Haskell’s GHC compiler has notoriously slow compile times. In a bid to drive the change it seeks in the world, Tweag is opening a call for applications for an intern to work for three months solely on improving GHC’s performance. This post describes a number of approaches the intern might take. Most of these ideas are broadly applicable to any Haskell program, and so even those of you who do not hack regularly on GHC might find this useful.
Do you have the interest and skills to improve compile times? Please apply! (More details about the application are at the bottom of this post.)
Data structures. Does GHC use the most efficient datatypes internally? For example, lists are very common in the GHC codebase. Are they effective? Maybe small arrays would be better. Maybe
Data.Seqwould be better. Ideally, we would come up with concrete guidelines for when to use each datatype.
Fusion. Much data — in particular, lists — is meant to be fused away. Does this happen in practice? Fusion is implemented within GHC by using rewrite rules. However, GHC defines a good deal of its own low-level list-manipulation functions. Do these fuse correctly? The intern would look to find out.
Case-of-known-constructor. A good deal of functions within GHC return a
Maybe. This means that the functions invariably must allocate their result freshly. Is that really necessary? If the function is inlined into a context that immediately matches against the
Maybe, then we can avoid this allocation, via the case-of-known-constructor optimization (already present). In the case of recursive functions, the allocation incurred by all the
Justnodes may be significant. Another place where case-of-known-constructor may help is around tuples, which also allocate. One possible direction here is to replace
Maybe (..., ...)with
Maybe2 ... ..., where
data Maybe2 a b = Just2 a b | Nothing2. This change would reduce allocations. But is it worth it?
Laziness. Many data structures in GHC are lazy. But perhaps some important data structures would be better with (some) strict fields. We could look at which data structures are in the most common use and see whether they can be optimized by adding strictness annotations.
Unboxed types. Normal, boxed types are convenient, but unboxed types (including unboxed sums
(# ... | ... #)) can be faster. If there is a tight loop, the speed of unboxed types may be worth the annoyance of working with them.
Data representation. There is often a multitude of ways of representing data. For example, we could store the expression
f a b cas
App (App (App f a) b) c, as we do now. But maybe
Apps f [a,b,c]would be better, or
Apps f (fromList [a,b,c])(for some appropriate data structure’s
fromListfunction) would be even better. The intern could explore these possibilities.
Algorithms. There are a number of algorithms that do unnecessary work within GHC. For example, we might want to know whether there are any free type variables in an expression. A simple way to do this is to gather all the free variables of an expression (there is a function for that) and then iterate through the set of variables (there is a function for that) detecting type variables (there is a function for that). But this is madness: it allocates a ton of space in building the set, when a straightforward traversal of the expression would yield what we want, without allocating a whit. The intern would hunt down and kill such mad examples, possibly using abstractions like the
foldTyCoidea recently added. This is a way of doing lightweight traversals fast.
Heap profiling. GHC has features enabling heap profiling, meaning that we can see what structures appear in the heap at regular time intervals. Either using existing performance tests or writing new ones, the intern could generate heap profiles, looking for something unexpected — for example, that an entire half of the heap is just
Ticky-ticky profiling. The ticky-ticky profiler is a low-level instrumentation of the optimized code, and it allows you to see exactly what is being allocated and when. Looking at this output is enlightening: a recent profile shows, for example, 321,081 entries of the recursive helper function in the
lengthfunction for lists. (Because it’s the recursive helper function, this result means that a total of 321,081 list nodes were counted, not that
lengthwas called 321,081 times.) Yet I would hope that we almost never need to find the length of a linked list — doing so is often a sign that you have the wrong data structure. Investigating this would likely lead to optimization opportunities.
Streams. GHC has its own Stream module added by Simon Marlow 8 years ago (commit) which allows the STG-to-CMM codegen to run in constant space. This is great, but not great enough, since
Streamisn’t more widely used throughout the GHC codebase to replace lists, and the implemented operations of
Streamare quite limited. The intern could look into chances to use
Streamin more places and possibly port the
vector, given that it implements much richer operations and is fine-tuned for fusion and performance.
Community-sourced opportunities. The list here is surely incomplete. We can community-source other ideas to pursue.
Success criterion: I would expect that three months of directed effort should yield at least a 10% improvement in compile times. This expectation could be way off, admittedly.
The intern would also be required to share their knowledge:
They would write a blog post on their work, which would hopefully offer readers insights in performance debugging beyond GHC.
They would offer a knowledge sharing within Tweag, helping other to performance-tune their own code.
Though the focus would remain on optimizing GHC, the intern may be able to make progress at automating their work. Here are some ideas in this vein:
New optimizations. One idea above is to change
Maybe (ty1, ty2)to
Maybe2 ty1 ty2. However, (ignoring laziness) these types are isomorphic, and GHC should be able to transform one into the other. Why should we do this manually? (Not ignoring laziness: we may need strictness annotations to make them isomorphic, but that’s not a problem.) Maybe the intern can design and implement a new optimization pass.
New data collection opportunities. In performance-tuning the linear types branch, I really wanted to know what constructors were being allocated on the heap. This data was not easily available, but some poking around (and help from others) showed how it could be done easily. GHC was thus augmented to keep track of this information when requested. Other, similar measurement ideas may arise, as improvements to GHC that would benefit anyone doing profiling work.
Inspection testing. There are often certain optimization expectations embodied in a chunk of code. For example, we might expect that inlining a certain function will eliminate some of its tests. However, how can we be sure these optimizations really take place? Joachim Breitner has written an inspection testing tool that permits exactly this sort of unit test, as described in this paper. We could imagine integrating inspection testing into the GHC and library testsuites.
Do you have what it takes to optimize GHC and become a folk hero in the Haskell world? I hope so, and I look forward to working with you! Apply using our online application. Include a cover letter describing your Haskell experience, any experience you already have with working on GHC (though none is required), and any experience you already have with performance work (in any language). We will collect applications until Thursday, March 4, 2021 (anywhere on Earth) and then review all collected applications; we hope to be able to update applicants on their status within days. The internship can start any time after the offer is made, subject to mutual availability. Internships typically last 12 weeks, although there may be some flexibility here. If you have any questions, feel free to email me directly. Please do apply!