On Mon, Oct 14, 2019 at 02:11:41PM -0500, Segher Boessenkool wrote: > On Mon, Oct 14, 2019 at 08:56:12AM -0700, Nick Desaulniers wrote: > > On Mon, Oct 14, 2019 at 2:35 AM Segher Boessenkool > > <seg...@kernel.crashing.org> wrote: > > > > > > On Sun, Oct 13, 2019 at 07:51:01PM -0700, Nathan Chancellor wrote: > > > > r374662 gives LLVM the ability to convert certain loops into a reference > > > > to bcmp as an optimization; this breaks prom_init_check.sh: > > > > > > When/why does LLVM think this is okay? This function has been removed > > > from POSIX over a decade ago (and before that it always was marked as > > > legacy). > > > > Segher, do you have links for any of the above? If so, that would be > > helpful to me. > > Sure! > > https://pubs.opengroup.org/onlinepubs/9699919799/xrat/V4_xsh_chap03.html > > Older versions are harder to find online, unfortunately. But there is > > https://kernel.org/pub/linux/docs/man-pages/man-pages-posix/ > > in which man3p/bcmp.3p says: > > FUTURE DIRECTIONS > This function may be withdrawn in a future version. > > Finally, the Linux man pages say (man bcmp): > > CONFORMING TO > 4.3BSD. This function is deprecated (marked as LEGACY in > POSIX.1-2001): use memcmp(3) in new programs. POSIX.1-2008 removes the > specification of bcmp(). > > > > I'm arguing against certain transforms that assume that > > one library function is faster than another, when such claims are > > based on measurements from one stdlib implementation. > > Wow. The difference between memcmp and bcmp is trivial (just the return > value is different, and that costs hardly anything to add). And memcmp > is guaranteed to exist since C89/C90 at least. > > > The rationale for why it was added was that memcmp takes a measurable > > amount of time in Google's fleet, and most calls to memcmp don't care > > about the position of the mismatch; bcmp is lower overhead (or at > > least for our libc implementation, not sure about others). > > You just have to do the read of the last words you compare as big-endian, > and then you can just subtract the two words, convert that to "int" (which > is very inconvenient to do, but hardly expensive), and there you go. > > Or on x86 use the bswap insn, or something like it. > > Or, if you use GCC, it has __builtin_memcmp but also __builtin_memcmp_eq, > and those are automatically used, too. > > > Segher
Just as an FYI, there was some more discussion around the availablity and use of bcmp in this LLVM bug which spawned commit 5f074f3e192f ("lib/string.c: implement a basic bcmp"). https://bugs.llvm.org/show_bug.cgi?id=41035#c13 I believe this is the proper solution but I am fine with whatever works, I just want our CI to be green without any out of tree patches again... Cheers, Nathan