Hi Michael, thanks for your reply. I've built from scratch the Chapel version that includes your changes on my laptop and it works correctly. I'm mostly interested in LLVM-based communication optimizations for Coarray Fortran; mainly because the LLVM optimization passes are easier to create/modify than gfortran's. I've already read the paper you mentioned (Ahayashi et al.) and I'm quite suspicious about the generality of the proposed solution. I'll send you a private email about this pretty soon.
Thanks for your support. Regards, Alessandro 2016-08-26 9:52 GMT-06:00 Michael Ferguson <[email protected]>: > Hi Alessandro - > > Thanks very much for bringing up this issue, and for your interest > in Chapel and --llvm-wide-opt [1]. > > I've prepared a pull request that resolves the issue for me. > I'm still testing this PR but I hope it will be merged sometime > next week. It is here: > > https://github.com/chapel-lang/chapel/pull/4412 > > I'd be interested to know if that change resolves the issue > for you and also what you plan to do with --llvm-wide-opt. > > I'll note that while --llvm-wide-opt is an exciting way to do > communication optimization, it is still considered experimental. > In particular, it doesn't run in nightly testing (although > it probably should). I hope to improve upon it in the future > but recently I have needed to focus on other efforts. > > Besides --llvm-wide-opt being experimental, there are a number > of tasks that need to be done to really get the --llvm mode > working as well as it could. I've listed these TODOs in the PR > linked above. Of course, if you are interested in working on any > of these, I'd be happy to help you get started. Also, if you're > interested in using these LLVM-based communication optimizations > in a different compiler, that's something I'd like to discuss > as well. Discussion of either of these topics should probably > move over to [email protected]. > >>I installed the trunk version of Chapel on Edison (NERSC) choosing >>llvm as back-end. Even the simplest helloworld case (without any >>parallelism) leads to the following error (after about 2 minutes and >>33 seconds of compilation process...): > > I can't tell if it took you 2+ minutes to run 'make' or the 'chpl' command. > I'm guessing that you're talking about the 'chpl' command. On my system, > the chpl command takes under 15 seconds in this case. Perhaps that > compilation > mode is particularly slow on Edison for some reason? Nonetheless we are > aware that chpl compiles can take a long time. > >>$> chpl --llvm --llvm-wide-opt hello.chpl -o hello > > This is reasonable for testing, but note that if you want --llvm-wide-opt > to do anything useful, you need to also pass --fast or another flag that > causes most LLVM optimizations to run. > > Cheers, > > -michael > > [1] A paper about --llvm-wide-opt and also some other LLVM-based > communication optimization is available here: > > http://ahayashi.blogs.rice.edu/files/2013/07/Chapel_LLVM_camera_ready-q6usv > 4.pdf > ------------------------------------------------------------------------------ _______________________________________________ Chapel-bugs mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/chapel-bugs
