[gentoo-amd64] FW: GCC4 and -ftree-vectorize

[gentoo-amd64] FW: GCC4 and -ftree-vectorize

[Duncan]

-ftree-vectorize just came up on the gentoo-dev list.  I've not even
checked the URLs yet, but since it came up here, I thought I'd ask about
it and having gotten a response, here's the gist.  (FWIW, I've been
running it experimentally for a bit now.  No issues with it have I seen,
but I've not done a full emerge --emptytree system/world yet.)  I'll
probably have a bit more to say about it after checking the tinyurls,
which at this point I don't even know where they go yet, let alone their
content.

<quote from Ryan Hill>

Duncan wrote:

> Could you point me at some info on this one (-ftree-vectorize)?  It came
> up on the amd64 list a week or so ago[]

http://tinyurl.com/l75we

They've fixed quite a few of the ICE's since last I looked, though there's
more than a couple that went in after 4.1.1.  4.2 is a little better, but
I'm having enough trouble getting it to build things properly _without_
using any fancy flags right now. ;p

See http://tinyurl.com/rt3aa for some real-world examples.

> Or does the problem not necessarily apply to amd64?

I'd say that there's more ICE's on i686-pc-linux-gnu than
x86_64-*-linux-gnu, but there's still enough.  Luckily Halcy0n was really
good for reducing testcases and pushing them upstream, so a lot of these
issues got fixed at the source.

<end of quote from Ryan>

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

-- 
gentoo-amd64@gentoo.org mailing list

[gentoo-amd64] Re: FW: GCC4 and -ftree-vectorize

[Duncan]

"Duncan" <1i5t5.duncan@cox.net> posted efqqvt$sqh$6@sea.gmane.org,
excerpted below, on  Mon, 02 Oct 2006 10:51:41 +0000:

> -ftree-vectorize just came up on the gentoo-dev list.

...  And here's something from Flame-eyes (and Donnie) on x86 (and PPC) and
amd64 specifically...

< quote from Flame-eyes >

On Sunday 01 October 2006 18:49, Donnie Berkholz wrote:
> I can't give you reasons, but I can tell you it totally broke my x86
> system a while back.
-ftree-vectorize on x86 and PowerPC is known to create broken executables
(when it comes to actually create the executable). I'm using it on AMD64
fine since 4.0 pre-releases though (suggested for the daredevils by
Halcy0n iirc), and never had problems with it. Most likely this is due
either to the usual limitations of x86 (registers, PIC, 387 fpmath), or to
not be able to ensure the availability of extra instructions on x86.

</quote>

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

-- 
gentoo-amd64@gentoo.org mailing list

[gentoo-amd64] Re: FW: GCC4 and -ftree-vectorize

[Duncan]

"Duncan" <1i5t5.duncan@cox.net> posted efqqvt$sqh$6@sea.gmane.org,
excerpted below, on  Mon, 02 Oct 2006 10:51:41 +0000:

> I'll probably have a bit more to say about it after checking the
> tinyurls, which at this point I don't even know where they go yet, let
> alone their content.
> 
> <quote from Ryan Hill>
> 
> Duncan wrote:
> 
>> Could you point me at some info on this one (-ftree-vectorize)?  It
>> came up on the amd64 list a week or so ago[]
> 
> http://tinyurl.com/l75we

OK, this one's the gcc bugtracker list of -ftree-vectorize bugs.  Lots of
reports, many of them reduced to skeleton programs with just the
triggering code. Of note, /most/ of the issues involve x86/ppc and/or size
increasing loop optimizations that aren't of interest to those whose
strategy targets -Os in any case. For example, one of the triggers seems
to be -funswitch-loops, enabled by default with -O3, another is
-funroll-loops, neither of which (again) are of interest to those
targeting size for cache efficiency among other reasons.

> See http://tinyurl.com/rt3aa for some real-world examples.

This one's the Gentoo bugtracker -ftree-vectorize related bug list. 
Again, most are x86/ppc and/or involve flags folks interested in -Os
won't be worrying about.

Looks like on amd64 it may be less trouble than -combine is, and Flameeyes
said he has been using it for awhile, so it looks like it'll be standard
in my cflags now, to be removed only where it seems to cause problems,
which shouldn't be that often, hopefully.

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

-- 
gentoo-amd64@gentoo.org mailing list