Of Lisp Macros and Washing Machines
loper-os.org
loper-os.org
1. Macros allow you to easily improve your syntax, but offer less guidance for improving your semantics. To use a mathematical analogy, it's nice to have a better notation, but what you really want are better definitions and theorems. Most macros are thin syntactic wrappers; few offer profound semantic insights. (Although there are some lovely examples of the latter in PG's excellent On Lisp: http://www.paulgraham.com/onlisp.html .)
2. Macros almost never call macroexpand recursively, except inside of code bodies. This means that macros like defclass usually cannot be extended in an incremental or modular fashion. Instead, you need to declare a def-your-class macro that can't usefully compose with any other def-her-class macro that you might encounter.
3. Weaker mechanisms than macros can accomplish nearly as much, with fewer issues similar to (2). For example, you can re-implement many Lisp macros using Ruby's block syntax, metaobject protocol, and ability to call class methods from inside a class body. And you can usually compose the resulting DSLs freely, even when they're analogous to defclass extensions—witness all the ActiveRecord add-ons.
So I like macros, but I no longer think that they're the ultimate high-level abstraction. They're mostly a nice way to incrementally improve your notation, except in the hands of programming language designers who are prepared to pull out all the stops.
Vladimir Sedach: The Haskell gang is primarily interested in advancing applied type theory.
If we're talking about high-level abstractions, this is probably not the best way to describe Haskell. (In fact, Haskell's type system is notoriously ad hoc, compared to languages like ML.)
Originally, Haskell started out a project to explore lazy evaluation and programming without side effects. But in recent years, the Haskell community has been investigating ways to build better domain-specific languages using ideas from mathematics: combinators, monads, co-monads, arrows, derivatives of types, and so on.
So far, many of these Haskell ideas have a steep learning curve, but the results are nonetheless impressive. If I were trying to design a "hundred-year language", I would definitely pay close attention to the Haskell community—they just bubble with fascinating ideas. And mathematical ideas tend to endure.
This IMO is the biggest problem with macro usage today and needs to be overcome. I've started a campaign against the "macros are syntactic sugar, and I have a sweet tooth!" camp (http://carcaddar.blogspot.com/2010/08/input-needed.html). I recommend reading this excellent parody blog post (written at the height of PG-inspired Lisp mania) for background: http://classic-web.archive.org/web/20070706135848/brucio.blo...
> Macros almost never call macroexpand recursively, except inside of code bodies.
This is by design. The expansion of a macro needs to be opaque - otherwise where is the abstraction? When you macroexpand, you're opening the black box. You can only rely on macroexpanding macros that you have some control over.
> This means that macros like defclass usually cannot be extended in an incremental or modular fashion.
This is because it's the same problem as extending an arbitrary grammar. You can't compose two grammars and expect the result to be non-ambiguous.
> Instead, you need to declare a def-your-class macro that can't usefully compose with any other def-her-class macro that you might encounter.
This is completely, totally wrong. Things like def-your-class is the reason I started my education campaign. You should never write a def-your-class macro. Macros are the wrong way to solve this problem, which is why CLOS and the CLOS metaobject protocol include all the facilities you could want for extending defclass in a composable way (Art of the Metaobject Protocol is a comprehensive, but confusing, reference).
I think we're looking at this from exactly opposite directions. I'm not talking about code-walking pre-existng macros, but rather providing hooks for other people to extend yours:
;; Completely hypothetical package management-system.
(define-module foo
(require 'bar)
(file "foo.ss")
(file "foo-tests.ss" :in-mode 'test))
(define-module foo
(require 'bar)
;; Doesn't work:
(tested-file "foo"))
Note that the actual advisability of this approach varies greatly between Lisp and Scheme dialects.
Contrast the following ActiveRecord example:
class Person < ActiveRecord::Base
belongs_to :organization
state_machine :initial => :starting do
state :starting
state :online
state :stopping
state :offline
end
end
Of course, macros are a very useful tool. But they have limitations, and other languages also provide interesting mechanisms for "writing programs that write programs", with different tradeoffs.
You should never write a def-your-class macro.
Well, even if the CLOS metaobject protocol does solve this problem, it doesn't fix defsystem, or any of the other thousand non-extensible defblah forms in the Lisp community.
I agree that you have some valid points but your overall argument holds little water. In good Lisps you have access to both runtime and compile time abstractions. Used wisely they will always trump a system that can only lean on runtime abstractions.
I'll also note that in my experience runtime metaprogramming is horribly painful. And many, many, many others have pointed out this fact. There are numerous cases where structural transformation is far simpler to reason about.
Overall, macros are a very useful tool. But there's a school of thought that sees them as the ultimate, universal abstraction. I'm no longer convinced by that argument.
However, I do agree with your remark, elsewhere in this discussion, that The Reasoned Schemer is an extraordinary fine example of what macros can be used for. It's one of my all-time favorite programming books.
If tested-file expanded to just file, it means that the only thing it can do is have side-effects at compile-time. Those side-effects won't be preserved in compiled files. So tested-file has to expand to something that includes file and the extra code, and define-module has to specify what that expansion can look like and how it will be handled. The design of this protocol can't be automated, because code-walking the expansion of tested-file is potentially Turing-complete.
So now you have two problems:
1. every define-foo macro needs a protocol to specify how it can be customized 2. this only works at compile-time
The need for this kind of customization is where the meta-object protocol arose from. Read up on Gregor Kiczales' work on open implementations (http://www2.parc.com/csl/groups/sda/projects/oi/ieee-softwar...). Putting in the required hooks into the interpreter (object system) using a restricted protocol provides a common system that all customizations can share (no need to design your own define-foo protocol) and is available at all stages (compile and run-time).
In what sense is Haskell's type system any more ad-hoc than ML's?
They're almost equivalent, except for Haskell's type-class extension, which is used for ad-hoc polymorphism. Support for ad-hoc polymorphism does not make the type system ad-hoc. It means you can use (\*) for multiplication between any type of number, which is very useful.
Many GHC extensions to the type systems are also based on sound theory (Rank 2/N types, Type families, GADTs) and not ad-hoc. ML lacks these too.
Haskell is one of the forefronts of applied type theory, AFAIK, ML seems to stagnate in this area.
According to Mark P Jones, the author of "Typing Haskell in Haskell":
Haskell benefits from a sophisticated type system, but implementors, programmers, and researchers suffer because it has no formal description. http://web.cecs.pdx.edu/~mpj/thih/
Haskell's type system was designed to make programming delightful. I vastly prefer it to some of the contortions of ML's type system, such as the separate + and +. operators. But by functional programming standards, I think that any type system without an official description can be fairly described as ad hoc.
Before muddying the water for future generations of potential Lisp programmers in this forum please go and master macros first instead of spewing unfounded bullshit about their limitations.
1. Macros were never meant to 'offer guidance for improving your semantics' however they are the mechanism that makes it possible to create your own new semantics (which is just not possible in any other family of languages). The semantics that you want to create are often specific to the domain you may be working in. Learn from those who have mastered the domain and created their own semantics and then come up with your own.
2. It's clear that you've been drinking the OO koolaid for a while. Let it go and try really learning macros. Leave the baggage at the door, it's just not required where you can go with macros.
3. Ruby is a badly diluted hack that borrowed some ideas from Lisp and other languages. No you cannot accomplish in Ruby what is possible with Lisp macros. For a start all the Ruby mechanisms that you suggest execute at runtime. Ruby does not have direct access to it's AST the way that Lisp does. Lisp and Ruby are not even in the same class of languages. Contrast Ruby with Java, and then yes it looks great.
If your talking about the "hundred-year language" then just forget Haskell. Lisp is the future, go and learn Qi. Qi has a turing complete type system, this is already light-years beyond what Haskell's type system will ever have.
This book is written by three true masters of macros and functional programming Oleg Kiselyov, William Byrd, and Daniel P. Friedman.
I know great C code to suggest, but draw a blank when it comes to Lisp macros done right.
The only thing I can gather is this guy is about to become a billionaire, because he has found a way to rid the world of spreadsheets and CRUD apps... apparently with steel mills.
What kind of evidence would you accept?
And if you won't accept it from R. Gabriel, P. Graham, P. Greenspun, and other "household" names, why should the rest of us bother sweating to assemble a watertight case for your wastebasket's eyes?
Code. Every author takes on himself the onerous to justify his claims. I am familiar with the works of the authors you mention; they present a good argument that lisp is a fine language for fine works. It is an entirely different argument to assert that, merely through the use of lisp, you will revolutionize the world.
That, as I read it, is the claim made by the original author. It is that claim which has gone unsubstantiated, for years. The rest of you--incidentally, no need to use exclusionary language: we are all friend here--need do nothing. Mr. Datskovskiy has some code to deliver. A bootloader in two years is no accomplishment, though Mr. Datskovskiy has excelled in punditry in this period.
If you actually read the blog (unbearable ordeal, I understand) you would know that Lisp (in available incarnations on available hardware) is not enough:
http://www.loper-os.org/?p=284
Let's see how much of an OS and hardware architecture you can design in several years - without using a single line of code written by another person. Starting with the NAND gate - because that is where you will have to start if you truly intend to avoid recapitulating the mistakes of the past three decades.
There was no sequel to the bootloader - I have been occupied with reverse-engineering the bitstream format of a major brand of FPGA (no naming names) in order to dispense with the proprietary x86 toolchain (and eventually with the abomination called Verilog.)
Publishing incremental progress in bite-sized chunks is vastly over-rated. It wastes everyone's time and contributes to the proliferation of hideously ragtag systems reminiscent of "Junkyard Wars" (aka much of the software industry.)
Punditry, on the other hand, is a relaxing and inexpensive hobby. The hosting bill for the past year of the blog was around $30.
Most of my time is spent working on unrelated efforts which I am not at liberty to make public. Currently, they leave precious-little time and energy for my long-term (think lifetime) project.
Feel free, of course, to imagine that I am an idler sitting in an armchair drinking the days away, if it makes you happy.
I sincerely hope you succeed. Verilog needs to be killed and replaced with something saner.
Then I read an article like this and realize I've been explaining it all wrong.
This is a great read and a real eye-opener.
In OOP languages such as Smalltalk, there is much less need for macros.
The author's point is that programming languages are great when you're cutting with the grain, when you're working in the domain the author optimized the language around. But when you leave the domain, macros are an important tool.
Don't get me wrong, I think that lazy evaluation is an incredibly valuable tool for separating what from how. But I wonder how it obviates the need for macros any more so than a well-defined OO system or any other non-trivial language feature.
However, you need monads to express other patterns of control flow (exceptions and dynamic scoping, for example) that are possible with macros. The problem is that you have to do everything through bind. Macros don't have that restriction.
But the one thing that makes macros irreplaceable is the fact that they can operate at compile-time. This is what I meant to emphasize when I wrote about automating programming.
> But the one thing that makes macros irreplaceable is the fact that they can operate at compile-time.
"Sufficiently smart compiler" is the standard strawman argument. It's not a sufficient response, but it is actually pretty astounding what can be handled at compile time.
Smalltalk has always done iteration this way with blocks (closures), and because of the syntax it works and looks nice.
Of course this needs the "sufficiently smart compiler" to be efficient. But there's no way to program domain-specific knowledge into the compiler like you can with macros. For example Gerald Sussman has done a lot of work in optimizing numerical code symbolically (http://repository.readscheme.org/ftp/papers/berlinsurati-pep...), I think _On Lisp_ has a chapter about doing partial evaluation on Bezier curves, and there's some regex-specific tricks in CL-PPCRE that I don't think can be matched by generic partial evaluation.
I loved the flexibility they allowed when I first used them, but these days they often feel to me like going down a level in that more needs to be expressed. Yeah, if your tools can't efficiently handle your current case, that escape hatch is incredibly valuable.
You can. It just won't be as syntactically elegant. Use expression trees and code emission the language exposes. The optimization happens at run-time, rather than compile time. But you'll actually, in most cases, get more optimization opportunity.
Monads are just a generalization of a commonly recurring pattern.
All imperative languages can be said to have "monads" in the sense of having a "semicolon operator" (in some languages, that's just newline) that is the monadic bind. You're programming in one "ambient monad" whose power is fixed by the language itself.
Haskell has an overridable semicolon, and that gives you a lot of DSL power. Instead of one "ambient monad", you have lots of user-defined monads useful for different purposes.
Besides overridable semicolons, monads are just a plain useful abstraction of data structures.
Monads are incredibly useful in a huge variety of programming tasks (non-determinism, effects, exceptions, parsing, and much more). They make Haskell code utilizing them far more broadly useful than code I've seen in any other language.