFX7-LABEL: flat_wavefront_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_unordered_store(
+; GFX7-LABEL: flat_wavefront_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_monotonic_store(
+; GFX7-LABEL: flat_wavefront_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_release_store(
+; GFX7-LABEL: flat_wavefront_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_store(
+; GFX7-LABEL: flat_wavefront_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_monotonic_atomicrmw(
+; GFX7-LABEL: flat_wavefront_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acquire_atomicrmw(
+; GFX7-LABEL: flat_wavefront_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_release_atomicrmw(
+; GFX7-LABEL: flat_wavefront_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_wavefront_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_wavefront_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_wavefront_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_wavefront_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_wavefront_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_unordered_load(
+; GFX7-LABEL: flat_wavefront_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_monotonic_load(
+; GFX7-LABEL: flat_wavefront_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acquire_load(
+; GFX7-LABEL: flat_wavefront_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_load(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_unordered_store(
+; GFX7-LABEL: flat_wavefront_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_monotonic_store(
+; GFX7-LABEL: flat_wavefront_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_release_store(
+; GFX7-LABEL: flat_wavefront_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_store(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_monotonic_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acquire_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_release_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("wavefront-one-as") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll
new file mode 100644
index 000000000000..942f574c9862
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll
@@ -0,0 +1,4725 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @flat_workgroup_unordered_load(
+; GFX7-LABEL: flat_workgroup_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_monotonic_load(
+; GFX7-LABEL: flat_workgroup_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acquire_load(
+; GFX7-LABEL: flat_workgroup_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_load(
+; GFX7-LABEL: flat_workgroup_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_unordered_store(
+; GFX7-LABEL: flat_workgroup_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_monotonic_store(
+; GFX7-LABEL: flat_workgroup_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_release_store(
+; GFX7-LABEL: flat_workgroup_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_store(
+; GFX7-LABEL: flat_workgroup_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_monotonic_atomicrmw(
+; GFX7-LABEL: flat_workgroup_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acquire_atomicrmw(
+; GFX7-LABEL: flat_workgroup_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_release_atomicrmw(
+; GFX7-LABEL: flat_workgroup_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_workgroup_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_workgroup_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_workgroup_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_workgroup_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_workgroup_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_unordered_load(
+; GFX7-LABEL: flat_workgroup_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") unordered, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_monotonic_load(
+; GFX7-LABEL: flat_workgroup_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") monotonic, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acquire_load(
+; GFX7-LABEL: flat_workgroup_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") acquire, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_load(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_load_dword v0, v[0:1]
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[2:3], v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %in, i32* %out) {
+entry:
+  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") seq_cst, align 4
+  store i32 %val, i32* %out
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_unordered_store(
+; GFX7-LABEL: flat_workgroup_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_monotonic_store(
+; GFX7-LABEL: flat_workgroup_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_release_store(
+; GFX7-LABEL: flat_workgroup_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_store(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s2, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32* %out) {
+entry:
+  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_monotonic_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acquire_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_release_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acq_rel_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acquire_ret_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acquire
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acq_rel_ret_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") acq_rel
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_ret_atomicrmw(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32* %out, i32 %in syncscope("workgroup-one-as") seq_cst
+  store i32 %val, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_monotonic_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acquire_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_release_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acq_rel_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_monotonic_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acquire_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_release_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acq_rel_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_acquire_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s0, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s1, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s2, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s3, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acquire_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_release_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @flat_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX7-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-WGP-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_add_u32 s4, s0, 16
+; GFX10-CU-NEXT:    s_addc_u32 s5, s1, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-CU-NEXT:    v_mov_b32_e32 v3, s3
+; GFX10-CU-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s2, 16
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s3, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v3, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s5
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    flat_store_dword v[0:1], v2
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32* %out, i32 4
+  %val = cmpxchg volatile i32* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll
new file mode 100644
index 000000000000..679a56b3805f
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll
@@ -0,0 +1,5547 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @global_agent_unordered_load(
+; GFX6-LABEL: global_agent_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_monotonic_load(
+; GFX6-LABEL: global_agent_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acquire_load(
+; GFX6-LABEL: global_agent_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_load(
+; GFX6-LABEL: global_agent_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_unordered_store(
+; GFX6-LABEL: global_agent_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_monotonic_store(
+; GFX6-LABEL: global_agent_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_release_store(
+; GFX6-LABEL: global_agent_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_store(
+; GFX6-LABEL: global_agent_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_monotonic_atomicrmw(
+; GFX6-LABEL: global_agent_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acquire_atomicrmw(
+; GFX6-LABEL: global_agent_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_release_atomicrmw(
+; GFX6-LABEL: global_agent_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") release
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acq_rel_atomicrmw(
+; GFX6-LABEL: global_agent_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_atomicrmw(
+; GFX6-LABEL: global_agent_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_agent_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_agent_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_agent_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_release_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_agent_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_agent_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_unordered_load(
+; GFX6-LABEL: global_agent_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_monotonic_load(
+; GFX6-LABEL: global_agent_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acquire_load(
+; GFX6-LABEL: global_agent_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_load(
+; GFX6-LABEL: global_agent_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("agent-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_unordered_store(
+; GFX6-LABEL: global_agent_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_monotonic_store(
+; GFX6-LABEL: global_agent_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_release_store(
+; GFX6-LABEL: global_agent_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_store(
+; GFX6-LABEL: global_agent_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("agent-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acquire_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_release_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("agent-one-as") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_agent_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll
new file mode 100644
index 000000000000..58d803761fa7
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll
@@ -0,0 +1,302 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @global_nontemporal_load_0(
+; GFX6-LABEL: global_nontemporal_load_0:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX6-NEXT:    s_mov_b32 s4, s2
+; GFX6-NEXT:    s_mov_b32 s5, s3
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_nontemporal_load_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s0
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_nontemporal_load_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_nontemporal_load_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_nontemporal_load_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0x0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load i32, i32 addrspace(1)* %in, align 4, !nontemporal !0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_nontemporal_load_1(
+; GFX6-LABEL: global_nontemporal_load_1:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, 0
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s6
+; GFX6-NEXT:    s_mov_b32 s1, s7
+; GFX6-NEXT:    s_mov_b32 s6, 0
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 glc slc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_nontemporal_load_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v3, s1
+; GFX7-NEXT:    v_add_i32_e32 v2, vcc, s0, v2
+; GFX7-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GFX7-NEXT:    flat_load_dword v2, v[2:3] glc slc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_nontemporal_load_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v0, v0, s[0:1] slc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v1, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_nontemporal_load_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v0, v0, s[0:1] slc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v1, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_nontemporal_load_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 glc slc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val.gep = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 %tid
+  %val = load i32, i32 addrspace(1)* %val.gep, align 4, !nontemporal !0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_nontemporal_store_0(
+; GFX6-LABEL: global_nontemporal_store_0:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX6-NEXT:    s_mov_b32 s4, s2
+; GFX6-NEXT:    s_mov_b32 s5, s3
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0 glc slc
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_nontemporal_store_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s0
+; GFX7-NEXT:    flat_store_dword v[0:1], v2 glc slc
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_nontemporal_store_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_nontemporal_store_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_nontemporal_store_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0x0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0 glc slc
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  store i32 %val, i32 addrspace(1)* %out, !nontemporal !0
+  ret void
+}
+
+define amdgpu_kernel void @global_nontemporal_store_1(
+; GFX6-LABEL: global_nontemporal_store_1:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, 0
+; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX6-NEXT:    s_mov_b64 s[4:5], s[2:3]
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64 glc slc
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_nontemporal_store_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s3
+; GFX7-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
+; GFX7-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s0
+; GFX7-NEXT:    flat_store_dword v[0:1], v2 glc slc
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_nontemporal_store_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_nontemporal_store_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_nontemporal_store_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, 0
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0x0
+; SKIP-CACHE-INV-NEXT:    s_mov_b64 s[4:5], s[2:3]
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s0
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64 glc slc
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 %tid
+  store i32 %val, i32 addrspace(1)* %out.gep, !nontemporal !0
+  ret void
+}
+
+!0 = !{i32 1}
+declare i32 @llvm.amdgcn.workitem.id.x()

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
new file mode 100644
index 000000000000..dd50b95e8691
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
@@ -0,0 +1,4837 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @global_singlethread_unordered_load(
+; GFX6-LABEL: global_singlethread_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_monotonic_load(
+; GFX6-LABEL: global_singlethread_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acquire_load(
+; GFX6-LABEL: global_singlethread_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_load(
+; GFX6-LABEL: global_singlethread_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_unordered_store(
+; GFX6-LABEL: global_singlethread_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_monotonic_store(
+; GFX6-LABEL: global_singlethread_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_release_store(
+; GFX6-LABEL: global_singlethread_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_store(
+; GFX6-LABEL: global_singlethread_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_monotonic_atomicrmw(
+; GFX6-LABEL: global_singlethread_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acquire_atomicrmw(
+; GFX6-LABEL: global_singlethread_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_release_atomicrmw(
+; GFX6-LABEL: global_singlethread_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") release
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acq_rel_atomicrmw(
+; GFX6-LABEL: global_singlethread_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_atomicrmw(
+; GFX6-LABEL: global_singlethread_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_singlethread_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_release_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_unordered_load(
+; GFX6-LABEL: global_singlethread_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_monotonic_load(
+; GFX6-LABEL: global_singlethread_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acquire_load(
+; GFX6-LABEL: global_singlethread_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_load(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("singlethread-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_unordered_store(
+; GFX6-LABEL: global_singlethread_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_monotonic_store(
+; GFX6-LABEL: global_singlethread_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_release_store(
+; GFX6-LABEL: global_singlethread_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_store(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("singlethread-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acquire_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_release_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("singlethread-one-as") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll
new file mode 100644
index 000000000000..e4b0b29e7437
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll
@@ -0,0 +1,5547 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @global_system_unordered_load(
+; GFX6-LABEL: global_system_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_monotonic_load(
+; GFX6-LABEL: global_system_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acquire_load(
+; GFX6-LABEL: global_system_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_load(
+; GFX6-LABEL: global_system_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_unordered_store(
+; GFX6-LABEL: global_system_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_monotonic_store(
+; GFX6-LABEL: global_system_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_release_store(
+; GFX6-LABEL: global_system_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_store(
+; GFX6-LABEL: global_system_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_monotonic_atomicrmw(
+; GFX6-LABEL: global_system_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acquire_atomicrmw(
+; GFX6-LABEL: global_system_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_release_atomicrmw(
+; GFX6-LABEL: global_system_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in release
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acq_rel_atomicrmw(
+; GFX6-LABEL: global_system_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_atomicrmw(
+; GFX6-LABEL: global_system_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_system_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_system_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_system_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_system_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_release_acquire_cmpxchg(
+; GFX6-LABEL: global_system_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_system_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_system_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_system_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_system_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_unordered_load(
+; GFX6-LABEL: global_system_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("one-as") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_monotonic_load(
+; GFX6-LABEL: global_system_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acquire_load(
+; GFX6-LABEL: global_system_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("one-as") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_load(
+; GFX6-LABEL: global_system_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1] glc dlc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_unordered_store(
+; GFX6-LABEL: global_system_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_monotonic_store(
+; GFX6-LABEL: global_system_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_release_store(
+; GFX6-LABEL: global_system_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_store(
+; GFX6-LABEL: global_system_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: global_system_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acquire_atomicrmw(
+; GFX6-LABEL: global_system_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_release_atomicrmw(
+; GFX6-LABEL: global_system_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: global_system_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: global_system_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_system_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_system_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_system_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("one-as") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: global_system_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_system_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll
new file mode 100644
index 000000000000..9aab2f3b6123
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll
@@ -0,0 +1,4837 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @global_wavefront_unordered_load(
+; GFX6-LABEL: global_wavefront_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_monotonic_load(
+; GFX6-LABEL: global_wavefront_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acquire_load(
+; GFX6-LABEL: global_wavefront_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_load(
+; GFX6-LABEL: global_wavefront_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_unordered_store(
+; GFX6-LABEL: global_wavefront_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_monotonic_store(
+; GFX6-LABEL: global_wavefront_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_release_store(
+; GFX6-LABEL: global_wavefront_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_store(
+; GFX6-LABEL: global_wavefront_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_monotonic_atomicrmw(
+; GFX6-LABEL: global_wavefront_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acquire_atomicrmw(
+; GFX6-LABEL: global_wavefront_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_release_atomicrmw(
+; GFX6-LABEL: global_wavefront_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") release
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acq_rel_atomicrmw(
+; GFX6-LABEL: global_wavefront_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_atomicrmw(
+; GFX6-LABEL: global_wavefront_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_wavefront_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_wavefront_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_wavefront_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_release_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_unordered_load(
+; GFX6-LABEL: global_wavefront_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_monotonic_load(
+; GFX6-LABEL: global_wavefront_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acquire_load(
+; GFX6-LABEL: global_wavefront_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_load(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("wavefront-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_unordered_store(
+; GFX6-LABEL: global_wavefront_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_monotonic_store(
+; GFX6-LABEL: global_wavefront_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_release_store(
+; GFX6-LABEL: global_wavefront_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_store(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("wavefront-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acquire_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_release_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("wavefront-one-as") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll
new file mode 100644
index 000000000000..1fd85e6acb9f
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll
@@ -0,0 +1,5135 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @global_workgroup_unordered_load(
+; GFX6-LABEL: global_workgroup_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_monotonic_load(
+; GFX6-LABEL: global_workgroup_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acquire_load(
+; GFX6-LABEL: global_workgroup_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_load(
+; GFX6-LABEL: global_workgroup_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_unordered_store(
+; GFX6-LABEL: global_workgroup_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_monotonic_store(
+; GFX6-LABEL: global_workgroup_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_release_store(
+; GFX6-LABEL: global_workgroup_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_store(
+; GFX6-LABEL: global_workgroup_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_monotonic_atomicrmw(
+; GFX6-LABEL: global_workgroup_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acquire_atomicrmw(
+; GFX6-LABEL: global_workgroup_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_release_atomicrmw(
+; GFX6-LABEL: global_workgroup_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") release
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acq_rel_atomicrmw(
+; GFX6-LABEL: global_workgroup_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_atomicrmw(
+; GFX6-LABEL: global_workgroup_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_workgroup_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_workgroup_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_workgroup_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_release_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_unordered_load(
+; GFX6-LABEL: global_workgroup_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_monotonic_load(
+; GFX6-LABEL: global_workgroup_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acquire_load(
+; GFX6-LABEL: global_workgroup_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_load(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s0, s4
+; GFX6-NEXT:    s_mov_b32 s1, s5
+; GFX6-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_mov_b32 s4, s6
+; GFX6-NEXT:    s_mov_b32 s5, s7
+; GFX6-NEXT:    s_mov_b32 s6, s2
+; GFX6-NEXT:    s_mov_b32 s7, s3
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_load_dword v0, v[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[2:3], v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_load_dword v1, v0, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s0, s4
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s1, s5
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s6
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s5, s7
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, s2
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(1)* %in syncscope("workgroup-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_unordered_store(
+; GFX6-LABEL: global_workgroup_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_monotonic_store(
+; GFX6-LABEL: global_workgroup_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_release_store(
+; GFX6-LABEL: global_workgroup_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_store(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(1)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(1)* %out syncscope("workgroup-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acquire_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_release_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") acquire
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") acq_rel
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(1)* %out, i32 %in syncscope("workgroup-one-as") seq_cst
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s0, s0, 16
+; GFX7-NEXT:    s_addc_u32 s1, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v2, v[0:1], s[2:3] offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @global_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: global_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: global_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_add_u32 s4, s0, 16
+; GFX7-NEXT:    s_addc_u32 s5, s1, 0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s4
+; GFX7-NEXT:    v_mov_b32_e32 v2, s2
+; GFX7-NEXT:    v_mov_b32_e32 v1, s5
+; GFX7-NEXT:    v_mov_b32_e32 v3, s3
+; GFX7-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: global_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: global_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[2:3] offset:16 glc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: global_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(1)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(1)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-load.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-load.ll
deleted file mode 100644
index 9832c24f42ec..000000000000
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-load.ll
+++ /dev/null
@@ -1,1019 +0,0 @@
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX8,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX8,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX9,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX9,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10,GFX10WGP %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10,GFX10CU %s
-
-declare i32 @llvm.amdgcn.workitem.id.x()
-
-; GCN-LABEL: {{^}}system_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("one-as") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("one-as") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("one-as") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX89-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("one-as") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread-one-as") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent-one-as") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent-one-as") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent-one-as") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX89-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NEXT:   s_waitcnt vmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent-one-as") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL:    {{^}}workgroup_one_as_monotonic:
-; GCN-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:        flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX10WGP:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10CU-NOT:  flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GCN-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT:    buffer_wbinvl1_vol
-; GFX10-NOT:    buffer_gl{{[01]}}_inv
-; GCN:          flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_acquire:
-; GCN-NOT:       s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:     s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:         flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX10WGP:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10CU-NOT:   flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX89-NOT:     buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GCN:           flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst:
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0
-; GFX89:         flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX10WGP:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10CU:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX89-NOT:     buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GCN:           flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup-one-as") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_one_as_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront-one-as") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_private_0:
-; GFX89: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen glc slc{{$}}
-; GFX10: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_private_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_private_0(
-    i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
-entry:
-  %val = load i32, i32 addrspace(5)* %in, align 4, !nontemporal !0
-  store i32 %val, i32 addrspace(1)* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_private_1:
-; GFX89: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen glc slc{{$}}
-; GFX10: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_private_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_private_1(
-    i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val.gep = getelementptr inbounds i32, i32 addrspace(5)* %in, i32 %tid
-  %val = load i32, i32 addrspace(5)* %val.gep, align 4, !nontemporal !0
-  store i32 %val, i32 addrspace(1)* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_global_0:
-; GCN: s_load_dword s{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0x0{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_global_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_global_0(
-    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
-entry:
-  %val = load i32, i32 addrspace(1)* %in, align 4, !nontemporal !0
-  store i32 %val, i32 addrspace(1)* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_global_1:
-; GFX8:  flat_load_dword v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}] glc slc{{$}}
-; GFX9:  global_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}] glc slc{{$}}
-; GFX10: global_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}] slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_global_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_global_1(
-    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val.gep = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 %tid
-  %val = load i32, i32 addrspace(1)* %val.gep, align 4, !nontemporal !0
-  store i32 %val, i32 addrspace(1)* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_local_0:
-; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_local_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_local_0(
-    i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
-entry:
-  %val = load i32, i32 addrspace(3)* %in, align 4, !nontemporal !0
-  store i32 %val, i32 addrspace(1)* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_local_1:
-; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_local_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_local_1(
-    i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val.gep = getelementptr inbounds i32, i32 addrspace(3)* %in, i32 %tid
-  %val = load i32, i32 addrspace(3)* %val.gep, align 4, !nontemporal !0
-  store i32 %val, i32 addrspace(1)* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_flat_0:
-; GFX89: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}] glc slc{{$}}
-; GFX10: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}] slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_flat_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_flat_0(
-    i32* %in, i32* %out) {
-entry:
-  %val = load i32, i32* %in, align 4, !nontemporal !0
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_flat_1:
-; GFX89: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}] glc slc{{$}}
-; GFX10: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+}}:{{[0-9]+}}] slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_flat_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_flat_1(
-    i32* %in, i32* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val.gep = getelementptr inbounds i32, i32* %in, i32 %tid
-  %val = load i32, i32* %val.gep, align 4, !nontemporal !0
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GFX89-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst:
-; GFX89:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GFX89-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GFX89-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel system_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel singlethread_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("singlethread") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_acquire:
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:  s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GFX89-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst:
-; GFX89:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GFX89-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10-NEXT: flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc dlc{{$}}
-; GFX89-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX89-NEXT: buffer_wbinvl1_vol
-; GFX10-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10-NEXT: buffer_gl0_inv
-; GFX10-NEXT: buffer_gl1_inv
-; GCN:        flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel agent_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("agent") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL:    {{^}}workgroup_monotonic:
-; GCN-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:        flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX10WGP:     flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10CU-NOT:  flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GCN-NOT:      s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT:    s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT:    buffer_wbinvl1_vol
-; GFX10-NOT:    buffer_gl{{[01]}}_inv
-; GCN:          flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_acquire:
-; GFX10-NOT:     s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89:         flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX10WGP:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10CU-NOT:   flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX89:         s_waitcnt lgkmcnt(0){{$}}
-; GFX89:         s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX89-NOT:     buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GCN:           flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst:
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0
-; GFX89:         flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX10WGP:      flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}] glc{{$}}
-; GFX10CU:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GFX89:         s_waitcnt lgkmcnt(0){{$}}
-; GFX89:         s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX89-NOT:     buffer_wbinvl1_vol
-; GFX10WGP-NEXT: buffer_gl0_inv
-; GFX10CU-NOT:   buffer_gl0_inv
-; GCN:           flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel workgroup_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("workgroup") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_unordered(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront") unordered, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_monotonic(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront") monotonic, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_acquire:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_acquire
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_acquire(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront") acquire, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_load_dword [[RET:v[0-9]+]], v[{{[0-9]+}}:{{[0-9]+}}]{{$}}
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GFX89-NOT: buffer_wbinvl1_vol
-; GFX10-NOT: buffer_gl{{[01]}}_inv
-; GCN:       flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RET]]
-; GFX10:         .amdhsa_kernel wavefront_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_seq_cst(
-    i32* %in, i32* %out) {
-entry:
-  %val = load atomic i32, i32* %in syncscope("wavefront") seq_cst, align 4
-  store i32 %val, i32* %out
-  ret void
-}
-
-!0 = !{i32 1}

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll
new file mode 100644
index 000000000000..49499b7570e9
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll
@@ -0,0 +1,4719 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @local_agent_unordered_load(
+; GFX6-LABEL: local_agent_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_monotonic_load(
+; GFX6-LABEL: local_agent_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acquire_load(
+; GFX6-LABEL: local_agent_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_load(
+; GFX6-LABEL: local_agent_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_unordered_store(
+; GFX6-LABEL: local_agent_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_monotonic_store(
+; GFX6-LABEL: local_agent_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_release_store(
+; GFX6-LABEL: local_agent_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_store(
+; GFX6-LABEL: local_agent_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_monotonic_atomicrmw(
+; GFX6-LABEL: local_agent_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acquire_atomicrmw(
+; GFX6-LABEL: local_agent_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_release_atomicrmw(
+; GFX6-LABEL: local_agent_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") release
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acq_rel_atomicrmw(
+; GFX6-LABEL: local_agent_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_atomicrmw(
+; GFX6-LABEL: local_agent_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_agent_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_agent_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_agent_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_release_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_agent_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_agent_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_unordered_load(
+; GFX6-LABEL: local_agent_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_monotonic_load(
+; GFX6-LABEL: local_agent_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acquire_load(
+; GFX6-LABEL: local_agent_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_load(
+; GFX6-LABEL: local_agent_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("agent-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_unordered_store(
+; GFX6-LABEL: local_agent_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_monotonic_store(
+; GFX6-LABEL: local_agent_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_release_store(
+; GFX6-LABEL: local_agent_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_store(
+; GFX6-LABEL: local_agent_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("agent-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acquire_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_release_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("agent-one-as") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_agent_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("agent-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll
new file mode 100644
index 000000000000..db3b8b2a5f22
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll
@@ -0,0 +1,313 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @local_nontemporal_load_0(
+; GFX6-LABEL: local_nontemporal_load_0:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_nontemporal_load_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s2
+; GFX7-NEXT:    ds_read_b32 v2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_nontemporal_load_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_nontemporal_load_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_nontemporal_load_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load i32, i32 addrspace(3)* %in, align 4, !nontemporal !0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_nontemporal_load_1(
+; GFX6-LABEL: local_nontemporal_load_1:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_nontemporal_load_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
+; GFX7-NEXT:    ds_read_b32 v2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_nontemporal_load_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_nontemporal_load_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_nontemporal_load_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val.gep = getelementptr inbounds i32, i32 addrspace(3)* %in, i32 %tid
+  %val = load i32, i32 addrspace(3)* %val.gep, align 4, !nontemporal !0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_nontemporal_store_0(
+; GFX6-LABEL: local_nontemporal_store_0:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_load_dword s1, s[2:3], 0x0
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_nontemporal_store_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX7-NEXT:    v_mov_b32_e32 v0, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_nontemporal_store_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_nontemporal_store_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_nontemporal_store_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[2:3], 0x0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  store i32 %val, i32 addrspace(3)* %out, !nontemporal !0
+  ret void
+}
+
+define amdgpu_kernel void @local_nontemporal_store_1(
+; GFX6-LABEL: local_nontemporal_store_1:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_load_dword s1, s[2:3], 0x0
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, s0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_nontemporal_store_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX7-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_nontemporal_store_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-WGP-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_nontemporal_store_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-CU-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_nontemporal_store_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[2:3], 0x0
+; SKIP-CACHE-INV-NEXT:    v_add_i32_e32 v0, vcc, s0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(3)* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %out.gep = getelementptr inbounds i32, i32 addrspace(3)* %out, i32 %tid
+  store i32 %val, i32 addrspace(3)* %out.gep, !nontemporal !0
+  ret void
+}
+
+!0 = !{i32 1}
+declare i32 @llvm.amdgcn.workitem.id.x()

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll
new file mode 100644
index 000000000000..cb45b58da5ac
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll
@@ -0,0 +1,4345 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @local_singlethread_unordered_load(
+; GFX6-LABEL: local_singlethread_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_monotonic_load(
+; GFX6-LABEL: local_singlethread_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acquire_load(
+; GFX6-LABEL: local_singlethread_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_load(
+; GFX6-LABEL: local_singlethread_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_unordered_store(
+; GFX6-LABEL: local_singlethread_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_monotonic_store(
+; GFX6-LABEL: local_singlethread_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_release_store(
+; GFX6-LABEL: local_singlethread_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_store(
+; GFX6-LABEL: local_singlethread_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_monotonic_atomicrmw(
+; GFX6-LABEL: local_singlethread_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acquire_atomicrmw(
+; GFX6-LABEL: local_singlethread_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_release_atomicrmw(
+; GFX6-LABEL: local_singlethread_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") release
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acq_rel_atomicrmw(
+; GFX6-LABEL: local_singlethread_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_atomicrmw(
+; GFX6-LABEL: local_singlethread_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_singlethread_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_singlethread_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_singlethread_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_release_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_unordered_load(
+; GFX6-LABEL: local_singlethread_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_monotonic_load(
+; GFX6-LABEL: local_singlethread_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acquire_load(
+; GFX6-LABEL: local_singlethread_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_load(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("singlethread-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_unordered_store(
+; GFX6-LABEL: local_singlethread_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_monotonic_store(
+; GFX6-LABEL: local_singlethread_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_release_store(
+; GFX6-LABEL: local_singlethread_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_store(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("singlethread-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acquire_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_release_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("singlethread-one-as") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("singlethread-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll
new file mode 100644
index 000000000000..7e500d4415e6
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll
@@ -0,0 +1,4719 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @local_system_unordered_load(
+; GFX6-LABEL: local_system_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_monotonic_load(
+; GFX6-LABEL: local_system_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acquire_load(
+; GFX6-LABEL: local_system_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_load(
+; GFX6-LABEL: local_system_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_unordered_store(
+; GFX6-LABEL: local_system_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_monotonic_store(
+; GFX6-LABEL: local_system_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_release_store(
+; GFX6-LABEL: local_system_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_store(
+; GFX6-LABEL: local_system_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_monotonic_atomicrmw(
+; GFX6-LABEL: local_system_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acquire_atomicrmw(
+; GFX6-LABEL: local_system_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_release_atomicrmw(
+; GFX6-LABEL: local_system_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in release
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acq_rel_atomicrmw(
+; GFX6-LABEL: local_system_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_atomicrmw(
+; GFX6-LABEL: local_system_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_system_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_system_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_system_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_system_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_release_acquire_cmpxchg(
+; GFX6-LABEL: local_system_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_system_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_system_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_system_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_system_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    buffer_wbinvl1
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    buffer_wbinvl1_vol
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    buffer_gl1_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-CU-NEXT:    buffer_gl0_inv
+; GFX10-CU-NEXT:    buffer_gl1_inv
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_unordered_load(
+; GFX6-LABEL: local_system_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("one-as") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_monotonic_load(
+; GFX6-LABEL: local_system_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acquire_load(
+; GFX6-LABEL: local_system_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("one-as") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_load(
+; GFX6-LABEL: local_system_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_unordered_store(
+; GFX6-LABEL: local_system_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_monotonic_store(
+; GFX6-LABEL: local_system_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_release_store(
+; GFX6-LABEL: local_system_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_store(
+; GFX6-LABEL: local_system_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: local_system_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acquire_atomicrmw(
+; GFX6-LABEL: local_system_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_release_atomicrmw(
+; GFX6-LABEL: local_system_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: local_system_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: local_system_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_system_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_system_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_system_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("one-as") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: local_system_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_system_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll
new file mode 100644
index 000000000000..1bc5c14992e4
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll
@@ -0,0 +1,4345 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @local_wavefront_unordered_load(
+; GFX6-LABEL: local_wavefront_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_monotonic_load(
+; GFX6-LABEL: local_wavefront_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acquire_load(
+; GFX6-LABEL: local_wavefront_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_load(
+; GFX6-LABEL: local_wavefront_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_unordered_store(
+; GFX6-LABEL: local_wavefront_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_monotonic_store(
+; GFX6-LABEL: local_wavefront_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_release_store(
+; GFX6-LABEL: local_wavefront_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_store(
+; GFX6-LABEL: local_wavefront_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_monotonic_atomicrmw(
+; GFX6-LABEL: local_wavefront_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acquire_atomicrmw(
+; GFX6-LABEL: local_wavefront_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_release_atomicrmw(
+; GFX6-LABEL: local_wavefront_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") release
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acq_rel_atomicrmw(
+; GFX6-LABEL: local_wavefront_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_atomicrmw(
+; GFX6-LABEL: local_wavefront_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_wavefront_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_wavefront_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_wavefront_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_release_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_unordered_load(
+; GFX6-LABEL: local_wavefront_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_monotonic_load(
+; GFX6-LABEL: local_wavefront_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acquire_load(
+; GFX6-LABEL: local_wavefront_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_load(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("wavefront-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_unordered_store(
+; GFX6-LABEL: local_wavefront_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_monotonic_store(
+; GFX6-LABEL: local_wavefront_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_release_store(
+; GFX6-LABEL: local_wavefront_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_store(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("wavefront-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acquire_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_release_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("wavefront-one-as") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("wavefront-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll
new file mode 100644
index 000000000000..1823f3ad8882
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll
@@ -0,0 +1,4564 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @local_workgroup_unordered_load(
+; GFX6-LABEL: local_workgroup_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_monotonic_load(
+; GFX6-LABEL: local_workgroup_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acquire_load(
+; GFX6-LABEL: local_workgroup_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_load(
+; GFX6-LABEL: local_workgroup_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_unordered_store(
+; GFX6-LABEL: local_workgroup_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_monotonic_store(
+; GFX6-LABEL: local_workgroup_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_release_store(
+; GFX6-LABEL: local_workgroup_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_store(
+; GFX6-LABEL: local_workgroup_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_monotonic_atomicrmw(
+; GFX6-LABEL: local_workgroup_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acquire_atomicrmw(
+; GFX6-LABEL: local_workgroup_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_release_atomicrmw(
+; GFX6-LABEL: local_workgroup_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") release
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acq_rel_atomicrmw(
+; GFX6-LABEL: local_workgroup_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_atomicrmw(
+; GFX6-LABEL: local_workgroup_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_workgroup_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_workgroup_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_workgroup_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_release_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-WGP-NEXT:    buffer_gl0_inv
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_unordered_load(
+; GFX6-LABEL: local_workgroup_one_as_unordered_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_unordered_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_unordered_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_unordered_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_unordered_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup-one-as") unordered, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_monotonic_load(
+; GFX6-LABEL: local_workgroup_one_as_monotonic_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_monotonic_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_monotonic_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_monotonic_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_monotonic_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup-one-as") monotonic, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acquire_load(
+; GFX6-LABEL: local_workgroup_one_as_acquire_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acquire_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acquire_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acquire_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acquire_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup-one-as") acquire, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_load(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_load:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    ds_read_b32 v0, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v1, v0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_load:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    ds_read_b32 v0, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v1, v0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_load:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_read_b32 v0, v0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v1, v0
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_load:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_read_b32 v0, v0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v1, v0
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_load:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    ds_read_b32 v0, v0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v1, v0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %in, i32 addrspace(3)* %out) {
+entry:
+  %val = load atomic i32, i32 addrspace(3)* %in syncscope("workgroup-one-as") seq_cst, align 4
+  store i32 %val, i32 addrspace(3)* %out
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_unordered_store(
+; GFX6-LABEL: local_workgroup_one_as_unordered_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_unordered_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_unordered_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_unordered_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_unordered_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup-one-as") unordered, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_monotonic_store(
+; GFX6-LABEL: local_workgroup_one_as_monotonic_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_monotonic_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_monotonic_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_monotonic_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_monotonic_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup-one-as") monotonic, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_release_store(
+; GFX6-LABEL: local_workgroup_one_as_release_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_release_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_release_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_release_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_release_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup-one-as") release, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_store(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_store:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s2
+; GFX6-NEXT:    v_mov_b32_e32 v0, s0
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_store:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s1
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_store:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_store:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s1
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_store:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 %in, i32 addrspace(3)* %out) {
+entry:
+  store atomic i32 %in, i32 addrspace(3)* %out syncscope("workgroup-one-as") seq_cst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_monotonic_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_monotonic_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_monotonic_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_monotonic_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_monotonic_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_monotonic_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acquire_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_acquire_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acquire_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acquire_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acquire_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acquire_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_release_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_release_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_release_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_release_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_release_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_release_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") release
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acq_rel_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_acq_rel_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acq_rel_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acq_rel_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acq_rel_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acq_rel_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") acq_rel
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v0, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acquire_ret_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acquire_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acquire_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") acquire
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acq_rel_ret_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acq_rel_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acq_rel_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") acq_rel
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_ret_atomicrmw(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-WGP-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_ret_atomicrmw:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX10-CU-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_ret_atomicrmw:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    ds_wrxchg_rtn_b32 v1, v0, v1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in) {
+entry:
+  %val = atomicrmw volatile xchg i32 addrspace(3)* %out, i32 %in syncscope("workgroup-one-as") seq_cst
+  store i32 %val, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_monotonic_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_monotonic_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") monotonic monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acquire_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acquire_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acquire_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_release_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_release_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_release_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acq_rel_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acq_rel_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_monotonic_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_monotonic_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acquire_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acquire_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acquire_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_release_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_release_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_release_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_release_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_release_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_release_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acq_rel_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acq_rel_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_acquire_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_acquire_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_seq_cst_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_seq_cst_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_b32 v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acquire_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst monotonic
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acquire_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acquire_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acquire acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_release_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_release_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_release_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") release acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_acq_rel_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") acq_rel acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_acquire_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst acquire
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+
+define amdgpu_kernel void @local_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg(
+; GFX6-LABEL: local_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_load_dword s2, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xa
+; GFX6-NEXT:    s_mov_b32 m0, -1
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s2
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    v_mov_b32_e32 v2, s0
+; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    ds_write_b32 v0, v1
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: local_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX7-NEXT:    s_mov_b32 m0, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    v_mov_b32_e32 v2, s1
+; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    ds_write_b32 v0, v1
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: local_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-WGP-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    ds_write_b32 v0, v1
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: local_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
+; GFX10-CU-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    ds_write_b32 v0, v1
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: local_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 m0, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s2
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v2, s3
+; SKIP-CACHE-INV-NEXT:    ds_cmpst_rtn_b32 v1, v0, v1, v2 offset:16
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    ds_write_b32 v0, v1
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(3)* %out, i32 %in, i32 %old) {
+entry:
+  %gep = getelementptr i32, i32 addrspace(3)* %out, i32 4
+  %val = cmpxchg volatile i32 addrspace(3)* %gep, i32 %old, i32 %in syncscope("workgroup-one-as") seq_cst seq_cst
+  %val0 = extractvalue { i32, i1 } %val, 0
+  store i32 %val0, i32 addrspace(3)* %out, align 4
+  ret void
+}
+

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-mesa3d.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-mesa3d.ll
deleted file mode 100644
index fc598a66849a..000000000000
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-mesa3d.ll
+++ /dev/null
@@ -1,440 +0,0 @@
-; RUN: llc -mtriple=amdgcn--mesa3d -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN %s
-; RUN: llc -mtriple=amdgcn--mesa3d -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN %s
-; RUN: llc -mtriple=amdgcn--mesa3d -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN %s
-; RUN: llc -mtriple=amdgcn--mesa3d -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=FUNC,GCN %s
-
-; FUNC-LABEL: {{^}}system_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN-NEXT:   buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_acquire() {
-entry:
-  fence syncscope("one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_release() {
-entry:
-  fence syncscope("one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_acq_rel() {
-entry:
-  fence syncscope("one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_one_as_seq_cst() {
-entry:
-  fence syncscope("one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_acquire() {
-entry:
-  fence syncscope("singlethread-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_release() {
-entry:
-  fence syncscope("singlethread-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_acq_rel() {
-entry:
-  fence syncscope("singlethread-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_one_as_seq_cst() {
-entry:
-  fence syncscope("singlethread-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN-NEXT:   buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_acquire() {
-entry:
-  fence syncscope("agent-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_release() {
-entry:
-  fence syncscope("agent-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_acq_rel() {
-entry:
-  fence syncscope("agent-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_one_as_seq_cst() {
-entry:
-  fence syncscope("agent-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_acquire() {
-entry:
-  fence syncscope("workgroup-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_release() {
-entry:
-  fence syncscope("workgroup-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_acq_rel() {
-entry:
-  fence syncscope("workgroup-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_one_as_seq_cst() {
-entry:
-  fence syncscope("workgroup-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_acquire() {
-entry:
-  fence syncscope("wavefront-one-as") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_release() {
-entry:
-  fence syncscope("wavefront-one-as") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_acq_rel() {
-entry:
-  fence syncscope("wavefront-one-as") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_one_as_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_one_as_seq_cst() {
-entry:
-  fence syncscope("wavefront-one-as") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN-NEXT:   buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_acquire() {
-entry:
-  fence acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_release() {
-entry:
-  fence release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_acq_rel() {
-entry:
-  fence acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}system_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @system_seq_cst() {
-entry:
-  fence seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_acquire() {
-entry:
-  fence syncscope("singlethread") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_release() {
-entry:
-  fence syncscope("singlethread") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_acq_rel() {
-entry:
-  fence syncscope("singlethread") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}singlethread_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @singlethread_seq_cst() {
-entry:
-  fence syncscope("singlethread") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN-NEXT:   buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_acquire() {
-entry:
-  fence syncscope("agent") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_release() {
-entry:
-  fence syncscope("agent") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_acq_rel() {
-entry:
-  fence syncscope("agent") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}agent_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN:        buffer_wbinvl1{{$}}
-; GCN:        s_endpgm
-define amdgpu_kernel void @agent_seq_cst() {
-entry:
-  fence syncscope("agent") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_acquire() {
-entry:
-  fence syncscope("workgroup") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_release:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_release() {
-entry:
-  fence syncscope("workgroup") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_acq_rel() {
-entry:
-  fence syncscope("workgroup") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}workgroup_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @workgroup_seq_cst() {
-entry:
-  fence syncscope("workgroup") seq_cst
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_acquire:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_acquire() {
-entry:
-  fence syncscope("wavefront") acquire
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_release:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_release() {
-entry:
-  fence syncscope("wavefront") release
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_acq_rel:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_acq_rel() {
-entry:
-  fence syncscope("wavefront") acq_rel
-  ret void
-}
-
-; FUNC-LABEL: {{^}}wavefront_seq_cst:
-; GCN:        %bb.0
-; GCN-NOT:    ATOMIC_FENCE
-; GCN:        s_endpgm
-define amdgpu_kernel void @wavefront_seq_cst() {
-entry:
-  fence syncscope("wavefront") seq_cst
-  ret void
-}

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll
new file mode 100644
index 000000000000..03977d855005
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll
@@ -0,0 +1,395 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx600 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX6 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX7 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-WGP %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx700 -amdgcn-skip-cache-invalidations -verify-machineinstrs < %s | FileCheck --check-prefixes=SKIP-CACHE-INV %s
+
+define amdgpu_kernel void @private_nontemporal_load_0(
+; GFX6-LABEL: private_nontemporal_load_0:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GFX6-NEXT:    s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s10, -1
+; GFX6-NEXT:    s_mov_b32 s11, 0xe8f000
+; GFX6-NEXT:    s_add_u32 s8, s8, s3
+; GFX6-NEXT:    s_addc_u32 s9, s9, 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s4
+; GFX6-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen glc slc
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: private_nontemporal_load_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_add_u32 s8, s8, s7
+; GFX7-NEXT:    s_addc_u32 s9, s9, 0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s2
+; GFX7-NEXT:    buffer_load_dword v2, v0, s[8:11], 0 offen glc slc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: private_nontemporal_load_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-WGP-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-WGP-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-WGP-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen slc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: private_nontemporal_load_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-CU-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-CU-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-CU-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen slc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: private_nontemporal_load_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_getpc_b64 s[8:9]
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s8, s0
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s8, s8, s3
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s9, s9, 0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s4
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen glc slc
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
+entry:
+  %val = load i32, i32 addrspace(5)* %in, align 4, !nontemporal !0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @private_nontemporal_load_1(
+; GFX6-LABEL: private_nontemporal_load_1:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GFX6-NEXT:    s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GFX6-NEXT:    s_load_dword s4, s[0:1], 0x9
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; GFX6-NEXT:    s_mov_b32 s10, -1
+; GFX6-NEXT:    s_mov_b32 s11, 0xe8f000
+; GFX6-NEXT:    s_add_u32 s8, s8, s3
+; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX6-NEXT:    s_addc_u32 s9, s9, 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
+; GFX6-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen glc slc
+; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: private_nontemporal_load_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2
+; GFX7-NEXT:    s_add_u32 s8, s8, s7
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_addc_u32 s9, s9, 0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
+; GFX7-NEXT:    buffer_load_dword v2, v0, s[8:11], 0 offen glc slc
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    flat_store_dword v[0:1], v2
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: private_nontemporal_load_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-WGP-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-WGP-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-WGP-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen slc
+; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-WGP-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: private_nontemporal_load_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-CU-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x8
+; GFX10-CU-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-CU-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-CU-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen slc
+; GFX10-CU-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-CU-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: private_nontemporal_load_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_getpc_b64 s[8:9]
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s8, s0
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x0
+; SKIP-CACHE-INV-NEXT:    s_load_dword s4, s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s2, -1
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s8, s8, s3
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s9, s9, 0
+; SKIP-CACHE-INV-NEXT:    v_add_i32_e32 v0, vcc, s4, v0
+; SKIP-CACHE-INV-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen glc slc
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s3, 0xf000
+; SKIP-CACHE-INV-NEXT:    s_waitcnt vmcnt(0)
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val.gep = getelementptr inbounds i32, i32 addrspace(5)* %in, i32 %tid
+  %val = load i32, i32 addrspace(5)* %val.gep, align 4, !nontemporal !0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+define amdgpu_kernel void @private_nontemporal_store_0(
+; GFX6-LABEL: private_nontemporal_store_0:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_mov_b32 s4, SCRATCH_RSRC_DWORD0
+; GFX6-NEXT:    s_mov_b32 s5, SCRATCH_RSRC_DWORD1
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_mov_b32 s7, 0xe8f000
+; GFX6-NEXT:    s_add_u32 s4, s4, s3
+; GFX6-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    s_addc_u32 s5, s5, 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_load_dword s1, s[2:3], 0x0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v0, s1
+; GFX6-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen glc slc
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: private_nontemporal_store_0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_add_u32 s8, s8, s7
+; GFX7-NEXT:    s_addc_u32 s9, s9, 0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX7-NEXT:    v_mov_b32_e32 v1, s2
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v0, s0
+; GFX7-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen glc slc
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: private_nontemporal_store_0:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-WGP-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-WGP-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-WGP-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen slc
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: private_nontemporal_store_0:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-CU-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-CU-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
+; GFX10-CU-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen slc
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: private_nontemporal_store_0:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_getpc_b64 s[4:5]
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s0
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s4, s3
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s5, 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[2:3], 0x0
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s1
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen glc slc
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
+entry:
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  store i32 %val, i32 addrspace(5)* %out, !nontemporal !0
+  ret void
+}
+
+define amdgpu_kernel void @private_nontemporal_store_1(
+; GFX6-LABEL: private_nontemporal_store_1:
+; GFX6:       ; %bb.0: ; %entry
+; GFX6-NEXT:    s_mov_b32 s4, SCRATCH_RSRC_DWORD0
+; GFX6-NEXT:    s_mov_b32 s5, SCRATCH_RSRC_DWORD1
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_mov_b32 s7, 0xe8f000
+; GFX6-NEXT:    s_add_u32 s4, s4, s3
+; GFX6-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; GFX6-NEXT:    s_load_dword s0, s[0:1], 0xb
+; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX6-NEXT:    s_addc_u32 s5, s5, 0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_load_dword s1, s[2:3], 0x0
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, s0, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v1, s1
+; GFX6-NEXT:    buffer_store_dword v1, v0, s[4:7], 0 offen glc slc
+; GFX6-NEXT:    s_endpgm
+;
+; GFX7-LABEL: private_nontemporal_store_1:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX7-NEXT:    s_load_dword s2, s[4:5], 0x2
+; GFX7-NEXT:    s_add_u32 s8, s8, s7
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_addc_u32 s9, s9, 0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX7-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s0
+; GFX7-NEXT:    buffer_store_dword v1, v0, s[8:11], 0 offen glc slc
+; GFX7-NEXT:    s_endpgm
+;
+; GFX10-WGP-LABEL: private_nontemporal_store_1:
+; GFX10-WGP:       ; %bb.0: ; %entry
+; GFX10-WGP-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-WGP-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-WGP-NEXT:    s_clause 0x1
+; GFX10-WGP-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-WGP-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-WGP-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-WGP-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-WGP-NEXT:    buffer_store_dword v1, v0, s[8:11], 0 offen slc
+; GFX10-WGP-NEXT:    s_endpgm
+;
+; GFX10-CU-LABEL: private_nontemporal_store_1:
+; GFX10-CU:       ; %bb.0: ; %entry
+; GFX10-CU-NEXT:    s_mov_b64 s[10:11], s[2:3]
+; GFX10-CU-NEXT:    s_mov_b64 s[8:9], s[0:1]
+; GFX10-CU-NEXT:    s_clause 0x1
+; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
+; GFX10-CU-NEXT:    s_add_u32 s8, s8, s7
+; GFX10-CU-NEXT:    s_addc_u32 s9, s9, 0
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
+; GFX10-CU-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
+; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-CU-NEXT:    buffer_store_dword v1, v0, s[8:11], 0 offen slc
+; GFX10-CU-NEXT:    s_endpgm
+;
+; SKIP-CACHE-INV-LABEL: private_nontemporal_store_1:
+; SKIP-CACHE-INV:       ; %bb.0: ; %entry
+; SKIP-CACHE-INV-NEXT:    s_getpc_b64 s[4:5]
+; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s0
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
+; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_add_u32 s4, s4, s3
+; SKIP-CACHE-INV-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x9
+; SKIP-CACHE-INV-NEXT:    s_load_dword s0, s[0:1], 0xb
+; SKIP-CACHE-INV-NEXT:    s_addc_u32 s5, s5, 0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    s_load_dword s1, s[2:3], 0x0
+; SKIP-CACHE-INV-NEXT:    v_add_i32_e32 v0, vcc, s0, v0
+; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
+; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1
+; SKIP-CACHE-INV-NEXT:    buffer_store_dword v1, v0, s[4:7], 0 offen glc slc
+; SKIP-CACHE-INV-NEXT:    s_endpgm
+    i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
+entry:
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %val = load i32, i32 addrspace(1)* %in, align 4
+  %out.gep = getelementptr inbounds i32, i32 addrspace(5)* %out, i32 %tid
+  store i32 %val, i32 addrspace(5)* %out.gep, !nontemporal !0
+  ret void
+}
+
+!0 = !{i32 1}
+declare i32 @llvm.amdgcn.workitem.id.x()

diff  --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-store.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-store.ll
deleted file mode 100644
index 3ab06280d7e5..000000000000
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-store.ll
+++ /dev/null
@@ -1,754 +0,0 @@
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX8,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX8,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd- -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX9,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX9,GFX89 %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10,GFX10WGP %s
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10,GFX10CU %s
-
-declare i32 @llvm.amdgcn.workitem.id.x()
-
-; GCN-LABEL: {{^}}system_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("one-as") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("one-as") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_release:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("one-as") release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_one_as_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_one_as_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("one-as") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_one_as_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_one_as_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread-one-as") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent-one-as") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent-one-as") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_release:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent-one-as") release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_one_as_seq_cst:
-; GCN:        s_waitcnt vmcnt(0){{$}}
-; GFX10-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_one_as_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent-one-as") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_release:
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") release, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_one_as_seq_cst:
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_one_as_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup-one-as") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_one_as_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_one_as_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_one_as_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront-one-as") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_private_0:
-; GFX89: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen glc slc{{$}}
-; GFX10: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_private_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_private_0(
-    i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
-entry:
-  %val = load i32, i32 addrspace(1)* %in, align 4
-  store i32 %val, i32 addrspace(5)* %out, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_private_1:
-; GFX89: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen glc slc{{$}}
-; GFX10: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], 0 offen slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_private_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_private_1(
-    i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val = load i32, i32 addrspace(1)* %in, align 4
-  %out.gep = getelementptr inbounds i32, i32 addrspace(5)* %out, i32 %tid
-  store i32 %val, i32 addrspace(5)* %out.gep, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_global_0:
-; GFX8:  flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc slc{{$}}
-; GFX9:  global_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}} glc slc{{$}}
-; GFX10: global_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}} slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_global_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_global_0(
-    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
-entry:
-  %val = load i32, i32 addrspace(1)* %in, align 4
-  store i32 %val, i32 addrspace(1)* %out, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_global_1:
-; GFX8:  flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc slc{{$}}
-; GFX9:  global_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}] glc slc{{$}}
-; GFX10: global_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}] slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_global_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_global_1(
-    i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val = load i32, i32 addrspace(1)* %in, align 4
-  %out.gep = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 %tid
-  store i32 %val, i32 addrspace(1)* %out.gep, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_local_0:
-; GCN: ds_write_b32 v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_local_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_local_0(
-    i32 addrspace(1)* %in, i32 addrspace(3)* %out) {
-entry:
-  %val = load i32, i32 addrspace(1)* %in, align 4
-  store i32 %val, i32 addrspace(3)* %out, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_local_1:
-; GCN: ds_write_b32 v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_local_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_local_1(
-    i32 addrspace(1)* %in, i32 addrspace(3)* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val = load i32, i32 addrspace(1)* %in, align 4
-  %out.gep = getelementptr inbounds i32, i32 addrspace(3)* %out, i32 %tid
-  store i32 %val, i32 addrspace(3)* %out.gep, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_flat_0:
-; GFX89: flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc slc{{$}}
-; GFX10: flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_flat_0
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_flat_0(
-    i32* %in, i32* %out) {
-entry:
-  %val = load i32, i32* %in, align 4
-  store i32 %val, i32* %out, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}nontemporal_flat_1:
-; GFX89: flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} glc slc{{$}}
-; GFX10: flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} slc{{$}}
-; GFX10:         .amdhsa_kernel nontemporal_flat_1
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @nontemporal_flat_1(
-    i32* %in, i32* %out) {
-entry:
-  %tid = call i32 @llvm.amdgcn.workitem.id.x()
-  %val = load i32, i32* %in, align 4
-  %out.gep = getelementptr inbounds i32, i32* %out, i32 %tid
-  store i32 %val, i32* %out.gep, !nontemporal !0
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_release:
-; GFX89:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}system_seq_cst:
-; GFX89:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel system_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @system_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread") release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}singlethread_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel singlethread_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @singlethread_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("singlethread") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_release:
-; GFX89:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent") release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}agent_seq_cst:
-; GFX89:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt lgkmcnt(0){{$}}
-; GFX10:      s_waitcnt_vscnt null, 0x0{{$}}
-; GCN-NEXT:   flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel agent_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @agent_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("agent") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}workgroup_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_release:
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup") release, align 4
-  ret void
-}
-
-; GCN-LABEL:     {{^}}workgroup_seq_cst:
-; GFX89-NOT:     s_waitcnt vmcnt(0){{$}}
-; GFX10WGP:      s_waitcnt vmcnt(0) lgkmcnt(0){{$}}
-; GFX10WGP-NEXT: s_waitcnt_vscnt null, 0x0{{$}}
-; GFX10CU-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10CU-NOT:   s_waitcnt_vscnt null, 0x0{{$}}
-; GCN:           flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel workgroup_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @workgroup_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("workgroup") seq_cst, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_unordered:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_unordered
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_unordered(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront") unordered, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_monotonic:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_monotonic
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_monotonic(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront") monotonic, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_release:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_release
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_release(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront") release, align 4
-  ret void
-}
-
-; GCN-LABEL: {{^}}wavefront_seq_cst:
-; GCN-NOT:   s_waitcnt vmcnt(0){{$}}
-; GFX10-NOT: s_waitcnt_v{{[ms]}}cnt {{[^,]+, (0x)*0$}}
-; GCN:       flat_store_dword v[{{[0-9]+}}:{{[0-9]+}}], {{v[0-9]+}}{{$}}
-; GFX10:         .amdhsa_kernel wavefront_seq_cst
-; GFX10WGP-NOT:  .amdhsa_workgroup_processor_mode 0
-; GFX10CU:       .amdhsa_workgroup_processor_mode 0
-; GFX10-NOT:     .amdhsa_memory_ordered 0
-define amdgpu_kernel void @wavefront_seq_cst(
-    i32 %in, i32* %out) {
-entry:
-  store atomic i32 %in, i32* %out syncscope("wavefront") seq_cst, align 4
-  ret void
-}
-
-!0 = !{i32 1}


        
