[gentoo-dev] CPU use flag detection

[gentoo-dev] CPU use flag detection

[yac]

Hi,

I was recently investigating what cpu flags do I have and how does it
work. I have put what I have so far at [1]. 

So I thought I let you know in case someone wants to chip in.

[1] https://github.com/yaccz/cufd

[gentoo-dev] Re: CPU use flag detection

[Ryan Hill]

[-- Attachment #1: Type: text/plain, Size: 893 bytes --]

On Wed, 15 May 2013 16:59:57 +0200
yac <yac@gentoo.org> wrote:

> Hi,
> 
> I was recently investigating what cpu flags do I have and how does it
> work. I have put what I have so far at [1]. 
> 
> So I thought I let you know in case someone wants to chip in.
> 
> [1] https://github.com/yaccz/cufd

I've seen gcc -Q --help=target giving false results before due to the way that
options are parsed and some flags that set other flags not getting processed on
the --help code path.  The second option you list is better.

-native doesn't set 3dnow, mmx, sse, sse2, sse3, ssse3 or sse4a.  I'm guessing
they're just handed through the -march setting.

sys-apps/cpuid is awesome.

MMX2/MMXEXT still confuses me.


-- 
Ryan Hill                        psn: dirtyepic_sk
   gcc-porting/toolchain/wxwidgets @ gentoo.org

47C3 6D62 4864 0E49 8E9E  7F92 ED38 BD49 957A 8463

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 490 bytes --]

Re: [gentoo-dev] Re: CPU use flag detection

[Matt Turner]

On Fri, May 17, 2013 at 9:39 PM, Ryan Hill <dirtyepic@gentoo.org> wrote:
> On Wed, 15 May 2013 16:59:57 +0200
> yac <yac@gentoo.org> wrote:
>
>> Hi,
>>
>> I was recently investigating what cpu flags do I have and how does it
>> work. I have put what I have so far at [1].
>>
>> So I thought I let you know in case someone wants to chip in.
>>
>> [1] https://github.com/yaccz/cufd
>
> I've seen gcc -Q --help=target giving false results before due to the way that
> options are parsed and some flags that set other flags not getting processed on
> the --help code path.  The second option you list is better.
>
> -native doesn't set 3dnow, mmx, sse, sse2, sse3, ssse3 or sse4a.  I'm guessing
> they're just handed through the -march setting.

Yes, they are.

> sys-apps/cpuid is awesome.
>
> MMX2/MMXEXT still confuses me.

SSE1 and /Enhanced/ 3DNow! added some extra MMX instructions. Some
(pshufw and pmulhuw particularly) turn out to be rather useful in
software compositing. I use them in the pixman MMX code.

See http://mattst88.com/programming/asmref/ (Sorry about the broken
search box. It works, you just don't see what you type)

Set Requires = Enhanced 3DNow!. The instructions that are listed as
'Enhanced 3DNow! or SSE1' are what are known as MMX2 or MMXEXT.

The particularly annoying thing about using them is that there's no
-mmmx2 or -mmmxext, but turning on -msse causes illegal instructions
to be generated for old AMD or Geode CPUs, while -m3dnow causes the
same problems for Intel CPUs. And, there's not actually even an
-m3dnowext flag (anymore?) anyway, so it's kind of a mess.

Re: [gentoo-dev] Re: CPU use flag detection

[Walter Dnes]

On Sat, May 18, 2013 at 12:14:35PM -0700, Matt Turner wrote

> The particularly annoying thing about using them is that there's no
> -mmmx2 or -mmmxext...

  Now that you mention it...

[i660][waltdnes][~] grep mmxext /usr/portage/profiles/use.local.desc 
media-libs/libpostproc:mmxext - Enable mmx2 support.
media-plugins/vdr-softdevice:mmxext - enables MMXExt support
media-video/ffmpeg:mmxext - Enables mmx2 support
media-video/libav:mmxext - Enable mmx2 support.
media-video/mplayer:mmxext - Enables mmx2 support
media-video/mplayer2:mmxext - Enables mmx2 support
x11-libs/pixman:mmxext - Enable MMX2 support.

...shouldn't "mmxext" be moved out of use.local.desc into use.desc?

-- 
Walter Dnes <waltdnes@waltdnes.org>
I don't run "desktop environments"; I run useful applications

Re: [gentoo-dev] Re: CPU use flag detection

[Alexis Ballier]

On Sat, 18 May 2013 22:31:11 -0400
"Walter Dnes" <waltdnes@waltdnes.org> wrote:

[...]
> ...shouldn't "mmxext" be moved out of use.local.desc into use.desc?


all the cpu flags should be global IMHO, regardless of how many packages
use them: we already mask/unmask them globally on arches where they are
irrelevant.

Alexis.

Re: [gentoo-dev] Re: CPU use flag detection

[vivo75]

On 05/19/13 17:47, Alexis Ballier wrote:
> On Sat, 18 May 2013 22:31:11 -0400
> "Walter Dnes" <waltdnes@waltdnes.org> wrote:
>
> [...]
>> ...shouldn't "mmxext" be moved out of use.local.desc into use.desc?
>
> all the cpu flags should be global IMHO, regardless of how many packages
> use them: we already mask/unmask them globally on arches where they are
> irrelevant.
>
> Alexis.
>
+1
also +1 they should be easily recognizable as cpu features USE flags, as
already mentioned recently by someone (sorry could not search)

Re: [gentoo-dev] Re: CPU use flag detection

[René Neumann]

Am 19.05.2013 18:03, schrieb vivo75@gmail.com:
> On 05/19/13 17:47, Alexis Ballier wrote:
>> On Sat, 18 May 2013 22:31:11 -0400
>> "Walter Dnes" <waltdnes@waltdnes.org> wrote:
>>
>> [...]
>>> ...shouldn't "mmxext" be moved out of use.local.desc into use.desc?
>>
>> all the cpu flags should be global IMHO, regardless of how many packages
>> use them: we already mask/unmask them globally on arches where they are
>> irrelevant.
>>
>> Alexis.
>>
> +1
> also +1 they should be easily recognizable as cpu features USE flags, as
> already mentioned recently by someone (sorry could not search)

So this sounds like one of the places, where a USE_EXPAND could really
be correct and worthwhile :).

- René

Re: [gentoo-dev] Re: CPU use flag detection

[Matt Turner]

On Sun, May 19, 2013 at 8:47 AM, Alexis Ballier <aballier@gentoo.org> wrote:
> On Sat, 18 May 2013 22:31:11 -0400
> "Walter Dnes" <waltdnes@waltdnes.org> wrote:
>
> [...]
>> ...shouldn't "mmxext" be moved out of use.local.desc into use.desc?
>
>
> all the cpu flags should be global IMHO, regardless of how many packages
> use them: we already mask/unmask them globally on arches where they are
> irrelevant.
>
> Alexis.

Yes, they should. And while we're at it, we should stop naming them
differently. (sse4_1 vs sse41)

Also, since SSE4 isn't a thing (there's SSE4.1, SSE4.2, and SSE4a) we
shouldn't have /any/ USE=sse4 in the tree.

And we should never set up the masks like this:
https://bugs.gentoo.org/show_bug.cgi?id=470220

[gentoo-dev] Re: CPU use flag detection

[Ryan Hill]

[-- Attachment #1: Type: text/plain, Size: 1310 bytes --]

On Sat, 18 May 2013 12:14:35 -0700
Matt Turner <mattst88@gentoo.org> wrote:

> > MMX2/MMXEXT still confuses me.
> 
> SSE1 and /Enhanced/ 3DNow! added some extra MMX instructions. Some
> (pshufw and pmulhuw particularly) turn out to be rather useful in
> software compositing. I use them in the pixman MMX code.
> 
> See http://mattst88.com/programming/asmref/ (Sorry about the broken
> search box. It works, you just don't see what you type)

Oh I am going to get a lot of use out of this.

> Set Requires = Enhanced 3DNow!. The instructions that are listed as
> 'Enhanced 3DNow! or SSE1' are what are known as MMX2 or MMXEXT.
> 
> The particularly annoying thing about using them is that there's no
> -mmmx2 or -mmmxext, but turning on -msse causes illegal instructions
> to be generated for old AMD or Geode CPUs, while -m3dnow causes the
> same problems for Intel CPUs. And, there's not actually even an
> -m3dnowext flag (anymore?) anyway, so it's kind of a mess.

Ah, that explains why I always thought they were Enh 3DNow options.  Looks like
I have them spread out between MMX and Enh 3DNow in analyze-x86.  Oops.


-- 
Ryan Hill                        psn: dirtyepic_sk
   gcc-porting/toolchain/wxwidgets @ gentoo.org

47C3 6D62 4864 0E49 8E9E  7F92 ED38 BD49 957A 8463

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 490 bytes --]