http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39102
--- Comment #2 from Jens Maurer <jens.maurer at gmx dot net> 2012-11-23 14:07:34 UTC --- Here is another test case: struct S { unsigned char * p; unsigned char * limit; }; S g(S s); inline __attribute__((always_inline)) S f(S s, unsigned int n); S f(S s, unsigned int n) { if (__builtin_expect(s.p + n < s.limit, 1)) { s.p += n; return s; } return g(s); } extern S sext; int main() { S s = sext; s = f(s, 4); s = f(s, 8); s = f(s, 2); return (int)(unsigned long)s.p; } leading to movq %rax, (%rsp) movq %rdx, 8(%rsp) in the code paths calling g(s), although "s" is kept in registers throughout.