davem at davemloft.net
Fri Apr 12 21:55:38 CEST 2013
From: Torbjorn Granlund <tg at gmplib.org>
Date: Fri, 12 Apr 2013 20:10:13 +0200
> If we could trust movXX to be silent, we should of course not bother to
> create that mask, but replace that and the logops with 4 moveq for a
> great speedup...
BTW, even if we could do this, this would be really bad on Ultra1&2
("moveq" is single issue and 3 cycle latency) and suboptimal on
Ultra3&4 (only one "moveq" dispatched per cycle).
Only T4&T5 would really benefit from this sort of approach.
More information about the gmp-devel