Make the synthesized class invariant a member

[TurkeyMan]

buildInv pushes the merged __invariant function into the members but never added it to the symbol table, so T.__invariant could not be named (user invariants get numbered names). It is now added, and listed by __traits(allMembers) like __xdtor, so a runtime can take its address.

This progresses towards runtime ClassInfo synthesis.

[thewilsonator]

Looks good, needs a spec update though

[dkorpel]

While it's good to mention this in the changelog, I wouldn't advertise it as something users should reach for (double underscore means internal reserved name). Maybe mention the specialized use case.

[TurkeyMan]

I'm actually tempted to just skip the changelog and spec entry for this... nobody wants to know about it, and in the future, allMembers will make it plainly discoverable to anyone.

[TurkeyMan]

I did delete it; a double-underscore symbol isn't public contract, and I don't think we need to advertise it.

[dkorpel]

While technically true, from experience with similar changes (like updating __unittest .mangleof) I have a strong feeling this will break someone's delicate introspection contraption when upgrading dmd, so it is courteous to mention it in that regard.

[TurkeyMan]

I'm sure it will, but that's the reason we make no guarantees.
So, a changelog mention? I wouldn't spec it...?

[github-actions]

DMD perf check

Metric	Base	PR	Δ
compile Phobos codegen (instr)	1,423.7 M	1,427.5 M	+0.269%

All measurements

Metric	Base	PR	Δ
compile hello.d (instr)	211.6 M	211.5 M	-0.041%
compile hello.d -O -release (instr)	229.2 M	229.1 M	-0.040%
compile Phobos (instr)	4,896.2 M	4,899.7 M	+0.071%
compile Phobos codegen (instr)	1,423.7 M	1,427.5 M	+0.269%
compile vibe.d (instr)	14,468.8 M	14,477.3 M	+0.058%
dmd binary size (stripped)	8.12 MB	8.12 MB	+0.05%
hello binary size (stripped)	0.72 MB	0.72 MB	0.00%
peak RSS (compile hello.d)	44.03 MB	44.12 MB	+0.22%
peak RSS (compile Phobos)	615.0 MB	614.9 MB	-0.01%
peak RSS (compile vibe.d)	1902 MB	1902 MB	0.00%
page faults (compile hello.d)	8,928	8,929	+0.01%
page faults (compile Phobos)	153,198	153,198	0.00%
page faults (compile vibe.d)	476,435	476,435	0.00%
compile dmd itself (wall)	12.2 s	12.3 s	+0.89%
compile hello.d (wall)	63.2 ms	63.0 ms	-0.27%
compile Phobos (wall)	1,458 ms	1,455 ms	-0.22%

_{f50dc28 vs merge-base 10adc6c · about these metrics}

[TurkeyMan]

needs a spec update though

I removed the changelog and spec update; I decided this is a private/internal symbol name, and shouldn't be part of the compiler contract. If you disagree, I can put it back, but I think this is correct; we should (theoretically) reserve the right to change __ things.

People that ever care about this in the future will easily discover it among allMembers, but we should make no commitments.

[thewilsonator]

I removed the changelog and spec update; I decided this is a private/internal symbol name, and shouldn't be part of the compiler contract. If you disagree, I can put it back, but I think this is correct; we should (theoretically) reserve the right to change __ things.

I see that __dtor is not very well spec'd, its only appearance is

ef S opAssign(ref S s)
{
    S tmp = this;   // bitcopy this into tmp
    this = s;       // bitcopy s into this
    tmp.__dtor();   // call destructor on tmp
    return this;
}

and in the output of __traits(allMembers,...) tests in traits.dd. I think you should document __invariant in the __traits(allMembers,...) example for classes and structs in traits.dd.

[TurkeyMan]

I think these should not appear in the spec personally. Are you expressing an opinion, or a ruling? Happy either way.