[Bug tree-optimization/111770] predicated loads inactive lane values not modelled

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111770

--- Comment #2 from Alex Coplan <acoplan at gcc dot gnu.org> ---
I think to progress this and related cases we need to have .MASK_LOAD defined
to zero in the case that the predicate is false (either unconditionally for all
targets if possible or otherwise conditionally for targets where that is safe).

Here is a related case:

int bar(int n, char *a, char *b, char *c) {
  int sum = 0;
  for (int i = 0; i < n; ++i)
    if (c[i] == 0)
      sum += a[i] * b[i];
  return sum;
}

in this case we get the missed optimization even before vectorization during
ifcvt (in some ways it is a simpler case to consider as only scalars are
involved).  Here with -O3 -march=armv9-a from ifcvt we get:

  <bb 3> [local count: 955630224]:
  # sum_23 = PHI <_ifc__41(8), 0(18)>
  # i_25 = PHI <i_20(8), 0(18)>
  _1 = (sizetype) i_25;
  _2 = c_16(D) + _1;
  _3 = *_2;
  _29 = _3 == 0;
  _43 = _42 + _1;
  _4 = (char *) _43;
  _5 = .MASK_LOAD (_4, 8B, _29);
  _6 = (int) _5;
  _45 = _44 + _1;
  _7 = (char *) _45;
  _8 = .MASK_LOAD (_7, 8B, _29);
  _9 = (int) _8;
  _46 = (unsigned int) _6;
  _47 = (unsigned int) _9;
  _48 = _46 * _47;
  _10 = (int) _48;
  _ifc__41 = .COND_ADD (_29, sum_23, _10, sum_23);

for this case it should be possible to use an unpredicated add instead of a
.COND_ADD.  We essentially need to show that this transformation is valid:

  _29 ? sum_23 + _10 : sum_23 --> sum_23 + _10

and this essentially boils down to showing that:

  _29 = false => _10 = 0

now I'm not sure if there's a way of match-and-simplifying some GIMPLE
expression under the assumption that a given SSA name takes a particular value;
but if there were, and we defined .MASK_LOAD to zero given a false predicate,
then we could evaluate _10 under the assumption that _29 = false, which if we
added some simple match.pd rule for .MASK_LOAD with a false predicate would
allow it to evaluate to zero, and thus we could establish _10 = 0 proving the
transformation is correct.  If such an approach is possible then I guess ifcvt
could use it to avoid conditionalizing statements unnecessarily.

Richi: any thoughts on the above or on how we should handle this sort of thing?