On Fri, Aug 25, 2023 at 4:35 PM Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > But for non-constant ones, the inline asm actually generates better > code: gcc generatea some disgusting mess with a 'bsf' followed by a > 'cmov' for the zero case, when we know better. > > See for example > > https://godbolt.org/z/jKKf48Wsf > > I don't understand why compiler people prefer a builtin that is an > untested special case that assumes that the compiler knows what is > going on (and often doesn't), over a generic escape facility that is > supported and needed anyway (inline asm). > > In other words: the statement "builtins generate better code" is > simply PROVABLY NOT TRUE. > > Builtins have often generated *worse* code than using inline asms, to > the point where "worse" is actively buggy crap. > > At least inline asms are reliable. That's a *big* deal. So 2 concerns where "I'll do it in inline asm" can pessimize codegen: 1. You alluded to this, but what happens when one of these functions is called with a constant? (Not just a literal value, but a value that can be proven constant at compile time via optimizations as well?) arch/x86/include/asm/bitops.h got this right for ffs(), but it did not for fls()! (I think that could be `if (__builtin_constant_p(x)) return x ? 32 - __builtin_clz(x) : 0;` but check my math; oh, good job arch/powerpc/include/asm/bitops.h). 2. by providing the definition of a symbol typically provided by libc (and then not building with -ffreestanding) pessimizes libcall optimization. example: https://godbolt.org/z/crrTKEf6G ffs() gets this right again by using a macro, and __always_inline can work around this somewhat (so fls() if off the hook here). But any attempt using `static inline` would be pessimized for constants. -- Thanks, ~Nick Desaulniers
