4e6d3722fc
VarLocBasedLDV::collectIDsForRegs sorts a SmallVector<Register> using array_pod_sort which is a thin wrapper around qsort. That shows up as a hotspot in compile-time profiles under __GI___qsort_r. Switching this to an explicit-comparator llvm::sort call, which takes the std::sort path instead improves compile-time with no change to code-size. CTMark geomean: - stage1-O0-g: -0.41% - stage1-aarch64-O0-g: -0.58% - stage2-O0-g: -0.40% http://llvm-compile-time-tracker.com/compare.php?from=347aa3f6fbcc48cd752d02aa581b74c33d18dd41&to=cca8df56a576682510733c4c1b6fc12556e2dd7c&stat=instructions%3Au