On 2/12/07, Ralf Wildenhues <[EMAIL PROTECTED]> wrote:
Hello Jason, all,
* Jason Kraftcheck wrote on Mon, Feb 12, 2007 at 07:34:25PM CET:
>
> This looks like a race to create .libs by concurrent libtool processes.
Nope. Unless the output is mangled in order, the link commands are
simply issued too early.
This is what I thought at first. However, another characteristic is
that I am _not_ able to cause an error through these steps:
configure
make
make clean
PARALLEL=4
export PARALLEL
while make -P ; do make clean ; done
In other words, _after_ config.status runs, it is safe to run in
parallel. But as you can see in the log, config.status is at times
running _again_ with HP make. Thus, if I do this in Makefile.am, I
also don't see errors:
$(PROGRAMS): Makefile
We also have a very large project, but have not seen this problem in
it. Perhaps the reason is that it has a lot of $(BUILT_SOURCES) to
generate before it starts compiling and linking, and by the time it
compiles and links, config.status has already finished running.
But I am not certain of my conclusions ... just reporting some
additional behavior that I've seen.
Looks like a bug in HP-UX make. Haven't
reproduced it yet, though. My guess is it doesn't understand inference
rules with separated dependencies, when targets and/or dependencies are
in subdirectories. But I haven't been able to confirm this either.
> Perhaps libtool should check again to see if the directory exists if mkdir
> fails?
It does already.
Cheers,
Ralf
