libstdc++/4150: catastrophic performance decrease in C++ code

Loren James Rittle
Thu Apr 18 04:58:00 GMT 2002

Jason Merrill <> writes:

> The current implementation seems to be trying to allow interleaved reads
> and writes on the same buffer, whereas the libio implementation requires a
> flush when switching modes.

Agreed (that it does work this way without your patch).

> Was this difference a deliberate design decision?

Unknown (since I came into this later).

> I'm a bit skeptical of it's usefulness; how often do people
> really read a few characters, write a few characters, and then read a few
> more characters without an intervening seek?

I agree with Nathan's read of the standard here.

[Skipping analysis that seems right.]

> Anyway, here's a patch to clean up filebuf somewhat.  It involves these
> design changes: [...]

> logauswerter.C times:
>           synced   not synced
> 2.96       0:19       0:19
> pre-patch  1:09       0:14
> post-patch 0:26       0:13

(I applied the patch to mainline while I looked at it; minor tweaking
 was required and tested it on my hardware/OS.)

logauswerter.C -O2 times, ./a.out <log:
2.95.2		1:00	0:50
pre-patch-main	1:35	1:35
post-patch-main	1:23	1:05

> Interesting that v3 is faster than v2 when not synced with stdio...
> I feel like I know my way around streambufs a lot better now.

> Tested i686-pc-linux-gnu, no regressions.

Tested i386-unknown-freebsd4.5, no regressions (both 3.1 and mainline).

Hand tested some additional cases.  test03 in 27_io/
(must be enabled and run interactively by hand) is now broken on this platform.
But it may be a latent issue since the cin buffer size was always constrained
to 1 on this platform.  Can you confirm that this case isn't broken for you
with your patch?

> Any objections?

Only the concern I have is about the remaining failing interactive
test case.  Assuming that is solved: If you commit it, then I will
commit the simple patch to remove documentation and pre-port
enablement of _GLIBCPP_AVOID_FSEEK.  Thus, finally removing the
conformance hack I installed some time ago.


More information about the Libstdc++ mailing list