> having just another "MathFloorOfDiv"-like mess in our code base
I agree that we should not pile optimization passes on top of each other instead of unifing them when possible. Folding this and similar passes into MathFloorOfDiv and turning MathFloorOfDiv into NarrowDiv might be a way to go. Unfortunately IR itself is standing between us and this optimization in it's generic form (see above about Simulate). So it seems for now we should let this problem be and return to it later. -- Vyacheslav Egorov On Fri, Aug 10, 2012 at 8:35 AM, Sven Panne <[email protected]> wrote: > Personally, I'd like to drop the patch for several reasons: It is adding a > considerable amount of complexity for very restricted use cases, and I > really don't like having just another "MathFloorOfDiv"-like mess in our code > base. Furthermore, the initial problem of undoing alpha pre-multiplication > can easily be done in a much more efficient way by using a small (64kB) > precomputed 2D table mapping a (color, alpha) pair to a color. Tables are > your best friends in any kind of signal processing. :-) > > The real underlying problem is that our general representation inference > algorithm is not very clever (a.k.a. naive), and it is currently > particularly bad for division (well, it's basically wrong there). Jakob has > already done some work in this area, and we should probably wait until his > stuff landed. > > Note that I am not saying that we should not improve our code generation for > e.g. floor-of-div or the division in your image processing code, it just > doesn't feel right to pile nearly unmaintainable hacks for special cases > onto each other. If we do things right, we should be able to do such > optimizations in a much more structured and cleaner way. > > > On Fri, Aug 10, 2012 at 1:50 AM, Evan Wallace <[email protected]> wrote: >> >> Just wanted to check up on this. Is the current decision to drop the >> patch? -- v8-dev mailing list [email protected] http://groups.google.com/group/v8-dev
