Mark if(__ctfe) blocks as SCOPEctfe

[jpf91]

Allows e.g. allocating in an if(__ctfe) block in a @nogc function

[jpf91]

ping @k-hara @WalterBright

[WalterBright]

Not a bad idea. But since it is an enhancement, it needs a bugzilla entry for it. It also apparently needs to be rebased.

[9rnsr]

@jpf91 @WalterBright A blocker issue of this idea is that currently the special variable __ctfe could appear everywhere.

void main() {
    int x = __ctfe;
    if (__ctfe + 1) {}   // this is CTFE-ed branch?
    if (__ctfe * 0) {}   // this is not CTFE-ed branch?
}

Therefore, compiler cannot determine the CTFE-specific code branch in semantic analysis stage.

An easy way to resolve the issue is, that defining the statement if (__ctfe) { ... } as a special form, and stop considering __ctfe variable in other places.

[jpf91]

@WalterBright OK.
@9rnsr I think it's good enough to support only this special case in SCOPEctfe.

A little background: I've hacked gdc to report all GC allocations to completely remove the GC dependencies from druntime. I'll use that for a first port to the NintendoDS were our GC probably wouldn't work. And there's some code in druntime which uses this if(__ctfe){//Use GC} else {complex pointer stuff} pattern. More complicated patters can always be rewritten into this form.

[WalterBright]

@9rnsr you're right about this being a kludge that doesn't work in the general case. Your solution has a lot of merit - and I think it merits a wider discussion. We need a bugzilla entry and a newsgroup discussion on it.

[WalterBright]

BTW, other unconsidered cases include:

a = __ctfe ? c : d;
a = (__ctfe && func());
switch (__ctfe) { ... }
case __ctfe:

where do we stop?

[9rnsr]

The form of conditional expression a = __ctfe ? c : d; is one more useful case for one-liner hack for CTFE-ability. Other cases are not generic pattern IMO.

[MartinNowak]

where do we stop?

Probably at function border?
If it were possible to mark a function as ctfe-only than we could relax the @nogc rules for calling it.

[andralex]

So... what's the status of this? Fix? Close? Please advise.

[ibuclaw]

I suggest close, however remind me again. Why isn't static if (__ctfe) allowed? I always returns true during the interpret pass, why not the ctfe?

[ibuclaw]

I always returns true during the interpret pass, why not the ctfe?

Ah right, because the false branch would always be removed.

[ibuclaw]

Closing for #7509