Kelly,
As well as looking into parallelizing (parts of) javac, javac can also
provide
all the dependency info you need if you know how to get it. (Hint:
this means writing
code.) I'm hoping to write up a description of how to do this sooner
rather than later.
With the right code, you could make a very simple wrapper for javac
that would
accurately compile the correct subset of files that needed to be
compiled.
-- Jon
On May 10, 2008, at 11:43 AM, Kelly O'Hair wrote:
We already use 'make -j' when building the .o files, and the
dependencies
should be ok for the .o files. So for native compilations I don't
think we have a problem.
But with Java building, it's not clear 'make -j' is the answer or will
even work.
First, a .java can create multiple .class files, and second all the
javac builds read and write to the classes directory.
If a bunch of javac compilations were done in parallel I don't think
it would be a pretty sight. In general, make and java compilation
has always been complicated or non-robust to say the least.
Most people don't realize that compiling an explicit list of sources
can sometimes compile more than that, and that often the output
directory
is also used in the classpath as an input directory.
You might think this is a bad idea, but it's done, has been for some
time.
There are multiple possibilities for improving the javac compile time,
but doing it via make is not on my list.
We would like to look into javamake (soon to be open sourced)
http://www.experimentalstuff.com/Technologies/JavaMake/index.html
which does the right thing in terms of incremental builds.
It could be used with ant, or run directly from the makefiles.
And the javac team may be looking into actually paralizing the
internals
of javac soon, which would help without changing any build procedures.
-kto
Andrew Haley wrote:
Martin Buchholz wrote:
Unfortunately, incremental builds by running make in subdirectories
have historically been quite unreliable, and it has never been
highest priority
to fix them. Developers often do (at least in the jdk repository)
- try incremental build; it might work...
- else back off to rebuilding everything;
cd jdk/make; make clobber all
You could try to fix all of the subdirectory makefiles, but that's
an
enormous job.
For bonus points, you could try to fix the architecture of the
recursive makes themselves
using the ideas in "Recursive Make Considered Harmful"
http://miller.emu.id.au/pmiller/books/rmch/
Mmm. Surely if anyone is going to hack on the makefiles the first
thing
to do is to fix the dependencies so that "make -j" works properly.
Andrew.
