r207433 - [driver] Disable the slp vectorizer at -O0, -O1, and -Oz. This mirrors the

[Chandler Carruth]

I think the real question is -Oz. My understanding is that the SLP
vectorizer should always make code smaller.

I don't think matching the loop vectorizer is important. I think the
expectations of -O flags are not about the vectorization of code, but about
the trade-offs between user experiences. While I've mentioned it on other
threads:

O1: quickly optimize what you can
O2: our general ideal tradeoff between compile time and optimizations
O3: (i wish we didn't have it) burn more compile time and/or binary size to
try and dig out some optimizations.
Os: optimize, but try to make the resulting code smaller where possible.
Oz: do anything necessary to make code smaller, even things that hurt
performance

My impression was that SLP vectorization made sense at least for 2, 3, s,
and z. If the compile time is low enough, maybe for O1 as well.



On Mon, Apr 28, 2014 at 1:09 PM, Chad Rosier <mcrosier at codeaurora.org>wrote:

> One argument for matching the behavior of the loop vectorizer is that it's
> probably what the user expects.
>
> > Hi Chandler,
> > The discussion was in the PR as well as via IRC.  I'm certainly not
> > against continuing the discussion.
> >
> > I believe everyone strongly agrees this shouldn't be on at -O0;  Let me
> > know if you feel otherwise.  The decision to match the behavior of the
> > loop vectorizer at -O1 and -Oz was mostly arbitrary and I have no strong
> > feelings either way.
> >
> > What are your thoughts?
> >
> >  Chad
> >
> >> On Mon, Apr 28, 2014 at 12:30 PM, Chad Rosier
> >> <mcrosier at codeaurora.org>wrote:
> >>
> >>> -O1, and -Oz
> >>
> >>
> >> Wait, why? Based on what discussion?
> >>
> >> This is a pretty significant change, I'd like to understand why it was
> >> made.
> >>
> >
> >
> > --
> > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> > hosted by The Linux Foundation
> >
>
>
> --
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> hosted by The Linux Foundation
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20140428/31081dbd/attachment.html>