Pure higher order functions

[KennyTM~]

On Jul 7, 11 19:34, bearophile wrote:
> With the latest beta update this compiles:
>
>
> @property bool empty(T)(in T[] a) pure nothrow {
>      return !a.length;
> }
>
> @property ref T front(T)(T[] a) pure nothrow {
>      assert(a.length);
>      return a[0];
> }
>
> void popFront(A)(ref A a) pure nothrow {
>      assert(a.length);
>      a = a[1 .. $];
> }
>
> struct Map(alias fun, R) {
>      R _input;
>
>      this(R input) nothrow pure {
>          _input = input;
>      }
>
>      @property bool empty() nothrow const pure {
>          return _input.empty;
>      }
>
>      @property auto ref front() nothrow const pure {
>          return fun(_input.front);
>      }
>
>      void popFront() nothrow pure {
>          _input.popFront();
>      }
> }
>
> template map(alias fun) {
>      auto pure map(R)(R range) {
>          return Map!(fun, R)(range);
>      }
> }
>
> int sqr(int x) pure nothrow { return x * x; }
>
> pure nothrow int foo(int n) {
>      int total;
>      foreach (x; map!(sqr)([1, 2, 3, 4]))
>          total += x;
>      return total;
> }
>
> void main() {
>      assert(foo(10) == 30);
> }
>
>
> But I have had to write, this I don't know why:
> auto pure map(R)(R range) {
>
> And despite Map is a template, its methods aren't pure, so I have had to write them with pure:
> @property bool empty() nothrow const pure {
>
> So I think currently map() can't be pure.
> I think we need the notion of pure structs/classes (instances).
>
> Bye,
> bearophile

No, I think it's a bug in pure-inference.

pure int h() {
     return 1;
}
int f(alias g)() {
     return g();
}
pure int i() {
     return f!h();  // pure function 'i' cannot call impure function 'f'
}