[Bug target/93720] New: [10 Regression] vector creation from two parts of two vectors produces TBL rather than ins

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93720

            Bug ID: 93720
           Summary: [10 Regression] vector creation from two parts of two
                    vectors produces TBL rather than ins
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64*-*-*

Take:
#define vector __attribute__((vector_size(2*sizeof(double) )))

vector double
test_vpasted2 (vector double low, vector double high)
{
          return (vector double){low[0], high[1]};
}

--- CUT ---
This produces on the trunk:
        adrp    x0, .LC0
        ldr     q2, [x0, #:lo12:.LC0]
        tbl     v0.16b, {v0.16b - v1.16b}, v2.16b
        ret
---- CUT ----
Where LC0 is {0..7,24..31}
When really this should produce just ins.

GCC 8.2 produces:
        dup     v0.2d, v0.d[0]
        ins     v0.d[1], v1.d[1]
But GCC 9 produces the correct thing of just an ins (though I don't have a
compiler to prove that).

The problem is GCC 10 converts the above code to:
  _4 = VEC_PERM_EXPR <low_1(D), high_2(D), { 0, 3 }>;

Which the back-end does not optimize to do an ins.