Thought: I like it :)

On 7/6/06, Diego 'Flameeyes' Pettenò <[EMAIL PROTECTED]> wrote:
So, I've been drafting this up in my blog[1], and it is a simple way to
replace the CPU feature useflags. Let's try to summarise:

Right now we have mmx, 3dnow, 3dnowex, sse, sse2 and so on useflags present in
the tree, almost never used to get new dependencies, but usually used to
supply econf switches.

This works as long as the user enable the flags, but for AMD64 the story was,
until now, that we simply enabled them when they worked, because we had some
minimum support available. Unfortunately this became a problem with the
introduction of nocona, because that is an amd64-like system but with no
3dnow. And there is the problem that sse3 is supported only in later versions
of Athlon64 and so on.

To try to clean up this mess, and to make it simpler to work in
cross-compilation, I thought of using the definitions created by the C
Preprocessor (CPP) by default for the given CFLAGS. Basically when
using -march=athlon64, the preprocessor will enable a few definitions that
tells the availability of MMX, 3dNOW, SSE and so on... if we wrap that
around, we can use it to know if it's the case of enabling something or not.
This is customisable by the user by setting the CFLAGS themselves. If one does
not want MMX instructions to be generated, but still want the rest of
Athlon64 optimisations, it's simply the matter of
using "-march=athlon64 -mno-mmx".

So, rather than the functions proposed in [1], I've sampled today the
following function:

has_cpuset() {
   local def hasfeat

   while [[ -n $1 ]]; do
       case $1 in
           mmx)
               def="__MMX__" ;;
           3dnow)
               def="__3dNOW__" ;;
           3dnowex)
               def="__3dNOW_A__" ;;
           sse)
               def="__SSE__" ;;
           sse2)
               def="__SSE2__" ;;
           sse3)
               def="__SSE3__" ;;
           altivec)
               def="__ALTIVEC__" ;;
           *)
               ewarn "Instruction set $1 not supported."
               die "Instruction set not supported."
       esac

       echo | $(tc-getCC) ${CFLAGS} -dM -E - 2>/dev/null | grep -q ${def} ||
hasfeat="no"
       shift
   done

   if [[ ${hasfeat} == "no" ]]; then
       return 1
   else
       return 0
   fi
}

that can be tested easily with the following code:

yesno() {
   if "$@"; then
       echo "yes"
   else
       echo "no"
   fi
}

echo "Does it have 3dnow mmx sse2?"
yesno has_cpuset 3dnow mmx sse2
echo "Does it have mmx sse sse3?"
yesno has_cpuset mmx sse sse3
echo "Does it have altivec?"
yesno has_cpuset altivec


Note that you  need to set your CFLAGS corretly or it will, by default, tell
you that everything is disabled.

Thoughts? Comments?

SPARC team: I'd like to know if VIS does a similar thing, would make simpler
for instance its handling in xine-lib ebuild.

[1]
http://farragut.flameeyes.is-a-geek.org/articles/2006/06/24/crazy-idea-an-alternative-to-cpu-features-useflags
--
Diego "Flameeyes" Pettenò - http://farragut.flameeyes.is-a-geek.org/
Gentoo/Alt lead, Gentoo/FreeBSD, Video, AMD64, Sound, PAM, KDE





--
Ioannis Aslanidis

<deathwing00[at]gentoo.org> 0xB9B11F4E

--
gentoo-dev@gentoo.org mailing list

Reply via email to