The following two patterns in MD5 intrinsic is inefficient.

1) Loading the updated hash values which were just written into the memory.
  md5_loop:
      __ ldrw(a, Address(state, 0)); // b) read the updated hash values which were just written
      ... // loop body
      __ ldrw(rscratch1, Address(state, 0));
      __ addw(rscratch1, rscratch1, a);
      __ strw(rscratch1, Address(state, 0)); // a) write the updated hash values into the memory
      ...
      __ br(Assembler::LE, md5_loop);

2) Redundant loads generated by md5_FF, md5_GG, md5_HH, and md5_II.
Loads generated by two md5_FFs and md5_GGs:
  __ ldrw(rscratch1, Address(buf, 0)); // from md5_FF(.., k = 0, ..)
  ...
  __ ldrw(rscratch1, Address(buf, 4)); // from md5_FF(.., k = 1, ..)
  ...
  __ ldrw(rscratch1, Address(buf, 4)); // from md5_GG(.., k = 1, ..)
  ...
  __ ldrw(rscratch1, Address(buf, 0)); // from md5_GG(.., k = 0, ..)