[llvm] [Attributor] Skip AS specialization for volatile memory instructions (PR #107250)

[Matt Arsenault via llvm-commits]

================
@@ -38,9 +38,15 @@ define amdgpu_kernel void @kern_indirect_use_queue_ptr(i32) #1 {
 }
 
 ; GCN-LABEL: {{^}}use_queue_ptr_addrspacecast:
-; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0
-; GCN-DAG: ds_write_b32 v[[LO]], v[[LO]] offset:16
+; CIVI: s_load_dword [[APERTURE_LOAD:s[0-9]+]], s[4:5], 0x0
+ ; CIVI: v_mov_b32_e32 v[[LO:[0-9]+]], 16
+ ; CIVI-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], [[APERTURE_LOAD]]
 
+ ; GFX9: s_mov_b64 s[{{[0-9]+}}:[[HI:[0-9]+]]], src_shared_base
+ ; GFX9-DAG: v_mov_b32_e32 v[[VGPR_HI:[0-9]+]], s[[HI]]
+ ; GFX9: {{flat|global}}_store_dword v{{\[[0-9]+}}:[[VGPR_HI]]]
+
+ ; CIVI: {{flat|global}}_store_dword v[[[LO]]:[[HI]]]
----------------
arsenm wrote:

This is un-breaking this test 

https://github.com/llvm/llvm-project/pull/107250