RFR: 8272573: Redundant unique_concrete_method_4 dependencies

[Vladimir Ivanov]

On Tue, 17 Aug 2021 19:39:19 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:

>> Equivalence checks in `Dependencies::assert_common_4()` are too strong which leads to `unique_concrete_method_4` dependencies being repeatedly recorded and duplicating nmethod dependencies registered.
>> 
>> The underlying problem is `note_dep_seen()` mutates internal bitmap, but the nested checks fail fast.
>> So, multiple requests are required to populate the bit-map for all the arguments in order to make the detection of duplicate assertions work. And it leads to duplicate dependencies registered.
>> 
>> It turns out `call_site_target_value` is also affected in a similar manner (see `Dependencies::assert_common_2` for dependencies without explicit context argument), so I fix it along the way.
>> 
>> Testing: hs-tier1 - hs-tier6
>
> src/hotspot/share/code/dependencies.cpp line 242:
> 
>> 240:     }
>> 241:   } else {
>> 242:     if (note_dep_seen(dept, x0)) {
> 
> Why `x0` is special in all these cases?
> Should we do `||` instead?
> 
>   if (note_dep_seen(dept, x0) || note_dep_seen(dept, x1)) {

It's just a fast path approximate check which guards linear search over recorded dependencies. 
`note_dep_seen(dept, x0) == false` says that such a dependency argument hasn't been seen before, but `note_dep_seen(dept, x0) == true` requires a search over recorded dependencies to prove the argument is in the right position.

Another way to fix it is to perform all the queries beforehand:

bool dep_seen_x0 = note_dep_seen(dept, x0);
bool dep_seen_x1 = note_dep_seen(dept, x1);
if (dep_seen_x0 && dep_seen_x1) {
  ...
}

It should dramatically reduce the rate of false positives.

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

PR: https://git.openjdk.java.net/jdk/pull/5141