[Issue 22092] [REG2.067] for loop over range with closure loops infinitely

[d-bugmail at puremagic.com]

https://issues.dlang.org/show_bug.cgi?id=22092

Mathias LANG <pro.mathias.lang at gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pro.mathias.lang at gmail.com

--- Comment #1 from Mathias LANG <pro.mathias.lang at gmail.com> ---
> Note that with -version=no_bug (i.e. adding a call to an empty function), the bug does not manifest.

Looking at the PR that introduced the regression, I think that is because it
prevents the compiler from inferring "pure" on `popFront`.

Currently `popFront` compiles to:
```
% nm bug.o| ddemangle | grep popFront
00000000000001d0 S const pure nothrow @nogc @safe void
bug.makeRange!().makeRange().R.popFront()
0000000000000130 S _bug.makeRange!().makeRange.R.popFront.eh
```

But it is *not* `const pure`, because it obviously modify its context!

Brainstorming, I think that, to solve this, we could:
1) Reject `const`: Consider that `const` applies to the captured context as
well.
  This is good old https://issues.dlang.org/show_bug.cgi?id=1983

2) Do not infer `pure`: Consider that the context as inherently non-pure.
  I don't think that's the right thing to do, however, the context should be
  non-pure if the function from which the context is taken isn't pure.
  Currently you can bypass `pure` by taking a pointer to a global,
  and accessing that pointer from `popFront`:
```D
void formatRange(T)(T val)
{
    for (size_t i; !val.empty; val.popFront, i) {}
}

int pos; // This changed

auto makeRange()()
{
    int* p = &pos; // This changed
    struct R
    {
        bool empty() { return pos == 3; }
        void popFront() const
        {
            ++(*p); // This changed
        }
    }
    return R();
}

void main()
{
    auto r = makeRange;
    formatRange(r);
}
```

--