Re: [Firebird-devel] Firebird 2.5 built on HP-UX IA64 - core dump before main() (and suggested fix)

[Alex P. <pes...@ma...>]

On 06/17/14 19:51, vv m wrote:
> Hi All,
>
> I've built  FB from the source code (v. 2.5.2.26540-0) on HP-UX IA64
> (Itanium).
>
> This is my setup: HP-UX ia64, HP C/aC++ B3910B A.06.25
>
> My settings:
> export CC=cc
> export CXX=aCC
> export CFLAGS=+DD64
> export CFLAGS=+DD64
>
> ./configure --enable-debug=yes
>
> Resolving a few build issues, I've been able to build create_db, but
> similar to a situation which I had building FB on AIX (see corresponding
> discussion) it produced a core dump:
>
> Pid 14664 received a SIGSEGV for stack growth failure.
> Possible causes: insufficient memory or swap space,
> or stack size exceeded maxssiz.
>
> I checked corresponding system settings using kctune and swapinfo, and it
> looked OK.
> I started debugging and figure out that during creation of static object
> GlobalPtr<IntlUtil::Utf8CharSet> IntlUtil::utf8CharSet
> (.../jrd/IntlUtil.cpp), the assert fb_assert(keySet) is met at
> .../common/classes/fb_tls.h:129 during a call of TLS_GET macro at
> .../common/classes/alloc.cpp:300.
> Then this assert is supposed to be stored into the log, and during that
> TLS_GET called again, then, in turn assert happens again, etc. recursively.
> Finally, it kills the stack (I've attached a file with stack details).
>
> In my understanding, it happens 'cause TLS_GET uses FB's own implementation
> of TLS instead of the one provided by the system compiler.

That's not FB's own, that's posix standard functions. But I agree that 
there was firebird bug, showing itself on RISC machines with compiler 
missing __thread support.

> In file .../src/include/gen/autoconfig.h, corresponding macro is defined:
>
> /* Define it if compiler supports ISO syntax for thread-local storage */
> #define HAVE___THREAD 1
>
> which means that the system supports it.
>
> Also, there is another comment in fb_tls.h:
>
> #if defined(HAVE___THREAD)
> // Recent GCC supports __thread keyword. Sun compiler and HP-UX should have
> it too
> # define TLS_DECLARE(TYPE, NAME) __thread TYPE NAME
> # define TLS_GET(NAME) NAME
>
> But all this is in contradiction with definition in file
> .../src/common/classes/common.h:
>
> #if defined (__HP_aCC)
> // aCC error, __thread can be used only with C-like structs
> #undef HAVE___THREAD
> #endif
>
> Having that, my suggestion was to comment the line "#undef HAVE___THREAD".
> After that I've been able to finalize a build successfully and pass basic
> tests.

I've committed a fix for initial problem to B2_5. When you will be able 
to access snapshot (BTW, tomorrow you can get it from FB site) please 
also try clean build on HPUX if possible.