FUNCTION_TRACER and parisc kernel

FUNCTION_TRACER and parisc kernel

[Helge Deller]

Hi Carlos, Dave & parisc-linux mailing list,

I just started looking into implementing function tracer functionality 
into the parisc kernel (http://cateee.net/lkddb/web-lkddb/FUNCTION_TRACER.html).

While doing so, I noticed that gcc issues lots of warnings like this one:
  CC      arch/parisc/lib/iomap.o
cc1: warning: -ffunction-sections disabled; it makes profiling impossible

So it seems that this statement in arch/parisc/Makefile:
# Without this, "ld -r" results in .text sections that are too big
# (> 0x40000) for branches to reach stubs.
cflags-y        += -ffunction-sections

breaks with the -pg compile option which is needed for function tracing.

Some searching with google brought up this thread:
http://gcc.gnu.org/ml/gcc-help/2008-11/msg00128.html
and esp. this answer:
http://gcc.gnu.org/ml/gcc-help/2008-11/msg00141.html

Or this thread:
http://osdir.com/ml/linux.debian.ports.hppa/2005-06/msg00045.html

Since it's late I felt too tired to look deeply into glibc and gcc myself and
just thought I should ask here: Do you see any chance that -pg and -ffunction-sections
could work together on hppa/linux?

Helge

PS: I assume my latest patches to fix loading of xfs modules only worked
because we compile the kernel with -ffunction-sections by default. Disabling
-ffunction-sections will then probably bring back the module loading bug... :-(

Re: FUNCTION_TRACER and parisc kernel

[John David Anglin]

> Hi Carlos, Dave & parisc-linux mailing list,
> 
> I just started looking into implementing function tracer functionality 
> into the parisc kernel (http://cateee.net/lkddb/web-lkddb/FUNCTION_TRACER.html).
> 
> While doing so, I noticed that gcc issues lots of warnings like this one:
>   CC      arch/parisc/lib/iomap.o
> cc1: warning: -ffunction-sections disabled; it makes profiling impossible
> 
> So it seems that this statement in arch/parisc/Makefile:
> # Without this, "ld -r" results in .text sections that are too big
> # (> 0x40000) for branches to reach stubs.
> cflags-y        += -ffunction-sections
> 
> breaks with the -pg compile option which is needed for function tracing.
> 
> Some searching with google brought up this thread:
> http://gcc.gnu.org/ml/gcc-help/2008-11/msg00128.html
> and esp. this answer:
> http://gcc.gnu.org/ml/gcc-help/2008-11/msg00141.html

Hmmm, we killed named section support on the HP-UX SOM port several years
ago because it wasn't very useful.  Jeff did it...  So, at this time, only
Solaris is an issue.  Of course, changing the code in toplev.c affects all
targets.  So, a change to remove this would have to go into 4.5 quite
early.

There was some work on the profiling implementation for both HP-UX and
Linux after the removal of named section support on HP-UX, so the label
issue may not be a problem.  Remove the code disabling function sections
when -pg is specified and see what happens.

[I have thought it would be useful to have named sections for dwarf2
debug support on HP-UX.  However, this is not useful for normal code.
I have a patch somewhere to do this, but gdb needs work which I've
never got around to.  So, I might bring named section support back
in a limited way.]

Regarding .text sections that are too big, see this thread:
http://sourceware.org/ml/binutils/2009-02/msg00061.html

It seems that it should be possible to do more fine grained stub
sections.  At this time, it's not fully clear where the problem
lies.  GCC may need to start a new .text section before each function,
or maybe the .text sections are getting merged.

Recently, I got pushed to look a bit at GNU ld.  It turns out there is no
long branch stub support in it at all.  As a result, it can't assemble
large applications like cc1plus.  It can just barely link cc1.  It
might have trouble with a large linux kernel.  I added a check recently
to generate an error if a branch target is out of range.  There's
a possibility that the current code for stub management in elf32-hppa.c
can be modified for hppa64 without too much work.

I don't know about the "ld -r" problem but I suspect we have to stop
.text sections from getting merged in partial links.  The other possiblity
would be to add long branch stubs for calls to undefined symbols.  The
latter approach would complicate final links.

Regards,
Dave
-- 
J. David Anglin                                  dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6602)

Re: FUNCTION_TRACER and parisc kernel

[John David Anglin]

> > So it seems that this statement in arch/parisc/Makefile:
> > # Without this, "ld -r" results in .text sections that are too big
> > # (> 0x40000) for branches to reach stubs.

This needs some debugging.  hppaelf_after_parse in ld/emultempl/hppaelf.em
is supposed to stop the 32-bit linker from merging .text sections on a
relocatable link.

GCC does not generate .text before each new function.  Even if it did,
gas currently merges all functions to a single .text section.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6602)

Re: FUNCTION_TRACER and parisc kernel

[Helge Deller]

John David Anglin wrote:
>> Hi Carlos, Dave & parisc-linux mailing list,
>>
>> I just started looking into implementing function tracer functionality 
>> into the parisc kernel (http://cateee.net/lkddb/web-lkddb/FUNCTION_TRACER.html).
>>
>> While doing so, I noticed that gcc issues lots of warnings like this one:
>>   CC      arch/parisc/lib/iomap.o
>> cc1: warning: -ffunction-sections disabled; it makes profiling impossible
>>
>> So it seems that this statement in arch/parisc/Makefile:
>> # Without this, "ld -r" results in .text sections that are too big
>> # (> 0x40000) for branches to reach stubs.
>> cflags-y        += -ffunction-sections
>>
>> breaks with the -pg compile option which is needed for function tracing.
>>
>> Some searching with google brought up this thread:
>> http://gcc.gnu.org/ml/gcc-help/2008-11/msg00128.html
>> and esp. this answer:
>> http://gcc.gnu.org/ml/gcc-help/2008-11/msg00141.html
> 
> Hmmm, we killed named section support on the HP-UX SOM port several years
> ago because it wasn't very useful.  Jeff did it...  So, at this time, only
> Solaris is an issue.  Of course, changing the code in toplev.c affects all
> targets.  So, a change to remove this would have to go into 4.5 quite
> early.
> 
> There was some work on the profiling implementation for both HP-UX and
> Linux after the removal of named section support on HP-UX, so the label
> issue may not be a problem.  Remove the code disabling function sections
> when -pg is specified and see what happens.

It does not link when building 32bit or PA20 code.
On 64bit it does build and link, although taking the patch from 
http://sourceware.org/ml/binutils/2009-02/msg00040.html
into account, as well as me running an old ld, it might be that
I just don't see it failing yet...

> [I have thought it would be useful to have named sections for dwarf2
> debug support on HP-UX.  However, this is not useful for normal code.
> I have a patch somewhere to do this, but gdb needs work which I've
> never got around to.  So, I might bring named section support back
> in a limited way.]

Ok. This means it would be possible to drop function-sections.

> Regarding .text sections that are too big, see this thread:
> http://sourceware.org/ml/binutils/2009-02/msg00061.html
> 
> It seems that it should be possible to do more fine grained stub
> sections.  At this time, it's not fully clear where the problem
> lies.  GCC may need to start a new .text section before each function,
> or maybe the .text sections are getting merged.
> 
> Recently, I got pushed to look a bit at GNU ld.  It turns out there is no
> long branch stub support in it at all.  As a result, it can't assemble
> large applications like cc1plus.  It can just barely link cc1.  It
> might have trouble with a large linux kernel.  I added a check recently
> to generate an error if a branch target is out of range.  There's
> a possibility that the current code for stub management in elf32-hppa.c
> can be modified for hppa64 without too much work.

That would be great.

> I don't know about the "ld -r" problem but I suspect we have to stop
> .text sections from getting merged in partial links.  The other possiblity
> would be to add long branch stubs for calls to undefined symbols.  The
> latter approach would complicate final links.

Thanks a lot for those explanations!!!
I see I need to dig in closer into ld/gcc as well too :-(

Helge