On Thu, 25 Apr 2024 13:34:50 GMT, Claes Redestad <redes...@openjdk.org> wrote:
>> When analyzing (startup) performance of the Classfile API I found this >> opportunity to further improve `MethodTypeDescImpl::descriptorString`. >> >> Performance improves across the board in existing microbenchmarks: >> >> Name >> (descString) Cnt Base Error Test Error Unit Change >> MethodTypeDescFactories.descriptorString >> (Ljava/lang/Object;Ljava/lang/String;)I 6 55,179 ± 2,027 32,920 ± 1,189 >> ns/op 1,68x (p = 0,000*) >> MethodTypeDescFactories.descriptorString >> ()V 6 17,689 ± 1,871 11,060 ± 0,331 ns/op 1,60x (p = 0,000*) >> MethodTypeDescFactories.descriptorString >> ([IJLjava/lang/String;Z)Ljava/util/List; 6 86,627 ± 1,646 41,035 ± 0,636 >> ns/op 2,11x (p = 0,000*) >> MethodTypeDescFactories.descriptorString >> ()[Ljava/lang/String; 6 18,305 ± 1,974 13,110 ± 0,089 ns/op 1,40x (p = >> 0,000*) >> * = significant >> >> >> The improvement is even more pronounced when running with `-Xint`, which is >> relevant for reducing startup overheads of early ClassFile API use: >> >> Name >> (descString) Cnt Base Error Test Error Unit Change >> MethodTypeDescFactories.descriptorString >> (Ljava/lang/Object;Ljava/lang/String;)I 6 5122,061 ± 81,335 2626,481 ± >> 101,466 ns/op 1,95x (p = 0,000*) >> MethodTypeDescFactories.descriptorString >> ()V 6 3481,316 ± 258,904 1489,267 ± 15,506 ns/op 2,34x (p = 0,000*) >> MethodTypeDescFactories.descriptorString >> ([IJLjava/lang/String;Z)Ljava/util/List; 6 7741,081 ± 1628,244 3281,778 ± >> 41,892 ns/op 2,36x (p = 0,000*) >> MethodTypeDescFactories.descriptorString >> ()[Ljava/lang/String; 6 3677,803 ± 63,432 1495,291 ± 8,995 ns/op >> 2,46x (p = 0,000*) >> * = significant >> ``` >> >> I also applied similar approach to `MethodTypeDesc::displayDescriptor`: >> while not performance sensitive I think these are so inter-related that it >> makes sense to implement them in a similar fashion. > > Claes Redestad has updated the pull request incrementally with one additional > commit since the last revision: > > comma-separated Do we have any research on the average length of method descriptor strings? I wonder if manual pre-allocation works better (iterating all descriptor strings, allocate the sum of their sizes plus 2 (for parentheses), as descriptor strings won't be re-calculated after initial allocation.) Especially in case of user code, as many user classes have very long package names that can easily make the string much longer than the 24-char default. ------------- PR Comment: https://git.openjdk.org/jdk/pull/18945#issuecomment-2077206533
