Lower PADLOCK_CHUNK till value, which doesn't affect the benchmark results.

Well, it's even contrary, 512 was observed to *improve* performance by 5%.
Excuse ourselves from treating C7 specially.

diff --git a/crypto/engine/eng_padlock.c b/crypto/engine/eng_padlock.c
index 4e1eae3172e1cfc726d11cb497363ed038f6a7c7..1ba0d4302a437a7d20ba3a15628b2cb89881259c 100644 (file)
--- a/crypto/engine/eng_padlock.c
+++ b/crypto/engine/eng_padlock.c
@@ -878,7 +878,7 @@ padlock_aes_cipher_omnivorous(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
 }
 
 #ifndef  PADLOCK_CHUNK
-# define PADLOCK_CHUNK 4096    /* Must be a power of 2 larger than 16 */
+# define PADLOCK_CHUNK 512     /* Must be a power of 2 larger than 16 */
 #endif
 #if PADLOCK_CHUNK<16 || PADLOCK_CHUNK&(PADLOCK_CHUNK-1)
 # error "insane PADLOCK_CHUNK..."
@@ -905,6 +905,11 @@ padlock_aes_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out_arg,
        /* VIA promises CPUs that won't require alignment in the future.
           For now padlock_aes_align_required is initialized to 1 and
           the condition is never met... */
+       /* C7 core is capable to manage unaligned input in non-ECB[!]
+          mode, but performance penalties appear to be approximately
+          same as for software alignment below or ~3x. They promise to
+          improve it in the future, but for now we can just as well
+          pretend that it can only handle aligned input... */
        if (!padlock_aes_align_required)
                return padlock_aes_cipher_omnivorous(ctx, out_arg, in_arg, nbytes);