Lucas Stach dev@lynxeye.de writes:
Am Montag, den 02.07.2012, 10:53 +0100 schrieb Måns Rullgård:
Tetsuyuki Kobayashi koba@kmckk.co.jp writes:
Recent compiler generates unaligned memory access in armv7 default. But current U-Boot does not allow unaligned memory access, so it causes data abort exception. This patch add compile option "-mno-unaligned-access" if it is available.
Why not allow unaligned accesses instead?
IMHO, our recent discussion showed that both ways are wrong. "-mno-unaligned-access" works around misaligned data on the software level, while allowing unaligned access does on the hardware level.
What we really want is no unaligned access in U-Boot at all. Just because "-mno-unaligned-access" is the default on ARMv5, we should not consider it a gold standard.
It's slightly more complicated than that. Data can be misaligned for a variety of reasons:
1. Errors in software. 2. Specified by a file format or communication protocol. 3. Deliberately misaligned by the compiler.
Misaligned data of type 1 should of course be fixed properly, not worked around in any way.
Type 2 happens all the time, and has to be dealt with one way or another. If the hardware supports unaligned accesses, this is usually faster than reading a byte at a time.
When targeting ARMv6 and later, recent gcc versions have started issuing deliberate unaligned accesses where previously byte by byte accesses would have been done. This happens with "packed" structs and sometimes to write multiple smaller values at once, typically when zero-initialising things. These unaligned accesses are *good*. They make code smaller and faster.
The real problem here is that u-boot is setting the strict alignment checking flag, invalidating the assumption of the compiler that the system allows unaligned accesses. For ARMv5 and earlier, setting this flag is usually advisable since it makes finding accidental unaligned accesses much easier.
This was debated in the context of the kernel a while ago, ultimately leading to strict alignment being disabled for ARMv6 and up [1].
[1] http://git.kernel.org/?p=linux/kernel/git/torvalds/linux.git;a=commitdiff;h=...