https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67937
--- Comment #2 from Joshua Cranmer <Pidgeot18 at gmail dot com> --- Created attachment 36486 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36486&action=edit test-case.gcno And the corresponding .gcno file. The testcase was minimized by bisecting the original .gcda/.gcno files by dropping functions and seeing where my attempt to write a new implementation that matched gcov's output failed. The attempt (I did several!) that failed this time was using Boost graph library's hawick_circuits detector, which matched the input incorrectly since this is double-counting loop edges if one of them is negative.