Tests for multiply (R←X×Y)

[sloorush]

• closes Tests for Multiply (X×Y) #74

[sloorush]

Todo:

• fix cases with model not working with complex multiplication of the order 1E10 J 1E10

[sloorush]

@pmikkelsen @definitelyprobably This PR is ready for review. Please LMK what you think about it and if I have missed anything. Thanks

Note: I have skipped ⎕FR←1287 for Hcmplx because the model is not appropriate for it. I had a chat with @pmikkelsen about it and we decided it is better to skip it.

[pmikkelsen]

Looks good to me 👍

[definitelyprobably]

Apologies for the delay. I think, in general, I just want more testing for certain types of inputs. Multiply is a trivial primitive, but still...

[definitelyprobably]

Could you use {*(⍟⍺)+⍟⍵} as the model (using logarithms instead)? This would remove the need to handle complex numbers specifically.

Or, alternatively, could you use multiple models for redundancy?

[sloorush]

There is a minor problem with the model that it gives complex answers from negative numbers. I changed the model to:

    model←{⍺{
            0≡⍵:0
            0≡⍺:0
            (⌹¨÷/×⍺ ⍵)×*(⍟|⍺)+⍟|⍵
        }¨⍵
    }

(it is a makeshift approach, i know the × should not be there to multiply the sign)

The problem with this now is the direction operator approach will not give me the sign of the for the complex numbers but the direction of the complex vector(i think). Is there a solution to this?

[definitelyprobably]

Again, I would be testing more "interesting" values too. For mantis 21743 the test was using 9007199254740992, which was the largest float that can be expressed as an integer.

Perhaps also use inputs that you know will produce (close to) the largest resulting float answer (~1e154)?

[sloorush]

The more bigger values are supposed to be tested by the random number generator in RunVariations but because of the limitations of the current model I had turned it off. I will move to your model and try it out.

[sloorush]

Is there a way I can easily find these 'interesting values' in mantis? For now, are there any specific values you would want to include other than this?

[definitelyprobably]

Again, in general, testing more small numbers would be good. Including denormal values (e.g. 1e¯310). Not that it should cause any issues here, but still.

[definitelyprobably]

Multiply is a trivial primitive, but I'd still like to see tests for more places where we expect failure. E.g., ×⍨1e155. We don't want this to generate a 1287 accidentally, for example (even though that's improbable, we should still treat the interpreter as a black box and create that test).

[sloorush]

Want to confirm:

We don't want this to generate a 1287 accidentally

Should I test if this gives a domain error or not?