GeezerVison

Since the Pratt rewrite (5.4), &, |, #, << and >> have the same precedence as + and -. That gives 1 & 7 - 2 as (1 & 7) - 2 and 1 << n - 1 as (1 << n) - 1. PostgreSQL, MySQL and SQLite all bind + and - tighter:

5.3 got this right: BitwiseAndOr took AdditiveExpression operands.

The change gives & | # << >> their own level (prec 4), below + - || (prec 5). With Dialect.POSTGRESQL, PostgreSQL's || and its JSON operators already use prec 4. So in that dialect, all of PostgreSQL's "other operators" now share a single level, which matches PostgreSQL. A space-separated | | concat keeps the level of +. The bitwise operators share one left-associative level, as in PostgreSQL and SQLite. MySQL separates |, & and the shifts into different levels, and this PR doesn't change how that case is handled.

One existing assertion changes. testGetSign expected 1&2||3 to parse as (1 & 2) || 3. Outside the PostgreSQL dialect, || now binds like + (since the Pratt rewrite), and & binds looser than +. So || binds tighter than & and the expression parses as 1 & (2 || 3). That is also how SQLite reads it. With Dialect.POSTGRESQL the result is still (1 & 2) || 3, as in PostgreSQL, where SELECT 1 & 2 || 'x' gives '0x'. The test now asserts both. Please tell me if you'd rather keep the old default for that mix.

Also unsure: in T-SQL, & and | sit at the + - level, so with Dialect.SQLSERVER the master behaviour was right for a & b + c. 5.3 behaved like this PR. If you'd like, I can add a Dialect.SQLSERVER branch.

Tests: three cases are added to ExpressionPrecedenceTest, covering &, |, <<, >>, # (PostgreSQL dialect) against + -, the shared bitwise level, | | concat, and a round trip. testGetSign is updated as described above.

What was run:

• ./gradlew test --tests ExpressionPrecedenceTest without the grammar change: testBitwiseOperatorsBindLooserThanAddition and testSpaceSeparatedConcatStillBindsLikeAddition fail (expected: <BitwiseAnd> but was: <Subtraction> / <Concat>). With the change, both pass.
• ./gradlew test with the change but before updating testGetSign: 10,747 tests, only testGetSign failed.
• ./gradlew check (JDK 21, with a JDK 17 toolchain): BUILD SUCCESSFUL. That covers 10,748 tests with 0 failures, plus checkstyle, PMD, spotless, spotbugs and jacoco coverage verification.
• A differential check against PostgreSQL 17. Random integer and boolean expressions using + - * / % ^ & | # << >>, comparisons, BETWEEN, IN, IS NULL, NOT, CASE and ::int were parsed with JSqlParser. Each tree was re-printed fully parenthesised, and PostgreSQL evaluated both forms. On master, a few percent of the expressions came out different. With this change, the only mismatches left involved ^, which fix: bind ^ tighter than *, / and % #2831 fixed. With both changes applied, 0 of 8,623 expressions differ.

Not run: gradle jmh. The only extra work is a getToken(2) check when the next token is |.

This is rebased on #2831, which touched the same lines. After the rebase, ExpressionPrecedenceTest was run again (without the grammar change, testGetSign, testBitwiseOperatorsBindLooserThanAddition and testSpaceSeparatedConcatStillBindsLikeAddition fail; with it, all 23 pass) along with the full ./gradlew test (10,750 tests, 0 failures).