[llvm-commits] [patch] Don't create neon moves in CopyRegToReg, instead trust NEONMoveFixPass to do the conversion

[Evan Cheng]

I have no problem with the patch. But I  would feel a lot better if we are sure NEONMoveFix pass does not miss some important cases. I thought Jakob pointed out (in an earlier) thread that it's missing a case?

Evan

On Jul 5, 2010, at 8:21 AM, Rafael Espindola wrote:

> A recent patch caused a regression on ARM: we would create NEON copy
> instructions when a VFP one would be more profitable. Two options were
> discussed:
> 
> *) Represent the NEON registers as distinct from the VFP registers and
> have explicit copies from one to the other.
> *) Figure out why NEONMoveFixPass was not fixing this.
> 
> Currently NEONMoveFixPass only converts VFP moves to NEON moves. This
> looks reasonable, we just have to avoid creating NEON moves early.
> This is what the attached patch does. It fixes the previous
> regression.
> 
> Is there a case I am missing? That is, a case where CopyRegToReg was
> correctly creating a NEON move but NEONMoveFixPass fails to convert?
> The tests found nothing.
> 
> Cheers,
> -- 
> Rafael Ávila de Espíndola
> <neon.patch>_______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits