Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-8257594

C2 compiled checkcast of non-null object triggers endless deoptimization/recompilation cycle

    XMLWordPrintable

Details

    • b28

    Backports

      Description

        When C2 determines at compile time that there is no subtype relation between the objects involved in a checkcast, it inserts a null assert because the checkcast can only pass if the value is null. This will trigger deoptimization and recompilation if the object is non-null:

           1151 636 ! 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes) made not entrant
           1151 638 !b 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes)
           1152 638 ! 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes) made not entrant
           1152 639 !b 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes)
           1152 639 ! 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes) made not entrant
           1153 640 !b 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes)

        [...]

           1475 1030 ! 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes) made not entrant
           1475 1031 !b 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes)
        made not compilable on level 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes) decompile_count > PerMethodRecompilationCutoff
           1476 1031 ! 4 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes) made not entrant
           1476 1032 !b 1 compiler.uncommontrap.TestNullAssertAtCheckCast::test (17 bytes)

        However, we can do better here if we already know that the object is non-null (i.e. the check always fails) or if we hit too_many_traps_or_recompiles (which suggests that the object is often null) to avoid an endless deoptimization/recompilation cycle until we hit the PerMethodRecompilationCutoff and starve at lower tiers.

        In the former case we can simply throw right away and in the latter case we should fall back to a full check in case there is an exception handler in compiled code that handles the failing (non-null) case. That way we avoid deoptimization.

        I've spotted this with our test framework for inline types in project Valhalla, while investigating an intermittent failure where several test methods suddenly became non-compilable at tier 4.

        Attachments

          Issue Links

            Activity

              People

                thartmann Tobias Hartmann
                thartmann Tobias Hartmann
                Votes:
                0 Vote for this issue
                Watchers:
                4 Start watching this issue

                Dates

                  Created:
                  Updated:
                  Resolved: