gdb crashes due to a gcc bug (on AIX)

gdb crashes due to a gcc bug (on AIX)

[Michael Veksler]

gcc-3.2.1 emits wrong stubs debug information. This causes an assertion
failure in gdb on AIX. This failure happens during "break main".

The version that crashes is a weekly snapshot:
        GNU gdb 20030729

The unsorted debug information (nm -Cpf) is as follows:

_ZNKSt5ctypeIcE5tableEv:f1934=*1785 -       6648


This is the first occurrence of symbol 1785, and here we see the bug.
There is a "=k-8" missing. When according to rs6000_builtin_type,
"-8" stands for "unsigned int".

The line should have been:
_ZNKSt5ctypeIcE5tableEv:f1934=*1785=k-8 -       6648

This does not cause the crash yet. After several simple and harmless
references to 1785, there is:
locale:Tt1784=s4none:/21785=k-8:_ZNSt6locale4noneE;ctype:/21785:_ZNSt6 
......

This line should have been the first reference to 1785 because it defines
it to be k-8 (unsigned int). When GDB gets to this part of the debug info
it crashes. Here is the analysis of the crash:
stabsread.c:read_type(.....):
Reads 1785 and gets previously read debug information into
typeptr. This debug info is wrong, because it is not marked as
a builtin type (due to original missing "=k-8")
Enters the case for 'k' with typeptr wrongly marked as defined in
this translation unit:  TYPE_OBJFILE(*typeptr)->name != NULL
and wrongly read a new type for -8 into 'type' for which:
TYPE_OBJFILE(*typeptr)->name == NULL
make_cv_type is called with a bad typeptr and type combination
that crashes in its gdb_assert call.

I know that this is a gcc bug, but I also think that gdb should not crash
on such gcc bugs. This happens for gcc-3.2, for which development
had stopped, so it is not reasonable to look for a fix there.

It is difficult to generate this gcc bug on purpose (using a simple
example), although it happens all the time by accident :-(
I'll report this bug to GCC team once a simple example is found...

Also note that there is a PR for this bug that lacks this analysis. It is
GDB PR 1170.