Date: Thu, 03 Aug 2006 15:52:18 -0700 From: Maxim Sobolev <sobomax@FreeBSD.org> To: David Edelsohn <dje@watson.ibm.com> Cc: Peter Grehan <grehan@FreeBSD.org>, freebsd-ppc@FreeBSD.org Subject: Re: Unaligned 64-bits access on FreeBSD/powerpc Message-ID: <44D27E22.8030909@FreeBSD.org> In-Reply-To: <44D275C4.5020407@FreeBSD.org> References: <44D23F02.9020709@FreeBSD.org> <200608031826.k73IQX835138@makai.watson.ibm.com> <44D2419F.1050100@FreeBSD.org> <44D245E5.4070102@freebsd.org> <44D24772.7080109@FreeBSD.org> <44D24A3F.5080407@freebsd.org> <200608031929.k73JTq834496@makai.watson.ibm.com> <44D275C4.5020407@FreeBSD.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Never mind, I have identified and fixed the problem.
Thanks to everybody for help!
-Maxim
Maxim Sobolev wrote:
> David Edelsohn wrote:
>>>>>>> Peter Grehan writes:
>>
>>>> atype = "64";
>>>> for (i = 0; i < 256; i++) {
>>>> apoint = (uint8_t *)&(buf[i]);
>>>> v64 = *(uint64_t *)apoint;
>>>> }
>>
>> Peter> Would you be able to do a disassembly on this ? I suspect gcc
>> is using Peter> floating-point regs to do the 64-bit loads.
>>
>> That does not matter. The OS is suppose to catch the misaligned
>> load and emulate the instruction in a handler. If FreeBSD cannot handle
>> that, then it needs to configure GCC for PowerPC in strict-alignment
>> embedded mode, which is ABI incompatible. If FreeBSD wants to use the
>> standard PPC SVR4 ABI or PPC32 Linux variant, it must support any
>> misaligned load. That is the OS's responsibility, period.
>
> David,
>
> Thank you for your input. I am trying to debug this problem right now.
> It appears that the FreeBSD kernel already handles this condition based
> on the DSISR/DAR contents(trap.c, function fix_unaligned()). However,
> for some reason when the trap happens the DSISR is set to 0x40000000
> (all bits but bit 30 are set to 0).
>
> IBM's Programming Environments Manual says the following regarding
> Alignment Exception in the big endian mode:
>
> <quote>
> Setting the DSISR and DAR as described below is optional for
> implementations on which alignment exceptions
> occur rarely, if ever, for cases that the alignment exception handler
> emulates. For such implementations,
> if the DSISR and DAR are not set as described below they are set to
> undefined values.
> </quote>
>
> Therefore, I am wondering if it's the case here.
>
> -Maxim
>
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?44D27E22.8030909>