For rounds 16--79, W[i] only depends on W[i - 2], W[i - 7], W[i - 15] and W[i - 16].
Consequently, keeping all W[80] array on stack is unnecessary,
only 16 values are really needed.

Using W[16] instead of W[80] greatly reduces stack usage
(~750 bytes to ~340 bytes on x86_64).

Line by line explanation:
* BLEND_OP
  array is "circular" now, all indexes have to be modulo 16.
  Round number is positive, so remainder operation should be
  without surprises.

* initial full message scheduling is trimmed to first 16 values which
  come from data block, the rest is calculated before it's needed.

* original loop body is unrolled version of new SHA512_0_15 and
  SHA512_16_79 macros, unrolling was done to not do explicit variable
  renaming. Otherwise it's the very same code after preprocessing.
  See sha1_transform() code which does the same trick.

Patch survives in-tree crypto test and original bugreport test
(ping flood with hmac(sha512).

See FIPS 180-2 for SHA-512 definition
http://csrc.nist.gov/publications/fips/fips180-2/fips180-2withchangenotice.pdf



Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>