Measure the whole type of a cast before inlining a chain onto it

[abashev]

Closes #60.

The bug

In sun.rmi.transport.tcp.TCPTransport the cast (PrivilegedAction<String>) came out cut between < and String, with the lambda it applies to on a third line, while the call around it stayed on one line:

    static final Log tcpLog = Log.getLog(
            "sun.rmi.transport.tcp", "tcp", LogStream.parseLevel(AccessController.doPrivileged((PrivilegedAction<
                            String>)
                    () -> System.getProperty("sun.rmi.transport.tcp.logLevel"))));

Why

The arguments of getLog are inlined when a prefix of the last one fits, and CountWidthUntilBreakVisitor measures that prefix up to the first break inside the last level. visitTypeCast declares its level ACCEPT_INLINE_CHAIN, which means that first break is the one before the operand. The first break was the one visitParameterizedType emits after <, so the prefix ended at (PrivilegedAction<, fit, and the type then had to take that break because String>) did not fit behind it. A wildcard bound (? extends List<Integer>), an intersection type (Runnable & Serializable) and an annotated type (@Nullable String) offer breaks of their own in the same place and split the same way.

The fix

The cast's level gets PartialInlineability.IF_FIRST_LEVEL_FITS, what visitDotWithPrefix uses for the same purpose: the visitor counts the whole first inner level, which is the rest of the type up to the break before the operand, when it fits. A chain is inlined onto a cast only if the type fits behind it. Here it does not, the chain is not inlined, and the arguments break one per line, as they do today when the type is not parameterized:

    static final Log tcpLog = Log.getLog(
            "sun.rmi.transport.tcp",
            "tcp",
            LogStream.parseLevel(AccessController.doPrivileged(
                    (PrivilegedAction<String>) () -> System.getProperty("sun.rmi.transport.tcp.logLevel"))));

The type stays breakable for a cast that does not fit on a line at all: laying the type out without breaks, the first option in #60, gave the same result here but turned the 231-character cast of the golden C into a 254-column line, so it was dropped.

Checked

• The golden ojf-issue-60-cast-type-in-inline-chain holds the statement above, the same cast one call shallower (which fit and does not change), and a method reference, two type arguments, a wildcard bound, a qualified name, an intersection type and an annotated type in the cast. It fails on main and passes with the change, and its output is stable on a second run.
• ./gradlew :open-java-format:test on JDK 21: 1545 tests, all green, the goldens C and B23749160 among them.
• Of the 15,747 files of the JDK 21 sources, that one statement in TCPTransport is the only one that formats differently, as A cast breaks inside its type when the call around it stays on one line #60 predicted; the 203 lambdas and 17 method references cast to a generic type elsewhere in the JDK are unchanged.

Version

A bug fix that changes the output of one shape, which the README allows in 2.x. The Migrate page on the website lists such fixes and should get a line for this one with the release that carries it.