PATCH: mips16 function attributes, version N+1

Sandra Loosemore sandra@codesourcery.com
Fri Sep 7 16:08:00 GMT 2007


David Daney wrote:

>>> It causes bootstrap failure due to an ICE building libstdc++ on 
>>> mipsel-linux.
>>>
>>> See: http://gcc.gnu.org/ml/gcc/2007-09/msg00143.html
>>
>> Hrrmmm.  I'm not sure I can duplicate this build environment.
> I did my regression hunt on a cross compiler.
> 
> If you would like, I could make some inquiries about making the 
> mipsel-linux box available...

It's OK; I've found a test case in the libc++ testsuite that is causing 
the error.

I think there most likely is someplace deep in the C++ template 
expansion code that really does have an incomplete or incorrect tree 
structure, that never normally sees the light of day.  But it's blowing 
up in the early RTL generation caused by mips_set_current_function_hook. 
  To tell the truth, I was always suspicious of that, and was somewhat 
shocked that it seemed to work!  :-P

Richard advises temporarily patching the mips_set_current_function_hook 
to be a no-op rather than reverting the whole patch (since there are 
already other changes that depend on it) in order to un-break bootstrap. 
  I'll commit that as soon as I verify that it builds OK.  Then I will 
rewrite the hook to avoid the early RTL generation and do a complete 
round of re-testing on that before re-submitting.  OK?

-Sandra



More information about the Gcc-patches mailing list