fix covariant behavior with 'return scope'

[WalterBright]

This is a significantly better way to check for covariance of function parameters that have combinations of ref, return and scope. It's reduced to an enumeration of 8 cases in enum SR which then gets fed into a matrix covariant[][] to produce a yay or nay on covariance.

It makes it feasible to reason about this without the various bits of it scattered all through the code.

[MartinNowak]

You could move the translation into the combine instead of introducing another enum.
combine(STCreturn | STCscope, 0)
The cases are CTFE'ed anyhow.

[MartinNowak]

The useful we can add scope, but not subtract it, we can subtract return but not add it comments got lost.

[mathias-lang-sociomantic]

What is this fixing ? There's neither description nor a test case.

[WalterBright]

There's neither description nor a test case.

That's why I marked it as "needs work". I'm working on finding a way to resolve an ambiguity I discovered, and needed to know what the existing test suite thought of my changes.

[WalterBright]

@MartinNowak what's going on with the cycle detection in druntime?

[WalterBright]

Blocking dlang/druntime#1733

[WalterBright]

With this, druntime now compiles without changes with -dip1000 yay!

[andralex]

Please improve coverage

[andralex]

this function seems uncovered

[WalterBright]

It seems the coverage data is messed up - the enums further on are not executable code.

[MartinNowak]

You're introducing a constructor cycles here.

./dmd/src/dmd -c -o- main --DRT-oncycle=print                                                                         20:19:50
Cyclic dependency between module ddmd.traits and ddmd.mtype
ddmd.traits* ->
ddmd.mtype* ->
ddmd.expression ->
ddmd.traits*

[MartinNowak]

You should be able to run this in CTFE instead of in a constructor.

[WalterBright]

Thanks. I thought the error came from running one of the test programs, not the compiler itself. But there's nothing platform specific there - why is the error happening only one some platforms?

[MartinNowak]

You should be able to run this in CTFE instead of in a constructor.

[WalterBright]

You should be able to run this in CTFE instead of in a constructor.

Good idea. Done.

[MartinNowak]

Always confuses me that Parameter.isCovariant is used when checking for contravariant parameters.

[WalterBright]

I never remember which is which, either.

[MartinNowak]

As it's an ordering relation (e.g. subtype) the same comparison operator can be used for both, just that covariant checks A <= B and contravariant A >= B (or vice versa 😉).