Re: [PATCH v4 08/23] ewah: compressed bitmap implementation

On Thu, Jan 23, 2014 at 12:23:42PM -0800, Jonathan Nieder wrote:

quoted

The memcpy solution is taken from read-cache.c, but as we noted, it
probably hasn't been used a lot. The blk_sha1 get_be may be faster, as
it converts as it reads.

I doubt there's much difference either way, especially after an
optimizer gets its hands on it.  According to [1] ARM has no fast
byte swap instruction so with -O0 the byte-at-a-time implementation is
probably faster there.  I can try a performance test if you like.

If you're curious and have time, go ahead and benchmark what I posted
against what you posted (with your fix). But you'll probably need a big
repo like the kernel to notice anything.

But I don't mind that much if we just use the memcpy trick for now. It's
nice and obvious, and we can always change it later if somebody has
numbers (I doubt it will be all that noticeable anyway; this isn't
nearly as tight a loop as the BLK_SHA1 code).

-Peff