#if defined(ARCH_X86)
// avoid +32 for shift optimization (gcc should do that ...)
static inline int32_t NEG_SSR32( int32_t a, int8_t s){
#if defined(ARCH_X86)
// avoid +32 for shift optimization (gcc should do that ...)
static inline int32_t NEG_SSR32( int32_t a, int8_t s){