[llvm] r268539 - [AArch64] Use the reciprocal estimation machinery

[Evandro Menezes via llvm-commits]

Renato,

You're looking at the initial version of the patch, before Matthias 
added the appropriate feature to remove the test for a specific core.

And, not unlike other features, yes, it's only used for specific 
subtargets.  Singling it out amidst a dozen of features seems arbitrary.

Regardless, I'll propose a cost model to be used in the DAG combiner 
generically by targets in the discussion you pointed out.

Thank you,

-- 
Evandro Menezes

On 09/20/16 07:36, Renato Golin wrote:
> On 20 September 2016 at 13:15, Evandro Menezes via llvm-commits
> <llvm-commits at lists.llvm.org> wrote:
>> Thus, if anything, the discussion should be broader than AArch64.  Until
>> that's settled, it's preferable to keep this patch until an universal
>> solution is found, IMO.
> Hi Evandro,
>
> I agree this needs a wider audience, but the patch is introducing
> "isCoreX" which we have extensively discussed is a pattern we want to
> be moving away from, not towards. Making it into
> "hasFeatureFastReciprocal" won't help either. If you read the
> discussion topics I pointed you, this is not just dependent on the
> architecture, but on the context.
>
> Rushing this feature in and leaving it there to push artificial
> benchmark numbers up (no one else will be building them with
> -mcpu=exynos-m1) seems tempting, but we much refrain from specific
> code when the behaviour is not deterministic in the overall case.
>
> cheers,
> --renato
>
>
>