Why GADTs matter for performance (2015)

https://blog.janestreet.com/why-gadts-matter-for-performance/

83•hyperbrainer•1y ago

Comments

rbjorklin•1y ago

Does anyone have some hard numbers on the expected performance uplift when using GADTs? Couldn't see any mentioned in the article.

ackfoobar•1y ago

The example here is basically an 8-fold memory saving going from `long[]` from `byte[]` - while still retaining polymorphism (whereas in Java the two are unrelated types).

Hard to say exactly how much performance one would get, as that depends on access patterns.

misja111•1y ago

The reason that a byte array is in reality layed out as a (mostly empty) long array in Java, is actually for performance. Computers tend to have their memory aligned at 8 byte intervals and accessing such an address is faster than accessing an address that's at an offset of an 8 byte interval.

Of course it depends on your use case, in some cases a compact byte array performs better anyway, for instance because now you're able to fit it in your CPU cache.

john-h-k•1y ago

But you can load any byte by loading 8 bytes and shift (v cheap)

ackfoobar•1y ago

> a byte array is in reality layed out as a (mostly empty) long array in Java

Are you saying each byte takes up a word? That is the case in the `char array` in OCaml, but not Java's `byte[]`. AFAIK The size of a byte array is rounded up to words. Byte arrays of length 1-8 all have the same size in a 64-bit machine, then length 7-16 take up one more word.

https://shipilev.net/jvm/objects-inside-out/

cosmic_quanta•1y ago

Interesting, thanks for posting.

I share the author's frustration with the lack of non-compiler-related examples of GADT uses. It seems like such a powerful idea, but I haven't been able to get a feel for when to reach for GADTs in Haskell

wyager•1y ago

I often find them handy for locking down admissible states at compile time. Maybe ~10 years ago in a processor design class, I wrote some CPUs in Haskell/Clash for FPGA usage. A nice thing I could do was write a single top-level instruction set, but then lock down the instructions based on what stages of the processor they could exist at.

For example, something like (not an actual example from my code, just conceptually - may be misremembering details):

  data Instruction stages where
   MovLit :: Word64 -> Register -> Instruction '[Fetch, Decode, Execute, Writeback]
   -- MovReg instruction gets rewritten to MovLit in Execute stage
   MovReg :: Register -> Register -> Instruction '[Fetch, Decode, Execute]
   ...

And then my CPU's writeback handler block could be something like:

  writeback :: (Writeback `member` stages) => Instruction stages -> WritebackState -> WritebackState
  writeback (MovLit v reg) = ...
  -- Compiler knows (MovReg _ _) is not required here

So you can use the type parameters to impose constraints on the allowed values, and the compiler is smart enough to use this data during exhaustiveness checks (cf "GADTs Meet Their Match")

anyfoo•1y ago

Wow, someone else who (used to be) using Clash. I still use it for everything I can in my (hobby) FPGA projects. I'm not sure I've used GADTs, but I've certainly made use of other more "advanced" parts of the type system, like type families.

What you're doing here is pretty cool, I think I will start doing so, too. I have a number of places where I use "undefined" instead. (The "undefined" from the Clash Prelude, which translates into a "don't care" signal state.)

hyperbrainer•1y ago

goldchainposse•1y ago

I know Jane Street love OCaml, but you have to wonder how much it's cost them in velocity and maintenance. This is a quant firm blogging about a programming language they're the most famous user of.

pjmlp•1y ago

It is thanks to the companies like Jane Street that believe there is something else beyond C, that we can have nice toys.

Remember if OCaml wasn't a mature programming language, maybe Rust would not have happened in first place.

kryptiskt•1y ago

Why do you assume it's a drag for them and not a competitive advantage? I don't know if it's such a terrible thing to use a slightly out of mainstream language, when the standard in the business is to accumulate tens of millions of lines of C++.

ackfoobar•1y ago

Agreed, indeed I believe they have mentioned that OCaml gets them to ship quicker because they are more confident with the correctness of changes.

But being outside of the mainstream may mean you need to occasionally debug more esoteric stuff: https://gallium.inria.fr/blog/intel-skylake-bug/ I'm sure Jane Street can afford doing that, but I'm not so sure if a small team can.

gjadi•1y ago

That was an interesting read, thanks. However I fail to see how it's an issue specific to ocaml. It was a bug in the Skylake processor triggered by a special pattern of instructions produced by gcc. Ocaml built with clang was ok because it doesn't used the same pattern. Did I miss something?

cryptonector•1y ago

What's not clear from reading TFA is whether the compiler monomorphizes TFA's `Compact_array` for the two special cases of it (array of bytes vs. array of anything else), but I'm assuming so. Perhaps if I was familiar with OCaml the answer would be blindingly obvious. What's happening here is that w/ GADTs you can have a _singular_ abstraction with multiple distinct implementations for specific types and others for generic types, and you don't have to think about it too much, except you have to remember to use these type hints in the interface definitions to get the compiler to do what you want.

> Yaron Minsky joined Jane Street back in 2002, and claims the dubious honor of having convinced the firm to start using OCaml.

That's pretty cool. And I guess Stephen Dolan ended up there due to his work on OCaml, which is pretty cool too. (I'd like to meet Stephen some day.)

fjwufjfa•1y ago

It's easier to reason in FP plus the python paradox [1] [2].

[1]: https://www.paulgraham.com/pypar.html

[2]: https://blog.janestreet.com/why-ocaml/

codr7•1y ago

For certain classes of programs, yes. I have a hunch finance is a pretty good fit.

AdieuToLogic•1y ago

I agree with your point about reasoning when employing Functional Programming (FP).

However, I very much disagree with Graham's 2004 assertion[0]:

  It's a lot of work to learn a new programming language. And 
  people don't learn Python because it will get them a job; 
  they learn it because they genuinely like to program and 
  aren't satisfied with the languages they already know.

It does not require "a lot of work to learn a new programming language" once a person has fluency with at least one. Actually, the difficulty of learning a new programming language is inversely proportional to how many programming languages the person has already learned. Especially if a new programming language is in the same paradigm category as those already known (Procedural, OOP, FP, etc.).

I was a professional software engineer in 2004, when the Graham post was written. To say, "people don't learn Python because it will get them a job ..." was bullshit then just as it is now. The remainder of the quoted sentence is unfounded extrapolation and has the value of same.

0 - https://www.paulgraham.com/pypar.html

DSpark: Speculative decoding accelerates LLM inference [pdf]

Fintech Engineering Handbook

OpenRA

If you can't hold it, you don't own it

Beer CSS – Build material design in record time

Previewing GPT‑5.6 Sol: a next-generation model

Long Wave radio era set to end with switch-off

Nox Metals (YC S25) Is Hiring SWE

Linux on Older Hardware: The Complete Revival Guide

The US Army Issued Ocarinas to Soldiers in World War II

WordStar: A Writer's Word Processor (1996)

Why does kinetic energy increase quadratically, not linearly, with speed? (2011)

Cultures of Making and Relating

Faster KNN search in Manticore: 2-pass HNSW, batched distances, and AVX-512

Task Failed Successfully: Saturating NIC and Disk Bandwidth

U.S. allows Anthropic to release Mythos AI to ‘trusted’ US organizations

MicroVMs: Run isolated sandboxes with full lifecycle control

AI in mathematics is forcing big questions

OpenTTD 16.0-Beta1

Jest/Vitest interactive course (runs in the browser)

Anatomy of a Failed (Nation-State?) Attack

Fusion Programming Language

Hellishly Slow Level 13 Deflate Compression

U.S. government will decide who gets to use GPT-5.6

International investment and local rules push prices up faster than supply

IBM MCGA Gate Array Reverse Engineering

The gap between open weights LLMs and closed source LLMs

Ultrasound imaging of the brain

Show HN: Hacker News on a train station-style flip board

Om

DSpark: Speculative decoding accelerates LLM inference [pdf]

Fintech Engineering Handbook

OpenRA

If you can't hold it, you don't own it

Beer CSS – Build material design in record time

Previewing GPT‑5.6 Sol: a next-generation model

Long Wave radio era set to end with switch-off

Nox Metals (YC S25) Is Hiring SWE

Linux on Older Hardware: The Complete Revival Guide

The US Army Issued Ocarinas to Soldiers in World War II

WordStar: A Writer's Word Processor (1996)

Why does kinetic energy increase quadratically, not linearly, with speed? (2011)

Cultures of Making and Relating

Faster KNN search in Manticore: 2-pass HNSW, batched distances, and AVX-512

Task Failed Successfully: Saturating NIC and Disk Bandwidth

U.S. allows Anthropic to release Mythos AI to ‘trusted’ US organizations

MicroVMs: Run isolated sandboxes with full lifecycle control

AI in mathematics is forcing big questions

OpenTTD 16.0-Beta1

Jest/Vitest interactive course (runs in the browser)

Anatomy of a Failed (Nation-State?) Attack

Fusion Programming Language

Hellishly Slow Level 13 Deflate Compression

U.S. government will decide who gets to use GPT-5.6

International investment and local rules push prices up faster than supply

IBM MCGA Gate Array Reverse Engineering

The gap between open weights LLMs and closed source LLMs

Ultrasound imaging of the brain

Show HN: Hacker News on a train station-style flip board

Om

Why GADTs matter for performance (2015)

Comments