Multi-line comments on one line double the indent each time, up to an OutOfMemoryError

[abashev]

What

Reported upstream as google/google-java-format#413 (open since 2019-11-05): a 6 kB file of commented-out code ran out of memory with a 12 GB heap, and its output would have been about 3 GB. It reproduces here on main (53ac7d8). The shape is block comments that end and start on the same line, as in commented-out code that had javadoc in it:

/*
class X {
	*//** javadoc *//*
	public void foo(Bar bar) {}

	*//** javadoc *//*
	public void foo(Bar bar) {}

	*//** javadoc *//*
	public void foo(Bar bar) {}

	*//** javadoc *//*
	public void foo(Bar bar) {}
}*/

The continuation lines of these comments come out indented by 17, 50, 116 and 248 columns: each comment on the line starts twice as far right as the one before it. The file size grows the same way:

comments	input	output
6	317 B	4.3 kB
10	517 B	68 kB
16	817 B	4.3 MB

Why it happens

Comment.computeBreaks moves the column past a comment by adding the length of the comment's last line:

int firstLineLength = text.length() - Iterators.getLast(Newlines.lineOffsetIterator(text));
return state.withColumn(state.column() + firstLineLength)

For a one-line comment that is right. For a multi-line one, JavaCommentsHelper.rewrite has just re-indented the continuation lines to start at state.column(), so the last line already contains the column, and it is counted twice. The next comment on that line is re-indented from the doubled column, and so on. Despite its name, firstLineLength holds the length of the last line.

google-java-format has the same code in Doc.Tok.computeBreaks.

Fix

After a comment that spans lines, the column is the length of its last line; after a one-line comment it stays column + length.

The output changes only where something follows a multi-line comment on that comment's last line, so a corpus run should show next to nothing.

Done when

• a golden of the file above indents each comment's continuation lines to the column where that comment starts, and a second run changes nothing;
• a test formats twenty such comments into a small output (the old code writes about 70 MB);
• a corpus run says how many files change.