Algorithm loops in PIII assembly: Help requested

Hello,

Here is a rather long question with several parts. Hopefully there are some patient experts out there...

I'm developing a multiplication algorithm in a mixed C++/ASM project. The critical inner loops are shown in the C++ language below. I would like to work out a hand-crafted PentiumIII assembler implementation of these inner loops.

I have moderate experience with such matters and would be in a position to design such loops using 386 assembler. However, I would very much like to find out if it is possible to improve run-time characteristics using MMX instructions and 64-bit registers. I have never programmed MMX assembler. I do not want to go all the way up to PentiumIV and SSE2, but rather stay with MMX since a lot of clients do not have support for SSE2.

The questions are then:

1) In the algorithm below, can the variables carry and sum be effectively manipulated using 64-bit MMX registers and are there any advantages here?
2) For the multiplication in the inner loop, is there a MMX method for u32xu32-->u64 multiply which superior to the old mul command with results in eax:edx?

The algorithm in C++ is shown here. Please take a look.

static void test_mul_loop(const UINT32* const pu,
                          const UINT32* const pv,
                                UINT32* const pw)
{
  // Standard Order(N^2) multiplication in C++, taken from
  // an empirical analysis of polynomial multiplication
  // created for several cases using Mathematica 4.1.
  // Constraint on assembler implementation: The array
  // 'number' can not be unlimited in size since the division
  // carry = sum / mask should remain u64/u32--> u32 result
  // with u32 remainder.
  UINT64 carry = 0;
  for(INT32 j = number - 1; j >= 0; j--)
  {
    UINT64 sum = carry;
    for(INT32 i = 0; i <= j; i++)
    {
      sum += pu[j - i] * static_cast<UINT64>(pv[i]);
    }
    pw[j + 1] = static_cast<UINT32>(sum % mask);
    carry     = sum / mask;
  }
}

Thank you very much for any assistance.

Sincerely, Chris.