opencilk/v2.0

Release notes for OpenCilk version 2.0

Major changes

OpenCilk 2.0 features the following major changes from OpenCilk 1.1.

• [Beta] Cilk reducer hyperobjects (a.k.a. reducers) are now supported through a new language syntax and implementation. A local or global variable in C or C++ can be made into a reducer by adding cilk_reducer(I,R) to its type, where I and R designate the identity and reduce functions for the reducer. For example, here is how a simple integer-summation reducer can be implemented using the new reducer syntax:

#include<cilk/cilk.h>

void zero(void *v) {
  *(int*)v = 0;
}

void plus(void *l, void *r) {
  *(int *)l += *(int *)r;
}

int foo(int *A, int n) {
  int cilk_reducer(zero, plus) sum = 0;
  cilk_for (int i = 0; i < n; ++i)
    sum += A[i];
  return sum;
}

• Support for the Intel Cilk Plus reducer library has been removed.
• The compiler has been upgraded to be based on LLVM 14.0.6.
• Support for pedigrees and built-in deterministic parallel random-number generation has been improved and optimized. In particular, pedigrees are now correctly updated at both spawns and syncs.
• Support for pedigrees has been streamlined. To enable pedigree support, simply link the Cilk program with the pedigree library, -lopencilk-pedigrees.
• Many bug fixes and performance improvements have been included compared to the previous version.

Known limitations

• We are writing new documentation for OpenCilk and posting it on the new website, https://www.opencilk.org/.
• With the compiler now based on LLVM 14, the default DWARF version is now DWARFv5. If you encounter an error when using a tool such as unhandled dwarf2 abbrev form code 0x25, then the tool does not support DWARFv5. You can opt back into using the old DWARF version by passing either -gdwarf-4 or -fdebug-default-version=4 to clang when you compile the program.
• The default setting of floating-point contraction is now -ffp-contract=on. As a result, floating-point computation may behave differently with this version of OpenCilk. You can opt back into the old floating-point-contraction behavior by passing the compiler flag -ffp-contract=off. See here for more details.
• There are some standard library functions and LLVM intrinsic functions that Cilksan does not recognize. When Cilksan fails to recognize such a function, it may produce a link-time error of the form, undefined reference to '__csan_FUNC' for some function name __csan_FUNC.
  • Please report these missing functions as issues on the OpenCilk issue tracker.
  • [Beta] You can work around this issue by passing the additional flag -mllvm -cilksan-bc-path=`find /path/to/opencilk/ -name "libcilksan.bc"` when compiling the Cilk program.

Acknowledgments

Thanks to everyone who contributed to the development of OpenCilk 2.0, via code contributions, testing, or in other ways, including:

• Brian Wheatman
• Teo Collin
• William Luo