A Horrible 8086 Star

Author:   atsampson
Category: Christmas Challenge
System:   IBM PC compatible
Language: 8086 assembler
Len source code: 761
Len exe file:    76
Len code only:   76
Instructions:
  dos1.com is a regular .COM file, which should in theory work on any PC DOS
  derivative (although it does make some assumptions about the initial values
  of registers that may not always hold). Run it with DOSBox using:
    dosbox dos1.com
Description:
  This was a late attempt, because I hadn't done an assembler version this year.
  I don't like x86 much, but it does have better code density than my preferred
  platforms! I'm sure there will be cleverer solutions, but I was happy enough
  to get it under 80 bytes.

  This uses the same approach as my UNIX v4 implementation, rearranged a bit to
  take advantage of the shorter x86 instructions: store one quarter of the
  pattern as an array of bit patterns, and mirror it horizontally and
  vertically, with a special case in the middle to draw 9 stars for the central
  row/column.

  In both axes it reuses the same code to scan in both directions: vertically
  it uses lodsb and flips the DF flag, and horizontally it uses some
  self-modifying code that toggles between rol and ror on alternate calls.
  Handily the bit that needs to toggle is 8, so I was able to tweak the inner
  loop to count from 8 down to 0 and xor it with the initial value of the loop
  counter, saving a byte.

  The choice between space and star bugs me. There's got to be a way of doing
  that in less than six bytes. I did wonder whether you could use aad to adjust
  by 10 somehow...

  Assembled using "nasm -f bin" on Linux.
Comments:
  Please can I have my RISC architectures back now, thanks.