MD5 ASM Optimized Implementation

Overview

This project contains an optimized implementation of the MD5 hashing algorithm written in x86-64 assembly.
A naive C implementation is included as a reference for benchmarking and performance comparison.

---

TODO

• Do not allocate the buffer on the heap
• Compute padding / preprocessing at the end of the pipeline

---

Optimizations

The main optimizations are based on low-level instruction and register usage improvements.

• Loop unrolling is used where possible
• Constant values computed per message are precomputed and hardcoded
• The implementation is fully register-based:
  • No stack saving/restoring during core rounds
  • No use of global variables (.data section is avoided)
• The computation of F = ( B & C) | (~B & D) in the first 16 rounds can be calculated with its equivalent expresion F = D ^ ( B & ( C ^ D)), which takes less asm instructions. Same for the rounds 16-32, where we compute F as F = C ^ ( D & (B ^ C)).

Time comparisons

Naive C

$ time bin/md5_c "$(python3 -c "print('A'*100000)")"
5793f7e3037448b250ae716b43ece2c2

real	0m0.026s
user	0m0.017s
sys	    0m0.009s

Optimzed x86-64

$ time bin/md5_asm "$(python3 -c "print('A'*100000)")"
5793f7e3037448b250ae716b43ece2c2

real	0m0.016s
user	0m0.009s
sys	    0m0.009s