Performance improvements#1621
Conversation
Types like "string", "DateTime<'Y-m-d'>", "array<MyClass>" are re-parsed on every serialize/deserialize call. Cache eliminates redudant lexer/parser work. Wired as default in SerializerBuilder; users calling setTypeParser() still override entirely.
Avoid creating new ReflectionClass and running array_search() on every call. Use a static flipped constants map for O(1) lookup.
Replace O(n) linear scan with O(1) hash lookup using a static flipped array.
| self::$constantNames = array_flip((new \ReflectionClass(Lexer::class))->getConstants()); | ||
| } | ||
|
|
||
| return array_search($value, $oClass->getConstants()); |
There was a problem hiding this comment.
array_search defaulted to false if not found and would have caused crash as function has expected output of string.
After this change error might change, maybe add ?? false, but again this will just cause same type related error in php...
In my opinion this is such edge case it should not happen as it would mean this library being broken internally and would just add extra check which should not be ever needed. Any opinion on this could be helpful, or just leave it "as is".
Avoid creating new DateTimeZone objects on every deserialization call for the same timezone string. Instance-level caching.
Isolates type parsing performance: 9 realistic type strings, 1000 revs x 5 iterations, comparing raw Parser vs CachingParser.
uniqid() is just slow and not even cryptographically secure. The prefix only needs to be unique within a single XPath registerXPathNamespace() call, not globally. Replaced with simpler int counter.
…zationVisitor Added instance-leven caches for resolveNamespacePrefix() avoiding repeated SHA1 calculation and isElementNameValid() avoiding repeated preg_match for the same keys.
Cache the handler result from the polymorphic type check and reuse it if the type hasn't changed after the pre_serialize event dispatch, eliminating a redundant hash lookup per object node.
…trategy When nestedGroups is enabled, shouldSkipProperty() used in_array() O(n) to check group membership. Build a hashmap from getGroupsFor() result and use isset() O(1) instead.
Use three fast paths based on stack delta analysis to avoid full scans on every shouldSkipProperty()/shouldSkipClass() call. Same stack count retunrs cached result O(1). Stack shrunk and was not too deep O(1). Stack grew, no maxDepth on stack, not too deep, check only deltas O(delta). hasMaxDepthOnStack flag tracks if any maxDepth attributes exist on the stack. When false, growth never triggers full scan. When true or cazched result is too deep, the logic falls back to original full scan.
This comment was marked as outdated.
This comment was marked as outdated.
In PHP 7.x/8.x arrays outperform SplStack with noticable difference.
|
After these changes I can see some improvements (avg from 5 runs):
Memory usage unchanged across all benchmarks (~39.2 MB for JSON, ~11.7 MB for XML). @scyzoryck What do you think, is this something to finalize and prep for merge? I might still go trough and try minimize changes, but at this stage this is what we would get if I decide to continue this PR. |
|
Hey! |
|
👍 Just a side note, I found this to be varying between envs. On my workstation I get clean results, but again with older laptop the difference is smaller. Especially PHP 7 vs PHP 8. |
|
@NikoGrano can you rebase to rerun tests with the PHP8.5? |
2 participants
Improve library performance. See commits for exact details.