[clang] [llvm] AMDGPU: Implement tensor load and store instructions for gfx1250 (PR #146636)

[Krzysztof Drewniak via llvm-commits]

krzysz00 wrote:

Very late review feedback:

I see that we've defined `_d2` variants of all the instructions (which, by not taking arguments for descriptor groups 2 and 3, make them NULL) but we haven't defined `_d3` variants (that let SGPR group 3 be null).

I'd like to suggest that we either do that or, more preferrably, get rid of these `_d2` variants and just recognize `zeroinitializer` as the NULL registers

https://github.com/llvm/llvm-project/pull/146636