futhark profile: incorrect source heatmap in some cases

[athas]

This nbody-bh.json produces this HTML output. The highlighting of this file seems wrong - it starts in the middle of a type definition, which should never contribute to a cost centre.

I have not investigated whether the problem is the cost centre data that is reported, or how the heatmap is produced.

[Hyalunar]

I remember experiencing this only when the source file changed in-between profiling and html-generation, which is hard to detect and/or handle.

[athas]

I am fairly sure this is not the case here, as everything is freshly generated.

[athas]

The raw timeline does seem to have a cost centre starting at that line, so it is probably not a bug in futhark profile.

[Hyalunar]

The example made me aware of a different problem: It is very hard to tell where one source range ends and another one starts.

[athas]

Oh no, I think it is because the source location of the constant 8 in the type definition of octnode is propagated to the uses of the octnode type, and our way of combining multiple source spans is just to take the minimum of starts and maximum of ends.

This is clearly not useful here, but it is not a bug in futhark profile per se.

[Hyalunar]

src/Futhark/Profile/SourceRange.hs contains a more sophisticated method for combining source ranges. But this is special to SourceRange and might require some plumbing to work with (Pos, Pos)