[clang] [llvm] [clang-tools-extra] [AMDGPU][GFX12] Add 16 bit atomic fadd instructions (PR #75917)

[Mariusz Sikora via cfe-commits]

================
@@ -362,24 +358,34 @@ define amdgpu_ps void @struct_buffer_atomic_add_v2f16_noret(<2 x half> %val, <4
   ret void
 }
 
-define amdgpu_ps float @struct_buffer_atomic_add_v2bf16_ret(<2 x i16> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+define amdgpu_ps float @struct_buffer_atomic_add_v2bf16_ret(<2 x bfloat> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
 ; GFX12-LABEL: struct_buffer_atomic_add_v2bf16_ret:
 ; GFX12:       ; %bb.0:
 ; GFX12-NEXT:    buffer_atomic_pk_add_bf16 v0, v[1:2], s[0:3], s4 idxen offen th:TH_ATOMIC_RETURN
+; GFX12-NEXT:    v_mov_b32_e32 v1, 0
+; GFX12-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX12-NEXT:    s_waitcnt vmcnt(0)
+; GFX12-NEXT:    flat_store_b32 v[1:2], v0
+; GFX12-NEXT:    v_mov_b32_e32 v0, 1.0
+; GFX12-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX12-NEXT:    ; return to shader part epilog
 ;
 ; GFX12-GISEL-LABEL: struct_buffer_atomic_add_v2bf16_ret:
 ; GFX12-GISEL:       ; %bb.0:
 ; GFX12-GISEL-NEXT:    buffer_atomic_pk_add_bf16 v0, v[1:2], s[0:3], s4 idxen offen th:TH_ATOMIC_RETURN
+; GFX12-GISEL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX12-GISEL-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX12-GISEL-NEXT:    s_waitcnt vmcnt(0)
+; GFX12-GISEL-NEXT:    flat_store_b32 v[1:2], v0
+; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, 1.0
+; GFX12-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX12-GISEL-NEXT:    ; return to shader part epilog
-  %orig = call <2 x i16> @llvm.amdgcn.struct.buffer.atomic.fadd.v2bf16(<2 x i16> %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0)
-  %r = bitcast <2 x i16> %orig to float
----------------
mariusz-sikora-at-amd wrote:

Found issue in GlobalISel and bitcast with bfloat type. I prepare fix and push in different change.

https://github.com/llvm/llvm-project/pull/75917