[llvm] AMDGPU: Legalize atomicrmw fadd for v2f16/v2bf16 for local memory (PR #95393)

[Matt Arsenault via llvm-commits]

================
@@ -301,6 +301,9 @@ static const LLT V10S16 = LLT::fixed_vector(10, 16);
 static const LLT V12S16 = LLT::fixed_vector(12, 16);
 static const LLT V16S16 = LLT::fixed_vector(16, 16);
 
+static const LLT V2F16 = LLT::fixed_vector(2, LLT::float16());
+static const LLT V2BF16 = V2F16; // FIXME
----------------
arsenm wrote:

bfloat currently has no globalisel representation 

https://github.com/llvm/llvm-project/pull/95393