On Sun, Mar 14, 2021 at 9:36 AM Kristoffer Brånemyr via GNU coreutils General Discussion <coreutils@gnu.org> wrote: > > Hi, > Ok, good to know it's progressing! > I think that it would be best not to compile the main cksum.c with the -mavx > .. flags. With those flags I think the compiler is free to insert AVX > instructions in the main cksum.o file, which would fail on a CPU not > supporting it. Think for instance of a package maintainer of a linux > distributions, the compiler that person used would support -mavx so the flag > is enabled by autoconf, but another person using the resulting binary might > have a CPU not supporting AVX. That's what I tried to avoid by only having > the cksum_pclmul.c enabling the compiler flags, plus runtime detection.
+1. Do the per-object file thing. Clang will even use AVX to zero memory buffers when AVX is in effect via -mavx. That's speaking from experience from a SIGILL crash on a Core2 machine. The underlying problem is GCC, Clang and friends conflate the user's ISA with ISA the compiler uses. They are not the same - they are distinct. Unfortunately, GCC and Clang never addressed the underlying problem. Only Microsoft compilers got it right. MS compilers allow a user to use any ISA the compiler supports without options. If an option is present, like /arch:AVX2, that signals the compiler it can use the ISA, too. Jeff
