[Bug tree-optimization/126079] New: BB vectorization of loads with gaps is not handled well

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126079

            Bug ID: 126079
           Summary: BB vectorization of loads with gaps is not handled
                    well
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: rguenth at gcc dot gnu.org
  Target Milestone: ---

Split out from PR126053 where we have roughly

double x[4];

void foo (double *p)
{
  int tem0 = x[0];
  int tem1 = x[1];
  int tem2 = x[2];
  int tem3 = x[3];
  tem0 = tem0 + p[0];
  tem1 = tem1 + p[1];
  tem2 = tem2 + p[2];
  tem3 = tem3 + p[2];
  x[0] = tem0;
  x[1] = tem1;
  x[2] = tem2;
  x[3] = tem3;
}

and

t2.c:13:8: note:   ==> examining statement: _6 = *p_25(D);
t2.c:13:8: missed:   BB vectorization with gaps at the end of a load is not
supported
t2.c:9:18: missed:   not vectorized: relevant stmt not supported: _6 =
*p_25(D);
t2.c:13:8: note:   Building vector operands of 0x9d6be10 from scalars instead

while we vectorize the above testcase because it's still profitable, we
miss out doing a SSE vector load {p[0], p[1]} plus a splat {p[2], p[2]}
which would be cheaper than three scalar loads(?).

We expand from

 _6 = *p_25(D);
  _9 = MEM[(double *)p_25(D) + 8B];
  _40 = {_6, _9};
  _12 = MEM[(double *)p_25(D) + 16B];
  _39 = {_12, _12};

but in the end manage

        movupd  (%rdi), %xmm0

        movsd   16(%rdi), %xmm2
        unpcklpd        %xmm2, %xmm2

so recover code-gen wise on the RTL side.

Adding

  p = __builtin_assume_aligned (p, 32);

also shows we are not exploiting alignment guarantees.

We have some related bugs (one with gap in the middle), but none exact
duplicate I think.