does linux -fPIC/eh do so poorly?

does linux -fPIC/eh do so poorly?

[Robert Lipe]

Can one of you guys with a current EGCS  on Linux try something for me?
I haven't been able to do a 'make check' for over a month, and I'm trying
to figure out if something is broken on other x86 elf/unix targets, or if
it's unique to OpenServer.   I also can't tell if I'm just the only one
masochistic enough to make another pass over all the testsuites with -fPIC
or not.  When I last ran the tests around mid Jan, this didn't happen.

In your gcc directory, type

	 make check-g++ RUNTESTFLAGS="--tool_opts -fPIC old-deja.exp=eh*"

                === g++ Summary ===

# of expected passes            66
# of unexpected failures        37
/home/play/negcs/gcc/testsuite/../xgcc version egcs-2.91.08 980214 (gcc-2.8.0 re
lease)



If I lose the -fPIC, I see much better numbers.

	                === g++ Summary ===

# of expected passes            103


It looks like all the EH tests with -fPIC drop core.


-- 
Robert Lipe       http://www.dgii.com/people/robertl       robertl@dgii.com

Re: does linux -fPIC/eh do so poorly?

[H.J. Lu]

> 
> In your gcc directory, type
> 
> 	 make check-g++ RUNTESTFLAGS="--tool_opts -fPIC old-deja.exp=eh*"
> 
>                 === g++ Summary ===
> 
> # of expected passes            66
> # of unexpected failures        37
> /home/play/negcs/gcc/testsuite/../xgcc version egcs-2.91.08 980214 (gcc-2.8.0 re
> lease)
> 
> 

# make check-g++ RUNTESTFLAGS="--tool_opts -fPIC old-deja.exp=eh*"

                === g++ Summary ===

# of expected passes            102
# of untested testcases         1
/home/work/gnu/bin/egcs/gcc/testsuite/../xgcc version egcs-2.90.24 980208 (egcs-1.0.2 prerelease)

It looks fine on linux. I have been working with egcs 1.0.2. I haven't
built egcs for a few days now.


-- 
H.J. Lu (hjl@gnu.org)

Re: does linux -fPIC/eh do so poorly?

[Robert Lipe]

[ OpenServer, EGCS from 0216 ] 
> > 	 make check-g++ RUNTESTFLAGS="--tool_opts -fPIC old-deja.exp=eh*"
> > 
> >                 === g++ Summary ===
> > 
> > # of expected passes            66
> > # of unexpected failures        37


> # make check-g++ RUNTESTFLAGS="--tool_opts -fPIC old-deja.exp=eh*"
> 
>                 === g++ Summary ===
> 
> # of expected passes            102
> # of untested testcases         1
> /home/work/gnu/bin/egcs/gcc/testsuite/../xgcc version egcs-2.90.24 980208 (egcs-1.0.2 prerelease)
> 
> It looks fine on linux. I have been working with egcs 1.0.2. I haven't
> built egcs for a few days now.


Thanx for the sanity test, H.J.


When I do the same in my 1.0.2 tree, I almost see the same thing you see.

                === g++ Summary ===

# of expected passes            103
/home/play/egcs-release/native/gcc/testsuite/../xgcc version egcs-2.90.24 980208
 (egcs-1.0.2 prerelease)



I better go do some digging here...

Re: does linux -fPIC/eh do so poorly?

[Alexandre Oliva]

Robert Lipe writes:

> I'm trying
> to figure out if something is broken on other x86 elf/unix targets, or if
> it's unique to OpenServer.   I also can't tell if I'm just the only one
> masochistic enough to make another pass over all the testsuites with -fPIC
> or not.

I used to do that myself, but when we got `make check', I started to
use it, and it would not run tests with `-fPIC'.  Perhaps a nice
addition to egcs's makefiles would be a full-check, that would test
all multilibs with and without -fPIC, if -fPIC makes a difference.  I
just don't know how to test whether -fPIC makes a difference in a
portable way, is there any way?

> 	 make check-g++ RUNTESTFLAGS="--tool_opts -fPIC old-deja.exp=eh*"

Done, all tests passed on i586-pc-linux-gnu, with egcs-980214,
binutils-2.8.1.0.18, out-of-the-box RedHat 5.0.

-- 
Alexandre Oliva
mailto:oliva@dcc.unicamp.br mailto:aoliva@acm.org
http://www.dcc.unicamp.br/~oliva
Universidade Estadual de Campinas, SP, Brasil

Re: does linux -fPIC/eh do so poorly?

[Robert Lipe]

> > I'm trying to figure out if something is broken on other x86
> > elf/unix targets, or if it's unique to OpenServer.  I also can't
> > tell if I'm just the only one masochistic enough to make another
> > pass over all the testsuites with -fPIC or not.
>
> I used to do that myself, but when we got `make check', I started
> to use it, and it would not run tests with `-fPIC'.  Perhaps a nice

It will do it, but it takes some non-obvious voodoo.

> addition to egcs's makefiles would be a full-check, that would test
> all multilibs with and without -fPIC, if -fPIC makes a difference.

The magic syntax is:

make -k check \
  RUNTESTFLAGS="--outdir ${OUT} --target_board 'unix{-mcoff,-melf,-fPIC}'"

the --outdir is just so that I can keep a history of when things passed
or failed.   The --target_board stuff loops over each of the three sets
of flags.   Even though -melf is the default on my target, I explictly
call it out here so that the reports are more readable.

I imagine that it would be possible to wire the output of a 
	(robertl) rjlhome:/play/negcs/gcc
	$ ./xgcc -B./ --print-multi-lib /tmp/x.c
	.;
	coff;@mcoff
	pic;@fPIC
into the list of options to be passed into target_board.


> I just don't know how to test whether -fPIC makes a difference in a
> portable way, is there any way?

If there is, the testsuites need to be taught about it becuase one of my
multilibs (ELF) will do -fPIC and one (COFF) won't and the testsuites
that explictly pass -fPIC fail.


> >      make check-g++ RUNTESTFLAGS="--tool_opts -fPIC
> >      old-deja.exp=eh*"
>
> Done, all tests passed on i586-pc-linux-gnu, with egcs-980214,
> binutils-2.8.1.0.18, out-of-the-box RedHat 5.0.

Bummer.  This means I'll get less sympathy that it's broken.  :-(

Thanx for checking, Alexandre.

RJL