Show closest matching methods on MethodError

[jhasse]

Hi! Often I run into no method matching errors without knowing what I'm passing wrong. To demonstrate with an example:

julia> function foo(l::Clong, i::Integer)
       println("$l $i")
       end
foo (generic function with 1 method)

julia> foo(0, 0)
ERROR: `foo` has no method matching foo(::Int64, ::Int64)

julia> foo(convert(Clong, 0), 0)
0 0

The error doesn't tell me which argument can't be implicitly converted.

The following in C gives me a better error message:

struct SomeType {};

void foo(SomeType, int) {
}

int main() {
    foo(1, 2);
}

test.cpp:10:13: error: could not convert '1' from 'int' to 'SomeType'
     foo(1, 2);
             ^

[simonster]

There are no implicit conversions in Julia method dispatch, and with multiple dispatch which argument doesn't match is not always well-defined. For example, given:

f(x::Float64, y::Float64) = ...
f(x::Int, y::Int) = ...

which argument doesn't match for f(::Float64, ::Int)? Still, it would be helpful to print the "closest" extant methods on MethodErrors, if we can come up with a metric. Related to #7512.

[JeffBezanson]

closing as dup of #7512.

[ivarne]

This is not a dup of #7512. This issue is about automatically suggesting different methods on the same function ordered by some measure of similarity. #7512 is about manually collecting a list of ambiguous and unexpectedly named functions as kind of an extension of the manual.

[jhasse]

I'm working on my idea I descriped here: #7512 (comment)

[ivarne]

Great!

I'm not yet totally convinced, because I can't really see how you can add much value by ordering the methods differently than methods(fun) already do. When there isn't a method that matches, you will probably have to look at the full method table (or documentation) in order to find the right solution. It will be exiting to see what you come up with.

There is already (#4957) about colorizing the terminal display of a MethodTable, which should make it easier to read (if it ever is going to recover from the discussion about readability of different colors).

[jhasse]

I'm using many typealiases and functions with a long signature. When I run into no method matching then it's very hard to find out what argument is wrong (I have to check every argument's type by myself). methods(fun) doesn't help me then, because it prints only one method anyway where ALL arguments are different to mine (because of typealiases).

[ivarne]

Ok, the problem definitely changes when you have long signatures and typealiases. Maybe we could use custom ordering and some sort of equality highlighting when the missing method has 3 or more arguments?

[jhasse]

Isn't that what I suggested in #7512 (comment)?

[ivarne]

As far as I can tell, your f function in that comment has only two arguments.

[jhasse]

Okay apart from only displaying it when the method has 3 or more arguments ;)
But even for methods with a short signature I think the more error info the better.

[jhasse]

Okay I've created a first patch which implements my idea. It's still very dirty and sorting isn't implemented. But apart from that it works :)
I will try to improve my patch in the next days before creating a PR. I'm still very new to Julia. Can someone tell me how I can hack on Julia's base library without having to invoke make on every change?

[ivarne]

You can redefine functions, so if you have a file replhack.jl you can just include it with the new definition

# replhack.jl
function Base.showerror(io::IO, e::MethodError)
     # snip method content.
end

#testing
try
    foo(0,0)
catch e
    Base.showerror(STDIN, e)
end

When you are finished, you can copy the code into the base/replutils.jl and run make again

In this case the important thing is actually just to create a function that takes IO and a MethodError and test with that. The REPL does not show backtraces when errors are encountered in error displaying code, so that will be far less annoying.

[ivarne]

You should also work on the latest master, as I changed this method when I merged #7688 earlier today. It will luckily be easy to copy your changes on top of mine.

[jhasse]

Thanks! I've used your replhack.jl to improve my patch.

I rebased against master and created a pull request :)