musl libc.so working and almost building with current pcc

musl libc.so working and almost building with current pcc

[Isaac Dunham]

Hello,
Current (20160403) pcc almost builds a working libc.so.

(Unfortunately, it's still necessary to fix the pcc-libs package
by updating config.guess/config.sub and adding protected visibility;
I didn't add protected visibility and forced -fPIC, though the build
system seems to add -fPIC for some files.)

However, PCC reports a major internal compiler error at line 1382
of ldso/dynlink.c (which is conditional on FDPIC support).
If I convert the "if (DL_FDPIC) ..." to an ifdef, I get a working libc.
(src/complex/catanf.c is now properly compiled.)

I'd like to find out what's causing the ICE and report it, but reducing
the dynamic linker to a testcase isn't something that I'm ready to do.
If someone has any further ideas, that would be appreciated.

Thanks,
Isaac Dunham

Re: musl libc.so working and almost building with current pcc

[Rich Felker]

On Sun, Apr 03, 2016 at 02:47:56PM -0700, Isaac Dunham wrote:
> Hello,
> Current (20160403) pcc almost builds a working libc.so.
> 
> (Unfortunately, it's still necessary to fix the pcc-libs package
> by updating config.guess/config.sub and adding protected visibility;
> I didn't add protected visibility and forced -fPIC, though the build
> system seems to add -fPIC for some files.)

Everything in libpcc.a needs to be _hidden_, not protected. The issue
is not whether it binds locally but whether it produces
wrongly-exported symbols in shared libs it gets linked into.

> However, PCC reports a major internal compiler error at line 1382
> of ldso/dynlink.c (which is conditional on FDPIC support).
> If I convert the "if (DL_FDPIC) ..." to an ifdef, I get a working libc.
> (src/complex/catanf.c is now properly compiled.)

It's always defined (to either 0 or 1) so #ifdef produces wrong code.
Did you try #if DL_FDPIC? That should work but there's obviously some
serious pcc bug going on here, probably preprocessor breakage again.

FWIW pcc used to work fine but they keep breaking it, usually the
preprocessor.

Rich

Re: musl libc.so working and almost building with current pcc

[Isaac Dunham]

On Sun, Apr 03, 2016 at 08:04:48PM -0400, Rich Felker wrote:
> On Sun, Apr 03, 2016 at 02:47:56PM -0700, Isaac Dunham wrote:
> > Hello,
> > Current (20160403) pcc almost builds a working libc.so.
> > 
> > (Unfortunately, it's still necessary to fix the pcc-libs package
> > by updating config.guess/config.sub and adding protected visibility;
> > I didn't add protected visibility and forced -fPIC, though the build
> > system seems to add -fPIC for some files.)
> 
> Everything in libpcc.a needs to be _hidden_, not protected. The issue
> is not whether it binds locally but whether it produces
> wrongly-exported symbols in shared libs it gets linked into.

I remembered the general problem, but not the exact solution (had hidden
and protected mixed up).

By the way, checking things, yes, adding '-fPIC' to CFLAGS is still
necessary.

> > However, PCC reports a major internal compiler error at line 1382
> > of ldso/dynlink.c (which is conditional on FDPIC support).
> > If I convert the "if (DL_FDPIC) ..." to an ifdef, I get a working libc.
> > (src/complex/catanf.c is now properly compiled.)
> 
> It's always defined (to either 0 or 1) so #ifdef produces wrong code.
> Did you try #if DL_FDPIC? That should work but there's obviously some
> serious pcc bug going on here, probably preprocessor breakage again.

Yes, I used '#if DL_FDPIC' (not #ifdef).

> FWIW pcc used to work fine but they keep breaking it, usually the
> preprocessor.

I know.

Thanks,
Isaac Dunham

Re: musl libc.so working and almost building with current pcc

[Isaac Dunham]

On Sun, Apr 03, 2016 at 08:04:48PM -0400, Rich Felker wrote:
> On Sun, Apr 03, 2016 at 02:47:56PM -0700, Isaac Dunham wrote:
> > However, PCC reports a major internal compiler error at line 1382
> > of ldso/dynlink.c (which is conditional on FDPIC support).
> > If I convert the "if (DL_FDPIC) ..." to an ifdef, I get a working libc.
> > (src/complex/catanf.c is now properly compiled.)
> 
> It's always defined (to either 0 or 1) so #ifdef produces wrong code.
> Did you try #if DL_FDPIC? That should work but there's obviously some
> serious pcc bug going on here, probably preprocessor breakage again.

Not a preprocessor bug; building dynlink.i by hand, I see that this line
	if (DL_FDPIC) ((stage3_func)&ldso.funcdescs[dls3_def.sym-ldso.syms])(sp);
becomes
	if (0) ((stage3_func)&ldso.funcdescs[dls3_def.sym-ldso.syms])(sp);

Which is what it should be.
So it's hitting a problem with understanding the code.

Thanks,
Isaac Dunham