LDC generates segfaulting code with nested functions

[Safety0ff]

See this code: http://dpaste.dzfl.pl/f2d04ad8
DMD HEAD compiles this fine, whereas LDC compiles it and it crashes on run.
Older DMD gives a compilation error.

[Safety0ff]

By the way, I've edited the OP to include a reduce test case (instead of the long code I had in the OP).

[dnadlinger]

Ah, thanks for the comment (editing an issue doesn't trigger a notification on GitHub). This should make the issue much easier to track down, even though there are still Phobos dependencies left.

Regarding "older" DMD versions not accepting the code: Does 2.063.2 accept it? If not, it might be a frontend issue that we need to mitigate by also producing an error message. But that's just speculation without having looked at the test case in any detail, it could be a bug in our nested context handling code just as well.

[Safety0ff]

I should have been more clear, I did mean DMD 2.063.2 rejected it (master works fine.)
The error was generated from src/toir.c in DMD's source, I don't know which commit fixed it though (maybe the fix for http://d.puremagic.com/issues/show_bug.cgi?id=9578 , I don't know.)

[Safety0ff]

The test for http://d.puremagic.com/issues/show_bug.cgi?id=9578 from dlang/dmd@98a398b
Also segfaults:

template t9578(alias f) { void tf()() { f(); } }

void g9578a(alias f)()  { f(); }        // Error -> OK
void g9578b(alias ti)() { ti.tf(); }    // Error -> OK

void test9578()
{
    int i = 0;
    int m() { return i; }

    g9578a!(t9578!m.tf)();
    g9578b!(t9578!m)();
}
void main() { test9578(); }

[dnadlinger]

Even better, thanks. Hope I will be able to look into this tomorrow or so.

[dnadlinger]

Okay, so this is a frontend bug fixed in dlang/dmd@98a398b. The reason that the code crashes with LDC is that we pass null for the nested context pointer to tf (which is then passed to m) because DMD doesn't provide the necessary information in the AST (see the linked commit). We output null instead of producing an error because the way our glue code is set up, this can also happen in some valid cases.

[dnadlinger]

As back-porting changes to central pieces of the frontend is rather risky, I don't think we will fix this until 2.064 is merged. It would be good if we could mitigate this by adding a check similar to that in DMD, though.

[dnadlinger]

Hm, I don't think I'll be able to resolve this (i.e. mitigate the wrong-code issue to an error messages similar to what DMD 2.063.2 gives) on the weekend – somebody else wants to tackle this?

[redstar]

Fixed in branch merge-2.064.

[redstar]

Available in release 0.13.0-alpha1.