[llvm] r207528 - Add optimization remarks to the loop unroller and vectorizer.

[David Blaikie]

On Tue, Apr 29, 2014 at 1:04 PM, Diego Novillo <dnovillo at google.com> wrote:
>
>
>
> On Tue, Apr 29, 2014 at 8:23 AM, David Blaikie <dblaikie at gmail.com> wrote:
>>
>>
>> Test coverage? (I assume we've plumbed the remark stuff through llc so
>> this can be tested in LLVM directly?)
>
>
> These are doable, but may be tricky in general. Not all passes allow their
> hand to be forced by overriding their cost model. This may mean additional
> burden in the future if the current models change.

Yeah, I'm assuming there are some "/obviously/ we should do this
optimization here" and "/obviously/ we should never do this
optimization here" that could be used as examples to make test cases
to verify the reporting that would be stable across tweaks to the
optimizers, etc.

Which, if I'm reading you correctly, it sounds like what you're describing \/

> I've added some tests for cases that should not change too much.  The
> coverage we have on the transformations is not complete either.  Some of the
> loop unrolling features don't seem to be exercised.  I'll add more as I find
> them in our code base.
>
>
> Diego.