But do “optimizations” actually produce speedups for benchmarks? Despite frequent claims by compiler maintainers that they do, they rarely present numbers to support these claims. E.g., Chris Lattner (from Clang) wrote a three-part blog posting8 about undefined behavior, with most of the first part devoted to “optimizations”, yet does not provide any speedup numbers. On the GCC side, when asked for numbers, one developer presented numbers he had from some unnamed source from IBM’s XLC compiler, not GCC; these numbers show a speedup factor 1.24 for SPECint from assuming that signed overflow does not happen (i.e., corresponding to the difference between -fwrapv and the default on GCC and Clang). Fortunately, Wang et al. [WCC+12] performed their own experiments compiling SPECint 2006 for AMD64 with both gcc-4.7 and clang-3.1 with default “optimizations” and with those “optimizations” disabled that they could identify, and running the results on a on a Core i7-980. They found speed differences on two out of the twelve SPECint benchmarks: 456.hmmer exhibits a speedup by 1.072 (GCC) or 1.09 (Clang) from assuming that signed overflow does not happen. For 462.libquantum there is a speedup by 1.063 (GCC) or 1.118 (Clang) from assuming that pointers to different types don’t alias. If the other benchmarks don’t show a speed difference, this is an overall SPECint improvement by a factor 1.011 (GCC) or 1.017 (Clang) from “optimizations”.

I mean, the speed-up of 7.2%, 9%, 6.3%, 11.8% for specific cases is huge. It will make real time and money difference for people. He is advocating against having that just because he thinks what signed integer overflow should do is obvious. Entitlement, ignorance, incompetence. I am not sure which one is it. It takes 1 minute to figure out that signed integer overflow is undefined and act accordingly.