[llvm] [AMDGPU] Fix immediate parsing for packed types (PR #202417)

[Shilei Tian via llvm-commits]

================
@@ -2053,7 +2053,7 @@ static const fltSemantics *getFltSemantics(unsigned Size) {
 }
 
 static const fltSemantics *getFltSemantics(MVT VT) {
-  return getFltSemantics(VT.getSizeInBits() / 8);
+  return getFltSemantics(VT.getScalarSizeInBits() / 8);
----------------
shiltian wrote:

I think there's some historical reason for this, dating back to before my time at AMD. We used to represent `half` as `i16` before we had proper support for it. Even today, I think we still use `vNi16` for some instructions.

https://github.com/llvm/llvm-project/pull/202417