frame access regression

[dlangBugzillaToGithub]

John Colvin (@John-Colvin) reported this on 2019-10-20T20:56:51Z

Transferred from https://issues.dlang.org/show_bug.cgi?id=20308

CC List

Description

template OverloadSet(Args ...)
{
    static foreach (Arg; Args)
        alias overloads = Arg;
}

struct S(handlers ...)
{
    static struct HandlerGroup
    {
        static opCall(int a)
        {
            return OverloadSet!(handlers).overloads(a);
        }
    }

    static auto ref call(alias f)()
    {
		return f(HandlerGroup.init);
    }
}

void main()
{
	assert(
		S!((double z) => z, y => y)
			.call!(r => r(1))()
        == 1);
}

onlineapp.d(13): Error: static function onlineapp.main.S!(function (double z) => z, (y) => y).S.HandlerGroup.opCall cannot access frame of function D main
onlineapp.d(26): Error: template instance `onlineapp.main.S!(function (double z) => z, (y) => y)` error instantiating

The regression was introduced by https://github.com/dlang/dmd/pull/10214 and was released in 2.087.1

[dlangBugzillaToGithub]

timon.gehr commented on 2019-12-13T01:45:19Z

Unfortunately the reason for this regression is that the bug I fixed was masking another unrelated bug. I.e., this is not my fault and I cannot fix it very easily. DMD doesn't know that your handler lambdas do not need a frame pointer at the time it checks for frame pointer accessibility. Due to the bug fixed in https://github.com/dlang/dmd/pull/10214, DMD used to think that the lambdas are not nested in a function.

The following workaround makes the code compile:

void main(){
    static handler0(double z){ return z; }
    static handler1(T)(T y){ return y; }
    assert(
        S!(handler0, handler1)
        .call!(r => r(1))()
    == 1);
}

This allows DMD to see that the handlers are static functions without analyzing their bodies.