[Haskell-cafe] New type of expressions containing (error ...) includes noisy implicit parameter

Sat Feb 13 20:52:58 UTC 2016

Replying to a few here.

Gigante:

>Just a question: how do you manage the type of simple arithmetic
expressions like 1 + 1? I mean, the type contains a constraint there.
> Prelude> :t 1 + 1
> 1 + 1 :: Num a => a

Earlier versions of the book intentionally juked this or concreted the type
of numbers to Integer, but HaskellForMac defenestrated that with the
always-on type showing so now we've had an explanation along the lines you
suggest for a few months now. We know how to adapt - we've tested the book
extensively. What we need is the implementation not to expose irrelevant
magic/noise in ways that could leak into the first chapter with Haskell
code in it.

Seidel:

>If your readers are using :t they must already know about simple types
like Integer, [], and, ->, so the new things are HasCallStack and =>.

You'd bloody think as we're careful when we introduce :t, but we get
tickets from people using HaskellForMac[1] that get confused because they
are shown the types of expressions too early. Now we're careful in how we
introduce _any_ expressions.

Fisking your attempt:

> => is just like -> except

They don't know what either of those things are or what they mean in the
second chapter because this is the _first_ chapter with any Haskell code.
They're just beginning to see how Haskell code might be kinda like the
lambdas in the lambda calculus chapter.

> the compiler fills in the argument by

We do explain what a compiler and interpreter are, but they won't know what
it means for it to fill in an argument. They don't know why it needs to
fill in an argument. Where did the argument come from?

>  HasCallStack tells the compiler

How? Why? Why do they need to care? What's a HasCallStack? Keep in mind
they don't know types, typeclasses, or anything else.

> that the expression needs a call-stack

Still don't know what a call, stack, or call-stack are.

> because it might crash.

Why does that change the type? We can construct bottoms like `let x in x`
that crash the program without changing the type.

> >So HasCallStack => [Integer] is a [Integer]

What makes this even more obnoxious is that when we finally do introduce
typeclasses and constraints, we talk about constraining a _type variable_
and now you've baked this magic in they cannot possibly be explained at all
in the book.

> that might crash and produce a stack-trace

First bit they might pick up from context, they don't know what a stack
trace is. Bonus round: when you explain things "from first principles", you
can't duck the fact that it's actually a call graph when explaining a
"stack" trace or call stack. Now you have to explain why/how it gets
flattened from one representation into the other.

Oliver had it when he said,

>"What's a call stack?"

They don't know what a stack, a call, or the combination thereof is. We had
planned to address these issues in the (much later) chapters that cover
profiling and errors. Because that's when they're relevant. This hasn't
been relevant the entire span of the book. It _never_ mattered that you
didn't get a stack trace from bottoms. In practice, does it suck? Sure! But
they're not practitioners yet! I am (I use Haskell for my 9-5 and have done
for a year and a half) and it still hasn't mattered to me. The only time
I've really wanted a stack trace is when this mechanism would not have been
available to me to begin with.

Gamari / amindfv
> I don't have a copy of GHC 8 atm to test this with: is an expression like
this now illegal?
>
> x :: Int
> x = undefined
>
>This is still valid. The change in GHC 8.0 is merely that GHC will infer
>a CallStack constraint instead of solving it in-place if asked to infer
>a type for a let binding whose RHS demands a callstack.

We have readers use the REPL _a lot_. Not only to load code but also
free-standing expressions in the REPL when experimenting and learning. Type
assignment in the REPL is noisy and we have to write around some pretty
gnarly width limitations (40-60 cols). This breaks the examples where we're
combining bottom and type inference to explore how terms and types interact.

I am less disturbed by `HasCallStack =>` than I was by the inferred type of
($).

I know designing around pedagogical limitations like this is tedious but
imagine doing it for 900-1,200 pages (formatting varies) of tutorial and
exercises, then getting unpleasant surprises right as the book is about to
be done.

Sorry about the messy thread all.

[1]: http://haskellformac.com/

On Sat, Feb 13, 2016 at 1:48 PM, Bryan Richter <b at chreekat.net> wrote:

> On Sat, Feb 13, 2016 at 09:18:07AM -0800, Eric Seidel wrote:
> > Here's what the GHCi session should look like.
> >
> > > $ ghci
> > > GHCi, version 8.0.0.20160204: http://www.haskell.org/ghc/  :? for help
> > > Loaded GHCi configuration from /home/callen/.ghci
> > > Prelude> let myList = [1..5 :: Integer]
> > > Prelude> let myList' = myList ++ undefined
> > > Prelude> :t myList'
> > > myList' :: HasCallStack => [Integer]
>
> What use case is satisfied by providing this information? How does it
> benefit the Haskell programmer? How do I use it?
>
> _______________________________________________
> Haskell-Cafe mailing list
> Haskell-Cafe at haskell.org
> http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe
>
>

-- 
Chris Allen
Currently working on http://haskellbook.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/haskell-cafe/attachments/20160213/c8213524/attachment-0001.html>