Re: [head tinderbox] failure on mips/mips

In message: <4CB77F1D-4235-47D8-B654-1C4F29B6C649_at_mac.com>
            Marcel Moolenaar <xcllnt_at_mac.com> writes:
: 
: On Feb 17, 2009, at 9:21 PM, M. Warner Losh wrote:
: 
: > In message: <20090217.203647.-1518647466.imp_at_bsdimp.com>
: >            "M. Warner Losh" <imp_at_bsdimp.com> writes:
: > : In message: <20090218023328.227617302F_at_freebsd-current.sentex.ca>
: > :             FreeBSD Tinderbox <tinderbox_at_freebsd.org> writes:
: > : : /src/usr.sbin/bsnmpd/modules/snmp_mibII/../../../../contrib/ 
: > bsnmp/snmp_mibII/mibII.c:1016: warning: cast increases required  
: > alignment of target type
: > :
: > : there's still 3 or 4 of these in the tree that I'm trying to track
: > : back to root cause.  A simple (void *) fixes the problem, but I want
: > : to understand the issues before I slap that bad-boy in there...
: >
: > The first one is:
: >
: > 	  case RTM_IFINFO:
: > 		ifm = (struct if_msghdr *)rtm;
: > 		mib_extract_addrs(ifm->ifm_addrs, (u_char *)(ifm + 1), addrs);
: > 		if ((ifp = mib_find_if_sys(ifm->ifm_index)) == NULL)
: > 			break;
: >
: > rtm is of type struct rt_msghdr.  This has an alignment requirement of
: > 4 on mips, at least on 32-bit mips (the biggest data element is a
: > u_long).  struct if_msghdr has an alignment requirement of 8, because
: > time_t is int64_t on MIPS, which is 8-bytes in size.
: 
: Normally on 32-bit architectures, 64-bit data types have a 32-bit
: alignment requirement by virtue of needing 2 32-bit loads/stores
: to read/write them. Does 32-bit MIPS use 64-bit wide registers?

MIPS isn't normal :).  MIPS really is a 64-bit architecture.  The only
way to make the warning go away would be to build -mmips32r2 or
something like that.  But since we're going to be supporting n32 ABI
(which really is a 64-bit ABI despite its name) and n64, the issue
won't be going away.  This is desirable for the Octeon support that I
hope to commit soon.

: > One way to fix this is to add __aligned(8) to struct rt_msghdr to
: > compensate for this.  Otherwise, if the time_t element is referenced
: > in ifm_data we'll core dump.
: 
: A safer approach is to mark ifi_epoch as packed or put differently,
: define time_t as a 64-bit integral with 32-bit alignment. This can
: avoid a lot of unexpected internal padding as well (e.g. struct
: timeval).

Marking it as packed won't help.  If the elements aren't properly
aligned, gcc won't access multi-word entities properly.  It might
eliminate the warning, but it will break at runtime.

Warner