5072961 Need an optimized MD5 implementation for amd64

*** 1,17 ****
  /*
!  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
   * Use is subject to license terms.
   */
  
  /*
   * Cleaned-up and optimized version of MD5, based on the reference
   * implementation provided in RFC 1321.  See RSA Copyright information
   * below.
   */
  
! #pragma ident   "@(#)md5.c      1.27    07/04/10 SMI"
  
  /*
   * MD5C.C - RSA Data Security, Inc., MD5 message-digest algorithm
   */
  
--- 1,17 ----
  /*
!  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
   * Use is subject to license terms.
   */
  
  /*
   * Cleaned-up and optimized version of MD5, based on the reference
   * implementation provided in RFC 1321.  See RSA Copyright information
   * below.
   */
  
! #pragma ident   "@(#)md5.c      1.28    08/01/02 SMI"
  
  /*
   * MD5C.C - RSA Data Security, Inc., MD5 message-digest algorithm
   */
  
*** 49,60 ****
--- 49,66 ----
  #ifdef _KERNEL
  #include <sys/systm.h>
  #endif /* _KERNEL */
  
  static void Encode(uint8_t *, const uint32_t *, size_t);
+ 
+ #if !defined(__amd64)
  static void MD5Transform(uint32_t, uint32_t, uint32_t, uint32_t, MD5_CTX *,
      const uint8_t [64]);
+ #else
+ void md5_block_asm_host_order(MD5_CTX *ctx, const void *inpp,
+     unsigned int input_length_in_blocks);
+ #endif /* !defined(__amd64) */
  
  static uint8_t PADDING[64] = { 0x80, /* all zeros */ };
  
  /*
   * F, G, H and I are the basic MD5 functions.
*** 241,250 ****
--- 247,259 ----
  {
          uint32_t                i, buf_index, buf_len;
  #ifdef  sun4v
          uint32_t                old_asi;
  #endif  /* sun4v */
+ #if defined(__amd64)
+         uint32_t                block_count;
+ #endif /* !defined(__amd64) */
          const unsigned char     *input = (const unsigned char *)inpp;
  
          /* compute (number of bytes computed so far) mod 64 */
          buf_index = (ctx->count[0] >> 3) & 0x3F;
  
*** 280,301 ****
--- 289,323 ----
  #endif /* sun4v */
  
                  if (buf_index) {
                          bcopy(input, &ctx->buf_un.buf8[buf_index], buf_len);
  
+ #if !defined(__amd64)
                          MD5Transform(ctx->state[0], ctx->state[1],
                              ctx->state[2], ctx->state[3], ctx,
                              ctx->buf_un.buf8);
+ #else
+                         md5_block_asm_host_order(ctx, ctx->buf_un.buf8, 1);
+ #endif /* !defined(__amd64) */
  
                          i = buf_len;
                  }
  
+ #if !defined(__amd64)
                  for (; i + 63 < input_len; i += 64)
                          MD5Transform(ctx->state[0], ctx->state[1],
                              ctx->state[2], ctx->state[3], ctx, &input[i]);
  
+ #else
+                 block_count = (input_len - i) >> 6;
+                 if (block_count > 0) {
+                         md5_block_asm_host_order(ctx, &input[i], block_count);
+                         i += block_count << 6;
+                 }
+ #endif /* !defined(__amd64) */
  
+ 
  #ifdef sun4v
                  /*
                   * Restore old %ASI value
                   */
                  set_little(old_asi);
*** 365,374 ****
--- 387,397 ----
          MD5Final(output, &context);
  }
  
  #endif  /* !_KERNEL */
  
+ #if !defined(__amd64)
  /*
   * sparc register window optimization:
   *
   * `a', `b', `c', and `d' are passed into MD5Transform explicitly
   * since it increases the number of registers available to the
*** 633,642 ****
--- 656,666 ----
           */
  
          x_0 = x_1  = x_2  = x_3  = x_4  = x_5  = x_6  = x_7 = x_8 = 0;
          x_9 = x_10 = x_11 = x_12 = x_13 = x_14 = x_15 = 0;
  }
+ #endif /* !defined(__amd64) */
  
  /*
   * Encode()
   *
   * purpose: to convert a list of numbers from big endian to little endian