Isaac, Austin and I have actually just been email introduced to Tom, so excellent timing :) Happily, we've got a direct contact with the apple folks working on llvm / clang via a friend of mine, and thus in turn Tom and some others. The most important piece is making sure the Clang patches land in Xcode 5's version of Clang (which is apple's variant that isn't strictly tied to a point release), because thats going to be the only C compiler for OS X by default very soon. So in some respects, because BSD/Linux distros provide system package managers by default, its easy for a user on those platforms to have a suitable GCC installed, less so with vanilla OS X, so the most important part is making sure that these patches land in the Clang that lives in Xcode 5. If it lands in mainline clang point release too, great! (ghc not being able to useable on macs without the end user needing to build their own ghc or figure out installing and switching to an older xcode would be a HUGE headache for the haskell community) point being: exciting things! (and many many hearty thanks to Austin for the work to make this move along) -Carter On Wed, Jun 26, 2013 at 12:47 AM, Isaac Dupree < ml@isaac.cedarswampstudios.org> wrote:

Tom Stellard is leading an effort to make dot releases for LLVM 3.3 [*]; perhaps LLVM folks could be convinced to fix the issues in a 3.3.1 release.

-Isaac (just happens to be on both email lists and doesn't know anything more than this)

[*] http://lists.cs.uiuc.edu/pipermail/llvmdev/2013-June/063189.html

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

Unfortunately the two machines are fairly wildly different in their hardware characteristics. The OS X machine has 4GB of RAM; 8 Core i7. The Linux machine has 16GB, but only a 4-core i5. And they have different clock speeds. I'll get GCC 4.8 on my OS X machine so I can force a build with it and compare, but that'll take a while. Also, as I said, the Linux/Clang build technically has this very same bug too (GCTDecl.h needs to be modified slightly to fix this, because it *always* falls back to pthread_getspecific/setspecific currently with an LLVM based compiler. On Linux, we can just use __thread instead.) There, compared to a Linux/GCC build the change is approximately a ~30% difference on -threaded applications like gc_bench. So the slowdown ratios *seem* relatively consistent. Anyway, I'll get around to some full nofib runs today if possible, but the OS X machine will be a little sluggish. I have to do some other stuff today for this, anyway (like actually getting my second patch to Clang accepted today, hopefully.) Interested parties who'd like to see this change, as it stands, can look at my diff here (it's the 'clang-fast-tls' branch on my GHC fork): https://github.com/thoughtpolice/ghc/commit/88f0a0b047ff67b40eeb4de940aca162... On Fri, Jun 28, 2013 at 4:37 AM, Simon Marlow wrote:

On 26/06/13 04:13, Austin Seipp wrote:

...

Thanks Manuel!

I have an update on this work (I am also CC'ing glasgow-haskell-users, as I forgot last time.) The TL;DR is this:

* HEAD will correctly work with Clang 3.4svn on both Linux, and OS X. * I have a small, 6-line patch to Clang to fix the build failure in primitive (Clang was too eager to stringify something.) Once this fix is integrated into Clang (hopefully very soon,) it will be possible to build GHC entirely including all stage2 libraries without any patches. The patch is here: http://llvm.org/bugs/show_bug.cgi?id=16371 - I am hoping this will also make it into XCode 5. * I still have to eliminate some warnings throughout the build, which will require fiddling and a bit of refactoring. The testsuite still probably won't run cleanly on Linux, at least, until this is done I'm afraid (but then again I haven't tried...)

As for the infamous ticket #7602, the large performance regression on Mac OS X, I have some numbers finally between my fast-TLS and slow-TLS approach.

./gc_bench.slow-tls 19 500000 5 22 +RTS -H180m -N7 -RTS 395.57s user 173.18s system 138% cpu 6:50.71 total

vs

./gc_bench.fast-tls 19 500000 5 22 +RTS -H180m -N7 -RTS 322.98s user 132.37s system 132% cpu 5:44.65 total

Now, this probably looks totally awful from a scalability POV. And, well, yeah, it is. But I am almost 100% certain there is something extremely screwy going on with my machine here. I base this on the fact that during gc_bench, kernel_task was eating up about ~600% of my CPU consistently, giving user threads no time to run. I've noticed this with other applications that were totally unrelated too (close tweetbot -> 800% CPU usage,) so I guess it's time to learn DTrace. Or turn it on and off again or something. Ugh.

Anyway, if you look at the user times, you get a nice 30% speedup which is about what we expect!

30% better than before is good, but we need some absolute figures. Can you validate that against the performance on Linux, or against the performance you get when the RTS is compiled with gcc? If it's hard to get a direct comparison on equivalent hardware. you could compare the slowdown with -threaded on Linux and OS X.

Cheers, Simon

...

On a related note, due to the source code structure at the moment, Linux/Clang hilariously suffers from this same bug. That's because while Clang on Linux supports extremely fast TLS via __thread (like GCC,) it falls back to pthread_getspecific/setspecific. I haven't fixed this yet. It'll happen after I fix #7602 and get it merged in. On my Linux machine, gc_bench also sees a consistent 30% speedup between these two approaches, so I think this is a relatively accurate measurement. Well, as accurate as I can be without running nofib just yet. So if you're just dying to have GHC HEAD built with Clang HEAD on Linux because you've got reasons, you should probably hold on.

I also may have a similar, better approach to fixing #7602 that is not entirely as evil and sneaky as crashing the WebKit party. I'll follow up on this soon when I have more info in a separate thread to confer with Simon. With nofib results. I hope.

But anyway, 7.8 will be shaping up quite nicely - in particular in the Mac OS X department, I hope. Please feel free to pester me with questions or if you attempt something and it doesn't work.

On Tue, Jun 25, 2013 at 7:34 PM, Manuel M T Chakravarty wrote:

...

Austin,

Thank you very much for taking care of all these clang issues — that is very helpful!

Cheers, Manuel

Austin Seipp :

...

Hi all,

As of commit 5dc74f it should now be possible to build a working stage1 and stage2 compiler with (an extremely recent) Clang. With some caveats.

You can just do:

$ CC=/path/to/clang ./configure --with-gcc=/path/to/clang $ make

I have done this work on Linux. I don't expect much difficulty on Mac OS X, but it needs testing. Ditto with Windows, although Clang/mingw is considered experimental.

The current caveats are:

* The testsuite will probably fail everywhere, because of some warnings that happen during the linking phase when you invoke the built compiler. So the testsuite runner will probably be unhappy. Clang is very noisy about unused options, unlike GCC. That needs to be fixed somewhere in DriverPipeline I'd guess, but with some refactoring. * Some of the stage2 libraries don't build due to a Clang bug. These are vector/primitive/dph so far. * There is no buildbot or anything to cover it.

You will need a very recent Clang. Due to this bug (preventing primitive etc from building,) you'll preferably want to use an SVN checkout from about 6 hours ago at latest:

http://llvm.org/bugs/show_bug.cgi?id=16363

Hilariously, this bug was tripped on primitive's Data.Primitive.Types module due to some CPP weirdness. But even with a proper bugfix and no segfault, it still fails to correctly parse this same module with the same CPP declarations. I'm fairly certain this is another bug in Clang, but I might be wrong. I'm trying to isolate it. Unfortunately Clang/LLVM 3.3 was just released and it won't see bugfix releases. But it will *probably* work if we just get rid of the CPP tomfoolery in primitive. I'll be testing it in the next few days to see if we can get 3.3 supported. (I'm sort of kicking myself in the head for not doing this a week or two ago...)

Anyway, there are some rough edges but it should be in shape for 7.8 I hope. It should be especially welcome for Mac users. (I'm also hoping modern Macs could even go all-clang-all-the-time if my fix for #7602 can go in soon...)

PS. If you use ASSERT, I just fixed a lot of instances of using that macro, involving whitespace between the name and arguments (commit d8ee2b.) Clang does not forgive you for this. Should I note this anywhere for the future in the wiki or something?

-- Regards, Austin - PGP: 4096R/0x91384671

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

-- Regards, Austin - PGP: 4096R/0x91384671

Err, GCC replacement. But, ironically, GHC [backend] replacement as well, as of the recent ICFP paper. On Mon, Jul 1, 2013 at 12:03 PM, Ryan Newton wrote:

By the way, on the topic of these preformance comparisons -- has anyone tried building the RTS with the Intel C compiler? (They try very hard at being a drop-in GHC replacement.)