UMAC C Source Technical Notes
=============================

Contents
--------
1. ANSI C source can be forced...
2. Alignment of data...

-----------------------------------------------------------------------------

1. ANSI C source can be forced by setting the ANSI_C_ONLY preprocessor macro.
   If the macro is not set then specialized code is compiled under a few
   different compilers for Intel x86 and PowerPC. The level of support is
   different under each compiler, as indicated in the following table
   
               Rotate     Endian     NH-Asm     PRF-Asm
              x86  ppc   x86  ppc   x86  ppc   x86   ppc
           
      GCC      Y    Y     Y    Y     Y    N*    Y     N
   
      MSVC     Y    -     Y    -     Y    -     N     -
   
      MWERKS   N    Y     N    Y     Y    N*    N     N
   
      MSVC   = Microsoft Visual C++ 6.0
      GCC    = GNU C Compiler (gcc) 2.7
      MWERKS = Metrowerks C Compiler 2.3
      N*     = UMAC-MMX for ppc uses AltiVec C intrinsics if available.
      
   "Rotate" indicates that the processor's rotate instruction is forced.
   "Endian" indicates that the processor's fast byteswapping instruction
   is forced. "NH-Asm" and "PRF-Asm" indicate that the hash or PRF functions
   are written in the compiler's inline-assembly dialect.
   
   UMAC-MMX for x86 assumes the target processor supports MMX instructions
   unless ANSI_C_ONLY has been set.
   
   The specialized code for each compiler may work on newer and/or older
   versions, but this has not been tested.

-----------------------------------------------------------------------------

2. Alignment of data to be processed is important for correctness and maximum
   performance. Register loads to MMX registers from memory which are not
   aligned to an 8 byte boundary incur a significant penalty. Register loads
   to AltiVec regsters from memory not aligned to a 16 byte boundary result
   in incorrect execution of UMAC.
   
   For best performance pass UMAC buffers which are aligned to 16 byte
   boundaries and when passing buffers to "umac_update()" pass buffers of
   length divisible by 16 for all buffers except the last. These precautions
   will also avoid potential misaligned access faults on processors that cannot
   handle misaligned data. 

-----------------------------------------------------------------------------

UMAC C Source Revision History
==============================

Version 0.25.05 -- 1999 JULY 28
  - Added support for Visual C++
  - Fixed bug in NH-STD where intermediate additions were not MOD 2^32
  - Corrected opad/ipad constants in HMAC-SHA for UMAC-STD PRF.
  - Eliminated most unsigned types. Now only used to force logical shifts.

-----------------------------------------------------------------------------

This file was last modified on 28 July 1999.