RFR: 8294062: Improve parsing performance of j.l.c.MethodTypeDesc

Adam Sotona asotona at openjdk.org
Wed Sep 21 08:41:43 UTC 2022


On Tue, 20 Sep 2022 16:54:27 GMT, Paul Sandoz <psandoz at openjdk.org> wrote:

>> Parsing performance of `j.l.c.MethodTypeDescriptor::ofDescriptor` can be significantly improved and add performance boost to intensive users.
>> 
>> Two main reasons have been identified and fixed in this patch:
>> 
>> First glitch is late insertion of return type to the first position of `j.u.ArrayList`, causing all previously parsed parameter types to shift using `System::arraycopy`.
>> 
>> Proposed patch inserts a placeholder for the return type first and replaces it with return type later.
>> 
>> Second performance degradation is in `MethodTypeDescriptor::skipOverFieldSignature`, where each identified class descriptor is exactly located by `String::indexOf`, extracted by `String::substring`, and passed for for the purpose of verification to `ConstantUtils::verifyUnqualifiedClassName`.
>> 
>> Proposed patch inlines the unqualified class name verification into `ConstantUtils::skipOverFieldSignature` parsing loop and so eliminates calls of `String::indexOf` and `String::substring`. Package private method `ConstantUtils::verifyUnqualifiedClassName` has no other use within the package, so it is removed.
>> 
>> Please review proposed performance improving patch.
>> 
>> Thanks,
>> Adam
>
> src/java.base/share/classes/java/lang/constant/ConstantUtils.java line 217:
> 
>> 215:                             }
>> 216:                             case '/' -> {
>> 217:                                 if (!legal) return 0;
> 
> IIUC that works to reject `//` and trailing `/`, but does not reject leading `/`? which the previous code rejected.

`legal` is initially set to false, so leading `/` is also rejected

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

PR: https://git.openjdk.org/jdk/pull/10358


More information about the core-libs-dev mailing list