On Wed, Apr 13, 2011 at 11:42 AM, Paul Richard Thomas <paul.richard.tho...@gmail.com> wrote: > Dear Dominique, > >> I think it is the automatic array in the subroutine trisolve. Note that the >> speedup is rather 27->19s and may be darwin specific (slow malloc). > > I saw a speed-up of similar order on FC9/x86_64. > > I strongly doubt that it is anything to do with the automatic array - > if it is there is an error somewhere, since none of the references to > trisolve need copy-in/copy-out. > >> >> Note also that -fstack-arrays prevents some optimizations on >> fatigue: 4.7->7s. This may be related to pr45810. > > Has PR45810 converged now? If I have understood properly, a patch has > been devised that cures the problem and does not cause slow-downs > anywhere else?
VLAs and malloc based arrays may behave differently with respect to alias analysis (I'd have to look at some examples). All effects other than malloc overhead I would attribute to that. That said, the general idea of the patch is sound and I see nothing wrong with it. Both performance improvements and regressions are worth looking at - they hint at possible improvements in the middle-end parts of the compiler. Richard. > Cheers > > Paul >