Does it sound a good idea to implement "backend plugins"?

Shao, Cheng cheng.shao at tweag.io
Fri Oct 5 00:02:23 UTC 2018


> A long time ago, I’ve tried to inject plugin logic to allows some control over the driver pipeline (phase ordering) and hooking various code gen related functions.
>
> See https://phabricator.haskell.org/D535

Cool! I haven't thoroughly read the history of that diff, but allowing
manipulation of a Hooks via a Plugin seems overkill in this case, and
even if one can do so, it still doesn't lead to the backend IR types;
one would need to use runPhaseHook and modify the behavior after a
CgGuts is generated, which unfortunately leads to quite some
boilerplate code.

> At that time I ran into issues that might simply not exist with plugins anymore today, but I haven’t looked.

Interesting. I'll make sure to consult you in case I'm bitten by some
hidden issues when I actually implement it :)

> The whole design wasn’t quite right and injects everything into the dynflags.  Also ghc wanted to be able to compile the plugin on the fly, but I needed the plugin to be loaded very early during the startup phase to exert enough control of the rest of the pipeline through the plugin.

Well, in the case of backend plugins, it isn't supposed to be a home
plugin to be compiled and used on the fly. A typical use case would be
compiling/installing the plugin to a standalone pkgdb, then used to
compile other packages.

>
> On 5 Oct 2018, at 1:52 AM, Shao, Cheng <cheng.shao at tweag.io> wrote:
>
> Adding "pluggable backends" to spin up new targets seems to require quite a bit of additional infrastructure for initialising a library directory and package database. But there are probably more specific use cases that need inspecting/modifying STG or Cmm where plugins would already be useful in practice.
>
>
> I think setting up a new global libdir/pkgdb is beyond the scope of
> backend plugins. The user shall implement his/her own boot script to
> configure for the new architecture, generate relevant headers, run
> Cabal's Setup program to launch GHC with the plugin loaded.
>
> Hooks (or rather their locations in the pipeline) are rather ad hoc by nature, but for Asterius a hook that takes Cmm and takes over from there seems like a reasonable approach given the current state of things. I think the Cmm hook you implemented (or something similar) would be perfectly acceptable to use for now.
>
>
> For the use case of asterius itself, indeed Hooks already fit the use
> case for now. But since we seek to upstream our newly added features
> in our ghc fork back to ghc hq, we should upstream those changes early
> and make them more principled. Compared to Hooks, I prefer to move to
> Plugins entirely since:
>
> * Plugins are more composable, you can load multiple plugins in one
> ghc invocation. Hooks are not.
> * If I implement the same mechanisms in Plugins, this can be
> beneficial to other projects. Currently, in asterius, everything works
> via a pile of hacks upon hacks in ghc-toolkit, and it's not good for
> reuse.
> * The newly added backend plugins shouldn't have visible
> correctness/performance impact if they're not used, and it's just a
> few local modifications in the ghc codebase.
>
> On Thu, Oct 4, 2018 at 3:56 PM Shao, Cheng <cheng.shao at tweag.io> wrote:
>
>
> Hi all,
>
>
> I'm thinking of adding "backend plugins" in the current Plugins
>
> mechanism which allows one to inspect/modify the IRs post simplifier
>
> pass (STG/Cmm), similar to the recently added source plugins for HsSyn
>
> IRs. This can be useful for anyone creating a custom GHC backend to
>
> target an experimental platform (e.g. the Asterius compiler which
>
> targets WebAssembly), and previously in order to retrieve those IRs
>
> from the regular pipeline, we need to use Hooks which is somewhat
>
> hacky.
>
>
> Does this sound a good idea to you? If so, I can open a trac ticket
>
> and a Phab diff for this feature.
>
>
> Best,
>
> Shao Cheng
>
> _______________________________________________
>
> ghc-devs mailing list
>
> ghc-devs at haskell.org
>
> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>
> _______________________________________________
> ghc-devs mailing list
> ghc-devs at haskell.org
> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs


More information about the ghc-devs mailing list