I think copy-less interfaces basically can be implemented on top of chunked(or block-oriented) read(or buffer) also w/o extra inefficiencies. Recently in phobos ML, someone has suggested a archetypal buffering interface allows us to build upper layers which have copy-less interfaces. If it would be polished more, probably we can find a point of compromise.
2010/10/14 Andrei Alexandrescu <seewebsiteforem...@erdani.org>: > On 10/13/2010 06:23 PM, Denis Koroskin wrote: >> >> On Thu, 14 Oct 2010 03:06:30 +0400, Andrei Alexandrescu >> <seewebsiteforem...@erdani.org> wrote: >>> >>> Gnu offers two specialized routines: >>> http://www.gnu.org/s/libc/manual/html_node/Line-Input.html. It is many >>> times more efficient than anything that can be done in client code >>> using the stdio API. I'm thinking along those lines. >>> >> >> I can easily implement similar interface on top of chunked read: >> ubyte[] readLine(ubyte[] lineBuffer); or bool readLine(ref ubyte[] >> lineBuffer); > > You can't. > >> I've quickly looked through an implementation, too, and it's still >> filling a buffer first, and then copying character byte-by-byte to the >> output string (making realloc when needed) until a delimiter is found. >> It is exactly as efficient as implemented externally. > > Except you don't have an interface to copy byte by byte. Oops... > >> It does the same >> amount of copying and memory allocations. "Many times more efficient" is >> just an overestimation. > > It's not. I measured because it was important in an application I was > working on. It's shocking how some seemingly minor changes can make a big > difference in throughput. > >> BTW, did you see my message about std.concurrency? > > Yes, but I'll need to leave the bulk of it to Sean. Thanks. > > > Andrei >
