RFR: 8329570: G1: Excessive is_obj_dead_cond calls in verification
Stefan Karlsson
stefank at openjdk.org
Wed Apr 3 11:15:11 UTC 2024
On Wed, 3 Apr 2024 10:42:27 GMT, Guoxiong Li <gli at openjdk.org> wrote:
>> While investigating [JDK-8329314](https://bugs.openjdk.org/browse/JDK-8329314) it was found that G1 was slower than Parallel when allocating objects because of the extra verification when running with fastdebug builds.
>>
>> This tiny tests takes 30s with G1:
>>
>> public class Test {
>> static int n = 847734685;
>>
>> public static void main(String[] args) {
>> String[] strs = new String[n + 1];
>> }
>> }
>>
>>
>> A large portion of that time is spent inside this assert:
>>
>> assert(!G1CollectedHeap::heap()->is_obj_dead_cond(_containing_obj, _vo), "Precondition");
>>
>> which is called for every single oop in the _containing_obj. If I move this assert to where _containing_obj is initialized, the test now completes in 4-5 seconds. I propose that we make this tiny change to improve the speed of the debug builds.
>
> src/hotspot/share/gc/g1/g1HeapRegion.cpp line 627:
>
>> 625: void set_containing_obj(oop const obj) {
>> 626: assert(!G1CollectedHeap::heap()->is_obj_dead_cond(obj, _vo), "Precondition");
>> 627: _containing_obj = obj;
>
> Should the method `set_containing_obj` judge whether the `obj` is `null`? Just like the `do_oop_work`.
Sounds like a reasonable thing to do. I'll move that assert as well.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/18595#discussion_r1549505155
More information about the hotspot-gc-dev
mailing list