[llvm] [RegAlloc] Scale the spill weight by target factor (PR #113675)

[Philip Reames via llvm-commits]

preames wrote:

> I still think we should be making active effort to just make this the default. I believe this is another case where the target hook is just being used to paper over issues in other areas (I'm not objecting to this as an incremental step).

I agree with this, but FYI, your last clause had gotten lost in previous review comments, at least for me.  I do think it's important to highlight that the thing being papered over is for the target keeping the old behavior.  Long term, I do think all targets should have this particular behavior.  

> In particular in the previous revision of the patch, the AMDGPU behavior showed rematerialize was totally broken. I also think that something is wrong with whatever tablegen is doing to compute the default class weight. We have quite a lot of code scattered in the allocator and tablegen that were never written to account for subregisters. The defaults are wrong for the allocation priorities (which RISCV also does not look like it's trying to set). We also lack a splitting strategy to evict only a subregister when it would help, and spilling induces new liveness to dead lanes.

I've been looking into some of this, and trying to find ways to split things apart into some actionable pieces.  Incrementalism here is critical for keeping the problem vaguely understandable.  However, it sounds like you have a couple cases identified that I haven't yet stumbled into to.  Could we compare notes offline?  



https://github.com/llvm/llvm-project/pull/113675