[PATCH] D12765: [LV] Allow vectorization of loops with induction post-inc expressions

[Sanjoy Das via llvm-commits]

Hi Mikhail,

[+CC Igor -- he implemented `findExistingExpansion`]

I agree with the overall direction of the change you've proposed.
`findExistingExpansion` was designed exactly for use cases like these.

 > My concern here is that while A==B isn't equivalent to
 > getMinusSCEV(A,B)==Zero due to possible overflows (or is it?)

Overflow or no overflow, you can always add the same value on both
sides of an *equality* in modulo arithmetic (with inequalities like
`slt` there are problems around overflow).

To be more concrete:

If you have, as machine integers, getMinusSCEV(A,B) = Zero, then in
normal unbounded integer arithmetic (`rem` is the remainder operation,
`N` is is 2^BitWidth, and k, l are integers), you have:

      (A + (((-1) * B) `rem` N)) `rem` N == 0
  <=> A + (((-1) * B) `rem` N)		== kN
  <=> (((-1) * B) `rem` N)		== kN - A
  <=> ((-1) * B)				== lN + kN - A
  <=> B                                  == A - (l + k) * N

And, going back to machine (wrapping) arithmetic

  <=> A == B

-- Sanjoy