On Sat, Mar 12, 2011 at 1:44 AM, Xinliang David Li <davi...@google.com> wrote: > On Fri, Mar 11, 2011 at 11:50 AM, Jeff Law <l...@redhat.com> wrote: >> -----BEGIN PGP SIGNED MESSAGE----- >> Hash: SHA1 >> >> On 03/09/11 09:24, Xinliang David Li wrote: >>> On Wed, Mar 9, 2011 at 6:03 AM, Jeff Law <l...@redhat.com> wrote: >>> On 03/09/11 02:45, Richard Guenther wrote: >>>>>> On Tue, Mar 8, 2011 at 11:04 PM, Jeff Law <l...@redhat.com> wrote: >>> >>>>>> True. I've been repeatedly thinking of building some on-the-side CFG >>>>>> with value-numbered predicates to also catch the CFG vs. scalar-code >>>>>> predicate combinations we have. On such on-the-side data structure >>>>>> we could do very aggressive jump-threading just for analysis purposes >>>>>> (experiments when working on separating conditions shows that >>>>>> a PRE-like algorithm could drive this). >>> I'm pondering the same kind of thing. PRE on the predicates with a more >>> structured approach to block copying to isolate the paths seems to be >>> the way to go. >>> >>> >>>> Any simple examples to show how your idea would work? >> Our current jump threading is just a special case of path isolation; the >> biggest problem with path isolation is exploding codesize. > > Jump threading can reduce size in some cases -- due to more aggressive > cleanups -- the -Os need to be more intelligent about it. > >> >> I'd like to see that code generalized in a few ways based on well known >> algorithms rather than the ad-hoc stuff we've got now keeping a >> reasonable knob on the codesize bloat. >> >> In cases where we want more accurate warnings, but aren't willing to >> accept the size bloat, on the side structures might be the way to go. >> >> I'm still looking a literature on the subject, but there's clearly >> things we can do to improve the path sensitivity of the optimizers and >> warning analysis. > > There are actually many papers on this (e.g. IMPACT group) as a > result of research for EPIC optimizations (predication and if > convert). Most of them talk about predicate query systems (efficient > and accurate), e.g PQS using partition graph, PAS using BDD etc. The > predicates relation can be built as side data structure, and the > following queries can be made: > > p1 = get_predicate (gimple_stmt1); > p2 = get_predicate (gimple_stmt2); > > if (is_disjoint (p1, p2) ) > .. > else if is_subset(p1, p2). > .. > > Keeping the information up to date with transformations can be difficult.
Sure, I'd say this is a clear candidate for something that would get re-computed if needed. Richard.
