Re: [PATCH] ARM: Do not allow unaligned accesses when CONFIG_ALIGNMENT_TRAP

[Catalin Marinas <catalin.marinas@arm.com>]

On Fri, May 27, 2011 at 09:54:14AM +0100, Russell King - ARM Linux wrote:
> On Fri, May 27, 2011 at 09:38:08AM +0100, Catalin Marinas wrote:
> > OK, I tried this now:
> >
> > -fconserve-stack: we get unaligned accesses on the stack because the
> > newer versions of gcc turned unaligned accesses on by default.
> >
> > -fconserve-stack -mno-unaligned-access: the stack variables are aligned.
> > We probably get the benefit of -fconserve-stack as well.
> >
> > So as per the initial post in this thread, we could have
> > -mno-unaligned-access on ARM always on (when CONFIG_ALIGNMENT_TRAP). As
> > Nicolas suggested, we could compile some files with -munaligned-access
> > (and maybe -fno-conserve-stack).
> >
> > I raised this with the gcc guys so they are looking into it. But it
> > really doesn't look like a gcc bug as long as -mno-unaligned-access is
> > taken into account.
> 
> Ok, we need to check one last thing, and that's what the behaviour is
> with -mno-unaligned-access and packed structures (such as the ethernet
> header).  If it makes no difference, then I suggest we always build
> with -mno-unaligned-access.

I tried some simple code below:

struct test {
	unsigned char a[6];
	unsigned long b;
} __attribute__((packed));

void set(struct test *t, unsigned long v)
{
	t->b = v;
}

int main(void)
{
	struct test t;
	
	set(&t, 10);

	return 0;
}

With -mno-unaligned-access in newer toolchains, the set() function looks
like this (compiled with -march=armv7):

00000000 <set>:
   0:   e7e7c451        ubfx    ip, r1, #8, #8
   4:   e7e72851        ubfx    r2, r1, #16, #8
   8:   e1a03c21        lsr     r3, r1, #24
   c:   e5c01006        strb    r1, [r0, #6]
  10:   e5c0c007        strb    ip, [r0, #7]
  14:   e5c02008        strb    r2, [r0, #8]
  18:   e5c03009        strb    r3, [r0, #9]
  1c:   e12fff1e        bx      lr

If I don't pass -mno-unaligned-access later toolchains use unaligned
accesses by default and the set() function is more efficient:

00000000 <set>:
   0:   e5801006        str     r1, [r0, #6]
   4:   e12fff1e        bx      lr

The problem is that in addition to that we also get unaligned stack
variables which are not really efficient. Either way we have a drawback
somewhere. We could argue that -fconserve-stack is badly implemented on
ARM.

-- 
Catalin