Willy Tarreau wrote:
+#if defined(__GNUC__) && (__GNUC__ >= 12) +__attribute__((optimize("no-tree-loop-distribute-patterns"))) +#endif static __attribute__((unused)) -size_t nolibc_strlen(const char *str
I'd suggest to use asm("") in the loop body. It worked in the past to prevent folding division loop back into division instruction.
Or switch to
size_t f(const char *s) { const char *s0 = s; while (*s++) ; return s - s0 - 1; }
which compiles to 1 branch, not 2.
But of course they could recognise this pattern too.