Fast Code and HAVE_EFFICIENT_UNALIGNED_ACCESS (was: [PATCH] poly1305: generic C can be faster on chips with slow unaligned access)

From: Jeffrey Walton
Date: Wed Nov 02 2016 - 17:37:02 EST

On Wed, Nov 2, 2016 at 5:25 PM, Jason A. Donenfeld <Jason@xxxxxxxxx> wrote:
> These architectures select HAVE_EFFICIENT_UNALIGNED_ACCESS:
> s390 arm arm64 powerpc x86 x86_64
> So, these will use the original old code.
> The architectures that will thus use the new code are:
> alpha arc avr32 blackfin c6x cris frv h7300 hexagon ia64 m32r m68k
> metag microblaze mips mn10300 nios2 openrisc parisc score sh sparc
> tile um unicore32 xtensa

What I have found in practice from helping maintain a security library
and running benchmarks until my eyes bled....

UNALIGNED_ACCESS is a kiss of death. It effectively prohibits -O3 and
above due to undefined behavior in C and problems with GCC
vectorization. In the bigger picture, it simply slows things down.

Once we moved away from UNALIGNED_ACCESS and started testing at -O3
and -O5, the benchmarks enjoyed non-trivial speedups on top of any
speedups we were trying to achieve with hand tuned assembly language
routines. Effectively, the best speedup was the sum of C-code and ASM;
they were not disjoint as they appear.

The one wrinkle for UNALIGNED_ACCESS is Bernstein's compressed tables
UNALIGNED_ACCESS meets some security goals. The techniques from
Bernstein's paper apply equally well to AES, Camellia and other
table-driven implementations. Painting with a broad brush (and as far
as I know), the kernel is not observing the recommendations. My
apologies if I parsed things incorrectly.