The following example:

#include <stdint.h>

int f( int16_t a[16] )
{
    int res = 0;
    for ( int i = 0; i < 16; i++ )
      res |= (a[i]);

    return res;
}

gets vectorized with the ORs happening as int32, but since | can't overflow or underflow the ORs could have been done as int16.  This saves 2 ORs and 4 widenings.

I would have expected overwidening detection to handle this, however that fails because VRP has no range information for `a` but in this case it's safe to do based on the type alone.