Hi Ludovic,
[email protected] (Ludovic Courtès) writes:
> As discussed on IRC, our R6 custom binary input ports (CBIPs) are
> currently buffered. The buffer is hard-coded and setvbuf doesn’t work
> on non-file ports.
[...]
> The patch below makes CBIPs unbuffered (see the tests for the properties
> it gives.) It works thanks to the optimization in ‘scm_c_read’ for
> unbuffered binary ports.
When you say "It works thanks to the optimization in 'scm_c_read'", do
you mean that something will be broken if another method of reading is
used?
I think it's important that everything that currently works keeps on
working, even if less efficiently.
> This is going to be a performance hit for applications that read things
> byte by byte, *or* via textual procedures (‘scm_getc’, ‘get-string’,
> etc.) But the assumption is that people rather use ‘get-bytevector-n’
> (or similar) to get a chunk of data.
>
> However! There are places in (web ...) where CBIPs are used for mixed
> binary/textual input. When that happens, all the accesses end up being
> unbuffered, which really sucks.
>
> So, what do we do? :-)
While I agree that CBIPs should be unbuffered by default, I think it's
important to have a way to enable buffering, even if it's not portable.
Also, I think we should make R6RS custom binary _output_ ports
unbuffered by default as well, and we should have a way to enable
buffering on them.
Thanks for looking into this!
Mark
