[InstCombine] Missed optimization for (a ^ -1) < -a

[vladimirradosavljevic]

This could be optimized to a simple a != 0.
Missed example and proof.

[pinskia]

What is interesting is GCC is able to catch this to do optimizing it first to:

  _2 = -a_4(D);
  _7 = .ADD_OVERFLOW (a_4(D), _2);
  _8 = IMAGPART_EXPR <_7>;

That is GCC is able to optimize __builtin_add_overflow(a, -a, &tmp) into (tmp = 0, a != 0).

[XChy]

LLVM(trunk) folds it now.