Hi, I'm trying to build a race detector for Cilk programs (as an open-source alternative to the closed-source Intel cilkscreen race detector).
I'm trying to use the -fsanitize=thread to instrument the loads and stores, and I'm also use -fcilkplus to generate cilk code at the same time. My plan is to replace the tsan library with my own library that uses a near-linear time race detection algorithm. (A race detector that works on nested fork-join parallel programs can be faster than a general race detector which must cope with arbitrary dependencies). So here's the problem: -fcilkplus introduces extra loads and stores into the code, to interact with Cilk's work stealing scheduler runtime system. I don't want those loads and stores to be instrumented with calls such as tsan_write4(). Is there some way to mark a memory access so that it's not instrumented by tsan? Cheers -Bradley
