https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68682

            Bug ID: 68682
           Summary: [6 Regression] [graphite] loop interchange no longer
                    working after r227277
           Product: gcc
           Version: 6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dominiq at lps dot ens.fr
                CC: spop at gcc dot gnu.org
  Target Milestone: ---

Up to revision r227276, compiling the test at
https://gcc.gnu.org/bugzilla/attachment.cgi?id=32703 (pr60997) with '-Ofast
-floop-interchange' gave the following timings

   94378416668672.000     
 Elapsed time =  0.55861600000000000      seconds
   94378416668672.000     
 Elapsed time =  0.56767100000000004      seconds
   94378416668672.000     
 Elapsed time =  0.54659700000000000      seconds
   94378416668672.000     
 Elapsed time =  0.54611900000000002      seconds

At r227277 this becomes with -floop-nest-optimize

   94378416668672.000     
 Elapsed time =   1.3123039999999999      seconds
   94378416668672.000     
 Elapsed time =  0.53941300000000003      seconds
   94378416668672.000     
 Elapsed time =   1.4528440000000000      seconds
   94378416668672.000     
 Elapsed time =  0.54144199999999998      seconds

then

   94378416668672.000     
 Elapsed time =   6.8884299999999996      seconds
   94378416668672.000     
 Elapsed time =  0.54318699999999998      seconds
   94378416668672.000     
 Elapsed time =   6.8123120000000004      seconds
   94378416668672.000     
 Elapsed time =  0.54194500000000001      seconds

with r228457 up to r230172 (bad loop exchange) and now (from r230325)

   94378416668672.000     
 Elapsed time =   3.7355589999999999      seconds
   94378416668672.000     
 Elapsed time =  0.54112400000000005      seconds
   94378416668672.000     
 Elapsed time =   6.8592209999999998      seconds
   94378416668672.000     
 Elapsed time =  0.53388500000000005      seconds

i.e., the same timings than with -Ofast alone (no loop exchange).

I am using ISL-0.15 for the latest tests (since Sep 29).

Reply via email to