[Bug tree-optimization/103037] [11/12 Regression] Wrong code with -O2 since r11-6100-gd41b097350d3c5d0

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103037

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
More reduced testcase - the min() obfuscation is to avoid recognizing a
MIN_EXPR before jump threading gets the chance to disrupt it.  A GIMPLE unit
testcase for
PRE is difficult since we are not supporting SSA annotations (ranges) at the
moment.

static inline const unsigned short *
min(unsigned short *d, const unsigned short *e)
{
  return *e < *d ? e : d;
}

unsigned short __attribute__((noipa))
test(unsigned short arr, unsigned short val)
{
  unsigned short tem = 1;
  unsigned short tem2 = *min(&arr, &tem);
  return tem2 / (arr ? arr : val);
}

int
main()
{
  if (test (2, 2) != 0)
    __builtin_abort();
  return 0;
}

The IL for the reduced testcase before PRE is

  <bb 2> [local count: 1073741824]:
  if (arr_6(D) > 1)
    goto <bb 4>; [50.00%]
  else
    goto <bb 3>; [50.00%]

  <bb 3> [local count: 536870912]:
  # RANGE [0, 1] NONZERO 1
  _1 = (int) arr_6(D);                 // (A)
  if (arr_6(D) != 0)
    goto <bb 4>; [20.00%]
  else
    goto <bb 5>; [80.00%]

  <bb 4> [local count: 536870913]:
  # RANGE [0, 1] NONZERO 1
  # _17 = PHI <_1(3), 1(2)>
  # RANGE [1, 65535] NONZERO 65535
  iftmp.0_9 = (int) arr_6(D);          // (B) 
  goto <bb 6>; [100.00%]

  <bb 5> [local count: 536870913]:
  # RANGE [0, 65535] NONZERO 65535
  iftmp.0_8 = (int) val_7(D);

  <bb 6> [local count: 1073741824]:
  # RANGE [0, 65535] NONZERO 65535
  # iftmp.0_3 = PHI <iftmp.0_9(4), iftmp.0_8(5)>
  # RANGE [0, 1] NONZERO 1
  # _18 = PHI <_17(4), 0(5)>
  # RANGE [0, 1] NONZERO 1
  _2 = _18 / iftmp.0_3;
  # RANGE [0, 1] NONZERO 1
  _10 = (short unsigned int) _2;
  return _10;

and the problematic value-numbering is

iftmp.0_9 = _1 (0001)

at (A) and (B).  By design we may only use _available_ expressions during
simplification but _1 is not available on the 4 -> 6 edge.  That's working
OK for the translation from 4 -> 6 but then we record a valueized
expression into the ANTIC set of 4 which when further translated and
simplified has the wrong representative inside.