On Tue, Aug 18, 2020 at 12:25 PM Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote: > > On Tue, Aug 18, 2020 at 12:19 PM Linus Torvalds > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > > > On Tue, Aug 18, 2020 at 12:03 PM H. Peter Anvin <hpa@xxxxxxxxx> wrote: > > > > > > I'm not saying "change the semantics", nor am I saying that playing > > > whack-a-mole *for a limited time* is unreasonable. But I would like to go back > > > to the compiler authors and get them to implement such a #pragma: "this > > > freestanding implementation *does* support *this specific library function*, > > > and you are free to call it." > > > > I'd much rather just see the library functions as builtins that always > > do the right thing (with the fallback being "just call the standard > > function"). > > > > IOW, there's nothing wrong with -ffreestanding if you then also have > > __builtin_memcpy() etc, and they do the sane compiler optimizations > > for memcpy(). > > > > What we want to avoid is the compiler making *assumptions* based on > > standard names, because we may implement some of those things > > differently. That's asking for trouble; please don't implement routines with identifiers from libc but with differing function signatures, and then proceed to *not* use -ffreestanding. You can't have it both ways (optimizations from *not* using -ffreestanding, then breaking all kinds of assumptions based on conventions used across userspace), at least not with the tools you currently have. > > > > And honestly, a compiler that uses 'bcmp' is just broken. WTH? It's > > the year 2020, we don't use bcmp. It's that simple. Fix your damn > > broken compiler and use memcmp. The argument that memcmp is more > > expensive than bcmp is garbage legacy thinking from four decades ago. > > > > It's likely the other way around, where people have actually spent > > time on memcmp, but not on bcmp. > > > > If somebody really *wants* to use bcmp, give them the "Get off my > > lawn" flag, I wrote a paper in college on the philosophy and symbolism in "Gran Torino." Would recommend (the movie, not the paper). > > and leave them alone. But never ever should "use bcmp" be > > any kind of default behavior. That's some batshit crazy stuff. > > > > Linus > > You'll have to ask Clement about that. I'm not sure I ever saw the > "faster bcmp than memcmp" implementation, but I was told "it exists" > when I asked for a revert when all of our kernel builds went red. Also, to Clement's credit, every patch I've ever seen from Clement is backed up by data; typically fleetwide profiles at Google. "we spend a lot of time in memcmp, particularly comparing the result against zero and no other value; hmm...how do we spend less time in memcmp...oh, well there's another library function with slightly different semantics we can call instead." I don't think anyone would consider the optimization batshit crazy given the number of cycles saved across the fleet. That an embedded project didn't provide an implementation, is a footnote that can be fixed in the embedded project, either by using -ffreestanding or -fno-builtin-bcmp, which is what this series proposes to do. -- Thanks, ~Nick Desaulniers
