[llvm-commits] PATCH: Teach the x86 backend to fold mask & shift patterns into an addressing mode computation

[Chandler Carruth]

On Tue, Jan 10, 2012 at 2:46 PM, Jakob Stoklund Olesen <stoklund at 2pi.dk>wrote:

>
> On Jan 10, 2012, at 11:14 AM, Chandler Carruth wrote:
>
> > Anyone else have comments?
>
> Did you actually see code run faster?
>
> Simple addressing modes (base+offset) can be faster on Sandy Bridge, for
> example.
>

Yes, as mentioned in my original message, I'm seeing noticable gains in the
hot loop of Snappy. I can benchmark it and some other software and report
more detailed performance numbers (and I'll do so either before or after
the commit, either way)

It's a bit hard to accurately measure the impact though, as there are more
bugs I've fixed in subsequent patches that prevent this (and other)
transformations from always firing on x86-64 compiles...
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20120110/6763937b/attachment.html>