[llvm] [AMDGPU][wmma] - Add tied wmma intrinsic (PR #69903)

[Nicolai Hähnle via llvm-commits]

nhaehnle wrote:

> Would "preserve" be a better suffix, as in "preserve the other half"?

I don't feel strongly about it, but in an SSA world I think "copy" is the better way to think about it. "Preserve" is the way to think about it in a non-SSA world, i.e. at the assembly level after register allocation.

> If this is strictly more expressive than what we already have for the same instructions, can we just auto-upgrade the existing intrinsic to the version with the tied operand?

I wouldn't say it's more expressive, it's just different. As Piotr said, using the tied versions when you don't need them can lead to worse codegen.

https://github.com/llvm/llvm-project/pull/69903