egcs-20000114 - internal compiler error compiling XEmacs

[Martin Buchholz]

>>>>> "AO" == Alexandre Oliva <oliva@lsd.ic.unicamp.br> writes:

AO> On Jan 28, 2000, Zack Weinberg <zack@wolery.cumb.org> wrote:
>> On Thu, Jan 27, 2000 at 10:13:35PM -0800, Martin Buchholz wrote:
>>> The combination gcc + --enable-cpplib when compiling XEmacs src/elhash.c
>>> crashes gcc in the following way:
>>> 
>>> gcc -O3 -Wreturn-type ==> crash
>>> gcc -O -Wreturn-type ==> no crash
>>> gcc -O3 ==> no crash

>> [...]
>>> (martin@lasker) /xemacs/build/moS/src $ /hack/import/egcs-20000124/inst/linux-libc2/bin/gcc -v -c -O3 -Wreturn-type -Demacs -I. -DHAVE_CONFIG_H -I/usr/local/include -I/usr/X11R6/include /xemacs/ws/dev/src/elhash.c

>> Please do this again with -save-temps added to the command line, and
>> mail us elhash.i (compress it first, it'll be huge).

AO> I don't think that's necessary.  I've already posted a patch for a bug
AO> that triggered on very similar conditions during bootstrap with -O3.

AO> Martin, will you please test the patch at
AO> http://gcc.gnu.org/ml/gcc-patches/2000-01/msg00978.html

This patch fixes the problem for me.  I am now again able to build an
optimized XEmacs binary using an egcs snapshot.

TODOs for the gcc maintainers:

- Apply a version of Alexandre's patch.

- Fix cpp to make --enable-cpplib unnecessary.

- Performance on my favorite benchmark (byte-compiling elisp) is worse
  by 5% on egcs-20000124 relative to gcc-2.95.2, with 
  -mcpu=pentiumpro -march=pentiumpro -fstrict-aliasing -O3

  Details of benchmark available by request.
  
Thanks.

Martin