> Too late.  Since this has existed in binary form, you can't change
> it; such would break binary compatibility badly, since the layout of already
> compiled structures would change.

It wasn't clear from the mail that went by, but none of the suggested
changes affect structure layout. The layout code was already fine, it
was the block move code later on that was incorrectly inferring that
since 4 shorts make a double, it could use a floating point op to move
4 shorts, even though the layout code had aligned them as shorts
rather than doubles.

So it's all good.

David committed his fix to the gcc mainline this morning (Thanks,
David!), so I'll see about applying the patch to our sources soon.

-- Dan