Some collected patches for gcc

• __attribute__((leafify)) support for
  • gcc-3.3
  • gcc-3.4
  • gcc-4.0
  hints the compiler to inline all calls inside a function (f.i. a math kernel). It was integrated to gcc-4.1 with a different name, flatten. Use like

  void __attribute__((leafify)) runfast(void)
  {
    for (;;)
      for (;;)
        {
          maybe_deep_template_hierarchy_call();
        }
  }
  

  such function should then compile into a straight loop without any calls left. The 3.3 and 3.4 versions are slightly hacks where leafification may not succeed in some circumstances (3.3) and inlining behavior outside of leafified functions may be affected (3.4). The mainline and lno versions of the patch work as long as unit-at-a-time mode is activated (-O1 and beyond).
• #pragma inline and #pragma inline complete are to be done as soon as the leafify stuff is accepted by gcc.

Favorite compile and runtime benchmarks

• Original tramp3d testcase. 3d monotonic transport finite difference code in Cartesian coordinates doing a Sedov explosion.
• Stripped down version tramp3d-v2.
• Even more stripped down version tramp3d-v3.
• Updated version, including OpenMP support tramp3d-v4.