This is the mail archive of the cygwin mailing list for the Cygwin project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: higher-level IO very slow with cygwin1.dll 5.10 (due to set_flags?)

On Sat, Jun 26, 2004 at 01:41:45PM -0400, Pierre A. Humblet wrote:
>On Sat, Jun 26, 2004 at 01:09:40PM -0400, Christopher Faylor wrote:
>> On Sat, Jun 26, 2004 at 12:05:54PM -0400, Pierre A. Humblet wrote:
>> >Beware, I found this:
>> >2000-05-19  DJ Delorie  <>
>> >	* libc/include/stdio.h: no getc/putc macros for cygwin; causes
>> >	compatibility issues with different dll versions
>> >so you may need to recompile when updating cygwin.
>> Also wouldn't that work around the file locks that were ostensibly put
>> there for a reason?
>That crossed my mind. But there is no file lock in the macro, which is
>used by systems other than cygwin. How do they manage it?
>I also assume that single threaded programs don't need the lock.
>I would be perfectly happy to have a thread-unsafe getc/putc macro,
>and to use fgetc/fputc when I care about multithreading.
>Digging deeper, I see there is a function getc_unlocked. Using it
>instead of getc improves the speed by a factor 5. 
>Now that I know about it, I will redefine getc to getc_unlocked when
>porting single threaded applications.
>That issue may be a factor in the legendary slowness of Cygwin.

FWIW, I added some code to this particular locking function to avoid
doing any locks if there is only one thread.  I'm generating a snapshot
now.  It will be interesting to see if it improves things.  I didn't
do any benchmarking.  I just verified that it worked the way I thought
it should by looking at strace output.

In the process of debugging this, I found that the version of cygwin
that I'd been building was not using any of the cygwin override functions
as are required for locking.  So, my personal cygwin was already pretty
fast -- at the expense of thread safety.  I'm going to be submitting
a patch to newlib to fix this problem.

So, please give today's snapshot a try.  I just uploaded a new one
a couple of minutes ago.


Unsubscribe info:
Problem reports:

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]