RFR: 8370947: Mitigate Neoverse-N1 erratum 1542419 negative impact on GCs and JIT performance [v28]

[Evgeny Astigeevich]

On Thu, 26 Feb 2026 10:14:48 GMT, Andrew Haley <aph at openjdk.org> wrote:

>> src/hotspot/share/code/relocInfo.cpp line 621:
>> 
>>> 619: 
>>> 620: 
>>> 621: bool metadata_Relocation::fix_metadata_relocation() {
>> 
>> I understand we return the status here, so that we avoid invalidation when there is no real patching work was done. Granted, it likely matches the behavior we have. But I wonder how much this buys us? If we are doing the deferred invalidation in a very broad scope, it stands to reason we would _almost definitely_ have to invalidate, and all this tracking would _nearly always_ give us the same answer, "Do invalidate"?
>> 
>> IOW, this might be an unnecessary complication of the interface.
>> 
>> _Dropping_ this change would also be more robust, in cases something somewhere _forgets_ to announce the code cache was changed? If we don't trust the downstream code about this and just summarily invalidate, it feels safer.
>
>> @shipilev
>> 
>> > But I wonder how much this buys us?
>> 
>> SPECjvm crypto.aes is the case for this change. When I ran it on Graviton 2, 64 cores I got the following scores:
>>
>>     * Baseline: 1047(min), 1058(max), 1052(geomean)
>> 
>>     * PR: 1058(min), 1078(max), 1069(geomean), +1.6% improvement
> 
> This improvement is almost in the noise. What's the average SPECjvm improvement?
> 
>> **OPTION 1**: If we don't track code modification in `fix_oop_relocations` and assume any call of `fix_oop_relocations` requires icache invalidation, results are the following:
>> 
>>     * 1032(min), 1057(max), 1045(geomean), -0.7% regression
> 
> So, Options 1 and 2 are actually worse than doing nothing?

@theRealAph 

> > SPECjvm crypto.aes is the case for this change. When I ran it on Graviton 2, 64 cores I got the following scores:
> > ```
> > * Baseline: 1047(min), 1058(max), 1052(geomean)
> > 
> > * PR: 1058(min), 1078(max), 1069(geomean), +1.6% improvement
> > ```
> 
> This improvement is almost in the noise. What's the average SPECjvm improvement?

I ran it 58 times with `-i 3 -ict -it 120 -wt 90`. The plan was to run 100 times but it would be taking a lot of time.

Baseline:

                    Mean
                     ↓
 ops/min        Frequency
 [1025,1030) │                              0
 [1030,1035) │ ██                           2
 [1035,1040) │ ████                         4
 [1040,1045) │ ████████                     8
 [1045,1050) │ ████████                     8
 [1050,1055) │ █████████                    9
 [1055,1060) │ █████████████               13  ← mode
 [1060,1065) │ █████                        5
 [1065,1070) │ ███                          3
 [1070,1075) │ ██                           2
 [1075,1080) │ ██                           2
 [1080,1085) │                              0
 [1085,1090) │ ██                           2
 [1090,1095) │                              0
              └──────────────────────────
              Mean = 1054.53    n = 58


Fix:

                      Mean
                       ↓
 ops/min        Frequency
 [1025,1030) │ █                            1
 [1030,1035) │                              0
 [1035,1040) │ █                            1
 [1040,1045) │ ████                         4
 [1045,1050) │ ██████                       6
 [1050,1055) │ ██████████                  10
 [1055,1060) │ ████████                     8
 [1060,1065) │ ████████                     8
 [1065,1070) │ ██████████                  10  ← mode (tied)
 [1070,1075) │ ███████                      7
 [1075,1080) │ █                            1
 [1080,1085) │                              0
 [1085,1090) │ █                            1
 [1090,1095) │ █                            1
              └──────────────────────────
              Mean = 1059.45    n = 58


Statistical tests, 4 out of 5 tests agree (p < 0.05) that the fix produces a statistically significant, small ~0.5% but consistent performance improvement.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/28328#discussion_r2872671390