[Bug tree-optimization/71343] missed optimization (can't "prove" shift and multiplication equivalence)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71343

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2016-05-30
          Component|rtl-optimization            |tree-optimization
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed and eventually not helped by factoring *4 early:

  return (a << 2) + (b << 2) == (a + b) * 4;

adjusted testcase for that:

unsigned int test12(unsigned int a , unsigned int b)
{
  unsigned int tem1 = a * 4;
  unsigned int tem2 = b * 4;
  return ((a << 2) + (b << 2)) ==  tem1 + tem2;
}

also fails to optimize.  In the end this is a missed canonicalization though
semantically shifts are not always equal to multiplication (they are for
unsigned arithmetic).

CSE could also handle this (it also "fails" to handle x + x as x * 2, but we
canonicalize that now).

I'd say the multiplication is canonical here.

We do optimize

unsigned int test12(unsigned int a , unsigned int b)
{
  unsigned int tem1 = a / 4;
  unsigned int tem2 = b / 4;
  return ((a >> 2) + (b >> 2)) ==  tem1 + tem2;
}

but pretty late in DOM2.  VRP1 is replacing the division by a shift.