[llvm] r182499 - This is an update to a previous commit (r181216).

[Rafael Espíndola]

On 28 May 2013 10:42, Jean-Luc Duprat <jduprat at apple.com> wrote:
> Rafael,
>
> I just wanted to circle back with you and summarize the conclusions of more internal discussions we came to.
> The select form of the operation is most likely the right cannonicalization for this expression, and it is in fact possible that the vectorizer could do something with that representation.  At the current time however, this would have a negative impacton performance because it impairs FMA formation.  The proper fix is to fix FMA formation in all relevant backends, then enable the select inst combine, and finally to get the vectorizer to make use of the canonical select in helping make decisions.  This is quite a bit of work for unquantifiable performance gains…
>
> In the short term, the fix I submitted maintains performance to the same level as it was prior to my initial commit.
> As the above changes roll in, performance may improve.  We are tracking the request for changes, and will get to them eventually; but this is fairly low priority for us at the current time.

Thanks for summary, it sounds reasonable. Could you just open a bug on
llvm.org/bugs to track it?

> JL

Thanks,
Rafael