Compiler Plugins For GHC: Week Two
I wasn’t quite as productive with my Summer Of Code project this week as I was last week. Let’s take a look at the big ticket items that were accomplished.
Phase Control System Implemented
I covered the design of in my last post, and most of my work over the week has been on implementing and refining that proposal. It’s not essentially done, with a remaining small but significant wrinkle. The new phase control system lets you write rules that make use of phases above and beyond the existing ontology of phase 0, 1 and 2. An example of such a rule is as follows:
That’s all very well, but the snag is that we actually have one of these phases for every compiler pass in GHC, so in order to ensure that we always fire the RULEs that may be set up we need to insert a full simplifier pass after almost every compiler pass - yikes! That’s a lot more simplification than we currently do and it can’t be good for compile times. I’m still thinking about how to resolve that one.
Compiler Pipeline Dynamically Constructed From Constraints This is the reason why we have one phase for every compiler pass: I’ve changed GHC so its entire core-to-core pipeline is now built up from the relative ordering of these phases and the phase tags I’ve attached to every pass. This is a prerequisite for allowing compiler plugin authors to insert their own core-to-core passes by specifying declaratively when they would like them to run.
Template Haskell Phase Integration So we have these phase pragmas, but how do plugin go about referring to phases in their actual code that talks to GHC? The answer is with the new support for phases in Template Haskell!
This code is using the new triple quote notation to get a Template Haskell Name for the compiler phase, which is converted to a GHC Name and finally given to GHC itself. Of course, the Template Haskell support allows a lot more than this, such as generating new phases and splicing them in to your code at compile time.
Conclusion The project is still coming steadily along. I’m starting this week with ancillary work on the Static Argument Transformation that isn’t directly related to the project, but then I hope to move on to the plugin annotations system that I called out last week as a looming and highly thorny issue: expect to see more on this topic soon!