Re: block-sha1: improve code on large-register-set machines

From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Aug 11, 2009, 22:57 UTC
Message-ID: <alpine.LFD.2.01.0908111553460.28882@localhost.localdomain>
In-Reply-To: <fLYKSyures_wcvAvAV9-MgKQlhk959HJpx-pKz7T1n-Mel7f2RBkMw@cipher.nrlssc.navy.mil>

On Tue, 11 Aug 2009, Brandon Casey wrote:

> 
> In that case, why not change the interface of blk_SHA1Block() so that its
> second argument is const unsigned char* and get rid of __d and the { } ?

Because on big-endian, or on architectures like x86 that have an efficient byte swap, that would be horrible.

You absoluetoy MUST NOT do things a byte at a time in those cases. The memory operations and the shifting just kills you.

The reason you want to do things a byte at a time on ARM is that ARM cannot do unaligned accesses well (very modern cores are better, but rare), and that ARM has no bswap instruction and has fairly cheap shifts.

On no sane architecture is that true. Unaligned loads are fast (and quite frankly, hardware where unaliged loads aren't fast is just crazy sh*t), and doing 'bswap' is way faster than doing many shifts and masks.

So everything should be fundamentally word-oriented. Then, broken architectures that can't handle it should split up the words, not the other way around.

			Linus