RFR: 8342692: C2: MemorySegment API slow with short running loops [v5]

[Roland Westrelin]

On Thu, 28 Nov 2024 14:42:23 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>> To optimize a long counted loop and long range checks in a long or int
>> counted loop, the loop is turned into a loop nest. When the loop has
>> few iterations, the overhead of having an outer loop whose backedge is
>> never taken, has a measurable cost. Furthermore, creating the loop
>> nest usually causes one iteration of the loop to be peeled so
>> predicates can be set up. If the loop is short running, then it's an
>> extra iteration that's run with range checks (compared to an int
>> counted loop with int range checks).
>> 
>> This change doesn't create a loop nest when:
>> 
>> 1- it can be determined statically at loop nest creation time that the
>>    loop runs for a short enough number of iterations
>>   
>> 2- profiling reports that the loop runs for no more than ShortLoopIter
>>    iterations (1000 by default).
>>   
>> For 2-, a guard is added which is implemented as yet another predicate.
>> 
>> While this change is in principle simple, I ran into a few
>> implementation issues:
>> 
>> - while c2 has a way to compute the number of iterations of an int
>>   counted loop, it doesn't have that for long counted loop. The
>>   existing logic for int counted loops promotes values to long to
>>   avoid overflows. I reworked it so it now works for both long and int
>>   counted loops.
>> 
>> - I added a new deoptimization reason (Reason_short_running_loop) for
>>   the new predicate. Given the number of iterations is narrowed down
>>   by the predicate, the limit of the loop after transformation is a
>>   cast node that's control dependent on the short running loop
>>   predicate. Because once the counted loop is transformed, it is
>>   likely that range check predicates will be inserted and they will
>>   depend on the limit, the short running loop predicate has to be the
>>   one that's further away from the loop entry. Now it is also possible
>>   that the limit before transformation depends on a predicate
>>   (TestShortRunningLongCountedLoopPredicatesClone is an example), we
>>   can have: new predicates inserted after the transformation that
>>   depend on the casted limit that itself depend on old predicates
>>   added before the transformation. To solve this cicular dependency,
>>   parse and assert predicates are cloned between the old predicates
>>   and the loop head. The cloned short running loop parse predicate is
>>   the one that's used to insert the short running loop predicate.
>> 
>> - In the case of a long counted loop, the loop is transformed into a
>>   regular loop with a ...
>
> Roland Westrelin has updated the pull request with a new target base due to a merge or a rebase. The pull request now contains 21 commits:
> 
>  - Merge branch 'master' into JDK-8342692
>  - whitespaces
>  - more
>  - merge
>  - more
>  - one more test
>  - Merge branch 'master' into JDK-8342692
>  - more
>  - more
>  - Merge branch 'master' into JDK-8342692
>  - ... and 11 more: https://git.openjdk.org/jdk/compare/3b21a298...74c38342

I pushed an update that should fix all test failures except the one in `compiler/escapeAnalysis/TestMissingAntiDependency.java` (covered by JDK-8341976). A lot of them were caused by the following part of the change:
> In the case of a long counted loop, the loop is transformed into a
regular loop with a new limit and transformed range checks that's
later turned into an in counted loop. The int counted loop doesn't
need loop limit checks because of the way it's constructed. There's
an assert that catches that we don't attempt to add one. I ran into
test failures where, by the time the int counted loop is created,
the fact that the number of iterations of the loop is small enough
to not need a loop limit check gets lost. I added a cast to make
sure the narrowed limit's type is not lost (I had to do something
similar for loop nests). But then, I ran into the same issue again
because the cast was pushed through a sub or add and the narrowed
type was lost. I propose that pushing casts through sub/add be only
done after loop opts are over (same as what's done for range check
CastII).

So I removed that part of the initial change and instead added some logic to pattern match the `CastLL` used by the loop nest for which the transformation of `(CastLL (AddL ...))` shouldn't be performed until the inner loop is turned into a counted loop.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/21630#issuecomment-2506391106