3.3.3: libstdc++ regression on ia64-hpux - messy EH problem - need help

[Jim Wilson]

Zack Weinberg wrote:
> GCC 3.3.3 shows one regression relative to 3.3.2 on ia64-hpux:
> libstdc++'s 27_io/ios_init.cc test.  The failure appears only in LP64
> mode (RUNTESTFLAGS="--target_board unix/-mlp64").  I have tracked this
> as far as I can, which is not far enough to identify the problem.

A useful question to ask is whether the same testcase fails on 
ia64-linux.  Checking the gcc-testresults mailing list, I see
     http://gcc.gnu.org/ml/gcc-testresults/2004-02/msg01250.html
which does not show this failure.  So perhaps this is some sort of 
incompatibility between the hpux unwinder and gcc.  You could try 
tracing the linux and hpux binaries in parallel, to see why they behave 
differently.  It might also be useful to known whether or not having the 
linux libunwind makes a difference.

There is one known compiler unwinder bug.  See bug 13158.  It does not 
seem to apply to this testcase though.  The only sibcalls I see are in 
__GLOBAL__{I,D}_buf.  They are the only ones with an alloc instruction 
in the middle of a function.  The problem could be in some other file 
though.  If you wanted to be sure, you could try compiling all of 
libstdc++ without sibcall optimizations.  I think this is unlikely to 
help though.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com