A quick architecture status report

On Mon, Jul 10, 2017 at 8:56 AM, Joseph Myers <joseph@codesourcery.com> wrote:
> On Mon, 10 Jul 2017, Zack Weinberg wrote:
>
>> Alas, this is significantly more research than I have time for.  I was
>> hoping for a simple matter of adding .note.GNU-stack annotations to
>> our own .S files :)
>
> Since we build with -Wa,--noexecstack if the compiler puts .note.GNU-stack
> sections in its output, that's never necessary.  What's needed in GCC is
> (a) a TARGET_ASM_FILE_END hook that uses file_end_indicate_exec_stack, (b)
> any assembly sources in libgcc need .note.GNU-stack annotations.  Easy to
> add, but architecture maintainers are best placed to know if that's the
> right thing to do for the particular architecture.

Hm, what do you think of this patch?  It makes elf/check-execstack an
expected failure on any target where the compiler doesn't put
.note.GNU-stack sections in its output.  (UNSUPPORTED might be more
appropriate, but I don't see a way to trigger that from a makefile
conditional.)

zw

On Mon, 10 Jul 2017, Zack Weinberg wrote:

> Hm, what do you think of this patch?  It makes elf/check-execstack an
> expected failure on any target where the compiler doesn't put
> .note.GNU-stack sections in its output.  (UNSUPPORTED might be more
> appropriate, but I don't see a way to trigger that from a makefile
> conditional.)

It's not clear it *should* be expected unless there's a good 
architecture-specific reason.

On Jul 10 2017, Zack Weinberg <zackw@panix.com> wrote:

> +# If the compiler does not support .note.GNU-stack for this
> +# architecture, check-execstack is expected to fail.
> +ifeq (,$(filter %noexecstack,$(ASFLAGS-config)))
> +test-xfail-check-execstack = yes
> +endif

An architecture that is noexec by default doesn't have -znoexecstack
either.

Andreas.

On Mon, Jul 10, 2017 at 11:59 AM, Joseph Myers <joseph@codesourcery.com> wrote:
> On Mon, 10 Jul 2017, Zack Weinberg wrote:
>> Hm, what do you think of this patch?  It makes elf/check-execstack an
>> expected failure on any target where the compiler doesn't put
>> .note.GNU-stack sections in its output.  (UNSUPPORTED might be more
>> appropriate, but I don't see a way to trigger that from a makefile
>> conditional.)
>
> It's not clear it *should* be expected unless there's a good
> architecture-specific reason.

This is why I said UNSUPPORTED might be more appropriate.

Regardless of whether it *should* work, the fact is that it doesn't,
for reasons outside the control of code in glibc, and therefore
leaving this as a FAIL is inappropriate in my book.  Would you be
satisfied with a patch along these lines (and a more specific
configure check, because of what Andreas said downthread) if I also
filed bugs on GCC for all affected architectures?  I didn't find any
existing such bugs, but I didn't look very hard.

zw

On Mon, 10 Jul 2017, Zack Weinberg wrote:

> On Mon, Jul 10, 2017 at 11:59 AM, Joseph Myers <joseph@codesourcery.com> wrote:
> > On Mon, 10 Jul 2017, Zack Weinberg wrote:
> >> Hm, what do you think of this patch?  It makes elf/check-execstack an
> >> expected failure on any target where the compiler doesn't put
> >> .note.GNU-stack sections in its output.  (UNSUPPORTED might be more
> >> appropriate, but I don't see a way to trigger that from a makefile
> >> conditional.)
> >
> > It's not clear it *should* be expected unless there's a good
> > architecture-specific reason.
> 
> This is why I said UNSUPPORTED might be more appropriate.
> 
> Regardless of whether it *should* work, the fact is that it doesn't,
> for reasons outside the control of code in glibc, and therefore
> leaving this as a FAIL is inappropriate in my book.  Would you be

Well, if you (a new architecture, say) have executable stacks there's 
something wrong with your toolchain port unless there's a good reason, and 
I think it's desirable for there to be something immediately visible 
showing there's something wrong.

That is, we should decide case-by-case whether something should be fixed 
or XFAILed.  Which for this test should be fix if it's just a 
straightforward GCC change needed (ask architecture maintainers), XFAIL 
with comments otherwise.  Or obsolete the port in the absence of 
architecture maintainers (MicroBlaze, for which email to the maintainer 
bounces).