wangpc-pp wrote:

> > > * `-0.9%` of vector spills.
> > > * `-1.1%` of vector reloads.
> > 
> > 
> > Do you know why this patch improve spills and reloads?
> 
> I was surprised by this result as well. I will do some researches and try to 
> find some cases. But TBH, this can be a noise as register allocation is 
> fragile and easy to be affected.

We only see the improvements of spills and reloads on a small group of tests 
(like `matrix-types-spec`). For these cases, LMUL/register pressure are large. 
I guess this new scheduling heuristics *accidentally* move some instructions to 
a place closer to its uses. The absolute value of reduced spilled/reloads are 
really small, let's consider them as a noise.

https://github.com/llvm/llvm-project/pull/95924
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits

Reply via email to