Yo..
 > When you are moving to an ISA device, such as most VGA's on 386 systems,
 > the
 > data is being sent to the card a WORD at a time (ISA is 16-bit max)
 >  (barely any 386's have EISA, and none have PCI or VLB)
Just a quick point: I have 386 M/B's that DO have VLB. The Opti-495SLC 
chipset supports this configuration. It's rare though, and aligned REP MOVSD 
instructions are definately faster than REP MOVSW on these systems, but not 
by a factor of 2.. it's about a 30% improvement by my observations. Still, 
it's well worth having!
 > Also, if your REP MOVSD is misaligned, you again pick up problems even
 > if you
 > are writing to system memory, which then must break up your dwords into
 > words
 > (or even bytes) before writing.
With system memory and non-alinged writes, you may find that the cache mostly 
absorbs problems there. Of course, in non-cached systems it's very 
important.. but hey, this is 1997, not 1987... if you are still coding for 
1987 vintage systems....
    Craig
--- FMail/386 1.0g
---------------