Add optimized wipememory for ARM

src/g10lib.h [arm] (fast_wipememory2_unaligned_head)

(fast_wipememory2): New macros.

Previous patch that removed _gcry_burn_stack optimization causes burn_stack
take over 30% CPU usage when looping 'benchmark cipher blowfish' on
ARM/Cortex-A8. Optimizing wipememory2 for ARM helps situation a lot.

Old vs new (Cortex-A8):

   ECB/Stream         CBC             CFB             OFB             CTR
--------------- --------------- --------------- --------------- ---------------

IDEA 1.20x 1.18x 1.16x 1.15x 1.16x 1.18x 1.18x 1.16x 1.16x 1.17x
3DES 1.14x 1.14x 1.12x 1.13x 1.12x 1.13x 1.12x 1.13x 1.13x 1.15x
CAST5 1.66x 1.67x 1.43x 1.00x 1.48x 1.00x 1.44x 1.44x 1.04x 0.96x
BLOWFISH 1.56x 1.66x 1.47x 1.00x 1.54x 1.05x 1.44x 1.47x 1.00x 1.00x
AES 1.52x 1.42x 1.04x 1.00x 1.00x 1.00x 1.38x 1.37x 1.00x 1.00x
AES192 1.36x 1.36x 1.00x 1.00x 1.00x 1.04x 1.26x 1.22x 1.00x 1.04x
AES256 1.32x 1.31x 1.03x 1.00x 1.00x 1.00x 1.24x 1.30x 1.03x 0.97x
TWOFISH 1.31x 1.26x 1.23x 1.00x 1.25x 1.00x 1.24x 1.23x 1.00x 1.03x
ARCFOUR 1.05x 0.96x
DES 1.31x 1.33x 1.26x 1.29x 1.28x 1.29x 1.26x 1.29x 1.27x 1.29x
TWOFISH128 1.27x 1.24x 1.23x 1.00x 1.28x 1.00x 1.21x 1.26x 0.97x 1.06x
SERPENT128 1.19x 1.19x 1.15x 1.00x 1.14x 1.00x 1.17x 1.17x 0.98x 1.00x
SERPENT192 1.19x 1.24x 1.17x 1.00x 1.14x 1.00x 1.15x 1.17x 1.00x 1.00x
SERPENT256 1.16x 1.19x 1.17x 1.00x 1.14x 1.00x 1.15x 1.15x 1.00x 1.00x
RFC2268_40 1.00x 0.99x 1.00x 1.01x 1.00x 1.00x 1.03x 1.00x 1.01x 1.00x
SEED 1.20x 1.20x 1.18x 1.17x 1.17x 1.19x 1.18x 1.16x 1.19x 1.19x
CAMELLIA128 1.38x 1.34x 1.31x 1.00x 1.31x 1.00x 1.29x 1.32x 1.00x 1.00x
CAMELLIA192 1.27x 1.27x 1.23x 1.00x 1.25x 1.03x 1.20x 1.23x 1.00x 1.00x
CAMELLIA256 1.27x 1.27x 1.26x 1.00x 1.25x 1.03x 1.20x 1.23x 1.00x 1.00x
SALSA20 1.04x 1.00x

(Note: bulk encryption/decryption do burn_stack after full buffer processing,
instead of after each block.)

• Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>