On May 2, 2014, at 9:05 AM, Simon Marlow <marlowsd@gmail.com> wrote:
On 02/05/2014 00:46, Edward Kmett wrote: With the old custom linker we weren't able to get our custom MPFR linked in properly on all platforms for use in ghci.
On Macs we ran into some rather interesting problems. We could get it to work for actual executables, but ghci would segfault with stuff resolved to clearly wrong addresses. If I recall correctly it may have been some kind of MachO symbol type that wasn't being resolved properly by the custom GHC linker, perhaps? We chased after it off and on for a long time to no avail.
It was complicated a great deal by having to build their library and merge it directly into ours, because we couldn't use system MPFR, due to the issues with the GHC garbage collector hooks into GMP.
I think this is the whole problem. Normally you would use an external dynamically-linked MPFR, but because you need custom changes to it, you have a local version of it which needs to get linked with your Haskell code. Presumably you have to arrange to build it with -fPIC somehow?
If you have C/C++ code in the same library as your Haskell code, then dynamic linking will work better (at least on OS X, where they tend to change the linking semantics more often than ELF, and we have to duplicate that in our linker). This is a good point, but rather than try to fix that I would instead look at whether you could build your custom MPFR as a shared library, which should work fine regardless of whether GHCi is dynamically linked or not. The only issue should be passing the right -rpath so that executables can find your custom MPFR, which is something we already do for dynamic linking.
Perhaps. We actually tried that originally, but had issues about where and how to get cabal to place it. We'd need it to go somewhere installed rather than the local build dir lest it not be there when we go to use the lib, but IIRC, cabal couldn't/wouldn't tell me where it was putting the final installed version and then there is the issue of the local in place runs vs. post cabal install runs and referencing that dir from subsequent projects. If we want a transparent ’cabal install rounded` It isn't clear to me how to get there down this path, but it is entirely possible I just missed something obvious. MPFR/rounded was just an off the cuff example. Another place where the dynamic linker really helps is external c++ libraries which should now actually get all of their initializers called in the right order when launched from ghci.
Cheers, Simon
Switching to the system dynamic linker fo ghci seems to have resolved all of that effortlessly.
Dan Peebles has been talking to the MPFR folks to see if we can get them to expose enough information about the 'hidden' allocations they use that we can make them visible to GHC or have them do what our local fix does and avoid using the MPFR allocator for their hidden constant cache.
If they do that then we can actually link to the library like normal rather than link it in directly, but it isn't clear to me what would happen even with those hooks if we rolled back to something like the old custom linker.
-Edward
On Thu, May 1, 2014 at 5:30 PM, Simon Marlow <marlowsd@gmail.com <mailto:marlowsd@gmail.com>> wrote:
On 01/05/14 15:27, Edward Kmett wrote:
Figured I'd make one case for dynamic linking:
https://github.com/ekmett/__rounded <https://github.com/ekmett/rounded>
Dynamic linking is finally enabling us to build a version of MPFR bindings for Haskell for scientific/high precision computing with 7.8. I would really hate to lose it after all of these years trying to get it work, as I have a rather large edifice being built atop that platform. We tried and failed due to limitations of the old linker for almost 3 years.
I understand the issues with MPFR. But how is dynamic linking helping?
That said, -dynamic-too seems to cause me all sorts of problems elsewhere. ^C'ing out of a build and restarting it will often make a .o but lose the .dyn_o, leading to GHC + cabal getting confused and refusing to build until I clean. This hits me several times a day.
We should fix this (or at least make it a lot less likely). Is there a ticket?
Cheers, Simon
-Edward
On Thu, May 1, 2014 at 3:29 AM, Simon Peyton Jones <simonpj@microsoft.com <mailto:simonpj@microsoft.com> <mailto:simonpj@microsoft.com <mailto:simonpj@microsoft.com>>__> wrote:
| Dynamic linking has been a huge headache in GHC, and it's not clear that | it's an overall improvement compared with the static linker. Now that | 7.8 is out of the way, it's time to have a conversation about whether we | want to do dynamic linking again for 7.10, or revert to static linking.
I echo this. Dynamic linking has had many un-anticipated costs and it is still very far from sorted out. It originally felt like a Fantastic Idea to give up our own linker and adopt the system linker, but it now feels to me like a black hole, endlessly sucking effort and increasing complexity.
My viewpoint is highly un-informed about details; I just watch the traffic going by. And of course it does have benefits that doubtless generate less traffic.
Simon
| | | | > | > On Tue, Apr 29, 2014 at 6:13 PM, Simon Peyton Jones | > <simonpj@microsoft.com <mailto:simonpj@microsoft.com> <mailto:simonpj@microsoft.com <mailto:simonpj@microsoft.com>> <mailto:simonpj@microsoft.com <mailto:simonpj@microsoft.com> <mailto:simonpj@microsoft.com <mailto:simonpj@microsoft.com>>__>> wrote: | > | > As Austin has told us, there's a draft of the *GHC Status Report | for | > the HCAR*, here:____ | > | > https://ghc.haskell.org/trac/__ghc/wiki/Status/May14____ <https://ghc.haskell.org/trac/ghc/wiki/Status/May14____> | > | > Have we missed out something you have been working hard on? Do | > take a moment to add a bullet in an appropriate place (it's a | > wiki). I'd like to be sure that we are giving credit to all the | > appropriate people, so please help us fix that too. GHC is a team | > effort.____ | > | > Deadline is 1 May I think.____ | > | > Thanks____ | > | > Simon____ | > | > __ __ | > | > | > _________________________________________________ | > ghc-devs mailing list | > ghc-devs@haskell.org <mailto:ghc-devs@haskell.org> <mailto:ghc-devs@haskell.org <mailto:ghc-devs@haskell.org>> <mailto:ghc-devs@haskell.org <mailto:ghc-devs@haskell.org> <mailto:ghc-devs@haskell.org <mailto:ghc-devs@haskell.org>>>
| > http://www.haskell.org/__mailman/listinfo/ghc-devs <http://www.haskell.org/mailman/listinfo/ghc-devs> | > | > | > | > | > _________________________________________________ | > ghc-devs mailing list | > ghc-devs@haskell.org <mailto:ghc-devs@haskell.org> <mailto:ghc-devs@haskell.org <mailto:ghc-devs@haskell.org>> | > http://www.haskell.org/__mailman/listinfo/ghc-devs <http://www.haskell.org/mailman/listinfo/ghc-devs> | > _________________________________________________ ghc-devs mailing list ghc-devs@haskell.org <mailto:ghc-devs@haskell.org> <mailto:ghc-devs@haskell.org <mailto:ghc-devs@haskell.org>> http://www.haskell.org/__mailman/listinfo/ghc-devs <http://www.haskell.org/mailman/listinfo/ghc-devs>