int32x4_t
const int32x4_t zero = vdupq_n_s32(0);
int32x4_t t, t_;
vcltq_s32(int32x4_t __v0, int32x4_t __v1)
static __inline int32x4_t
return (int32x4_t) { __x, __x, __x, __x };
return (uint32_t)__builtin_neon_vget_laneuv4si((int32x4_t)__v, __i);
(uint32_t)__builtin_neon_vgetq_lane_i32((int32x4_t)(__v), \
static __inline int32x4_t
return (int32x4_t)__v;
vreinterpretq_u8_s32(int32x4_t __v)
(uint32x4_t)__builtin_neon_vsetq_lane_i32((__x), (int32x4_t)(__v), \
static __inline int32x4_t
vshlq_n_s32(int32x4_t __v, uint8_t __bits)
return (int32x4_t)__builtin_aarch64_ashlv4si(__v, __bits);
return (int32x4_t)__builtin_neon_vshl_nv4si(__v, __bits);
(int32x4_t)__builtin_neon_vshlq_n_v((int32x4_t)(__v), (__bits), 34)
return (uint32x4_t)__builtin_aarch64_ashlv4si((int32x4_t)__v, __bits);
return (uint32x4_t)__builtin_neon_vshl_nv4si((int32x4_t)__v, __bits);
(uint32x4_t)__builtin_neon_vshlq_n_v((int32x4_t)(__v), (__bits), 50)
return (uint32x4_t)__builtin_aarch64_lshrv4si((int32x4_t)__v, __bits);
return (uint32x4_t)__builtin_neon_vshru_nv4si((int32x4_t)__v, __bits);
(uint32x4_t)__builtin_neon_vshrq_n_v((int32x4_t)(__v), (__bits), 50)
static __inline int32x4_t
vsliq_n_s32(int32x4_t __vins, int32x4_t __vsh, uint8_t __bits)
return (int32x4_t)__builtin_aarch64_ssli_nv4si(__vins, __vsh, __bits);
return (int32x4_t)__builtin_neon_vsli_nv4si(__vins, __vsh, __bits);
(int32x4_t)__builtin_neon_vsliq_n_v((int32x4_t)(__vins), \
(int32x4_t)(__vsh), (__bits), 34)
int32x4_t __tvins = (__vins); \
int32x4_t __tvsh = (__vsh); \
int32x4_t __vins_r = __builtin_shufflevector(__tvins, __tvins, \
int32x4_t __vsh_r = __builtin_shufflevector(__tvsh, __tvsh, \
int32x4_t __r = __builtin_neon_vsliq_n_v(__tvins, __tvsh, __tbits, \
return (uint32x4_t)__builtin_neon_vsri_nv4si((int32x4_t)__vins,
(int32x4_t)__vsh, __bits);
(int32x4_t)__builtin_neon_vsriq_n_v((int32x4_t)(__vins), \
(int32x4_t)(__vsh), (__bits), 34)
int32x4_t __tvins = (__vins); \
int32x4_t __tvsh = (__vsh); \
int32x4_t __vins_r = __builtin_shufflevector(__tvins, __tvins, \
int32x4_t __vsh_r = __builtin_shufflevector(__tvsh, __tvsh, \
int32x4_t __r = __builtin_neon_vsriq_n_v(__tvins, __tvsh, __tbits, \
__builtin_aarch64_st1v4si(__p, (int32x4_t)__v);
__builtin_neon_vst1v4si(__p, (int32x4_t)__v);