Cilkscreen showing races on TBB mutex locks

[Daniel_Faken]

Hello,

When I ran cilkscreen on my program I saw some apparent data races, so I mutex-protected the read/write (with a TBB spin_rw_mutex) to see if this resolved the race report.
But instead I get a race on the TBB mutex, as shown below. Is this correct (i.e. some problem in how I'm using TBB) or is my installation somehow faulty?
Obviously this makes the race-detection hard to use..

This is with update 2 of the XE compiler on windows, and the commercial version of update 5 of TBB 3 (both released just recently).
(icl output: Intel C++ Compiler XE for applications running on IA-32, Version 12.0.2.154 Build 20110112)

Cilkscreen output:
Race condition on location 001FC98C
write access at 05DD5825: (.....\\TBB\\tbb30_u5com\\ComposerXE-2011.1.131\\tbb\\include\\tbb\\machine\\windows_ia32.h:130, modeler-core.dll!_TBB_machine_AND+0x9)
read access at 0604ABEB: (tbb.dll!tbb::spin_rw_mutex_v3::internal_acquire_reader+0x1b)
called by 05DDF9CA: (.....\\TBB\\tbb30_u5com\\ComposerXE-2011.1.131\\tbb\\include\\tbb\\spin_rw_mutex.h:105, modeler-core.dll!tbb::spin_rw_mutex_v3::scoped_lock::acquire+0x3e)
called by 05DDF961: (.....\\TBB\\tbb30_u5com\\ComposerXE-2011.1.131\\tbb\\include\\tbb\\spin_rw_mutex.h:91, modeler-core.dll!tbb::spin_rw_mutex_v3::scoped_lock::scoped_lock+0x35)
[........ lots more calls elided]

thanks very much,
Daniel Faken

[Brandon_H_Intel]

I can reproduce this, Daniel. Something specific to the spin_rw_mutex. Plain TBB mutexes work fine. I'll get this to our runtime folks and see what they can come up with.

[Daniel_Faken]

Thanks very much, Brandon.

[Brandon_H_Intel]

This works now with the 12.1 compiler (available in C++ Composer XE 2011 update 6 and later) and the applicable version of Cilkscreen for 12.1 available from the Cilk Plus SDK website. Note that for TBB locks, you will need to define TBB_USE_THREADING_TOOLS=1 for your build, probably best done using the -D option at compile time. Let me know if there's any problems.