Re: LLVM and dynamic linking

Personally I'd be in favor of that to keep it easy, but there hasn't really been any poll about what to do. For the most part it tends to work fine, but I think it's the wrong thing to do in any case. IMO the truly 'correct' thing to do, is not to rely on the system LLVM at all, but a version specifically tested with and distributed with GHC. This can be a private binary only we use. We already do this with MinGW on Windows actually, because in practice, relying on versions 'in the wild' is somewhat troublesome. In our case, we really just need the bitcode compiler and optimizer, which are pretty small pieces. Relying on a moving target like the system install or whatever possible random XYZ install from SVN or (some derivative forked toolchain!) is problematic for developers, and users invariably want to try new combinations, which can break in subtle or odd ways. I think it's more sensible and straightforward - for the vast majority of users and use-cases - for us to pick version that is tested, reliably works and optimizes code well, and ship that. Then users just know '-fasm is faster for compiling, -fllvm will optimize better for some code.' That's all they really need to know. If LLVM is to be considered 'stable' for Tier 1 GHC platforms, I'm sympathetic to Aaron's argument, and I'd say it should be held to the same standards as the NCG. That means it should be considered a reliable option and we should vet it to reasonable standards, even if it's a bit more work. It's just really hard to do that right now. But I think implementing this wouldn't be difficult, it just has some sticky bits about how to do it. We can of course upgrade it over time - but I think trying to hit moving targets in the wild is a bad long-term solution. On Tue, Jan 7, 2014 at 3:07 PM, George Colpitts wrote:

wrt

We support a wide range of LLVM versions

Why can't we stop doing that and only support one or two, e.g. GHC 7.8 would only support llvm 3.3 and perhaps 3.4?

On Tue, Jan 7, 2014 at 4:54 PM, Austin Seipp wrote:

...

Hi all,

Apologies for the late reply.

First off, one thing to note wrt GMP: GMP is an LGPL library which we link against. Technically, we need to allow relinking to be compliant and free of of the LGPL for our own executables, but this should be reasonably possible - on systems where there is a system-wide GMP installed, we use that copy (this occurs mostly on OSX and Linux.) And so do executables compiled by GHC. Even when GHC uses static linking or dynamic linking for haskell code in this case, it will still always dynamically link to libgmp - meaning replacing the shared object should be possible. This is just the way modern Linux/OSX systems distribute system-wide C libraries, as you expect.

In the case where we don't have this, we build our own copy of libgmp inside the source tree and use that instead. That said there are other reasons why we might want to be free of GMP entirely, but that's neither here nor there. In any case, the issue is pretty orthogonal to LLVM, dynamic haskell linking, etc - on a Linux system, you should reasonably be able to swap out a `libgmp.so` for another modified copy[1], and your Haskell programs should be compliant in this regard.[2]

Now, as for LLVM.

For one, LLVM actually is a 'relatively' cheap backend to have around. I say LLVM is 'relatively' cheap because All External Dependencies Have A Cost. The code is reasonably small, and in any case GHC still does most of the heavy lifting - the LLVM backend and native code generator share a very large amount of code. We don't really duplicate optimizations ourselves, for example, and some optimizations we do perform on our IR can't be done by LLVM anyway (it doesn't have enough information.)

But LLVM has some very notable costs for GHC developers:

* It's slower to compile with, because it tries to re-optimize the code we give it, but it mostly accomplishes nothing beyond advanced optimizations like vectorization/scalar evolution. * We support a wide range of LLVM versions (a nightmare IMO) which means pinning down specific versions and supporting them all is rather difficult. Combined with e.g. distro maintainers who may patch bugs themselves, and the things you're depending on in the wild (or what users might report bugs with) aren't as solid or well understood. * LLVM is extremely large, extremely complex, and the amount of people who can sensibly work on both GHC and LLVM are few and far inbetween. So fixing these issues is time consuming, difficult, and mostly tedious grunt work.

All this basically sums up to the fact that dealing with LLVM comes with complications all on its own that makes it a different kind of beast to handle.

So, the LLVM backend definitely needs some love. All of these things are solveable (and I have some ideas for solving most of them,) but none of them will quite come for free. But there are some real improvements that can be made here I think, and make LLVM much more smoothly supported for GHC itself. If you'd like to help it'd be really appreciated - I'd like to see LLVM have more love put forth, but it's a lot of work of course!.

(Finally, in reference to the last point: I am in the obvious minority, but I am favorable to having the native code generator around, even if it's a bit old and crufty these days - at least it's small, fast and simple enough to be grokked and hacked on, and I don't think it fragments development all that much. By comparison, LLVM is a mammoth beast of incredible size with a sizeable entry barrier IMO. I think there's merit to having both a simple, 'obviously working' option in addition to the heavy duty one.)

[1] Relevant tool: http://nixos.org/patchelf.html [2] Of course, IANAL, but there you go.

On Wed, Jan 1, 2014 at 9:03 PM, Aaron Friel wrote:

...

Because I think it’s going to be an organizational issue and a duplication of effort if GHC is built one way but the future direction of LLVM is another.

Imagine if GCC started developing a new engine and it didn’t work with one of the biggest, most regular consumers of GCC. Say, the Linux kernel, or itself. At first, the situation is optimistic - if this engine doesn’t work for the project that has the smartest, brightest GCC hackers potentially looking at it, then it should fix itself soon enough. Suppose the situation lingers though, and continues for months without fix. The new GCC backend starts to become the default, and the community around GCC advocates for end-users to use it to optimize code for their projects and it even becomes the default for some platforms, such as ARM.

What I’ve described is analogous to the GHC situation - and the result is that GHC isn’t self-hosting on some platforms and the inertia that used to be behind the LLVM backend seems to have stagnated. Whereas LLVM used to be the “new hotness”, I’ve noticed that issues like Trac #7787 no longer have a lot of eyes on them and externally it seems like GHC has accepted a bifurcated approach for development.

I dramatize the situation above, but there’s some truth to it. The LLVM backend needs some care and attention and if the majority of GHC devs can’t build GHC with LLVM, then that means the smartest, brightest GHC hackers won’t have their attention turned toward fixing those problems. If a patch to GHC-HEAD broke compilation for every backend, it would be fixed in short order. If a new version of GCC did not work with GHC, I can imagine it would be only hours before the first patches came in resolving the issue. On OS X Mavericks, an incompatibility with GHC has led to a swift reaction and strong support for resolving platform issues. The attention to the LLVM backend is visibly smaller, but I don’t know enough about the people working on GHC to know if it is actually smaller.

The way I am trying to change this is by making it easier for people to start using GHC (by putting images on Docker.io) and, in the process, learning about GHC’s build process and trying to make things work for my own projects. The Docker image allows anyone with a Linux kernel to build and play with GHC HEAD. The information about building GHC yourself is difficult to approach and I found it hard to get started, and I want to improve that too, so I’m learning and asking questions.

From: Carter Schonwald Sent: Wednesday, January 1, 2014 5:54 PM

...

To: Aaron Friel Cc: ghc-devs@haskell.org

7.8 should have working dylib support on the llvm backend. (i believe some of the relevant patches are in head already, though Ben Gamari can opine on that)

why do you want ghc to be built with llvm? (i know i've tried myself in the past, and it should be doable with 7.8 using 7.8 soon too)

On Wed, Jan 1, 2014 at 5:38 PM, Aaron Friel wrote:

...

Replying to include the email list. You’re right, the llvm backend and the gmp licensing issues are orthogonal - or should be. The problem is I get build errors when trying to build GHC with LLVM and dynamic libraries.

The result is that I get a few different choices when producing a platform image for development, with some uncomfortable tradeoffs:

LLVM-built GHC, dynamic libs - doesn’t build. LLVM-built GHC, static libs - potential licensing oddities with me shipping a statically linked ghc binary that is now gpled. I am not a lawyer, but the situation makes me uncomfortable. GCC/ASM-built GHC, dynamic libs - this is the *standard* for most platforms shipping ghc binaries, but it means that one of the biggest and most critical users of the LLVM backend is neglecting it. It also bifurcates development resources for GHC. Optimization work is duplicated and already devs are getting into the uncomfortable position of suggesting to users that they should trust GHC to build your programs in a particular way, but not itself. GCC/ASM-built GHC, static libs - worst of all possible worlds.

Because of this, the libgmp and llvm-backend issues aren’t entirely orthogonal. Trac ticket #7885 is exactly the issue I get when trying to compile #1.

From: Carter Schonwald Sent: Monday, December 30, 2013 1:05 PM

...

...

To: Aaron Friel

Good question but you forgot to email the mailing list too :-)

Using llvm has nothing to do with Gmp. Use the native code gen (it's simper) and integer-simple.

That said, standard ghc dylinks to a system copy of Gmp anyways (I think ). Building ghc as a Dylib is orthogonal.

-Carter

On Dec 30, 2013, at 1:58 PM, Aaron Friel wrote:

Excellent research - I’m curious if this is the right thread to inquire about the status of trying to link GHC itself dynamically.

I’ve been attempting to do so with various LLVM versions (3.2, 3.3, 3.4) using snapshot builds of GHC (within the past week) from git, and I hit ticket #7885 [https://ghc.haskell.org/trac/ghc/ticket/7885] every time (even the exact same error message).

I’m interested in dynamically linking GHC with LLVM to avoid the entanglement with libgmp’s license.

If this is the wrong thread or if I should reply instead to the trac item, please let me know.

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

-- Regards,

Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/ _______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

-- Regards, Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/