Re: gcj 4.3.0 on mpc860 kernel 2.4.24

[Andrew Haley <aph at redhat dot com>]

Ben Gardiner wrote:
> Andrew Haley wrote:
>> Linuxthreads?  I don't like the sound of that.  gcj doesn't have any
>> explicit dependencies on NPTL, but it doesn't get tested on Linuxthreads.
>>   
> I see. That's important information, thank you.
>> Where does it segfault?  I mean, where *exactly*?
>>   
> It segfaults in _Jv_MonitorEnter at locks.h:44 (compare_and_swap); here
> is a full bt up to prims.cc:1666 ( _Jv_InitClass
> (&java::lang::ClassLoader::class$); ) :
> 
> -------------------------------------------------------------------------
> #0  _Jv_MonitorEnter (obj=<value optimized out>) at sysdep/locks.h:44
>        addr = 264790936
>        address = 0
>        he = (hash_entry *) 0xff83458
>        self = 16384
>        was_heavy = <value optimized out>

Aha!  At the segfault, do something like

x/4i $pc

> ------------------------------------------------------------------------
> 
> and here is my locks.h around 44:
> ------------------------------------------------------------------------
> 30      compare_and_swap (volatile obj_addr_t *addr, obj_addr_t old,
> 31                        obj_addr_t new_val)
> 32      {
> 33        obj_addr_t ret;
> 34
> 35        __asm__ __volatile__ (
> 36                 "0:    " _LARX "%0,0,%1 \n"
> 37                 "      xor. %0,%3,%0\n"
> 38                 "      bne 1f\n"
> 39                 "      " _STCX "%2,0,%1\n"
> 40                 "      bne- 0b\n"
> 41                 "1:   \n"
> 42              : "=&r" (ret)
> 43              : "r" (addr), "r" (new_val), "r" (old)
> 44              : "cr0", "memory");
> 45
> 46        /* This version of __compare_and_swap is to be used when
> acquiring
> 47           a lock, so we don't need to worry about whether other memory
> 48           operations have completed, but we do need to be sure that
> any loads
> 49           after this point really occur after we have acquired the
> lock.  */
> 50        __asm__ __volatile__ ("isync" : : : "memory");
> 51        return ret == 0;
> 52      }
> -------------------------------------------------------------------------
> 
> I did a diff and discovered that this file is the
> libjava/sysdep/powerpc/locks.h file (no surprise). I'm not an expert in
> PPC assembly and I don't know if those instructions are supported on the
> mpc860; I will look into that. In the meantime, I may try to get the '
> --enable-portable-native-sync' configure option from 4.2.3 to work with
> 4.3.0. I'm thinking I might have better luck with sysdep/generic/locks.h.

I doubt it.  gcj needs something that really works.

It really, really, really is worth trying to get a real compare_and_swap
that is written for your hardware.  

Andrew.