Repository navigation
[BUG] JSQLParser 5.4-SNAPSHOT : performance.sql contains unparseable @Prompt macros; parseStatements used to hide this by returning null (JSQLParserBenchmark measured incomplete parses) #2599
Description
Activity
- changed the title
[-][BUG] JSQLParser 5.4-SNAPSHOT : benchmark corpus no longer parses (IN @Prompt(...)), JSQLParserBenchmark fails on master[/-][+][BUG] JSQLParser 5.4-SNAPSHOT : performance.sql contains unparseable @Prompt macros; parseStatements used to hide this by returning null (JSQLParserBenchmark measured incomplete parses)[/+]on Sep 11, 2026 Correction after re-verifying with a probe that checks the return value: the corpus statement with the
@Promptmacros fails at the same position on both6c726d8and current master. The only change is that #2568 / #2594 replaced the silentnullreturn ofparseStatementswith a proper exception, so the benchmark now fails instead of silently timing an incomplete parse. Title and description are updated accordingly.Thank you for bisecting this, in my opinion we should go this route: "should the @prompt statement be repaired".
Thank you, fixed -- with interesting outcome:
Benchmark (version) Mode Cnt Score Error Units JSQLParserBenchmark.parseSQLStatements latest avgt 15 32.135 ± 0.820 ms/op JSQLParserBenchmark.parseSQLStatements 5.3 avgt 15 965.707 ± 15.947 ms/op JSQLParserBenchmark.parseSQLStatements 5.1 avgt 15 314.454 ± 3.717 ms/opThe 5.1 and 5.3 numbers measure incomplete parses and cannot be compared with latest: both versions still return null silently from
parseStatementswhen a statement fails (#2576), and the repaired corpus still fails on them.- Per-version result on the repaired corpus (null/size-checked, 40 warmup rounds, median of 30, same host):
- 5.1: null after 283 ms (3 statements fail, incl. the trailing
WITH) - 5.2: complete, 54 statements, 619 ms
- 5.3: null after 1089 ms (2 statements fail)
- master
c7f8ff89: complete, 54 statements, ~35 ms (consistent with the 32.1 above)
- 5.1: null after 283 ms (3 statements fail, incl. the trailing
- On 5.3 the ~80 KB statement at line 194 fails with
TimeoutException: that is the super-linear lookahead from [BUG] Parse time grows super-linearly on long dotted-name chains and nested parentheses (phase-2 lookahead), StackOverflowError at nesting depth ~2000 #2519. 5.2 parses it fine, so the cliff landed between 5.2 and 5.3, not gradually. - The comparable pair is 5.2 vs master: 619 ms -> ~35 ms (~18x), mostly fix: memoize long-chain predicate walks behind a size threshold #2520 (memoized long-chain predicate walks); 5.2 barely moves between cold and warm runs because the cliff dominates its profile.
- The benchmark could assert non-null and the expected statement count per version, so truncated timings cannot pass as results: this is also what had hidden the corpus damage for years.
- Per-version result on the repaired corpus (null/size-checked, 40 warmup rounds, median of 30, same host):
Thank you, noted, but almost irrelevant now:
- we got way faster
- and we got way more feature complete and correct
The fact, that there is even more story to tell about the details does not matter for most of the users.
AI disclosure: this issue was prepared with AI coding agents, reviewed and revised line by line by me.
Failing SQL Feature:
src/test/resources/net/sf/jsqlparser/performance.sqlcontains BusinessObjects@Prompt(...)macros that JSQLParser cannot parse (and, as far as I can tell, never could: the statement fails at the same position on6c726d8and on current master, standalone or as part of the file).parseStatementssilently returnednullwhen a statement failed. As a consequenceJSQLParserBenchmark.parseSQLStatementsproduced timing numbers while the corpus was not fully parsed (thenullresult was simply consumed). With the fix, the benchmark now correctly fails fast with theParseExceptionbelow.@Promptstatement be repaired or supported?SQL Example:
Excerpt of the failing spot (corpus line 497):
Repro on both commits (statement alone, corpus lines 196-1688):
Software Information:
7cc86386; same parse failure on6c726d8)@Promptmacros)Tips:
x IN @Prompt(p1)orSELECT @f(1)fail on both commits; the corpus statement is the only place this surfaces in the repo.6c726d8the JMH score described an incomplete parse of the corpus.