[PATCH] D33172: [InstCombine] Simpify inverted predicates in 'or'

[Davide Italiano via Phabricator via llvm-commits]

davide requested changes to this revision.
davide added a comment.
This revision now requires changes to proceed.

After giving some more careful thought, I'm not necessarily sure `InstCombine` is the right pass to solve this problem (or at least, there's an alternative solution).
I've been convincing myself this is a value-range propagation/copy-propagation problem, and in fact, this is how GCC optimizes this code.
If I understand correctly, you start with some C code that looks like this: https://godbolt.org/g/0HBds4

Now you have some pseudo-IR that looks like this (actually, this is GIMPLE, so we can map to something real :))

  <bb 2>:
  if (a_3(D) < b_4(D))
    goto <bb 4>;
  else
    goto <bb 3>;
  
  <bb 3>:
  
  <bb 4>:
  # iftmp.0_1 = PHI <c_5(D)(2), 0(3)>
  if (a_3(D) >= b_4(D))
    goto <bb 6>;
  else
    goto <bb 5>;
  
  <bb 5>:
  
  <bb 6>:
  # iftmp.1_2 = PHI <d_6(D)(4), 0(5)>
  _7 = iftmp.0_1 | iftmp.1_2;
  return _7;

your VRP algorithm can prove after SSA replacement (where LHS replaces RHS)

  a_9 -> { a_3(D) }
  a_10 -> { a_3(D) }
  b_11 -> { b_4(D) }
  b_12 -> { b_4(D) }

that your SSA values have the following ranges:

  a_9: [-INF, b_4(D) + -1]
  a_10: [b_4(D), +INF]
  b_11: [a_9 + 1, +INF]
  b_12: [-INF, a_10]

therefore removing `bb4` and transforming your IR into:

  <bb 2>:
  if (a_3(D) < b_4(D))
    goto <bb 5>;
  else
    goto <bb 4>;
  
  <bb 3>:
  # iftmp.1_2 = PHI <0(5), d_6(D)(4)>
  # iftmp.0_15 = PHI <iftmp.0_14(5), iftmp.0_13(4)>
  _7 = iftmp.0_15 | iftmp.1_2;
  return _7;
  
  <bb 4>:
  # iftmp.0_13 = PHI <0(2)>
  goto <bb 3>;
  
  <bb 5>:
  # iftmp.0_14 = PHI <c_5(D)(2)>
  goto <bb 3>;

(at that point you have a bunch of trivial phi nodes that probably copy propagation or something else could clean up, you get the idea).


https://reviews.llvm.org/D33172