Plugins: Accessing unexported bindings

Mon Dec 7 15:44:24 UTC 2015

The problem, as I recall, is that GHC does an initial bit of dead-code
elimination in the desugarer, before the plugins have a chance to run.
(I believe this is part of simpleOptPgm, but may be mistaken)

I'm not sure why this is done in the desugarer, it seems to be out of
place there. 

On Mon, Dec 7, 2015, at 05:14, Simon Peyton Jones wrote:
> Plugins get to edit the entire core-to-core pipeline!  There is no magic.
>  At least I don’t think so
> 
> file:///Z:/tmp/users_guide/compiler-plugins.html
> 
> S
> 
> From: Levent Erkok [mailto:erkokl at gmail.com]
> Sent: 07 December 2015 13:11
> To: Simon Peyton Jones <simonpj at microsoft.com>
> Cc: Eric Seidel <eric at seidel.io>; omeragacan at gmail.com; ezyang at mit.edu;
> ghc-devs at haskell.org
> Subject: Re: Plugins: Accessing unexported bindings
> 
> That's a good point; keeping all annotated bindings alive seems to be an
> overkill..
> 
> Regarding implementing "pass at the start." I'm not sure if plugin
> authors have any freedom as to decide when their plugin actually runs. It
> seems GHC magically determines the order and runs them. Can you point me
> to some code/docs that tells me how to go "first" in that sense? (Or at
> least before the pass that drops dead code.)
> 
> On Dec 7, 2015, at 4:45 AM, Simon Peyton Jones
> <simonpj at microsoft.com<mailto:simonpj at microsoft.com>> wrote:
> Indeed. How about this: if there's an ANN on a binder (any ANN), then GHC
> should keep it alive.
> 
> Really?  It might be something like “don’t give warnings for this
> binding” or “don’t inline me” or something.   To say *any* annotation
> seems a bit brutal doesn’t it?    Mind you I don’t have a better idea.
> 
> One thought: your plugin could add a pass right at the start, which marks
> everything you want as keep-alive.
> 
> S
> 
> From: Levent Erkok [mailto:erkokl at gmail.com]
> Sent: 07 December 2015 12:42
> To: Simon Peyton Jones
> <simonpj at microsoft.com<mailto:simonpj at microsoft.com>>
> Cc: Eric Seidel <eric at seidel.io<mailto:eric at seidel.io>>;
> omeragacan at gmail.com<mailto:omeragacan at gmail.com>;
> ezyang at mit.edu<mailto:ezyang at mit.edu>;
> ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
> Subject: Re: Plugins: Accessing unexported bindings
> 
> Indeed. How about this: if there's an ANN on a binder (any ANN), then GHC
> should keep it alive.
> 
> Is that something one of the core-developers can implement? Happy to open
> a ticket if that helps.
> 
> On Dec 7, 2015, at 4:14 AM, Simon Peyton Jones
> <simonpj at microsoft.com<mailto:simonpj at microsoft.com>> wrote:
> If it's "dead" in this sense, it's already removed from ModGuts, no?
> 
> Yes, if it’s dead it’s gone.   That’s not too surprising, is it?
> 
> So you need a way to keep it alive. Maybe we need a pragma for that.   Or
> how would you like to signal it in the source code?
> 
> Simon
> 
> From: Levent Erkok [mailto:erkokl at gmail.com]
> Sent: 07 December 2015 12:05
> To: Simon Peyton Jones
> <simonpj at microsoft.com<mailto:simonpj at microsoft.com>>
> Cc: Eric Seidel <eric at seidel.io<mailto:eric at seidel.io>>;
> omeragacan at gmail.com<mailto:omeragacan at gmail.com>;
> ezyang at mit.edu<mailto:ezyang at mit.edu>;
> ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
> Subject: Re: Plugins: Accessing unexported bindings
> 
> Thanks Simon.. But I remain utterly confused. As a "plugin" author, how
> do I get my hands on the Id associated with a top-level binder? If it's
> "dead" in this sense, it's already removed from ModGuts, no?
> 
> That is, by the time GHC runs my plugin, the Id has already disappeared
> for me to mark it "Local Exported." Is that not correct?
> 
> On Dec 7, 2015, at 2:28 AM, Simon Peyton Jones
> <simonpj at microsoft.com<mailto:simonpj at microsoft.com>> wrote:
> In the mean time, I'm still looking for a solution that doesn't involve
> exporting such identifiers from modules. As Eric pointed out, that seems
> to be the only current work-around for the time being.
> 
> “Exported” in this context only means “keep alive”. It does not mean
> exported in the Haskell source code sense. I’ve just added this comment
> to Var.hs.
> 
> So I think it does just what you want.
> 
> Simon
> 
> data ExportFlag   -- See Note [ExportFlag on binders]
>   = NotExported   -- ^ Not exported: may be discarded as dead code.
>   | Exported      -- ^ Exported: kept alive
> 
> {- Note [ExportFlag on binders]
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> An ExportFlag of "Exported" on a top-level binder says "keep this
> binding alive; do not drop it as dead code".  This transititively
> keeps alive all the other top-level bindings that this binding refers
> to.  This property is persisted all the way down the pipeline, so that
> the binding will be compiled all the way to object code, and its
> symbols will appear in the linker symbol table.
> 
> However, note that this use of "exported" is quite different to the
> export list on a Haskell module.  Setting the ExportFlag on an Id does
> /not/ mean that if you import the module (in Haskell source code you
> will see this Id.  Of course, things that appear in the export list
> of the source Haskell module do indeed have their ExportFlag set.
> But many other things, such as dictionary functions, are kept alive
> by having their ExportFlag set, even though they are not exported
> in the source-code sense.
> 
> We should probably use a different term for ExportFlag, like
> KeepAlive.
> 
> From: ghc-devs [mailto:ghc-devs-bounces at haskell.org] On Behalf Of Levent
> Erkok
> Sent: 06 December 2015 20:32
> To: Eric Seidel <eric at seidel.io<mailto:eric at seidel.io>>;
> omeragacan at gmail.com<mailto:omeragacan at gmail.com>;
> ezyang at mit.edu<mailto:ezyang at mit.edu>
> Cc: ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
> Subject: Re: Plugins: Accessing unexported bindings
> 
> Omer, Eric, Ed: Thanks for the comments.
> 
> Omer: I think Eric's observation is at play here. We're talking about
> "dead-code," i.e., a binding that is neither exported, nor used by any
> binding inside the module. Those seem to be getting dropped by the time
> user-plugins are run. Unfortunately, this is precisely what one would do
> with "properties" embedded in code. They serve as documentation perhaps,
> but are otherwise not needed by any other binding nor it makes sense to
> export them.
> 
> Edward: Can you provide some more info into your solution? Sounds like a
> chicken-egg issue to me: As a plugin author, I need the bindings to
> access the Ids, and looks like I need the Ids to access the binders?
> 
> A simple solution would be to simply keep all top-level bindings around
> while the plugin are running, but that obviously can lead to unnecessary
> work if the code is truly dead. A compromise could be that the
> annotations can serve as entry points as well: I.e., if there's an
> annotation on a top-level binder, then it should *not* be considered
> dead-code at least until after all the plugins are run. That would
> definitely simplify life. Would that be an acceptable alternative?
> 
> In the mean time, I'm still looking for a solution that doesn't involve
> exporting such identifiers from modules. As Eric pointed out, that seems
> to be the only current work-around for the time being.
> 
> Thanks,
> 
> -Levent.
> 
> On Sun, Dec 6, 2015 at 11:08 AM, Eric Seidel
> <eric at seidel.io<mailto:eric at seidel.io>> wrote:
> GHC should only drop un-exported bindings from the ModGuts if they're
> also unused, ie *dead code*.
> 
> The only way I know to get around this is to use the bindings somewhere,
> or just export them.
> 
> On Sat, Dec 5, 2015, at 23:01, Levent Erkok wrote:
> > Hello,
> >
> > The mg_binds field of the ModGuts seem to only contain the bindings that
> > are exported from the module being compiled.
> >
> > I guess GHC must be running user-plugins after it drops the bindings that
> > are not exported, which makes perfect sense for most use cases. However,
> > I'm working on a plugin where the end-programmer embeds "properties" in
> > the
> > form of functions inside his/her code, which are not necessarily exported
> > from the module under consideration.
> >
> > Is there a way to access all top-level bindings in a module from a
> > plugin,
> > even if those bindings are not exported?
> >
> > Thanks,
> >
> > -Levent.
> > _______________________________________________
> > ghc-devs mailing list
> > ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
> > http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fmail.haskell.org%2fcgi-bin%2fmailman%2flistinfo%2fghc-devs&data=01%7c01%7csimonpj%40064d.mgd.microsoft.com%7cac4cbfe22e314080909908d2fe7c4ed8%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=1z6DcZxjIAKj0PcsLeALphRLWJ3i%2fxvyaPtq0qo6elY%3d>
> _______________________________________________
> ghc-devs mailing list
> ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fmail.haskell.org%2fcgi-bin%2fmailman%2flistinfo%2fghc-devs&data=01%7c01%7csimonpj%40064d.mgd.microsoft.com%7cac4cbfe22e314080909908d2fe7c4ed8%7c72f988bf86f141af91ab2d7cd011db47%7c1&sdata=1z6DcZxjIAKj0PcsLeALphRLWJ3i%2fxvyaPtq0qo6elY%3d>
> 