[llvm] [AMDGPU] Add baseline test to show spilling of wmma scale. NFC (PR #168163)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

> > This is to show the spilling of WMMA scale values which are limited to low 256 VGPRs. We have free registers, just RA allocates low 256 first. The only problem, this test is large.
> 
> Can you maybe capture the test at the mir level and constraint the liveness using KILL instructions with a bunch of defs and uses? Alternatively you could play with clobber list with inline asm to artificially constraint the available register. Finally, you may be able to leverage the `stress-regalloc` option to burn some registers without having a long sequence of instruction.

I have tried inline asm with clobber list to no avail, it is not so easy to fool greedy. Stress regalloc does not help because in this case we shall not reduce number of available registers. Contrarily we have more than needed but still spilling. Reducing the test to mir can be a next step, yes.

https://github.com/llvm/llvm-project/pull/168163