exp-drd/TODO.txt

Last updated February 17, 2008
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~


Data-race detection algorithm
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- Implement segment merging, such that the number of segments per thread
  remains limited even when there is no synchronization between threads.
- Add locking order checking.
- Discuss on the Valgrind mailing list how to make it possible to call
  POSIX thread synchronization primitives from client code (drd_intercepts.c)
  without triggering Valgrind's redirection mechanism.
- Make sure -v works again with exp-drd.
- Continue the discussion on the Valgrind mailing list about docbook and
  'make dist'.
- Continue the discussion on the Valgrind mailing list about -Wformat.
- Enable format string checking for DRD during compilation.
- Propose to modify tl_assert() / vg_assert() such that it uses
  __builtin_expect.
- Explain on the Valgrind mailing list the difference between a bus lock
  and acquire / release labels.
- Add configure tests for the name of the 'kind' member in pthread_mutex_t.
- Find out why a race is reported on std::string::string(std::string const&)
  (stc test case 16).
- Find out why no races are reported on matinv_openmp.c when "private(k)" is
  removed.
- Add support for objects that are shared over threads and that use reference
  counting, e.g. std::string -- no races should be reported on the reference
  count.
- Add test programs for destroying a condition variable / semaphore / barrier
  being waited upon.
- Make sure an error message is printed if no debug information is found in
  libopenmp.so.1.
- Make sure tc14_laog_dinphils is run during exp-drd regression tests.
- Find out why no line number or variable name information is printed for
  races detected in parallel sections of OpenMP programs. An example:
  ./vg-in-place --tool=exp-drd exp-drd/tests/omp_prime 4 -t 2
- Find a way for suppressing races reported on _IO_2_1_stdout. This race is
  triggered by calling printf() from more than one thread. Examples (AMD64):
  ./vg-in-place --tool=exp-drd exp-drd/tests/tc21_pthonce
- Find out why 'matinv 400' runs faster in 32-bit mode than in 64-bit mode
  on the same CPU.
- testing on PPC and AIX (current implementation is only tested on X86 and
  AMD64).
- Convert the array in drd_thread.c with thread information into an OSet.
- Performance testing and tuning.
- Optimize run time of the following test case:
  ./vg-in-place --tool=exp-drd exp-drd/tests/matinv 100 

Testing
~~~~~~~
- testing with more complex multithreaded test programs.
- test exp-drd's performance with the SPLASH-2 software, e.g. fft
  (http://www-flash.stanford.edu/apps/SPLASH/) or the PARSEC benchmark
  (http://parsec.cs.princeton.edu/license.htm).
- test with Synfig Studio (see also http://bugs.kde.org/show_bug.cgi?id=158555)


Documentation
~~~~~~~~~~~~~
- Document the command-line options of the exp-drd tool.
- Explain that happens-before race detectors are also suited for certain
  classes of programs that do not follow a locking discipline
  (e.g. exp-drd/test/matinv.c).


Limitations
~~~~~~~~~~~
- Does not work with a glibc library compiled with linuxthreads -- NPTL is
  required for proper operation.


Known bugs
~~~~~~~~~~
- Gets killed by the OOM handler for realistically sized applications,
  e.g. knode and OpenOffice.
- [x86_64] Reports "Allocation context: unknown" for BSS symbols on AMD64
  (works fine on i386). This is a bug in Valgrind's debug info reader
  -- VG_(find_seginfo)() returns NULL for BSS symbols on x86_64. Not yet in
  the KDE bug tracking system.