With global optimization enabled, Microsoft Visual C++ 4.2 may change
the following inline asm sequence:

    cmp DL,0
    jne rarely_taken
  continue:
    nop
    jmp [pointer]
  rarely_taken:
    nop
    jmp continue

to:

    cmp DL,0
    jne rarely_taken
    jmp continue
    // some other stuff
  continue:
    nop
    jmp [pointer]
  rarely_taken:
    nop
    jmp continue

On the Pentium, this will cost 1 cycle. There may be an additional
performance loss because the unconditional jump needs an entry in the
Branch Target Buffer.

To prevent the re-arrangement of inline asm, insert:

  #pragma optimize("g", off)

before the function which contains inline assembler.