[PATCH] D34471: [Inliner] Boost inlining of an indirect call to always_inline function.

[Chandler Carruth via Phabricator via llvm-commits]

chandlerc added a comment.

FWIW, I'm not really a fan of this behavior in the inliner.

I know the test case that motivates this, and as was discussed previously I suspect using a template argument would be better. I think this is actually a general pattern we can recommend rather than using alway_inline for this:

  __attribute__((always_inline)) inline void f(int, int) {
    volatile int x;
    x = 42; x = 42; x = 42; x = 42;
  }
  
  template <typename FunctionT, FunctionT *Function>
  void g(int *a, int *b, int count)  {
    for (int i = 0; i < count; ++i)
      Function(a[i], b[i]);
  }
  
  void h(int *a, int *b, int count) {
    g<decltype(f), f>(a, b, count);
  }

Even at `-O0` this doesn't allows the `always_inline` to have its effect: https://godbolt.org/g/S4HzTD

It also makes it explicit in the source code that the programmer *wants* code duplication. For something this powerful, that seems like the right general approach.


https://reviews.llvm.org/D34471