Type inference for cat

[simonster]

The return type of cat is determined by the value of the dim argument as well as the types of the inputs, which means we can't infer the dimensionality of the returned array. OTOH dim is usually constant, so we usually have all of the information necessary to determine the return type at compile time, just not in a way that type inference can use. What is the best approach here?

[pao]

Could use a macro, same as we do for @printf?

[simonster]

A macro is an interesting idea, but I'm not sure how we'd implement it. The dimensionality is max(dim, map(ndims, args)...), but we don't know the dimensionality of the arguments before type inference, so we can't just add a typeassert.

[mbauman]

You could do the same sort of pseudo-typing that NumericExtensions does with its functors, couldn't you? E.g., @cat(n,X...) would emit a call to cat{n}(DimT{n}(),X...), where DimT is an empty type whose sole purpose is to carry around an integer as parameterized type information. Would that be sufficient for getting the type inference up and running?

[simonster]

Even if you encapsulate dim in a type, you still need to do the max operation across dim and the number of dims in the input arrays, which is not so simple to code in a way that type inference can make sense of. One approach is to make tuple of dim length and do what I did in #4741, but I wasn't really comfortable with how ugly that was at the time, and this would be even uglier because we'd be creating a tuple just for the sake of being able to perform the max operation.

[rikhuijzer]

Related PR at #45028.

[aviatesk]

Fixed by #45028, thanks for mentioning this issue on the PR, @rikhuijzer !