Repository navigation
Description
What happens
Reflowing a long string literal fails the whole file when the literal sits under something that binds tighter than +: a cast, the enclosing expression of a qualified .new, or a method reference. StringWrapper splits the literal into "a" + "b", which changes what the cast, the .new or the :: applies to:
(Object) "ab"becomes(Object) "a" + "b", that is((Object) "a") + "b""ab".new Inner()becomes"a" + "b".new Inner(), that is"a" + ("b".new Inner())"ab"::lengthbecomes"a" + "b"::length
The AST check at the end of wrapOnce notices this and throws, so the formatter prints "Something has gone terribly wrong. Please file a bug" and exits with 2. Nothing in the file is formatted, the rest of it included. With the 21.0.2-29 jar:
class T {
Object o = (Object) "one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt or to say anything";
Object p = "one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt or to say anything".new Inner();
java.util.function.IntSupplier s = "one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt or to say anything"::length;
String fine = "one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt or to say anything";
}open-java-format T.java fails on each of the first three fields on its own. --skip-reflowing-long-strings formats the file.
The scanner that picks the literals to reflow, LongStringsAndTextBlockScanner.visitLiteral, already leaves alone a literal that is the expression of a member select ("…".length()), for the same reason. It does not check the other contexts.
Why it matters
The check does its job: no wrong code comes out. But a single cast of a long literal is enough to make the formatter refuse the file, in the IDE as on the command line. (CharSequence) "…" as a method argument does it too.
Upstream has the same behaviour. google-java-format#1533, open, keeps the exception and pins it with wrap_typeCastParentFallsBackToFullAstCheck and wrap_newClassEnclosingExpressionFallsBackToFullAstCheck. The change for #250 ports both tests as stopsAReflowThatWouldMoveACastToTheFirstPiece and stopsAReflowThatWouldMoveAnInnerClassCreationToTheLastPiece. They pin the safety check, not the behaviour we want.
Proposal
LongStringsAndTextBlockScanner.visitLiteralskips a literal whose parent is a cast, aNewClassTreeit is the enclosing expression of, or a method reference, as it skips the expression of a member select. The literal stays on its line, as the formatter lays it out.- Wrapping the pieces in parentheses,
(Object) ("a" + "b"), would also keep the meaning. But it adds code that nobody wrote, and the member-select case already chooses to leave the literal alone. - The two ported tests then expect the input to come back as it is. A case file covers the three contexts plus a literal that is still reflowed.
After (the three literals as --skip-reflowing-long-strings lays them out, the last one reflowed):
class T {
Object o = (Object)
"one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt or to say anything";
Object p =
"one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt or to say anything"
.new Inner();
java.util.function.IntSupplier s =
"one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt or to say anything"
::length;
String fine =
"one long incredibly unbroken sentence moving from topic to topic so that no one had a chance to interrupt"
+ " or to say anything";
}
Activity