------- Additional Comments From rakdver at atrey dot karlin dot mff dot cuni dot cz 2005-01-25 01:57 ------- Subject: Re: [4.0 Regression] IV-OPTS is O(N^3)
> >> * tree-data-ref.c (analyze_overlapping_iterations): chrecs that > >> are equal overlap on every iteration. > >> > >> This stuff is just simple symbolic differencing, and checking of > >> invariantness of symbols, > >> But it is indeed starting to se the symbolic scev info > > > > Ugh... OK, if you think it is a right way... Anyway, I am seriously > > considering resurrecting the simple iv analysis for purposes of passes > > that are not interested in this fancy stuff. Overhead of scev is > > probably acceptable if it is only used for this type of analysis, > > but for other purposes it is clearly overkill. > > Uh, almost all high level loop optimizations are going to do very badly in > the depth of the loop nest. Remove "almost" and "high level" from this sentence and I will agree with you. This really does not worry me that much. > The fact that scev doesn't do so well on a 100 deep loop nest doesn't > really concern me at all. Me neither. Its time and memory consumption on normal testcases however is not entirely satisfactory (not critical, but there definitely is a place for improvement), and implementation could (should?) also be significantly cleaned up. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18595
