[llvm] [AMDGPU] Fix folding of v2i16/v2f16 splat imms (PR #72709)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -41,9 +41,11 @@ entry:
   ret void
 }
 
+; FIXME: GFX940 fails to fold 0x1001 with op_sel
+
 ; GCN-LABEL: {{^}}test_llvm_amdgcn_udot2_op_sel:
 ; GFX906: v_dot2_u32_u16 v{{[0-9]+}}, 1, v{{[0-9]+}}, s{{[0-9]+}} op_sel:[0,1,0] op_sel_hi:[0,0,1]{{$}}
-; GFX940: v_dot2_u32_u16 v{{[0-9]+}}, 1, v{{[0-9]+}}, s{{[0-9]+}}{{$}}
----------------
rampitec wrote:

And this is still a bug even if bf16 needs a different handling.

https://github.com/llvm/llvm-project/pull/72709