[gentoo-dev] RFC: LD_AS_NEEDED="1" in profiles/targets/dev..

Ciaran McCreesh wrote:
> On Sat, 03 Oct 2009 22:13:59 +0300
> Samuli Suominen <ssuominen.RemoveThis@gentoo.org> wrote:
>> Since new binutils will support LD_AS_NEEDED="1" to force ld behave
>> asneeded we could use this for the developer -target in profiles?
>>
>> Speak up if you think it's a terrible idea.
>
> Well, it does break correct code, so it's about on par with forcing
> -ffast-math or -fno-exceptions or -fvisibility-inlines-hidden on
> globally...
>

Be reasonable. We can add -Wl,--no-as-needed for those packages with
valid code. And I'm certainly not suggesting adding it to base/, _only_
for the developer target.

On 03-10-2009 22:35:58 +0300, Samuli Suominen wrote:
> Ciaran McCreesh wrote:
> > On Sat, 03 Oct 2009 22:13:59 +0300
> > Samuli Suominen <ssuominen.TakeThisOut@gentoo.org> wrote:
> >> Since new binutils will support LD_AS_NEEDED="1" to force ld behave
> >> asneeded we could use this for the developer -target in profiles?
> >>
> >> Speak up if you think it's a terrible idea.
> >
> > Well, it does break correct code, so it's about on par with forcing
> > -ffast-math or -fno-exceptions or -fvisibility-inlines-hidden on
> > globally...
> >
>
> Be reasonable. We can add -Wl,--no-as-needed for those packages with
> valid code. And I'm certainly not suggesting adding it to base/, _only_
> for the developer target.

Please recall that this is a GNU binutils flag. I prefer the
environment flag over the command-line option, because it doesn't break
other linkers, but if it implies many ebuilds getting -Wl,--no-as-needed
append-ldflagged, then I'm not getting happy as that breaks my builds.


--
Fabian Groffen
Gentoo on a different level

Fabian Groffen wrote:
> On 03-10-2009 22:35:58 +0300, Samuli Suominen wrote:
>> Ciaran McCreesh wrote:
>>> On Sat, 03 Oct 2009 22:13:59 +0300
>>> Samuli Suominen <ssuominen.TakeThisOut@gentoo.org> wrote:
>>>> Since new binutils will support LD_AS_NEEDED="1" to force ld behave
>>>> asneeded we could use this for the developer -target in profiles?
>>>>
>>>> Speak up if you think it's a terrible idea.
>>> Well, it does break correct code, so it's about on par with forcing
>>> -ffast-math or -fno-exceptions or -fvisibility-inlines-hidden on
>>> globally...
>>>
>> Be reasonable. We can add -Wl,--no-as-needed for those packages with
>> valid code. And I'm certainly not suggesting adding it to base/, _only_
>> for the developer target.
>
> Please recall that this is a GNU binutils flag. I prefer the
> environment flag over the command-line option, because it doesn't break
> other linkers, but if it implies many ebuilds getting -Wl,--no-as-needed
> append-ldflagged, then I'm not getting happy as that breaks my builds.
>
>

The amount of pkgs in tree with valid code that fails with asneeded is
close to zero. We can use this,

if use userland_GNU; then
append-ldflags -Wl,--no-as-needed
fi

logic so it won't cause you troubles. Or the obvious shorter && version
of it.

On 03-10-2009 23:27:41 +0300, Samuli Suominen wrote:
> The amount of pkgs in tree with valid code that fails with asneeded is
> close to zero. We can use this,
>
> if use userland_GNU; then
> append-ldflags -Wl,--no-as-needed
> fi
>
> logic so it won't cause you troubles. Or the obvious shorter && version
> of it.

Not really, because the userland is GNU, the compiler too, but the
linker is not. We once played with a version of append-ldflags that
tested the flag against the linker to see if it would work, but that's
expensive as well as allowing real typos/bugs to go unnoticed.

We currently do either a CHOST check, or call ld to see if it's GNU for
all packages that hardcode -Wl,--{,no-}as-needed usage.


--
Fabian Groffen
Gentoo on a different level

Fabian Groffen wrote:
> On 03-10-2009 23:27:41 +0300, Samuli Suominen wrote:
>> The amount of pkgs in tree with valid code that fails with asneeded is
>> close to zero. We can use this,
>>
>> if use userland_GNU; then
>> append-ldflags -Wl,--no-as-needed
>> fi
>>
>> logic so it won't cause you troubles. Or the obvious shorter && version
>> of it.
>
> Not really, because the userland is GNU, the compiler too, but the
> linker is not. We once played with a version of append-ldflags that
> tested the flag against the linker to see if it would work, but that's
> expensive as well as allowing real typos/bugs to go unnoticed.
>
> We currently do either a CHOST check, or call ld to see if it's GNU for
> all packages that hardcode -Wl,--{,no-}as-needed usage.
>

Ah. I wasn't aware of that.

Perhaps we could add a new function to the flag-o-matic that does the
CHOST check, and appends the flag, so the check code wouldn't have to be
duplicated in ebuilds? It should be rather trivial.

On 04-10-2009 00:11:06 +0300, Samuli Suominen wrote:
> Perhaps we could add a new function to the flag-o-matic that does the
> > CHOST check, and appends the flag, so the check code wouldn't have to be
> > duplicated in ebuilds? It should be rather trivial.
> >
>
> ok, chost check would be cheap. how about,
>
> no-as-needed() {
> case $($(tc-getLD) -v 2>&1 </dev/null) in
> *GNU*) # GNU ld
> echo "-Wl,--no-as-needed" ;;
> esac
> }
>
> usage: append-ldflags $(no-as-needed)

that would work fine for me, and all possible future cases I can think
of


--
Fabian Groffen
Gentoo on a different level

Samuli Suominen wrote:
> Fabian Groffen wrote:
>> On 03-10-2009 23:27:41 +0300, Samuli Suominen wrote:
>>> The amount of pkgs in tree with valid code that fails with asneeded is
>>> close to zero. We can use this,
>>>
>>> if use userland_GNU; then
>>> append-ldflags -Wl,--no-as-needed
>>> fi
>>>
>>> logic so it won't cause you troubles. Or the obvious shorter && version
>>> of it.
>> Not really, because the userland is GNU, the compiler too, but the
>> linker is not. We once played with a version of append-ldflags that
>> tested the flag against the linker to see if it would work, but that's
>> expensive as well as allowing real typos/bugs to go unnoticed.
>>
>> We currently do either a CHOST check, or call ld to see if it's GNU for
>> all packages that hardcode -Wl,--{,no-}as-needed usage.
>>
>
> Ah. I wasn't aware of that.
>
> Perhaps we could add a new function to the flag-o-matic that does the
> CHOST check, and appends the flag, so the check code wouldn't have to be
> duplicated in ebuilds? It should be rather trivial.
>

ok, chost check would be cheap. how about,

no-as-needed() {
case $($(tc-getLD) -v 2>&1 </dev/null) in
*GNU*) # GNU ld
echo "-Wl,--no-as-needed" ;;
esac
}

usage: append-ldflags $(no-as-needed)

Ryan Hill wrote:
> On Sat, 03 Oct 2009 22:13:59 +0300
> Samuli Suominen <ssuominen DeleteThis @gentoo.org> wrote:
>
>> Since new binutils will support LD_AS_NEEDED="1" to force ld behave
>> asneeded we could use this for the developer -target in profiles?
>>
>> Speak up if you think it's a terrible idea.
>>
>> Thanks, Samuli
>>
>>
>
> I think it's a not terrible idea.
>
>
> Only barely related: can we enable FEATURES=test too?
>

*nod*, I think we can enable "test" too.

Fabian Groffen wrote:
>>> Perhaps we could add a new function to the flag-o-matic that does the
>>> CHOST check, and appends the flag, so the check code wouldn't have to be
>>> duplicated in ebuilds? It should be rather trivial.
>>
>> ok, chost check would be cheap. how about,
>> usage: append-ldflags $(no-as-needed)
>
> that would work fine for me, and all possible future cases I can think
> of

+# @FUNCTION: no-as-needed
+# @RETURN: Flag to disable asneeded behavior for use with append-ldflags.
+no-as-needed() {
+ case $($(tc-getLD) -v 2>&1 </dev/null) in
+ *GNU*) # GNU ld
+ echo "-Wl,--no-as-needed" ;;
+ esac
+}

It's committed.

To avoid the already caused trouble for you (prefix), these should be
converted:

(It's a ugly grep -r from gentoo-x86. I will handle this also in a QA
bug, opening one later today.)

app-admin/apache-tools/apache-tools-2.2.11.ebuild: # Instead of
filtering --as-needed (bug #128505), append --no-as-needed
app-admin/apache-tools/apache-tools-2.2.11.ebuild: append-ldflags
-Wl,--no-as-needed
app-admin/apache-tools/apache-tools-2.2.13.ebuild: # Instead of
filtering --as-needed (bug #128505), append --no-as-needed
app-admin/apache-tools/apache-tools-2.2.13.ebuild: append-ldflags
-Wl,--no-as-needed
app-admin/apache-tools/apache-tools-2.2.12.ebuild: # Instead of
filtering --as-needed (bug #128505), append --no-as-needed
app-admin/apache-tools/apache-tools-2.2.12.ebuild: append-ldflags
-Wl,--no-as-needed
app-editors/mlview/mlview-0.8-r1.ebuild: append-ldflags
-Wl,--no-as-needed
dev-games/ogre/files/ogre-1.4.9-as-needed.patch:+ ,-Xlinker
--no-as-needed -lstdc++)
dev-lang/gdl/gdl-0.9_rc2.ebuild: use imagemagick &&
append-ldflags -Wl,--no-as-needed
dev-libs/yaz++/yaz++-1.1.0.ebuild: append-ldflags
-Wl,--no-as-needed # FIXME.
dev-libs/dvmysql/dvmysql-1.0.2.ebuild: append-ldflags -Wl,--no-as-needed
dev-libs/dvutil/dvutil-0.15.5.ebuild: append-ldflags -Wl,--no-as-needed
dev-libs/dvutil/dvutil-1.0.5.ebuild: append-ldflags -Wl,--no-as-needed
dev-lisp/sbcl/sbcl-1.0.19.ebuild: append-ldflags
-Wl,--no-as-needed # see Bug #132992
dev-lisp/sbcl/sbcl-1.0.31.ebuild: append-ldflags
-Wl,--no-as-needed # see Bug #132992
dev-lisp/sbcl/sbcl-1.0.27-r10.ebuild: append-ldflags
-Wl,--no-as-needed # see Bug #132992
dev-lisp/sbcl/sbcl-1.0.26-r10.ebuild: append-ldflags
-Wl,--no-as-needed # see Bug #132992
dev-lisp/sbcl/sbcl-1.0.28.ebuild: append-ldflags
-Wl,--no-as-needed # see Bug #132992
eclass/apache-2.eclass: # Instead of filtering --as-needed (bug
#128505), append --no-as-needed
eclass/apache-2.eclass: append-ldflags -Wl,--no-as-needed
media-libs/SoQt/SoQt-1.4.1.ebuild: append-ldflags -Wl,--no-as-needed
media-sound/pulseaudio/pulseaudio-0.9.15-r2.ebuild: append-ldflags
-Wl,--no-as-needed
media-sound/pulseaudio/pulseaudio-0.9.18-r50.ebuild: append-ldflags
-Wl,--no-as-needed
media-sound/pulseaudio/pulseaudio-0.9.19.ebuild: append-ldflags
-Wl,--no-as-needed
media-sound/pulseaudio/pulseaudio-0.9.19-r50.ebuild: append-ldflags
-Wl,--no-as-needed
media-sound/pulseaudio/pulseaudio-0.9.18.ebuild: append-ldflags
-Wl,--no-as-needed
net-firewall/ebtables/ebtables-2.0.9.1.ebuild: append-ldflags
-Wl,--no-as-needed
net-firewall/ebtables/ebtables-2.0.8.2-r2.ebuild: append-ldflags
-Wl,--no-as-needed
net-libs/rb_libtorrent/rb_libtorrent-0.13-r1.ebuild: append-ldflags
-Wl,--no-as-needed
net-mail/email/email-3.0.5.ebuild: append-ldflags -Wl,--no-as-needed
sci-libs/mkl/mkl-10.0.2.018-r2.ebuild: Libs: -Wl,--no-as-needed
-L\${libdir} ${2} ${3} -lmkl_core ${4} -lpthread
sci-libs/mkl/mkl-10.0.2.018-r2.ebuild: Libs: -Wl,--no-as-needed
-L\${libdir} ${2} ${3} -lmkl_core ${4} -lpthread
sci-libs/mkl/mkl-10.0.2.018-r2.ebuild: Libs: -Wl,--no-as-needed
-L\${libdir} ${2} ${3} -lmkl_core -lmkl_lapack ${4} -lpthread
sci-libs/mkl/mkl-10.0.5.025.ebuild: Libs: -Wl,--no-as-needed
-L\${libdir} ${2} ${3} -lmkl_core ${4} -lpthread
sci-libs/mkl/mkl-10.0.5.025.ebuild: Libs: -Wl,--no-as-needed
-L\${libdir} ${2} ${3} -lmkl_core ${4} -lpthread
sci-libs/mkl/mkl-10.0.5.025.ebuild: Libs: -Wl,--no-as-needed
-L\${libdir} ${2} ${3} -lmkl_core -lmkl_lapack ${4} -lpthread
sys-devel/libtool/libtool-2.2.6a.ebuild: append-ldflags
-Wl,--no-as-needed
sys-libs/gwenhywfar/gwenhywfar-3.7.2.ebuild:append-ldflags
-Wl,--no-as-needed
sys-libs/gwenhywfar/gwenhywfar-3.8.0.ebuild:append-ldflags
-Wl,--no-as-needed
x11-wm/matchbox/matchbox-0.7.1.ebuild: append-ldflags -Wl,--no-as-needed

On 04-10-2009 13:13:30 +0300, Samuli Suominen wrote:
> +# @FUNCTION: no-as-needed
> +# @RETURN: Flag to disable asneeded behavior for use with append-ldflags.
> +no-as-needed() {
> + case $($(tc-getLD) -v 2>&1 </dev/null) in
> + *GNU*) # GNU ld
> + echo "-Wl,--no-as-needed" ;;
> + esac
> +}
>
> It's committed.
>
> To avoid the already caused trouble for you (prefix), these should be
> converted:

thanks!


--
Fabian Groffen
Gentoo on a different level

lör 2009-10-03 klockan 14:21 -0600 skrev Ryan Hill:
> On Sat, 03 Oct 2009 22:13:59 +0300
> Samuli Suominen <ssuominen RemoveThis @gentoo.org> wrote:
>
> > Since new binutils will support LD_AS_NEEDED="1" to force ld behave
> > asneeded we could use this for the developer -target in profiles?
> >
> > Speak up if you think it's a terrible idea.
> >
> > Thanks, Samuli
> >
> >
>
> I think it's a not terrible idea.
>
>
> Only barely related: can we enable FEATURES=test too?
>

For FEATURES="test" a policy for how to handle stuff like:

if use test; then
elog
ewarn "You have unit tests enabled, this results in an insecure library"
ewarn "It is recommended that you reinstall *without* FEATURES=test"
fi
[1]


needs to be formed before enabling it anywhere by default. Should this test really be run by default, or should it be shielded by USE="unsafetests" or something of just restricted?

My 2 cents.

[1] http://sources.gentoo.org/viewcvs.py/gentoo-x86/sys-apps/dbus/dbus-1.3...ebuild? and scroll to the bottom.

//Peter Hjalmarsson