On Tue, Sep 15, 2015 at 8:54 AM, Richard Biener <richard.guent...@gmail.com> wrote: > On Thu, Sep 3, 2015 at 5:22 PM, Sebastian Pop <seb...@gmail.com> wrote: >> Richard Biener wrote: >>> > * gcc.dg/graphite/uns-interchange-12.c: Adjust pattern to >>> > pass with >>> > both isl-0.12 and isl-0.15. >>> >>> Does it mean with 0.15 we now "time out" on some of the cases? >> >> "time out" will not trigger on the testcases modified in this patch. >> >>> Or is this >>> just a general difference between 0.12 and 0.15? In which case, like for >>> this testcase, is there a better way to verify whether the loops J and K >>> were >>> interchanged? >> >> We have more "tiled by" with isl-0.15 than with isl-0.12, so that means that >> the >> pattern we are looking for is not stable enough between isl versions: I will >> have to find and test for another pattern to check that loops have been >> blocked, >> interchanged, etc., which in my opinion is hard as we currently use different >> schedulers for different versions of isl. >> >> I have tuned the time out such that it will not trigger on the interchange >> testcases. It will trigger on a fortran testcase pr42334-1.f on which I have >> seen warnings of dejagnu timing out, and I have also tried on the reduced >> testcase attached to PR53852 which will time out with isl-0.15. I have not >> added PR53852's testcase as there still are people using isl-0.12 that would >> get >> another testcase that uses large amounts of memory and compile time. > > Btw, as all versions from GCC 4.8 on support ISL 0.14 it is reasonable to > remove > support for older ISL versions from trunk (so require 0.14 at least > which has the > timeout mechanism IIRC)
I will prepare a patch for trunk. Thanks for the idea. Sebastian
