Re: [musl] Initial xtensa/fdpic port review

[Rich Felker <dalias@libc.org>]

On Wed, Feb 28, 2024 at 03:26:45PM -0500, Rich Felker wrote:
> On Wed, Feb 28, 2024 at 03:14:12PM -0500, Rich Felker wrote:
> > On Wed, Feb 28, 2024 at 11:41:30AM -0800, Max Filippov wrote:
> > > On Wed, Feb 28, 2024 at 11:34 AM Max Filippov <jcmvbkbc@gmail.com> wrote:
> > > >
> > > > On Wed, Feb 28, 2024 at 10:36 AM Rich Felker <dalias@libc.org> wrote:
> > > > > On Wed, Feb 28, 2024 at 01:30:32PM -0500, Rich Felker wrote:
> > > > > > On Wed, Feb 28, 2024 at 09:20:33AM -0800, Max Filippov wrote:
> > > > > > >
> > > > > > > > >               p->relocated = 1;
> > > > > > > > >       }
> > > > > > > > > @@ -1485,7 +1487,7 @@ void __libc_exit_fini()
> > > > > > > > >               if (dyn[0] & (1<<DT_FINI_ARRAY)) {
> > > > > > > > >                       size_t n = dyn[DT_FINI_ARRAYSZ]/sizeof(size_t);
> > > > > > > > >                       size_t *fn = (size_t *)laddr(p, dyn[DT_FINI_ARRAY])+n;
> > > > > > > > > -                     while (n--) ((void (*)(void))*--fn)();
> > > > > > > > > +                     while (n--) fpaddr(p, *--fn)();
> > > > > > > >
> > > > > > > > If this is fixable on the tooling side it really should be fixed
> > > > > > > > there. init/fini arrays should have actual language-level function
> > > > > > > > addresses (descriptor addresses on fdpic), not instruction addresses.
> > > > > > >
> > > > > > > I read libgcc code at
> > > > > > >   https://github.com/jcmvbkbc/gcc-xtensa/blob/xtensa-14-8789-fdpic/libgcc/crtstuff.c#L498-L503
> > > > > > > and the way it's written suggests that this was done on purpose.
> > > > > > > I put it into the WIP pile to figure out later what the purpose was..
> > > > > > > I thought that SH might not have this issue because it just didn't
> > > > > > > use the .array_init/.array_fini.
> > > > > >
> > > > > > I'm pretty sure we're using it -- musl-cross-make always forces it on
> > > > > > via the gcc configure command line -- but it's possible there's some
> > > > > > override disabling it for sh. I'll try some test cases and confirm
> > > > > > whether sh is doing it right. Maybe the arm folks will have input on
> > > > > > this too..?
> > > > >
> > > > > Confirmed both that it works, and that it's working via init_array.
> > > > > GCC emits:
> > > > >
> > > > >         .section        .init_array,"aw"
> > > > >         .align 2
> > > > >         .long   foo@FUNCDESC
> > > > >
> > > > > for
> > > > >
> > > > >         __attribute__((__constructor__))
> > > > >         void foo() { ... }
> > > > >
> > > >
> > > > Oh, no doubt that that C code generates a function descriptor, it
> > > > works for xtensa too. But the piece of libgcc quoted above specifically
> > > > puts a pointer to an object, not to a function into the .init_array.
> > > 
> > > It was introduced to gcc by the ARM FDPIC series:
> > > https://github.com/jcmvbkbc/gcc-xtensa/commit/11189793b6ef60645d5d1126d0bd9d0dd83e6583
> > > 
> > > This is the second change that I find made by the ARM FDPIC
> > > series that appears to be not right for other FDPIC ports, first
> > > being this change to the C++ unwinding code:
> > > https://github.com/jcmvbkbc/gcc-xtensa/commit/67b0605494f32811364e25328d3522467aaf0638
> > 
> > OK, so the arm folks put explicitly wrong/broken code here. That needs
> > to be reverted, and they can work out the mess they created on glibc.
> > 
> > There is probably wrong arm target code too whereby gcc is generating
> > instruction addresses for __attribute__((__constructor__)) rather than
> > function addresses.
> > 
> > If they have compat to worry about with glibc binaries, that's going
> > to be a mess for them to fix, but we can just patch it out for musl
> > target regardless of what they do since we have no existing broken
> > binaries.
> 
> OK, really good news! They didn't actually botch it. This test on
> godbolt shows GCC is doing the right thing:
> 
> https://godbolt.org/z/b53ExYoPf
> 
> (GCC 13.1.0 ARM, -O2 -mfdpic, following code)
> 
> #include <stdio.h>
> __attribute__((__constructor__))
> void foo()
> {
> 	printf("hello ");
> }
> 
> int main()
> {
> 	printf("world\n");
> }
> 
> Rather, someone just made the crtstuff gratuitously do the wrong
> thing, then later hard-coded disabling initfiniarray on fdpic because
> it was doing the wrong thing:
> 
> https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=9c560cf23996271ee26dfc4a1d8484b85173cd12;hp=6bcbf80c6e2bd8a60d88bbcac3d70ffb67f4888f
> 
> So all that's needed is to revert the wrong patch to crtstuff and then
> xtensa and arm initfiniarray will work as they should.
> 
> Anyone want to open a GCC bug tracker item for this?

Filed as https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114158