glibc build failure for Hurd with mainline GCC

glibc build failure for Hurd with mainline GCC

[Joseph Myers]

The build failures in the compilers stage of a build-many-glibcs.py build 
are now down to just a failure for i686-gnu, which I hope the Hurd 
maintainers might look at.

The error reported is:

In file included from intr-msg.c:25:
intr-msg.c: In function '_hurd_intr_rpc_mach_msg':
../sysdeps/mach/hurd/i386/intr-msg.h:48:21: error: array subscript -1 is outside array bounds of 'mach_msg_header_t[0]' [-Werror=array-bounds]
   48 |        : "m" ((&msg)[-1]), "m" (*cancel_p), "i" (EINTR)                       \
      |               ~~~~~~^~~~
intr-msg.c:118:9: note: in expansion of macro 'INTR_MSG_TRAP'
  118 |   err = INTR_MSG_TRAP (msg, option, send_size,
      |         ^~~~~~~~~~~~~
intr-msg.c:32:45: note: at offset -4 into object 'msg' of size 4
   32 | _hurd_intr_rpc_mach_msg (mach_msg_header_t *msg,
      |                          ~~~~~~~~~~~~~~~~~~~^~~

As far as I can tell, this seems to be reporting a genuine issue.  The asm 
in INTR_MSG_TRAP has inputs saying it accesses ((&msg)[-1]).  And in the 
call to that macro from _hurd_intr_rpc_mach_msg, msg is a function 
parameter.  That is, this code seems to be taking the address of a 
function parameter, then looking at whatever comes before that function 
parameter in memory, which doesn't make much sense.

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: glibc build failure for Hurd with mainline GCC

[Samuel Thibault]

Hello,

Joseph Myers, le jeu. 14 oct. 2021 21:15:59 +0000, a ecrit:
> The build failures in the compilers stage of a build-many-glibcs.py build 
> are now down to just a failure for i686-gnu, which I hope the Hurd 
> maintainers might look at.

Yep, that was an old hack to pass parameters over.  I have now reworked
it, it should be much safer now.

Samuel