Currently, attributes such as @safe, pure and nothrow are automatically inferred for function templates.  They should also be inferred for methods of struct/class templates and for Voldemort types.

Test case:

  struct S1
  {
      void foo(T)() { }
  }

  struct S2(T)
  {
      void foo() { }
  }

  auto makeS3(T)()
  {
      struct S3
      {
          void foo() { }
      }
      return S3();
  }

  void main() @safe pure nothrow
  {
      // Works
      S1 s1;
      s1.foo!int();

      // Compilation failure
      S2!int s2;
      s2.foo();

      // Compilation failure
      auto s3 = makeS3!int();
      s3.foo();
  }

This currently prevents large parts of Phobos from being used in @safe/pure/nothrow contexts, and it prevents parts of Phobos from being marked as such.  It also blocks some changes to std.path that I have in the pipeline, because I want to reimplement a function in terms of std.algorithm functions without removing its current attributes (as this would be a breaking change).

A workaround (which is *not* feasible for use in Phobos) is to make the methods themselves trivial templates:

  struct S2(T)
  {
      void foo()() { }
  }