On Wed, Jun 07, 2017 at 02:06:13PM +0200, Peter Zijlstra wrote: > On Tue, Jun 06, 2017 at 04:00:03PM -0700, Palmer Dabbelt wrote: > What pretty much all the other architectures do is something like: > > #define ATOMIC_OP(op, asm_op, c_op) \ > static __always_inline void atomic_##op(int i, atomic_t *v) \ > { \ > __asm__ __volatile__ ( \ > "amo" #asm_op ".w zero, %1, %0" \ > : "+A" (v->counter) \ > : "r" (i)); \ > } > > #define ATOMIC_FETCH_OP(op, asm_op, c_op) \ > static __always_inline int atomic_fetch_##op(int i, atomic_t *v)\ > { \ > register int ret; \ > __asm__ __volatile__ ( \ > "amo" #asm_op ".w %2, %1, %0" \ > : "+A" (v->counter), "=r" (ret) \ > : "r" (mask)); \ > return ret; \ > } > > #define ATOMIC_OP_RETURN(op, asm_op, c_op) \ > static __always_inline int atomic_##op##_return(int i, atomic_t *v) \ > { \ > return atomic_fetch_##op(i, v) c_op i; \ > } > > #define ATOMIC_OPS(op, asm_op, c_op) \ > ATOMIC_OP(op, asm_op, c_op) \ > ATOMIC_OP_RETURN(op, asm_op, c_op) \ > ATOMIC_FETCH_OP(op, asm_op, c_op) > > ATOMIC_OPS(add, add, +) > ATOMIC_OPS(sub, sub, -) > > #undef ATOMIC_OPS > > #define ATOMIC_OPS(op, asm_op, c_op) \ > ATOMIC_OP(op, asm_op, c_op) \ > ATOMIC_FETCH_OP(op, asm_op, c_op) > > ATOMIC_OPS(and, and, &) > ATOMIC_OPS(or, or, |) > ATOMIC_OPS(xor, xor, ^) > > #undef ATOMIC_OPS > > Which is much simpler no? In fact, after having read your manual you'd want something like: #define ATOMIC_OP(op, asm_op, c_op) \ static __always_inline void atomic_##op(int i, atomic_t *v) \ { \ __asm__ __volatile__ ( \ "amo" #asm_op ".w zero, %1, %0" \ : "+A" (v->counter) \ : "r" (i)); \ } #define ATOMIC_FETCH_OP(op, asm_op, c_op, asm_or, order) \ static __always_inline int atomic_fetch_##op##order(int i, atomic_t *v)\ { \ register int ret; \ __asm__ __volatile__ ( \ "amo" #asm_op ".w" #asm_or " %2, %1, %0" \ : "+A" (v->counter), "=r" (ret) \ : "r" (mask)); \ return ret; \ } #define ATOMIC_OP_RETURN(op, asm_op, c_op, asm_or, order) \ static __always_inline int atomic_##op##_return##order(int i, atomic_t *v) \ { \ return atomic_fetch_##op##order(i, v) c_op i; \ } #define ATOMIC_OPS(op, asm_op, c_op) \ ATOMIC_OP(op, asm_op, c_op, , _relaxed) \ ATOMIC_OP_RETURN(op, asm_op, c_op, , _relaxed) \ ATOMIC_FETCH_OP(op, asm_op, c_op, , _relaxed) ATOMIC_OPS(add, add, +) ATOMIC_OPS(sub, sub, -) #undef ATOMIC_OPS #define ATOMIC_OPS(op, asm_op, c_op, asm_or, order) \ ATOMIC_OP_RETURN(op, asm_op, c_op, , _relaxed) \ ATOMIC_FETCH_OP(op, asm_op, c_op, , _relaxed) ATOMIC_OPS(add, add, +, ".aq", _acquire) ATOMIC_OPS(add, add, +, ".rl", _release) ATOMIC_OPS(add, add, +, ".aq.rl", ) ATOMIC_OPS(sub, sub, -, ".aq", _acquire) ATOMIC_OPS(sub, sub, -, ".rl", _release) ATOMIC_OPS(sub, sub, -, ".aq.rl", ) #undef ATOMIC_OPS #define ATOMIC_OPS(op, asm_op, c_op) \ ATOMIC_OP(op, asm_op, c_op) \ ATOMIC_FETCH_OP(op, asm_op, c_op, , _relaxed) ATOMIC_OPS(and, and, &) ATOMIC_OPS(or, or, |) ATOMIC_OPS(xor, xor, ^) #undef ATOMIC_OPS ATOMIC_FETCH_OP(and, and, &, ".aq", _acquire) ATOMIC_FETCH_OP(and, and, &, ".rl", _release) ATOMIC_FETCH_OP(and, and, &, ".aq.rl", ) ATOMIC_FETCH_OP(or, or, |, ".aq", _acquire) ATOMIC_FETCH_OP(or, or, |, ".rl", _release) ATOMIC_FETCH_OP(or, or, |, ".aq.rl", ) ATOMIC_FETCH_OP(xor, xor, ^, ".aq", _acquire) ATOMIC_FETCH_OP(xor, xor, ^, ".rl", _release) ATOMIC_FETCH_OP(xor, xor, ^, ".aq.rl", ) #define smp_mb__before_atomic() smp_mb() #define smp_mb__after_atomic() smp_mb() Which (pending the sub confusion) will generate the entire set of: atomic_add, atomic_add_return{_relaxed,_acquire,_release,} atomic_fetch_add{_relaxed,_acquire,_release,} atomic_sub, atomic_sub_return{_relaxed,_acquire,_release,} atomic_fetch_sub{_relaxed,_acquire,_release,} atomic_and, atomic_fetch_and{_relaxed,_acquire,_release,} atomic_or, atomic_fetch_or{_relaxed,_acquire,_release,} atomic_xor, atomic_fetch_xor{_relaxed,_acquire,_release,}