[Gnash-dev] Bitwise stream reading performance

[Udo Giacomozzi]

For an introduction to this discussion, see
http://www.gnashdev.org/wiki/index.php/NinjaProfile

I feel uncomfortable *discussing* in the Wiki, so I do it here.

Ok, since there are 18,5 million calls to read_uint() it's no wonder
it's slow.

First, it causes some million calls to read_byte() which in the end
means the equal number of calls to fread() with 1 byte counts (in case
of a local file). Maybe it would be better to implement a 
internal buffered memory stream (sort of a read-ahead buffer for the
"stream" class.

The latter would also allow faster bit-reading operations when mapping
the first two/four bytes of the buffer as a 16/32 bit unsigned integer
and just shifting as needed (instead of ORing and LOOPing). Of course,
that would depend on the host byte order.

Secondly, the read_uint() function masks the used bits after having
consumed them:

  m_current_byte &= ((1 << (m_unused_bits - bits_needed)) - 1);

Maybe it would be faster to leave m_current_byte as-is and mask the
needed bits when updating "value". Should avoid a write access to
memory and we don't need masking at all in the "consume all the unused
bits" case.

Setting m_current_byte = 0 should be unnecessary, too. And instead of
setting bits_needed = 0 a "break" would do as well (ok, not very nice,
but sligtly faster).

Udo


BTW, we have another "tu_xxxxx" with "tu_file"...