Dynamic linker name

Dynamic linker name

[Lei Zhang]

Hi,

I'm trying to make a few changes to clang/LLVM to make it work with
musl, i.e. link executables against musl instead of glibc.

One of the changes involved is to tell clang where to find musl's
dynamic linker. My question is: is musl's dynamic linker always named
"ld-musl-$ARCH.so.1" and put under /lib?

More specifically, on a multilib system, which of the following is encouraged?

1. /lib/ld-musl-{x86,x86_64}.so.1
2. /lib32/ld-musl-x86.so.1 and /lib64/ld-musl-x86_64.so.1


Thanks,
Lei

Re: Dynamic linker name

[Rich Felker]

On Mon, Jun 06, 2016 at 09:47:10AM +0800, Lei Zhang wrote:
> Hi,
> 
> I'm trying to make a few changes to clang/LLVM to make it work with
> musl, i.e. link executables against musl instead of glibc.
> 
> One of the changes involved is to tell clang where to find musl's
> dynamic linker. My question is: is musl's dynamic linker always named
> "ld-musl-$ARCH.so.1" and put under /lib?
> 
> More specifically, on a multilib system, which of the following is encouraged?
> 
> 1. /lib/ld-musl-{x86,x86_64}.so.1
> 2. /lib32/ld-musl-x86.so.1 and /lib64/ld-musl-x86_64.so.1

It's always /lib/, not /lib32/ or /lib64/.

Rich

Re: Dynamic linker name

[Khem Raj]

[-- Attachment #1: Type: text/plain, Size: 729 bytes --]

On Jun 5, 2016 6:47 PM, "Lei Zhang" <zhanglei.april@gmail.com> wrote:
>
> Hi,
>
> I'm trying to make a few changes to clang/LLVM to make it work with
> musl, i.e. link executables against musl instead of glibc.
>
> One of the changes involved is to tell clang where to find musl's
> dynamic linker. My question is: is musl's dynamic linker always named
> "ld-musl-$ARCH.so.1" and put under /lib?
>
> More specifically, on a multilib system, which of the following is
encouraged?
>
> 1. /lib/ld-musl-{x86,x86_64}.so.1
> 2. /lib32/ld-musl-x86.so.1 and /lib64/ld-musl-x86_64.so.1

There is a llvm clang patch for this already. Look at

https://github.com/kraj/meta-clang/tree/master/recipes-devtools/clang/clang
>
>
> Thanks,
> Lei

[-- Attachment #2: Type: text/html, Size: 1088 bytes --]

Re: Dynamic linker name

[Szabolcs Nagy]

* Khem Raj <raj.khem@gmail.com> [2016-06-14 07:43:34 -0700]:
> On Jun 5, 2016 6:47 PM, "Lei Zhang" <zhanglei.april@gmail.com> wrote:
> > One of the changes involved is to tell clang where to find musl's
> > dynamic linker. My question is: is musl's dynamic linker always named
> > "ld-musl-$ARCH.so.1" and put under /lib?
> 
> There is a llvm clang patch for this already. Look at
> 
> https://github.com/kraj/meta-clang/tree/master/recipes-devtools/clang/clang

i think these should be upstreamed into clang and
the recent r272662 clang commit should be reverted
and the existing mips musl hacks should be fixed too.

Re: Dynamic linker name

[Khem Raj]

[-- Attachment #1: Type: text/plain, Size: 756 bytes --]

On Jun 14, 2016 8:04 AM, "Szabolcs Nagy" <nsz@port70.net> wrote:
>
> * Khem Raj <raj.khem@gmail.com> [2016-06-14 07:43:34 -0700]:
> > On Jun 5, 2016 6:47 PM, "Lei Zhang" <zhanglei.april@gmail.com> wrote:
> > > One of the changes involved is to tell clang where to find musl's
> > > dynamic linker. My question is: is musl's dynamic linker always named
> > > "ld-musl-$ARCH.so.1" and put under /lib?
> >
> > There is a llvm clang patch for this already. Look at
> >
> >
https://github.com/kraj/meta-clang/tree/master/recipes-devtools/clang/clang
>
> i think these should be upstreamed into clang and
> the recent r272662 clang commit should be reverted
> and the existing mips musl hacks should be fixed too.

I planned to do that but was off for few weeks

[-- Attachment #2: Type: text/html, Size: 1198 bytes --]

Re: Dynamic linker name

[Rafael Espíndola]

[-- Attachment #1: Type: text/plain, Size: 891 bytes --]

Sorry, but my understanding is that r272662 is correct for x86, no?

We can definitely add support for more arches with musl if someone sends a
patch to the list.

Cheers,
Rafael
On Jun 14, 2016 11:04 AM, "Szabolcs Nagy" <nsz@port70.net> wrote:

> * Khem Raj <raj.khem@gmail.com> [2016-06-14 07:43:34 -0700]:
> > On Jun 5, 2016 6:47 PM, "Lei Zhang" <zhanglei.april@gmail.com> wrote:
> > > One of the changes involved is to tell clang where to find musl's
> > > dynamic linker. My question is: is musl's dynamic linker always named
> > > "ld-musl-$ARCH.so.1" and put under /lib?
> >
> > There is a llvm clang patch for this already. Look at
> >
> >
> https://github.com/kraj/meta-clang/tree/master/recipes-devtools/clang/clang
>
> i think these should be upstreamed into clang and
> the recent r272662 clang commit should be reverted
> and the existing mips musl hacks should be fixed too.
>

[-- Attachment #2: Type: text/html, Size: 1488 bytes --]

Re: Dynamic linker name

[Lei Zhang]

2016-06-14 23:11 GMT+08:00 Rafael Espíndola <rafael.espindola@gmail.com>:
> Sorry, but my understanding is that r272662 is correct for x86, no?
>
> We can definitely add support for more arches with musl if someone sends a
> patch to the list.

If no one intends to do this, I can prepare further patches and
relevant test cases for other archs.


Lei

Re: Dynamic linker name

[Lei Zhang]

2016-06-14 23:56 GMT+08:00 Lei Zhang <zhanglei.april@gmail.com>:
> 2016-06-14 23:11 GMT+08:00 Rafael Espíndola <rafael.espindola@gmail.com>:
>> Sorry, but my understanding is that r272662 is correct for x86, no?
>>
>> We can definitely add support for more arches with musl if someone sends a
>> patch to the list.
>
> If no one intends to do this, I can prepare further patches and
> relevant test cases for other archs.

Now I'm determining the $ARCH field in musl's dynamic linker name for
non-x86 archs. After reading these two patches (thanks to their
authors):

http://git.alpinelinux.org/cgit/aports/plain/main/clang/clang-0004-Add-musl-targets-and-dynamic-linker.patch
https://github.com/kraj/meta-clang/blob/master/recipes-devtools/clang/clang/0001-clang-driver-Add-musl-ldso-support.patch

I got the following list:

i386
x86_64
arm
armhf
armeb
armebhf
aarch64
aarch64_be
mips
mipsel
mips64
mipsel64el // this one looks weird
powerpc
powerpc64

Is this list comprehensive? Is there anything wrong? I have no access
to non-x86 machines at the moment, thus no way to verify them.

BTW, I leaved out x32 on purpose, since musl's support for it is
experimental (right?), and it requires yet another environment type
"muslx32" in LLVM.


Thanks,
Lei

Re: Dynamic linker name

[Felix Janda]

Lei Zhang wrote:
> 2016-06-14 23:56 GMT+08:00 Lei Zhang <zhanglei.april@gmail.com>:
> > 2016-06-14 23:11 GMT+08:00 Rafael Espíndola <rafael.espindola@gmail.com>:
> >> Sorry, but my understanding is that r272662 is correct for x86, no?
> >>
> >> We can definitely add support for more arches with musl if someone sends a
> >> patch to the list.
> >
> > If no one intends to do this, I can prepare further patches and
> > relevant test cases for other archs.
> 
> Now I'm determining the $ARCH field in musl's dynamic linker name for
> non-x86 archs. After reading these two patches (thanks to their
> authors):
> 
> http://git.alpinelinux.org/cgit/aports/plain/main/clang/clang-0004-Add-musl-targets-and-dynamic-linker.patch
> https://github.com/kraj/meta-clang/blob/master/recipes-devtools/clang/clang/0001-clang-driver-Add-musl-ldso-support.patch

You should really take a look at musl's configure script.

> I got the following list:
> 
> i386
> x86_64
> arm
> armhf
> armeb
> armebhf
> aarch64
> aarch64_be
> mips
> mipsel
> mips64
> mipsel64el // this one looks weird
> powerpc
> powerpc64
> 
> Is this list comprehensive? Is there anything wrong? I have no access
> to non-x86 machines at the moment, thus no way to verify them.

As mentioned by Szabolcs Nagy, the patches are not correct for mips:
It should be "mips64el".

For mips* and powerpc* there are also "-sf" variants. There are also
"mipsr6*" and "mipsn32*" which might or might not be supported by llvm.

> BTW, I leaved out x32 on purpose, since musl's support for it is
> experimental (right?), and it requires yet another environment type
> "muslx32" in LLVM.

x32 support has been in musl for many releases, whereas mips64 has
not landed in a released version of musl, yet...

> 
> Thanks,
> Lei

Thanks for moving forward and getting actually something merged.

Felix

Re: Dynamic linker name

[Lei Zhang]

2016-06-15 15:10 GMT+08:00 Felix Janda <felix.janda@posteo.de>:
> Lei Zhang wrote:
>> 2016-06-14 23:56 GMT+08:00 Lei Zhang <zhanglei.april@gmail.com>:
>> > 2016-06-14 23:11 GMT+08:00 Rafael Espíndola <rafael.espindola@gmail.com>:
>> >> Sorry, but my understanding is that r272662 is correct for x86, no?
>> >>
>> >> We can definitely add support for more arches with musl if someone sends a
>> >> patch to the list.
>> >
>> > If no one intends to do this, I can prepare further patches and
>> > relevant test cases for other archs.
>>
>> Now I'm determining the $ARCH field in musl's dynamic linker name for
>> non-x86 archs. After reading these two patches (thanks to their
>> authors):
>>
>> http://git.alpinelinux.org/cgit/aports/plain/main/clang/clang-0004-Add-musl-targets-and-dynamic-linker.patch
>> https://github.com/kraj/meta-clang/blob/master/recipes-devtools/clang/clang/0001-clang-driver-Add-musl-ldso-support.patch
>
> You should really take a look at musl's configure script.

I should've done that. Sorry for my laziness...

> For mips* and powerpc* there are also "-sf" variants. There are also
> "mipsr6*" and "mipsn32*" which might or might not be supported by llvm.

From LLVM's code, it seems these variants aren't supported.

>> BTW, I leaved out x32 on purpose, since musl's support for it is
>> experimental (right?), and it requires yet another environment type
>> "muslx32" in LLVM.
>
> x32 support has been in musl for many releases, whereas mips64 has
> not landed in a released version of musl, yet...

So now I have this list:

i386
x86_64
powerpc
powerpc64
mips
mipsel
mips64
mips64el

which should just work out the box with my previous patch. I'll take
care of x32 and ARM (subarchs and ABIs) with further patches.

Lei

Re: Dynamic linker name

[Luca Barbato]

On 15/06/16 10:16, Lei Zhang wrote:
> 2016-06-15 15:10 GMT+08:00 Felix Janda <felix.janda@posteo.de>:
>> Lei Zhang wrote:
>>> 2016-06-14 23:56 GMT+08:00 Lei Zhang <zhanglei.april@gmail.com>:
>>>> 2016-06-14 23:11 GMT+08:00 Rafael Espíndola <rafael.espindola@gmail.com>:
>>>>> Sorry, but my understanding is that r272662 is correct for x86, no?
>>>>>
>>>>> We can definitely add support for more arches with musl if someone sends a
>>>>> patch to the list.
>>>>
>>>> If no one intends to do this, I can prepare further patches and
>>>> relevant test cases for other archs.
>>>
>>> Now I'm determining the $ARCH field in musl's dynamic linker name for
>>> non-x86 archs. After reading these two patches (thanks to their
>>> authors):
>>>
>>> http://git.alpinelinux.org/cgit/aports/plain/main/clang/clang-0004-Add-musl-targets-and-dynamic-linker.patch
>>> https://github.com/kraj/meta-clang/blob/master/recipes-devtools/clang/clang/0001-clang-driver-Add-musl-ldso-support.patch
>>
>> You should really take a look at musl's configure script.
> 
> I should've done that. Sorry for my laziness...
> 
>> For mips* and powerpc* there are also "-sf" variants. There are also
>> "mipsr6*" and "mipsn32*" which might or might not be supported by llvm.
> 
> From LLVM's code, it seems these variants aren't supported.
> 
>>> BTW, I leaved out x32 on purpose, since musl's support for it is
>>> experimental (right?), and it requires yet another environment type
>>> "muslx32" in LLVM.
>>
>> x32 support has been in musl for many releases, whereas mips64 has
>> not landed in a released version of musl, yet...
> 
> So now I have this list:
> 
> i386
> x86_64
> powerpc
> powerpc64
> mips
> mipsel
> mips64
> mips64el
> 
> which should just work out the box with my previous patch. I'll take
> care of x32 and ARM (subarchs and ABIs) with further patches.

Powerpc can be little endian as well but it can wait the arm patch as well.

lu

Re: Dynamic linker name

[Lei Zhang]

2016-06-15 17:31 GMT+08:00 Luca Barbato <lu_zero@gentoo.org>:
> On 15/06/16 10:16, Lei Zhang wrote:
>> i386
>> x86_64
>> powerpc
>> powerpc64
>> mips
>> mipsel
>> mips64
>> mips64el
>>
>> which should just work out the box with my previous patch. I'll take
>> care of x32 and ARM (subarchs and ABIs) with further patches.
>
> Powerpc can be little endian as well but it can wait the arm patch as well.

I've only seen powerpc64le; don't know if 32-bit powerpc doesn't
support little endian.


Lei

Re: Dynamic linker name

[Bobby Bingham]

On Wed, Jun 15, 2016 at 05:39:51PM +0800, Lei Zhang wrote:
> 2016-06-15 17:31 GMT+08:00 Luca Barbato <lu_zero@gentoo.org>:
> > On 15/06/16 10:16, Lei Zhang wrote:
> >> i386
> >> x86_64
> >> powerpc
> >> powerpc64
> >> mips
> >> mipsel
> >> mips64
> >> mips64el
> >>
> >> which should just work out the box with my previous patch. I'll take
> >> care of x32 and ARM (subarchs and ABIs) with further patches.
> >
> > Powerpc can be little endian as well but it can wait the arm patch as well.
> 
> I've only seen powerpc64le; don't know if 32-bit powerpc doesn't
> support little endian.

There are a couple things you should know about musl's support for
powerpc64 which mean you probably shouldn't add it.

Two different ABIs exist for powerpc64.  The new one is sometimes called
powerpc64le.  While it's typically used on little endian, there's no
reason it can't be used on big endian systems as well.  Musl only
supports the new ABI, but supports it for both big and little endian.
The arch names used by musl for them are:

powerpc64
powerpc64le

Some software out there assumes that big endian powerpc means the old ABI.
I'm not sure if clang supports big endian powerpc64 with the new ABI or
not, but it needs to if it's to target big endian powerpc64 musl.

The larger problem for both endians is long double support.  The new ABI
calls for a 128-bit long double, either IEEE quad, or IBM double-double.
Musl will not support double-double due to the different floating point
semantics it provides, and gcc's support for IEEE quad is not fully usable
on powerpc64 yet.  Therefore, musl currently uses a 64-bit long double on
powerpc64.

Last I checked, clang only supported using IBM double-double for long
double on powerpc64.  If this is still true, then clang cannot match
musl's ABI on powerpc64.

Bobby

Re: Dynamic linker name

[Lei Zhang]

2016-06-15 21:00 GMT+08:00 Bobby Bingham <koorogi@koorogi.info>:
> On Wed, Jun 15, 2016 at 05:39:51PM +0800, Lei Zhang wrote:
>> 2016-06-15 17:31 GMT+08:00 Luca Barbato <lu_zero@gentoo.org>:
>> > On 15/06/16 10:16, Lei Zhang wrote:
>> >> i386
>> >> x86_64
>> >> powerpc
>> >> powerpc64
>> >> mips
>> >> mipsel
>> >> mips64
>> >> mips64el
>> >>
>> >> which should just work out the box with my previous patch. I'll take
>> >> care of x32 and ARM (subarchs and ABIs) with further patches.
>> >
>> > Powerpc can be little endian as well but it can wait the arm patch as well.
>>
>> I've only seen powerpc64le; don't know if 32-bit powerpc doesn't
>> support little endian.
>
> There are a couple things you should know about musl's support for
> powerpc64 which mean you probably shouldn't add it.
>
> Two different ABIs exist for powerpc64.  The new one is sometimes called
> powerpc64le.  While it's typically used on little endian, there's no
> reason it can't be used on big endian systems as well.  Musl only
> supports the new ABI, but supports it for both big and little endian.
> The arch names used by musl for them are:

I just did a quick search; are these two ABIs called "ppc64 elf" and
"ppc64 elf v2" respectively?

> Last I checked, clang only supported using IBM double-double for long
> double on powerpc64.  If this is still true, then clang cannot match
> musl's ABI on powerpc64.

I think powerpc64 needs no special handling anyway (unlike ARM); it
just happens to be supported by the patch. OTOH, do you think I should
*explicitly* disallow clang being configured against musl when
targeting powerpc64? I can put that logic in my next patch if it makes
sense.


Thanks,
Lei

Re: Dynamic linker name

[Khem Raj]

On Wed, Jun 15, 2016 at 6:00 AM, Bobby Bingham <koorogi@koorogi.info> wrote:
> On Wed, Jun 15, 2016 at 05:39:51PM +0800, Lei Zhang wrote:
>> 2016-06-15 17:31 GMT+08:00 Luca Barbato <lu_zero@gentoo.org>:
>> > On 15/06/16 10:16, Lei Zhang wrote:
>> >> i386
>> >> x86_64
>> >> powerpc
>> >> powerpc64
>> >> mips
>> >> mipsel
>> >> mips64
>> >> mips64el
>> >>
>> >> which should just work out the box with my previous patch. I'll take
>> >> care of x32 and ARM (subarchs and ABIs) with further patches.
>> >
>> > Powerpc can be little endian as well but it can wait the arm patch as well.
>>
>> I've only seen powerpc64le; don't know if 32-bit powerpc doesn't
>> support little endian.
>
> There are a couple things you should know about musl's support for
> powerpc64 which mean you probably shouldn't add it.
>
> Two different ABIs exist for powerpc64.  The new one is sometimes called
> powerpc64le.  While it's typically used on little endian, there's no
> reason it can't be used on big endian systems as well.  Musl only
> supports the new ABI, but supports it for both big and little endian.
> The arch names used by musl for them are:
>
> powerpc64
> powerpc64le
>
> Some software out there assumes that big endian powerpc means the old ABI.
> I'm not sure if clang supports big endian powerpc64 with the new ABI or
> not, but it needs to if it's to target big endian powerpc64 musl.
>
> The larger problem for both endians is long double support.  The new ABI
> calls for a 128-bit long double, either IEEE quad, or IBM double-double.
> Musl will not support double-double due to the different floating point
> semantics it provides, and gcc's support for IEEE quad is not fully usable
> on powerpc64 yet.  Therefore, musl currently uses a 64-bit long double on
> powerpc64.
>
> Last I checked, clang only supported using IBM double-double for long
> double on powerpc64.  If this is still true, then clang cannot match
> musl's ABI on powerpc64.

Thats right. I have fixed it in clang here
https://github.com/kraj/clang/commit/3c3627dbd4358c8e97c95a614e63803cc9048fa6

>
> Bobby