[PATCH] D17203: [LICM] Hoist and sink entire inner loops.

[Chris Diamand via llvm-commits]

chrisdiamand_arm added a comment.

In http://reviews.llvm.org/D17203#354312, @reames wrote:

> Taking a step back, can you give a motivating example on why we might want to do this?  Your tests look like they'd be caught by loop-unswitch and LICM together, but I suspect that's just because the tests are (correctly) simple.


The motivation behind this is to handle things like the following:

  void expensive_loop(...) {
    for (...)
      do_stuff;
  }
  
  for (int i = 0; i < big_number; ++i)
    expensive_loop(...);

I noticed that setting `__attribute__((noinline))` on `expensive_loop()` made it run much faster. If inlining is disabled, the call to `expensive_loop()` can be sunk and is only run once. With inlining, the loop in `expensive_loop()` gets turned into a subloop, and can't be moved, so the inner loop is run `big_number` times. This patch would allow the inlined contents of `expensive_loop()` to be removed from the outer loop.

Cheers!
Chris


http://reviews.llvm.org/D17203