There are lots of places in the VM that assume that if a double is
{big,little}-endian, then so is a jlong/int64_t. Apparently this is not true
for all platforms (ARM reported this problem), so the code is not portable if
the ll_* functions are operating on the native long long type. It is
undesirable to rewrite all the ll_* functions just to change the endianness,
because it would slow down math operations on Java long values.

Example from classload.c:

           case CONSTANT_Long:
           case CONSTANT_Double: {
               unsigned high = get4bytes(context);
               unsigned low = get4bytes(context);
               int64_t value;
               value = ll_add(ll_shl(uint2ll(high), 32), uint2ll(low));
               SET_INT64(t1, &constant_pool[i], value);

dean.long@Eng 1997-07-31