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

[Renato Golin via llvm-commits]

On 20 September 2016 at 15:00, Evandro Menezes <e.menezes at samsung.com> wrote:
> And, not unlike other features, yes, it's only used for specific subtargets.
> Singling it out amidst a dozen of features seems arbitrary.

Eric has identified a problem with the implementation, this may be
more serious than we can see from here.

Eric, the review was https://reviews.llvm.org/D19426.

Can you be more specific as to what could go wrong when linking
functions with different options?

AFAICS, this is a problem that will happen with *any* feature flag,
possibly even scheduler choices, if sub-arch can change for different
functions.

I imagine LTO could be particularly bad if you don't pass all the options.


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

This is a *much* larger implementation than just the scheduler. If
something is really broken, we can't wait for that to happen.

cheers,
--renato