John Regehr

@johnregehr The aliasing? I plan to fix it with an __attribute__((__may_alias__)) type conditional on __GNUC__, fallback to naive if !GNUC.

@johnregehr Actually, for 32-bit words, a naive 16x unrolled byte-based strlen blows away everything else in performance.

@johnregehr It consist of 100% correctly-predicted branches until end-of-string, and achieves ~1 cycle/byte even on my slow Atoms.

@RichFelker cool!

@johnregehr The bad news: compilers won't generate this unrolled code even if you ask them to. You have to write it by hand...

@RichFelker gcc 5.3 gives me a nice 8-way unroll with -funroll-all-loops -O3

@johnregehr What's your idea of a "nice unroll"? I get cmpb;leaq;je per iteration. Instead it should just do cmpb;je.

@johnregehr This change makes a 2x performance difference, from "slowest reasonable strlen" to "fastest strlen".

@johnregehr don't forget musl's beauties such as this 64-bit atomic primitive which totally won't tear: http://git.musl-libc.org/cgit/musl/tree/arch/arm/atomic.h#n247 …

@jfbastien I don't even understand

@johnregehr @jfbastien Do you have more examples of questionable musl stuff? OpenWrt (many SOHO routers & IoT things) just switched to it...

@johnregehr just the read past the end or something else I don't see?

@whitequark yep, reads past the end-- probably not a big deal

@johnregehr can you even implement fast strlen() without UB?

@whitequark I can't see any way to do it w/o padding all strings

@johnregehr @whitequark either write asm, or write the naive loop and depend on a Smart Compiler.

@johnregehr @whitequark (significantly less compiler smarts required for the naive loop than this clever implementation, BTW).

@stephentyrone @johnregehr there's something incestous in compiler recognizing this idiom and replacing it with optimized SIMD

@stephentyrone @johnregehr hell, let it also recognize bubble sort and replace with qsort. why not.