I had not anticipated this being really worthwhile, but it turns out

to have been so! Change MDx_HashFunction::hash to a new compress_n
which hashes an arbitrary number of blocks. I had a thought this might
reduce a bit of loop overhead but the results were far better than I
anticipated. Speedup across the board of about 2%, and very
noticable (+10%) increases for MD4 and Tiger (probably b/c both
of those have so few instructions in each iteration of the
compression function).

Before:
SHA-1:
  amd64: 211.9 MiB/s
  core: 210.0 MiB/s
  sse2: 295.2 MiB/s

MD4: 476.2 MiB/s
MD5: 355.2 MiB/s
SHA-256: 99.8 MiB/s
SHA-512: 151.4 MiB/s
RIPEMD-128: 326.9 MiB/s
RIPEMD-160: 225.1 MiB/s
Tiger: 214.8 MiB/s
Whirlpool: 38.4 MiB/s

After:
SHA-1:
  amd64: 215.6 MiB/s
  core: 213.8 MiB/s
  sse2: 299.9 MiB/s

MD4: 528.4 MiB/s
MD5: 368.8 MiB/s
SHA-256: 103.9 MiB/s
SHA-512: 156.8 MiB/s
RIPEMD-128: 334.8 MiB/s
RIPEMD-160: 229.7 MiB/s
Tiger: 240.7 MiB/s
Whirlpool: 38.6 MiB/s

1 files changed, 6 insertions, 2 deletions

diff --git a/src/hash/sha1_sse2/sha1_sse2.cpp b/src/hash/sha1_sse2/sha1_sse2.cpp
index 1d47cb1df..e2e0352fe 100644
--- a/src/hash/sha1_sse2/sha1_sse2.cpp
+++ b/src/hash/sha1_sse2/sha1_sse2.cpp
@@ -10,9 +10,13 @@ namespace Botan {
 /*************************************************
 * SHA-160 Compression Function                   *
 *************************************************/
-void SHA_160_SSE2::hash(const byte input[])
+void SHA_160_SSE2::compress_n(const byte input[], u32bit blocks)
    {
-   botan_sha1_sse2_compress(digest, reinterpret_cast<const u32bit*>(input));
+   for(u32bit i = 0; i != blocks; ++i)
+      {
+      botan_sha1_sse2_compress(digest, reinterpret_cast<const u32bit*>(input));
+      input += HASH_BLOCK_SIZE;
+      }
    }
 
 }