[Bug rtl-optimization/81019] [6/7/8 Regression] wrong code with -O -fno-tree-ccp

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81019

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2017-11-27
                 CC|                            |aoliva at gcc dot gnu.org,
                   |                            |segher at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
/* PR rtl-optimization/81019 */
/* { dg-do run } */
/* { dg-options "-O -fno-tree-ccp" } */

unsigned long long __attribute__((noinline, noclone))
foo (unsigned char a, unsigned short b, unsigned c, unsigned long long d,
     unsigned char e, unsigned short f, unsigned g, unsigned long long h)
{
  g = e;
  c &= 0 < d;
  b *= d;
  g ^= -1;
  g &= 1;
  c |= 1;
  a -= 0 < g;
  g >>= 1;
  f = b | (f >> b);
  return a + c + d + f + g + h;
}

int
main (void)
{
  if (foo (0, 0, 0, 0, 0, 0, 0, 0) != 0x100)
    __builtin_abort ();
  return 0;
}

This is indeed miscompiled in the combiner if the r249850 change is reverted
(which changes costs and thus combiner decides to try different optimizations).
With the change reverted, the g &= 1 masking is lost.

Alex/Segher, what is the current state of the patch?
I see https://gcc.gnu.org/ml/gcc-patches/2017-07/msg00363.html as the last mail
on the subject.