On 09/19/10 05:49 PM, Andrew Haley wrote:
On 19/09/10 14:01, Dr. David Kirkby wrote:
We have noted on the Sage maths project that we get occasional build
failures when the constituent parts are built in parallel. Each time a
bug is found, the Makefile gets edited, only for another to be found. I
built one component called "Singular" 110 times and it worked in all 110
times. I did this in 4 directories and started the builds at different
times, but still I got the problem.
Others had built this package in parallel too. Then a Linux user
suddenly gets a a build failure due to a race condition.
This got me thinking if there was a way to make such race conditions
more easily found. One idea I had was for gcc to have an option to sleep
for a random time before it actually did anything. So if file A takes
longer to compiler than file B, by introducing some random element, this
would change. This should make uncovering race conditions much easier.
Does that sound a sensible idea?
Sure, but it's a trivial wrapper script.
Andrew.
I always find wrapper scripts annoying though. For one reason or another, they
have often presented problems. I've seen others have problems with wrapper
scripts too. I know for a fact there's a wrapper script for Fortran built into
this code.
Can you give me your idea of what is the perfect wrapper script for
gcc/g++/gfortran? Creating a program to sleep for short delays would be quite
easy - just read from /dev/random and use nanosleep. The use of a pseduo random
number generator might be possible - though how to seed them, and save the state
needs a bit of care. But I've seen too often where a wrapper script fails under
one set of circumstances or another.
As a matter of interest, if I wanted to add such a delay into gcc myself, what
file would be best to do this in? I actually think reading an environment
variable might be better than a command line switch, as often code sets switches
and its more difficult to fiddle them. In contrast, if I could get gcc to check
for a couple of environment variables, such as
GCC_MIN_SLEEP
GCC_MAX_SLEEP
and set to some number of microseconds, it would help I think. I just don't know
where would be the best place to put such code. I'm sure if I search the source
code for "main" I'll find lots of mains, and will now know what one is the best
place.
I need this to work for gcc, gfortran and g++, so guess doing this in the
compiler, rather than the very front end, would be the best place.
Dave
