[gentoo-dev] filter-mfpmath: has Gentoo out grown it ?

[gentoo-dev] filter-mfpmath: has Gentoo out grown it ?

[Mike Frysinger]

[-- Attachment #1: Type: text/plain, Size: 1028 bytes --]

i'd like to think Gentoo has grown up now to the point where we don't bother 
with trivial ricer behavior.  to that end, i'd like to EOL `filter-mfpmath`.

my main beef with filtering -mfpmath is that we use this only when someone 
actually notices and reports misbehavior with the package in question, the 
behavior can fluctuate between gcc versions, and it's questionable whether the 
flag makes a significant difference in performance.  considering this is x86-
only, and our main user base is amd64 based, it doesn't see nearly the amount 
of attention that it did in the past.  i'd prefer we leave this flag to the 
respective upstream packages to validate when it should be used (i.e. the 
mplayer's and ffmpeg's and such in the world).

Ryan did a check and it seems we've got all of 5 packages (and one eclass) 
using this.  so i say it's time to scrub the tree, punt the func, and then 
punt people who attempt to report bugs when building their whole system with -
mfpmath and see misbehavior.
-mike

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: [gentoo-dev] filter-mfpmath: has Gentoo out grown it ?

[Mike Gilbert]

[-- Attachment #1: Type: text/plain, Size: 1611 bytes --]

On 12/02/2011 05:57 PM, Mike Frysinger wrote:
> i'd like to think Gentoo has grown up now to the point where we don't bother 
> with trivial ricer behavior.  to that end, i'd like to EOL `filter-mfpmath`.
> 
> my main beef with filtering -mfpmath is that we use this only when someone 
> actually notices and reports misbehavior with the package in question, the 
> behavior can fluctuate between gcc versions, and it's questionable whether the 
> flag makes a significant difference in performance.  considering this is x86-
> only, and our main user base is amd64 based, it doesn't see nearly the amount 
> of attention that it did in the past.  i'd prefer we leave this flag to the 
> respective upstream packages to validate when it should be used (i.e. the 
> mplayer's and ffmpeg's and such in the world).
> 
> Ryan did a check and it seems we've got all of 5 packages (and one eclass) 
> using this.  so i say it's time to scrub the tree, punt the func, and then 
> punt people who attempt to report bugs when building their whole system with -
> mfpmath and see misbehavior.
> -mike

I've never thought of -mfpmath=sse as a "ricer" flag. If anything, it
should make floating point calculations more consistent. Back when I
used x86, I had it in my global CFLAGS and can't remember it causing any
issues.

The varying behavior between gcc versions seems like a possibly good
reason to not "support" it, but I'm not familiar with this. Has it
stabilized recently or is it still in flux?

I would rather leave this in the hands of package maintainers than punt
it globally.


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 230 bytes --]

[gentoo-dev] Re: filter-mfpmath: has Gentoo out grown it ?

[Ryan Hill]

[-- Attachment #1: Type: text/plain, Size: 1071 bytes --]

On Fri, 02 Dec 2011 20:36:16 -0500
Mike Gilbert <floppym@gentoo.org> wrote:

> I've never thought of -mfpmath=sse as a "ricer" flag. If anything, it
> should make floating point calculations more consistent.

Not really.  It's it's extremely target processor dependent and generally the
defaults are the best settings.  It can vary a lot based on the SSE
instructions and number of registers available and the register allocator and
scheduler are optimized with that in mind.

> The varying behavior between gcc versions seems like a possibly good
> reason to not "support" it, but I'm not familiar with this. Has it
> stabilized recently or is it still in flux?

I'd say it's still in flux.  The register allocator and instruction scheduler
were both rewritten for 4.4 and have seen a lot of work since then.


-- 
fonts, gcc-porting,                  it makes no sense how it makes no sense
toolchain, wxwidgets                           but i'll take it free anytime
@ gentoo.org                EFFD 380E 047A 4B51 D2BD C64F 8AA8 8346 F9A4 0662

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 198 bytes --]