[PATCH] Add optimization remarks to the loop unroller and vectorizer.
Diego Novillo
dnovillo at google.com
Thu Apr 24 10:37:26 PDT 2014
On Thu, Apr 24, 2014 at 9:33 AM, Tyler Nowicki <tnowicki at apple.com> wrote:
> Hi Diego,
>
> Would a missed or failed optimization use a different type of diagnostic
> remark? Perhaps a MissedOptimizationRemark?
>
>
Yeah, in the original design document (
https://docs.google.com/document/d/1FYUatSjZZO-zmFBxjOiuOzAy9mhHA8hqdvklZv68WuQ/edit?usp=sharing),
we would say things like -Rpass=pattern=kind, where "kind" would specify
what type of remark we want.
I briefly discussed offline this with Richard. The things I have in mind
for "kind" are:
- kind=applied: what we have today. -Rpass=inline is the same as
-Rpass=inline=applied. It tells you when the pass named "inline" applies
its transformation.
- kind=missed: -Rpass=inline=missed tells you when the pass tried to apply
the transformation but it failed. The diagnostic would include a short
explanation.
- kind=analysis: -Rpass=inline=analysis is fairly verbose set of remarks
that give more detail on what analysis went on behind the decision to apply
or miss the transformation.
I am still not completely convinced that this is a reasonable interface.
Suggestions more than welcome.
Diego.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140424/3762ccdf/attachment.html>
More information about the llvm-commits
mailing list