[llvm-commits] [PATCH] Execution domain support for VMOV and VLDR

[Jakob Stoklund Olesen]

On Aug 31, 2012, at 1:17 PM, Tim Northover <t.p.northover at gmail.com> wrote:

> Thanks for your reassurances! I should be able to get a patch ready to
> make things correct again tomorrow.
> 
>> If you find that you do need to determine if a lane is live at a given point, it would be nice to have a general-purpose function to do it.
> 
> That would obviously be the comprehensive way to do it, but I was
> unsure how it would be viewed by others. It's a general liveness
> calculation with performance implications (the X86 one you mentioned
> earlier seemed to limit itself to a few instructions before giving
> up).
> 
> I'd sort of thought that if LLVM has existed so long without it, that
> was probably because of a policy decision rather than lack of
> interest. Was I wrong there?

We haven't had much need for liveness computations after RA, except for PEI which uses the register scavenger.

For something like this, I think it would make sense to perform a limited scan of the neighborhood, and to have the option of returning "don't know" instead of scanning all of a giant basic block.

If you can't find a mention of the other lane or the super-register in a 25-instruction neighborhood, chances are the NEON conversion won't help much anyway.

/jakob