Skip to content

Latest commit

 

History

History
36 lines (30 loc) · 1.57 KB

compiling.adoc

File metadata and controls

36 lines (30 loc) · 1.57 KB

Compiling and running Cilk

Compiling a Cilk program is similar to compiling an ordinary C or C++ program. To compile a Cilk program using Tapir/LLVM, add the -fcilkplus flag to the clang or clang++ command you would use to compile and link an ordinary serial program. When run, the resulting executable will use the Cilk runtime system to execute in parallel on whatever parallel processors are available on the machine.

Testing for determinism

We recommend checking your Cilk program for race bugs to ensure that the program is correct. The Tapir/LLVM compiler includes the Cilksan race detector for finding race bugs in Cilk programs in a provably good way. To use Cilksan, compile and link your Cilk program with the additional flag -fsanitize=cilk, and then run it normally.

Unlike many race detectors, Cilksan offers a guarantee to find a race bug if one exists. Cilksan executes a Cilk program serially, but tracks the logical parallelism between all executed instructions to determine when two instructions can race on a memory location.

Best practices

In practice, memory errors from ordinary serial programming can create race bugs in Cilk programs. The Tapir/LLVM compiler supports the AddressSanitizer tool in Clang/LLVM version 5 for finding memory errors. To use AddressSanitizer, compile and link your Cilk program with the flag -fsanitize=address, and then run the executable on a single Cilk worker as follows:

$ CILK_NWORKERS=1 ./my_cilk_program