On Mon, Mar 23, 2015 at 9:25 AM, David Miller <davem@xxxxxxxxxxxxx> wrote:
>
> Ok, here is what I committed.
So I wonder - looking at that assembly, I get the feeling that it
isn't any better code than gcc could generate from simple C code.
Would it perhaps be better to turn memmove() into C?
That's particularly true because if I read this code right, it now
seems to seriously pessimise non-overlapping memmove's, in that it now
*always* uses that slow downward copy if the destination is below the
source.
Now, admittedly, the kernel doesn't use a lot of memmov's, but this
still falls back on the "byte at a time" model for a lot of cases (all
non-64-bit-aligned ones). I could imagine those existing. And some
people (reasonably) hate memcpy because they've been burnt by the
overlapping case and end up using memmove as a "safe alternative", so
it's not necessarily just the overlapping case that might trigger
this.
Maybe the code could be something like
void *memmove(void *dst, const void *src, size_t n);
{
// non-overlapping cases
if (src + n <= dst)
return memcpy(dst, src, n);
if (dst + n <= src)
return memcpy(dst, src, n);
// overlapping, but we know we
// (a) copy upwards
// (b) initialize the result in at most chunks of 64
if (dst+64 <= src)
return memcpy(dst, src, n);
.. do the backwards thing ..
}
(ok, maybe I got it wrong, but you get the idea).
I *think* gcc should do ok on the above kind of code, and not generate
wildly different code from your handcoded version.
Linus
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>