[llvm] [AMDGPU] Support image atomic no return instructions (PR #150742)

[Harrison Hao via llvm-commits]

================
@@ -92,8 +92,7 @@ define amdgpu_ps void @atomic_swap_1d_agpr_noret(<8 x i32> inreg %rsrc, i32 %s)
 ; GFX90A-NEXT:    ;;#ASMSTART
 ; GFX90A-NEXT:    ; def a0
 ; GFX90A-NEXT:    ;;#ASMEND
-; GFX90A-NEXT:    v_accvgpr_read_b32 v1, a0
-; GFX90A-NEXT:    image_atomic_swap v1, v0, s[0:7] dmask:0x1 unorm glc
+; GFX90A-NEXT:    image_atomic_swap a0, v0, s[0:7] dmask:0x1 unorm
----------------
harrisonGPU wrote:

I’m still thinking about why it doesn’t remove v_accvgpr_read_b32.

https://github.com/llvm/llvm-project/pull/150742