On Mon, Jan 27, 2014 at 09:43:49PM +0000, Nicholas Clark wrote:
> I don't think that there are that many places where this happens (but I
> haven't rigged the build to count them yet), so I don't think that it's
> worth adding to the C structures to remember how many slots are used, and
> updating the serialisation to serialise this. Particularly as the new
> variable length integers store 0 as 1 byte, it's likely that the overhead of
> serialising the count will be more than the saving from not serialising a
> few zeros.
Exactly 2 places. Just in compiling the setting, and just leaving 2 slots
free. The revised "count" code is this:
diff --git a/src/6model/reprs/P6opaque.c b/src/6model/reprs/P6opaque.c
index 4edbcf1..24d4392 100644
--- a/src/6model/reprs/P6opaque.c
+++ b/src/6model/reprs/P6opaque.c
@@ -893,11 +893,19 @@ static void serialize_repr_data(MVMThreadContext *tc,
MVMSTable *st, MVMSerializ
writer->write_varint(tc, writer, repr_data->unbox_str_slot);
if (repr_data->unbox_slots) {
+ unsigned int count = 0;
writer->write_varint(tc, writer, 1);
for (i = 0; i < repr_data->num_attributes; i++) {
+ if (repr_data->unbox_slots[i].repr_id == 0)
+ ++count;
writer->write_varint(tc, writer,
repr_data->unbox_slots[i].repr_id);
writer->write_varint(tc, writer, repr_data->unbox_slots[i].slot);
}
+ if (1) {
+ int fd = open("/tmp/p6o", O_WRONLY|O_APPEND|O_CREAT, 0600);
+ write(fd, &count, sizeof(count));
+ close(fd);
+ }
}
else {
writer->write_varint(tc, writer, 0);
and counting all the places where that code is reached:
$ od -i /tmp/p6o
0000000 0 0 0 0
0000020 0 0 0 2
0000040 0 0 2 0
0000060
Exactly two aren't fully used.
And I'm clearly not reading the code carefully enough, as the length is
already (effectively) serialised [that writer->write_varint(tc, writer, 1);]
so it would be possible to store the used length instead.
But for saving 4 bytes on disk, and (I think) 16 in memory, is it worth it?
Nicholas Clark
