On Sun, May 11, 2008 at 09:06:19AM +0000, Bernhard R. Link wrote:
> I've identified one cause of BusError on startup. (I'm not entirely sure
> it is the only one, but it definitely is the cause for the easily
> reproduceable one, and while I think I saw one with this worked around
> in the early tests, I cannot reproduce the other, so there might either
> be another issue left or I once forget to save or recompile).
>
> The cause is corruption at the following piece of code in igcref.c:715:
>
> 715: ref rtemp;
> 716:
> 717: if_debug2('8', " [8]ref 0x%lx copied to 0x%lx\n",
> 718: (ulong) src, (ulong) dest);
> 719: /* We can't just use ref_assign_inline, */
> 720: /* because the source and destination */
> 721: /* might overlap! */
> 722: ref_assign_inline(&rtemp, (ref *) src);
> 723: r_clear_attrs(&rtemp, l_mark);
> 724: ref_assign_inline((ref *) dest, &rtemp);
>
> As far as I understand, the problem is that gcc knows that two (ref*)
> pointers cannot possibly overlap, so it optimizes away the temporary
> variable (at least partially), causing corruption of the packed array.
> (Which then again causes bus errors in other code like the garbage
> collection).
>
> Here is the assember generated by gcc with -O2:
> .loc 1 722 0
> ld [%g3+4], %g2
> .loc 1 724 0
> st %g2, [%i1+4]
> sth %g1, [%i1]
> .loc 1 722 0
> st %g2, [%fp-20]
> lduh [%g3+2], %g1
> .loc 1 725 0
> add %g3, 8, %g3
> .loc 1 724 0
> sth %g1, [%i1+2]
>
> It first does part of the move from src to register, then from register to
> dest,
> only then copying another part of src to register, clearing the bit and
> storing
> that part to dest. As the clearing is done in the part with the lower address,
> it even copies the second part first, thus if compressing an array quite
> likely
> overwriting a part it just reads after that.
>
> The attached patch fixes it for me, and I think it also is the correct
> solution
> that will survive more inteligent compilers.If that's really the problem, and that you believe there are other in the code, then the proper solution with GCC is to build gs with -fno-strict-aliasing. -- ·O· Pierre Habouzit ··O [EMAIL PROTECTED] OOO http://www.madism.org
pgpwPmV16VZqd.pgp
Description: PGP signature
