Version 50 (modified by guest, 10 years ago) (diff) |
---|

# Maintaining an explicit call stack

There has been a vigorous thread on error attribution ("I get a `head []` error; but who called `head`?"). This page summarises some half baked ideas that Simon and I have been discussing. Do by all means edit this page to add comments and further ideas or pointers. (As usual, *discussion* is best done by email; but this page could be a place to record ideas, design alternatives, list pros and cons, pointers to related work etc.)

See also

## The basic idea

- GHC's 'assert' magically injects the current file location. One could imagine generalising this a bit so that you could say
...(f $currentLocation)...

to pass a string describing the current location to f.

- But that doesn't help with 'head'. We want to pass head's
*call site*to head. That's what jhc does when you give 'head' the a magic SRCLOC_ANNOTATE pragma:- every call to
`head`gets replaced with`head_check $currentLocation` - in jhc, you get to write
`head_check`yourself, with typehead_check :: String -> [a] -> a

- every call to

It'd be nicer if you didn't have to write `head_check` yourself, but instead the compiler wrote it.

- But what about the caller of the function that calls head? Obviously we'd like to pass that on too!
foo :: [Int] -> Int {-# SRCLOC_ANNOTATE foo #-} foo xs = head (filter odd xs) ===> foo_check :: String -> [Int] -> Int foo_check s xs = head_check ("line 5 in Bar.hs" ++ s) xs

Now in effect, we build up a call stack. Now we *really* want the compiler to write `foo_check`.

- In fact, it's very similar to the "cost-centre stack" that GHC builds for profiling, except that it's explicit rather than implicit. (Which is good. Of course the stack should be a proper data type, not a String.)

However, unlike GHC's profiling stuff, it is *selective*. You can choose to annotate just one function, or 10, or all. If call an annotated function from an unannotated one, you get only the information that it was called from the unannotated one:

foo :: [Int] -> Int -- No SRCLOC_ANNOTATE foo xs = head (filter odd xs) ===> foo:: [Int] -> Int foo xs = head_check ("line 5 in Bar.hs") xs

This selectiveness makes it much less heavyweight than GHC's currrent "recompile everything" story.

- The dynamic hpc tracer will allow reverse time-travel, from an exception to the call site, by keeping a small queue of recently ticked locations. This will make it easier to find out what called the error calling function (head, !, !!, etc.), but will require a hpc-trace compiled prelude if we want to find places in the prelude that called the error. (A standard prelude would find the prelude function that was called that called the error inducing function).

## Open questions

Lots of open questions

- It would be great to use the exact same stack value for profiling. Not so easy...for example, time profiling uses sampling based on timer interrupts that expect to find the current cost centre stack in a particular register. But a big pay-off; instead of having magic rules in GHC to handle SCC annotations, we could throw the full might of the Simplifier at it.

- CAFs are a nightmare. Here's a nasty case:
foo :: Int -> Int -> Int foo = \x. if fac x > 111 then \y. stuff else \y. other-stuff bad :: Int -> Int bad = foo 77

How would you like to transform this?

## An example to consider

Suppose we are to transform the following code. The line numbers in comments are useful later when we consider what hat-stack does.

main = print d d :: Int d = e [] {- line 4 -} e :: [Int] -> Int e = f 10 f :: Int -> [Int] -> Int f = \x -> case fac x < 10 of True -> \_ -> 3 False -> hd hd :: [a] -> a hd = \x -> case x of {- line 15 -} [] -> error "hd: empty list" {- line 16 -} (x:_) -> x fac :: Int -> Int fac = \n -> case n == 0 of True -> 1 False -> n * fac (n - 1)

People probably don't write code like this very much, but nonetheless, it does expose some of the issues we must deal with.

Here is a basic term-reduction for the program:

main => print d => print (e []) => print ((f 10) []) => print ((case fac 10 < 10 of { True -> \_ -> 3; False -> hd }) []) => .. some reductions of fac, and then select the False branch of the case ... => print (hd []) => print (error "hd: empty list") => <uncaught exception>

By the time that `hd` is called on the empty list, the only thing remaining on the dynamic evaluation stack is `print`.

The question is what stack would you *like* to see in this case?

One option is this:

main -> d -> e -> f -> hd

Another option is this:

main -> d -> hd

The difference between these two stacks is how we determine *when* `hd` is called. Is it in the context where `hd` is first mentioned by name (in the body of `f`), or is it when `hd` becomes fully saturated (in the body of `d`)? Both contexts seem reasonable. Does it really matter which one is chosen? At the moment I can't say for sure.

There are more possibilities, for instance, we treat CAFs as roots of the stack, thus dropping `main` and `d` from the first of the options above:

e -> f -> hd

And so on.

## What does Hat do?

As a starting point it is useful to see what Hat does, in particular hat-stack, which is the tool for generating stack traces. The example of "`hd []`" is exactly the kind of problem that hat-stack is designed to tackle.

Here is the stack trace generated for the example program. You can see the relevant line numbers in comments in the source code above. I've added comments to the hat output on the RHS to emphasise what the entries mean, when it is not immediately clear.

Program terminated with error: hd: empty list Virtual stack trace: (unknown) {?} (H.hs:16) error "hd: empty..." (H.hs:15) (\..) [] | case [] {- case analysis of [] in hd -} (H.hs:4) (\..) [] {- hd applied to [] in the body of d -} (unknown) d

More-or-less this stack resembles:

d -> hd

Curiously, if we change the definition of `hd` to a function binding instead of a pattern binding we get this stack trace instead:

Program terminated with error: hd: empty list Virtual stack trace: (unknown) {?} (I.hs:16) error "hd: empty..." (I.hs:15) hd [] | case [] (I.hs:4) hd [] (unknown) d

Which is a little clearer, but still represents:

d -> hd

So hat-stack considers the context in which a function is saturated to be the place where it is called. To make this even more apparent we can eta-expand `e`:

e :: [Int] -> Int e x = f 10 x

Now we get this stack trace:

Program terminated with error: hd: empty list Virtual stack trace: (unknown) {?} (J.hs:16) error "hd: empty..." (J.hs:15) hd [] | case [] (J.hs:7) hd [] (J.hs:4) e [] (unknown) d

Which is:

d -> e -> hd

This is because `hd` now becomes saturated in the body of `e`.

## Transformation rules

The key issues are:

**Higher-order calls**. Where does a partially applied function receive its stack trace from? Possible options include:- The lexical call site (corresponding to where the function is mentioned in the source code).
- The context in which the function receives a particular argument, for instance the one where it is saturated.

**CAFs**. The problem with CAFs is that, at least for expensive ones, we want to preserve the sharing of their evaluation. Therefore we cannot simply extend them into functions which take a stack trace as an argument - this would cause the CAF to be recomputed at each place where it is called. The simplest solution is to make CAFs the roots of call stacks, but it seems like there will be situations where it would be useful to know more about the context in which a CAF was evaluated.

**Recursion**. Obviously the call stack will grow in size proportional to the depth of recursion. This could lead to prohibitive space usage, thus it is desirable that the size of the stack be kept within reasonable bounds. We will probably need some way to dynamically prune the stack.

**Extent**. It is desirable to have a scheme where only some functions in the program are transformed for stack tracing, whilst others remain in their original form. We want to avoid the all-or-nothing situation, where the whole program has to be recompiled before tracing can be done. For example, with the current state of profiling in GHC, the whole program has to be recompiled, and special profiling libraries must be linked against. This is a nuisance which reduces the usability of the system. Similar problems occur with other debugging tools, such as Hat and buddha, and this really hampers their acceptance by programmers.

### An abstract syntax

For the purpose of exploring the rules we need an abstract syntax. Below is one for a simple core functional language:

Decls(D) --> x :: T | x = E | data f a1 .. an = K1 .. Km Constructors(K) --> k T1 .. Tn Types(T) --> f | a | T1 T2 Expressions(E) --> x | k | E1 E2 | let D1 .. Dn in E | case E of A1 .. An | \y1 .. yn -> E Alts(A) --> p -> E Pats(P) --> x | k P1 .. Pn

### Stack representation

For simplicity we assume:

type Stack = [String]

which is just a list of function names.

### Notation

Double square brackets denote the transformation function, which has either one or two arguments, depending on what type of entity it is applied to. In most cases it has one argument, which is just a syntactic construct, but for expressions it has an additional argument which represents the current stack value.

For instance:

[[ E ]]_k

means transform expression E with k as the current stack value.

### Transformation option 1

This is probably the simplest transformation style possible. Stack traces are passed to (let bound) functions at their lexical call sites, which correspond to the places where the function is mentioned in the source code. CAF bindings are treated as roots of stacks, so only function bindings receive stack arguments. In this transformation we can get away with simply passing one stack argument for each function, regardless of how many regular arguments it has. In contrast, other transformation styles might pass one stack argument for every regular argument of the function.

Declarations: [[ x :: T ]] ==> x :: Trace -> T , x is function bound, and transformed for tracing [[ x :: T ]] ==> x :: T , x does not match the above rule [[ x = \y1 .. yn -> E ]] ==> x = \t y1 .. yn -> [[ E ]]_("x":t) , x is transformed for tracing [[ x = \y1 .. yn -> E ]] ==> x = \y1 .. yn -> [[ E ]]_["x"] , x is not transformed for tracing [[ x = E ]] ==> x = [[ E ]]_["x"] [[ data f a1 .. an = K1 .. Km ]] ==> data f a1 .. an = K1 .. Km Expressions: [[ x ]]_t ==> x t , x is function bound, and transformed for tracing [[ x ]]_t ==> x , x does not match the above rule [[ k ]]_t ==> k [[ E1 E2 ]]_t ==> [[ E1 ]]_t [[ E2 ]]_t [[ let D1 .. Dn in E ]]_t ==> let [[ D1 ]] .. [[ Dn ]] in [[ E ]]_t [[ case E of A1 .. An ]]_t ==> case [[ E ]]_t of [[ A1 ]]_t .. [[ An ]]_t [[ \y1 .. yn -> E ]]_t ==> \y1 .. yn -> [[ E ]]_t Alternatives: [[ p -> E ]]_t ==> p -> [[ E ]]_t

An advantage of this transformation style is that it handles combinations of transformed and untransformed functions easily. When variable expressions are transformed we simply check to see if the variable corresponds to a transformed function. If it does, we pass it the current stack value as an argument, otherwise we don't.

If you apply this transformation to the example above you get:

main = print d d :: Int d = e [] e :: [Int] -> Int e = f 10 f :: Int -> [Int] -> Int f = \x -> case fac x < 10 of True -> \_ -> 3 False -> hd hd :: [a] -> a hd = \x -> case x of [] -> error "hd: empty list" (x:_) -> x fac :: Int -> Int fac = \n -> case n == 0 of True -> 1 False -> n * fac (n - 1)

A problem with this transformation style is that it is sensitive to the position of lambdas in the body of a declaration. For example, it transforms these two functions differently, even though they are semantically equivalent:

f1 = let x = EXP in (\y -> head (foo x)) f2 = \y -> head (foo (let x = EXP in x))

Here is the output of the two different transformations:

f1 = let x = EXP in (\y -> head ["f1"] (foo ["f1"] x)) f2 = \t y -> head ("f2":t) (foo ("f2":t) (let x = EXP in x))

Notice that in the first case the stack passed to `head` and `foo` is simply `["f1"]`, but in the second case it is `"f2":t`. One *might* expect the same stack trace to be generated for each declaration.