Change use of long double to double.

[brenthuisman]

Change from long double to double for storing floating point values in modcc.

WSL, and Windows generally, treats long double differently than Linux, which leads to inconsistencies in literal numeric values in code generated by modcc on the two platforms.

Fixes #1245

[halfflat]

We'd have to change the uses of long double to double in modcc too.

[brenthuisman]

What about long long? I think those are OK, but please confirm :)

[halfflat]

I think we can consider changing long long to an explicit width type as something to consider in the scope of a different PR.

[bcumming]

Does this fix the test on WSL? I thought that the problem was that on WSL the long double was being treated as a double. If so, using double will still create the incorrect result.

How about we instead update the test to test for relative error, and not absolute error? I would like to keep the explicit, more accurate, compile time long double if possible.

[halfflat]

My issue with long double is that it bakes in a platform-dependent numerical interpretation in generated code, making reproducibility harder; it's not just that the specific numerical environment would need to be recreated (if this is indeed possible), but this has to be done on the platform that runs modcc as well as on the execution platform.

It's not too stringent a requirement, I think, that we demand IEEE double precision semantics everywhere for correctness and reproducibility, but expecting long double behaviour to be the same everywhere is unreasonable.

[thorstenhater]

Question: What was the rationale for using long double in the first place?

[bcumming]

Rationale was that it was at least as accurate as double, likely more so. If we were going to do folding, might as well make it accurate. That was before I knew what I know now.

[brenthuisman]

The PR at present does make the unit-modcc test pass on Windows WSL.