[Bug rtl-optimization/53652] New: *andn* isn't used for vectorization

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53652

             Bug #: 53652
           Summary: *andn* isn't used for vectorization
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassig...@gcc.gnu.org
        ReportedBy: ja...@gcc.gnu.org
                CC: u...@gcc.gnu.org
            Target: x86_64-linux


#define N 1024
long a[N], b[N], c[N];
int d[N], e[N], f[N];

void
foo (void)
{
  int i;
  for (i = 0; i < N; i++)
    a[i] = b[i] & ~c[i];
}

void
bar (void)
{
  int i;
  for (i = 0; i < N; i++)
    d[i] = e[i] & ~f[i];
}

doesn't use *andn* insns (e.g. vandnp[sd] for -O3 -mavx).  The problem is that
combiner doesn't help here, because
(insn 42 18 33 2 (set (reg:V4DI 94)
        (mem/u/c:V4DI (symbol_ref/u:DI ("*.LC0") [flags 0x2]) [2 S32 A256])) -1
     (expr_list:REG_EQUAL (const_vector:V4DI [
                (const_int -1 [0xffffffffffffffff])
                (const_int -1 [0xffffffffffffffff])
                (const_int -1 [0xffffffffffffffff])
                (const_int -1 [0xffffffffffffffff])
            ])
        (nil)))
is before the loop and thus in a different bb,
so the combiner doesn't substitute the all ones constant into the xor (which
should fail, i?86 doesn't have a *not* SSE/AVX insn) and later on when the xor
is substituted into the and (at that point it could figure that and (xor x -1)
y
is andn).  Wonder if we should change the combiner somehow for the cases where
REG_N_SETS == 1 pseudo has REG_EQUAL note, or if we want instead to handle this
during expansion (introduce optional andnotM3 standard patterns?).