Proposed javac argument processing performance improvement

From David Schlosnagle, via compiler-dev
http://mail.openjdk.java.net/pipermail/compiler-dev/2011-June/003335.html
For full details, see the mail thread.

I'd like to propose a minor change for javac to improve performance
processing large numbers of filename arguments, especially on Windows.
The main issue is that com.sun.tools.javac.main.Main currently uses a
ListBuffer for the collection of filenames. Main.addFile calls
ListBuffer.contains before calling ListBuffer.add, both of which are
linear searches leading to O(N^2) performance. Additionally, the
implementation of java.io.File.equals uses the underlying filesystem's
compare method which on Windows requires an expensive case insensitive
string comparison. For large numbers of files (I work with several
modules of approximately 10,000 Java input files for a single javac
invocation) so this is a big performance hit. The simple fix is to use
a LinkedHashSet<File> instead of ListBuffer<File> for the filenames
field in com.sun.tools.javac.main.Main (see attached patch).

My preliminary tests show that execution time of
com.sun.tools.javac.main.Main.processArgs method for 10,000 input
files on my Windows machine went from around 15 seconds to around 300
milliseconds with the patch. The performance improvement on
case-sensitive filesystems isn't as good, but still seems to be an
order of magnitude as seen in the following results on my main Mac OS
X machine for com.sun.tools.javac.main.Main.processArgs method
execution time.

# files Before (ms) After (ms) Delta (ms)
------- ----------- ---------- ----------
1 0.6 1.1 0.5
1000 39.4 39.7 0.3
2000 94.4 41.4 (53.1)
3000 184.8 61.4 (123.4)
4000 185.2 77.2 (108.0)
5000 257.9 89.1 (168.7)
6000 358.5 119.9 (238.6)
7000 422.1 126.7 (295.4)
8000 624.9 137.8 (487.2)
9000 620.1 139.8 (480.4)
10000 1,054.9 149.4 (905.5)


Thanks,
Dave