[Buildroot] Problem building latest arm

[Buildroot] Problem building latest arm

[michael.miller at prga.com]

I'm trying to build for an Xscale (ARM), EABI, NPTL with binutils
2.17.50.0.16, software floating point, and target options bash, bzip2
and diffutils.  All other options are default (gcc 4.2.0).  I'm using
the latest snapshot for uclibc, and have tried buildroot snapshots from
6/20, 6/21 and 6/23.

If I use 6/20, 6/21 or 6/23 buildroot snapshots then the build fails in
pthreads because sysdep-cancel.h is not found.  Looking in the
linuxthreads/sysdeps/unix/sysv/linux directory, I see subdirectories for
i386, mips, etc. but not for arm.

Searching recent mailing list archives, Bernhard says things that make
me think this configuration should work.  He hints that helptext will
give me great insight, probably in the GNU_TARGET_SUFFIX area.  I
gleaned this from a reply to an inquiry by Stuart Wood.  Unfortunately,
I don't seem to be able to make use of the hint.  My target suffix is
linux-uclibcgnueabi, which is the default.  If I interpret Bernhard's
statement correctly, I should be using the default?

The build system is a Mandriva 2007 workstation, which uses gcc 4.1.2.

Any ideas how to proceed?  Should I create a sysdep-cancel.h for ARM?

	Regards,
	Mike

[Buildroot] Problem building latest arm

[Bernhard Fischer]

On Mon, Jun 25, 2007 at 10:04:44PM -0500, michael.miller at prga.com wrote:
>I'm trying to build for an Xscale (ARM), EABI, NPTL with binutils
>2.17.50.0.16, software floating point, and target options bash, bzip2
>and diffutils.  All other options are default (gcc 4.2.0).  I'm using
>the latest snapshot for uclibc, and have tried buildroot snapshots from
>6/20, 6/21 and 6/23.
>
>If I use 6/20, 6/21 or 6/23 buildroot snapshots then the build fails in
>pthreads because sysdep-cancel.h is not found.  Looking in the
>linuxthreads/sysdeps/unix/sysv/linux directory, I see subdirectories for
>i386, mips, etc. but not for arm.
>
>Searching recent mailing list archives, Bernhard says things that make
>me think this configuration should work.  He hints that helptext will

I was talking about eabi. Threading is a different issue (but even
that one has a nice, IMHO descriptive help-text).

[Buildroot] Problem building latest arm

[michael.miller at prga.com]

>-----Original Message-----
>From: Bernhard Fischer [mailto:rep.dot.nop at gmail.com] 
>Sent: Tuesday, June 26, 2007 1:22 AM
>To: Michael D. Miller
>Cc: buildroot at uclibc.org
>Subject: Re: [Buildroot] Problem building latest arm
>
>On Mon, Jun 25, 2007 at 10:04:44PM -0500, michael.miller at prga.com
wrote:
>>I'm trying to build for an Xscale (ARM), EABI, NPTL with binutils 
>>2.17.50.0.16, software floating point, and target options bash, bzip2 
>>and diffutils.  All other options are default (gcc 4.2.0).  I'm using 
>>the latest snapshot for uclibc, and have tried buildroot snapshots
from 
>>6/20, 6/21 and 6/23.
>>
>>If I use 6/20, 6/21 or 6/23 buildroot snapshots then the build fails
in 
>>pthreads because sysdep-cancel.h is not found.  Looking in the 
>>linuxthreads/sysdeps/unix/sysv/linux directory, I see subdirectories 
>>for i386, mips, etc. but not for arm.
>>
>>Searching recent mailing list archives, Bernhard says things that make

>>me think this configuration should work.  He hints that helptext will
>
>I was talking about eabi. Threading is a different issue (but even that
one has a nice, >IMHO descriptive help-text).

I tried just building for "old" Linux threads instead of POSIX threads.
In that case I got an error because build_arm_nofpu/staging_dir/include
didn't exist.  I manually created the directory and restarted the build.
Things continued without further incident and completed with a tool
chain.  A quick test seems to indicate the generated cross compiler
works, checked with a simple "Hello world" compiled after exporting the
staging_dir/usr/bin.

So, NPTL seems broken for ARM.  I've tried using an old ecos version of
sysdep-cancel.h and things reach the point that during the gcc-final
stage libgomp blows up because the pthread library can't be found.  This
is a configure error saying that pthreads are required for libgomp.  I'd
say I don't even want libgomp.  Is there an easy way to prevent this
from even being attempted?  I'm thinking of some additional build flag I
can pass in to buildroot.

	Regards,
	Mike

[Buildroot] Problem building latest arm

[Bernhard Fischer]

On Wed, Jun 27, 2007 at 07:45:02AM -0500, michael.miller at prga.com wrote:
>>-----Original Message-----
>>From: Bernhard Fischer [mailto:rep.dot.nop at gmail.com] 
>>Sent: Tuesday, June 26, 2007 1:22 AM
>>To: Michael D. Miller
>>Cc: buildroot at uclibc.org
>>Subject: Re: [Buildroot] Problem building latest arm
>>
>>On Mon, Jun 25, 2007 at 10:04:44PM -0500, michael.miller at prga.com
>wrote:
>>>I'm trying to build for an Xscale (ARM), EABI, NPTL with binutils 
>>>2.17.50.0.16, software floating point, and target options bash, bzip2 
>>>and diffutils.  All other options are default (gcc 4.2.0).  I'm using 
>>>the latest snapshot for uclibc, and have tried buildroot snapshots
>from 
>>>6/20, 6/21 and 6/23.
>>>
>>>If I use 6/20, 6/21 or 6/23 buildroot snapshots then the build fails
>in 
>>>pthreads because sysdep-cancel.h is not found.  Looking in the 
>>>linuxthreads/sysdeps/unix/sysv/linux directory, I see subdirectories 
>>>for i386, mips, etc. but not for arm.
>>>
>>>Searching recent mailing list archives, Bernhard says things that make
>
>>>me think this configuration should work.  He hints that helptext will
>>
>>I was talking about eabi. Threading is a different issue (but even that
>one has a nice, >IMHO descriptive help-text).
>
>I tried just building for "old" Linux threads instead of POSIX threads.
>In that case I got an error because build_arm_nofpu/staging_dir/include

this area changed recently. You may need to erase build_* toolchain_*

>didn't exist.  I manually created the directory and restarted the build.
>Things continued without further incident and completed with a tool
>chain.  A quick test seems to indicate the generated cross compiler
>works, checked with a simple "Hello world" compiled after exporting the
>staging_dir/usr/bin.
>
>So, NPTL seems broken for ARM.  I've tried using an old ecos version of

you're better off to ask the uclibc list about the state of NTPL
support.
>sysdep-cancel.h and things reach the point that during the gcc-final
>stage libgomp blows up because the pthread library can't be found.  This
>is a configure error saying that pthreads are required for libgomp.  I'd
>say I don't even want libgomp.  Is there an easy way to prevent this
>from even being attempted?  I'm thinking of some additional build flag I
>can pass in to buildroot.

See menuconfig. ()  Additional gcc options

[Buildroot] Problem building latest arm

[michael.miller at prga.com]

>-----Original Message-----
>From: Bernhard Fischer [mailto:rep.dot.nop at gmail.com] 
>Sent: Wednesday, June 27, 2007 11:20 AM
>To: Michael D. Miller
>Cc: buildroot at uclibc.org
>Subject: Re: [Buildroot] Problem building latest arm
>
>On Wed, Jun 27, 2007 at 07:45:02AM -0500, michael.miller at prga.com
wrote:
>>>-----Original Message-----
>>>From: Bernhard Fischer [mailto:rep.dot.nop at gmail.com]
>>>Sent: Tuesday, June 26, 2007 1:22 AM
>>>To: Michael D. Miller
>>>Cc: buildroot at uclibc.org
>>>Subject: Re: [Buildroot] Problem building latest arm
>>>
>>>On Mon, Jun 25, 2007 at 10:04:44PM -0500, michael.miller at prga.com
>>wrote:
>>>>I'm trying to build for an Xscale (ARM), EABI, NPTL with binutils 
>>>>2.17.50.0.16, software floating point, and target options bash,
bzip2 
>>>>and diffutils.  All other options are default (gcc 4.2.0).  I'm
using 
>>>>the latest snapshot for uclibc, and have tried buildroot snapshots
>>>from
>>>>6/20, 6/21 and 6/23.
>>>>
>>>>If I use 6/20, 6/21 or 6/23 buildroot snapshots then the build fails
>>in
>>>>pthreads because sysdep-cancel.h is not found.  Looking in the 
>>>>linuxthreads/sysdeps/unix/sysv/linux directory, I see subdirectories

>>>>for i386, mips, etc. but not for arm.
>>>>
>>>>Searching recent mailing list archives, Bernhard says things that 
>>>>make
>>
>>>>me think this configuration should work.  He hints that helptext
will
>>>
>>>I was talking about eabi. Threading is a different issue (but even 
>>>that
>>one has a nice, >IMHO descriptive help-text).
>>
>>I tried just building for "old" Linux threads instead of POSIX
threads.
>>In that case I got an error because
build_arm_nofpu/staging_dir/include
>
>this area changed recently. You may need to erase build_* toolchain_*

I've been doing clean builds by removing the entire buildroot directory
and starting over each time.

>>didn't exist.  I manually created the directory and restarted the
build.
>>Things continued without further incident and completed with a tool 
>>chain.  A quick test seems to indicate the generated cross compiler 
>>works, checked with a simple "Hello world" compiled after exporting
the 
>>staging_dir/usr/bin.
>>
>>So, NPTL seems broken for ARM.  I've tried using an old ecos version
of
>
>you're better off to ask the uclibc list about the state of NTPL
support.

This really seems more of a buildroot issue since the problem is the
gcc-final stage.  I'll query the uclibc list if you think it might help,
but I expect I'll get one of Rob's famous flames for being off topic.

I'm not sure why the pthread test is failing, so I'll poke around a
little more there.  A check via -dumpspecs of xgcc indicates that the
pthread support is there.  Maybe configure is looking for the library.

>>sysdep-cancel.h and things reach the point that during the gcc-final 
>>stage libgomp blows up because the pthread library can't be found.  
>>This is a configure error saying that pthreads are required for 
>>libgomp.  I'd say I don't even want libgomp.  Is there an easy way to 
>>prevent this from even being attempted?  I'm thinking of some 
>>additional build flag I can pass in to buildroot.
>
>See menuconfig. ()  Additional gcc options

On more reflection, I can't decide for sure if libgomp support is needed
in the final gcc.  It seems to me that OpenMP and POSIX threads are
different beasts.  I can understand OpenMP requiring pthreads, but it
doesn't' seem like pthreads should really require OpenMP.  That's a vote
for getting libgomp out of the final gcc.  We're talking about the cross
tool that runs on a x686 platform here, not the final ARM target.  I'll
give --disable-libgomp a try and see if that gets me further.

	Regards,
	Mike

[Buildroot] Problem building latest arm

[michael.miller at prga.com]

Bernhard:

[a lot of snips]

I had reported problems building pthreads/ARM/EABI.

Short summary:
1) it looks like an actual test program is compiled/linked to determine
pthread support and the pthread library wasn't found in the final gcc
step.  I don't know why, since the library seems to have build OK.

2) adding --disable-libgomp to the gcc flags in buildroot resulted in a
successful build of a tool chain.  This is really more of a workaround
than addressing the underlying build issue.
  
	Regards,
	Mike