DImode speedup on i386

[Bernd Schmidt]

> One general note -- you have to be *extremely* careful defining DImode
> patterns on the x86 due to the bad interactions with register
> reloading.

When the patch was written, I simply used the constraints from the addi3/subdi3
patterns to avoid problems.
Since it's about a year old (I've tried a couple of times to get it into
gcc2, without success), it's quite possible that things have changed. I'll
check that and implement Richard Henderson's suggestion, and send an updated
patch.

Bernd