3 ms·
If something is parallelized that must not be, the result typically won't be something "close enough", which is the point in the article. Not correctness as in
by kazinator 4y ago
If something is parallelized that must not be, the result typically won't be something "close enough", which is the point in the article. Not correctness as in numerical accuracy, where we can establish tolerances.
However, the irony here is that the specific target audience for this tool actually will cheerfully accept incorrect parallelization for a speedup. Evidence for this is the use of parallel make, which does nothing to assure correctness; correctness is up to the Makefile rules expressing the correct, complete dependency graph.
It's a fairly common practice to unleash parallel make on Makefiles that were written by other people (e.g. authors of free software), where those authors provided not a shred of documented assurance that their Makefiles were designed and tested for parallel use.