On Thu, Aug 17, 2017 at 11:55 AM, Wilco Dijkstra <wilco.dijks...@arm.com> wrote: > Richard Biener wrote: >> On Tue, Aug 15, 2017 at 4:11 PM, Wilco Dijkstra <wilco.dijks...@arm.com> >> wrote: >> > Richard Biener wrote: >>>> > We also change the association of >>>> > >>>> > x / (y * C) -> (x / C) / y >>>> > >>>> > If C is a constant. >>>> >>>> Why's that profitable? >>> >>> It enables (x * C1) / (y * C2) -> (x * C1/C2) / y for example. >>> Also 1/y is now available to the reciprocal optimization, see >>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71026 for details. >> >> Sure, but on its own it's going to be slower. So this isn't the >> correct way to enable those followup transforms. > > How can it be any slower? It's one division and one multiply in both cases.
(x / C) / y is two divisions. If you restrict it to the case where we can transform this to (x * C') / y then it's indeed one. >>>> > x / (- y) -> (-x) / y >>>> >>>> Why? (it's only one of the possible canonicalizations) >>> >>> Same here, y is now available for reciprocal optimization. The >>> negate may now be optimized, for example (a * b) / -y -> (-a*b) / y >>> will use a negated multiple on various targets. >> >> Fair enough. Though if it were x / -(a*b) you'd regress that case. > > Possibly. You might still be able to merge the negate if the result is used > in an > addition or multiply, which wouldn't be possible if it were hiding in a > subexpression. > Without global analysis it seems best to move constants/negates to the > toplevel > if they can't be trivially removed in a subexpression. Eg. -x / (a * b * -c). Sure. So both patterns are canonicalization which is fine for match.pd. Those followup transforms should be done at a place that can look at more complicated patterns. We have the reassoc pass, then backprop (not exactly matching), and the recip pattern matching / cse pass. Richard. > Wilco
