## Parallelism and Concurrency, Revisited

April 9, 2014

To my delight, I still get compliments on and criticisms of my post from three years ago (can it possibly be that long?) on parallelism and concurrency.  In that post I offered a “top down” argument to the effect that these are different abstractions with different goals: parallelism is about exploiting computational resources to maximize efficiency, concurrency is about non-deterministic composition of components in a system.  Parallelism never introduces bugs (the semantics is identical to the sequential execution), but concurrency could be said to be the mother lode of all bugs (the semantics of a component changes drastically, without careful provision, when composed concurrently with other components).  The two concepts just aren’t comparable, yet somehow the confusion between them persists.  (Not everyone agrees with me on this distinction, but neither have I seen a comparable analysis that shows them to be the same concept.  Most complaints seem to be about my use of the words “parallelism” and “concurrency” , which is an unavoidable problem, or about my temerity in trying to define two somewhat ill-defined concepts, a criticism that I’ll just have to accept.)

I’ve recently gotten an inkling of why it might be that many people equate the two concepts (or see no point in distinguishing them).  This post is an attempt to clear up what I perceive to be a common misunderstanding that seems to explain it.  It’s hard for me to say whether it really is all that common of a misunderstanding, but it’s the impression I’ve gotten, so forgive me if I’m over-stressing an obvious point.  In any case I’m going to try for a “bottom up” explanation that might make more sense to some people.

The issue is scheduling.

The naive view of parallelism is that it’s just talk for concurrency, because all you do when you’re programming in parallel is fork off some threads, and then do something with their results when they’re done.  I’ve previously argued that this is the wrong way to think about parallelism (it’s really about cost), but let’s just let that pass.  It’s unarguably true that a parallel computation does consist of a bunch of, well, parallel computations.  So, the argument goes, it’s nothing but concurrency, because concurrency is, supposedly, all about forking off some threads and waiting to see what they do, and then doing something with it.  I’ve argued that that’s not a good way to think about concurrency either, but we’ll let that pass too.  So, the story goes, concurrency and parallelism are synonymous, and bullshitters like me are just trying to confuse people and make trouble.

Being the troublemaker that I am, my response is, predictably, nojust no.  Sure, it’s kinda sorta right, as I’ve already acknowledged, but not really, and here’s why: scheduling as you learned about it in OS class (for example) is an altogether different thing than scheduling for parallelism.  And this is the heart of the matter, from a “bottom-up” perspective.

There are two aspects of OS-like scheduling that I think are relevant here.  First, it is non-deterministic, and second, it is competitive.  Non-deterministic, because you have little or no control over what runs when or for how long.  A beast like the Linux scheduler is controlled by a zillion “voodoo parameters” (a turn of phrase borrowed from my queueing theory colleague, Mor Harchol-Balter), and who the hell knows what is going to happen to your poor threads once they’re in its clutches.  Second, and more importantly, an OS-like scheduler is allocating resources competitively.  You’ve got your threads, I’ve got my threads, and we both want ours to get run as soon as possible.  We’ll even pay for the privilege (priorities) if necessary.  The scheduler, and the queueing theory behind it (he says optimistically) is designed to optimize resource usage on a competitive basis, taking account of quality of service guarantees purchased by the participants.  It does not matter whether there is one processor or one thousand processors, the schedule is unpredictable.  That’s what makes concurrent programming hard: you have to program against all possible schedules.  And that’s why you can’t prove much about the time or space complexity of your program when it’s implemented concurrently.

Parallel scheduling is a whole ‘nother ball of wax.  It is (usually, but not necessarily) deterministic, so that you can prove bounds on its efficiency (Brent-type theorems, as I discussed in my previous post and in PFPL).  And, more importantly, it is cooperative in the sense that all threads are working together for the same computation towards the same ends.  The threads are scheduled so as to get the job (there’s only one) done as quickly and as efficiently as possible.  Deterministic schedulers for parallelism are the most common, because they are the easiest to analyze with respect to their time and space bounds.  Greedy schedulers, which guarantee to maximize use of available processors, never leaving any idle when there is work to be done, form an important class for which the simple form of Brent’s Theorem is obvious.

Many deterministic greedy scheduling algorithms are known, of which I will mention p-DFS and p-BFS, which do p-at-a-time depth- and breadth-first search of the dependency graph, and various forms of work-stealing schedulers, pioneered by Charles Leiserson at MIT.  (Incidentally, if you don’t already know what p-DFS or p-BFS are, I’ll warn you that they are a little trickier than they sound.  In particular p-DFS uses a data structure that is sort of like a stack but is not a stack.)  These differ significantly in their time bounds (for example, work stealing usually involves expectation over a random variable, whereas the depth- and breadth traversals do not), and differ dramatically in their space complexity.  For example, p-BFS is absolutely dreadful in its space complexity.  For a full discussion of these issues in parallel scheduling, I recommend Dan Spoonhower’s PhD Dissertation.  (His semantic profiling diagrams are amazingly beautiful and informative!)

So here’s the thing: when you’re programming in parallel, you don’t just throw some threads at some non-deterministic competitive scheduler.  Rather, you generate an implicit dependency graph that a cooperative scheduler uses to maximize efficiency, end-to-end.  At the high level you do an asymptotic cost analysis without considering platform parameters such as the number of processors or the nature of the interconnect.  At the low level the implementation has to validate that cost analysis by using clever techniques to ensure that, once the platform parameters are known, maximum use is made of the computational resources to get your job done for you as fast as possible.  Not only are there no bugs introduced by the mere fact of being scheduled in parallel, but even better, you can prove a theorem that tells you how fast your program is going to run on a real platform.  Now how cool is that?

## Old neglected theorems are still theorems

March 20, 2014

I have very recently been thinking about the question of partiality vs totality in programming languages, a perennial topic in PL’s that every generation thinks it discovers for itself.  And this got me to remembering an old theorem that, it seems, hardly anyone knows ever existed in the first place.  What I like about the theorem is that it says something specific and technically accurate about the sizes of programs in total languages compared to those in partial languages.  The theorem provides some context for discussion that does not just amount to opinion or attitude (and attitude alway seems to abound when this topic arises).

The advantage of a total programming language such as Goedel’s T is that it ensures, by type checking, that every program terminates, and that every function is total. There is simply no way to have a well-typed program that goes into an infinite loop. This may seem appealing, until one considers that the upper bound on the time to termination can be quite large, so large that some terminating programs might just as well diverge as far as we humans are concerned. But never mind that, let us grant that it is a virtue of  T that it precludes divergence.

Why, then, bother with a language such as PCF that does not rule out divergence? After all, infinite loops are invariably bugs, so why not rule them out by type checking? (Don’t be fooled by glib arguments about useful programs, such as operating systems, that “run forever”. After all, infinite streams are programmable in the language M of inductive and coinductive types in which all functions terminate. Computing infinitely does not mean running forever, it just means “for as long as one wishes, without bound.”)  The notion does seem appealing until one actually tries to write a program in a language such as T.

Consider computing the greatest common divisor (GCD) of two natural numbers. This can be easily programmed in PCF by solving the following equations using general recursion:

$\begin{array}{rcl} \textit{gcd}(m,0) & = & m \\ \textit{gcd}(0,m) & = & m \\ \textit{gcd}(m,n) & = & \textit{gcd}(m-n,n) \quad \text{if}\ m>n \\ \textit{gcd}(m,n) & = & \textit{gcd}(m,n-m) \quad \text{if}\ m

The type of $\textit{gcd}$ defined in this manner has partial function type $(\mathbb{N}\times \mathbb{N})\rightharpoonup \mathbb{N}$, which suggests that it may not terminate for some inputs. But we may prove by induction on the sum of the pair of arguments that it is, in fact, a total function.

Now consider programming this function in T. It is, in fact, programmable using only primitive recursion, but the code to do it is rather painful (try it!). One way to see the problem is that in T the only form of looping is one that reduces a natural number by one on each recursive call; it is not (directly) possible to make a recursive call on a smaller number other than the immediate predecessor. In fact one may code up more general patterns of terminating recursion using only primitive recursion as a primitive, but if you examine the details, you will see that doing so comes at a significant price in performance and program complexity. Program complexity can be mitigated by building libraries that codify standard patterns of reasoning whose cost of development should be amortized over all programs, not just one in particular. But there is still the problem of performance. Indeed, the encoding of more general forms of recursion into primitive recursion means that, deep within the encoding, there must be “timer” that “goes down by ones” to ensure that the program terminates. The result will be that programs written with such libraries will not be nearly as fast as they ought to be.  (It is actually quite fun to derive “course of values” recursion from primitive recursion, and then to observe with horror what is actually going on, computationally, when using this derived notion.)

But, one may argue, T is simply not a serious language. A more serious total programming language would admit sophisticated patterns of control without performance penalty. Indeed, one could easily envision representing the natural numbers in binary, rather than unary, and allowing recursive calls to be made by halving to achieve logarithmic complexity. This is surely possible, as are numerous other such techniques. Could we not then have a practical language that rules out divergence?

We can, but at a cost.  One limitation of total programming languages is that they are not universal: you cannot write an interpreter for T within T (see Chapter 9 of PFPL for a proof).  More importantly, this limitation extends to any total language whatever.  If this limitation does not seem important, then consider the Blum Size Theorem (BST) (from 1967), which places a very different limitation on total languages.  Fix any total language, L, that permits writing functions on the natural numbers. Pick any blowup factor, say $2^{2^n}$, or however expansive you wish to be.  The BST states that there is a total function on the natural numbers that is programmable in L, but whose shortest program in L is larger by the given blowup factor than its shortest program in PCF!

The underlying idea of the proof is that in a total language the proof of termination of a program must be baked into the code itself, whereas in a partial language the termination proof is an external verification condition left to the programmer. Roughly speaking, there are, and always will be, programs whose termination proof is rather complicated to express, if you fix in advance the means by which it may be proved total. (In T it was primitive recursion, but one can be more ambitious, yet still get caught by the BST.)  But if you leave room for ingenuity, then programs can be short, precisely because they do not have to embed the proof of their termination in their own running code.

There are ways around the BST, of course, and I am not saying otherwise.  For example, the BST merely guarantees the existence of a bad case, so one can always argue that such a case will never arise in practice.  Could be, but I did mention the GCD in T problem for a reason: there are natural problems that are difficult to express in a language such as T.  By fixing the possible termination arguments in advance, one is tempting fate, for there are many problems, such as the Collatz Conjecture, for which the termination proof of a very simple piece of code has been an open problem for decades, and has resisted at least some serious attempts on it.  One could argue that such a function is of no practical use.  I agree, but I point out the example not to say that it is useful, but to say that it is likely that its eventual termination proof will be quite nasty, and that this will have to be reflected in the program itself if you are limited to a T-like language (rendering it, once again, useless).  For another example, there is no inherent reason why termination need be assured by means similar to that used in T.  We got around this issue in NuPRL by separating the code from the proof, using a type theory based on a partial programming language, not a total one.  The proof of termination is still required for typing in the core theory (but not in the theory with “bar types” for embracing partiality).  But it’s not baked into the code itself, affecting its run-time; it is “off to the side”, large though it may be).

Updates: word smithing, fixed bad link, corrected gcd, removed erroneous parenthetical reference to Coq, fixed LaTeX problems.

## Exceptions are shared secrets

December 3, 2012

It’s quite obvious to me that the treatment of exceptions in Haskell is wrong. Setting aside the example I gave before of an outright unsoundness, exceptions in Haskell are nevertheless done improperly, even if they happen to be sound. One reason is that the current formulation is not stable under seemingly mild extensions to Haskell that one might well want to consider, notably any form of parameterized module or any form of shadowing of exception declarations. For me this is enough to declare the whole thing wrong, but as it happens Haskell is too feeble to allow full counterexamples to be formulated, so one may still claim that what is there now is ok … for now.

But I wish to argue that Haskell exceptions are nevertheless poorly designed, because it misses the crucial point that exception values are shared secrets. Let us distinguish two parties, the raiser of the exception, and the handler of it. The fundamental idea of exceptions is to transfer a value from the raiser to the handler without the possibility of interception by another party. While the language of secrecy seems appropriately evocative, I hasten to add that I am not here concerned with “attackers” or suchlike, but merely with the difficulties of ensuring modular composition of programs from components. In such a setting the “attacker” is yourself, who is not malicious, but who is fallible.

By raising an exception the raiser is “contacting” a handler with a message. The raiser wishes to limit which components of a program may intercept that message. More precisely, the raiser wishes to ensure that only certain previously agreed-upon components may handle that exception, perhaps only one. This property should remain stable under extension to the program or composition with any other component. It should not be possible for an innocent third party to accidentally intercept a message that was not intended for it.

Achieving this requires a secrecy mechanism that allows the raiser and the handler(s) to agree upon their cooperation. This is accomplished by dynamic classification, exactly as it is done properly in Standard ML (but not O’Caml). The idea is that the raiser has access to a dynamically generated constructor for exception values, and any handler has access to the corresponding dynamically generated matcher for exception values. This means that the handler, and only the handler, can decode the message sent by the raiser; no other party can do anything with it other than pass it along unexamined. It is “perfectly encrypted” and cannot be deciphered by any unintended component.

The usual exception mechanisms, as distinct from exception values, allow for “wild-card handlers”, which means that an exception can be intercepted by a third party. This means that the raiser cannot ensure that the handler actually receives the message, but it can ensure, using dynamic classification, that only a legitimate handler may decipher it. Decades of experience with Standard ML shows that this is a very useful thing indeed, and has application far beyond just the simple example considered here. For full details, see my forthcoming book, for a full discussion of dynamic classification and its role for ensuring integrity and confidentiality in a program. Dynamic classification is not just for “security”, but is rather a good tool for everyday programming.

So why does Haskell not do it this way? Well, I’m not the one to answer that question, but my guess is that doing so conflicts with the monadic separation of effects. To do exceptions properly requires dynamic allocation, and this would force code that is otherwise functional into the IO monad. Alternatively, one would have to use unsafePerformIO—as in ezyang’s implementation—to “hide” the effects of exception allocation. But this would then be further evidence that the strict monadic separation of effects is untenable.

Update: Reworked last paragraph to clarify the point I am making; the previous formulation appears to have invited misinterpretation.

Update: This account of exceptions also makes clear why the perennial suggestion to put exception-raising information into types makes no sense to me. I will write more about this in a future post, but meanwhile contemplate that a computation may raise an exception that is not even in principle nameable in the type. That is, it is not conservativity that’s at issue, it’s the very idea.

## PFPL is out!

December 3, 2012

Practical Foundations for Programming Languages, published by Cambridge University Press, is now available in print! It can be ordered from the usual sources, and maybe some unusual ones as well. If you order directly from Cambridge using this link, you will get a 20% discount on the cover price (pass it on).

Since going to press I have, inevitably, been informed of some (so far minor) errors that are corrected in the online edition. These corrections will make their way into the second printing. If you see something fishy-looking, compare it with the online edition first to see whether I may have already corrected the mistake. Otherwise, send your comments to me.

By the way, the cover artwork is by Scott Draves, a former student in my group, who is now a professional artist as well as a researcher at Google in NYC. Thanks, Scott!

Update: The very first author’s copy hit my desk today!

## Yet Another Reason Not To Be Lazy Or Imperative

August 26, 2012

In an earlier post I argued that, contrary to much of the literature in the area, parallelism is all about efficiency, and has little or nothing to do with concurrency.  Concurrency is concerned with controlling non-determinism, which can arise in all sorts of situations having nothing to do with parallelism.  Process calculi, for example, are best viewed as expressing substructural composition of programs, and have very little to do with parallel computing.  (See my PFPL and Rob Simmons’ forthcoming Ph.D. dissertation for more on this perspective.)  Parallelism, on the other hand, is captured by analyzing the cost of a computation whose meaning is independent of its parallel execution.  A cost semantics specifies the abstract cost of a program that is validated by a provable implementation that transfers the abstract cost to a precise concrete cost on a particular platform.  The cost of parallel execution is neatly captured by the concept of a cost graph that captures the dynamic data dependencies among subcomputations.  Details such as the number of processors or the nature of the interconnect are factored into the provable implementation, which predicts the asymptotic behavior of a program on a hardware platform based on its cost graph.  One advantage of cost semantics for parallelism is that it is easy to teach freshmen how to write parallel programs; we’ve been doing this successfully for two years now, with little fuss or bother.

This summer Guy Blelloch and I began thinking about other characterizations of the complexity of programs besides the familiar abstractions of execution time and space requirements of a computation.  One important measure, introduced by Jeff Vitter, is called I/O Complexity.  It measures the efficiency of algorithms with respect to memory traffic, a very significant determiner of performance of programs.  The model is sufficiently abstract as to encompass several different interpretations of I/O complexity.  Basically, the model assumes an unbounded main memory in which all data is ultimately stored, and considers a cache of $M=c\times B$ blocked into chunks of size $B$ that provides quick access to main memory.  The complexity of algorithms is analyzed in terms of these parameters, under the assumption that in-cache accesses are cost-free, so that the only significant costs are those incurred by loading and flushing the cache.  You may interpret the abstract concepts of main memory and cache in the standard way as a two-level hierarchy representing, say, on- and off-chip memory access, or instead as representing a disk (or other storage medium) loaded into memory for processing.  The point is that the relative costs of processing cached versus uncached data is huge, and worth considering as a measure of the efficiency of an algorithm.

As usual in the algorithms world Vitter makes use of a low-level machine model in which to express and evaluate algorithms.  Using this model Vitter obtains a lower-bound for sorting in the I/O model, and a matching upper bound using a $k$-way merge sort, where $k$ is chosen as a function of $M$ and $B$ (that is, it is not cache oblivious in the sense of Leiserson, et al.)  Although such models provide a concrete, and well-understood, basis for analyzing algorithms, we all know full well that programming at such a low-level is at best a tedious exercise.  Worse, machine models provide no foundation for composition of programs, the single most important characteristic of higher-level language models.  (Indeed, the purpose of types is to mediate composition of components; without types, you’re toast.)

The point of Guy’s and my work this summer is to adapt the I/O model to functional programs, avoiding the mess, bother, and futility of trying to work at the machine level.  You might think that it would be impossible to reason about the cache complexity of a functional program (especially if you’re under the impression that functional programming necessarily has something to do with Haskell, which it does not, though you may perhaps say that Haskell has something to do with functional programming).  Traditional algorithms work, particularly as concerns cache complexity, is extremely finicky about memory management in order to ensure that reasonable bounds are met, and you might reasonably suspect that it will ever be thus.  The point of our paper, however, is to show that the same asymptotic bounds obtained by Vitter in the I/O model may be met using purely functional programming, provided that the functional language is (a) non-lazy (of course), and (b) implemented properly (as we describe).

Specifically, we give a cost semantics for functional programs (in the paper, a fragment of ML) that takes account of the memory traffic engendered by evaluation, and a provable implementation that validates the cost semantics by describing how to implement it on a machine-like model.  The crux of the matter is to account for the cache effects that arise from maintaining a control stack during evaluation, even though the abstract semantics has no concept of a stack (it’s part of the implementation, and cannot be avoided).  The cost semantics makes explicit the reading and allocation of values in the store (using Felleisen, Morrisett, and H’s “Abstract Models of Memory Management”), and imposes enough structure on the store to capture the critical concept of locality that is required to ensure good cache (or I/O) behavior.  The abstract model is parameterized by $M$ and $B$ described above, but interpreted as representing the number of objects in the cache and the neighborhood of an object in memory (the objects that are allocated near it, and that are therefore fetched along with the object whenever the cache is loaded).

The provable implementation is given in two steps.  First, we show how to transfer the abstract cost assigned to a computation into the amount of memory traffic incurred on an abstract machine with an explicit control stack.  The key idea here is an amortization argument that allows us to obtain tight bounds on the overhead required to maintain the stack.  Second, we show how to implement the crucial read and allocate operations that underpin the abstract semantics and the abstract machine.  Here we rely on a competitive analysis, given by Sleator, et al., of the ideal cache model, and on an amortization of the cost of garbage collection in the style of Appel.  We also make use of an original (as far as I know) technique for implementing the control stack so as to avoid unnecessary interference with the data objects in cache.  The net result is that the cost semantics provides an accurate asymptotic analysis of the I/O complexity of a functional algorithm, provided that it is implemented in the manner we describe in the paper (which, in fact, is not far from standard implementation techniques, the only trick being how to manage the control stack properly).  We then use the model to derive bounds for several algorithms that are comparable to those obtained by Vitter using a low-level machine model.

The upshot of all of this is that we can reason about the I/O or cache complexity of functional algorithms, much as we can reason about the parallel complexity of functional algorithms, namely by using a cost semantics.  There is no need to drop down to a low-level machine model to get a handle on this important performance metric for your programs, provided, of course, that you’re not stuck with a lazy language (for those poor souls, there is no hope).

## Polarity in Type Theory

August 25, 2012

There has recently arisen some misguided claims about a supposed opposition between functional and object-oriented programming.  The claims amount to a belated recognition of a fundamental structure in type theory first elucidated by Jean-Marc Andreoli, and developed in depth by Jean-Yves Girard in the context of logic, and by Paul Blain-Levy and Noam Zeilberger in the context of programming languages.  In keeping with the general principle of computational trinitarianism, the concept of polarization has meaning in proof theory, category theory, and type theory, a sure sign of its fundamental importance.

Polarization is not an issue of language design, it is an issue of type structure.  The main idea is that types may be classified as being positive or negative, with the positive being characterized by their structure and the negative being characterized by their behavior.  In a sufficiently rich type system one may consider, and make effective use of, both positive and negative types.  There is nothing remarkable or revolutionary about this, and, truly, there is nothing really new about it, other than the terminology.  But through the efforts of the above-mentioned researchers, and others, we have learned quite a lot about the importance of polarization in logic, languages, and semantics.  I find it particularly remarkable that Andreoli’s work on proof search turned out to also be of deep significance for programming languages.  This connection was developed and extended by Zeilberger, on whose dissertation I am basing this post.

The simplest and most direct way to illustrate the ideas is to consider the product type, which corresponds to conjunction in logic.  There are two possible ways that one can formulate the rules for the product type that from the point of view of inhabitation are completely equivalent, but from the point of view of computation are quite distinct.  Let us first state them as rules of logic, then equip these rules with proof terms so that we may study their operational behavior.  For the time being I will refer to these as Method 1 and Method 2, but after we examine them more carefully, we will find more descriptive names for them.

Method 1 of defining conjunction is perhaps the most familiar.  It consists of this introduction rule

$\displaystyle\frac{\Gamma\vdash A\;\textsf{true}\quad\Gamma\vdash B\;\textsf{true}}{\Gamma\vdash A\wedge B\;\textsf{true}}$

and the following two elimination rules

$\displaystyle\frac{\Gamma\vdash A\wedge B\;\textsf{true}}{\Gamma\vdash A\;\textsf{true}}\qquad\frac{\Gamma\vdash A\wedge B\;\textsf{true}}{\Gamma\vdash B\;\textsf{true}}$.

Method 2 of defining conjunction is only slightly different.  It consists of the same introduction

$\displaystyle \frac{\Gamma\vdash A\;\textsf{true}\quad\Gamma\vdash B\;\textsf{true}}{\Gamma\vdash A\wedge B\;\textsf{true}}$

and one elimination rule

$\displaystyle\frac{\Gamma\vdash A\wedge B\;\textsf{true} \quad \Gamma,A\;\textsf{true},B\;\textsf{true}\vdash C\;\textsf{true}}{\Gamma\vdash C\;\textsf{true}}$.

From a logical point of view the two formulations are interchangeable in that the rules of the one are admissible with respect to the rules of the other, given the usual structural properties of entailment, specifically reflexivity and transitivity.  However, one can discern a difference in “attitude” in the two formulations that will turn out to be a manifestation of the concept of polarity.

Method 1 is a formulation of the idea that a proof of a conjunction is anything that behaves conjunctively, which means that it supports the two elimination rules given in the definition.  There is no commitment to the internal structure of a proof, nor to the details of how projection operates; as long as there are projections, then we are satisfied that the connective is indeed conjunction.  We may consider that the elimination rules define the connective, and that the introduction rule is derived from that requirement.  Equivalently we may think of the proofs of conjunction as being coinductively defined to be as large as possible, as long as the projections are available.  Zeilberger calls this the pragmatist interpretation, following Count Basie’s principle, “if it sounds good, it is good.”

Method 2 is a direct formulation of the idea that the proofs of a conjunction are inductively defined to be as small as possible, as long as the introduction rule is valid.  Specifically, the single introduction rule may be understood as defining the structure of the sole form of proof of a conjunction, and the single elimination rule expresses the induction, or recursion, principle associated with that viewpoint.  Specifically, to reason from the fact that $A\wedge B\;\textsf{true}$ to derive $C\;\textsf{true}$, it is enough to reason from the data that went into the proof of the conjunction to derive $C\;\textsf{true}$.  We may consider that the introduction rule defines the connective, and that the elimination rule is derived from that definition.  Zeilberger calls this the verificationist interpretation.

These two perspectives may be clarified by introducing proof terms, and the associated notions of reduction that give rise to a dynamics of proofs.

When reformulated with explicit proofs, the rules of Method 1 are the familiar rules for ordered pairs:

$\displaystyle\frac{\Gamma\vdash M:A\quad\Gamma\vdash N:B}{\Gamma\vdash \langle M, N\rangle:A\wedge B}$

$\displaystyle\frac{\Gamma\vdash M:A\wedge B}{\Gamma\vdash \textsf{fst}(M):A}\qquad\frac{\Gamma\vdash M:A\wedge B}{\Gamma\vdash \textsf{snd}(M):B}$.

The associated reduction rules specify that the elimination rules are post-inverse to the introduction rules:

$\displaystyle\textsf{fst}(\langle M,N\rangle)\mapsto M \qquad \textsf{snd}(\langle M,N\rangle)\mapsto N$.

In this formulation the proposition $A\wedge B$ is often written $A\times B$, since it behaves like a Cartesian product of proofs.

When formulated with explicit proofs, Method 2 looks like this:

$\displaystyle \frac{\Gamma\vdash M:A\quad\Gamma\vdash M:B}{\Gamma\vdash M\otimes N:A\wedge B}$

$\displaystyle\frac{\Gamma\vdash M:A\wedge B \quad \Gamma,x:A,y:B\vdash N:C}{\Gamma\vdash \textsf{split}(M;x,y.N):C}$

with the reduction rule

$\displaystyle\textsf{split}(M\otimes N;x,y.P)\mapsto [M,N/x,y]P$.

With this formulation it is natural to write $A\wedge B$ as $A\otimes B$, since it behaves like a tensor product of proofs.

Since the two formulations of “conjunction” have different internal structure, we may consider them as two different connectives.  This may, at first, seem pointless, because it is easily seen that $x:A\times B\vdash M:A\otimes B$ for some $M$ and that $x:A\otimes B\vdash N:A\times B$ for some $N$, so that the two connectives are logically equivalent, and hence interchangeable in any proof.  But there is nevertheless a reason to draw the distinction, namely that they have different dynamics.

It is easy to see why.  From the pragmatic perspective, since the projections act independently of one another, there is no reason to insist that the components of a pair be evaluated before they are used.  Quite possibly we may only ever project the first component, so why bother with the second?  From the verificationist perspective, however, we are pattern matching against the proof of the conjunction, and are demanding both components at once, so it makes sense to evaluate both components of a pair in anticipation of future pattern matching.  (Admittedly, in a structural type theory one may immediately drop one of the variables on the floor and never use it, but then why give it a name at all?  In a substructural type theory such as linear type theory, this is not a possibility, and the interpretation is forced.)  Thus, the verficationist formulation corresponds to eager evaluation of pairing, and the pragmatist formulation to lazy evaluation of pairing.

Having distinguished the two forms of conjunction by their operational behavior, it is immediately clear that both forms are useful, and are by no means opposed to one another.  This is why, for example, the concept of a lazy language makes no sense, rather one should instead speak of lazy types, which are perfectly useful, but by no means the only types one should ever consider.  Similarly, the concept of an object-oriented language seems misguided, because it emphasizes the pragmatist conception, to the exclusion of the verificationist, by insisting that everything be an object characterized by its methods.

More broadly, it is useful to classify types into two polarities, the positive and the negative, corresponding to the verificationist and pragmatist perspectives.  Positive types are inductively defined by their introduction forms; they correspond to colimits, or direct limits, in category theory.  Negative types are coinductively defined by their elimination forms; they correspond to limits, or inverse limits, in category theory.  The concept of polarity is intimately related to the concept of focusing, which in logic sharpens the concept of a cut-free proof and elucidates the distinction between synchronous and asynchronous connectives, and which in programming languages provides an elegant account of pattern matching, continuations, and effects.

As ever, enduring principles emerge from the interplay between proof theory, category theory, and type theory.  Such concepts are found in nature, and do not depend on cults of personality or the fads of the computer industry for their existence or importance.

Update: word-smithing.

August 14, 2012

It is well known that Haskell is not type safe.  The most blatant violation is the all too necessary, but aptly named, unsafePerformIO operation.  You are enjoined not to use this in an unsafe manner, and must be careful to ensure that the encapsulated computation may be executed at any time because of the inherent unpredictability of lazy evaluation.  (The analogous operation in monadic ML, safePerformIO, is safe, because of the value restriction on polymorphism.)  A less blatant violation is that the equational theory of the language with seq is different from the equational theory without it, and the last I knew the GHC compiler was willing to make transformations that are valid only in the absence of this construct.  This too is well known.  A proper reformulation of the equational theory was given by Patricia Johann a few years ago as a step towards solving it.  (If the GHC compiler is no longer unsafe in this respect, it would be good to know.)

I’ve asked around a little bit, including some of the Haskell insiders, whether it is equally well known that the typed exception mechanism in Haskell is unsound.  From what I can tell this seems not to be as well-understood, so let me explain the situation as I see it, and I am sure I will be quickly corrected if I am wrong.  The starting point for me was the realization that in Haskell pure code (outside of the IO monad) may raise an exception, and, importantly for my point, the exceptions are user-defined.  Based on general semantic considerations, it seemed to me that this cannot possibly be sound, and Karl Crary helped me to isolate the source of the problem.

The difficulty is really nothing to do with exceptions per se, but rather with exception values.  (So my point has nothing whatsoever to do with imprecise exceptions.)  It seems to me that the root cause is an all-too-common misunderstanding of the concept of typed exceptions as they occur, for example, in Standard ML.  The mistake is a familiar one, the confusion of types with classes; it arises often in discussions related to oop.  To clarify the situation let me begin with a few remarks exception values in ML, and then move on to the issue at hand.

The first point, which is not particularly relevant to the present discussion, is that exceptions have nothing to do with dynamic binding.  The full details are in my book, so I will only summarize the situation here.  Many seem to have the idea that an exception handler, which typically looks like a sequence of clauses consisting of an exception and an action, amounts to a dynamic binding of each exception to its associated handler.  This is not so.  In fact the left-hand side of an exception clause is a pattern, including a variable (of which a wild card is one example), or nested patterns of exceptions and values.  I realize that one may implement the exception mechanism using a single dynamically bound symbol holding the current exception handler, but this implementation is but one of many possible ones, and does not in any case define the abstraction of exceptions.  Exceptions have no more to do with dynamic binding than does the familiar if-then-else available in any language.

The second point, which is pertinent to this discussion, is that exceptions have only one type.  You read that right: the typed exception mechanism in Standard ML is, in fact, uni-typed (where I have I heard that before?).  It is perfectly true that in Standard ML one may associate a value of any type you like with an exception, and this value may be communicated to the handler of that exception.  But it is perfectly false to say that there are multiple types of exceptions; there is only one, but it has many (in fact, “dynamically many”) classes.  When you declare, say, exception Error of string in Standard ML you are introducing a new class of type string->exn, so that Error s is an exception value of type exn carrying a string.  An exception handler associates to a computation of type α a function of type exn->α that handles any exceptions thrown by that computation.  To propagate uncaught exceptions, the handler is  implicitly post-composed with the handler fn x => raise x.  Pattern matching recovers the type of the associated value in the branch that handles a particular exception.  So, for example, a handler of the form Error x => exp propagates the fact that x has type string into the expression exp, and similarly for any other exception carrying any other type.  (Incidentally, another perfectly good handler is Error “abcdef” => exp, which handles only an Error exception with associated value “abcdef”, and no other.  This debunks the dynamic binding interpretation mentioned earlier.)

The third point is that exception classes are dynamically generated in the sense that each evaluation of an exception declaration generates a fresh exception.  This is absolutely essential for modularity, and is extremely useful for a wide variety of other purposes, including managing information flow security within a program. (See Chapter 34 of my book for a fuller discussion.)  O’Caml attempted to impose a static form of exceptions, but it is now recognized that this does not work properly in the presence of functors or separate compilation.  This means that exception declarations are inherently stateful and cannot be regarded as pure.

This got me to wondering how Haskell could get away with user-defined typed exceptions in “pure” code.  The answer seems to be “it can’t”, as the following example illustrates:

import Control.Exception

import Data.Typeable

newtype Foo = Foo (() -> IO ())

{- set Foo’s TypeRep to be the same as ErrorCall’s -}

instance Typeable Foo where

typeOf _ = typeOf (undefined :: ErrorCall)

instance Show Foo where  show _ = “”

instance Exception Foo

main = Control.Exception.catch (error “kaboom”) (\ (Foo f) -> f ())

If you run this code in GHC, you get “internal error: stg_ap_v_ret”, which amounts to “going wrong.”  The use of exceptions is really incidental, except insofar as it forces the use of the class Typeable, which is exploited here.

How are we to understand this unsoundness?  My own diagnosis is that typed exceptions are mis-implemented in Haskell using types, rather than classes.  There is no need in ML for any form of type casting to implement exceptions, because there is exactly one type of exception values, albeit one with many classes.  Haskell, on the other hand, tries to have exceptions with different types.  This immediately involves us in type casting, and hilarity ensues.

The problem appears to be difficult to fix, because to do so would require that exception values be declared within the IO monad, which runs against the design principle (unavoidable, in my opinion) that exceptions are permissible in pure code.  (Exceptions are, according to Aleks Nanevski, co-monadic, not monadic.)  Alternatively, since Haskell lacks modularity, a possible fix is to permit only static exceptions in essentially the style proposed in O’Caml.  This amounts to collecting up the exception declarations on a whole-program basis, and implicitly making a global data declaration that declares all the exception constructors “up front”.  Exception handlers would then work by pattern matching, and all would be well.  Why this is not done already is a mystery to me; the current strategy looks a lot like a kludge once you see the problem with safety.

Insofar as my interpretation of what is going on here is correct, the example once again calls into question the dogma of the separation of Church from state as it is realized in Haskell.  As beguiling as it is, the idea, in its current form, simply does not work.  Dave MacQueen recently described it to me as a “tar baby”, from the Brer Rabbit story: the more you get involved with it, the more entangled you become.

## для моих русских читателей

February 28, 2012

Владислав Натаров <4kcheshirecat@gmail.com> has kindly translated my blog posts to date into Russian.  I cannot vouch for the accuracy of the translation, but I am grateful for the effort, and hope that this will be of interest to my Russian readers.

## Words matter

February 1, 2012

Yesterday, during a very nice presentation by Ohad Kammar at Carnegie Mellon, the discussion got derailed, in part, because of a standard, and completely needless, terminological confusion involving the word “variable”.  I’m foolish enough to try to correct it.

The problem is that we’ve all been taught to confuse variables with variables—that is, program variables with mathematical variables.  The distinction is basic.  Since time immemorial (well, at least since al Khwarizmi) we have had the notion of a variable, properly so-called, which is given meaning by substitution.  A variable is an unknown, or indeterminate, quantity that can be replaced by any value of its type (a type being, at least since Russell, the range of significance of a variable).  Frege gave the first systematic study of the quantifiers, and Church exploited the crucial concept of a variable to give the most sharply original and broadly applicable model of computation, the $\lambda$-calculus.

Since the dawn of Fortran something that is not a variable has come to be called a variable.  A program variable, in the sense of Fortran and every imperative language since, is not given meaning by substitution.  Rather, it is given meaning by (at least) two operations associated with it, one to get its contents and one to put new contents into it.  (And, maybe, an operation to form a reference to it, as in C or even Algol.)  Now as many of you know, I think that the concept of a program variable in this sense is by and large a bad idea, or at any rate not nearly as important as it has been made out to be in conventional (including object-oriented) languages, but that’s an argument for another occasion.

Instead, I’m making a plea.  Let’s continue to call variables variables.  It’s a perfectly good name, and refers to what is perhaps one of the greatest achievements of the human mind, the fundamental concept of algebra, the variable.  But let’s stop calling those other things variables!  In my Practical Foundations for Programming Languages I coined (as far as I know) a word that seems perfectly serviceable, namely an assignable.  The things called variables in imperative languages should, rather, be called assignables.  The word is only a tad longer than variable, and rolls off the tongue just as easily, and has the advantage of being an accurate description of what it really is.  What’s not to like?

Why bother?  For one thing, some languages have both concepts, a necessity if you want your language to be mathematically civilized (and you do).  For another, in the increasingly important world of program verification, the specification formalisms, being mathematical in nature, make use of variables, which most definitely are not assignables!  But the real reason to make the distinction is, after all, because words matter.  Two different things deserve to have two different names, and it only confuses matters to use the same word for both.  This week’s confusion was only one example of many that I have seen over the years.

So, my suggestion: let’s call variables variables, and let’s call those other things assignables.  In the fullnesss of time (i.e., once the scourge of imperative programming has been lifted) we may not need the distinction any longer.  But until then, why not draw the distinction properly?

Addendum: It seems worth mentioning that in PFPL I have a novel (afaik) treatment of the concept of a reference, which is clarified in a subsequent post.

## Heartbeat

September 26, 2011

Just a note in response to several requests to say that I am still around, and plan to resume blogging soon.   I have been distracted by more pressing concerns, including teaching at the Oregon Programming Languages Summer School (OPLSS), preparing a submission to  POPL, visiting the Max Planck Institute for Software Systems, and trying hard to finish my book (on which I’ve made substantial progress over the summer).

Follow