On Thu, 16 Feb 2023 12:55:06 GMT, Maurizio Cimadamore <mcimadam...@openjdk.org> wrote:
>> Adam Sotona has updated the pull request incrementally with one additional >> commit since the last revision: >> >> added 4-byte Unicode text to Utf8EntryTest > > src/java.base/share/classes/jdk/internal/classfile/impl/SplitConstantPool.java > line 217: > >> 215: } >> 216: }; >> 217: // Doing a full scan here yields fall-off-the-cliff >> performance results, > > Understanding checkpoint: the parent is a class reader, which holds some > class bytes. The class reader is set up in a way that constant pool entries > are read "on-demand" - that is, if I ask the class reader "give me CP entry > 42", the classreader will only build that entry (assuming 42 is a valid > index) and give me back that entry. Also, the class reader will save the read > constant in an internal array, to avoid re-reading the classfile bytes > multiple times. > > So, what this code here does is: when we create an entry map, we populate the > map with the entries from the parent reader - but we only add entries that > were already looked up (e.g. we do not bother reading _all_ entries, and > adding all of them to the entry map). > > But, when `findEntry` is called, if we see that we can't find the entry, and > we know that there might be constant pool entries from the parent reader > still unread, we force a full scan of the parent reader pool and try again. > > Correct? Right, it is a multi-stage inflation, where partial-only inflation improves performance of frequent use cases significantly and full scan is a fall-back. ------------- PR: https://git.openjdk.org/jdk/pull/10982