Bug 15985 – [REG2.068/2.069] Code doesn't link unless compiled with -debug

/////////////////////// test.d /////////////////////// struct S() { void delegate() d; } S!() f_Ds(U=D*)() { static if (is(U == struct)) return S!() ( { foreach (i, field; U.init.tupleof) f_Ds!(typeof(field))(); } ); static if (is(U V : V*)) return f_Ds!V(); } void f_E()() { auto f = S!() ( { enum x = is(typeof(f_Ds!(D*)())); f_Ds!(D*)(); } ); } struct A { D* d; } struct B { } struct C { A a; B b; } struct D { C c; } void main() { f_E(); f_Ds!D(); } ////////////////////////////////////////////////////// Linux , DMD 2.068.0, with -debug: OK Linux , DMD 2.068.0, without -debug: OK Linux , DMD 2.069.0, with -debug: OK Linux , DMD 2.069.0, without -debug: Link error Windows, DMD 2.067.0, with -debug: OK Windows, DMD 2.067.0, without -debug: OK Windows, DMD 2.068.0, with -debug: OK Windows, DMD 2.068.0, without -debug: Link error Bisection attempts point towards master/stable branch merges (i.e. nothing useful). It is probably a latent heisenbug triggered by other changes. Downstream issue: https://github.com/CyberShadow/Digger/issues/36

I confirmed this with dmd 2.068 and -m64 mode in Windows. So -allinst option works.

(In reply to Vladimir Panteleev from comment #0) > Bisection attempts point towards master/stable branch merges (i.e. nothing > useful). > > It is probably a latent heisenbug triggered by other changes. Sorry, that was wrong. (Digger could not bisect changes in the stable branch properly) Introduced in https://github.com/D-Programming-Language/dmd/pull/4944.

Reduced it a bit to: struct S() { void delegate() d; } S!() f_Ds(U)() { static if (is(U == struct)) return S!() ( { foreach (i, field; U.init.tupleof) f_Ds!(typeof(field))(); } ); else return f_Ds!(D)(); } void f_E() { auto f = S!() ( { enum x = is(typeof(f_Ds!(D*)())); f_Ds!(D*)(); } ); } struct A { D* d; } struct D { A a; int b; } void main() { f_E(); f_Ds!(D*)(); }

Reduced further: ---- void f()() { h!()(); } void g()() { f!()(); } void h()() { g!()(); } enum x = is(typeof(g!()())); alias my_h = h!(); void main() {} ---- And here's a variant that fails both with and without -debug: ---- void f()() { g!()(); h!()(); } void g()() { f!()(); } void h()() { f!()(); } enum x = is(typeof(f!()())); alias my_g = g!(); void main() {} ----

Original test was fixed by https://github.com/dlang/dmd/pull/13224 All other reductions still fail, and now fail *with* -debug ever since this change. https://github.com/dlang/dmd/pull/10968

(In reply to ag0aep6g from comment #4) > void f()() > { > g!()(); > h!()(); > } > > void g()() { f!()(); } > void h()() { f!()(); } > > enum x = is(typeof(f!()())); > alias my_g = g!(); > > void main() {} Here's what I suspect is going on with this. Let's follow the semantic analsysis: 1. x is the first to get semantically analyzed. The typeof instantiates f!() 2. f!() instantiates g!() and h!() 3. my_g instantiates g!() directly h!() is only instantiated indirectly, and since that instantiation happens during a typeof, it never gets marked as needing code to be generated for it.

(In reply to Walter Bright from comment #6) > > Here's what I suspect is going on with this. Let's follow the semantic > analsysis: > DMD's default behaviour is to eagerly prune templates it expects to be instantiated elsewhere. Undefined references are expected as a default behaviour. If undefined errors still persist with -allinst, then that's a bug that needs fixing.

@WalterBright created dlang/dmd pull request #14855 "fix Issue 15985 - [REG2.068/2.069] Code doesn't link unless compiled …" fixing this issue: - fix Issue 15985 - [REG2.068/2.069] Code doesn't link unless compiled with -debug https://github.com/dlang/dmd/pull/14855

dlang/dmd pull request #14855 "fix Issue 15985 - [REG2.068/2.069] Code doesn't link unless compiled …" was merged into master: - a476bc101e5c3172b8f18479fbf4557ba35a5a8a by Walter Bright: fix Issue 15985 - [REG2.068/2.069] Code doesn't link unless compiled with -debug https://github.com/dlang/dmd/pull/14855