#ifndef __NEON_IS

#ifdef __ghs__
#pragma ghs startnomisra
#pragma ghs nowarning 1544
#endif

#define __NEON_IS

#include <stdint.h>

#ifdef __LLONG_BIT

#ifdef __cplusplus
extern "C" {
#endif

#ifndef __EDG__
typedef int __ghs_c_int__; /* const int argument only */
#endif


#define __NEONSAFE __attribute__((const,pure))
#define __NEONPURE __attribute__((pure))

#if (!defined(__ghs_alignment) || (__ghs_alignment < 8))
#define __NEONALIGN64 __attribute__((aligned(4)))
#define __NEONALIGN128 __attribute__((aligned(4)))
#else
#define __NEONALIGN64 __attribute__((aligned(8)))
#if (__ghs_alignment < 16)
#define __NEONALIGN128 __attribute__((aligned(8)))
#else
#define __NEONALIGN128 __attribute__((aligned(16)))
#endif
#endif


typedef float float32_t;
typedef double float64_t;
typedef uint8_t poly8_t;
typedef uint16_t poly16_t;

struct __simd64_int8_t { int8_t a[8]; } __NEONALIGN64;
struct __simd64_int16_t { int16_t a[4]; } __NEONALIGN64;
struct __simd64_int32_t { int32_t a[2]; } __NEONALIGN64;
struct __simd64_int64_t { int64_t a[1]; } __NEONALIGN64;
struct __simd64_uint8_t { uint8_t a[8]; } __NEONALIGN64;
struct __simd64_uint16_t { uint16_t a[4]; } __NEONALIGN64;
struct __simd64_uint32_t { uint32_t a[2]; } __NEONALIGN64;
struct __simd64_uint64_t { uint64_t a[1]; } __NEONALIGN64;
struct __simd64_float32_t { float32_t a[2]; } __NEONALIGN64;
struct __simd64_poly8_t { poly8_t a[8]; } __NEONALIGN64;
struct __simd64_poly16_t { poly16_t a[4]; } __NEONALIGN64;
struct __simd128_int8_t { int8_t v[16]; } __NEONALIGN128;
struct __simd128_int16_t { int16_t a[8]; } __NEONALIGN128;
struct __simd128_int32_t { int32_t a[4]; } __NEONALIGN128;
struct __simd128_int64_t { int64_t a[2]; } __NEONALIGN128;
struct __simd128_uint8_t { uint8_t a[16]; } __NEONALIGN128;
struct __simd128_uint16_t { uint16_t a[8]; } __NEONALIGN128;
struct __simd128_uint32_t { uint32_t a[4]; } __NEONALIGN128;
struct __simd128_uint64_t { uint64_t a[2]; } __NEONALIGN128;
struct __simd128_float32_t { float32_t a[4]; } __NEONALIGN128;
struct __simd128_float64_t { float64_t a[2]; } __NEONALIGN128;
struct __simd128_poly8_t { poly8_t a[16]; } __NEONALIGN128;
struct __simd128_poly16_t { poly16_t a[8]; } __NEONALIGN128;

typedef struct __simd64_int8_t int8x8_t;
typedef struct __simd64_int16_t int16x4_t;
typedef struct __simd64_int32_t int32x2_t;
typedef struct __simd64_int64_t int64x1_t;
typedef struct __simd64_uint8_t uint8x8_t;
typedef struct __simd64_uint16_t uint16x4_t;
typedef struct __simd64_uint32_t uint32x2_t;
typedef struct __simd64_uint64_t uint64x1_t;
typedef struct __simd64_float32_t float32x2_t;
typedef struct __simd64_poly8_t poly8x8_t;
typedef struct __simd64_poly16_t poly16x4_t;
typedef struct __simd128_int8_t int8x16_t;
typedef struct __simd128_int16_t int16x8_t;
typedef struct __simd128_int32_t int32x4_t;
typedef struct __simd128_int64_t int64x2_t;
typedef struct __simd128_uint8_t uint8x16_t;
typedef struct __simd128_uint16_t uint16x8_t;
typedef struct __simd128_uint32_t uint32x4_t;
typedef struct __simd128_uint64_t uint64x2_t;
typedef struct __simd128_float32_t float32x4_t;
typedef struct __simd128_float64_t float64x2_t;
typedef struct __simd128_poly8_t poly8x16_t;
typedef struct __simd128_poly16_t poly16x8_t;
typedef float64_t float64x1_t;


struct __simd_array2_int8x8_t { struct __simd64_int8_t val[2]; };
struct __simd_array3_int8x8_t { struct __simd64_int8_t val[3]; };
struct __simd_array4_int8x8_t { struct __simd64_int8_t val[4]; };
struct __simd_array2_int8x16_t { struct __simd128_int8_t val[2]; };
struct __simd_array3_int8x16_t { struct __simd128_int8_t val[3]; };
struct __simd_array4_int8x16_t { struct __simd128_int8_t val[4]; };
struct __simd_array2_int16x4_t { struct __simd64_int16_t val[2]; };
struct __simd_array3_int16x4_t { struct __simd64_int16_t val[3]; };
struct __simd_array4_int16x4_t { struct __simd64_int16_t val[4]; };
struct __simd_array2_int16x8_t { struct __simd128_int16_t val[2]; };
struct __simd_array3_int16x8_t { struct __simd128_int16_t val[3]; };
struct __simd_array4_int16x8_t { struct __simd128_int16_t val[4]; };
struct __simd_array2_int32x2_t { struct __simd64_int32_t val[2]; };
struct __simd_array3_int32x2_t { struct __simd64_int32_t val[3]; };
struct __simd_array4_int32x2_t { struct __simd64_int32_t val[4]; };
struct __simd_array2_int32x4_t { struct __simd128_int32_t val[2]; };
struct __simd_array3_int32x4_t { struct __simd128_int32_t val[3]; };
struct __simd_array4_int32x4_t { struct __simd128_int32_t val[4]; };
struct __simd_array2_int64x1_t { struct __simd64_int64_t val[2]; };
struct __simd_array3_int64x1_t { struct __simd64_int64_t val[3]; };
struct __simd_array4_int64x1_t { struct __simd64_int64_t val[4]; };
struct __simd_array2_int64x2_t { struct __simd128_int64_t val[2]; };
struct __simd_array3_int64x2_t { struct __simd128_int64_t val[3]; };
struct __simd_array4_int64x2_t { struct __simd128_int64_t val[4]; };
struct __simd_array2_uint8x8_t { struct __simd64_uint8_t val[2]; };
struct __simd_array3_uint8x8_t { struct __simd64_uint8_t val[3]; };
struct __simd_array4_uint8x8_t { struct __simd64_uint8_t val[4]; };
struct __simd_array2_uint8x16_t { struct __simd128_uint8_t val[2]; };
struct __simd_array3_uint8x16_t { struct __simd128_uint8_t val[3]; };
struct __simd_array4_uint8x16_t { struct __simd128_uint8_t val[4]; };
struct __simd_array2_uint16x4_t { struct __simd64_uint16_t val[2]; };
struct __simd_array3_uint16x4_t { struct __simd64_uint16_t val[3]; };
struct __simd_array4_uint16x4_t { struct __simd64_uint16_t val[4]; };
struct __simd_array2_uint16x8_t { struct __simd128_uint16_t val[2]; };
struct __simd_array3_uint16x8_t { struct __simd128_uint16_t val[3]; };
struct __simd_array4_uint16x8_t { struct __simd128_uint16_t val[4]; };
struct __simd_array2_uint32x2_t { struct __simd64_uint32_t val[2]; };
struct __simd_array3_uint32x2_t { struct __simd64_uint32_t val[3]; };
struct __simd_array4_uint32x2_t { struct __simd64_uint32_t val[4]; };
struct __simd_array2_uint32x4_t { struct __simd128_uint32_t val[2]; };
struct __simd_array3_uint32x4_t { struct __simd128_uint32_t val[3]; };
struct __simd_array4_uint32x4_t { struct __simd128_uint32_t val[4]; };
struct __simd_array2_uint64x1_t { struct __simd64_uint64_t val[2]; };
struct __simd_array3_uint64x1_t { struct __simd64_uint64_t val[3]; };
struct __simd_array4_uint64x1_t { struct __simd64_uint64_t val[4]; };
struct __simd_array2_uint64x2_t { struct __simd128_uint64_t val[2]; };
struct __simd_array3_uint64x2_t { struct __simd128_uint64_t val[3]; };
struct __simd_array4_uint64x2_t { struct __simd128_uint64_t val[4]; };
struct __simd_array2_float32x2_t { struct __simd64_float32_t val[2]; };
struct __simd_array3_float32x2_t { struct __simd64_float32_t val[3]; };
struct __simd_array4_float32x2_t { struct __simd64_float32_t val[4]; };
struct __simd_array2_float32x4_t { struct __simd128_float32_t val[2]; };
struct __simd_array3_float32x4_t { struct __simd128_float32_t val[3]; };
struct __simd_array4_float32x4_t { struct __simd128_float32_t val[4]; };
struct __simd_array2_float64x2_t { struct __simd128_float64_t val[2]; };
struct __simd_array3_float64x2_t { struct __simd128_float64_t val[3]; };
struct __simd_array4_float64x2_t { struct __simd128_float64_t val[4]; };
struct __simd_array2_poly8x8_t { struct __simd64_poly8_t val[2]; };
struct __simd_array3_poly8x8_t { struct __simd64_poly8_t val[3]; };
struct __simd_array4_poly8x8_t { struct __simd64_poly8_t val[4]; };
struct __simd_array2_poly8x16_t { struct __simd128_poly8_t val[2]; };
struct __simd_array3_poly8x16_t { struct __simd128_poly8_t val[3]; };
struct __simd_array4_poly8x16_t { struct __simd128_poly8_t val[4]; };
struct __simd_array2_poly16x4_t { struct __simd64_poly16_t val[2]; };
struct __simd_array3_poly16x4_t { struct __simd64_poly16_t val[3]; };
struct __simd_array4_poly16x4_t { struct __simd64_poly16_t val[4]; };
struct __simd_array2_poly16x8_t { struct __simd128_poly16_t val[2]; };
struct __simd_array3_poly16x8_t { struct __simd128_poly16_t val[3]; };
struct __simd_array4_poly16x8_t { struct __simd128_poly16_t val[4]; };

typedef struct __simd_array2_int8x8_t int8x8x2_t;
typedef struct __simd_array3_int8x8_t int8x8x3_t;
typedef struct __simd_array4_int8x8_t int8x8x4_t;
typedef struct __simd_array2_int8x16_t int8x16x2_t;
typedef struct __simd_array3_int8x16_t int8x16x3_t;
typedef struct __simd_array4_int8x16_t int8x16x4_t;
typedef struct __simd_array2_int16x4_t int16x4x2_t;
typedef struct __simd_array3_int16x4_t int16x4x3_t;
typedef struct __simd_array4_int16x4_t int16x4x4_t;
typedef struct __simd_array2_int16x8_t int16x8x2_t;
typedef struct __simd_array3_int16x8_t int16x8x3_t;
typedef struct __simd_array4_int16x8_t int16x8x4_t;
typedef struct __simd_array2_int32x2_t int32x2x2_t;
typedef struct __simd_array3_int32x2_t int32x2x3_t;
typedef struct __simd_array4_int32x2_t int32x2x4_t;
typedef struct __simd_array2_int32x4_t int32x4x2_t;
typedef struct __simd_array3_int32x4_t int32x4x3_t;
typedef struct __simd_array4_int32x4_t int32x4x4_t;
typedef struct __simd_array2_int64x1_t int64x1x2_t;
typedef struct __simd_array3_int64x1_t int64x1x3_t;
typedef struct __simd_array4_int64x1_t int64x1x4_t;
typedef struct __simd_array2_int64x2_t int64x2x2_t;
typedef struct __simd_array3_int64x2_t int64x2x3_t;
typedef struct __simd_array4_int64x2_t int64x2x4_t;
typedef struct __simd_array2_uint8x8_t uint8x8x2_t;
typedef struct __simd_array3_uint8x8_t uint8x8x3_t;
typedef struct __simd_array4_uint8x8_t uint8x8x4_t;
typedef struct __simd_array2_uint8x16_t uint8x16x2_t;
typedef struct __simd_array3_uint8x16_t uint8x16x3_t;
typedef struct __simd_array4_uint8x16_t uint8x16x4_t;
typedef struct __simd_array2_uint16x4_t uint16x4x2_t;
typedef struct __simd_array3_uint16x4_t uint16x4x3_t;
typedef struct __simd_array4_uint16x4_t uint16x4x4_t;
typedef struct __simd_array2_uint16x8_t uint16x8x2_t;
typedef struct __simd_array3_uint16x8_t uint16x8x3_t;
typedef struct __simd_array4_uint16x8_t uint16x8x4_t;
typedef struct __simd_array2_uint32x2_t uint32x2x2_t;
typedef struct __simd_array3_uint32x2_t uint32x2x3_t;
typedef struct __simd_array4_uint32x2_t uint32x2x4_t;
typedef struct __simd_array2_uint32x4_t uint32x4x2_t;
typedef struct __simd_array3_uint32x4_t uint32x4x3_t;
typedef struct __simd_array4_uint32x4_t uint32x4x4_t;
typedef struct __simd_array2_uint64x1_t uint64x1x2_t;
typedef struct __simd_array3_uint64x1_t uint64x1x3_t;
typedef struct __simd_array4_uint64x1_t uint64x1x4_t;
typedef struct __simd_array2_uint64x2_t uint64x2x2_t;
typedef struct __simd_array3_uint64x2_t uint64x2x3_t;
typedef struct __simd_array4_uint64x2_t uint64x2x4_t;
typedef struct __simd_array2_float32x2_t float32x2x2_t;
typedef struct __simd_array3_float32x2_t float32x2x3_t;
typedef struct __simd_array4_float32x2_t float32x2x4_t;
typedef struct __simd_array2_float32x4_t float32x4x2_t;
typedef struct __simd_array3_float32x4_t float32x4x3_t;
typedef struct __simd_array4_float32x4_t float32x4x4_t;
typedef struct __simd_array2_float64x2_t float64x2x2_t;
typedef struct __simd_array3_float64x2_t float64x2x3_t;
typedef struct __simd_array4_float64x2_t float64x2x4_t;
typedef struct __simd_array2_poly8x8_t poly8x8x2_t;
typedef struct __simd_array3_poly8x8_t poly8x8x3_t;
typedef struct __simd_array4_poly8x8_t poly8x8x4_t;
typedef struct __simd_array2_poly8x16_t poly8x16x2_t;
typedef struct __simd_array3_poly8x16_t poly8x16x3_t;
typedef struct __simd_array4_poly8x16_t poly8x16x4_t;
typedef struct __simd_array2_poly16x4_t poly16x4x2_t;
typedef struct __simd_array3_poly16x4_t poly16x4x3_t;
typedef struct __simd_array4_poly16x4_t poly16x4x4_t;
typedef struct __simd_array2_poly16x8_t poly16x8x2_t;
typedef struct __simd_array3_poly16x8_t poly16x8x3_t;
typedef struct __simd_array4_poly16x8_t poly16x8x4_t;


#ifdef __cplusplus
/* Detect if this header is included from within a user defined namespace. */
typedef int __MUST_INCLUDE_NEON_HEADER_AT_GLOBAL_NAMESPACE_SCOPE[sizeof(::__simd64_int8_t)];
#endif


uint8x8_t __simd_vld1_u8(uint8_t const * a) __NEONPURE;
int8x8_t __simd_vld1_s8(int8_t const * a) __NEONPURE;
poly8x8_t __simd_vld1_p8(poly8_t const * a) __NEONPURE;
uint16x4_t __simd_vld1_u16(uint16_t const * a) __NEONPURE;
int16x4_t __simd_vld1_s16(int16_t const * a) __NEONPURE;
poly16x4_t __simd_vld1_p16(poly16_t const * a) __NEONPURE;
uint32x2_t __simd_vld1_u32(uint32_t const * a) __NEONPURE;
int32x2_t __simd_vld1_s32(int32_t const * a) __NEONPURE;
float32x2_t __simd_vld1_f32(float32_t const * a) __NEONPURE;
uint64x1_t __simd_vld1_u64(uint64_t const * a) __NEONPURE;
int64x1_t __simd_vld1_s64(int64_t const * a) __NEONPURE;
uint8x16_t __simd_vld1q_u8(uint8_t const * a) __NEONPURE;
int8x16_t __simd_vld1q_s8(int8_t const * a) __NEONPURE;
poly8x16_t __simd_vld1q_p8(poly8_t const * a) __NEONPURE;
uint16x8_t __simd_vld1q_u16(uint16_t const * a) __NEONPURE;
int16x8_t __simd_vld1q_s16(int16_t const * a) __NEONPURE;
poly16x8_t __simd_vld1q_p16(poly16_t const * a) __NEONPURE;
uint32x4_t __simd_vld1q_u32(uint32_t const * a) __NEONPURE;
int32x4_t __simd_vld1q_s32(int32_t const * a) __NEONPURE;
float32x4_t __simd_vld1q_f32(float32_t const * a) __NEONPURE;
uint64x2_t __simd_vld1q_u64(uint64_t const * a) __NEONPURE;
int64x2_t __simd_vld1q_s64(int64_t const * a) __NEONPURE;
float64x2_t __simd_vld1q_f64(float64_t const * a) __NEONPURE;
uint8x8x2_t __simd_vld1_u8_x2(uint8_t const * a) __NEONPURE;
int8x8x2_t __simd_vld1_s8_x2(int8_t const * a) __NEONPURE;
poly8x8x2_t __simd_vld1_p8_x2(poly8_t const * a) __NEONPURE;
uint16x4x2_t __simd_vld1_u16_x2(uint16_t const * a) __NEONPURE;
int16x4x2_t __simd_vld1_s16_x2(int16_t const * a) __NEONPURE;
poly16x4x2_t __simd_vld1_p16_x2(poly16_t const * a) __NEONPURE;
uint32x2x2_t __simd_vld1_u32_x2(uint32_t const * a) __NEONPURE;
int32x2x2_t __simd_vld1_s32_x2(int32_t const * a) __NEONPURE;
float32x2x2_t __simd_vld1_f32_x2(float32_t const * a) __NEONPURE;
uint64x1x2_t __simd_vld1_u64_x2(uint64_t const * a) __NEONPURE;
int64x1x2_t __simd_vld1_s64_x2(int64_t const * a) __NEONPURE;
uint8x16x2_t __simd_vld1q_u8_x2(uint8_t const * a) __NEONPURE;
int8x16x2_t __simd_vld1q_s8_x2(int8_t const * a) __NEONPURE;
poly8x16x2_t __simd_vld1q_p8_x2(poly8_t const * a) __NEONPURE;
uint16x8x2_t __simd_vld1q_u16_x2(uint16_t const * a) __NEONPURE;
int16x8x2_t __simd_vld1q_s16_x2(int16_t const * a) __NEONPURE;
poly16x8x2_t __simd_vld1q_p16_x2(poly16_t const * a) __NEONPURE;
uint32x4x2_t __simd_vld1q_u32_x2(uint32_t const * a) __NEONPURE;
int32x4x2_t __simd_vld1q_s32_x2(int32_t const * a) __NEONPURE;
float32x4x2_t __simd_vld1q_f32_x2(float32_t const * a) __NEONPURE;
uint64x2x2_t __simd_vld1q_u64_x2(uint64_t const * a) __NEONPURE;
int64x2x2_t __simd_vld1q_s64_x2(int64_t const * a) __NEONPURE;
float64x2x2_t __simd_vld1q_f64_x2(float64_t const * a) __NEONPURE;
uint8x8x3_t __simd_vld1_u8_x3(uint8_t const * a) __NEONPURE;
int8x8x3_t __simd_vld1_s8_x3(int8_t const * a) __NEONPURE;
poly8x8x3_t __simd_vld1_p8_x3(poly8_t const * a) __NEONPURE;
uint16x4x3_t __simd_vld1_u16_x3(uint16_t const * a) __NEONPURE;
int16x4x3_t __simd_vld1_s16_x3(int16_t const * a) __NEONPURE;
poly16x4x3_t __simd_vld1_p16_x3(poly16_t const * a) __NEONPURE;
uint32x2x3_t __simd_vld1_u32_x3(uint32_t const * a) __NEONPURE;
int32x2x3_t __simd_vld1_s32_x3(int32_t const * a) __NEONPURE;
float32x2x3_t __simd_vld1_f32_x3(float32_t const * a) __NEONPURE;
uint64x1x3_t __simd_vld1_u64_x3(uint64_t const * a) __NEONPURE;
int64x1x3_t __simd_vld1_s64_x3(int64_t const * a) __NEONPURE;
uint8x16x3_t __simd_vld1q_u8_x3(uint8_t const * a) __NEONPURE;
int8x16x3_t __simd_vld1q_s8_x3(int8_t const * a) __NEONPURE;
poly8x16x3_t __simd_vld1q_p8_x3(poly8_t const * a) __NEONPURE;
uint16x8x3_t __simd_vld1q_u16_x3(uint16_t const * a) __NEONPURE;
int16x8x3_t __simd_vld1q_s16_x3(int16_t const * a) __NEONPURE;
poly16x8x3_t __simd_vld1q_p16_x3(poly16_t const * a) __NEONPURE;
uint32x4x3_t __simd_vld1q_u32_x3(uint32_t const * a) __NEONPURE;
int32x4x3_t __simd_vld1q_s32_x3(int32_t const * a) __NEONPURE;
float32x4x3_t __simd_vld1q_f32_x3(float32_t const * a) __NEONPURE;
uint64x2x3_t __simd_vld1q_u64_x3(uint64_t const * a) __NEONPURE;
int64x2x3_t __simd_vld1q_s64_x3(int64_t const * a) __NEONPURE;
float64x2x3_t __simd_vld1q_f64_x3(float64_t const * a) __NEONPURE;
uint8x8x4_t __simd_vld1_u8_x4(uint8_t const * a) __NEONPURE;
int8x8x4_t __simd_vld1_s8_x4(int8_t const * a) __NEONPURE;
poly8x8x4_t __simd_vld1_p8_x4(poly8_t const * a) __NEONPURE;
uint16x4x4_t __simd_vld1_u16_x4(uint16_t const * a) __NEONPURE;
int16x4x4_t __simd_vld1_s16_x4(int16_t const * a) __NEONPURE;
poly16x4x4_t __simd_vld1_p16_x4(poly16_t const * a) __NEONPURE;
uint32x2x4_t __simd_vld1_u32_x4(uint32_t const * a) __NEONPURE;
int32x2x4_t __simd_vld1_s32_x4(int32_t const * a) __NEONPURE;
float32x2x4_t __simd_vld1_f32_x4(float32_t const * a) __NEONPURE;
uint64x1x4_t __simd_vld1_u64_x4(uint64_t const * a) __NEONPURE;
int64x1x4_t __simd_vld1_s64_x4(int64_t const * a) __NEONPURE;
uint8x16x4_t __simd_vld1q_u8_x4(uint8_t const * a) __NEONPURE;
int8x16x4_t __simd_vld1q_s8_x4(int8_t const * a) __NEONPURE;
poly8x16x4_t __simd_vld1q_p8_x4(poly8_t const * a) __NEONPURE;
uint16x8x4_t __simd_vld1q_u16_x4(uint16_t const * a) __NEONPURE;
int16x8x4_t __simd_vld1q_s16_x4(int16_t const * a) __NEONPURE;
poly16x8x4_t __simd_vld1q_p16_x4(poly16_t const * a) __NEONPURE;
uint32x4x4_t __simd_vld1q_u32_x4(uint32_t const * a) __NEONPURE;
int32x4x4_t __simd_vld1q_s32_x4(int32_t const * a) __NEONPURE;
float32x4x4_t __simd_vld1q_f32_x4(float32_t const * a) __NEONPURE;
uint64x2x4_t __simd_vld1q_u64_x4(uint64_t const * a) __NEONPURE;
int64x2x4_t __simd_vld1q_s64_x4(int64_t const * a) __NEONPURE;
float64x2x4_t __simd_vld1q_f64_x4(float64_t const * a) __NEONPURE;
uint8x8x2_t __simd_vld2_u8(uint8_t const * a) __NEONPURE;
int8x8x2_t __simd_vld2_s8(int8_t const * a) __NEONPURE;
poly8x8x2_t __simd_vld2_p8(poly8_t const * a) __NEONPURE;
uint16x4x2_t __simd_vld2_u16(uint16_t const * a) __NEONPURE;
int16x4x2_t __simd_vld2_s16(int16_t const * a) __NEONPURE;
poly16x4x2_t __simd_vld2_p16(poly16_t const * a) __NEONPURE;
uint32x2x2_t __simd_vld2_u32(uint32_t const * a) __NEONPURE;
int32x2x2_t __simd_vld2_s32(int32_t const * a) __NEONPURE;
float32x2x2_t __simd_vld2_f32(float32_t const * a) __NEONPURE;
uint8x16x2_t __simd_vld2q_u8(uint8_t const * a) __NEONPURE;
int8x16x2_t __simd_vld2q_s8(int8_t const * a) __NEONPURE;
poly8x16x2_t __simd_vld2q_p8(poly8_t const * a) __NEONPURE;
uint16x8x2_t __simd_vld2q_u16(uint16_t const * a) __NEONPURE;
int16x8x2_t __simd_vld2q_s16(int16_t const * a) __NEONPURE;
poly16x8x2_t __simd_vld2q_p16(poly16_t const * a) __NEONPURE;
uint32x4x2_t __simd_vld2q_u32(uint32_t const * a) __NEONPURE;
int32x4x2_t __simd_vld2q_s32(int32_t const * a) __NEONPURE;
float32x4x2_t __simd_vld2q_f32(float32_t const * a) __NEONPURE;
uint64x2x2_t __simd_vld2q_u64(uint64_t const * a) __NEONPURE;
int64x2x2_t __simd_vld2q_s64(int64_t const * a) __NEONPURE;
float64x2x2_t __simd_vld2q_f64(float64_t const * a) __NEONPURE;
uint8x8x3_t __simd_vld3_u8(uint8_t const * a) __NEONPURE;
int8x8x3_t __simd_vld3_s8(int8_t const * a) __NEONPURE;
poly8x8x3_t __simd_vld3_p8(poly8_t const * a) __NEONPURE;
uint16x4x3_t __simd_vld3_u16(uint16_t const * a) __NEONPURE;
int16x4x3_t __simd_vld3_s16(int16_t const * a) __NEONPURE;
poly16x4x3_t __simd_vld3_p16(poly16_t const * a) __NEONPURE;
uint32x2x3_t __simd_vld3_u32(uint32_t const * a) __NEONPURE;
int32x2x3_t __simd_vld3_s32(int32_t const * a) __NEONPURE;
float32x2x3_t __simd_vld3_f32(float32_t const * a) __NEONPURE;
uint8x16x3_t __simd_vld3q_u8(uint8_t const * a) __NEONPURE;
int8x16x3_t __simd_vld3q_s8(int8_t const * a) __NEONPURE;
poly8x16x3_t __simd_vld3q_p8(poly8_t const * a) __NEONPURE;
uint16x8x3_t __simd_vld3q_u16(uint16_t const * a) __NEONPURE;
int16x8x3_t __simd_vld3q_s16(int16_t const * a) __NEONPURE;
poly16x8x3_t __simd_vld3q_p16(poly16_t const * a) __NEONPURE;
uint32x4x3_t __simd_vld3q_u32(uint32_t const * a) __NEONPURE;
int32x4x3_t __simd_vld3q_s32(int32_t const * a) __NEONPURE;
float32x4x3_t __simd_vld3q_f32(float32_t const * a) __NEONPURE;
uint64x2x3_t __simd_vld3q_u64(uint64_t const * a) __NEONPURE;
int64x2x3_t __simd_vld3q_s64(int64_t const * a) __NEONPURE;
float64x2x3_t __simd_vld3q_f64(float64_t const * a) __NEONPURE;
uint8x8x4_t __simd_vld4_u8(uint8_t const * a) __NEONPURE;
int8x8x4_t __simd_vld4_s8(int8_t const * a) __NEONPURE;
poly8x8x4_t __simd_vld4_p8(poly8_t const * a) __NEONPURE;
uint16x4x4_t __simd_vld4_u16(uint16_t const * a) __NEONPURE;
int16x4x4_t __simd_vld4_s16(int16_t const * a) __NEONPURE;
poly16x4x4_t __simd_vld4_p16(poly16_t const * a) __NEONPURE;
uint32x2x4_t __simd_vld4_u32(uint32_t const * a) __NEONPURE;
int32x2x4_t __simd_vld4_s32(int32_t const * a) __NEONPURE;
float32x2x4_t __simd_vld4_f32(float32_t const * a) __NEONPURE;
uint8x16x4_t __simd_vld4q_u8(uint8_t const * a) __NEONPURE;
int8x16x4_t __simd_vld4q_s8(int8_t const * a) __NEONPURE;
poly8x16x4_t __simd_vld4q_p8(poly8_t const * a) __NEONPURE;
uint16x8x4_t __simd_vld4q_u16(uint16_t const * a) __NEONPURE;
int16x8x4_t __simd_vld4q_s16(int16_t const * a) __NEONPURE;
poly16x8x4_t __simd_vld4q_p16(poly16_t const * a) __NEONPURE;
uint32x4x4_t __simd_vld4q_u32(uint32_t const * a) __NEONPURE;
int32x4x4_t __simd_vld4q_s32(int32_t const * a) __NEONPURE;
float32x4x4_t __simd_vld4q_f32(float32_t const * a) __NEONPURE;
uint64x2x4_t __simd_vld4q_u64(uint64_t const * a) __NEONPURE;
int64x2x4_t __simd_vld4q_s64(int64_t const * a) __NEONPURE;
float64x2x4_t __simd_vld4q_f64(float64_t const * a) __NEONPURE;
void __simd_vst1_u8(uint8_t * a, uint8x8_t b);
void __simd_vst1_s8(int8_t * a, int8x8_t b);
void __simd_vst1_p8(poly8_t * a, poly8x8_t b);
void __simd_vst1_u16(uint16_t * a, uint16x4_t b);
void __simd_vst1_s16(int16_t * a, int16x4_t b);
void __simd_vst1_p16(poly16_t * a, poly16x4_t b);
void __simd_vst1_u32(uint32_t * a, uint32x2_t b);
void __simd_vst1_s32(int32_t * a, int32x2_t b);
void __simd_vst1_f32(float32_t * a, float32x2_t b);
void __simd_vst1_u64(uint64_t * a, uint64x1_t b);
void __simd_vst1_s64(int64_t * a, int64x1_t b);
void __simd_vst1q_u8(uint8_t * a, uint8x16_t b);
void __simd_vst1q_s8(int8_t * a, int8x16_t b);
void __simd_vst1q_p8(poly8_t * a, poly8x16_t b);
void __simd_vst1q_u16(uint16_t * a, uint16x8_t b);
void __simd_vst1q_s16(int16_t * a, int16x8_t b);
void __simd_vst1q_p16(poly16_t * a, poly16x8_t b);
void __simd_vst1q_u32(uint32_t * a, uint32x4_t b);
void __simd_vst1q_s32(int32_t * a, int32x4_t b);
void __simd_vst1q_f32(float32_t * a, float32x4_t b);
void __simd_vst1q_u64(uint64_t * a, uint64x2_t b);
void __simd_vst1q_s64(int64_t * a, int64x2_t b);
void __simd_vst1q_f64(float64_t * a, float64x2_t b);
void __simd_vst1_u8_x2(uint8_t * a, uint8x8x2_t b);
void __simd_vst1_s8_x2(int8_t * a, int8x8x2_t b);
void __simd_vst1_p8_x2(poly8_t * a, poly8x8x2_t b);
void __simd_vst1_u16_x2(uint16_t * a, uint16x4x2_t b);
void __simd_vst1_s16_x2(int16_t * a, int16x4x2_t b);
void __simd_vst1_p16_x2(poly16_t * a, poly16x4x2_t b);
void __simd_vst1_u32_x2(uint32_t * a, uint32x2x2_t b);
void __simd_vst1_s32_x2(int32_t * a, int32x2x2_t b);
void __simd_vst1_f32_x2(float32_t * a, float32x2x2_t b);
void __simd_vst1_u64_x2(uint64_t * a, uint64x1x2_t b);
void __simd_vst1_s64_x2(int64_t * a, int64x1x2_t b);
void __simd_vst1q_u8_x2(uint8_t * a, uint8x16x2_t b);
void __simd_vst1q_s8_x2(int8_t * a, int8x16x2_t b);
void __simd_vst1q_p8_x2(poly8_t * a, poly8x16x2_t b);
void __simd_vst1q_u16_x2(uint16_t * a, uint16x8x2_t b);
void __simd_vst1q_s16_x2(int16_t * a, int16x8x2_t b);
void __simd_vst1q_p16_x2(poly16_t * a, poly16x8x2_t b);
void __simd_vst1q_u32_x2(uint32_t * a, uint32x4x2_t b);
void __simd_vst1q_s32_x2(int32_t * a, int32x4x2_t b);
void __simd_vst1q_f32_x2(float32_t * a, float32x4x2_t b);
void __simd_vst1q_u64_x2(uint64_t * a, uint64x2x2_t b);
void __simd_vst1q_s64_x2(int64_t * a, int64x2x2_t b);
void __simd_vst1q_f64_x2(float64_t * a, float64x2x2_t b);
void __simd_vst1_u8_x3(uint8_t * a, uint8x8x3_t b);
void __simd_vst1_s8_x3(int8_t * a, int8x8x3_t b);
void __simd_vst1_p8_x3(poly8_t * a, poly8x8x3_t b);
void __simd_vst1_u16_x3(uint16_t * a, uint16x4x3_t b);
void __simd_vst1_s16_x3(int16_t * a, int16x4x3_t b);
void __simd_vst1_p16_x3(poly16_t * a, poly16x4x3_t b);
void __simd_vst1_u32_x3(uint32_t * a, uint32x2x3_t b);
void __simd_vst1_s32_x3(int32_t * a, int32x2x3_t b);
void __simd_vst1_f32_x3(float32_t * a, float32x2x3_t b);
void __simd_vst1_u64_x3(uint64_t * a, uint64x1x3_t b);
void __simd_vst1_s64_x3(int64_t * a, int64x1x3_t b);
void __simd_vst1q_u8_x3(uint8_t * a, uint8x16x3_t b);
void __simd_vst1q_s8_x3(int8_t * a, int8x16x3_t b);
void __simd_vst1q_p8_x3(poly8_t * a, poly8x16x3_t b);
void __simd_vst1q_u16_x3(uint16_t * a, uint16x8x3_t b);
void __simd_vst1q_s16_x3(int16_t * a, int16x8x3_t b);
void __simd_vst1q_p16_x3(poly16_t * a, poly16x8x3_t b);
void __simd_vst1q_u32_x3(uint32_t * a, uint32x4x3_t b);
void __simd_vst1q_s32_x3(int32_t * a, int32x4x3_t b);
void __simd_vst1q_f32_x3(float32_t * a, float32x4x3_t b);
void __simd_vst1q_u64_x3(uint64_t * a, uint64x2x3_t b);
void __simd_vst1q_s64_x3(int64_t * a, int64x2x3_t b);
void __simd_vst1q_f64_x3(float64_t * a, float64x2x3_t b);
void __simd_vst1_u8_x4(uint8_t * a, uint8x8x4_t b);
void __simd_vst1_s8_x4(int8_t * a, int8x8x4_t b);
void __simd_vst1_p8_x4(poly8_t * a, poly8x8x4_t b);
void __simd_vst1_u16_x4(uint16_t * a, uint16x4x4_t b);
void __simd_vst1_s16_x4(int16_t * a, int16x4x4_t b);
void __simd_vst1_p16_x4(poly16_t * a, poly16x4x4_t b);
void __simd_vst1_u32_x4(uint32_t * a, uint32x2x4_t b);
void __simd_vst1_s32_x4(int32_t * a, int32x2x4_t b);
void __simd_vst1_f32_x4(float32_t * a, float32x2x4_t b);
void __simd_vst1_u64_x4(uint64_t * a, uint64x1x4_t b);
void __simd_vst1_s64_x4(int64_t * a, int64x1x4_t b);
void __simd_vst1q_u8_x4(uint8_t * a, uint8x16x4_t b);
void __simd_vst1q_s8_x4(int8_t * a, int8x16x4_t b);
void __simd_vst1q_p8_x4(poly8_t * a, poly8x16x4_t b);
void __simd_vst1q_u16_x4(uint16_t * a, uint16x8x4_t b);
void __simd_vst1q_s16_x4(int16_t * a, int16x8x4_t b);
void __simd_vst1q_p16_x4(poly16_t * a, poly16x8x4_t b);
void __simd_vst1q_u32_x4(uint32_t * a, uint32x4x4_t b);
void __simd_vst1q_s32_x4(int32_t * a, int32x4x4_t b);
void __simd_vst1q_f32_x4(float32_t * a, float32x4x4_t b);
void __simd_vst1q_u64_x4(uint64_t * a, uint64x2x4_t b);
void __simd_vst1q_s64_x4(int64_t * a, int64x2x4_t b);
void __simd_vst1q_f64_x4(float64_t * a, float64x2x4_t b);
void __simd_vst2_u8(uint8_t * a, uint8x8x2_t b);
void __simd_vst2_s8(int8_t * a, int8x8x2_t b);
void __simd_vst2_p8(poly8_t * a, poly8x8x2_t b);
void __simd_vst2_u16(uint16_t * a, uint16x4x2_t b);
void __simd_vst2_s16(int16_t * a, int16x4x2_t b);
void __simd_vst2_p16(poly16_t * a, poly16x4x2_t b);
void __simd_vst2_u32(uint32_t * a, uint32x2x2_t b);
void __simd_vst2_s32(int32_t * a, int32x2x2_t b);
void __simd_vst2_f32(float32_t * a, float32x2x2_t b);
void __simd_vst2q_u8(uint8_t * a, uint8x16x2_t b);
void __simd_vst2q_s8(int8_t * a, int8x16x2_t b);
void __simd_vst2q_p8(poly8_t * a, poly8x16x2_t b);
void __simd_vst2q_u16(uint16_t * a, uint16x8x2_t b);
void __simd_vst2q_s16(int16_t * a, int16x8x2_t b);
void __simd_vst2q_p16(poly16_t * a, poly16x8x2_t b);
void __simd_vst2q_u32(uint32_t * a, uint32x4x2_t b);
void __simd_vst2q_s32(int32_t * a, int32x4x2_t b);
void __simd_vst2q_f32(float32_t * a, float32x4x2_t b);
void __simd_vst2q_u64(uint64_t * a, uint64x2x2_t b);
void __simd_vst2q_s64(int64_t * a, int64x2x2_t b);
void __simd_vst2q_f64(float64_t * a, float64x2x2_t b);
void __simd_vst3_u8(uint8_t * a, uint8x8x3_t b);
void __simd_vst3_s8(int8_t * a, int8x8x3_t b);
void __simd_vst3_p8(poly8_t * a, poly8x8x3_t b);
void __simd_vst3_u16(uint16_t * a, uint16x4x3_t b);
void __simd_vst3_s16(int16_t * a, int16x4x3_t b);
void __simd_vst3_p16(poly16_t * a, poly16x4x3_t b);
void __simd_vst3_u32(uint32_t * a, uint32x2x3_t b);
void __simd_vst3_s32(int32_t * a, int32x2x3_t b);
void __simd_vst3_f32(float32_t * a, float32x2x3_t b);
void __simd_vst3q_u8(uint8_t * a, uint8x16x3_t b);
void __simd_vst3q_s8(int8_t * a, int8x16x3_t b);
void __simd_vst3q_p8(poly8_t * a, poly8x16x3_t b);
void __simd_vst3q_u16(uint16_t * a, uint16x8x3_t b);
void __simd_vst3q_s16(int16_t * a, int16x8x3_t b);
void __simd_vst3q_p16(poly16_t * a, poly16x8x3_t b);
void __simd_vst3q_u32(uint32_t * a, uint32x4x3_t b);
void __simd_vst3q_s32(int32_t * a, int32x4x3_t b);
void __simd_vst3q_f32(float32_t * a, float32x4x3_t b);
void __simd_vst3q_u64(uint64_t * a, uint64x2x3_t b);
void __simd_vst3q_s64(int64_t * a, int64x2x3_t b);
void __simd_vst3q_f64(float64_t * a, float64x2x3_t b);
void __simd_vst4_u8(uint8_t * a, uint8x8x4_t b);
void __simd_vst4_s8(int8_t * a, int8x8x4_t b);
void __simd_vst4_p8(poly8_t * a, poly8x8x4_t b);
void __simd_vst4_u16(uint16_t * a, uint16x4x4_t b);
void __simd_vst4_s16(int16_t * a, int16x4x4_t b);
void __simd_vst4_p16(poly16_t * a, poly16x4x4_t b);
void __simd_vst4_u32(uint32_t * a, uint32x2x4_t b);
void __simd_vst4_s32(int32_t * a, int32x2x4_t b);
void __simd_vst4_f32(float32_t * a, float32x2x4_t b);
void __simd_vst4q_u8(uint8_t * a, uint8x16x4_t b);
void __simd_vst4q_s8(int8_t * a, int8x16x4_t b);
void __simd_vst4q_p8(poly8_t * a, poly8x16x4_t b);
void __simd_vst4q_u16(uint16_t * a, uint16x8x4_t b);
void __simd_vst4q_s16(int16_t * a, int16x8x4_t b);
void __simd_vst4q_p16(poly16_t * a, poly16x8x4_t b);
void __simd_vst4q_u32(uint32_t * a, uint32x4x4_t b);
void __simd_vst4q_s32(int32_t * a, int32x4x4_t b);
void __simd_vst4q_f32(float32_t * a, float32x4x4_t b);
void __simd_vst4q_u64(uint64_t * a, uint64x2x4_t b);
void __simd_vst4q_s64(int64_t * a, int64x2x4_t b);
void __simd_vst4q_f64(float64_t * a, float64x2x4_t b);
uint8x8_t __simd_vld1_lane_u8(uint8_t const * a, uint8x8_t b, __ghs_c_int__ c) __NEONPURE;
int8x8_t __simd_vld1_lane_s8(int8_t const * a, int8x8_t b, __ghs_c_int__ c) __NEONPURE;
poly8x8_t __simd_vld1_lane_p8(poly8_t const * a, poly8x8_t b, __ghs_c_int__ c) __NEONPURE;
uint8x16_t __simd_vld1q_lane_u8(uint8_t const * a, uint8x16_t b, __ghs_c_int__ c) __NEONPURE;
int8x16_t __simd_vld1q_lane_s8(int8_t const * a, int8x16_t b, __ghs_c_int__ c) __NEONPURE;
poly8x16_t __simd_vld1q_lane_p8(poly8_t const * a, poly8x16_t b, __ghs_c_int__ c) __NEONPURE;
uint16x4_t __simd_vld1_lane_u16(uint16_t const * a, uint16x4_t b, __ghs_c_int__ c) __NEONPURE;
int16x4_t __simd_vld1_lane_s16(int16_t const * a, int16x4_t b, __ghs_c_int__ c) __NEONPURE;
poly16x4_t __simd_vld1_lane_p16(poly16_t const * a, poly16x4_t b, __ghs_c_int__ c) __NEONPURE;
uint16x8_t __simd_vld1q_lane_u16(uint16_t const * a, uint16x8_t b, __ghs_c_int__ c) __NEONPURE;
int16x8_t __simd_vld1q_lane_s16(int16_t const * a, int16x8_t b, __ghs_c_int__ c) __NEONPURE;
poly16x8_t __simd_vld1q_lane_p16(poly16_t const * a, poly16x8_t b, __ghs_c_int__ c) __NEONPURE;
uint32x2_t __simd_vld1_lane_u32(uint32_t const * a, uint32x2_t b, __ghs_c_int__ c) __NEONPURE;
int32x2_t __simd_vld1_lane_s32(int32_t const * a, int32x2_t b, __ghs_c_int__ c) __NEONPURE;
float32x2_t __simd_vld1_lane_f32(float32_t const * a, float32x2_t b, __ghs_c_int__ c) __NEONPURE;
uint32x4_t __simd_vld1q_lane_u32(uint32_t const * a, uint32x4_t b, __ghs_c_int__ c) __NEONPURE;
int32x4_t __simd_vld1q_lane_s32(int32_t const * a, int32x4_t b, __ghs_c_int__ c) __NEONPURE;
float32x4_t __simd_vld1q_lane_f32(float32_t const * a, float32x4_t b, __ghs_c_int__ c) __NEONPURE;
uint64x1_t __simd_vld1_lane_u64(uint64_t const * a, uint64x1_t b, __ghs_c_int__ c) __NEONPURE;
int64x1_t __simd_vld1_lane_s64(int64_t const * a, int64x1_t b, __ghs_c_int__ c) __NEONPURE;
uint64x2_t __simd_vld1q_lane_u64(uint64_t const * a, uint64x2_t b, __ghs_c_int__ c) __NEONPURE;
int64x2_t __simd_vld1q_lane_s64(int64_t const * a, int64x2_t b, __ghs_c_int__ c) __NEONPURE;
float64x2_t __simd_vld1q_lane_f64(float64_t const * a, float64x2_t b, __ghs_c_int__ c) __NEONPURE;
uint8x8x2_t __simd_vld2_lane_u8(uint8_t const * a, uint8x8x2_t b, __ghs_c_int__ c) __NEONPURE;
int8x8x2_t __simd_vld2_lane_s8(int8_t const * a, int8x8x2_t b, __ghs_c_int__ c) __NEONPURE;
poly8x8x2_t __simd_vld2_lane_p8(poly8_t const * a, poly8x8x2_t b, __ghs_c_int__ c) __NEONPURE;
uint8x16x2_t __simd_vld2q_lane_u8(uint8_t const * a, uint8x16x2_t b, __ghs_c_int__ c) __NEONPURE;
int8x16x2_t __simd_vld2q_lane_s8(int8_t const * a, int8x16x2_t b, __ghs_c_int__ c) __NEONPURE;
poly8x16x2_t __simd_vld2q_lane_p8(poly8_t const * a, poly8x16x2_t b, __ghs_c_int__ c) __NEONPURE;
uint16x4x2_t __simd_vld2_lane_u16(uint16_t const * a, uint16x4x2_t b, __ghs_c_int__ c) __NEONPURE;
int16x4x2_t __simd_vld2_lane_s16(int16_t const * a, int16x4x2_t b, __ghs_c_int__ c) __NEONPURE;
poly16x4x2_t __simd_vld2_lane_p16(poly16_t const * a, poly16x4x2_t b, __ghs_c_int__ c) __NEONPURE;
uint16x8x2_t __simd_vld2q_lane_u16(uint16_t const * a, uint16x8x2_t b, __ghs_c_int__ c) __NEONPURE;
int16x8x2_t __simd_vld2q_lane_s16(int16_t const * a, int16x8x2_t b, __ghs_c_int__ c) __NEONPURE;
poly16x8x2_t __simd_vld2q_lane_p16(poly16_t const * a, poly16x8x2_t b, __ghs_c_int__ c) __NEONPURE;
uint32x2x2_t __simd_vld2_lane_u32(uint32_t const * a, uint32x2x2_t b, __ghs_c_int__ c) __NEONPURE;
int32x2x2_t __simd_vld2_lane_s32(int32_t const * a, int32x2x2_t b, __ghs_c_int__ c) __NEONPURE;
float32x2x2_t __simd_vld2_lane_f32(float32_t const * a, float32x2x2_t b, __ghs_c_int__ c) __NEONPURE;
uint32x4x2_t __simd_vld2q_lane_u32(uint32_t const * a, uint32x4x2_t b, __ghs_c_int__ c) __NEONPURE;
int32x4x2_t __simd_vld2q_lane_s32(int32_t const * a, int32x4x2_t b, __ghs_c_int__ c) __NEONPURE;
float32x4x2_t __simd_vld2q_lane_f32(float32_t const * a, float32x4x2_t b, __ghs_c_int__ c) __NEONPURE;
uint64x1x2_t __simd_vld2_lane_u64(uint64_t const * a, uint64x1x2_t b, __ghs_c_int__ c) __NEONPURE;
int64x1x2_t __simd_vld2_lane_s64(int64_t const * a, int64x1x2_t b, __ghs_c_int__ c) __NEONPURE;
uint64x2x2_t __simd_vld2q_lane_u64(uint64_t const * a, uint64x2x2_t b, __ghs_c_int__ c) __NEONPURE;
int64x2x2_t __simd_vld2q_lane_s64(int64_t const * a, int64x2x2_t b, __ghs_c_int__ c) __NEONPURE;
float64x2x2_t __simd_vld2q_lane_f64(float64_t const * a, float64x2x2_t b, __ghs_c_int__ c) __NEONPURE;
uint8x8x3_t __simd_vld3_lane_u8(uint8_t const * a, uint8x8x3_t b, __ghs_c_int__ c) __NEONPURE;
int8x8x3_t __simd_vld3_lane_s8(int8_t const * a, int8x8x3_t b, __ghs_c_int__ c) __NEONPURE;
poly8x8x3_t __simd_vld3_lane_p8(poly8_t const * a, poly8x8x3_t b, __ghs_c_int__ c) __NEONPURE;
uint8x16x3_t __simd_vld3q_lane_u8(uint8_t const * a, uint8x16x3_t b, __ghs_c_int__ c) __NEONPURE;
int8x16x3_t __simd_vld3q_lane_s8(int8_t const * a, int8x16x3_t b, __ghs_c_int__ c) __NEONPURE;
poly8x16x3_t __simd_vld3q_lane_p8(poly8_t const * a, poly8x16x3_t b, __ghs_c_int__ c) __NEONPURE;
uint16x4x3_t __simd_vld3_lane_u16(uint16_t const * a, uint16x4x3_t b, __ghs_c_int__ c) __NEONPURE;
int16x4x3_t __simd_vld3_lane_s16(int16_t const * a, int16x4x3_t b, __ghs_c_int__ c) __NEONPURE;
poly16x4x3_t __simd_vld3_lane_p16(poly16_t const * a, poly16x4x3_t b, __ghs_c_int__ c) __NEONPURE;
uint16x8x3_t __simd_vld3q_lane_u16(uint16_t const * a, uint16x8x3_t b, __ghs_c_int__ c) __NEONPURE;
int16x8x3_t __simd_vld3q_lane_s16(int16_t const * a, int16x8x3_t b, __ghs_c_int__ c) __NEONPURE;
poly16x8x3_t __simd_vld3q_lane_p16(poly16_t const * a, poly16x8x3_t b, __ghs_c_int__ c) __NEONPURE;
uint32x2x3_t __simd_vld3_lane_u32(uint32_t const * a, uint32x2x3_t b, __ghs_c_int__ c) __NEONPURE;
int32x2x3_t __simd_vld3_lane_s32(int32_t const * a, int32x2x3_t b, __ghs_c_int__ c) __NEONPURE;
float32x2x3_t __simd_vld3_lane_f32(float32_t const * a, float32x2x3_t b, __ghs_c_int__ c) __NEONPURE;
uint32x4x3_t __simd_vld3q_lane_u32(uint32_t const * a, uint32x4x3_t b, __ghs_c_int__ c) __NEONPURE;
int32x4x3_t __simd_vld3q_lane_s32(int32_t const * a, int32x4x3_t b, __ghs_c_int__ c) __NEONPURE;
float32x4x3_t __simd_vld3q_lane_f32(float32_t const * a, float32x4x3_t b, __ghs_c_int__ c) __NEONPURE;
uint64x1x3_t __simd_vld3_lane_u64(uint64_t const * a, uint64x1x3_t b, __ghs_c_int__ c) __NEONPURE;
int64x1x3_t __simd_vld3_lane_s64(int64_t const * a, int64x1x3_t b, __ghs_c_int__ c) __NEONPURE;
uint64x2x3_t __simd_vld3q_lane_u64(uint64_t const * a, uint64x2x3_t b, __ghs_c_int__ c) __NEONPURE;
int64x2x3_t __simd_vld3q_lane_s64(int64_t const * a, int64x2x3_t b, __ghs_c_int__ c) __NEONPURE;
float64x2x3_t __simd_vld3q_lane_f64(float64_t const * a, float64x2x3_t b, __ghs_c_int__ c) __NEONPURE;
uint8x8x4_t __simd_vld4_lane_u8(uint8_t const * a, uint8x8x4_t b, __ghs_c_int__ c) __NEONPURE;
int8x8x4_t __simd_vld4_lane_s8(int8_t const * a, int8x8x4_t b, __ghs_c_int__ c) __NEONPURE;
poly8x8x4_t __simd_vld4_lane_p8(poly8_t const * a, poly8x8x4_t b, __ghs_c_int__ c) __NEONPURE;
uint8x16x4_t __simd_vld4q_lane_u8(uint8_t const * a, uint8x16x4_t b, __ghs_c_int__ c) __NEONPURE;
int8x16x4_t __simd_vld4q_lane_s8(int8_t const * a, int8x16x4_t b, __ghs_c_int__ c) __NEONPURE;
poly8x16x4_t __simd_vld4q_lane_p8(poly8_t const * a, poly8x16x4_t b, __ghs_c_int__ c) __NEONPURE;
uint16x4x4_t __simd_vld4_lane_u16(uint16_t const * a, uint16x4x4_t b, __ghs_c_int__ c) __NEONPURE;
int16x4x4_t __simd_vld4_lane_s16(int16_t const * a, int16x4x4_t b, __ghs_c_int__ c) __NEONPURE;
poly16x4x4_t __simd_vld4_lane_p16(poly16_t const * a, poly16x4x4_t b, __ghs_c_int__ c) __NEONPURE;
uint16x8x4_t __simd_vld4q_lane_u16(uint16_t const * a, uint16x8x4_t b, __ghs_c_int__ c) __NEONPURE;
int16x8x4_t __simd_vld4q_lane_s16(int16_t const * a, int16x8x4_t b, __ghs_c_int__ c) __NEONPURE;
poly16x8x4_t __simd_vld4q_lane_p16(poly16_t const * a, poly16x8x4_t b, __ghs_c_int__ c) __NEONPURE;
uint32x2x4_t __simd_vld4_lane_u32(uint32_t const * a, uint32x2x4_t b, __ghs_c_int__ c) __NEONPURE;
int32x2x4_t __simd_vld4_lane_s32(int32_t const * a, int32x2x4_t b, __ghs_c_int__ c) __NEONPURE;
float32x2x4_t __simd_vld4_lane_f32(float32_t const * a, float32x2x4_t b, __ghs_c_int__ c) __NEONPURE;
uint32x4x4_t __simd_vld4q_lane_u32(uint32_t const * a, uint32x4x4_t b, __ghs_c_int__ c) __NEONPURE;
int32x4x4_t __simd_vld4q_lane_s32(int32_t const * a, int32x4x4_t b, __ghs_c_int__ c) __NEONPURE;
float32x4x4_t __simd_vld4q_lane_f32(float32_t const * a, float32x4x4_t b, __ghs_c_int__ c) __NEONPURE;
uint64x1x4_t __simd_vld4_lane_u64(uint64_t const * a, uint64x1x4_t b, __ghs_c_int__ c) __NEONPURE;
int64x1x4_t __simd_vld4_lane_s64(int64_t const * a, int64x1x4_t b, __ghs_c_int__ c) __NEONPURE;
uint64x2x4_t __simd_vld4q_lane_u64(uint64_t const * a, uint64x2x4_t b, __ghs_c_int__ c) __NEONPURE;
int64x2x4_t __simd_vld4q_lane_s64(int64_t const * a, int64x2x4_t b, __ghs_c_int__ c) __NEONPURE;
float64x2x4_t __simd_vld4q_lane_f64(float64_t const * a, float64x2x4_t b, __ghs_c_int__ c) __NEONPURE;
void __simd_vst1_lane_u8(uint8_t * a, uint8x8_t b, __ghs_c_int__ c);
void __simd_vst1_lane_s8(int8_t * a, int8x8_t b, __ghs_c_int__ c);
void __simd_vst1_lane_p8(poly8_t * a, poly8x8_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_u8(uint8_t * a, uint8x16_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_s8(int8_t * a, int8x16_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_p8(poly8_t * a, poly8x16_t b, __ghs_c_int__ c);
void __simd_vst1_lane_u16(uint16_t * a, uint16x4_t b, __ghs_c_int__ c);
void __simd_vst1_lane_s16(int16_t * a, int16x4_t b, __ghs_c_int__ c);
void __simd_vst1_lane_p16(poly16_t * a, poly16x4_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_u16(uint16_t * a, uint16x8_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_s16(int16_t * a, int16x8_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_p16(poly16_t * a, poly16x8_t b, __ghs_c_int__ c);
void __simd_vst1_lane_u32(uint32_t * a, uint32x2_t b, __ghs_c_int__ c);
void __simd_vst1_lane_s32(int32_t * a, int32x2_t b, __ghs_c_int__ c);
void __simd_vst1_lane_f32(float32_t * a, float32x2_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_u32(uint32_t * a, uint32x4_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_s32(int32_t * a, int32x4_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_f32(float32_t * a, float32x4_t b, __ghs_c_int__ c);
void __simd_vst1_lane_u64(uint64_t * a, uint64x1_t b, __ghs_c_int__ c);
void __simd_vst1_lane_s64(int64_t * a, int64x1_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_u64(uint64_t * a, uint64x2_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_s64(int64_t * a, int64x2_t b, __ghs_c_int__ c);
void __simd_vst1q_lane_f64(float64_t * a, float64x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_u8(uint8_t * a, uint8x8x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_s8(int8_t * a, int8x8x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_p8(poly8_t * a, poly8x8x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_u8(uint8_t * a, uint8x16x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_s8(int8_t * a, int8x16x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_p8(poly8_t * a, poly8x16x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_u16(uint16_t * a, uint16x4x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_s16(int16_t * a, int16x4x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_p16(poly16_t * a, poly16x4x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_u16(uint16_t * a, uint16x8x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_s16(int16_t * a, int16x8x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_p16(poly16_t * a, poly16x8x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_u32(uint32_t * a, uint32x2x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_s32(int32_t * a, int32x2x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_f32(float32_t * a, float32x2x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_u32(uint32_t * a, uint32x4x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_s32(int32_t * a, int32x4x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_f32(float32_t * a, float32x4x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_u64(uint64_t * a, uint64x1x2_t b, __ghs_c_int__ c);
void __simd_vst2_lane_s64(int64_t * a, int64x1x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_u64(uint64_t * a, uint64x2x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_s64(int64_t * a, int64x2x2_t b, __ghs_c_int__ c);
void __simd_vst2q_lane_f64(float64_t * a, float64x2x2_t b, __ghs_c_int__ c);
void __simd_vst3_lane_u8(uint8_t * a, uint8x8x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_s8(int8_t * a, int8x8x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_p8(poly8_t * a, poly8x8x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_u8(uint8_t * a, uint8x16x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_s8(int8_t * a, int8x16x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_p8(poly8_t * a, poly8x16x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_u16(uint16_t * a, uint16x4x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_s16(int16_t * a, int16x4x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_p16(poly16_t * a, poly16x4x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_u16(uint16_t * a, uint16x8x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_s16(int16_t * a, int16x8x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_p16(poly16_t * a, poly16x8x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_u32(uint32_t * a, uint32x2x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_s32(int32_t * a, int32x2x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_f32(float32_t * a, float32x2x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_u32(uint32_t * a, uint32x4x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_s32(int32_t * a, int32x4x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_f32(float32_t * a, float32x4x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_u64(uint64_t * a, uint64x1x3_t b, __ghs_c_int__ c);
void __simd_vst3_lane_s64(int64_t * a, int64x1x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_u64(uint64_t * a, uint64x2x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_s64(int64_t * a, int64x2x3_t b, __ghs_c_int__ c);
void __simd_vst3q_lane_f64(float64_t * a, float64x2x3_t b, __ghs_c_int__ c);
void __simd_vst4_lane_u8(uint8_t * a, uint8x8x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_s8(int8_t * a, int8x8x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_p8(poly8_t * a, poly8x8x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_u8(uint8_t * a, uint8x16x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_s8(int8_t * a, int8x16x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_p8(poly8_t * a, poly8x16x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_u16(uint16_t * a, uint16x4x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_s16(int16_t * a, int16x4x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_p16(poly16_t * a, poly16x4x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_u16(uint16_t * a, uint16x8x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_s16(int16_t * a, int16x8x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_p16(poly16_t * a, poly16x8x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_u32(uint32_t * a, uint32x2x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_s32(int32_t * a, int32x2x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_f32(float32_t * a, float32x2x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_u32(uint32_t * a, uint32x4x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_s32(int32_t * a, int32x4x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_f32(float32_t * a, float32x4x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_u64(uint64_t * a, uint64x1x4_t b, __ghs_c_int__ c);
void __simd_vst4_lane_s64(int64_t * a, int64x1x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_u64(uint64_t * a, uint64x2x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_s64(int64_t * a, int64x2x4_t b, __ghs_c_int__ c);
void __simd_vst4q_lane_f64(float64_t * a, float64x2x4_t b, __ghs_c_int__ c);
uint8x8_t __simd_vld1_dup_u8(uint8_t const * a) __NEONPURE;
int8x8_t __simd_vld1_dup_s8(int8_t const * a) __NEONPURE;
poly8x8_t __simd_vld1_dup_p8(poly8_t const * a) __NEONPURE;
uint16x4_t __simd_vld1_dup_u16(uint16_t const * a) __NEONPURE;
int16x4_t __simd_vld1_dup_s16(int16_t const * a) __NEONPURE;
poly16x4_t __simd_vld1_dup_p16(poly16_t const * a) __NEONPURE;
uint32x2_t __simd_vld1_dup_u32(uint32_t const * a) __NEONPURE;
int32x2_t __simd_vld1_dup_s32(int32_t const * a) __NEONPURE;
float32x2_t __simd_vld1_dup_f32(float32_t const * a) __NEONPURE;
uint64x1_t __simd_vld1_dup_u64(uint64_t const * a) __NEONPURE;
int64x1_t __simd_vld1_dup_s64(int64_t const * a) __NEONPURE;
uint8x16_t __simd_vld1q_dup_u8(uint8_t const * a) __NEONPURE;
int8x16_t __simd_vld1q_dup_s8(int8_t const * a) __NEONPURE;
poly8x16_t __simd_vld1q_dup_p8(poly8_t const * a) __NEONPURE;
uint16x8_t __simd_vld1q_dup_u16(uint16_t const * a) __NEONPURE;
int16x8_t __simd_vld1q_dup_s16(int16_t const * a) __NEONPURE;
poly16x8_t __simd_vld1q_dup_p16(poly16_t const * a) __NEONPURE;
uint32x4_t __simd_vld1q_dup_u32(uint32_t const * a) __NEONPURE;
int32x4_t __simd_vld1q_dup_s32(int32_t const * a) __NEONPURE;
float32x4_t __simd_vld1q_dup_f32(float32_t const * a) __NEONPURE;
uint64x2_t __simd_vld1q_dup_u64(uint64_t const * a) __NEONPURE;
int64x2_t __simd_vld1q_dup_s64(int64_t const * a) __NEONPURE;
float64x2_t __simd_vld1q_dup_f64(float64_t const * a) __NEONPURE;
uint8x8x2_t __simd_vld2_dup_u8(uint8_t const * a) __NEONPURE;
int8x8x2_t __simd_vld2_dup_s8(int8_t const * a) __NEONPURE;
poly8x8x2_t __simd_vld2_dup_p8(poly8_t const * a) __NEONPURE;
uint16x4x2_t __simd_vld2_dup_u16(uint16_t const * a) __NEONPURE;
int16x4x2_t __simd_vld2_dup_s16(int16_t const * a) __NEONPURE;
poly16x4x2_t __simd_vld2_dup_p16(poly16_t const * a) __NEONPURE;
uint32x2x2_t __simd_vld2_dup_u32(uint32_t const * a) __NEONPURE;
int32x2x2_t __simd_vld2_dup_s32(int32_t const * a) __NEONPURE;
float32x2x2_t __simd_vld2_dup_f32(float32_t const * a) __NEONPURE;
uint64x1x2_t __simd_vld2_dup_u64(uint64_t const * a) __NEONPURE;
int64x1x2_t __simd_vld2_dup_s64(int64_t const * a) __NEONPURE;
uint8x16x2_t __simd_vld2q_dup_u8(uint8_t const * a) __NEONPURE;
int8x16x2_t __simd_vld2q_dup_s8(int8_t const * a) __NEONPURE;
poly8x16x2_t __simd_vld2q_dup_p8(poly8_t const * a) __NEONPURE;
uint16x8x2_t __simd_vld2q_dup_u16(uint16_t const * a) __NEONPURE;
int16x8x2_t __simd_vld2q_dup_s16(int16_t const * a) __NEONPURE;
poly16x8x2_t __simd_vld2q_dup_p16(poly16_t const * a) __NEONPURE;
uint32x4x2_t __simd_vld2q_dup_u32(uint32_t const * a) __NEONPURE;
int32x4x2_t __simd_vld2q_dup_s32(int32_t const * a) __NEONPURE;
float32x4x2_t __simd_vld2q_dup_f32(float32_t const * a) __NEONPURE;
uint64x2x2_t __simd_vld2q_dup_u64(uint64_t const * a) __NEONPURE;
int64x2x2_t __simd_vld2q_dup_s64(int64_t const * a) __NEONPURE;
float64x2x2_t __simd_vld2q_dup_f64(float64_t const * a) __NEONPURE;
uint8x8x3_t __simd_vld3_dup_u8(uint8_t const * a) __NEONPURE;
int8x8x3_t __simd_vld3_dup_s8(int8_t const * a) __NEONPURE;
poly8x8x3_t __simd_vld3_dup_p8(poly8_t const * a) __NEONPURE;
uint16x4x3_t __simd_vld3_dup_u16(uint16_t const * a) __NEONPURE;
int16x4x3_t __simd_vld3_dup_s16(int16_t const * a) __NEONPURE;
poly16x4x3_t __simd_vld3_dup_p16(poly16_t const * a) __NEONPURE;
uint32x2x3_t __simd_vld3_dup_u32(uint32_t const * a) __NEONPURE;
int32x2x3_t __simd_vld3_dup_s32(int32_t const * a) __NEONPURE;
float32x2x3_t __simd_vld3_dup_f32(float32_t const * a) __NEONPURE;
uint64x1x3_t __simd_vld3_dup_u64(uint64_t const * a) __NEONPURE;
int64x1x3_t __simd_vld3_dup_s64(int64_t const * a) __NEONPURE;
uint8x16x3_t __simd_vld3q_dup_u8(uint8_t const * a) __NEONPURE;
int8x16x3_t __simd_vld3q_dup_s8(int8_t const * a) __NEONPURE;
poly8x16x3_t __simd_vld3q_dup_p8(poly8_t const * a) __NEONPURE;
uint16x8x3_t __simd_vld3q_dup_u16(uint16_t const * a) __NEONPURE;
int16x8x3_t __simd_vld3q_dup_s16(int16_t const * a) __NEONPURE;
poly16x8x3_t __simd_vld3q_dup_p16(poly16_t const * a) __NEONPURE;
uint32x4x3_t __simd_vld3q_dup_u32(uint32_t const * a) __NEONPURE;
int32x4x3_t __simd_vld3q_dup_s32(int32_t const * a) __NEONPURE;
float32x4x3_t __simd_vld3q_dup_f32(float32_t const * a) __NEONPURE;
uint64x2x3_t __simd_vld3q_dup_u64(uint64_t const * a) __NEONPURE;
int64x2x3_t __simd_vld3q_dup_s64(int64_t const * a) __NEONPURE;
float64x2x3_t __simd_vld3q_dup_f64(float64_t const * a) __NEONPURE;
uint8x8x4_t __simd_vld4_dup_u8(uint8_t const * a) __NEONPURE;
int8x8x4_t __simd_vld4_dup_s8(int8_t const * a) __NEONPURE;
poly8x8x4_t __simd_vld4_dup_p8(poly8_t const * a) __NEONPURE;
uint16x4x4_t __simd_vld4_dup_u16(uint16_t const * a) __NEONPURE;
int16x4x4_t __simd_vld4_dup_s16(int16_t const * a) __NEONPURE;
poly16x4x4_t __simd_vld4_dup_p16(poly16_t const * a) __NEONPURE;
uint32x2x4_t __simd_vld4_dup_u32(uint32_t const * a) __NEONPURE;
int32x2x4_t __simd_vld4_dup_s32(int32_t const * a) __NEONPURE;
float32x2x4_t __simd_vld4_dup_f32(float32_t const * a) __NEONPURE;
uint64x1x4_t __simd_vld4_dup_u64(uint64_t const * a) __NEONPURE;
int64x1x4_t __simd_vld4_dup_s64(int64_t const * a) __NEONPURE;
uint8x16x4_t __simd_vld4q_dup_u8(uint8_t const * a) __NEONPURE;
int8x16x4_t __simd_vld4q_dup_s8(int8_t const * a) __NEONPURE;
poly8x16x4_t __simd_vld4q_dup_p8(poly8_t const * a) __NEONPURE;
uint16x8x4_t __simd_vld4q_dup_u16(uint16_t const * a) __NEONPURE;
int16x8x4_t __simd_vld4q_dup_s16(int16_t const * a) __NEONPURE;
poly16x8x4_t __simd_vld4q_dup_p16(poly16_t const * a) __NEONPURE;
uint32x4x4_t __simd_vld4q_dup_u32(uint32_t const * a) __NEONPURE;
int32x4x4_t __simd_vld4q_dup_s32(int32_t const * a) __NEONPURE;
float32x4x4_t __simd_vld4q_dup_f32(float32_t const * a) __NEONPURE;
uint64x2x4_t __simd_vld4q_dup_u64(uint64_t const * a) __NEONPURE;
int64x2x4_t __simd_vld4q_dup_s64(int64_t const * a) __NEONPURE;
float64x2x4_t __simd_vld4q_dup_f64(float64_t const * a) __NEONPURE;
uint8x8_t __simd_vadd_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vadd_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vadd_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vadd_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vadd_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vadd_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vaddq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vaddq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vaddq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vaddq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vaddq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vaddq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vaddq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vaddq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vsub_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vsub_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vsub_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vsub_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vsub_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vsub_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vsubq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vsubq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vsubq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vsubq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vsubq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vsubq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vsubq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vsubq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vpadd_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vpadd_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vpadd_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vpadd_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vpadd_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vpadd_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vpaddq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vpaddq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vpaddq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vpaddq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vpaddq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vpaddq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vpaddq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vpaddq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
int8x8_t __simd_vhadd_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vhadd_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vhadd_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vhaddq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vhaddq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vhaddq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vhadd_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vhadd_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vhadd_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vhaddq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vhaddq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vhaddq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vqadd_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vqadd_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vqadd_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vqaddq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vqaddq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vqaddq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vqaddq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vqadd_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vqadd_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vqadd_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vqaddq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vqaddq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vqaddq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vqaddq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int8x8_t __simd_vrhadd_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vrhadd_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vrhadd_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vrhaddq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vrhaddq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vrhaddq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vrhadd_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vrhadd_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vrhadd_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vrhaddq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vrhaddq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vrhaddq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vhsub_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vhsub_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vhsub_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vhsubq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vhsubq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vhsubq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vhsub_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vhsub_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vhsub_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vhsubq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vhsubq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vhsubq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vqsub_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vqsub_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vqsub_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vqsubq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vqsubq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vqsubq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vqsubq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vqsub_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vqsub_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vqsub_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vqsubq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vqsubq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vqsubq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vqsubq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
uint8x8_t __simd_vmul_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vmul_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vmul_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vmul_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vmul_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vmul_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vmulq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vmulq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vmulq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vmulq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vmulq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vmulq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vmla_u8(uint8x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vmla_s8(int8x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
uint16x4_t __simd_vmla_u16(uint16x4_t a, uint16x4_t b, uint16x4_t c) __NEONSAFE;
int16x4_t __simd_vmla_s16(int16x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
uint32x2_t __simd_vmla_u32(uint32x2_t a, uint32x2_t b, uint32x2_t c) __NEONSAFE;
int32x2_t __simd_vmla_s32(int32x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
uint8x16_t __simd_vmlaq_u8(uint8x16_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
int8x16_t __simd_vmlaq_s8(int8x16_t a, int8x16_t b, int8x16_t c) __NEONSAFE;
uint16x8_t __simd_vmlaq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
int16x8_t __simd_vmlaq_s16(int16x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
uint32x4_t __simd_vmlaq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int32x4_t __simd_vmlaq_s32(int32x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint8x8_t __simd_vmls_u8(uint8x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vmls_s8(int8x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
uint16x4_t __simd_vmls_u16(uint16x4_t a, uint16x4_t b, uint16x4_t c) __NEONSAFE;
int16x4_t __simd_vmls_s16(int16x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
uint32x2_t __simd_vmls_u32(uint32x2_t a, uint32x2_t b, uint32x2_t c) __NEONSAFE;
int32x2_t __simd_vmls_s32(int32x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
uint8x16_t __simd_vmlsq_u8(uint8x16_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
int8x16_t __simd_vmlsq_s8(int8x16_t a, int8x16_t b, int8x16_t c) __NEONSAFE;
uint16x8_t __simd_vmlsq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
int16x8_t __simd_vmlsq_s16(int16x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
uint32x4_t __simd_vmlsq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int32x4_t __simd_vmlsq_s32(int32x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
int16x4_t __simd_vqdmulh_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vqdmulh_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vqdmulhq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vqdmulhq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int16x4_t __simd_vqrdmulh_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vqrdmulh_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vqrdmulhq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vqrdmulhq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
poly8x8_t __simd_vmul_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
poly8x16_t __simd_vmulq_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
int8x8_t __simd_vshl_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vshl_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vshl_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vshlq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vshlq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vshlq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vshlq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vshl_u8(uint8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vshl_u16(uint16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vshl_u32(uint32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vshlq_u8(uint8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vshlq_u16(uint16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vshlq_u32(uint32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vshlq_u64(uint64x2_t a, int64x2_t b) __NEONSAFE;
int8x8_t __simd_vrshl_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vrshl_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vrshl_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vrshlq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vrshlq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vrshlq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vrshlq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vrshl_u8(uint8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vrshl_u16(uint16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vrshl_u32(uint32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vrshlq_u8(uint8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vrshlq_u16(uint16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vrshlq_u32(uint32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vrshlq_u64(uint64x2_t a, int64x2_t b) __NEONSAFE;
int8x8_t __simd_vqshl_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vqshl_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vqshl_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vqshlq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vqshlq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vqshlq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vqshlq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vqshl_u8(uint8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vqshl_u16(uint16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vqshl_u32(uint32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vqshlq_u8(uint8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vqshlq_u16(uint16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vqshlq_u32(uint32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vqshlq_u64(uint64x2_t a, int64x2_t b) __NEONSAFE;
int8x8_t __simd_vqrshl_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vqrshl_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vqrshl_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vqrshlq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vqrshlq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vqrshlq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vqrshlq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vqrshl_u8(uint8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vqrshl_u16(uint16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vqrshl_u32(uint32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vqrshlq_u8(uint8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vqrshlq_u16(uint16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vqrshlq_u32(uint32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vqrshlq_u64(uint64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vceq_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vceq_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint8x8_t __simd_vceq_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vceq_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint16x4_t __simd_vceq_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint16x4_t __simd_vceq_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vceq_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint32x2_t __simd_vceq_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vceqq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint8x16_t __simd_vceqq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint8x16_t __simd_vceqq_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vceqq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint16x8_t __simd_vceqq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint16x8_t __simd_vceqq_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vceqq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint32x4_t __simd_vceqq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vceqq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
uint64x2_t __simd_vceqq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vcge_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vcge_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vcge_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcgeq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcgeq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcgeq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcgeq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vcgt_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vcgt_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vcgt_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcgtq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcgtq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcgtq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcgtq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vcge_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vcge_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vcge_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcgeq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcgeq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcgeq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcgeq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
uint8x8_t __simd_vcgt_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vcgt_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vcgt_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcgtq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcgtq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcgtq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcgtq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
uint8x8_t __simd_vtst_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vtst_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint8x8_t __simd_vtst_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vtst_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint16x4_t __simd_vtst_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint16x4_t __simd_vtst_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vtst_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint32x2_t __simd_vtst_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vtstq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint8x16_t __simd_vtstq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint8x16_t __simd_vtstq_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vtstq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint16x8_t __simd_vtstq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint16x8_t __simd_vtstq_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vtstq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint32x4_t __simd_vtstq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vtstq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
uint64x2_t __simd_vtstq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vclt_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vclt_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vclt_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcltq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcltq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcltq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcltq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vcle_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vcle_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vcle_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcleq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcleq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcleq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcleq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vclt_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vclt_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vclt_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcltq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcltq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcltq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcltq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
uint8x8_t __simd_vcle_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vcle_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vcle_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vcleq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vcleq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vcleq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcleq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int8x8_t __simd_vmax_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vmax_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vmax_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vmaxq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vmaxq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vmaxq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vmax_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vmax_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vmax_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vmaxq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vmaxq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vmaxq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vmin_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vmin_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vmin_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vminq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vminq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vminq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vmin_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vmin_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vmin_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vminq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vminq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vminq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vpmax_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vpmax_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vpmax_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vpmaxq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vpmaxq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vpmaxq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vpmax_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vpmax_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vpmax_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vpmaxq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vpmaxq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vpmaxq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vpmin_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vpmin_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vpmin_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vpminq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vpminq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vpminq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vpmin_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vpmin_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vpmin_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vpminq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vpminq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vpminq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vaba_s8(int8x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
int16x4_t __simd_vaba_s16(int16x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int32x2_t __simd_vaba_s32(int32x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int8x16_t __simd_vabaq_s8(int8x16_t a, int8x16_t b, int8x16_t c) __NEONSAFE;
int16x8_t __simd_vabaq_s16(int16x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int32x4_t __simd_vabaq_s32(int32x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint8x8_t __simd_vaba_u8(uint8x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
uint16x4_t __simd_vaba_u16(uint16x4_t a, uint16x4_t b, uint16x4_t c) __NEONSAFE;
uint32x2_t __simd_vaba_u32(uint32x2_t a, uint32x2_t b, uint32x2_t c) __NEONSAFE;
uint8x16_t __simd_vabaq_u8(uint8x16_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
uint16x8_t __simd_vabaq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
uint32x4_t __simd_vabaq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int8x8_t __simd_vabd_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vabd_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vabd_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int8x16_t __simd_vabdq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vabdq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vabdq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint8x8_t __simd_vabd_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vabd_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vabd_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint8x16_t __simd_vabdq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vabdq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vabdq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_u8(uint8x8_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_u16(uint16x4_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_u32(uint32x2_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_u64(uint64x1_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s8(int8x8_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s16(int16x4_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s32(int32x2_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s64(int64x1_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_p8(poly8x8_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_p16(poly16x4_t a) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_f32(float32x2_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_u8(uint8x8_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_u16(uint16x4_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_u32(uint32x2_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_u64(uint64x1_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s8(int8x8_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s16(int16x4_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s32(int32x2_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s64(int64x1_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_p8(poly8x8_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_p16(poly16x4_t a) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_f32(float32x2_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_u8(uint8x8_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_u16(uint16x4_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_u32(uint32x2_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_u64(uint64x1_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s8(int8x8_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s16(int16x4_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s32(int32x2_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s64(int64x1_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_p8(poly8x8_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_p16(poly16x4_t a) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_f32(float32x2_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_u8(uint8x8_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_u16(uint16x4_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_u32(uint32x2_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_u64(uint64x1_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s8(int8x8_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s16(int16x4_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s32(int32x2_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s64(int64x1_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_p8(poly8x8_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_p16(poly16x4_t a) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_f32(float32x2_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u8(uint8x8_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u16(uint16x4_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u32(uint32x2_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u64(uint64x1_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_s8(int8x8_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_s16(int16x4_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_s32(int32x2_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_s64(int64x1_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_p8(poly8x8_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_p16(poly16x4_t a) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_f32(float32x2_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u8(uint8x8_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u16(uint16x4_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u32(uint32x2_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u64(uint64x1_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_s8(int8x8_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_s16(int16x4_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_s32(int32x2_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_s64(int64x1_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_p8(poly8x8_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_p16(poly16x4_t a) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_f32(float32x2_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u8(uint8x8_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u16(uint16x4_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u32(uint32x2_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u64(uint64x1_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_s8(int8x8_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_s16(int16x4_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_s32(int32x2_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_s64(int64x1_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_p8(poly8x8_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_p16(poly16x4_t a) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_f32(float32x2_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u8(uint8x8_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u16(uint16x4_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u32(uint32x2_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u64(uint64x1_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_s8(int8x8_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_s16(int16x4_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_s32(int32x2_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_s64(int64x1_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_p8(poly8x8_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_p16(poly16x4_t a) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_f32(float32x2_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u8(uint8x8_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u16(uint16x4_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u32(uint32x2_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u64(uint64x1_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s8(int8x8_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s16(int16x4_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s32(int32x2_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s64(int64x1_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_p8(poly8x8_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_p16(poly16x4_t a) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_f32(float32x2_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u8(uint8x8_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u16(uint16x4_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u32(uint32x2_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u64(uint64x1_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s8(int8x8_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s16(int16x4_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s32(int32x2_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s64(int64x1_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_p8(poly8x8_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_p16(poly16x4_t a) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_f32(float32x2_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u8(uint8x8_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u16(uint16x4_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u32(uint32x2_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u64(uint64x1_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s8(int8x8_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s16(int16x4_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s32(int32x2_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s64(int64x1_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_p8(poly8x8_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_p16(poly16x4_t a) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_f32(float32x2_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_u8(uint8x16_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_u16(uint16x8_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_u32(uint32x4_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_u64(uint64x2_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s8(int8x16_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s16(int16x8_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s32(int32x4_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s64(int64x2_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_p8(poly8x16_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_p16(poly16x8_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_f32(float32x4_t a) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_f64(float64x2_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_u8(uint8x16_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_u16(uint16x8_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_u32(uint32x4_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_u64(uint64x2_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s8(int8x16_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s16(int16x8_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s32(int32x4_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s64(int64x2_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_p8(poly8x16_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_p16(poly16x8_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_f32(float32x4_t a) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_f64(float64x2_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_u8(uint8x16_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_u16(uint16x8_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_u32(uint32x4_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_u64(uint64x2_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s8(int8x16_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s16(int16x8_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s32(int32x4_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s64(int64x2_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_p8(poly8x16_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_p16(poly16x8_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_f32(float32x4_t a) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_f64(float64x2_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_u8(uint8x16_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_u16(uint16x8_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_u32(uint32x4_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_u64(uint64x2_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s8(int8x16_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s16(int16x8_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s32(int32x4_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s64(int64x2_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_p8(poly8x16_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_p16(poly16x8_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_f64(float64x2_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u8(uint8x16_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u16(uint16x8_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u32(uint32x4_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u64(uint64x2_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_s8(int8x16_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_s16(int16x8_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_s32(int32x4_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_s64(int64x2_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_p8(poly8x16_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_p16(poly16x8_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_f32(float32x4_t a) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_f64(float64x2_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u8(uint8x16_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u16(uint16x8_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u32(uint32x4_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u64(uint64x2_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_s8(int8x16_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_s16(int16x8_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_s32(int32x4_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_s64(int64x2_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_p8(poly8x16_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_p16(poly16x8_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_f32(float32x4_t a) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_f64(float64x2_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u8(uint8x16_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u16(uint16x8_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u32(uint32x4_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u64(uint64x2_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_s8(int8x16_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_s16(int16x8_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_s32(int32x4_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_s64(int64x2_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_p8(poly8x16_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_p16(poly16x8_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_f32(float32x4_t a) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_f64(float64x2_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u8(uint8x16_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u16(uint16x8_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u32(uint32x4_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u64(uint64x2_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_s8(int8x16_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_s16(int16x8_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_s32(int32x4_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_s64(int64x2_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_p8(poly8x16_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_p16(poly16x8_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_f32(float32x4_t a) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_f64(float64x2_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u8(uint8x16_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u16(uint16x8_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u32(uint32x4_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u64(uint64x2_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s8(int8x16_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s16(int16x8_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s32(int32x4_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s64(int64x2_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_p8(poly8x16_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_p16(poly16x8_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_f32(float32x4_t a) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_f64(float64x2_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u8(uint8x16_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u16(uint16x8_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u32(uint32x4_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u64(uint64x2_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s8(int8x16_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s16(int16x8_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s32(int32x4_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s64(int64x2_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_p8(poly8x16_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_p16(poly16x8_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_f32(float32x4_t a) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_f64(float64x2_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u8(uint8x16_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u16(uint16x8_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u32(uint32x4_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u64(uint64x2_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s8(int8x16_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s16(int16x8_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s32(int32x4_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s64(int64x2_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_p8(poly8x16_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_p16(poly16x8_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_f32(float32x4_t a) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_f64(float64x2_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_u8(uint8x16_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_u16(uint16x8_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_u32(uint32x4_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_u64(uint64x2_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_s8(int8x16_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_s16(int16x8_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_s32(int32x4_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_s64(int64x2_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_p8(poly8x16_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_p16(poly16x8_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vreinterpretq_f64_f64(float64x2_t a) __NEONSAFE;
uint8x8_t __simd_vand_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vand_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vand_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint64x1_t __simd_vand_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int8x8_t __simd_vand_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vand_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vand_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int64x1_t __simd_vand_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint8x16_t __simd_vandq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vandq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vandq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vandq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int8x16_t __simd_vandq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vandq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vandq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vandq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vbic_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vbic_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vbic_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint64x1_t __simd_vbic_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int8x8_t __simd_vbic_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vbic_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vbic_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int64x1_t __simd_vbic_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint8x16_t __simd_vbicq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vbicq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vbicq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vbicq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int8x16_t __simd_vbicq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vbicq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vbicq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vbicq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vorr_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vorr_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vorr_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint64x1_t __simd_vorr_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int8x8_t __simd_vorr_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vorr_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vorr_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int64x1_t __simd_vorr_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint8x16_t __simd_vorrq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vorrq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vorrq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vorrq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int8x16_t __simd_vorrq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vorrq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vorrq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vorrq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vorn_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_vorn_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_vorn_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint64x1_t __simd_vorn_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int8x8_t __simd_vorn_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_vorn_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_vorn_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int64x1_t __simd_vorn_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint8x16_t __simd_vornq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_vornq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_vornq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_vornq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int8x16_t __simd_vornq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_vornq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_vornq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_vornq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_veor_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint16x4_t __simd_veor_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint32x2_t __simd_veor_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint64x1_t __simd_veor_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int8x8_t __simd_veor_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int16x4_t __simd_veor_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int32x2_t __simd_veor_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int64x1_t __simd_veor_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint8x16_t __simd_veorq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint16x8_t __simd_veorq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint32x4_t __simd_veorq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint64x2_t __simd_veorq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int8x16_t __simd_veorq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int16x8_t __simd_veorq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int32x4_t __simd_veorq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
int64x2_t __simd_veorq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vbsl_u8(uint8x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
uint16x4_t __simd_vbsl_u16(uint16x4_t a, uint16x4_t b, uint16x4_t c) __NEONSAFE;
uint32x2_t __simd_vbsl_u32(uint32x2_t a, uint32x2_t b, uint32x2_t c) __NEONSAFE;
uint64x1_t __simd_vbsl_u64(uint64x1_t a, uint64x1_t b, uint64x1_t c) __NEONSAFE;
int8x8_t __simd_vbsl_s8(uint8x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
int16x4_t __simd_vbsl_s16(uint16x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int32x2_t __simd_vbsl_s32(uint32x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int64x1_t __simd_vbsl_s64(uint64x1_t a, int64x1_t b, int64x1_t c) __NEONSAFE;
float32x2_t __simd_vbsl_f32(uint32x2_t a, float32x2_t b, float32x2_t c) __NEONSAFE;
poly8x8_t __simd_vbsl_p8(uint8x8_t a, poly8x8_t b, poly8x8_t c) __NEONSAFE;
poly16x4_t __simd_vbsl_p16(uint16x4_t a, poly16x4_t b, poly16x4_t c) __NEONSAFE;
uint8x16_t __simd_vbslq_u8(uint8x16_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
uint16x8_t __simd_vbslq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
uint32x4_t __simd_vbslq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
uint64x2_t __simd_vbslq_u64(uint64x2_t a, uint64x2_t b, uint64x2_t c) __NEONSAFE;
int8x16_t __simd_vbslq_s8(uint8x16_t a, int8x16_t b, int8x16_t c) __NEONSAFE;
int16x8_t __simd_vbslq_s16(uint16x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int32x4_t __simd_vbslq_s32(uint32x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
int64x2_t __simd_vbslq_s64(uint64x2_t a, int64x2_t b, int64x2_t c) __NEONSAFE;
float32x4_t __simd_vbslq_f32(uint32x4_t a, float32x4_t b, float32x4_t c) __NEONSAFE;
float64x2_t __simd_vbslq_f64(uint64x2_t a, float64x2_t b, float64x2_t c) __NEONSAFE;
poly8x16_t __simd_vbslq_p8(uint8x16_t a, poly8x16_t b, poly8x16_t c) __NEONSAFE;
poly16x8_t __simd_vbslq_p16(uint16x8_t a, poly16x8_t b, poly16x8_t c) __NEONSAFE;
int8_t __simd_vqaddb_s8(int8_t a, int8_t b) __NEONSAFE;
int16_t __simd_vqaddh_s16(int16_t a, int16_t b) __NEONSAFE;
int32_t __simd_vqadds_s32(int32_t a, int32_t b) __NEONSAFE;
int64x1_t __simd_vqadd_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
int64_t __simd_vqaddd_s64(int64_t a, int64_t b) __NEONSAFE;
uint8_t __simd_vqaddb_u8(uint8_t a, uint8_t b) __NEONSAFE;
uint16_t __simd_vqaddh_u16(uint16_t a, uint16_t b) __NEONSAFE;
uint32_t __simd_vqadds_u32(uint32_t a, uint32_t b) __NEONSAFE;
uint64x1_t __simd_vqadd_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64_t __simd_vqaddd_u64(uint64_t a, uint64_t b) __NEONSAFE;
int8_t __simd_vqsubb_s8(int8_t a, int8_t b) __NEONSAFE;
int16_t __simd_vqsubh_s16(int16_t a, int16_t b) __NEONSAFE;
int32_t __simd_vqsubs_s32(int32_t a, int32_t b) __NEONSAFE;
int64x1_t __simd_vqsub_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
int64_t __simd_vqsubd_s64(int64_t a, int64_t b) __NEONSAFE;
uint8_t __simd_vqsubb_u8(uint8_t a, uint8_t b) __NEONSAFE;
uint16_t __simd_vqsubh_u16(uint16_t a, uint16_t b) __NEONSAFE;
uint32_t __simd_vqsubs_u32(uint32_t a, uint32_t b) __NEONSAFE;
uint64x1_t __simd_vqsub_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64_t __simd_vqsubd_u64(uint64_t a, uint64_t b) __NEONSAFE;
uint64x1_t __simd_vadd_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int64x1_t __simd_vadd_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vaddd_u64(uint64_t a, uint64_t b) __NEONSAFE;
int64_t __simd_vaddd_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vsub_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int64x1_t __simd_vsub_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vsubd_u64(uint64_t a, uint64_t b) __NEONSAFE;
int64_t __simd_vsubd_s64(int64_t a, int64_t b) __NEONSAFE;
int16_t __simd_vqdmulhh_s16(int16_t a, int16_t b) __NEONSAFE;
int32_t __simd_vqdmulhs_s32(int32_t a, int32_t b) __NEONSAFE;
int16_t __simd_vqrdmulhh_s16(int16_t a, int16_t b) __NEONSAFE;
int32_t __simd_vqrdmulhs_s32(int32_t a, int32_t b) __NEONSAFE;
int64x1_t __simd_vshl_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
int64_t __simd_vshld_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vshl_u64(uint64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vshld_u64(uint64_t a, int64_t b) __NEONSAFE;
int64x1_t __simd_vrshl_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
int64_t __simd_vrshld_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vrshl_u64(uint64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vrshld_u64(uint64_t a, int64_t b) __NEONSAFE;
int8_t __simd_vqshlb_s8(int8_t a, int8_t b) __NEONSAFE;
int16_t __simd_vqshlh_s16(int16_t a, int16_t b) __NEONSAFE;
int32_t __simd_vqshls_s32(int32_t a, int32_t b) __NEONSAFE;
int64x1_t __simd_vqshl_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
int64_t __simd_vqshld_s64(int64_t a, int64_t b) __NEONSAFE;
uint8_t __simd_vqshlb_u8(uint8_t a, int8_t b) __NEONSAFE;
uint16_t __simd_vqshlh_u16(uint16_t a, int16_t b) __NEONSAFE;
uint32_t __simd_vqshls_u32(uint32_t a, int32_t b) __NEONSAFE;
uint64x1_t __simd_vqshl_u64(uint64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vqshld_u64(uint64_t a, int64_t b) __NEONSAFE;
int8_t __simd_vqrshlb_s8(int8_t a, int8_t b) __NEONSAFE;
int16_t __simd_vqrshlh_s16(int16_t a, int16_t b) __NEONSAFE;
int32_t __simd_vqrshls_s32(int32_t a, int32_t b) __NEONSAFE;
int64x1_t __simd_vqrshl_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
int64_t __simd_vqrshld_s64(int64_t a, int64_t b) __NEONSAFE;
uint8_t __simd_vqrshlb_u8(uint8_t a, int8_t b) __NEONSAFE;
uint16_t __simd_vqrshlh_u16(uint16_t a, int16_t b) __NEONSAFE;
uint32_t __simd_vqrshls_u32(uint32_t a, int32_t b) __NEONSAFE;
uint64x1_t __simd_vqrshl_u64(uint64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vqrshld_u64(uint64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vceq_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64x1_t __simd_vceq_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vceqd_u64(uint64_t a, uint64_t b) __NEONSAFE;
uint64_t __simd_vceqd_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vcge_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vcged_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vcgt_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vcgtd_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vcge_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64_t __simd_vcged_u64(uint64_t a, uint64_t b) __NEONSAFE;
uint64x1_t __simd_vcgt_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64_t __simd_vcgtd_u64(uint64_t a, uint64_t b) __NEONSAFE;
uint64x1_t __simd_vtst_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64x1_t __simd_vtst_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vtstd_u64(uint64_t a, uint64_t b) __NEONSAFE;
uint64_t __simd_vtstd_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vcle_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vcled_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vclt_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vcltd_s64(int64_t a, int64_t b) __NEONSAFE;
uint64x1_t __simd_vcle_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64_t __simd_vcled_u64(uint64_t a, uint64_t b) __NEONSAFE;
uint64x1_t __simd_vclt_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
uint64_t __simd_vcltd_u64(uint64_t a, uint64_t b) __NEONSAFE;
float32x2_t __simd_vadd_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vaddq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vaddq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vpadd_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vpaddq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vpaddq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vsub_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vsubq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vsubq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vmul_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vmulq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vmulq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vmulx_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vmulxq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vmulxq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vfma_f32(float32x2_t a, float32x2_t b, float32x2_t c) __NEONSAFE;
float32x4_t __simd_vfmaq_f32(float32x4_t a, float32x4_t b, float32x4_t c) __NEONSAFE;
float64x2_t __simd_vfmaq_f64(float64x2_t a, float64x2_t b, float64x2_t c) __NEONSAFE;
float32x2_t __simd_vfms_f32(float32x2_t a, float32x2_t b, float32x2_t c) __NEONSAFE;
float32x4_t __simd_vfmsq_f32(float32x4_t a, float32x4_t b, float32x4_t c) __NEONSAFE;
float64x2_t __simd_vfmsq_f64(float64x2_t a, float64x2_t b, float64x2_t c) __NEONSAFE;
float32x2_t __simd_vdiv_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vdivq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vdivq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vceq_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vceqq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vceqq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vcge_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcgeq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcgeq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vcgt_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcgtq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcgtq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vcage_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcageq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcageq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vcagt_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcagtq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcagtq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vcle_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcleq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcleq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vclt_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcltq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcltq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vcale_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcaleq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcaleq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint32x2_t __simd_vcalt_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint32x4_t __simd_vcaltq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vcaltq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vrecps_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vrecpsq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vrecpsq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vrsqrts_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vrsqrtsq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vrsqrtsq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vabd_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vabdq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vabdq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vmax_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vmaxq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vmaxq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vmaxnm_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vmaxnmq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vmaxnmq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vmin_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vminq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vminq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vminnm_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vminnmq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vminnmq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vpmax_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vpmaxq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vpmaxq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vpmaxnm_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vpmaxnmq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vpmaxnmq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vpmin_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vpminq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vpminq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vpminnm_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
float32x4_t __simd_vpminnmq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
float64x2_t __simd_vpminnmq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
float32_t __simd_vmulxs_f32(float32_t a, float32_t b) __NEONSAFE;
float64_t __simd_vmulxd_f64(float64_t a, float64_t b) __NEONSAFE;
float32_t __simd_vabds_f32(float32_t a, float32_t b) __NEONSAFE;
float64_t __simd_vabdd_f64(float64_t a, float64_t b) __NEONSAFE;
float32_t __simd_vrecpss_f32(float32_t a, float32_t b) __NEONSAFE;
float64_t __simd_vrecpsd_f64(float64_t a, float64_t b) __NEONSAFE;
float32_t __simd_vrsqrtss_f32(float32_t a, float32_t b) __NEONSAFE;
float64_t __simd_vrsqrtsd_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vceqs_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vceqd_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vcges_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcged_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vcgts_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcgtd_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vcages_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcaged_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vcagts_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcagtd_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vcles_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcled_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vclts_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcltd_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vcales_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcaled_f64(float64_t a, float64_t b) __NEONSAFE;
uint32_t __simd_vcalts_f32(float32_t a, float32_t b) __NEONSAFE;
uint64_t __simd_vcaltd_f64(float64_t a, float64_t b) __NEONSAFE;
int16x8_t __simd_vaddl_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int32x4_t __simd_vaddl_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int64x2_t __simd_vaddl_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vaddl_high_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int32x4_t __simd_vaddl_high_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vaddl_high_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint16x8_t __simd_vaddl_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint32x4_t __simd_vaddl_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint64x2_t __simd_vaddl_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint16x8_t __simd_vaddl_high_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint32x4_t __simd_vaddl_high_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint64x2_t __simd_vaddl_high_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int16x8_t __simd_vaddw_s8(int16x8_t a, int8x8_t b) __NEONSAFE;
int32x4_t __simd_vaddw_s16(int32x4_t a, int16x4_t b) __NEONSAFE;
int64x2_t __simd_vaddw_s32(int64x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vaddw_high_s8(int16x8_t a, int8x16_t b) __NEONSAFE;
int32x4_t __simd_vaddw_high_s16(int32x4_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vaddw_high_s32(int64x2_t a, int32x4_t b) __NEONSAFE;
uint16x8_t __simd_vaddw_u8(uint16x8_t a, uint8x8_t b) __NEONSAFE;
uint32x4_t __simd_vaddw_u16(uint32x4_t a, uint16x4_t b) __NEONSAFE;
uint64x2_t __simd_vaddw_u32(uint64x2_t a, uint32x2_t b) __NEONSAFE;
uint16x8_t __simd_vaddw_high_u8(uint16x8_t a, uint8x16_t b) __NEONSAFE;
uint32x4_t __simd_vaddw_high_u16(uint32x4_t a, uint16x8_t b) __NEONSAFE;
uint64x2_t __simd_vaddw_high_u32(uint64x2_t a, uint32x4_t b) __NEONSAFE;
int16x8_t __simd_vsubl_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int32x4_t __simd_vsubl_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int64x2_t __simd_vsubl_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vsubl_high_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int32x4_t __simd_vsubl_high_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vsubl_high_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint16x8_t __simd_vsubl_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint32x4_t __simd_vsubl_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint64x2_t __simd_vsubl_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint16x8_t __simd_vsubl_high_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint32x4_t __simd_vsubl_high_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint64x2_t __simd_vsubl_high_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int16x8_t __simd_vsubw_s8(int16x8_t a, int8x8_t b) __NEONSAFE;
int32x4_t __simd_vsubw_s16(int32x4_t a, int16x4_t b) __NEONSAFE;
int64x2_t __simd_vsubw_s32(int64x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vsubw_high_s8(int16x8_t a, int8x16_t b) __NEONSAFE;
int32x4_t __simd_vsubw_high_s16(int32x4_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vsubw_high_s32(int64x2_t a, int32x4_t b) __NEONSAFE;
uint16x8_t __simd_vsubw_u8(uint16x8_t a, uint8x8_t b) __NEONSAFE;
uint32x4_t __simd_vsubw_u16(uint32x4_t a, uint16x4_t b) __NEONSAFE;
uint64x2_t __simd_vsubw_u32(uint64x2_t a, uint32x2_t b) __NEONSAFE;
uint16x8_t __simd_vsubw_high_u8(uint16x8_t a, uint8x16_t b) __NEONSAFE;
uint32x4_t __simd_vsubw_high_u16(uint32x4_t a, uint16x8_t b) __NEONSAFE;
uint64x2_t __simd_vsubw_high_u32(uint64x2_t a, uint32x4_t b) __NEONSAFE;
uint8x8_t __simd_vaddhn_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int8x8_t __simd_vaddhn_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint16x4_t __simd_vaddhn_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int16x4_t __simd_vaddhn_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint32x2_t __simd_vaddhn_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int32x2_t __simd_vaddhn_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x16_t __simd_vaddhn_high_u16(uint8x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
int8x16_t __simd_vaddhn_high_s16(int8x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
uint16x8_t __simd_vaddhn_high_u32(uint16x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int16x8_t __simd_vaddhn_high_s32(int16x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint32x4_t __simd_vaddhn_high_u64(uint32x2_t a, uint64x2_t b, uint64x2_t c) __NEONSAFE;
int32x4_t __simd_vaddhn_high_s64(int32x2_t a, int64x2_t b, int64x2_t c) __NEONSAFE;
uint8x8_t __simd_vraddhn_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int8x8_t __simd_vraddhn_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint16x4_t __simd_vraddhn_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int16x4_t __simd_vraddhn_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint32x2_t __simd_vraddhn_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int32x2_t __simd_vraddhn_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x16_t __simd_vraddhn_high_u16(uint8x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
int8x16_t __simd_vraddhn_high_s16(int8x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
uint16x8_t __simd_vraddhn_high_u32(uint16x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int16x8_t __simd_vraddhn_high_s32(int16x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint32x4_t __simd_vraddhn_high_u64(uint32x2_t a, uint64x2_t b, uint64x2_t c) __NEONSAFE;
int32x4_t __simd_vraddhn_high_s64(int32x2_t a, int64x2_t b, int64x2_t c) __NEONSAFE;
uint8x8_t __simd_vsubhn_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int8x8_t __simd_vsubhn_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint16x4_t __simd_vsubhn_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int16x4_t __simd_vsubhn_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint32x2_t __simd_vsubhn_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int32x2_t __simd_vsubhn_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x16_t __simd_vsubhn_high_u16(uint8x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
int8x16_t __simd_vsubhn_high_s16(int8x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
uint16x8_t __simd_vsubhn_high_u32(uint16x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int16x8_t __simd_vsubhn_high_s32(int16x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint32x4_t __simd_vsubhn_high_u64(uint32x2_t a, uint64x2_t b, uint64x2_t c) __NEONSAFE;
int32x4_t __simd_vsubhn_high_s64(int32x2_t a, int64x2_t b, int64x2_t c) __NEONSAFE;
uint8x8_t __simd_vrsubhn_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int8x8_t __simd_vrsubhn_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
uint16x4_t __simd_vrsubhn_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int16x4_t __simd_vrsubhn_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint32x2_t __simd_vrsubhn_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int32x2_t __simd_vrsubhn_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
uint8x16_t __simd_vrsubhn_high_u16(uint8x8_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
int8x16_t __simd_vrsubhn_high_s16(int8x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
uint16x8_t __simd_vrsubhn_high_u32(uint16x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int16x8_t __simd_vrsubhn_high_s32(int16x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint32x4_t __simd_vrsubhn_high_u64(uint32x2_t a, uint64x2_t b, uint64x2_t c) __NEONSAFE;
int32x4_t __simd_vrsubhn_high_s64(int32x2_t a, int64x2_t b, int64x2_t c) __NEONSAFE;
int16x8_t __simd_vabal_s8(int16x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
int32x4_t __simd_vabal_s16(int32x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int64x2_t __simd_vabal_s32(int64x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int16x8_t __simd_vabal_high_s8(int16x8_t a, int8x16_t b, int8x16_t c) __NEONSAFE;
int32x4_t __simd_vabal_high_s16(int32x4_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int64x2_t __simd_vabal_high_s32(int64x2_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint16x8_t __simd_vabal_u8(uint16x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
uint32x4_t __simd_vabal_u16(uint32x4_t a, uint16x4_t b, uint16x4_t c) __NEONSAFE;
uint64x2_t __simd_vabal_u32(uint64x2_t a, uint32x2_t b, uint32x2_t c) __NEONSAFE;
uint16x8_t __simd_vabal_high_u8(uint16x8_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
uint32x4_t __simd_vabal_high_u16(uint32x4_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
uint64x2_t __simd_vabal_high_u32(uint64x2_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int16x8_t __simd_vabdl_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int32x4_t __simd_vabdl_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int64x2_t __simd_vabdl_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vabdl_high_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int32x4_t __simd_vabdl_high_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vabdl_high_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint16x8_t __simd_vabdl_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint32x4_t __simd_vabdl_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint64x2_t __simd_vabdl_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint16x8_t __simd_vabdl_high_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint32x4_t __simd_vabdl_high_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint64x2_t __simd_vabdl_high_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int16x8_t __simd_vmlal_s8(int16x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
int32x4_t __simd_vmlal_s16(int32x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int64x2_t __simd_vmlal_s32(int64x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int16x8_t __simd_vmlal_high_s8(int16x8_t a, int8x16_t b, int8x16_t c) __NEONSAFE;
int32x4_t __simd_vmlal_high_s16(int32x4_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int64x2_t __simd_vmlal_high_s32(int64x2_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint16x8_t __simd_vmlal_u8(uint16x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
uint32x4_t __simd_vmlal_u16(uint32x4_t a, uint16x4_t b, uint16x4_t c) __NEONSAFE;
uint64x2_t __simd_vmlal_u32(uint64x2_t a, uint32x2_t b, uint32x2_t c) __NEONSAFE;
uint16x8_t __simd_vmlal_high_u8(uint16x8_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
uint32x4_t __simd_vmlal_high_u16(uint32x4_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
uint64x2_t __simd_vmlal_high_u32(uint64x2_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int16x8_t __simd_vmlsl_s8(int16x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
int32x4_t __simd_vmlsl_s16(int32x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int64x2_t __simd_vmlsl_s32(int64x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int16x8_t __simd_vmlsl_high_s8(int16x8_t a, int8x16_t b, int8x16_t c) __NEONSAFE;
int32x4_t __simd_vmlsl_high_s16(int32x4_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int64x2_t __simd_vmlsl_high_s32(int64x2_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
uint16x8_t __simd_vmlsl_u8(uint16x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
uint32x4_t __simd_vmlsl_u16(uint32x4_t a, uint16x4_t b, uint16x4_t c) __NEONSAFE;
uint64x2_t __simd_vmlsl_u32(uint64x2_t a, uint32x2_t b, uint32x2_t c) __NEONSAFE;
uint16x8_t __simd_vmlsl_high_u8(uint16x8_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
uint32x4_t __simd_vmlsl_high_u16(uint32x4_t a, uint16x8_t b, uint16x8_t c) __NEONSAFE;
uint64x2_t __simd_vmlsl_high_u32(uint64x2_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
int16x8_t __simd_vmull_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
int32x4_t __simd_vmull_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int64x2_t __simd_vmull_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vmull_high_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
int32x4_t __simd_vmull_high_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vmull_high_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
uint16x8_t __simd_vmull_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
uint32x4_t __simd_vmull_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
uint64x2_t __simd_vmull_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
uint16x8_t __simd_vmull_high_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint32x4_t __simd_vmull_high_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
uint64x2_t __simd_vmull_high_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vqdmlal_s16(int32x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int64x2_t __simd_vqdmlal_s32(int64x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int32x4_t __simd_vqdmlal_high_s16(int32x4_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int64x2_t __simd_vqdmlal_high_s32(int64x2_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
int32x4_t __simd_vqdmlsl_s16(int32x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int64x2_t __simd_vqdmlsl_s32(int64x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int32x4_t __simd_vqdmlsl_high_s16(int32x4_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int64x2_t __simd_vqdmlsl_high_s32(int64x2_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
int32x4_t __simd_vqdmull_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
int64x2_t __simd_vqdmull_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
int32x4_t __simd_vqdmull_high_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vqdmull_high_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
poly16x8_t __simd_vmull_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
poly16x8_t __simd_vmull_high_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
int32_t __simd_vqdmlalh_s16(int32_t a, int16_t b, int16_t c) __NEONSAFE;
int64_t __simd_vqdmlals_s32(int64_t a, int32_t b, int32_t c) __NEONSAFE;
int32_t __simd_vqdmlslh_s16(int32_t a, int16_t b, int16_t c) __NEONSAFE;
int64_t __simd_vqdmlsls_s32(int64_t a, int32_t b, int32_t c) __NEONSAFE;
int32_t __simd_vqdmullh_s16(int16_t a, int16_t b) __NEONSAFE;
int64_t __simd_vqdmulls_s32(int32_t a, int32_t b) __NEONSAFE;
int32x4_t __simd_vmlal_lane_s16(int32x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlal_laneq_s16(int32x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlal_n_s16(int32x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vmlal_lane_s32(int64x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlal_laneq_s32(int64x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlal_n_s32(int64x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
int32x4_t __simd_vmlal_high_lane_s16(int32x4_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlal_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlal_high_n_s16(int32x4_t a, int16x8_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vmlal_high_lane_s32(int64x2_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlal_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlal_high_n_s32(int64x2_t a, int32x4_t b, int32_t c) __NEONSAFE;
uint32x4_t __simd_vmlal_lane_u16(uint32x4_t a, uint16x4_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlal_laneq_u16(uint32x4_t a, uint16x4_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlal_n_u16(uint32x4_t a, uint16x4_t b, uint16_t c) __NEONSAFE;
uint64x2_t __simd_vmlal_lane_u32(uint64x2_t a, uint32x2_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlal_laneq_u32(uint64x2_t a, uint32x2_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlal_n_u32(uint64x2_t a, uint32x2_t b, uint32_t c) __NEONSAFE;
uint32x4_t __simd_vmlal_high_lane_u16(uint32x4_t a, uint16x8_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlal_high_laneq_u16(uint32x4_t a, uint16x8_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlal_high_n_u16(uint32x4_t a, uint16x8_t b, uint16_t c) __NEONSAFE;
uint64x2_t __simd_vmlal_high_lane_u32(uint64x2_t a, uint32x4_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlal_high_laneq_u32(uint64x2_t a, uint32x4_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlal_high_n_u32(uint64x2_t a, uint32x4_t b, uint32_t c) __NEONSAFE;
int32x4_t __simd_vqdmlal_lane_s16(int32x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlal_laneq_s16(int32x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlal_n_s16(int32x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vqdmlal_lane_s32(int64x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlal_laneq_s32(int64x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlal_n_s32(int64x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
int32x4_t __simd_vqdmlal_high_lane_s16(int32x4_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlal_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlal_high_n_s16(int32x4_t a, int16x8_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vqdmlal_high_lane_s32(int64x2_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlal_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlal_high_n_s32(int64x2_t a, int32x4_t b, int32_t c) __NEONSAFE;
int32x4_t __simd_vmlsl_lane_s16(int32x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlsl_laneq_s16(int32x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlsl_n_s16(int32x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vmlsl_lane_s32(int64x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlsl_laneq_s32(int64x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlsl_n_s32(int64x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
int32x4_t __simd_vmlsl_high_lane_s16(int32x4_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlsl_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlsl_high_n_s16(int32x4_t a, int16x8_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vmlsl_high_lane_s32(int64x2_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlsl_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vmlsl_high_n_s32(int64x2_t a, int32x4_t b, int32_t c) __NEONSAFE;
uint32x4_t __simd_vmlsl_lane_u16(uint32x4_t a, uint16x4_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlsl_laneq_u16(uint32x4_t a, uint16x4_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlsl_n_u16(uint32x4_t a, uint16x4_t b, uint16_t c) __NEONSAFE;
uint64x2_t __simd_vmlsl_lane_u32(uint64x2_t a, uint32x2_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlsl_laneq_u32(uint64x2_t a, uint32x2_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlsl_n_u32(uint64x2_t a, uint32x2_t b, uint32_t c) __NEONSAFE;
uint32x4_t __simd_vmlsl_high_lane_u16(uint32x4_t a, uint16x8_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlsl_high_laneq_u16(uint32x4_t a, uint16x8_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlsl_high_n_u16(uint32x4_t a, uint16x8_t b, uint16_t c) __NEONSAFE;
uint64x2_t __simd_vmlsl_high_lane_u32(uint64x2_t a, uint32x4_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlsl_high_laneq_u32(uint64x2_t a, uint32x4_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vmlsl_high_n_u32(uint64x2_t a, uint32x4_t b, uint32_t c) __NEONSAFE;
int32x4_t __simd_vqdmlsl_lane_s16(int32x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlsl_laneq_s16(int32x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlsl_n_s16(int32x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vqdmlsl_lane_s32(int64x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlsl_laneq_s32(int64x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlsl_n_s32(int64x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
int32x4_t __simd_vqdmlsl_high_lane_s16(int32x4_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlsl_high_laneq_s16(int32x4_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqdmlsl_high_n_s16(int32x4_t a, int16x8_t b, int16_t c) __NEONSAFE;
int64x2_t __simd_vqdmlsl_high_lane_s32(int64x2_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlsl_high_laneq_s32(int64x2_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vqdmlsl_high_n_s32(int64x2_t a, int32x4_t b, int32_t c) __NEONSAFE;
int32x4_t __simd_vmull_lane_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vmull_laneq_s16(int16x4_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vmull_n_s16(int16x4_t a, int16_t b) __NEONSAFE;
int64x2_t __simd_vmull_lane_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vmull_laneq_s32(int32x2_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vmull_n_s32(int32x2_t a, int32_t b) __NEONSAFE;
int32x4_t __simd_vmull_high_lane_s16(int16x8_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vmull_high_laneq_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vmull_high_n_s16(int16x8_t a, int16_t b) __NEONSAFE;
int64x2_t __simd_vmull_high_lane_s32(int32x4_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vmull_high_laneq_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vmull_high_n_s32(int32x4_t a, int32_t b) __NEONSAFE;
uint32x4_t __simd_vmull_lane_u16(uint16x4_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vmull_laneq_u16(uint16x4_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vmull_n_u16(uint16x4_t a, uint16_t b) __NEONSAFE;
uint64x2_t __simd_vmull_lane_u32(uint32x2_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vmull_laneq_u32(uint32x2_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vmull_n_u32(uint32x2_t a, uint32_t b) __NEONSAFE;
uint32x4_t __simd_vmull_high_lane_u16(uint16x8_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vmull_high_laneq_u16(uint16x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vmull_high_n_u16(uint16x8_t a, uint16_t b) __NEONSAFE;
uint64x2_t __simd_vmull_high_lane_u32(uint32x4_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vmull_high_laneq_u32(uint32x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vmull_high_n_u32(uint32x4_t a, uint32_t b) __NEONSAFE;
int32x4_t __simd_vqdmull_lane_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqdmull_laneq_s16(int16x4_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqdmull_n_s16(int16x4_t a, int16_t b) __NEONSAFE;
int64x2_t __simd_vqdmull_lane_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vqdmull_laneq_s32(int32x2_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vqdmull_n_s32(int32x2_t a, int32_t b) __NEONSAFE;
int32x4_t __simd_vqdmull_high_lane_s16(int16x8_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqdmull_high_laneq_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqdmull_high_n_s16(int16x8_t a, int16_t b) __NEONSAFE;
int64x2_t __simd_vqdmull_high_lane_s32(int32x4_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vqdmull_high_laneq_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vqdmull_high_n_s32(int32x4_t a, int32_t b) __NEONSAFE;
uint16x4_t __simd_vmul_lane_u16(uint16x4_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vmul_lane_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vmul_laneq_u16(uint16x4_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vmul_laneq_s16(int16x4_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vmul_n_u16(uint16x4_t a, uint16_t b) __NEONSAFE;
int16x4_t __simd_vmul_n_s16(int16x4_t a, int16_t b) __NEONSAFE;
uint32x2_t __simd_vmul_lane_u32(uint32x2_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vmul_lane_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vmul_laneq_u32(uint32x2_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vmul_laneq_s32(int32x2_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vmul_n_u32(uint32x2_t a, uint32_t b) __NEONSAFE;
int32x2_t __simd_vmul_n_s32(int32x2_t a, int32_t b) __NEONSAFE;
uint16x8_t __simd_vmulq_lane_u16(uint16x8_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vmulq_lane_s16(int16x8_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vmulq_laneq_u16(uint16x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vmulq_laneq_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vmulq_n_u16(uint16x8_t a, uint16_t b) __NEONSAFE;
int16x8_t __simd_vmulq_n_s16(int16x8_t a, int16_t b) __NEONSAFE;
uint32x4_t __simd_vmulq_lane_u32(uint32x4_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vmulq_lane_s32(int32x4_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vmulq_laneq_u32(uint32x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vmulq_laneq_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vmulq_n_u32(uint32x4_t a, uint32_t b) __NEONSAFE;
int32x4_t __simd_vmulq_n_s32(int32x4_t a, int32_t b) __NEONSAFE;
uint16x4_t __simd_vmla_lane_u16(uint16x4_t a, uint16x4_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vmla_lane_s16(int16x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x4_t __simd_vmla_laneq_u16(uint16x4_t a, uint16x4_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vmla_laneq_s16(int16x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x4_t __simd_vmla_n_u16(uint16x4_t a, uint16x4_t b, uint16_t c) __NEONSAFE;
int16x4_t __simd_vmla_n_s16(int16x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
uint32x2_t __simd_vmla_lane_u32(uint32x2_t a, uint32x2_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vmla_lane_s32(int32x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x2_t __simd_vmla_laneq_u32(uint32x2_t a, uint32x2_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vmla_laneq_s32(int32x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x2_t __simd_vmla_n_u32(uint32x2_t a, uint32x2_t b, uint32_t c) __NEONSAFE;
int32x2_t __simd_vmla_n_s32(int32x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
uint16x8_t __simd_vmlaq_lane_u16(uint16x8_t a, uint16x8_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vmlaq_lane_s16(int16x8_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x8_t __simd_vmlaq_laneq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vmlaq_laneq_s16(int16x8_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x8_t __simd_vmlaq_n_u16(uint16x8_t a, uint16x8_t b, uint16_t c) __NEONSAFE;
int16x8_t __simd_vmlaq_n_s16(int16x8_t a, int16x8_t b, int16_t c) __NEONSAFE;
uint32x4_t __simd_vmlaq_lane_u32(uint32x4_t a, uint32x4_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlaq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlaq_laneq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlaq_laneq_s32(int32x4_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlaq_n_u32(uint32x4_t a, uint32x4_t b, uint32_t c) __NEONSAFE;
int32x4_t __simd_vmlaq_n_s32(int32x4_t a, int32x4_t b, int32_t c) __NEONSAFE;
uint16x4_t __simd_vmls_lane_u16(uint16x4_t a, uint16x4_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vmls_lane_s16(int16x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x4_t __simd_vmls_laneq_u16(uint16x4_t a, uint16x4_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vmls_laneq_s16(int16x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x4_t __simd_vmls_n_u16(uint16x4_t a, uint16x4_t b, uint16_t c) __NEONSAFE;
int16x4_t __simd_vmls_n_s16(int16x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
uint32x2_t __simd_vmls_lane_u32(uint32x2_t a, uint32x2_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vmls_lane_s32(int32x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x2_t __simd_vmls_laneq_u32(uint32x2_t a, uint32x2_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vmls_laneq_s32(int32x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x2_t __simd_vmls_n_u32(uint32x2_t a, uint32x2_t b, uint32_t c) __NEONSAFE;
int32x2_t __simd_vmls_n_s32(int32x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
uint16x8_t __simd_vmlsq_lane_u16(uint16x8_t a, uint16x8_t b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vmlsq_lane_s16(int16x8_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x8_t __simd_vmlsq_laneq_u16(uint16x8_t a, uint16x8_t b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vmlsq_laneq_s16(int16x8_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x8_t __simd_vmlsq_n_u16(uint16x8_t a, uint16x8_t b, uint16_t c) __NEONSAFE;
int16x8_t __simd_vmlsq_n_s16(int16x8_t a, int16x8_t b, int16_t c) __NEONSAFE;
uint32x4_t __simd_vmlsq_lane_u32(uint32x4_t a, uint32x4_t b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlsq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlsq_laneq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vmlsq_laneq_s32(int32x4_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vmlsq_n_u32(uint32x4_t a, uint32x4_t b, uint32_t c) __NEONSAFE;
int32x4_t __simd_vmlsq_n_s32(int32x4_t a, int32x4_t b, int32_t c) __NEONSAFE;
int16x4_t __simd_vqdmulh_lane_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vqdmulh_laneq_s16(int16x4_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vqdmulh_n_s16(int16x4_t a, int16_t b) __NEONSAFE;
int32x2_t __simd_vqdmulh_lane_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vqdmulh_laneq_s32(int32x2_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vqdmulh_n_s32(int32x2_t a, int32_t b) __NEONSAFE;
int16x8_t __simd_vqdmulhq_lane_s16(int16x8_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vqdmulhq_laneq_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vqdmulhq_n_s16(int16x8_t a, int16_t b) __NEONSAFE;
int32x4_t __simd_vqdmulhq_lane_s32(int32x4_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqdmulhq_laneq_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqdmulhq_n_s32(int32x4_t a, int32_t b) __NEONSAFE;
int16x4_t __simd_vqrdmulh_lane_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vqrdmulh_laneq_s16(int16x4_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vqrdmulh_n_s16(int16x4_t a, int16_t b) __NEONSAFE;
int32x2_t __simd_vqrdmulh_lane_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vqrdmulh_laneq_s32(int32x2_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vqrdmulh_n_s32(int32x2_t a, int32_t b) __NEONSAFE;
int16x8_t __simd_vqrdmulhq_lane_s16(int16x8_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vqrdmulhq_laneq_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vqrdmulhq_n_s16(int16x8_t a, int16_t b) __NEONSAFE;
int32x4_t __simd_vqrdmulhq_lane_s32(int32x4_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqrdmulhq_laneq_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqrdmulhq_n_s32(int32x4_t a, int32_t b) __NEONSAFE;
float32x2_t __simd_vmul_lane_f32(float32x2_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32x2_t __simd_vmul_laneq_f32(float32x2_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float32x2_t __simd_vmul_n_f32(float32x2_t a, float32_t b) __NEONSAFE;
float32x4_t __simd_vmulq_lane_f32(float32x4_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32x4_t __simd_vmulq_laneq_f32(float32x4_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float32x4_t __simd_vmulq_n_f32(float32x4_t a, float32_t b) __NEONSAFE;
float64x2_t __simd_vmulq_laneq_f64(float64x2_t a, float64x2_t b, __ghs_c_int__ c) __NEONSAFE;
float64x2_t __simd_vmulq_n_f64(float64x2_t a, float64_t b) __NEONSAFE;
float32x2_t __simd_vfma_lane_f32(float32x2_t a, float32x2_t b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32x2_t __simd_vfma_laneq_f32(float32x2_t a, float32x2_t b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float32x2_t __simd_vfma_n_f32(float32x2_t a, float32x2_t b, float32_t c) __NEONSAFE;
float32x4_t __simd_vfmaq_lane_f32(float32x4_t a, float32x4_t b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32x4_t __simd_vfmaq_laneq_f32(float32x4_t a, float32x4_t b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float32x4_t __simd_vfmaq_n_f32(float32x4_t a, float32x4_t b, float32_t c) __NEONSAFE;
float64x2_t __simd_vfmaq_laneq_f64(float64x2_t a, float64x2_t b, float64x2_t c, __ghs_c_int__ d) __NEONSAFE;
float64x2_t __simd_vfmaq_n_f64(float64x2_t a, float64x2_t b, float64_t c) __NEONSAFE;
float32x2_t __simd_vfms_lane_f32(float32x2_t a, float32x2_t b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32x2_t __simd_vfms_laneq_f32(float32x2_t a, float32x2_t b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float32x2_t __simd_vfms_n_f32(float32x2_t a, float32x2_t b, float32_t c) __NEONSAFE;
float32x4_t __simd_vfmsq_lane_f32(float32x4_t a, float32x4_t b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32x4_t __simd_vfmsq_laneq_f32(float32x4_t a, float32x4_t b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float32x4_t __simd_vfmsq_n_f32(float32x4_t a, float32x4_t b, float32_t c) __NEONSAFE;
float64x2_t __simd_vfmsq_laneq_f64(float64x2_t a, float64x2_t b, float64x2_t c, __ghs_c_int__ d) __NEONSAFE;
float64x2_t __simd_vfmsq_n_f64(float64x2_t a, float64x2_t b, float64_t c) __NEONSAFE;
float32x2_t __simd_vmulx_lane_f32(float32x2_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32x2_t __simd_vmulx_laneq_f32(float32x2_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float32x2_t __simd_vmulx_n_f32(float32x2_t a, float32_t b) __NEONSAFE;
float32x4_t __simd_vmulxq_lane_f32(float32x4_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32x4_t __simd_vmulxq_laneq_f32(float32x4_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float32x4_t __simd_vmulxq_n_f32(float32x4_t a, float32_t b) __NEONSAFE;
float64x2_t __simd_vmulxq_laneq_f64(float64x2_t a, float64x2_t b, __ghs_c_int__ c) __NEONSAFE;
float64x2_t __simd_vmulxq_n_f64(float64x2_t a, float64_t b) __NEONSAFE;
int32_t __simd_vqdmlalh_lane_s16(int32_t a, int16_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqdmlalh_laneq_s16(int32_t a, int16_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int64_t __simd_vqdmlals_lane_s32(int64_t a, int32_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64_t __simd_vqdmlals_laneq_s32(int64_t a, int32_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqdmlslh_lane_s16(int32_t a, int16_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqdmlslh_laneq_s16(int32_t a, int16_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int64_t __simd_vqdmlsls_lane_s32(int64_t a, int32_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64_t __simd_vqdmlsls_laneq_s32(int64_t a, int32_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqdmullh_lane_s16(int16_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32_t __simd_vqdmullh_laneq_s16(int16_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int64_t __simd_vqdmulls_lane_s32(int32_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64_t __simd_vqdmulls_laneq_s32(int32_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16_t __simd_vqdmulhh_lane_s16(int16_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16_t __simd_vqdmulhh_laneq_s16(int16_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32_t __simd_vqdmulhs_lane_s32(int32_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32_t __simd_vqdmulhs_laneq_s32(int32_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16_t __simd_vqrdmulhh_lane_s16(int16_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16_t __simd_vqrdmulhh_laneq_s16(int16_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32_t __simd_vqrdmulhs_lane_s32(int32_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32_t __simd_vqrdmulhs_laneq_s32(int32_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float32_t __simd_vfmas_lane_f32(float32_t a, float32_t b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32_t __simd_vfmas_laneq_f32(float32_t a, float32_t b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float64_t __simd_vfmad_laneq_f64(float64_t a, float64_t b, float64x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32_t __simd_vfmss_lane_f32(float32_t a, float32_t b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32_t __simd_vfmss_laneq_f32(float32_t a, float32_t b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float64_t __simd_vfmsd_laneq_f64(float64_t a, float64_t b, float64x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32_t __simd_vmuls_lane_f32(float32_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32_t __simd_vmuls_laneq_f32(float32_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float64_t __simd_vmuld_laneq_f64(float64_t a, float64x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32_t __simd_vmulxs_lane_f32(float32_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32_t __simd_vmulxs_laneq_f32(float32_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float64_t __simd_vmulxd_laneq_f64(float64_t a, float64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vrev64_u8(uint8x8_t a) __NEONSAFE;
int8x8_t __simd_vrev64_s8(int8x8_t a) __NEONSAFE;
poly8x8_t __simd_vrev64_p8(poly8x8_t a) __NEONSAFE;
uint16x4_t __simd_vrev64_u16(uint16x4_t a) __NEONSAFE;
int16x4_t __simd_vrev64_s16(int16x4_t a) __NEONSAFE;
poly16x4_t __simd_vrev64_p16(poly16x4_t a) __NEONSAFE;
uint32x2_t __simd_vrev64_u32(uint32x2_t a) __NEONSAFE;
int32x2_t __simd_vrev64_s32(int32x2_t a) __NEONSAFE;
float32x2_t __simd_vrev64_f32(float32x2_t a) __NEONSAFE;
uint8x16_t __simd_vrev64q_u8(uint8x16_t a) __NEONSAFE;
int8x16_t __simd_vrev64q_s8(int8x16_t a) __NEONSAFE;
poly8x16_t __simd_vrev64q_p8(poly8x16_t a) __NEONSAFE;
uint16x8_t __simd_vrev64q_u16(uint16x8_t a) __NEONSAFE;
int16x8_t __simd_vrev64q_s16(int16x8_t a) __NEONSAFE;
poly16x8_t __simd_vrev64q_p16(poly16x8_t a) __NEONSAFE;
uint32x4_t __simd_vrev64q_u32(uint32x4_t a) __NEONSAFE;
int32x4_t __simd_vrev64q_s32(int32x4_t a) __NEONSAFE;
float32x4_t __simd_vrev64q_f32(float32x4_t a) __NEONSAFE;
uint8x8_t __simd_vrev32_u8(uint8x8_t a) __NEONSAFE;
int8x8_t __simd_vrev32_s8(int8x8_t a) __NEONSAFE;
poly8x8_t __simd_vrev32_p8(poly8x8_t a) __NEONSAFE;
uint16x4_t __simd_vrev32_u16(uint16x4_t a) __NEONSAFE;
int16x4_t __simd_vrev32_s16(int16x4_t a) __NEONSAFE;
poly16x4_t __simd_vrev32_p16(poly16x4_t a) __NEONSAFE;
uint8x16_t __simd_vrev32q_u8(uint8x16_t a) __NEONSAFE;
int8x16_t __simd_vrev32q_s8(int8x16_t a) __NEONSAFE;
poly8x16_t __simd_vrev32q_p8(poly8x16_t a) __NEONSAFE;
uint16x8_t __simd_vrev32q_u16(uint16x8_t a) __NEONSAFE;
int16x8_t __simd_vrev32q_s16(int16x8_t a) __NEONSAFE;
poly16x8_t __simd_vrev32q_p16(poly16x8_t a) __NEONSAFE;
uint8x8_t __simd_vrev16_u8(uint8x8_t a) __NEONSAFE;
int8x8_t __simd_vrev16_s8(int8x8_t a) __NEONSAFE;
poly8x8_t __simd_vrev16_p8(poly8x8_t a) __NEONSAFE;
uint8x16_t __simd_vrev16q_u8(uint8x16_t a) __NEONSAFE;
int8x16_t __simd_vrev16q_s8(int8x16_t a) __NEONSAFE;
poly8x16_t __simd_vrev16q_p8(poly8x16_t a) __NEONSAFE;
int16x4_t __simd_vpaddl_s8(int8x8_t a) __NEONSAFE;
int32x2_t __simd_vpaddl_s16(int16x4_t a) __NEONSAFE;
int64x1_t __simd_vpaddl_s32(int32x2_t a) __NEONSAFE;
int16x8_t __simd_vpaddlq_s8(int8x16_t a) __NEONSAFE;
int32x4_t __simd_vpaddlq_s16(int16x8_t a) __NEONSAFE;
int64x2_t __simd_vpaddlq_s32(int32x4_t a) __NEONSAFE;
uint16x4_t __simd_vpaddl_u8(uint8x8_t a) __NEONSAFE;
uint32x2_t __simd_vpaddl_u16(uint16x4_t a) __NEONSAFE;
uint64x1_t __simd_vpaddl_u32(uint32x2_t a) __NEONSAFE;
uint16x8_t __simd_vpaddlq_u8(uint8x16_t a) __NEONSAFE;
uint32x4_t __simd_vpaddlq_u16(uint16x8_t a) __NEONSAFE;
uint64x2_t __simd_vpaddlq_u32(uint32x4_t a) __NEONSAFE;
int16x4_t __simd_vpadal_s8(int16x4_t a, int8x8_t b) __NEONSAFE;
int32x2_t __simd_vpadal_s16(int32x2_t a, int16x4_t b) __NEONSAFE;
int64x1_t __simd_vpadal_s32(int64x1_t a, int32x2_t b) __NEONSAFE;
int16x8_t __simd_vpadalq_s8(int16x8_t a, int8x16_t b) __NEONSAFE;
int32x4_t __simd_vpadalq_s16(int32x4_t a, int16x8_t b) __NEONSAFE;
int64x2_t __simd_vpadalq_s32(int64x2_t a, int32x4_t b) __NEONSAFE;
uint16x4_t __simd_vpadal_u8(uint16x4_t a, uint8x8_t b) __NEONSAFE;
uint32x2_t __simd_vpadal_u16(uint32x2_t a, uint16x4_t b) __NEONSAFE;
uint64x1_t __simd_vpadal_u32(uint64x1_t a, uint32x2_t b) __NEONSAFE;
uint16x8_t __simd_vpadalq_u8(uint16x8_t a, uint8x16_t b) __NEONSAFE;
uint32x4_t __simd_vpadalq_u16(uint32x4_t a, uint16x8_t b) __NEONSAFE;
uint64x2_t __simd_vpadalq_u32(uint64x2_t a, uint32x4_t b) __NEONSAFE;
int8x8_t __simd_vuqadd_s8(int8x8_t a, uint8x8_t b) __NEONSAFE;
int16x4_t __simd_vuqadd_s16(int16x4_t a, uint16x4_t b) __NEONSAFE;
int32x2_t __simd_vuqadd_s32(int32x2_t a, uint32x2_t b) __NEONSAFE;
int8x16_t __simd_vuqaddq_s8(int8x16_t a, uint8x16_t b) __NEONSAFE;
int16x8_t __simd_vuqaddq_s16(int16x8_t a, uint16x8_t b) __NEONSAFE;
int32x4_t __simd_vuqaddq_s32(int32x4_t a, uint32x4_t b) __NEONSAFE;
int64x2_t __simd_vuqaddq_s64(int64x2_t a, uint64x2_t b) __NEONSAFE;
uint8x8_t __simd_vsqadd_u8(uint8x8_t a, int8x8_t b) __NEONSAFE;
uint16x4_t __simd_vsqadd_u16(uint16x4_t a, int16x4_t b) __NEONSAFE;
uint32x2_t __simd_vsqadd_u32(uint32x2_t a, int32x2_t b) __NEONSAFE;
uint8x16_t __simd_vsqaddq_u8(uint8x16_t a, int8x16_t b) __NEONSAFE;
uint16x8_t __simd_vsqaddq_u16(uint16x8_t a, int16x8_t b) __NEONSAFE;
uint32x4_t __simd_vsqaddq_u32(uint32x4_t a, int32x4_t b) __NEONSAFE;
uint64x2_t __simd_vsqaddq_u64(uint64x2_t a, int64x2_t b) __NEONSAFE;
int8x8_t __simd_vcls_s8(int8x8_t a) __NEONSAFE;
int16x4_t __simd_vcls_s16(int16x4_t a) __NEONSAFE;
int32x2_t __simd_vcls_s32(int32x2_t a) __NEONSAFE;
int8x16_t __simd_vclsq_s8(int8x16_t a) __NEONSAFE;
int16x8_t __simd_vclsq_s16(int16x8_t a) __NEONSAFE;
int32x4_t __simd_vclsq_s32(int32x4_t a) __NEONSAFE;
uint8x8_t __simd_vclz_u8(uint8x8_t a) __NEONSAFE;
int8x8_t __simd_vclz_s8(int8x8_t a) __NEONSAFE;
uint16x4_t __simd_vclz_u16(uint16x4_t a) __NEONSAFE;
int16x4_t __simd_vclz_s16(int16x4_t a) __NEONSAFE;
uint32x2_t __simd_vclz_u32(uint32x2_t a) __NEONSAFE;
int32x2_t __simd_vclz_s32(int32x2_t a) __NEONSAFE;
uint8x16_t __simd_vclzq_u8(uint8x16_t a) __NEONSAFE;
int8x16_t __simd_vclzq_s8(int8x16_t a) __NEONSAFE;
uint16x8_t __simd_vclzq_u16(uint16x8_t a) __NEONSAFE;
int16x8_t __simd_vclzq_s16(int16x8_t a) __NEONSAFE;
uint32x4_t __simd_vclzq_u32(uint32x4_t a) __NEONSAFE;
int32x4_t __simd_vclzq_s32(int32x4_t a) __NEONSAFE;
uint8x8_t __simd_vcnt_u8(uint8x8_t a) __NEONSAFE;
int8x8_t __simd_vcnt_s8(int8x8_t a) __NEONSAFE;
poly8x8_t __simd_vcnt_p8(poly8x8_t a) __NEONSAFE;
uint8x16_t __simd_vcntq_u8(uint8x16_t a) __NEONSAFE;
int8x16_t __simd_vcntq_s8(int8x16_t a) __NEONSAFE;
poly8x16_t __simd_vcntq_p8(poly8x16_t a) __NEONSAFE;
int8x8_t __simd_vqabs_s8(int8x8_t a) __NEONSAFE;
int16x4_t __simd_vqabs_s16(int16x4_t a) __NEONSAFE;
int32x2_t __simd_vqabs_s32(int32x2_t a) __NEONSAFE;
int8x16_t __simd_vqabsq_s8(int8x16_t a) __NEONSAFE;
int16x8_t __simd_vqabsq_s16(int16x8_t a) __NEONSAFE;
int32x4_t __simd_vqabsq_s32(int32x4_t a) __NEONSAFE;
int64x2_t __simd_vqabsq_s64(int64x2_t a) __NEONSAFE;
int8x8_t __simd_vqneg_s8(int8x8_t a) __NEONSAFE;
int16x4_t __simd_vqneg_s16(int16x4_t a) __NEONSAFE;
int32x2_t __simd_vqneg_s32(int32x2_t a) __NEONSAFE;
int8x16_t __simd_vqnegq_s8(int8x16_t a) __NEONSAFE;
int16x8_t __simd_vqnegq_s16(int16x8_t a) __NEONSAFE;
int32x4_t __simd_vqnegq_s32(int32x4_t a) __NEONSAFE;
int64x2_t __simd_vqnegq_s64(int64x2_t a) __NEONSAFE;
int8x8_t __simd_vabs_s8(int8x8_t a) __NEONSAFE;
int16x4_t __simd_vabs_s16(int16x4_t a) __NEONSAFE;
int32x2_t __simd_vabs_s32(int32x2_t a) __NEONSAFE;
int8x16_t __simd_vabsq_s8(int8x16_t a) __NEONSAFE;
int16x8_t __simd_vabsq_s16(int16x8_t a) __NEONSAFE;
int32x4_t __simd_vabsq_s32(int32x4_t a) __NEONSAFE;
int64x2_t __simd_vabsq_s64(int64x2_t a) __NEONSAFE;
int8x8_t __simd_vneg_s8(int8x8_t a) __NEONSAFE;
int16x4_t __simd_vneg_s16(int16x4_t a) __NEONSAFE;
int32x2_t __simd_vneg_s32(int32x2_t a) __NEONSAFE;
int8x16_t __simd_vnegq_s8(int8x16_t a) __NEONSAFE;
int16x8_t __simd_vnegq_s16(int16x8_t a) __NEONSAFE;
int32x4_t __simd_vnegq_s32(int32x4_t a) __NEONSAFE;
int64x2_t __simd_vnegq_s64(int64x2_t a) __NEONSAFE;
uint8x8_t __simd_vmovn_u16(uint16x8_t a) __NEONSAFE;
int8x8_t __simd_vmovn_s16(int16x8_t a) __NEONSAFE;
uint16x4_t __simd_vmovn_u32(uint32x4_t a) __NEONSAFE;
int16x4_t __simd_vmovn_s32(int32x4_t a) __NEONSAFE;
uint32x2_t __simd_vmovn_u64(uint64x2_t a) __NEONSAFE;
int32x2_t __simd_vmovn_s64(int64x2_t a) __NEONSAFE;
uint8x16_t __simd_vmovn_high_u16(uint8x8_t a, uint16x8_t b) __NEONSAFE;
int8x16_t __simd_vmovn_high_s16(int8x8_t a, int16x8_t b) __NEONSAFE;
uint16x8_t __simd_vmovn_high_u32(uint16x4_t a, uint32x4_t b) __NEONSAFE;
int16x8_t __simd_vmovn_high_s32(int16x4_t a, int32x4_t b) __NEONSAFE;
uint32x4_t __simd_vmovn_high_u64(uint32x2_t a, uint64x2_t b) __NEONSAFE;
int32x4_t __simd_vmovn_high_s64(int32x2_t a, int64x2_t b) __NEONSAFE;
int8x8_t __simd_vqmovn_s16(int16x8_t a) __NEONSAFE;
int16x4_t __simd_vqmovn_s32(int32x4_t a) __NEONSAFE;
int32x2_t __simd_vqmovn_s64(int64x2_t a) __NEONSAFE;
int8x16_t __simd_vqmovn_high_s16(int8x8_t a, int16x8_t b) __NEONSAFE;
int16x8_t __simd_vqmovn_high_s32(int16x4_t a, int32x4_t b) __NEONSAFE;
int32x4_t __simd_vqmovn_high_s64(int32x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vqmovn_u16(uint16x8_t a) __NEONSAFE;
uint16x4_t __simd_vqmovn_u32(uint32x4_t a) __NEONSAFE;
uint32x2_t __simd_vqmovn_u64(uint64x2_t a) __NEONSAFE;
uint8x16_t __simd_vqmovn_high_u16(uint8x8_t a, uint16x8_t b) __NEONSAFE;
uint16x8_t __simd_vqmovn_high_u32(uint16x4_t a, uint32x4_t b) __NEONSAFE;
uint32x4_t __simd_vqmovn_high_u64(uint32x2_t a, uint64x2_t b) __NEONSAFE;
uint8x8_t __simd_vqmovun_s16(int16x8_t a) __NEONSAFE;
uint16x4_t __simd_vqmovun_s32(int32x4_t a) __NEONSAFE;
uint32x2_t __simd_vqmovun_s64(int64x2_t a) __NEONSAFE;
uint8x16_t __simd_vqmovun_high_s16(uint8x8_t a, int16x8_t b) __NEONSAFE;
uint16x8_t __simd_vqmovun_high_s32(uint16x4_t a, int32x4_t b) __NEONSAFE;
uint32x4_t __simd_vqmovun_high_s64(uint32x2_t a, int64x2_t b) __NEONSAFE;
uint8x8_t __simd_vmvn_u8(uint8x8_t a) __NEONSAFE;
uint16x4_t __simd_vmvn_u16(uint16x4_t a) __NEONSAFE;
uint32x2_t __simd_vmvn_u32(uint32x2_t a) __NEONSAFE;
uint64x1_t __simd_vmvn_u64(uint64x1_t a) __NEONSAFE;
int8x8_t __simd_vmvn_s8(int8x8_t a) __NEONSAFE;
int16x4_t __simd_vmvn_s16(int16x4_t a) __NEONSAFE;
int32x2_t __simd_vmvn_s32(int32x2_t a) __NEONSAFE;
int64x1_t __simd_vmvn_s64(int64x1_t a) __NEONSAFE;
poly8x8_t __simd_vmvn_p8(poly8x8_t a) __NEONSAFE;
poly16x4_t __simd_vmvn_p16(poly16x4_t a) __NEONSAFE;
uint8x16_t __simd_vmvnq_u8(uint8x16_t a) __NEONSAFE;
uint16x8_t __simd_vmvnq_u16(uint16x8_t a) __NEONSAFE;
uint32x4_t __simd_vmvnq_u32(uint32x4_t a) __NEONSAFE;
uint64x2_t __simd_vmvnq_u64(uint64x2_t a) __NEONSAFE;
int8x16_t __simd_vmvnq_s8(int8x16_t a) __NEONSAFE;
int16x8_t __simd_vmvnq_s16(int16x8_t a) __NEONSAFE;
int32x4_t __simd_vmvnq_s32(int32x4_t a) __NEONSAFE;
int64x2_t __simd_vmvnq_s64(int64x2_t a) __NEONSAFE;
poly8x16_t __simd_vmvnq_p8(poly8x16_t a) __NEONSAFE;
poly16x8_t __simd_vmvnq_p16(poly16x8_t a) __NEONSAFE;
uint8x8_t __simd_vrbit_u8(uint8x8_t a) __NEONSAFE;
int8x8_t __simd_vrbit_s8(int8x8_t a) __NEONSAFE;
poly8x8_t __simd_vrbit_p8(poly8x8_t a) __NEONSAFE;
uint8x16_t __simd_vrbitq_u8(uint8x16_t a) __NEONSAFE;
int8x16_t __simd_vrbitq_s8(int8x16_t a) __NEONSAFE;
poly8x16_t __simd_vrbitq_p8(poly8x16_t a) __NEONSAFE;
uint8x8_t __simd_vceqz_u8(uint8x8_t a) __NEONSAFE;
uint8x8_t __simd_vceqz_s8(int8x8_t a) __NEONSAFE;
uint8x8_t __simd_vceqz_p8(poly8x8_t a) __NEONSAFE;
uint16x4_t __simd_vceqz_u16(uint16x4_t a) __NEONSAFE;
uint16x4_t __simd_vceqz_s16(int16x4_t a) __NEONSAFE;
uint16x4_t __simd_vceqz_p16(poly16x4_t a) __NEONSAFE;
uint32x2_t __simd_vceqz_u32(uint32x2_t a) __NEONSAFE;
uint32x2_t __simd_vceqz_s32(int32x2_t a) __NEONSAFE;
uint8x16_t __simd_vceqzq_u8(uint8x16_t a) __NEONSAFE;
uint8x16_t __simd_vceqzq_s8(int8x16_t a) __NEONSAFE;
uint8x16_t __simd_vceqzq_p8(poly8x16_t a) __NEONSAFE;
uint16x8_t __simd_vceqzq_u16(uint16x8_t a) __NEONSAFE;
uint16x8_t __simd_vceqzq_s16(int16x8_t a) __NEONSAFE;
uint16x8_t __simd_vceqzq_p16(poly16x8_t a) __NEONSAFE;
uint32x4_t __simd_vceqzq_u32(uint32x4_t a) __NEONSAFE;
uint32x4_t __simd_vceqzq_s32(int32x4_t a) __NEONSAFE;
uint64x2_t __simd_vceqzq_u64(uint64x2_t a) __NEONSAFE;
uint64x2_t __simd_vceqzq_s64(int64x2_t a) __NEONSAFE;
uint8x8_t __simd_vcgtz_s8(int8x8_t a) __NEONSAFE;
uint16x4_t __simd_vcgtz_s16(int16x4_t a) __NEONSAFE;
uint32x2_t __simd_vcgtz_s32(int32x2_t a) __NEONSAFE;
uint8x16_t __simd_vcgtzq_s8(int8x16_t a) __NEONSAFE;
uint16x8_t __simd_vcgtzq_s16(int16x8_t a) __NEONSAFE;
uint32x4_t __simd_vcgtzq_s32(int32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcgtzq_s64(int64x2_t a) __NEONSAFE;
uint8x8_t __simd_vcltz_s8(int8x8_t a) __NEONSAFE;
uint16x4_t __simd_vcltz_s16(int16x4_t a) __NEONSAFE;
uint32x2_t __simd_vcltz_s32(int32x2_t a) __NEONSAFE;
uint8x16_t __simd_vcltzq_s8(int8x16_t a) __NEONSAFE;
uint16x8_t __simd_vcltzq_s16(int16x8_t a) __NEONSAFE;
uint32x4_t __simd_vcltzq_s32(int32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcltzq_s64(int64x2_t a) __NEONSAFE;
uint8x8_t __simd_vcgez_s8(int8x8_t a) __NEONSAFE;
uint16x4_t __simd_vcgez_s16(int16x4_t a) __NEONSAFE;
uint32x2_t __simd_vcgez_s32(int32x2_t a) __NEONSAFE;
uint8x16_t __simd_vcgezq_s8(int8x16_t a) __NEONSAFE;
uint16x8_t __simd_vcgezq_s16(int16x8_t a) __NEONSAFE;
uint32x4_t __simd_vcgezq_s32(int32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcgezq_s64(int64x2_t a) __NEONSAFE;
uint8x8_t __simd_vclez_s8(int8x8_t a) __NEONSAFE;
uint16x4_t __simd_vclez_s16(int16x4_t a) __NEONSAFE;
uint32x2_t __simd_vclez_s32(int32x2_t a) __NEONSAFE;
uint8x16_t __simd_vclezq_s8(int8x16_t a) __NEONSAFE;
uint16x8_t __simd_vclezq_s16(int16x8_t a) __NEONSAFE;
uint32x4_t __simd_vclezq_s32(int32x4_t a) __NEONSAFE;
uint64x2_t __simd_vclezq_s64(int64x2_t a) __NEONSAFE;
uint32x2_t __simd_vrsqrte_u32(uint32x2_t a) __NEONSAFE;
uint32x4_t __simd_vrsqrteq_u32(uint32x4_t a) __NEONSAFE;
uint32x2_t __simd_vrecpe_u32(uint32x2_t a) __NEONSAFE;
uint32x4_t __simd_vrecpeq_u32(uint32x4_t a) __NEONSAFE;
float32x2_t __simd_vcvt_f32_f64(float64x2_t a) __NEONSAFE;
float32x4_t __simd_vcvt_high_f32_f64(float32x2_t a, float64x2_t b) __NEONSAFE;
float64x2_t __simd_vcvt_f64_f32(float32x2_t a) __NEONSAFE;
float64x2_t __simd_vcvt_high_f64_f32(float32x4_t a) __NEONSAFE;
float32x2_t __simd_vcvtx_f32_f64(float64x2_t a) __NEONSAFE;
float32x4_t __simd_vcvtx_high_f32_f64(float32x2_t a, float64x2_t b) __NEONSAFE;
float32x2_t __simd_vrndn_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrndnq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrndnq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrndm_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrndmq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrndmq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrndp_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrndpq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrndpq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrnd_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrndq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrndq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrnda_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrndaq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrndaq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrndx_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrndxq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrndxq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrndi_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrndiq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrndiq_f64(float64x2_t a) __NEONSAFE;
int32x2_t __simd_vcvtn_s32_f32(float32x2_t a) __NEONSAFE;
int32x4_t __simd_vcvtnq_s32_f32(float32x4_t a) __NEONSAFE;
int64x2_t __simd_vcvtnq_s64_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcvtn_u32_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcvtnq_u32_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcvtnq_u64_f64(float64x2_t a) __NEONSAFE;
int32x2_t __simd_vcvtm_s32_f32(float32x2_t a) __NEONSAFE;
int32x4_t __simd_vcvtmq_s32_f32(float32x4_t a) __NEONSAFE;
int64x2_t __simd_vcvtmq_s64_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcvtm_u32_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcvtmq_u32_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcvtmq_u64_f64(float64x2_t a) __NEONSAFE;
int32x2_t __simd_vcvta_s32_f32(float32x2_t a) __NEONSAFE;
int32x4_t __simd_vcvtaq_s32_f32(float32x4_t a) __NEONSAFE;
int64x2_t __simd_vcvtaq_s64_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcvta_u32_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcvtaq_u32_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcvtaq_u64_f64(float64x2_t a) __NEONSAFE;
int32x2_t __simd_vcvtp_s32_f32(float32x2_t a) __NEONSAFE;
int32x4_t __simd_vcvtpq_s32_f32(float32x4_t a) __NEONSAFE;
int64x2_t __simd_vcvtpq_s64_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcvtp_u32_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcvtpq_u32_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcvtpq_u64_f64(float64x2_t a) __NEONSAFE;
int32x2_t __simd_vcvt_s32_f32(float32x2_t a) __NEONSAFE;
int32x4_t __simd_vcvtq_s32_f32(float32x4_t a) __NEONSAFE;
int64x2_t __simd_vcvtq_s64_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcvt_u32_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcvtq_u32_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcvtq_u64_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vcvt_f32_s32(int32x2_t a) __NEONSAFE;
float32x4_t __simd_vcvtq_f32_s32(int32x4_t a) __NEONSAFE;
float64x2_t __simd_vcvtq_f64_s64(int64x2_t a) __NEONSAFE;
float32x2_t __simd_vcvt_f32_u32(uint32x2_t a) __NEONSAFE;
float32x4_t __simd_vcvtq_f32_u32(uint32x4_t a) __NEONSAFE;
float64x2_t __simd_vcvtq_f64_u64(uint64x2_t a) __NEONSAFE;
uint32x2_t __simd_vceqz_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vceqzq_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vceqzq_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcgtz_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcgtzq_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcgtzq_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcltz_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcltzq_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcltzq_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vcgez_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vcgezq_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vcgezq_f64(float64x2_t a) __NEONSAFE;
uint32x2_t __simd_vclez_f32(float32x2_t a) __NEONSAFE;
uint32x4_t __simd_vclezq_f32(float32x4_t a) __NEONSAFE;
uint64x2_t __simd_vclezq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vabs_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vabsq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vabsq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vneg_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vnegq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vnegq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrecpe_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrecpeq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrecpeq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vrsqrte_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vrsqrteq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vrsqrteq_f64(float64x2_t a) __NEONSAFE;
float32x2_t __simd_vsqrt_f32(float32x2_t a) __NEONSAFE;
float32x4_t __simd_vsqrtq_f32(float32x4_t a) __NEONSAFE;
float64x2_t __simd_vsqrtq_f64(float64x2_t a) __NEONSAFE;
int8_t __simd_vuqaddb_s8(int8_t a, uint8_t b) __NEONSAFE;
int16_t __simd_vuqaddh_s16(int16_t a, uint16_t b) __NEONSAFE;
int32_t __simd_vuqadds_s32(int32_t a, uint32_t b) __NEONSAFE;
int64x1_t __simd_vuqadd_s64(int64x1_t a, uint64x1_t b) __NEONSAFE;
int64_t __simd_vuqaddd_s64(int64_t a, uint64_t b) __NEONSAFE;
uint8_t __simd_vsqaddb_u8(uint8_t a, int8_t b) __NEONSAFE;
uint16_t __simd_vsqaddh_u16(uint16_t a, int16_t b) __NEONSAFE;
uint32_t __simd_vsqadds_u32(uint32_t a, int32_t b) __NEONSAFE;
uint64x1_t __simd_vsqadd_u64(uint64x1_t a, int64x1_t b) __NEONSAFE;
uint64_t __simd_vsqaddd_u64(uint64_t a, int64_t b) __NEONSAFE;
int8_t __simd_vqabsb_s8(int8_t a) __NEONSAFE;
int16_t __simd_vqabsh_s16(int16_t a) __NEONSAFE;
int32_t __simd_vqabss_s32(int32_t a) __NEONSAFE;
int64x1_t __simd_vqabs_s64(int64x1_t a) __NEONSAFE;
int64_t __simd_vqabsd_s64(int64_t a) __NEONSAFE;
int8_t __simd_vqnegb_s8(int8_t a) __NEONSAFE;
int16_t __simd_vqnegh_s16(int16_t a) __NEONSAFE;
int32_t __simd_vqnegs_s32(int32_t a) __NEONSAFE;
int64x1_t __simd_vqneg_s64(int64x1_t a) __NEONSAFE;
int64_t __simd_vqnegd_s64(int64_t a) __NEONSAFE;
int64x1_t __simd_vabs_s64(int64x1_t a) __NEONSAFE;
int64_t __simd_vabsd_s64(int64_t a) __NEONSAFE;
int64x1_t __simd_vneg_s64(int64x1_t a) __NEONSAFE;
int64_t __simd_vnegd_s64(int64_t a) __NEONSAFE;
uint64x1_t __simd_vceqz_u64(uint64x1_t a) __NEONSAFE;
uint64x1_t __simd_vceqz_s64(int64x1_t a) __NEONSAFE;
uint64_t __simd_vceqzd_u64(uint64_t a) __NEONSAFE;
uint64_t __simd_vceqzd_s64(int64_t a) __NEONSAFE;
uint64x1_t __simd_vcgtz_s64(int64x1_t a) __NEONSAFE;
uint64_t __simd_vcgtzd_s64(int64_t a) __NEONSAFE;
uint64x1_t __simd_vcltz_s64(int64x1_t a) __NEONSAFE;
uint64_t __simd_vcltzd_s64(int64_t a) __NEONSAFE;
uint64x1_t __simd_vcgez_s64(int64x1_t a) __NEONSAFE;
uint64_t __simd_vcgezd_s64(int64_t a) __NEONSAFE;
uint64x1_t __simd_vclez_s64(int64x1_t a) __NEONSAFE;
uint64_t __simd_vclezd_s64(int64_t a) __NEONSAFE;
int8_t __simd_vqmovnh_s16(int16_t a) __NEONSAFE;
int16_t __simd_vqmovns_s32(int32_t a) __NEONSAFE;
int32_t __simd_vqmovnd_s64(int64_t a) __NEONSAFE;
uint8_t __simd_vqmovnh_u16(uint16_t a) __NEONSAFE;
uint16_t __simd_vqmovns_u32(uint32_t a) __NEONSAFE;
uint32_t __simd_vqmovnd_u64(uint64_t a) __NEONSAFE;
uint8_t __simd_vqmovunh_s16(int16_t a) __NEONSAFE;
uint16_t __simd_vqmovuns_s32(int32_t a) __NEONSAFE;
uint32_t __simd_vqmovund_s64(int64_t a) __NEONSAFE;
int32_t __simd_vcvtns_s32_f32(float32_t a) __NEONSAFE;
int64_t __simd_vcvtnd_s64_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcvtns_u32_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcvtnd_u64_f64(float64_t a) __NEONSAFE;
int32_t __simd_vcvtms_s32_f32(float32_t a) __NEONSAFE;
int64_t __simd_vcvtmd_s64_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcvtms_u32_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcvtmd_u64_f64(float64_t a) __NEONSAFE;
int32_t __simd_vcvtas_s32_f32(float32_t a) __NEONSAFE;
int64_t __simd_vcvtad_s64_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcvtas_u32_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcvtad_u64_f64(float64_t a) __NEONSAFE;
int32_t __simd_vcvtps_s32_f32(float32_t a) __NEONSAFE;
int64_t __simd_vcvtpd_s64_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcvtps_u32_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcvtpd_u64_f64(float64_t a) __NEONSAFE;
int32_t __simd_vcvts_s32_f32(float32_t a) __NEONSAFE;
int64_t __simd_vcvtd_s64_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcvts_u32_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcvtd_u64_f64(float64_t a) __NEONSAFE;
float32_t __simd_vcvts_f32_s32(int32_t a) __NEONSAFE;
float64_t __simd_vcvtd_f64_s64(int64_t a) __NEONSAFE;
float32_t __simd_vcvts_f32_u32(uint32_t a) __NEONSAFE;
float64_t __simd_vcvtd_f64_u64(uint64_t a) __NEONSAFE;
float32_t __simd_vcvtxd_f32_f64(float64_t a) __NEONSAFE;
float32_t __simd_vrecpes_f32(float32_t a) __NEONSAFE;
float64_t __simd_vrecped_f64(float64_t a) __NEONSAFE;
float32_t __simd_vrecpxs_f32(float32_t a) __NEONSAFE;
float64_t __simd_vrecpxd_f64(float64_t a) __NEONSAFE;
float32_t __simd_vrsqrtes_f32(float32_t a) __NEONSAFE;
float64_t __simd_vrsqrted_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vceqzs_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vceqzd_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcgtzs_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcgtzd_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcltzs_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcltzd_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vcgezs_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vcgezd_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vclezs_f32(float32_t a) __NEONSAFE;
uint64_t __simd_vclezd_f64(float64_t a) __NEONSAFE;
int8x8_t __simd_vshr_n_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vshr_n_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vshr_n_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vshrq_n_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int16x8_t __simd_vshrq_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vshrq_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vshrq_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vshr_n_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vshr_n_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vshr_n_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vshrq_n_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vshrq_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vshrq_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vshrq_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vrshr_n_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vrshr_n_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vrshr_n_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vrshrq_n_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int16x8_t __simd_vrshrq_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vrshrq_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vrshrq_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vrshr_n_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vrshr_n_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vrshr_n_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vrshrq_n_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vrshrq_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vrshrq_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vrshrq_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vsra_n_s8(int8x8_t a, int8x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vsra_n_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vsra_n_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vsraq_n_s8(int8x16_t a, int8x16_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vsraq_n_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vsraq_n_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vsraq_n_s64(int64x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vsra_n_u8(uint8x8_t a, uint8x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vsra_n_u16(uint16x4_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vsra_n_u32(uint32x2_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x16_t __simd_vsraq_n_u8(uint8x16_t a, uint8x16_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vsraq_n_u16(uint16x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vsraq_n_u32(uint32x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vsraq_n_u64(uint64x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int8x8_t __simd_vrsra_n_s8(int8x8_t a, int8x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vrsra_n_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vrsra_n_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vrsraq_n_s8(int8x16_t a, int8x16_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vrsraq_n_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vrsraq_n_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vrsraq_n_s64(int64x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vrsra_n_u8(uint8x8_t a, uint8x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vrsra_n_u16(uint16x4_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vrsra_n_u32(uint32x2_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x16_t __simd_vrsraq_n_u8(uint8x16_t a, uint8x16_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vrsraq_n_u16(uint16x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vrsraq_n_u32(uint32x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vrsraq_n_u64(uint64x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vshl_n_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vshl_n_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vshl_n_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vshl_n_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vshl_n_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vshl_n_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vshlq_n_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vshlq_n_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vshlq_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x8_t __simd_vshlq_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vshlq_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vshlq_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vshlq_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vshlq_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vqshl_n_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vqshl_n_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vqshl_n_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vqshlq_n_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int16x8_t __simd_vqshlq_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vqshlq_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vqshlq_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vqshl_n_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vqshl_n_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vqshl_n_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vqshlq_n_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vqshlq_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vqshlq_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vqshlq_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vqshlu_n_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vqshlu_n_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vqshlu_n_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vqshluq_n_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vqshluq_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vqshluq_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vqshluq_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vsri_n_u8(uint8x8_t a, uint8x8_t b, __ghs_c_int__ c) __NEONSAFE;
int8x8_t __simd_vsri_n_s8(int8x8_t a, int8x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x8_t __simd_vsri_n_p8(poly8x8_t a, poly8x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vsri_n_u16(uint16x4_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vsri_n_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x4_t __simd_vsri_n_p16(poly16x4_t a, poly16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vsri_n_u32(uint32x2_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vsri_n_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x16_t __simd_vsriq_n_u8(uint8x16_t a, uint8x16_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vsriq_n_s8(int8x16_t a, int8x16_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x16_t __simd_vsriq_n_p8(poly8x16_t a, poly8x16_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vsriq_n_u16(uint16x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vsriq_n_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x8_t __simd_vsriq_n_p16(poly16x8_t a, poly16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vsriq_n_u32(uint32x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vsriq_n_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vsriq_n_u64(uint64x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vsriq_n_s64(int64x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vsli_n_u8(uint8x8_t a, uint8x8_t b, __ghs_c_int__ c) __NEONSAFE;
int8x8_t __simd_vsli_n_s8(int8x8_t a, int8x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x8_t __simd_vsli_n_p8(poly8x8_t a, poly8x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vsli_n_u16(uint16x4_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vsli_n_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x4_t __simd_vsli_n_p16(poly16x4_t a, poly16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vsli_n_u32(uint32x2_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vsli_n_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x16_t __simd_vsliq_n_u8(uint8x16_t a, uint8x16_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vsliq_n_s8(int8x16_t a, int8x16_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x16_t __simd_vsliq_n_p8(poly8x16_t a, poly8x16_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vsliq_n_u16(uint16x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vsliq_n_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x8_t __simd_vsliq_n_p16(poly16x8_t a, poly16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vsliq_n_u32(uint32x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vsliq_n_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vsliq_n_u64(uint64x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vsliq_n_s64(int64x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vshrn_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vshrn_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vshrn_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vshrn_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vshrn_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vshrn_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vshrn_high_n_u16(uint8x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vshrn_high_n_s16(int8x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vshrn_high_n_u32(uint16x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vshrn_high_n_s32(int16x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vshrn_high_n_u64(uint32x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vshrn_high_n_s64(int32x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vrshrn_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vrshrn_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vrshrn_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vrshrn_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vrshrn_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vrshrn_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vrshrn_high_n_u16(uint8x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vrshrn_high_n_s16(int8x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vrshrn_high_n_u32(uint16x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vrshrn_high_n_s32(int16x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vrshrn_high_n_u64(uint32x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vrshrn_high_n_s64(int32x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int8x8_t __simd_vqshrn_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vqshrn_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vqshrn_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vqshrn_high_n_s16(int8x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vqshrn_high_n_s32(int16x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqshrn_high_n_s64(int32x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vqshrn_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vqshrn_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vqshrn_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vqshrn_high_n_u16(uint8x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vqshrn_high_n_u32(uint16x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vqshrn_high_n_u64(uint32x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int8x8_t __simd_vqrshrn_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vqrshrn_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vqrshrn_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vqrshrn_high_n_s16(int8x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vqrshrn_high_n_s32(int16x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vqrshrn_high_n_s64(int32x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vqrshrn_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vqrshrn_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vqrshrn_n_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vqrshrn_high_n_u16(uint8x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vqrshrn_high_n_u32(uint16x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vqrshrn_high_n_u64(uint32x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vqshrun_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vqshrun_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vqshrun_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vqshrun_high_n_s16(uint8x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vqshrun_high_n_s32(uint16x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vqshrun_high_n_s64(uint32x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vqrshrun_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vqrshrun_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vqrshrun_n_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vqrshrun_high_n_s16(uint8x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vqrshrun_high_n_s32(uint16x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vqrshrun_high_n_s64(uint32x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vmovl_s8(int8x8_t a) __NEONSAFE;
int32x4_t __simd_vmovl_s16(int16x4_t a) __NEONSAFE;
int64x2_t __simd_vmovl_s32(int32x2_t a) __NEONSAFE;
int16x8_t __simd_vmovl_high_s8(int8x16_t a) __NEONSAFE;
int32x4_t __simd_vmovl_high_s16(int16x8_t a) __NEONSAFE;
int64x2_t __simd_vmovl_high_s32(int32x4_t a) __NEONSAFE;
uint16x8_t __simd_vmovl_u8(uint8x8_t a) __NEONSAFE;
uint32x4_t __simd_vmovl_u16(uint16x4_t a) __NEONSAFE;
uint64x2_t __simd_vmovl_u32(uint32x2_t a) __NEONSAFE;
uint16x8_t __simd_vmovl_high_u8(uint8x16_t a) __NEONSAFE;
uint32x4_t __simd_vmovl_high_u16(uint16x8_t a) __NEONSAFE;
uint64x2_t __simd_vmovl_high_u32(uint32x4_t a) __NEONSAFE;
int16x8_t __simd_vshll_n_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vshll_n_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vshll_n_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int16x8_t __simd_vshll_high_n_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vshll_high_n_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vshll_high_n_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vshll_n_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vshll_n_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vshll_n_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vshll_high_n_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vshll_high_n_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vshll_high_n_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
float32x2_t __simd_vcvt_n_f32_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
float32x4_t __simd_vcvtq_n_f32_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
float64x2_t __simd_vcvtq_n_f64_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
float32x2_t __simd_vcvt_n_f32_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
float32x4_t __simd_vcvtq_n_f32_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
float64x2_t __simd_vcvtq_n_f64_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vcvt_n_s32_f32(float32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vcvtq_n_s32_f32(float32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vcvtq_n_s64_f64(float64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vcvt_n_u32_f32(float32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vcvtq_n_u32_f32(float32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vcvtq_n_u64_f64(float64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int64x1_t __simd_vshr_n_s64(int64x1_t a, __ghs_c_int__ b) __NEONSAFE;
int64_t __simd_vshrd_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x1_t __simd_vshr_n_u64(uint64x1_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vshrd_n_u64(uint64_t a, __ghs_c_int__ b) __NEONSAFE;
int64x1_t __simd_vsra_n_s64(int64x1_t a, int64x1_t b, __ghs_c_int__ c) __NEONSAFE;
int64_t __simd_vsrad_n_s64(int64_t a, int64_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x1_t __simd_vsra_n_u64(uint64x1_t a, uint64x1_t b, __ghs_c_int__ c) __NEONSAFE;
uint64_t __simd_vsrad_n_u64(uint64_t a, uint64_t b, __ghs_c_int__ c) __NEONSAFE;
int64x1_t __simd_vrshr_n_s64(int64x1_t a, __ghs_c_int__ b) __NEONSAFE;
int64_t __simd_vrshrd_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x1_t __simd_vrshr_n_u64(uint64x1_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vrshrd_n_u64(uint64_t a, __ghs_c_int__ b) __NEONSAFE;
int64x1_t __simd_vrsra_n_s64(int64x1_t a, int64x1_t b, __ghs_c_int__ c) __NEONSAFE;
int64_t __simd_vrsrad_n_s64(int64_t a, int64_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x1_t __simd_vrsra_n_u64(uint64x1_t a, uint64x1_t b, __ghs_c_int__ c) __NEONSAFE;
uint64_t __simd_vrsrad_n_u64(uint64_t a, uint64_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x1_t __simd_vshl_n_u64(uint64x1_t a, __ghs_c_int__ b) __NEONSAFE;
int64x1_t __simd_vshl_n_s64(int64x1_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vshld_n_u64(uint64_t a, __ghs_c_int__ b) __NEONSAFE;
int64_t __simd_vshld_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
int8_t __simd_vqshlb_n_s8(int8_t a, __ghs_c_int__ b) __NEONSAFE;
int16_t __simd_vqshlh_n_s16(int16_t a, __ghs_c_int__ b) __NEONSAFE;
int32_t __simd_vqshls_n_s32(int32_t a, __ghs_c_int__ b) __NEONSAFE;
int64x1_t __simd_vqshl_n_s64(int64x1_t a, __ghs_c_int__ b) __NEONSAFE;
int64_t __simd_vqshld_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vqshlb_n_u8(uint8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vqshlh_n_u16(uint16_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vqshls_n_u32(uint32_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x1_t __simd_vqshl_n_u64(uint64x1_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vqshld_n_u64(uint64_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vqshlub_n_s8(int8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vqshluh_n_s16(int16_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vqshlus_n_s32(int32_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x1_t __simd_vqshlu_n_s64(int64x1_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vqshlud_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x1_t __simd_vsri_n_u64(uint64x1_t a, uint64x1_t b, __ghs_c_int__ c) __NEONSAFE;
int64x1_t __simd_vsri_n_s64(int64x1_t a, int64x1_t b, __ghs_c_int__ c) __NEONSAFE;
uint64_t __simd_vsrid_n_u64(uint64_t a, uint64_t b, __ghs_c_int__ c) __NEONSAFE;
int64_t __simd_vsrid_n_s64(int64_t a, int64_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x1_t __simd_vsli_n_u64(uint64x1_t a, uint64x1_t b, __ghs_c_int__ c) __NEONSAFE;
int64x1_t __simd_vsli_n_s64(int64x1_t a, int64x1_t b, __ghs_c_int__ c) __NEONSAFE;
uint64_t __simd_vslid_n_u64(uint64_t a, uint64_t b, __ghs_c_int__ c) __NEONSAFE;
int64_t __simd_vslid_n_s64(int64_t a, int64_t b, __ghs_c_int__ c) __NEONSAFE;
int8_t __simd_vqshrnh_n_s16(int16_t a, __ghs_c_int__ b) __NEONSAFE;
int16_t __simd_vqshrns_n_s32(int32_t a, __ghs_c_int__ b) __NEONSAFE;
int32_t __simd_vqshrnd_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vqshrnh_n_u16(uint16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vqshrns_n_u32(uint32_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vqshrnd_n_u64(uint64_t a, __ghs_c_int__ b) __NEONSAFE;
int8_t __simd_vqrshrnh_n_s16(int16_t a, __ghs_c_int__ b) __NEONSAFE;
int16_t __simd_vqrshrns_n_s32(int32_t a, __ghs_c_int__ b) __NEONSAFE;
int32_t __simd_vqrshrnd_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vqrshrnh_n_u16(uint16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vqrshrns_n_u32(uint32_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vqrshrnd_n_u64(uint64_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vqshrunh_n_s16(int16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vqshruns_n_s32(int32_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vqshrund_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vqrshrunh_n_s16(int16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vqrshruns_n_s32(int32_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vqrshrund_n_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
float32_t __simd_vcvts_n_f32_s32(int32_t a, __ghs_c_int__ b) __NEONSAFE;
float64_t __simd_vcvtd_n_f64_s64(int64_t a, __ghs_c_int__ b) __NEONSAFE;
float32_t __simd_vcvts_n_f32_u32(uint32_t a, __ghs_c_int__ b) __NEONSAFE;
float64_t __simd_vcvtd_n_f64_u64(uint64_t a, __ghs_c_int__ b) __NEONSAFE;
int32_t __simd_vcvts_n_s32_f32(float32_t a, __ghs_c_int__ b) __NEONSAFE;
int64_t __simd_vcvtd_n_s64_f64(float64_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vcvts_n_u32_f32(float32_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vcvtd_n_u64_f64(float64_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x1_t __simd_vpadd_u64(uint64x2_t a) __NEONSAFE;
int64x1_t __simd_vpadd_s64(int64x2_t a) __NEONSAFE;
uint64_t __simd_vpaddd_u64(uint64x2_t a) __NEONSAFE;
int64_t __simd_vpaddd_s64(int64x2_t a) __NEONSAFE;
float32_t __simd_vpadds_f32(float32x2_t a) __NEONSAFE;
float64_t __simd_vpaddd_f64(float64x2_t a) __NEONSAFE;
float32_t __simd_vpmaxs_f32(float32x2_t a) __NEONSAFE;
float64_t __simd_vpmaxd_f64(float64x2_t a) __NEONSAFE;
float32_t __simd_vpmaxnms_f32(float32x2_t a) __NEONSAFE;
float64_t __simd_vpmaxnmd_f64(float64x2_t a) __NEONSAFE;
float32_t __simd_vpmins_f32(float32x2_t a) __NEONSAFE;
float64_t __simd_vpmind_f64(float64x2_t a) __NEONSAFE;
float32_t __simd_vpminnms_f32(float32x2_t a) __NEONSAFE;
float64_t __simd_vpminnmd_f64(float64x2_t a) __NEONSAFE;
float32_t __simd_vget_lane_f32(float32x2_t a, __ghs_c_int__ b) __NEONSAFE;
float32_t __simd_vgetq_lane_f32(float32x4_t a, __ghs_c_int__ b) __NEONSAFE;
float64_t __simd_vgetq_lane_f64(float64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vdup_lane_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vdup_lane_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
poly8x8_t __simd_vdup_lane_p8(poly8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vdup_laneq_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int8x8_t __simd_vdup_laneq_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
poly8x8_t __simd_vdup_laneq_p8(poly8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vdup_n_u8(uint8_t a) __NEONSAFE;
int8x8_t __simd_vdup_n_s8(int8_t a) __NEONSAFE;
poly8x8_t __simd_vdup_n_p8(poly8_t a) __NEONSAFE;
uint8x16_t __simd_vdupq_lane_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vdupq_lane_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
poly8x16_t __simd_vdupq_lane_p8(poly8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vdupq_laneq_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int8x16_t __simd_vdupq_laneq_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
poly8x16_t __simd_vdupq_laneq_p8(poly8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x16_t __simd_vdupq_n_u8(uint8_t a) __NEONSAFE;
int8x16_t __simd_vdupq_n_s8(int8_t a) __NEONSAFE;
poly8x16_t __simd_vdupq_n_p8(poly8_t a) __NEONSAFE;
uint16x4_t __simd_vdup_lane_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vdup_lane_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
poly16x4_t __simd_vdup_lane_p16(poly16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vdup_laneq_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x4_t __simd_vdup_laneq_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
poly16x4_t __simd_vdup_laneq_p16(poly16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x4_t __simd_vdup_n_u16(uint16_t a) __NEONSAFE;
int16x4_t __simd_vdup_n_s16(int16_t a) __NEONSAFE;
poly16x4_t __simd_vdup_n_p16(poly16_t a) __NEONSAFE;
uint16x8_t __simd_vdupq_lane_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int16x8_t __simd_vdupq_lane_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
poly16x8_t __simd_vdupq_lane_p16(poly16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vdupq_laneq_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int16x8_t __simd_vdupq_laneq_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
poly16x8_t __simd_vdupq_laneq_p16(poly16x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint16x8_t __simd_vdupq_n_u16(uint16_t a) __NEONSAFE;
int16x8_t __simd_vdupq_n_s16(int16_t a) __NEONSAFE;
poly16x8_t __simd_vdupq_n_p16(poly16_t a) __NEONSAFE;
uint32x2_t __simd_vdup_lane_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vdup_lane_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
float32x2_t __simd_vdup_lane_f32(float32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vdup_laneq_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x2_t __simd_vdup_laneq_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
float32x2_t __simd_vdup_laneq_f32(float32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x2_t __simd_vdup_n_u32(uint32_t a) __NEONSAFE;
int32x2_t __simd_vdup_n_s32(int32_t a) __NEONSAFE;
float32x2_t __simd_vdup_n_f32(float32_t a) __NEONSAFE;
uint32x4_t __simd_vdupq_lane_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vdupq_lane_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
float32x4_t __simd_vdupq_lane_f32(float32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vdupq_laneq_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32x4_t __simd_vdupq_laneq_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
float32x4_t __simd_vdupq_laneq_f32(float32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint32x4_t __simd_vdupq_n_u32(uint32_t a) __NEONSAFE;
int32x4_t __simd_vdupq_n_s32(int32_t a) __NEONSAFE;
float32x4_t __simd_vdupq_n_f32(float32_t a) __NEONSAFE;
uint64x2_t __simd_vdupq_lane_u64(uint64x1_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vdupq_lane_s64(int64x1_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vdupq_laneq_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int64x2_t __simd_vdupq_laneq_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
float64x2_t __simd_vdupq_laneq_f64(float64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint64x2_t __simd_vdupq_n_u64(uint64_t a) __NEONSAFE;
int64x2_t __simd_vdupq_n_s64(int64_t a) __NEONSAFE;
float64x2_t __simd_vdupq_n_f64(float64_t a) __NEONSAFE;
uint32_t __simd_vget_lane_u32(uint32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint32_t __simd_vgetq_lane_u32(uint32x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vget_lane_u8(uint8x8_t a, __ghs_c_int__ b) __NEONSAFE;
poly8_t __simd_vget_lane_p8(poly8x8_t a, __ghs_c_int__ b) __NEONSAFE;
uint8_t __simd_vgetq_lane_u8(uint8x16_t a, __ghs_c_int__ b) __NEONSAFE;
poly8_t __simd_vgetq_lane_p8(poly8x16_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vget_lane_u16(uint16x4_t a, __ghs_c_int__ b) __NEONSAFE;
poly16_t __simd_vget_lane_p16(poly16x4_t a, __ghs_c_int__ b) __NEONSAFE;
uint16_t __simd_vgetq_lane_u16(uint16x8_t a, __ghs_c_int__ b) __NEONSAFE;
poly16_t __simd_vgetq_lane_p16(poly16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int8_t __simd_vget_lane_s8(int8x8_t a, __ghs_c_int__ b) __NEONSAFE;
int8_t __simd_vgetq_lane_s8(int8x16_t a, __ghs_c_int__ b) __NEONSAFE;
int16_t __simd_vget_lane_s16(int16x4_t a, __ghs_c_int__ b) __NEONSAFE;
int16_t __simd_vgetq_lane_s16(int16x8_t a, __ghs_c_int__ b) __NEONSAFE;
int32_t __simd_vgetq_lane_s32(int32x4_t a, __ghs_c_int__ b) __NEONSAFE;
int32_t __simd_vget_lane_s32(int32x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vget_lane_u64(uint64x1_t a, __ghs_c_int__ b) __NEONSAFE;
int64_t __simd_vget_lane_s64(int64x1_t a, __ghs_c_int__ b) __NEONSAFE;
uint64_t __simd_vgetq_lane_u64(uint64x2_t a, __ghs_c_int__ b) __NEONSAFE;
int64_t __simd_vgetq_lane_s64(int64x2_t a, __ghs_c_int__ b) __NEONSAFE;
uint8x8_t __simd_vset_lane_u8(uint8_t a, uint8x8_t b, __ghs_c_int__ c) __NEONSAFE;
int8x8_t __simd_vset_lane_s8(int8_t a, int8x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x8_t __simd_vset_lane_p8(poly8_t a, poly8x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x16_t __simd_vsetq_lane_u8(uint8_t a, uint8x16_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vsetq_lane_s8(int8_t a, int8x16_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x16_t __simd_vsetq_lane_p8(poly8_t a, poly8x16_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vset_lane_u16(uint16_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vset_lane_s16(int16_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x4_t __simd_vset_lane_p16(poly16_t a, poly16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vsetq_lane_u16(uint16_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vsetq_lane_s16(int16_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x8_t __simd_vsetq_lane_p16(poly16_t a, poly16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vset_lane_u32(uint32_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vset_lane_s32(int32_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vsetq_lane_u32(uint32_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vsetq_lane_s32(int32_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x1_t __simd_vset_lane_u64(uint64_t a, uint64x1_t b, __ghs_c_int__ c) __NEONSAFE;
int64x1_t __simd_vset_lane_s64(int64_t a, int64x1_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vsetq_lane_u64(uint64_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vsetq_lane_s64(int64_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vcreate_u8(uint64_t a) __NEONSAFE;
uint16x4_t __simd_vcreate_u16(uint64_t a) __NEONSAFE;
uint32x2_t __simd_vcreate_u32(uint64_t a) __NEONSAFE;
uint64x1_t __simd_vcreate_u64(uint64_t a) __NEONSAFE;
int8x8_t __simd_vcreate_s8(uint64_t a) __NEONSAFE;
int16x4_t __simd_vcreate_s16(uint64_t a) __NEONSAFE;
int32x2_t __simd_vcreate_s32(uint64_t a) __NEONSAFE;
int64x1_t __simd_vcreate_s64(uint64_t a) __NEONSAFE;
poly8x8_t __simd_vcreate_p8(uint64_t a) __NEONSAFE;
poly16x4_t __simd_vcreate_p16(uint64_t a) __NEONSAFE;
float32x2_t __simd_vcreate_f32(uint64_t a) __NEONSAFE;
uint8x8_t __simd_vcopy_lane_u8(uint8x8_t a, __ghs_c_int__ b, uint8x8_t c, __ghs_c_int__ d) __NEONSAFE;
int8x8_t __simd_vcopy_lane_s8(int8x8_t a, __ghs_c_int__ b, int8x8_t c, __ghs_c_int__ d) __NEONSAFE;
poly8x8_t __simd_vcopy_lane_p8(poly8x8_t a, __ghs_c_int__ b, poly8x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint8x8_t __simd_vcopy_laneq_u8(uint8x8_t a, __ghs_c_int__ b, uint8x16_t c, __ghs_c_int__ d) __NEONSAFE;
int8x8_t __simd_vcopy_laneq_s8(int8x8_t a, __ghs_c_int__ b, int8x16_t c, __ghs_c_int__ d) __NEONSAFE;
poly8x8_t __simd_vcopy_laneq_p8(poly8x8_t a, __ghs_c_int__ b, poly8x16_t c, __ghs_c_int__ d) __NEONSAFE;
uint8x16_t __simd_vcopyq_lane_u8(uint8x16_t a, __ghs_c_int__ b, uint8x8_t c, __ghs_c_int__ d) __NEONSAFE;
int8x16_t __simd_vcopyq_lane_s8(int8x16_t a, __ghs_c_int__ b, int8x8_t c, __ghs_c_int__ d) __NEONSAFE;
poly8x16_t __simd_vcopyq_lane_p8(poly8x16_t a, __ghs_c_int__ b, poly8x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint8x16_t __simd_vcopyq_laneq_u8(uint8x16_t a, __ghs_c_int__ b, uint8x16_t c, __ghs_c_int__ d) __NEONSAFE;
int8x16_t __simd_vcopyq_laneq_s8(int8x16_t a, __ghs_c_int__ b, int8x16_t c, __ghs_c_int__ d) __NEONSAFE;
poly8x16_t __simd_vcopyq_laneq_p8(poly8x16_t a, __ghs_c_int__ b, poly8x16_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x4_t __simd_vcopy_lane_u16(uint16x4_t a, __ghs_c_int__ b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vcopy_lane_s16(int16x4_t a, __ghs_c_int__ b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
poly16x4_t __simd_vcopy_lane_p16(poly16x4_t a, __ghs_c_int__ b, poly16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x4_t __simd_vcopy_laneq_u16(uint16x4_t a, __ghs_c_int__ b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vcopy_laneq_s16(int16x4_t a, __ghs_c_int__ b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
poly16x4_t __simd_vcopy_laneq_p16(poly16x4_t a, __ghs_c_int__ b, poly16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x8_t __simd_vcopyq_lane_u16(uint16x8_t a, __ghs_c_int__ b, uint16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vcopyq_lane_s16(int16x8_t a, __ghs_c_int__ b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
poly16x8_t __simd_vcopyq_lane_p16(poly16x8_t a, __ghs_c_int__ b, poly16x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint16x8_t __simd_vcopyq_laneq_u16(uint16x8_t a, __ghs_c_int__ b, uint16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vcopyq_laneq_s16(int16x8_t a, __ghs_c_int__ b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
poly16x8_t __simd_vcopyq_laneq_p16(poly16x8_t a, __ghs_c_int__ b, poly16x8_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x2_t __simd_vcopy_lane_u32(uint32x2_t a, __ghs_c_int__ b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vcopy_lane_s32(int32x2_t a, __ghs_c_int__ b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32x2_t __simd_vcopy_lane_f32(float32x2_t a, __ghs_c_int__ b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x2_t __simd_vcopy_laneq_u32(uint32x2_t a, __ghs_c_int__ b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vcopy_laneq_s32(int32x2_t a, __ghs_c_int__ b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float32x2_t __simd_vcopy_laneq_f32(float32x2_t a, __ghs_c_int__ b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vcopyq_lane_u32(uint32x4_t a, __ghs_c_int__ b, uint32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vcopyq_lane_s32(int32x4_t a, __ghs_c_int__ b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32x4_t __simd_vcopyq_lane_f32(float32x4_t a, __ghs_c_int__ b, float32x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint32x4_t __simd_vcopyq_laneq_u32(uint32x4_t a, __ghs_c_int__ b, uint32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vcopyq_laneq_s32(int32x4_t a, __ghs_c_int__ b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
float32x4_t __simd_vcopyq_laneq_f32(float32x4_t a, __ghs_c_int__ b, float32x4_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x1_t __simd_vcopy_lane_u64(uint64x1_t a, __ghs_c_int__ b, uint64x1_t c, __ghs_c_int__ d) __NEONSAFE;
int64x1_t __simd_vcopy_lane_s64(int64x1_t a, __ghs_c_int__ b, int64x1_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x1_t __simd_vcopy_laneq_u64(uint64x1_t a, __ghs_c_int__ b, uint64x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x1_t __simd_vcopy_laneq_s64(int64x1_t a, __ghs_c_int__ b, int64x2_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vcopyq_lane_u64(uint64x2_t a, __ghs_c_int__ b, uint64x1_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vcopyq_lane_s64(int64x2_t a, __ghs_c_int__ b, int64x1_t c, __ghs_c_int__ d) __NEONSAFE;
uint64x2_t __simd_vcopyq_laneq_u64(uint64x2_t a, __ghs_c_int__ b, uint64x2_t c, __ghs_c_int__ d) __NEONSAFE;
int64x2_t __simd_vcopyq_laneq_s64(int64x2_t a, __ghs_c_int__ b, int64x2_t c, __ghs_c_int__ d) __NEONSAFE;
float64x2_t __simd_vcopyq_laneq_f64(float64x2_t a, __ghs_c_int__ b, float64x2_t c, __ghs_c_int__ d) __NEONSAFE;
float32x2_t __simd_vset_lane_f32(float32_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32x4_t __simd_vsetq_lane_f32(float32_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float64x2_t __simd_vsetq_lane_f64(float64_t a, float64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int16_t __simd_vaddlv_s8(int8x8_t a) __NEONSAFE;
int32_t __simd_vaddlv_s16(int16x4_t a) __NEONSAFE;
int16_t __simd_vaddlvq_s8(int8x16_t a) __NEONSAFE;
int32_t __simd_vaddlvq_s16(int16x8_t a) __NEONSAFE;
int64_t __simd_vaddlvq_s32(int32x4_t a) __NEONSAFE;
uint16_t __simd_vaddlv_u8(uint8x8_t a) __NEONSAFE;
uint32_t __simd_vaddlv_u16(uint16x4_t a) __NEONSAFE;
uint16_t __simd_vaddlvq_u8(uint8x16_t a) __NEONSAFE;
uint32_t __simd_vaddlvq_u16(uint16x8_t a) __NEONSAFE;
uint64_t __simd_vaddlvq_u32(uint32x4_t a) __NEONSAFE;
uint8_t __simd_vaddv_u8(uint8x8_t a) __NEONSAFE;
int8_t __simd_vaddv_s8(int8x8_t a) __NEONSAFE;
uint16_t __simd_vaddv_u16(uint16x4_t a) __NEONSAFE;
int16_t __simd_vaddv_s16(int16x4_t a) __NEONSAFE;
uint8_t __simd_vaddvq_u8(uint8x16_t a) __NEONSAFE;
int8_t __simd_vaddvq_s8(int8x16_t a) __NEONSAFE;
uint16_t __simd_vaddvq_u16(uint16x8_t a) __NEONSAFE;
int16_t __simd_vaddvq_s16(int16x8_t a) __NEONSAFE;
uint32_t __simd_vaddvq_u32(uint32x4_t a) __NEONSAFE;
int32_t __simd_vaddvq_s32(int32x4_t a) __NEONSAFE;
int8_t __simd_vmaxv_s8(int8x8_t a) __NEONSAFE;
int16_t __simd_vmaxv_s16(int16x4_t a) __NEONSAFE;
int8_t __simd_vmaxvq_s8(int8x16_t a) __NEONSAFE;
int16_t __simd_vmaxvq_s16(int16x8_t a) __NEONSAFE;
int32_t __simd_vmaxvq_s32(int32x4_t a) __NEONSAFE;
uint8_t __simd_vmaxv_u8(uint8x8_t a) __NEONSAFE;
uint16_t __simd_vmaxv_u16(uint16x4_t a) __NEONSAFE;
uint8_t __simd_vmaxvq_u8(uint8x16_t a) __NEONSAFE;
uint16_t __simd_vmaxvq_u16(uint16x8_t a) __NEONSAFE;
uint32_t __simd_vmaxvq_u32(uint32x4_t a) __NEONSAFE;
int8_t __simd_vminv_s8(int8x8_t a) __NEONSAFE;
int16_t __simd_vminv_s16(int16x4_t a) __NEONSAFE;
int8_t __simd_vminvq_s8(int8x16_t a) __NEONSAFE;
int16_t __simd_vminvq_s16(int16x8_t a) __NEONSAFE;
int32_t __simd_vminvq_s32(int32x4_t a) __NEONSAFE;
uint8_t __simd_vminv_u8(uint8x8_t a) __NEONSAFE;
uint16_t __simd_vminv_u16(uint16x4_t a) __NEONSAFE;
uint8_t __simd_vminvq_u8(uint8x16_t a) __NEONSAFE;
uint16_t __simd_vminvq_u16(uint16x8_t a) __NEONSAFE;
uint32_t __simd_vminvq_u32(uint32x4_t a) __NEONSAFE;
float32_t __simd_vmaxvq_f32(float32x4_t a) __NEONSAFE;
float32_t __simd_vmaxnmvq_f32(float32x4_t a) __NEONSAFE;
float32_t __simd_vminvq_f32(float32x4_t a) __NEONSAFE;
float32_t __simd_vminnmvq_f32(float32x4_t a) __NEONSAFE;
uint8x8_t __simd_vext_u8(uint8x8_t a, uint8x8_t b, __ghs_c_int__ c) __NEONSAFE;
int8x8_t __simd_vext_s8(int8x8_t a, int8x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x8_t __simd_vext_p8(poly8x8_t a, poly8x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x16_t __simd_vextq_u8(uint8x16_t a, uint8x16_t b, __ghs_c_int__ c) __NEONSAFE;
int8x16_t __simd_vextq_s8(int8x16_t a, int8x16_t b, __ghs_c_int__ c) __NEONSAFE;
poly8x16_t __simd_vextq_p8(poly8x16_t a, poly8x16_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x4_t __simd_vext_u16(uint16x4_t a, uint16x4_t b, __ghs_c_int__ c) __NEONSAFE;
int16x4_t __simd_vext_s16(int16x4_t a, int16x4_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x4_t __simd_vext_p16(poly16x4_t a, poly16x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x2_t __simd_vext_u32(uint32x2_t a, uint32x2_t b, __ghs_c_int__ c) __NEONSAFE;
int32x2_t __simd_vext_s32(int32x2_t a, int32x2_t b, __ghs_c_int__ c) __NEONSAFE;
float32x2_t __simd_vext_f32(float32x2_t a, float32x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x1_t __simd_vext_u64(uint64x1_t a, uint64x1_t b, __ghs_c_int__ c) __NEONSAFE;
int64x1_t __simd_vext_s64(int64x1_t a, int64x1_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x1_t __simd_vextd_u64(uint64x1_t a, uint64x1_t b, __ghs_c_int__ c) __NEONSAFE;
int64x1_t __simd_vextd_s64(int64x1_t a, int64x1_t b, __ghs_c_int__ c) __NEONSAFE;
uint16x8_t __simd_vextq_u16(uint16x8_t a, uint16x8_t b, __ghs_c_int__ c) __NEONSAFE;
int16x8_t __simd_vextq_s16(int16x8_t a, int16x8_t b, __ghs_c_int__ c) __NEONSAFE;
poly16x8_t __simd_vextq_p16(poly16x8_t a, poly16x8_t b, __ghs_c_int__ c) __NEONSAFE;
uint32x4_t __simd_vextq_u32(uint32x4_t a, uint32x4_t b, __ghs_c_int__ c) __NEONSAFE;
int32x4_t __simd_vextq_s32(int32x4_t a, int32x4_t b, __ghs_c_int__ c) __NEONSAFE;
float32x4_t __simd_vextq_f32(float32x4_t a, float32x4_t b, __ghs_c_int__ c) __NEONSAFE;
uint64x2_t __simd_vextq_u64(uint64x2_t a, uint64x2_t b, __ghs_c_int__ c) __NEONSAFE;
int64x2_t __simd_vextq_s64(int64x2_t a, int64x2_t b, __ghs_c_int__ c) __NEONSAFE;
float64x2_t __simd_vextq_f64(float64x2_t a, float64x2_t b, __ghs_c_int__ c) __NEONSAFE;
uint8x8_t __simd_vtbl1_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vtbl1_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vtbl1_p8(poly8x8_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vtbl2_u8(uint8x8x2_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vtbl2_s8(int8x8x2_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vtbl2_p8(poly8x8x2_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vqtbl1_u8(uint8x16_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vqtbl1_s8(int8x16_t a, uint8x8_t b) __NEONSAFE;
poly8x8_t __simd_vqtbl1_p8(poly8x16_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vtbl3_u8(uint8x8x3_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vtbl3_s8(int8x8x3_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vtbl3_p8(poly8x8x3_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vtbl4_u8(uint8x8x4_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vtbl4_s8(int8x8x4_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vtbl4_p8(poly8x8x4_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vqtbl2_u8(uint8x16x2_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vqtbl2_s8(int8x16x2_t a, uint8x8_t b) __NEONSAFE;
poly8x8_t __simd_vqtbl2_p8(poly8x16x2_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vqtbl3_u8(uint8x16x3_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vqtbl3_s8(int8x16x3_t a, uint8x8_t b) __NEONSAFE;
poly8x8_t __simd_vqtbl3_p8(poly8x16x3_t a, uint8x8_t b) __NEONSAFE;
uint8x8_t __simd_vqtbl4_u8(uint8x16x4_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vqtbl4_s8(int8x16x4_t a, uint8x8_t b) __NEONSAFE;
poly8x8_t __simd_vqtbl4_p8(poly8x16x4_t a, uint8x8_t b) __NEONSAFE;
uint8x16_t __simd_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vqtbl1q_s8(int8x16_t a, uint8x16_t b) __NEONSAFE;
poly8x16_t __simd_vqtbl1q_p8(poly8x16_t a, uint8x16_t b) __NEONSAFE;
uint8x16_t __simd_vqtbl2q_u8(uint8x16x2_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vqtbl2q_s8(int8x16x2_t a, uint8x16_t b) __NEONSAFE;
poly8x16_t __simd_vqtbl2q_p8(poly8x16x2_t a, uint8x16_t b) __NEONSAFE;
uint8x16_t __simd_vqtbl3q_u8(uint8x16x3_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vqtbl3q_s8(int8x16x3_t a, uint8x16_t b) __NEONSAFE;
poly8x16_t __simd_vqtbl3q_p8(poly8x16x3_t a, uint8x16_t b) __NEONSAFE;
uint8x16_t __simd_vqtbl4q_u8(uint8x16x4_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vqtbl4q_s8(int8x16x4_t a, uint8x16_t b) __NEONSAFE;
poly8x16_t __simd_vqtbl4q_p8(poly8x16x4_t a, uint8x16_t b) __NEONSAFE;
uint8x8_t __simd_vtbx1_u8(uint8x8_t a, uint8x8_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vtbx1_s8(int8x8_t a, int8x8_t b, int8x8_t c) __NEONSAFE;
poly8x8_t __simd_vtbx1_p8(poly8x8_t a, poly8x8_t b, uint8x8_t c) __NEONSAFE;
uint8x8_t __simd_vtbx2_u8(uint8x8_t a, uint8x8x2_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vtbx2_s8(int8x8_t a, int8x8x2_t b, int8x8_t c) __NEONSAFE;
poly8x8_t __simd_vtbx2_p8(poly8x8_t a, poly8x8x2_t b, uint8x8_t c) __NEONSAFE;
uint8x8_t __simd_vqtbx1_u8(uint8x8_t a, uint8x16_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vqtbx1_s8(int8x8_t a, int8x16_t b, uint8x8_t c) __NEONSAFE;
poly8x8_t __simd_vqtbx1_p8(poly8x8_t a, poly8x16_t b, uint8x8_t c) __NEONSAFE;
uint8x8_t __simd_vtbx3_u8(uint8x8_t a, uint8x8x3_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vtbx3_s8(int8x8_t a, int8x8x3_t b, int8x8_t c) __NEONSAFE;
poly8x8_t __simd_vtbx3_p8(poly8x8_t a, poly8x8x3_t b, uint8x8_t c) __NEONSAFE;
uint8x8_t __simd_vtbx4_u8(uint8x8_t a, uint8x8x4_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vtbx4_s8(int8x8_t a, int8x8x4_t b, int8x8_t c) __NEONSAFE;
poly8x8_t __simd_vtbx4_p8(poly8x8_t a, poly8x8x4_t b, uint8x8_t c) __NEONSAFE;
uint8x8_t __simd_vqtbx2_u8(uint8x8_t a, uint8x16x2_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vqtbx2_s8(int8x8_t a, int8x16x2_t b, uint8x8_t c) __NEONSAFE;
poly8x8_t __simd_vqtbx2_p8(poly8x8_t a, poly8x16x2_t b, uint8x8_t c) __NEONSAFE;
uint8x8_t __simd_vqtbx3_u8(uint8x8_t a, uint8x16x3_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vqtbx3_s8(int8x8_t a, int8x16x3_t b, uint8x8_t c) __NEONSAFE;
poly8x8_t __simd_vqtbx3_p8(poly8x8_t a, poly8x16x3_t b, uint8x8_t c) __NEONSAFE;
uint8x8_t __simd_vqtbx4_u8(uint8x8_t a, uint8x16x4_t b, uint8x8_t c) __NEONSAFE;
int8x8_t __simd_vqtbx4_s8(int8x8_t a, int8x16x4_t b, uint8x8_t c) __NEONSAFE;
poly8x8_t __simd_vqtbx4_p8(poly8x8_t a, poly8x16x4_t b, uint8x8_t c) __NEONSAFE;
uint8x16_t __simd_vqtbx1q_u8(uint8x16_t a, uint8x16_t b, uint8x16_t c) __NEONSAFE;
int8x16_t __simd_vqtbx1q_s8(int8x16_t a, int8x16_t b, uint8x16_t c) __NEONSAFE;
poly8x16_t __simd_vqtbx1q_p8(poly8x16_t a, poly8x16_t b, uint8x16_t c) __NEONSAFE;
uint8x16_t __simd_vqtbx2q_u8(uint8x16_t a, uint8x16x2_t b, uint8x16_t c) __NEONSAFE;
int8x16_t __simd_vqtbx2q_s8(int8x16_t a, int8x16x2_t b, uint8x16_t c) __NEONSAFE;
poly8x16_t __simd_vqtbx2q_p8(poly8x16_t a, poly8x16x2_t b, uint8x16_t c) __NEONSAFE;
uint8x16_t __simd_vqtbx3q_u8(uint8x16_t a, uint8x16x3_t b, uint8x16_t c) __NEONSAFE;
int8x16_t __simd_vqtbx3q_s8(int8x16_t a, int8x16x3_t b, uint8x16_t c) __NEONSAFE;
poly8x16_t __simd_vqtbx3q_p8(poly8x16_t a, poly8x16x3_t b, uint8x16_t c) __NEONSAFE;
uint8x16_t __simd_vqtbx4q_u8(uint8x16_t a, uint8x16x4_t b, uint8x16_t c) __NEONSAFE;
int8x16_t __simd_vqtbx4q_s8(int8x16_t a, int8x16x4_t b, uint8x16_t c) __NEONSAFE;
poly8x16_t __simd_vqtbx4q_p8(poly8x16_t a, poly8x16x4_t b, uint8x16_t c) __NEONSAFE;
uint8x16_t __simd_vcombine_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x16_t __simd_vcombine_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x16_t __simd_vcombine_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x8_t __simd_vcombine_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x8_t __simd_vcombine_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x8_t __simd_vcombine_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x4_t __simd_vcombine_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x4_t __simd_vcombine_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x4_t __simd_vcombine_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint64x2_t __simd_vcombine_u64(uint64x1_t a, uint64x1_t b) __NEONSAFE;
int64x2_t __simd_vcombine_s64(int64x1_t a, int64x1_t b) __NEONSAFE;
float64x2_t __simd_vcombine_f64(float64x1_t a, float64x1_t b) __NEONSAFE;
uint8x8_t __simd_vget_high_u8(uint8x16_t a) __NEONSAFE;
int8x8_t __simd_vget_high_s8(int8x16_t a) __NEONSAFE;
poly8x8_t __simd_vget_high_p8(poly8x16_t a) __NEONSAFE;
uint16x4_t __simd_vget_high_u16(uint16x8_t a) __NEONSAFE;
int16x4_t __simd_vget_high_s16(int16x8_t a) __NEONSAFE;
poly16x4_t __simd_vget_high_p16(poly16x8_t a) __NEONSAFE;
uint32x2_t __simd_vget_high_u32(uint32x4_t a) __NEONSAFE;
int32x2_t __simd_vget_high_s32(int32x4_t a) __NEONSAFE;
float32x2_t __simd_vget_high_f32(float32x4_t a) __NEONSAFE;
uint64x1_t __simd_vget_high_u64(uint64x2_t a) __NEONSAFE;
int64x1_t __simd_vget_high_s64(int64x2_t a) __NEONSAFE;
float64x1_t __simd_vget_high_f64(float64x2_t a) __NEONSAFE;
uint8x8_t __simd_vget_low_u8(uint8x16_t a) __NEONSAFE;
int8x8_t __simd_vget_low_s8(int8x16_t a) __NEONSAFE;
poly8x8_t __simd_vget_low_p8(poly8x16_t a) __NEONSAFE;
uint16x4_t __simd_vget_low_u16(uint16x8_t a) __NEONSAFE;
int16x4_t __simd_vget_low_s16(int16x8_t a) __NEONSAFE;
poly16x4_t __simd_vget_low_p16(poly16x8_t a) __NEONSAFE;
uint32x2_t __simd_vget_low_u32(uint32x4_t a) __NEONSAFE;
int32x2_t __simd_vget_low_s32(int32x4_t a) __NEONSAFE;
float32x2_t __simd_vget_low_f32(float32x4_t a) __NEONSAFE;
uint64x1_t __simd_vget_low_u64(uint64x2_t a) __NEONSAFE;
int64x1_t __simd_vget_low_s64(int64x2_t a) __NEONSAFE;
float64x1_t __simd_vget_low_f64(float64x2_t a) __NEONSAFE;
uint8x8_t __simd_vtrn1_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vtrn1_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vtrn1_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vtrn1_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vtrn1_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4_t __simd_vtrn1_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vtrn1_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vtrn1_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2_t __simd_vtrn1_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16_t __simd_vtrn1q_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vtrn1q_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16_t __simd_vtrn1q_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vtrn1q_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vtrn1q_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8_t __simd_vtrn1q_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vtrn1q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vtrn1q_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4_t __simd_vtrn1q_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vtrn1q_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vtrn1q_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2_t __simd_vtrn1q_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8_t __simd_vtrn2_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vtrn2_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vtrn2_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vtrn2_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vtrn2_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4_t __simd_vtrn2_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vtrn2_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vtrn2_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2_t __simd_vtrn2_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16_t __simd_vtrn2q_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vtrn2q_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16_t __simd_vtrn2q_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vtrn2q_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vtrn2q_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8_t __simd_vtrn2q_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vtrn2q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vtrn2q_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4_t __simd_vtrn2q_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vtrn2q_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vtrn2q_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2_t __simd_vtrn2q_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8_t __simd_vuzp1_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vuzp1_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vuzp1_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vuzp1_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vuzp1_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4_t __simd_vuzp1_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vuzp1_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vuzp1_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2_t __simd_vuzp1_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16_t __simd_vuzp1q_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vuzp1q_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16_t __simd_vuzp1q_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vuzp1q_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vuzp1q_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8_t __simd_vuzp1q_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vuzp1q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vuzp1q_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4_t __simd_vuzp1q_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vuzp1q_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vuzp1q_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2_t __simd_vuzp1q_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8_t __simd_vuzp2_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vuzp2_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vuzp2_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vuzp2_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vuzp2_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4_t __simd_vuzp2_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vuzp2_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vuzp2_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2_t __simd_vuzp2_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16_t __simd_vuzp2q_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vuzp2q_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16_t __simd_vuzp2q_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vuzp2q_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vuzp2q_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8_t __simd_vuzp2q_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vuzp2q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vuzp2q_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4_t __simd_vuzp2q_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vuzp2q_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vuzp2q_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2_t __simd_vuzp2q_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8_t __simd_vzip1_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vzip1_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vzip1_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vzip1_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vzip1_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4_t __simd_vzip1_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vzip1_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vzip1_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2_t __simd_vzip1_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16_t __simd_vzip1q_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vzip1q_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16_t __simd_vzip1q_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vzip1q_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vzip1q_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8_t __simd_vzip1q_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vzip1q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vzip1q_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4_t __simd_vzip1q_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vzip1q_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vzip1q_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2_t __simd_vzip1q_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8_t __simd_vzip2_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8_t __simd_vzip2_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8_t __simd_vzip2_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4_t __simd_vzip2_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4_t __simd_vzip2_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4_t __simd_vzip2_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2_t __simd_vzip2_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2_t __simd_vzip2_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2_t __simd_vzip2_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16_t __simd_vzip2q_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16_t __simd_vzip2q_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16_t __simd_vzip2q_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8_t __simd_vzip2q_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8_t __simd_vzip2q_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8_t __simd_vzip2q_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4_t __simd_vzip2q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4_t __simd_vzip2q_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4_t __simd_vzip2q_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2_t __simd_vzip2q_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2_t __simd_vzip2q_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2_t __simd_vzip2q_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8x2_t __simd_vtrn_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8x2_t __simd_vtrn_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8x2_t __simd_vtrn_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4x2_t __simd_vtrn_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4x2_t __simd_vtrn_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4x2_t __simd_vtrn_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2x2_t __simd_vtrn_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2x2_t __simd_vtrn_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2x2_t __simd_vtrn_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16x2_t __simd_vtrnq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16x2_t __simd_vtrnq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16x2_t __simd_vtrnq_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8x2_t __simd_vtrnq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8x2_t __simd_vtrnq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8x2_t __simd_vtrnq_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4x2_t __simd_vtrnq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4x2_t __simd_vtrnq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4x2_t __simd_vtrnq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2x2_t __simd_vtrnq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2x2_t __simd_vtrnq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2x2_t __simd_vtrnq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8x2_t __simd_vzip_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8x2_t __simd_vzip_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8x2_t __simd_vzip_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4x2_t __simd_vzip_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4x2_t __simd_vzip_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4x2_t __simd_vzip_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2x2_t __simd_vzip_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2x2_t __simd_vzip_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2x2_t __simd_vzip_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16x2_t __simd_vzipq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16x2_t __simd_vzipq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16x2_t __simd_vzipq_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8x2_t __simd_vzipq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8x2_t __simd_vzipq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8x2_t __simd_vzipq_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4x2_t __simd_vzipq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4x2_t __simd_vzipq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4x2_t __simd_vzipq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2x2_t __simd_vzipq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2x2_t __simd_vzipq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2x2_t __simd_vzipq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
uint8x8x2_t __simd_vuzp_u8(uint8x8_t a, uint8x8_t b) __NEONSAFE;
int8x8x2_t __simd_vuzp_s8(int8x8_t a, int8x8_t b) __NEONSAFE;
poly8x8x2_t __simd_vuzp_p8(poly8x8_t a, poly8x8_t b) __NEONSAFE;
uint16x4x2_t __simd_vuzp_u16(uint16x4_t a, uint16x4_t b) __NEONSAFE;
int16x4x2_t __simd_vuzp_s16(int16x4_t a, int16x4_t b) __NEONSAFE;
poly16x4x2_t __simd_vuzp_p16(poly16x4_t a, poly16x4_t b) __NEONSAFE;
uint32x2x2_t __simd_vuzp_u32(uint32x2_t a, uint32x2_t b) __NEONSAFE;
int32x2x2_t __simd_vuzp_s32(int32x2_t a, int32x2_t b) __NEONSAFE;
float32x2x2_t __simd_vuzp_f32(float32x2_t a, float32x2_t b) __NEONSAFE;
uint8x16x2_t __simd_vuzpq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
int8x16x2_t __simd_vuzpq_s8(int8x16_t a, int8x16_t b) __NEONSAFE;
poly8x16x2_t __simd_vuzpq_p8(poly8x16_t a, poly8x16_t b) __NEONSAFE;
uint16x8x2_t __simd_vuzpq_u16(uint16x8_t a, uint16x8_t b) __NEONSAFE;
int16x8x2_t __simd_vuzpq_s16(int16x8_t a, int16x8_t b) __NEONSAFE;
poly16x8x2_t __simd_vuzpq_p16(poly16x8_t a, poly16x8_t b) __NEONSAFE;
uint32x4x2_t __simd_vuzpq_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
int32x4x2_t __simd_vuzpq_s32(int32x4_t a, int32x4_t b) __NEONSAFE;
float32x4x2_t __simd_vuzpq_f32(float32x4_t a, float32x4_t b) __NEONSAFE;
uint64x2x2_t __simd_vuzpq_u64(uint64x2_t a, uint64x2_t b) __NEONSAFE;
int64x2x2_t __simd_vuzpq_s64(int64x2_t a, int64x2_t b) __NEONSAFE;
float64x2x2_t __simd_vuzpq_f64(float64x2_t a, float64x2_t b) __NEONSAFE;
int16x4_t __simd_vqrdmlah_s16(int16x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int32x2_t __simd_vqrdmlah_s32(int32x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int16x8_t __simd_vqrdmlahq_s16(int16x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int32x4_t __simd_vqrdmlahq_s32(int32x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
int16_t __simd_vqrdmlahh_s16(int16_t a, int16_t b, int16_t c) __NEONSAFE;
int32_t __simd_vqrdmlahs_s32(int32_t a, int32_t b, int32_t c) __NEONSAFE;
int16x4_t __simd_vqrdmlah_lane_s16(int16x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vqrdmlah_laneq_s16(int16x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vqrdmlah_n_s16(int16x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
int32x2_t __simd_vqrdmlah_lane_s32(int32x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vqrdmlah_laneq_s32(int32x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vqrdmlah_n_s32(int32x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
int16x8_t __simd_vqrdmlahq_lane_s16(int16x8_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vqrdmlahq_laneq_s16(int16x8_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vqrdmlahq_n_s16(int16x8_t a, int16x8_t b, int16_t c) __NEONSAFE;
int32x4_t __simd_vqrdmlahq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqrdmlahq_laneq_s32(int32x4_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqrdmlahq_n_s32(int32x4_t a, int32x4_t b, int32_t c) __NEONSAFE;
int16_t __simd_vqrdmlahh_lane_s16(int16_t a, int16_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16_t __simd_vqrdmlahh_laneq_s16(int16_t a, int16_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqrdmlahs_lane_s32(int32_t a, int32_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqrdmlahs_laneq_s32(int32_t a, int32_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vqrdmlsh_s16(int16x4_t a, int16x4_t b, int16x4_t c) __NEONSAFE;
int32x2_t __simd_vqrdmlsh_s32(int32x2_t a, int32x2_t b, int32x2_t c) __NEONSAFE;
int16x8_t __simd_vqrdmlshq_s16(int16x8_t a, int16x8_t b, int16x8_t c) __NEONSAFE;
int32x4_t __simd_vqrdmlshq_s32(int32x4_t a, int32x4_t b, int32x4_t c) __NEONSAFE;
int16_t __simd_vqrdmlshh_s16(int16_t a, int16_t b, int16_t c) __NEONSAFE;
int32_t __simd_vqrdmlshs_s32(int32_t a, int32_t b, int32_t c) __NEONSAFE;
int16x4_t __simd_vqrdmlsh_lane_s16(int16x4_t a, int16x4_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vqrdmlsh_laneq_s16(int16x4_t a, int16x4_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x4_t __simd_vqrdmlsh_n_s16(int16x4_t a, int16x4_t b, int16_t c) __NEONSAFE;
int32x2_t __simd_vqrdmlsh_lane_s32(int32x2_t a, int32x2_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vqrdmlsh_laneq_s32(int32x2_t a, int32x2_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x2_t __simd_vqrdmlsh_n_s32(int32x2_t a, int32x2_t b, int32_t c) __NEONSAFE;
int16x8_t __simd_vqrdmlshq_lane_s16(int16x8_t a, int16x8_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vqrdmlshq_laneq_s16(int16x8_t a, int16x8_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int16x8_t __simd_vqrdmlshq_n_s16(int16x8_t a, int16x8_t b, int16_t c) __NEONSAFE;
int32x4_t __simd_vqrdmlshq_lane_s32(int32x4_t a, int32x4_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqrdmlshq_laneq_s32(int32x4_t a, int32x4_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;
int32x4_t __simd_vqrdmlshq_n_s32(int32x4_t a, int32x4_t b, int32_t c) __NEONSAFE;
int16_t __simd_vqrdmlshh_lane_s16(int16_t a, int16_t b, int16x4_t c, __ghs_c_int__ d) __NEONSAFE;
int16_t __simd_vqrdmlshh_laneq_s16(int16_t a, int16_t b, int16x8_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqrdmlshs_lane_s32(int32_t a, int32_t b, int32x2_t c, __ghs_c_int__ d) __NEONSAFE;
int32_t __simd_vqrdmlshs_laneq_s32(int32_t a, int32_t b, int32x4_t c, __ghs_c_int__ d) __NEONSAFE;

uint8x16_t __simd_vaesdq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint8x16_t __simd_vaeseq_u8(uint8x16_t a, uint8x16_t b) __NEONSAFE;
uint8x16_t __simd_vaesimcq_u8(uint8x16_t a) __NEONSAFE;
uint8x16_t __simd_vaesmcq_u8(uint8x16_t a) __NEONSAFE;
uint32x4_t __simd_vsha1cq_u32(uint32x4_t a, uint32_t b, uint32x4_t c) __NEONSAFE;
uint32_t __simd_vsha1h_u32(uint32_t a) __NEONSAFE;
uint32x4_t __simd_vsha1mq_u32(uint32x4_t a, uint32_t b, uint32x4_t c) __NEONSAFE;
uint32x4_t __simd_vsha1pq_u32(uint32x4_t a, uint32_t b, uint32x4_t c) __NEONSAFE;
uint32x4_t __simd_vsha1su0q_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
uint32x4_t __simd_vsha1su1q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint32x4_t __simd_vsha256hq_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
uint32x4_t __simd_vsha256h2q_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;
uint32x4_t __simd_vsha256su0q_u32(uint32x4_t a, uint32x4_t b) __NEONSAFE;
uint32x4_t __simd_vsha256su1q_u32(uint32x4_t a, uint32x4_t b, uint32x4_t c) __NEONSAFE;

#define vld1_u8(a) __simd_vld1_u8(a)
#define vld1_s8(a) __simd_vld1_s8(a)
#define vld1_p8(a) __simd_vld1_p8(a)
#define vld1_u16(a) __simd_vld1_u16(a)
#define vld1_s16(a) __simd_vld1_s16(a)
#define vld1_p16(a) __simd_vld1_p16(a)
#define vld1_u32(a) __simd_vld1_u32(a)
#define vld1_s32(a) __simd_vld1_s32(a)
#define vld1_f32(a) __simd_vld1_f32(a)
#define vld1_u64(a) __simd_vld1_u64(a)
#define vld1_s64(a) __simd_vld1_s64(a)
#define vld1q_u8(a) __simd_vld1q_u8(a)
#define vld1q_s8(a) __simd_vld1q_s8(a)
#define vld1q_p8(a) __simd_vld1q_p8(a)
#define vld1q_u16(a) __simd_vld1q_u16(a)
#define vld1q_s16(a) __simd_vld1q_s16(a)
#define vld1q_p16(a) __simd_vld1q_p16(a)
#define vld1q_u32(a) __simd_vld1q_u32(a)
#define vld1q_s32(a) __simd_vld1q_s32(a)
#define vld1q_f32(a) __simd_vld1q_f32(a)
#define vld1q_u64(a) __simd_vld1q_u64(a)
#define vld1q_s64(a) __simd_vld1q_s64(a)
#define vld1q_f64(a) __simd_vld1q_f64(a)
#define vld1_u8_x2(a) __simd_vld1_u8_x2(a)
#define vld1_s8_x2(a) __simd_vld1_s8_x2(a)
#define vld1_p8_x2(a) __simd_vld1_p8_x2(a)
#define vld1_u16_x2(a) __simd_vld1_u16_x2(a)
#define vld1_s16_x2(a) __simd_vld1_s16_x2(a)
#define vld1_p16_x2(a) __simd_vld1_p16_x2(a)
#define vld1_u32_x2(a) __simd_vld1_u32_x2(a)
#define vld1_s32_x2(a) __simd_vld1_s32_x2(a)
#define vld1_f32_x2(a) __simd_vld1_f32_x2(a)
#define vld1_u64_x2(a) __simd_vld1_u64_x2(a)
#define vld1_s64_x2(a) __simd_vld1_s64_x2(a)
#define vld1q_u8_x2(a) __simd_vld1q_u8_x2(a)
#define vld1q_s8_x2(a) __simd_vld1q_s8_x2(a)
#define vld1q_p8_x2(a) __simd_vld1q_p8_x2(a)
#define vld1q_u16_x2(a) __simd_vld1q_u16_x2(a)
#define vld1q_s16_x2(a) __simd_vld1q_s16_x2(a)
#define vld1q_p16_x2(a) __simd_vld1q_p16_x2(a)
#define vld1q_u32_x2(a) __simd_vld1q_u32_x2(a)
#define vld1q_s32_x2(a) __simd_vld1q_s32_x2(a)
#define vld1q_f32_x2(a) __simd_vld1q_f32_x2(a)
#define vld1q_u64_x2(a) __simd_vld1q_u64_x2(a)
#define vld1q_s64_x2(a) __simd_vld1q_s64_x2(a)
#define vld1q_f64_x2(a) __simd_vld1q_f64_x2(a)
#define vld1_u8_x3(a) __simd_vld1_u8_x3(a)
#define vld1_s8_x3(a) __simd_vld1_s8_x3(a)
#define vld1_p8_x3(a) __simd_vld1_p8_x3(a)
#define vld1_u16_x3(a) __simd_vld1_u16_x3(a)
#define vld1_s16_x3(a) __simd_vld1_s16_x3(a)
#define vld1_p16_x3(a) __simd_vld1_p16_x3(a)
#define vld1_u32_x3(a) __simd_vld1_u32_x3(a)
#define vld1_s32_x3(a) __simd_vld1_s32_x3(a)
#define vld1_f32_x3(a) __simd_vld1_f32_x3(a)
#define vld1_u64_x3(a) __simd_vld1_u64_x3(a)
#define vld1_s64_x3(a) __simd_vld1_s64_x3(a)
#define vld1q_u8_x3(a) __simd_vld1q_u8_x3(a)
#define vld1q_s8_x3(a) __simd_vld1q_s8_x3(a)
#define vld1q_p8_x3(a) __simd_vld1q_p8_x3(a)
#define vld1q_u16_x3(a) __simd_vld1q_u16_x3(a)
#define vld1q_s16_x3(a) __simd_vld1q_s16_x3(a)
#define vld1q_p16_x3(a) __simd_vld1q_p16_x3(a)
#define vld1q_u32_x3(a) __simd_vld1q_u32_x3(a)
#define vld1q_s32_x3(a) __simd_vld1q_s32_x3(a)
#define vld1q_f32_x3(a) __simd_vld1q_f32_x3(a)
#define vld1q_u64_x3(a) __simd_vld1q_u64_x3(a)
#define vld1q_s64_x3(a) __simd_vld1q_s64_x3(a)
#define vld1q_f64_x3(a) __simd_vld1q_f64_x3(a)
#define vld1_u8_x4(a) __simd_vld1_u8_x4(a)
#define vld1_s8_x4(a) __simd_vld1_s8_x4(a)
#define vld1_p8_x4(a) __simd_vld1_p8_x4(a)
#define vld1_u16_x4(a) __simd_vld1_u16_x4(a)
#define vld1_s16_x4(a) __simd_vld1_s16_x4(a)
#define vld1_p16_x4(a) __simd_vld1_p16_x4(a)
#define vld1_u32_x4(a) __simd_vld1_u32_x4(a)
#define vld1_s32_x4(a) __simd_vld1_s32_x4(a)
#define vld1_f32_x4(a) __simd_vld1_f32_x4(a)
#define vld1_u64_x4(a) __simd_vld1_u64_x4(a)
#define vld1_s64_x4(a) __simd_vld1_s64_x4(a)
#define vld1q_u8_x4(a) __simd_vld1q_u8_x4(a)
#define vld1q_s8_x4(a) __simd_vld1q_s8_x4(a)
#define vld1q_p8_x4(a) __simd_vld1q_p8_x4(a)
#define vld1q_u16_x4(a) __simd_vld1q_u16_x4(a)
#define vld1q_s16_x4(a) __simd_vld1q_s16_x4(a)
#define vld1q_p16_x4(a) __simd_vld1q_p16_x4(a)
#define vld1q_u32_x4(a) __simd_vld1q_u32_x4(a)
#define vld1q_s32_x4(a) __simd_vld1q_s32_x4(a)
#define vld1q_f32_x4(a) __simd_vld1q_f32_x4(a)
#define vld1q_u64_x4(a) __simd_vld1q_u64_x4(a)
#define vld1q_s64_x4(a) __simd_vld1q_s64_x4(a)
#define vld1q_f64_x4(a) __simd_vld1q_f64_x4(a)
#define vld2_u8(a) __simd_vld2_u8(a)
#define vld2_s8(a) __simd_vld2_s8(a)
#define vld2_p8(a) __simd_vld2_p8(a)
#define vld2_u16(a) __simd_vld2_u16(a)
#define vld2_s16(a) __simd_vld2_s16(a)
#define vld2_p16(a) __simd_vld2_p16(a)
#define vld2_u32(a) __simd_vld2_u32(a)
#define vld2_s32(a) __simd_vld2_s32(a)
#define vld2_f32(a) __simd_vld2_f32(a)
#define vld2q_u8(a) __simd_vld2q_u8(a)
#define vld2q_s8(a) __simd_vld2q_s8(a)
#define vld2q_p8(a) __simd_vld2q_p8(a)
#define vld2q_u16(a) __simd_vld2q_u16(a)
#define vld2q_s16(a) __simd_vld2q_s16(a)
#define vld2q_p16(a) __simd_vld2q_p16(a)
#define vld2q_u32(a) __simd_vld2q_u32(a)
#define vld2q_s32(a) __simd_vld2q_s32(a)
#define vld2q_f32(a) __simd_vld2q_f32(a)
#define vld2q_u64(a) __simd_vld2q_u64(a)
#define vld2q_s64(a) __simd_vld2q_s64(a)
#define vld2q_f64(a) __simd_vld2q_f64(a)
#define vld3_u8(a) __simd_vld3_u8(a)
#define vld3_s8(a) __simd_vld3_s8(a)
#define vld3_p8(a) __simd_vld3_p8(a)
#define vld3_u16(a) __simd_vld3_u16(a)
#define vld3_s16(a) __simd_vld3_s16(a)
#define vld3_p16(a) __simd_vld3_p16(a)
#define vld3_u32(a) __simd_vld3_u32(a)
#define vld3_s32(a) __simd_vld3_s32(a)
#define vld3_f32(a) __simd_vld3_f32(a)
#define vld3q_u8(a) __simd_vld3q_u8(a)
#define vld3q_s8(a) __simd_vld3q_s8(a)
#define vld3q_p8(a) __simd_vld3q_p8(a)
#define vld3q_u16(a) __simd_vld3q_u16(a)
#define vld3q_s16(a) __simd_vld3q_s16(a)
#define vld3q_p16(a) __simd_vld3q_p16(a)
#define vld3q_u32(a) __simd_vld3q_u32(a)
#define vld3q_s32(a) __simd_vld3q_s32(a)
#define vld3q_f32(a) __simd_vld3q_f32(a)
#define vld3q_u64(a) __simd_vld3q_u64(a)
#define vld3q_s64(a) __simd_vld3q_s64(a)
#define vld3q_f64(a) __simd_vld3q_f64(a)
#define vld4_u8(a) __simd_vld4_u8(a)
#define vld4_s8(a) __simd_vld4_s8(a)
#define vld4_p8(a) __simd_vld4_p8(a)
#define vld4_u16(a) __simd_vld4_u16(a)
#define vld4_s16(a) __simd_vld4_s16(a)
#define vld4_p16(a) __simd_vld4_p16(a)
#define vld4_u32(a) __simd_vld4_u32(a)
#define vld4_s32(a) __simd_vld4_s32(a)
#define vld4_f32(a) __simd_vld4_f32(a)
#define vld4q_u8(a) __simd_vld4q_u8(a)
#define vld4q_s8(a) __simd_vld4q_s8(a)
#define vld4q_p8(a) __simd_vld4q_p8(a)
#define vld4q_u16(a) __simd_vld4q_u16(a)
#define vld4q_s16(a) __simd_vld4q_s16(a)
#define vld4q_p16(a) __simd_vld4q_p16(a)
#define vld4q_u32(a) __simd_vld4q_u32(a)
#define vld4q_s32(a) __simd_vld4q_s32(a)
#define vld4q_f32(a) __simd_vld4q_f32(a)
#define vld4q_u64(a) __simd_vld4q_u64(a)
#define vld4q_s64(a) __simd_vld4q_s64(a)
#define vld4q_f64(a) __simd_vld4q_f64(a)
#define vst1_u8(a,b) __simd_vst1_u8(a,b)
#define vst1_s8(a,b) __simd_vst1_s8(a,b)
#define vst1_p8(a,b) __simd_vst1_p8(a,b)
#define vst1_u16(a,b) __simd_vst1_u16(a,b)
#define vst1_s16(a,b) __simd_vst1_s16(a,b)
#define vst1_p16(a,b) __simd_vst1_p16(a,b)
#define vst1_u32(a,b) __simd_vst1_u32(a,b)
#define vst1_s32(a,b) __simd_vst1_s32(a,b)
#define vst1_f32(a,b) __simd_vst1_f32(a,b)
#define vst1_u64(a,b) __simd_vst1_u64(a,b)
#define vst1_s64(a,b) __simd_vst1_s64(a,b)
#define vst1q_u8(a,b) __simd_vst1q_u8(a,b)
#define vst1q_s8(a,b) __simd_vst1q_s8(a,b)
#define vst1q_p8(a,b) __simd_vst1q_p8(a,b)
#define vst1q_u16(a,b) __simd_vst1q_u16(a,b)
#define vst1q_s16(a,b) __simd_vst1q_s16(a,b)
#define vst1q_p16(a,b) __simd_vst1q_p16(a,b)
#define vst1q_u32(a,b) __simd_vst1q_u32(a,b)
#define vst1q_s32(a,b) __simd_vst1q_s32(a,b)
#define vst1q_f32(a,b) __simd_vst1q_f32(a,b)
#define vst1q_u64(a,b) __simd_vst1q_u64(a,b)
#define vst1q_s64(a,b) __simd_vst1q_s64(a,b)
#define vst1q_f64(a,b) __simd_vst1q_f64(a,b)
#define vst1_u8_x2(a,b) __simd_vst1_u8_x2(a,b)
#define vst1_s8_x2(a,b) __simd_vst1_s8_x2(a,b)
#define vst1_p8_x2(a,b) __simd_vst1_p8_x2(a,b)
#define vst1_u16_x2(a,b) __simd_vst1_u16_x2(a,b)
#define vst1_s16_x2(a,b) __simd_vst1_s16_x2(a,b)
#define vst1_p16_x2(a,b) __simd_vst1_p16_x2(a,b)
#define vst1_u32_x2(a,b) __simd_vst1_u32_x2(a,b)
#define vst1_s32_x2(a,b) __simd_vst1_s32_x2(a,b)
#define vst1_f32_x2(a,b) __simd_vst1_f32_x2(a,b)
#define vst1_u64_x2(a,b) __simd_vst1_u64_x2(a,b)
#define vst1_s64_x2(a,b) __simd_vst1_s64_x2(a,b)
#define vst1q_u8_x2(a,b) __simd_vst1q_u8_x2(a,b)
#define vst1q_s8_x2(a,b) __simd_vst1q_s8_x2(a,b)
#define vst1q_p8_x2(a,b) __simd_vst1q_p8_x2(a,b)
#define vst1q_u16_x2(a,b) __simd_vst1q_u16_x2(a,b)
#define vst1q_s16_x2(a,b) __simd_vst1q_s16_x2(a,b)
#define vst1q_p16_x2(a,b) __simd_vst1q_p16_x2(a,b)
#define vst1q_u32_x2(a,b) __simd_vst1q_u32_x2(a,b)
#define vst1q_s32_x2(a,b) __simd_vst1q_s32_x2(a,b)
#define vst1q_f32_x2(a,b) __simd_vst1q_f32_x2(a,b)
#define vst1q_u64_x2(a,b) __simd_vst1q_u64_x2(a,b)
#define vst1q_s64_x2(a,b) __simd_vst1q_s64_x2(a,b)
#define vst1q_f64_x2(a,b) __simd_vst1q_f64_x2(a,b)
#define vst1_u8_x3(a,b) __simd_vst1_u8_x3(a,b)
#define vst1_s8_x3(a,b) __simd_vst1_s8_x3(a,b)
#define vst1_p8_x3(a,b) __simd_vst1_p8_x3(a,b)
#define vst1_u16_x3(a,b) __simd_vst1_u16_x3(a,b)
#define vst1_s16_x3(a,b) __simd_vst1_s16_x3(a,b)
#define vst1_p16_x3(a,b) __simd_vst1_p16_x3(a,b)
#define vst1_u32_x3(a,b) __simd_vst1_u32_x3(a,b)
#define vst1_s32_x3(a,b) __simd_vst1_s32_x3(a,b)
#define vst1_f32_x3(a,b) __simd_vst1_f32_x3(a,b)
#define vst1_u64_x3(a,b) __simd_vst1_u64_x3(a,b)
#define vst1_s64_x3(a,b) __simd_vst1_s64_x3(a,b)
#define vst1q_u8_x3(a,b) __simd_vst1q_u8_x3(a,b)
#define vst1q_s8_x3(a,b) __simd_vst1q_s8_x3(a,b)
#define vst1q_p8_x3(a,b) __simd_vst1q_p8_x3(a,b)
#define vst1q_u16_x3(a,b) __simd_vst1q_u16_x3(a,b)
#define vst1q_s16_x3(a,b) __simd_vst1q_s16_x3(a,b)
#define vst1q_p16_x3(a,b) __simd_vst1q_p16_x3(a,b)
#define vst1q_u32_x3(a,b) __simd_vst1q_u32_x3(a,b)
#define vst1q_s32_x3(a,b) __simd_vst1q_s32_x3(a,b)
#define vst1q_f32_x3(a,b) __simd_vst1q_f32_x3(a,b)
#define vst1q_u64_x3(a,b) __simd_vst1q_u64_x3(a,b)
#define vst1q_s64_x3(a,b) __simd_vst1q_s64_x3(a,b)
#define vst1q_f64_x3(a,b) __simd_vst1q_f64_x3(a,b)
#define vst1_u8_x4(a,b) __simd_vst1_u8_x4(a,b)
#define vst1_s8_x4(a,b) __simd_vst1_s8_x4(a,b)
#define vst1_p8_x4(a,b) __simd_vst1_p8_x4(a,b)
#define vst1_u16_x4(a,b) __simd_vst1_u16_x4(a,b)
#define vst1_s16_x4(a,b) __simd_vst1_s16_x4(a,b)
#define vst1_p16_x4(a,b) __simd_vst1_p16_x4(a,b)
#define vst1_u32_x4(a,b) __simd_vst1_u32_x4(a,b)
#define vst1_s32_x4(a,b) __simd_vst1_s32_x4(a,b)
#define vst1_f32_x4(a,b) __simd_vst1_f32_x4(a,b)
#define vst1_u64_x4(a,b) __simd_vst1_u64_x4(a,b)
#define vst1_s64_x4(a,b) __simd_vst1_s64_x4(a,b)
#define vst1q_u8_x4(a,b) __simd_vst1q_u8_x4(a,b)
#define vst1q_s8_x4(a,b) __simd_vst1q_s8_x4(a,b)
#define vst1q_p8_x4(a,b) __simd_vst1q_p8_x4(a,b)
#define vst1q_u16_x4(a,b) __simd_vst1q_u16_x4(a,b)
#define vst1q_s16_x4(a,b) __simd_vst1q_s16_x4(a,b)
#define vst1q_p16_x4(a,b) __simd_vst1q_p16_x4(a,b)
#define vst1q_u32_x4(a,b) __simd_vst1q_u32_x4(a,b)
#define vst1q_s32_x4(a,b) __simd_vst1q_s32_x4(a,b)
#define vst1q_f32_x4(a,b) __simd_vst1q_f32_x4(a,b)
#define vst1q_u64_x4(a,b) __simd_vst1q_u64_x4(a,b)
#define vst1q_s64_x4(a,b) __simd_vst1q_s64_x4(a,b)
#define vst1q_f64_x4(a,b) __simd_vst1q_f64_x4(a,b)
#define vst2_u8(a,b) __simd_vst2_u8(a,b)
#define vst2_s8(a,b) __simd_vst2_s8(a,b)
#define vst2_p8(a,b) __simd_vst2_p8(a,b)
#define vst2_u16(a,b) __simd_vst2_u16(a,b)
#define vst2_s16(a,b) __simd_vst2_s16(a,b)
#define vst2_p16(a,b) __simd_vst2_p16(a,b)
#define vst2_u32(a,b) __simd_vst2_u32(a,b)
#define vst2_s32(a,b) __simd_vst2_s32(a,b)
#define vst2_f32(a,b) __simd_vst2_f32(a,b)
#define vst2q_u8(a,b) __simd_vst2q_u8(a,b)
#define vst2q_s8(a,b) __simd_vst2q_s8(a,b)
#define vst2q_p8(a,b) __simd_vst2q_p8(a,b)
#define vst2q_u16(a,b) __simd_vst2q_u16(a,b)
#define vst2q_s16(a,b) __simd_vst2q_s16(a,b)
#define vst2q_p16(a,b) __simd_vst2q_p16(a,b)
#define vst2q_u32(a,b) __simd_vst2q_u32(a,b)
#define vst2q_s32(a,b) __simd_vst2q_s32(a,b)
#define vst2q_f32(a,b) __simd_vst2q_f32(a,b)
#define vst2q_u64(a,b) __simd_vst2q_u64(a,b)
#define vst2q_s64(a,b) __simd_vst2q_s64(a,b)
#define vst2q_f64(a,b) __simd_vst2q_f64(a,b)
#define vst3_u8(a,b) __simd_vst3_u8(a,b)
#define vst3_s8(a,b) __simd_vst3_s8(a,b)
#define vst3_p8(a,b) __simd_vst3_p8(a,b)
#define vst3_u16(a,b) __simd_vst3_u16(a,b)
#define vst3_s16(a,b) __simd_vst3_s16(a,b)
#define vst3_p16(a,b) __simd_vst3_p16(a,b)
#define vst3_u32(a,b) __simd_vst3_u32(a,b)
#define vst3_s32(a,b) __simd_vst3_s32(a,b)
#define vst3_f32(a,b) __simd_vst3_f32(a,b)
#define vst3q_u8(a,b) __simd_vst3q_u8(a,b)
#define vst3q_s8(a,b) __simd_vst3q_s8(a,b)
#define vst3q_p8(a,b) __simd_vst3q_p8(a,b)
#define vst3q_u16(a,b) __simd_vst3q_u16(a,b)
#define vst3q_s16(a,b) __simd_vst3q_s16(a,b)
#define vst3q_p16(a,b) __simd_vst3q_p16(a,b)
#define vst3q_u32(a,b) __simd_vst3q_u32(a,b)
#define vst3q_s32(a,b) __simd_vst3q_s32(a,b)
#define vst3q_f32(a,b) __simd_vst3q_f32(a,b)
#define vst3q_u64(a,b) __simd_vst3q_u64(a,b)
#define vst3q_s64(a,b) __simd_vst3q_s64(a,b)
#define vst3q_f64(a,b) __simd_vst3q_f64(a,b)
#define vst4_u8(a,b) __simd_vst4_u8(a,b)
#define vst4_s8(a,b) __simd_vst4_s8(a,b)
#define vst4_p8(a,b) __simd_vst4_p8(a,b)
#define vst4_u16(a,b) __simd_vst4_u16(a,b)
#define vst4_s16(a,b) __simd_vst4_s16(a,b)
#define vst4_p16(a,b) __simd_vst4_p16(a,b)
#define vst4_u32(a,b) __simd_vst4_u32(a,b)
#define vst4_s32(a,b) __simd_vst4_s32(a,b)
#define vst4_f32(a,b) __simd_vst4_f32(a,b)
#define vst4q_u8(a,b) __simd_vst4q_u8(a,b)
#define vst4q_s8(a,b) __simd_vst4q_s8(a,b)
#define vst4q_p8(a,b) __simd_vst4q_p8(a,b)
#define vst4q_u16(a,b) __simd_vst4q_u16(a,b)
#define vst4q_s16(a,b) __simd_vst4q_s16(a,b)
#define vst4q_p16(a,b) __simd_vst4q_p16(a,b)
#define vst4q_u32(a,b) __simd_vst4q_u32(a,b)
#define vst4q_s32(a,b) __simd_vst4q_s32(a,b)
#define vst4q_f32(a,b) __simd_vst4q_f32(a,b)
#define vst4q_u64(a,b) __simd_vst4q_u64(a,b)
#define vst4q_s64(a,b) __simd_vst4q_s64(a,b)
#define vst4q_f64(a,b) __simd_vst4q_f64(a,b)
#define vld1_lane_u8(a,b,c) __simd_vld1_lane_u8(a,b,c)
#define vld1_lane_s8(a,b,c) __simd_vld1_lane_s8(a,b,c)
#define vld1_lane_p8(a,b,c) __simd_vld1_lane_p8(a,b,c)
#define vld1q_lane_u8(a,b,c) __simd_vld1q_lane_u8(a,b,c)
#define vld1q_lane_s8(a,b,c) __simd_vld1q_lane_s8(a,b,c)
#define vld1q_lane_p8(a,b,c) __simd_vld1q_lane_p8(a,b,c)
#define vld1_lane_u16(a,b,c) __simd_vld1_lane_u16(a,b,c)
#define vld1_lane_s16(a,b,c) __simd_vld1_lane_s16(a,b,c)
#define vld1_lane_p16(a,b,c) __simd_vld1_lane_p16(a,b,c)
#define vld1q_lane_u16(a,b,c) __simd_vld1q_lane_u16(a,b,c)
#define vld1q_lane_s16(a,b,c) __simd_vld1q_lane_s16(a,b,c)
#define vld1q_lane_p16(a,b,c) __simd_vld1q_lane_p16(a,b,c)
#define vld1_lane_u32(a,b,c) __simd_vld1_lane_u32(a,b,c)
#define vld1_lane_s32(a,b,c) __simd_vld1_lane_s32(a,b,c)
#define vld1_lane_f32(a,b,c) __simd_vld1_lane_f32(a,b,c)
#define vld1q_lane_u32(a,b,c) __simd_vld1q_lane_u32(a,b,c)
#define vld1q_lane_s32(a,b,c) __simd_vld1q_lane_s32(a,b,c)
#define vld1q_lane_f32(a,b,c) __simd_vld1q_lane_f32(a,b,c)
#define vld1_lane_u64(a,b,c) __simd_vld1_lane_u64(a,b,c)
#define vld1_lane_s64(a,b,c) __simd_vld1_lane_s64(a,b,c)
#define vld1q_lane_u64(a,b,c) __simd_vld1q_lane_u64(a,b,c)
#define vld1q_lane_s64(a,b,c) __simd_vld1q_lane_s64(a,b,c)
#define vld1q_lane_f64(a,b,c) __simd_vld1q_lane_f64(a,b,c)
#define vld2_lane_u8(a,b,c) __simd_vld2_lane_u8(a,b,c)
#define vld2_lane_s8(a,b,c) __simd_vld2_lane_s8(a,b,c)
#define vld2_lane_p8(a,b,c) __simd_vld2_lane_p8(a,b,c)
#define vld2q_lane_u8(a,b,c) __simd_vld2q_lane_u8(a,b,c)
#define vld2q_lane_s8(a,b,c) __simd_vld2q_lane_s8(a,b,c)
#define vld2q_lane_p8(a,b,c) __simd_vld2q_lane_p8(a,b,c)
#define vld2_lane_u16(a,b,c) __simd_vld2_lane_u16(a,b,c)
#define vld2_lane_s16(a,b,c) __simd_vld2_lane_s16(a,b,c)
#define vld2_lane_p16(a,b,c) __simd_vld2_lane_p16(a,b,c)
#define vld2q_lane_u16(a,b,c) __simd_vld2q_lane_u16(a,b,c)
#define vld2q_lane_s16(a,b,c) __simd_vld2q_lane_s16(a,b,c)
#define vld2q_lane_p16(a,b,c) __simd_vld2q_lane_p16(a,b,c)
#define vld2_lane_u32(a,b,c) __simd_vld2_lane_u32(a,b,c)
#define vld2_lane_s32(a,b,c) __simd_vld2_lane_s32(a,b,c)
#define vld2_lane_f32(a,b,c) __simd_vld2_lane_f32(a,b,c)
#define vld2q_lane_u32(a,b,c) __simd_vld2q_lane_u32(a,b,c)
#define vld2q_lane_s32(a,b,c) __simd_vld2q_lane_s32(a,b,c)
#define vld2q_lane_f32(a,b,c) __simd_vld2q_lane_f32(a,b,c)
#define vld2_lane_u64(a,b,c) __simd_vld2_lane_u64(a,b,c)
#define vld2_lane_s64(a,b,c) __simd_vld2_lane_s64(a,b,c)
#define vld2q_lane_u64(a,b,c) __simd_vld2q_lane_u64(a,b,c)
#define vld2q_lane_s64(a,b,c) __simd_vld2q_lane_s64(a,b,c)
#define vld2q_lane_f64(a,b,c) __simd_vld2q_lane_f64(a,b,c)
#define vld3_lane_u8(a,b,c) __simd_vld3_lane_u8(a,b,c)
#define vld3_lane_s8(a,b,c) __simd_vld3_lane_s8(a,b,c)
#define vld3_lane_p8(a,b,c) __simd_vld3_lane_p8(a,b,c)
#define vld3q_lane_u8(a,b,c) __simd_vld3q_lane_u8(a,b,c)
#define vld3q_lane_s8(a,b,c) __simd_vld3q_lane_s8(a,b,c)
#define vld3q_lane_p8(a,b,c) __simd_vld3q_lane_p8(a,b,c)
#define vld3_lane_u16(a,b,c) __simd_vld3_lane_u16(a,b,c)
#define vld3_lane_s16(a,b,c) __simd_vld3_lane_s16(a,b,c)
#define vld3_lane_p16(a,b,c) __simd_vld3_lane_p16(a,b,c)
#define vld3q_lane_u16(a,b,c) __simd_vld3q_lane_u16(a,b,c)
#define vld3q_lane_s16(a,b,c) __simd_vld3q_lane_s16(a,b,c)
#define vld3q_lane_p16(a,b,c) __simd_vld3q_lane_p16(a,b,c)
#define vld3_lane_u32(a,b,c) __simd_vld3_lane_u32(a,b,c)
#define vld3_lane_s32(a,b,c) __simd_vld3_lane_s32(a,b,c)
#define vld3_lane_f32(a,b,c) __simd_vld3_lane_f32(a,b,c)
#define vld3q_lane_u32(a,b,c) __simd_vld3q_lane_u32(a,b,c)
#define vld3q_lane_s32(a,b,c) __simd_vld3q_lane_s32(a,b,c)
#define vld3q_lane_f32(a,b,c) __simd_vld3q_lane_f32(a,b,c)
#define vld3_lane_u64(a,b,c) __simd_vld3_lane_u64(a,b,c)
#define vld3_lane_s64(a,b,c) __simd_vld3_lane_s64(a,b,c)
#define vld3q_lane_u64(a,b,c) __simd_vld3q_lane_u64(a,b,c)
#define vld3q_lane_s64(a,b,c) __simd_vld3q_lane_s64(a,b,c)
#define vld3q_lane_f64(a,b,c) __simd_vld3q_lane_f64(a,b,c)
#define vld4_lane_u8(a,b,c) __simd_vld4_lane_u8(a,b,c)
#define vld4_lane_s8(a,b,c) __simd_vld4_lane_s8(a,b,c)
#define vld4_lane_p8(a,b,c) __simd_vld4_lane_p8(a,b,c)
#define vld4q_lane_u8(a,b,c) __simd_vld4q_lane_u8(a,b,c)
#define vld4q_lane_s8(a,b,c) __simd_vld4q_lane_s8(a,b,c)
#define vld4q_lane_p8(a,b,c) __simd_vld4q_lane_p8(a,b,c)
#define vld4_lane_u16(a,b,c) __simd_vld4_lane_u16(a,b,c)
#define vld4_lane_s16(a,b,c) __simd_vld4_lane_s16(a,b,c)
#define vld4_lane_p16(a,b,c) __simd_vld4_lane_p16(a,b,c)
#define vld4q_lane_u16(a,b,c) __simd_vld4q_lane_u16(a,b,c)
#define vld4q_lane_s16(a,b,c) __simd_vld4q_lane_s16(a,b,c)
#define vld4q_lane_p16(a,b,c) __simd_vld4q_lane_p16(a,b,c)
#define vld4_lane_u32(a,b,c) __simd_vld4_lane_u32(a,b,c)
#define vld4_lane_s32(a,b,c) __simd_vld4_lane_s32(a,b,c)
#define vld4_lane_f32(a,b,c) __simd_vld4_lane_f32(a,b,c)
#define vld4q_lane_u32(a,b,c) __simd_vld4q_lane_u32(a,b,c)
#define vld4q_lane_s32(a,b,c) __simd_vld4q_lane_s32(a,b,c)
#define vld4q_lane_f32(a,b,c) __simd_vld4q_lane_f32(a,b,c)
#define vld4_lane_u64(a,b,c) __simd_vld4_lane_u64(a,b,c)
#define vld4_lane_s64(a,b,c) __simd_vld4_lane_s64(a,b,c)
#define vld4q_lane_u64(a,b,c) __simd_vld4q_lane_u64(a,b,c)
#define vld4q_lane_s64(a,b,c) __simd_vld4q_lane_s64(a,b,c)
#define vld4q_lane_f64(a,b,c) __simd_vld4q_lane_f64(a,b,c)
#define vst1_lane_u8(a,b,c) __simd_vst1_lane_u8(a,b,c)
#define vst1_lane_s8(a,b,c) __simd_vst1_lane_s8(a,b,c)
#define vst1_lane_p8(a,b,c) __simd_vst1_lane_p8(a,b,c)
#define vst1q_lane_u8(a,b,c) __simd_vst1q_lane_u8(a,b,c)
#define vst1q_lane_s8(a,b,c) __simd_vst1q_lane_s8(a,b,c)
#define vst1q_lane_p8(a,b,c) __simd_vst1q_lane_p8(a,b,c)
#define vst1_lane_u16(a,b,c) __simd_vst1_lane_u16(a,b,c)
#define vst1_lane_s16(a,b,c) __simd_vst1_lane_s16(a,b,c)
#define vst1_lane_p16(a,b,c) __simd_vst1_lane_p16(a,b,c)
#define vst1q_lane_u16(a,b,c) __simd_vst1q_lane_u16(a,b,c)
#define vst1q_lane_s16(a,b,c) __simd_vst1q_lane_s16(a,b,c)
#define vst1q_lane_p16(a,b,c) __simd_vst1q_lane_p16(a,b,c)
#define vst1_lane_u32(a,b,c) __simd_vst1_lane_u32(a,b,c)
#define vst1_lane_s32(a,b,c) __simd_vst1_lane_s32(a,b,c)
#define vst1_lane_f32(a,b,c) __simd_vst1_lane_f32(a,b,c)
#define vst1q_lane_u32(a,b,c) __simd_vst1q_lane_u32(a,b,c)
#define vst1q_lane_s32(a,b,c) __simd_vst1q_lane_s32(a,b,c)
#define vst1q_lane_f32(a,b,c) __simd_vst1q_lane_f32(a,b,c)
#define vst1_lane_u64(a,b,c) __simd_vst1_lane_u64(a,b,c)
#define vst1_lane_s64(a,b,c) __simd_vst1_lane_s64(a,b,c)
#define vst1q_lane_u64(a,b,c) __simd_vst1q_lane_u64(a,b,c)
#define vst1q_lane_s64(a,b,c) __simd_vst1q_lane_s64(a,b,c)
#define vst1q_lane_f64(a,b,c) __simd_vst1q_lane_f64(a,b,c)
#define vst2_lane_u8(a,b,c) __simd_vst2_lane_u8(a,b,c)
#define vst2_lane_s8(a,b,c) __simd_vst2_lane_s8(a,b,c)
#define vst2_lane_p8(a,b,c) __simd_vst2_lane_p8(a,b,c)
#define vst2q_lane_u8(a,b,c) __simd_vst2q_lane_u8(a,b,c)
#define vst2q_lane_s8(a,b,c) __simd_vst2q_lane_s8(a,b,c)
#define vst2q_lane_p8(a,b,c) __simd_vst2q_lane_p8(a,b,c)
#define vst2_lane_u16(a,b,c) __simd_vst2_lane_u16(a,b,c)
#define vst2_lane_s16(a,b,c) __simd_vst2_lane_s16(a,b,c)
#define vst2_lane_p16(a,b,c) __simd_vst2_lane_p16(a,b,c)
#define vst2q_lane_u16(a,b,c) __simd_vst2q_lane_u16(a,b,c)
#define vst2q_lane_s16(a,b,c) __simd_vst2q_lane_s16(a,b,c)
#define vst2q_lane_p16(a,b,c) __simd_vst2q_lane_p16(a,b,c)
#define vst2_lane_u32(a,b,c) __simd_vst2_lane_u32(a,b,c)
#define vst2_lane_s32(a,b,c) __simd_vst2_lane_s32(a,b,c)
#define vst2_lane_f32(a,b,c) __simd_vst2_lane_f32(a,b,c)
#define vst2q_lane_u32(a,b,c) __simd_vst2q_lane_u32(a,b,c)
#define vst2q_lane_s32(a,b,c) __simd_vst2q_lane_s32(a,b,c)
#define vst2q_lane_f32(a,b,c) __simd_vst2q_lane_f32(a,b,c)
#define vst2_lane_u64(a,b,c) __simd_vst2_lane_u64(a,b,c)
#define vst2_lane_s64(a,b,c) __simd_vst2_lane_s64(a,b,c)
#define vst2q_lane_u64(a,b,c) __simd_vst2q_lane_u64(a,b,c)
#define vst2q_lane_s64(a,b,c) __simd_vst2q_lane_s64(a,b,c)
#define vst2q_lane_f64(a,b,c) __simd_vst2q_lane_f64(a,b,c)
#define vst3_lane_u8(a,b,c) __simd_vst3_lane_u8(a,b,c)
#define vst3_lane_s8(a,b,c) __simd_vst3_lane_s8(a,b,c)
#define vst3_lane_p8(a,b,c) __simd_vst3_lane_p8(a,b,c)
#define vst3q_lane_u8(a,b,c) __simd_vst3q_lane_u8(a,b,c)
#define vst3q_lane_s8(a,b,c) __simd_vst3q_lane_s8(a,b,c)
#define vst3q_lane_p8(a,b,c) __simd_vst3q_lane_p8(a,b,c)
#define vst3_lane_u16(a,b,c) __simd_vst3_lane_u16(a,b,c)
#define vst3_lane_s16(a,b,c) __simd_vst3_lane_s16(a,b,c)
#define vst3_lane_p16(a,b,c) __simd_vst3_lane_p16(a,b,c)
#define vst3q_lane_u16(a,b,c) __simd_vst3q_lane_u16(a,b,c)
#define vst3q_lane_s16(a,b,c) __simd_vst3q_lane_s16(a,b,c)
#define vst3q_lane_p16(a,b,c) __simd_vst3q_lane_p16(a,b,c)
#define vst3_lane_u32(a,b,c) __simd_vst3_lane_u32(a,b,c)
#define vst3_lane_s32(a,b,c) __simd_vst3_lane_s32(a,b,c)
#define vst3_lane_f32(a,b,c) __simd_vst3_lane_f32(a,b,c)
#define vst3q_lane_u32(a,b,c) __simd_vst3q_lane_u32(a,b,c)
#define vst3q_lane_s32(a,b,c) __simd_vst3q_lane_s32(a,b,c)
#define vst3q_lane_f32(a,b,c) __simd_vst3q_lane_f32(a,b,c)
#define vst3_lane_u64(a,b,c) __simd_vst3_lane_u64(a,b,c)
#define vst3_lane_s64(a,b,c) __simd_vst3_lane_s64(a,b,c)
#define vst3q_lane_u64(a,b,c) __simd_vst3q_lane_u64(a,b,c)
#define vst3q_lane_s64(a,b,c) __simd_vst3q_lane_s64(a,b,c)
#define vst3q_lane_f64(a,b,c) __simd_vst3q_lane_f64(a,b,c)
#define vst4_lane_u8(a,b,c) __simd_vst4_lane_u8(a,b,c)
#define vst4_lane_s8(a,b,c) __simd_vst4_lane_s8(a,b,c)
#define vst4_lane_p8(a,b,c) __simd_vst4_lane_p8(a,b,c)
#define vst4q_lane_u8(a,b,c) __simd_vst4q_lane_u8(a,b,c)
#define vst4q_lane_s8(a,b,c) __simd_vst4q_lane_s8(a,b,c)
#define vst4q_lane_p8(a,b,c) __simd_vst4q_lane_p8(a,b,c)
#define vst4_lane_u16(a,b,c) __simd_vst4_lane_u16(a,b,c)
#define vst4_lane_s16(a,b,c) __simd_vst4_lane_s16(a,b,c)
#define vst4_lane_p16(a,b,c) __simd_vst4_lane_p16(a,b,c)
#define vst4q_lane_u16(a,b,c) __simd_vst4q_lane_u16(a,b,c)
#define vst4q_lane_s16(a,b,c) __simd_vst4q_lane_s16(a,b,c)
#define vst4q_lane_p16(a,b,c) __simd_vst4q_lane_p16(a,b,c)
#define vst4_lane_u32(a,b,c) __simd_vst4_lane_u32(a,b,c)
#define vst4_lane_s32(a,b,c) __simd_vst4_lane_s32(a,b,c)
#define vst4_lane_f32(a,b,c) __simd_vst4_lane_f32(a,b,c)
#define vst4q_lane_u32(a,b,c) __simd_vst4q_lane_u32(a,b,c)
#define vst4q_lane_s32(a,b,c) __simd_vst4q_lane_s32(a,b,c)
#define vst4q_lane_f32(a,b,c) __simd_vst4q_lane_f32(a,b,c)
#define vst4_lane_u64(a,b,c) __simd_vst4_lane_u64(a,b,c)
#define vst4_lane_s64(a,b,c) __simd_vst4_lane_s64(a,b,c)
#define vst4q_lane_u64(a,b,c) __simd_vst4q_lane_u64(a,b,c)
#define vst4q_lane_s64(a,b,c) __simd_vst4q_lane_s64(a,b,c)
#define vst4q_lane_f64(a,b,c) __simd_vst4q_lane_f64(a,b,c)
#define vld1_dup_u8(a) __simd_vld1_dup_u8(a)
#define vld1_dup_s8(a) __simd_vld1_dup_s8(a)
#define vld1_dup_p8(a) __simd_vld1_dup_p8(a)
#define vld1_dup_u16(a) __simd_vld1_dup_u16(a)
#define vld1_dup_s16(a) __simd_vld1_dup_s16(a)
#define vld1_dup_p16(a) __simd_vld1_dup_p16(a)
#define vld1_dup_u32(a) __simd_vld1_dup_u32(a)
#define vld1_dup_s32(a) __simd_vld1_dup_s32(a)
#define vld1_dup_f32(a) __simd_vld1_dup_f32(a)
#define vld1_dup_u64(a) __simd_vld1_dup_u64(a)
#define vld1_dup_s64(a) __simd_vld1_dup_s64(a)
#define vld1q_dup_u8(a) __simd_vld1q_dup_u8(a)
#define vld1q_dup_s8(a) __simd_vld1q_dup_s8(a)
#define vld1q_dup_p8(a) __simd_vld1q_dup_p8(a)
#define vld1q_dup_u16(a) __simd_vld1q_dup_u16(a)
#define vld1q_dup_s16(a) __simd_vld1q_dup_s16(a)
#define vld1q_dup_p16(a) __simd_vld1q_dup_p16(a)
#define vld1q_dup_u32(a) __simd_vld1q_dup_u32(a)
#define vld1q_dup_s32(a) __simd_vld1q_dup_s32(a)
#define vld1q_dup_f32(a) __simd_vld1q_dup_f32(a)
#define vld1q_dup_u64(a) __simd_vld1q_dup_u64(a)
#define vld1q_dup_s64(a) __simd_vld1q_dup_s64(a)
#define vld1q_dup_f64(a) __simd_vld1q_dup_f64(a)
#define vld2_dup_u8(a) __simd_vld2_dup_u8(a)
#define vld2_dup_s8(a) __simd_vld2_dup_s8(a)
#define vld2_dup_p8(a) __simd_vld2_dup_p8(a)
#define vld2_dup_u16(a) __simd_vld2_dup_u16(a)
#define vld2_dup_s16(a) __simd_vld2_dup_s16(a)
#define vld2_dup_p16(a) __simd_vld2_dup_p16(a)
#define vld2_dup_u32(a) __simd_vld2_dup_u32(a)
#define vld2_dup_s32(a) __simd_vld2_dup_s32(a)
#define vld2_dup_f32(a) __simd_vld2_dup_f32(a)
#define vld2_dup_u64(a) __simd_vld2_dup_u64(a)
#define vld2_dup_s64(a) __simd_vld2_dup_s64(a)
#define vld2q_dup_u8(a) __simd_vld2q_dup_u8(a)
#define vld2q_dup_s8(a) __simd_vld2q_dup_s8(a)
#define vld2q_dup_p8(a) __simd_vld2q_dup_p8(a)
#define vld2q_dup_u16(a) __simd_vld2q_dup_u16(a)
#define vld2q_dup_s16(a) __simd_vld2q_dup_s16(a)
#define vld2q_dup_p16(a) __simd_vld2q_dup_p16(a)
#define vld2q_dup_u32(a) __simd_vld2q_dup_u32(a)
#define vld2q_dup_s32(a) __simd_vld2q_dup_s32(a)
#define vld2q_dup_f32(a) __simd_vld2q_dup_f32(a)
#define vld2q_dup_u64(a) __simd_vld2q_dup_u64(a)
#define vld2q_dup_s64(a) __simd_vld2q_dup_s64(a)
#define vld2q_dup_f64(a) __simd_vld2q_dup_f64(a)
#define vld3_dup_u8(a) __simd_vld3_dup_u8(a)
#define vld3_dup_s8(a) __simd_vld3_dup_s8(a)
#define vld3_dup_p8(a) __simd_vld3_dup_p8(a)
#define vld3_dup_u16(a) __simd_vld3_dup_u16(a)
#define vld3_dup_s16(a) __simd_vld3_dup_s16(a)
#define vld3_dup_p16(a) __simd_vld3_dup_p16(a)
#define vld3_dup_u32(a) __simd_vld3_dup_u32(a)
#define vld3_dup_s32(a) __simd_vld3_dup_s32(a)
#define vld3_dup_f32(a) __simd_vld3_dup_f32(a)
#define vld3_dup_u64(a) __simd_vld3_dup_u64(a)
#define vld3_dup_s64(a) __simd_vld3_dup_s64(a)
#define vld3q_dup_u8(a) __simd_vld3q_dup_u8(a)
#define vld3q_dup_s8(a) __simd_vld3q_dup_s8(a)
#define vld3q_dup_p8(a) __simd_vld3q_dup_p8(a)
#define vld3q_dup_u16(a) __simd_vld3q_dup_u16(a)
#define vld3q_dup_s16(a) __simd_vld3q_dup_s16(a)
#define vld3q_dup_p16(a) __simd_vld3q_dup_p16(a)
#define vld3q_dup_u32(a) __simd_vld3q_dup_u32(a)
#define vld3q_dup_s32(a) __simd_vld3q_dup_s32(a)
#define vld3q_dup_f32(a) __simd_vld3q_dup_f32(a)
#define vld3q_dup_u64(a) __simd_vld3q_dup_u64(a)
#define vld3q_dup_s64(a) __simd_vld3q_dup_s64(a)
#define vld3q_dup_f64(a) __simd_vld3q_dup_f64(a)
#define vld4_dup_u8(a) __simd_vld4_dup_u8(a)
#define vld4_dup_s8(a) __simd_vld4_dup_s8(a)
#define vld4_dup_p8(a) __simd_vld4_dup_p8(a)
#define vld4_dup_u16(a) __simd_vld4_dup_u16(a)
#define vld4_dup_s16(a) __simd_vld4_dup_s16(a)
#define vld4_dup_p16(a) __simd_vld4_dup_p16(a)
#define vld4_dup_u32(a) __simd_vld4_dup_u32(a)
#define vld4_dup_s32(a) __simd_vld4_dup_s32(a)
#define vld4_dup_f32(a) __simd_vld4_dup_f32(a)
#define vld4_dup_u64(a) __simd_vld4_dup_u64(a)
#define vld4_dup_s64(a) __simd_vld4_dup_s64(a)
#define vld4q_dup_u8(a) __simd_vld4q_dup_u8(a)
#define vld4q_dup_s8(a) __simd_vld4q_dup_s8(a)
#define vld4q_dup_p8(a) __simd_vld4q_dup_p8(a)
#define vld4q_dup_u16(a) __simd_vld4q_dup_u16(a)
#define vld4q_dup_s16(a) __simd_vld4q_dup_s16(a)
#define vld4q_dup_p16(a) __simd_vld4q_dup_p16(a)
#define vld4q_dup_u32(a) __simd_vld4q_dup_u32(a)
#define vld4q_dup_s32(a) __simd_vld4q_dup_s32(a)
#define vld4q_dup_f32(a) __simd_vld4q_dup_f32(a)
#define vld4q_dup_u64(a) __simd_vld4q_dup_u64(a)
#define vld4q_dup_s64(a) __simd_vld4q_dup_s64(a)
#define vld4q_dup_f64(a) __simd_vld4q_dup_f64(a)
#define vadd_u8(a,b) __simd_vadd_u8(a,b)
#define vadd_s8(a,b) __simd_vadd_s8(a,b)
#define vadd_u16(a,b) __simd_vadd_u16(a,b)
#define vadd_s16(a,b) __simd_vadd_s16(a,b)
#define vadd_u32(a,b) __simd_vadd_u32(a,b)
#define vadd_s32(a,b) __simd_vadd_s32(a,b)
#define vaddq_u8(a,b) __simd_vaddq_u8(a,b)
#define vaddq_s8(a,b) __simd_vaddq_s8(a,b)
#define vaddq_u16(a,b) __simd_vaddq_u16(a,b)
#define vaddq_s16(a,b) __simd_vaddq_s16(a,b)
#define vaddq_u32(a,b) __simd_vaddq_u32(a,b)
#define vaddq_s32(a,b) __simd_vaddq_s32(a,b)
#define vaddq_u64(a,b) __simd_vaddq_u64(a,b)
#define vaddq_s64(a,b) __simd_vaddq_s64(a,b)
#define vsub_u8(a,b) __simd_vsub_u8(a,b)
#define vsub_s8(a,b) __simd_vsub_s8(a,b)
#define vsub_u16(a,b) __simd_vsub_u16(a,b)
#define vsub_s16(a,b) __simd_vsub_s16(a,b)
#define vsub_u32(a,b) __simd_vsub_u32(a,b)
#define vsub_s32(a,b) __simd_vsub_s32(a,b)
#define vsubq_u8(a,b) __simd_vsubq_u8(a,b)
#define vsubq_s8(a,b) __simd_vsubq_s8(a,b)
#define vsubq_u16(a,b) __simd_vsubq_u16(a,b)
#define vsubq_s16(a,b) __simd_vsubq_s16(a,b)
#define vsubq_u32(a,b) __simd_vsubq_u32(a,b)
#define vsubq_s32(a,b) __simd_vsubq_s32(a,b)
#define vsubq_u64(a,b) __simd_vsubq_u64(a,b)
#define vsubq_s64(a,b) __simd_vsubq_s64(a,b)
#define vpadd_u8(a,b) __simd_vpadd_u8(a,b)
#define vpadd_s8(a,b) __simd_vpadd_s8(a,b)
#define vpadd_u16(a,b) __simd_vpadd_u16(a,b)
#define vpadd_s16(a,b) __simd_vpadd_s16(a,b)
#define vpadd_u32(a,b) __simd_vpadd_u32(a,b)
#define vpadd_s32(a,b) __simd_vpadd_s32(a,b)
#define vpaddq_u8(a,b) __simd_vpaddq_u8(a,b)
#define vpaddq_s8(a,b) __simd_vpaddq_s8(a,b)
#define vpaddq_u16(a,b) __simd_vpaddq_u16(a,b)
#define vpaddq_s16(a,b) __simd_vpaddq_s16(a,b)
#define vpaddq_u32(a,b) __simd_vpaddq_u32(a,b)
#define vpaddq_s32(a,b) __simd_vpaddq_s32(a,b)
#define vpaddq_u64(a,b) __simd_vpaddq_u64(a,b)
#define vpaddq_s64(a,b) __simd_vpaddq_s64(a,b)
#define vhadd_s8(a,b) __simd_vhadd_s8(a,b)
#define vhadd_s16(a,b) __simd_vhadd_s16(a,b)
#define vhadd_s32(a,b) __simd_vhadd_s32(a,b)
#define vhaddq_s8(a,b) __simd_vhaddq_s8(a,b)
#define vhaddq_s16(a,b) __simd_vhaddq_s16(a,b)
#define vhaddq_s32(a,b) __simd_vhaddq_s32(a,b)
#define vhadd_u8(a,b) __simd_vhadd_u8(a,b)
#define vhadd_u16(a,b) __simd_vhadd_u16(a,b)
#define vhadd_u32(a,b) __simd_vhadd_u32(a,b)
#define vhaddq_u8(a,b) __simd_vhaddq_u8(a,b)
#define vhaddq_u16(a,b) __simd_vhaddq_u16(a,b)
#define vhaddq_u32(a,b) __simd_vhaddq_u32(a,b)
#define vqadd_s8(a,b) __simd_vqadd_s8(a,b)
#define vqadd_s16(a,b) __simd_vqadd_s16(a,b)
#define vqadd_s32(a,b) __simd_vqadd_s32(a,b)
#define vqaddq_s8(a,b) __simd_vqaddq_s8(a,b)
#define vqaddq_s16(a,b) __simd_vqaddq_s16(a,b)
#define vqaddq_s32(a,b) __simd_vqaddq_s32(a,b)
#define vqaddq_s64(a,b) __simd_vqaddq_s64(a,b)
#define vqadd_u8(a,b) __simd_vqadd_u8(a,b)
#define vqadd_u16(a,b) __simd_vqadd_u16(a,b)
#define vqadd_u32(a,b) __simd_vqadd_u32(a,b)
#define vqaddq_u8(a,b) __simd_vqaddq_u8(a,b)
#define vqaddq_u16(a,b) __simd_vqaddq_u16(a,b)
#define vqaddq_u32(a,b) __simd_vqaddq_u32(a,b)
#define vqaddq_u64(a,b) __simd_vqaddq_u64(a,b)
#define vrhadd_s8(a,b) __simd_vrhadd_s8(a,b)
#define vrhadd_s16(a,b) __simd_vrhadd_s16(a,b)
#define vrhadd_s32(a,b) __simd_vrhadd_s32(a,b)
#define vrhaddq_s8(a,b) __simd_vrhaddq_s8(a,b)
#define vrhaddq_s16(a,b) __simd_vrhaddq_s16(a,b)
#define vrhaddq_s32(a,b) __simd_vrhaddq_s32(a,b)
#define vrhadd_u8(a,b) __simd_vrhadd_u8(a,b)
#define vrhadd_u16(a,b) __simd_vrhadd_u16(a,b)
#define vrhadd_u32(a,b) __simd_vrhadd_u32(a,b)
#define vrhaddq_u8(a,b) __simd_vrhaddq_u8(a,b)
#define vrhaddq_u16(a,b) __simd_vrhaddq_u16(a,b)
#define vrhaddq_u32(a,b) __simd_vrhaddq_u32(a,b)
#define vhsub_s8(a,b) __simd_vhsub_s8(a,b)
#define vhsub_s16(a,b) __simd_vhsub_s16(a,b)
#define vhsub_s32(a,b) __simd_vhsub_s32(a,b)
#define vhsubq_s8(a,b) __simd_vhsubq_s8(a,b)
#define vhsubq_s16(a,b) __simd_vhsubq_s16(a,b)
#define vhsubq_s32(a,b) __simd_vhsubq_s32(a,b)
#define vhsub_u8(a,b) __simd_vhsub_u8(a,b)
#define vhsub_u16(a,b) __simd_vhsub_u16(a,b)
#define vhsub_u32(a,b) __simd_vhsub_u32(a,b)
#define vhsubq_u8(a,b) __simd_vhsubq_u8(a,b)
#define vhsubq_u16(a,b) __simd_vhsubq_u16(a,b)
#define vhsubq_u32(a,b) __simd_vhsubq_u32(a,b)
#define vqsub_s8(a,b) __simd_vqsub_s8(a,b)
#define vqsub_s16(a,b) __simd_vqsub_s16(a,b)
#define vqsub_s32(a,b) __simd_vqsub_s32(a,b)
#define vqsubq_s8(a,b) __simd_vqsubq_s8(a,b)
#define vqsubq_s16(a,b) __simd_vqsubq_s16(a,b)
#define vqsubq_s32(a,b) __simd_vqsubq_s32(a,b)
#define vqsubq_s64(a,b) __simd_vqsubq_s64(a,b)
#define vqsub_u8(a,b) __simd_vqsub_u8(a,b)
#define vqsub_u16(a,b) __simd_vqsub_u16(a,b)
#define vqsub_u32(a,b) __simd_vqsub_u32(a,b)
#define vqsubq_u8(a,b) __simd_vqsubq_u8(a,b)
#define vqsubq_u16(a,b) __simd_vqsubq_u16(a,b)
#define vqsubq_u32(a,b) __simd_vqsubq_u32(a,b)
#define vqsubq_u64(a,b) __simd_vqsubq_u64(a,b)
#define vmul_u8(a,b) __simd_vmul_u8(a,b)
#define vmul_s8(a,b) __simd_vmul_s8(a,b)
#define vmul_u16(a,b) __simd_vmul_u16(a,b)
#define vmul_s16(a,b) __simd_vmul_s16(a,b)
#define vmul_u32(a,b) __simd_vmul_u32(a,b)
#define vmul_s32(a,b) __simd_vmul_s32(a,b)
#define vmulq_u8(a,b) __simd_vmulq_u8(a,b)
#define vmulq_s8(a,b) __simd_vmulq_s8(a,b)
#define vmulq_u16(a,b) __simd_vmulq_u16(a,b)
#define vmulq_s16(a,b) __simd_vmulq_s16(a,b)
#define vmulq_u32(a,b) __simd_vmulq_u32(a,b)
#define vmulq_s32(a,b) __simd_vmulq_s32(a,b)
#define vmla_u8(a,b,c) __simd_vmla_u8(a,b,c)
#define vmla_s8(a,b,c) __simd_vmla_s8(a,b,c)
#define vmla_u16(a,b,c) __simd_vmla_u16(a,b,c)
#define vmla_s16(a,b,c) __simd_vmla_s16(a,b,c)
#define vmla_u32(a,b,c) __simd_vmla_u32(a,b,c)
#define vmla_s32(a,b,c) __simd_vmla_s32(a,b,c)
#define vmlaq_u8(a,b,c) __simd_vmlaq_u8(a,b,c)
#define vmlaq_s8(a,b,c) __simd_vmlaq_s8(a,b,c)
#define vmlaq_u16(a,b,c) __simd_vmlaq_u16(a,b,c)
#define vmlaq_s16(a,b,c) __simd_vmlaq_s16(a,b,c)
#define vmlaq_u32(a,b,c) __simd_vmlaq_u32(a,b,c)
#define vmlaq_s32(a,b,c) __simd_vmlaq_s32(a,b,c)
#define vmls_u8(a,b,c) __simd_vmls_u8(a,b,c)
#define vmls_s8(a,b,c) __simd_vmls_s8(a,b,c)
#define vmls_u16(a,b,c) __simd_vmls_u16(a,b,c)
#define vmls_s16(a,b,c) __simd_vmls_s16(a,b,c)
#define vmls_u32(a,b,c) __simd_vmls_u32(a,b,c)
#define vmls_s32(a,b,c) __simd_vmls_s32(a,b,c)
#define vmlsq_u8(a,b,c) __simd_vmlsq_u8(a,b,c)
#define vmlsq_s8(a,b,c) __simd_vmlsq_s8(a,b,c)
#define vmlsq_u16(a,b,c) __simd_vmlsq_u16(a,b,c)
#define vmlsq_s16(a,b,c) __simd_vmlsq_s16(a,b,c)
#define vmlsq_u32(a,b,c) __simd_vmlsq_u32(a,b,c)
#define vmlsq_s32(a,b,c) __simd_vmlsq_s32(a,b,c)
#define vqdmulh_s16(a,b) __simd_vqdmulh_s16(a,b)
#define vqdmulh_s32(a,b) __simd_vqdmulh_s32(a,b)
#define vqdmulhq_s16(a,b) __simd_vqdmulhq_s16(a,b)
#define vqdmulhq_s32(a,b) __simd_vqdmulhq_s32(a,b)
#define vqrdmulh_s16(a,b) __simd_vqrdmulh_s16(a,b)
#define vqrdmulh_s32(a,b) __simd_vqrdmulh_s32(a,b)
#define vqrdmulhq_s16(a,b) __simd_vqrdmulhq_s16(a,b)
#define vqrdmulhq_s32(a,b) __simd_vqrdmulhq_s32(a,b)
#define vmul_p8(a,b) __simd_vmul_p8(a,b)
#define vmulq_p8(a,b) __simd_vmulq_p8(a,b)
#define vshl_s8(a,b) __simd_vshl_s8(a,b)
#define vshl_s16(a,b) __simd_vshl_s16(a,b)
#define vshl_s32(a,b) __simd_vshl_s32(a,b)
#define vshlq_s8(a,b) __simd_vshlq_s8(a,b)
#define vshlq_s16(a,b) __simd_vshlq_s16(a,b)
#define vshlq_s32(a,b) __simd_vshlq_s32(a,b)
#define vshlq_s64(a,b) __simd_vshlq_s64(a,b)
#define vshl_u8(a,b) __simd_vshl_u8(a,b)
#define vshl_u16(a,b) __simd_vshl_u16(a,b)
#define vshl_u32(a,b) __simd_vshl_u32(a,b)
#define vshlq_u8(a,b) __simd_vshlq_u8(a,b)
#define vshlq_u16(a,b) __simd_vshlq_u16(a,b)
#define vshlq_u32(a,b) __simd_vshlq_u32(a,b)
#define vshlq_u64(a,b) __simd_vshlq_u64(a,b)
#define vrshl_s8(a,b) __simd_vrshl_s8(a,b)
#define vrshl_s16(a,b) __simd_vrshl_s16(a,b)
#define vrshl_s32(a,b) __simd_vrshl_s32(a,b)
#define vrshlq_s8(a,b) __simd_vrshlq_s8(a,b)
#define vrshlq_s16(a,b) __simd_vrshlq_s16(a,b)
#define vrshlq_s32(a,b) __simd_vrshlq_s32(a,b)
#define vrshlq_s64(a,b) __simd_vrshlq_s64(a,b)
#define vrshl_u8(a,b) __simd_vrshl_u8(a,b)
#define vrshl_u16(a,b) __simd_vrshl_u16(a,b)
#define vrshl_u32(a,b) __simd_vrshl_u32(a,b)
#define vrshlq_u8(a,b) __simd_vrshlq_u8(a,b)
#define vrshlq_u16(a,b) __simd_vrshlq_u16(a,b)
#define vrshlq_u32(a,b) __simd_vrshlq_u32(a,b)
#define vrshlq_u64(a,b) __simd_vrshlq_u64(a,b)
#define vqshl_s8(a,b) __simd_vqshl_s8(a,b)
#define vqshl_s16(a,b) __simd_vqshl_s16(a,b)
#define vqshl_s32(a,b) __simd_vqshl_s32(a,b)
#define vqshlq_s8(a,b) __simd_vqshlq_s8(a,b)
#define vqshlq_s16(a,b) __simd_vqshlq_s16(a,b)
#define vqshlq_s32(a,b) __simd_vqshlq_s32(a,b)
#define vqshlq_s64(a,b) __simd_vqshlq_s64(a,b)
#define vqshl_u8(a,b) __simd_vqshl_u8(a,b)
#define vqshl_u16(a,b) __simd_vqshl_u16(a,b)
#define vqshl_u32(a,b) __simd_vqshl_u32(a,b)
#define vqshlq_u8(a,b) __simd_vqshlq_u8(a,b)
#define vqshlq_u16(a,b) __simd_vqshlq_u16(a,b)
#define vqshlq_u32(a,b) __simd_vqshlq_u32(a,b)
#define vqshlq_u64(a,b) __simd_vqshlq_u64(a,b)
#define vqrshl_s8(a,b) __simd_vqrshl_s8(a,b)
#define vqrshl_s16(a,b) __simd_vqrshl_s16(a,b)
#define vqrshl_s32(a,b) __simd_vqrshl_s32(a,b)
#define vqrshlq_s8(a,b) __simd_vqrshlq_s8(a,b)
#define vqrshlq_s16(a,b) __simd_vqrshlq_s16(a,b)
#define vqrshlq_s32(a,b) __simd_vqrshlq_s32(a,b)
#define vqrshlq_s64(a,b) __simd_vqrshlq_s64(a,b)
#define vqrshl_u8(a,b) __simd_vqrshl_u8(a,b)
#define vqrshl_u16(a,b) __simd_vqrshl_u16(a,b)
#define vqrshl_u32(a,b) __simd_vqrshl_u32(a,b)
#define vqrshlq_u8(a,b) __simd_vqrshlq_u8(a,b)
#define vqrshlq_u16(a,b) __simd_vqrshlq_u16(a,b)
#define vqrshlq_u32(a,b) __simd_vqrshlq_u32(a,b)
#define vqrshlq_u64(a,b) __simd_vqrshlq_u64(a,b)
#define vceq_u8(a,b) __simd_vceq_u8(a,b)
#define vceq_s8(a,b) __simd_vceq_s8(a,b)
#define vceq_p8(a,b) __simd_vceq_p8(a,b)
#define vceq_u16(a,b) __simd_vceq_u16(a,b)
#define vceq_s16(a,b) __simd_vceq_s16(a,b)
#define vceq_p16(a,b) __simd_vceq_p16(a,b)
#define vceq_u32(a,b) __simd_vceq_u32(a,b)
#define vceq_s32(a,b) __simd_vceq_s32(a,b)
#define vceqq_u8(a,b) __simd_vceqq_u8(a,b)
#define vceqq_s8(a,b) __simd_vceqq_s8(a,b)
#define vceqq_p8(a,b) __simd_vceqq_p8(a,b)
#define vceqq_u16(a,b) __simd_vceqq_u16(a,b)
#define vceqq_s16(a,b) __simd_vceqq_s16(a,b)
#define vceqq_p16(a,b) __simd_vceqq_p16(a,b)
#define vceqq_u32(a,b) __simd_vceqq_u32(a,b)
#define vceqq_s32(a,b) __simd_vceqq_s32(a,b)
#define vceqq_u64(a,b) __simd_vceqq_u64(a,b)
#define vceqq_s64(a,b) __simd_vceqq_s64(a,b)
#define vcge_s8(a,b) __simd_vcge_s8(a,b)
#define vcge_s16(a,b) __simd_vcge_s16(a,b)
#define vcge_s32(a,b) __simd_vcge_s32(a,b)
#define vcgeq_s8(a,b) __simd_vcgeq_s8(a,b)
#define vcgeq_s16(a,b) __simd_vcgeq_s16(a,b)
#define vcgeq_s32(a,b) __simd_vcgeq_s32(a,b)
#define vcgeq_s64(a,b) __simd_vcgeq_s64(a,b)
#define vcgt_s8(a,b) __simd_vcgt_s8(a,b)
#define vcgt_s16(a,b) __simd_vcgt_s16(a,b)
#define vcgt_s32(a,b) __simd_vcgt_s32(a,b)
#define vcgtq_s8(a,b) __simd_vcgtq_s8(a,b)
#define vcgtq_s16(a,b) __simd_vcgtq_s16(a,b)
#define vcgtq_s32(a,b) __simd_vcgtq_s32(a,b)
#define vcgtq_s64(a,b) __simd_vcgtq_s64(a,b)
#define vcge_u8(a,b) __simd_vcge_u8(a,b)
#define vcge_u16(a,b) __simd_vcge_u16(a,b)
#define vcge_u32(a,b) __simd_vcge_u32(a,b)
#define vcgeq_u8(a,b) __simd_vcgeq_u8(a,b)
#define vcgeq_u16(a,b) __simd_vcgeq_u16(a,b)
#define vcgeq_u32(a,b) __simd_vcgeq_u32(a,b)
#define vcgeq_u64(a,b) __simd_vcgeq_u64(a,b)
#define vcgt_u8(a,b) __simd_vcgt_u8(a,b)
#define vcgt_u16(a,b) __simd_vcgt_u16(a,b)
#define vcgt_u32(a,b) __simd_vcgt_u32(a,b)
#define vcgtq_u8(a,b) __simd_vcgtq_u8(a,b)
#define vcgtq_u16(a,b) __simd_vcgtq_u16(a,b)
#define vcgtq_u32(a,b) __simd_vcgtq_u32(a,b)
#define vcgtq_u64(a,b) __simd_vcgtq_u64(a,b)
#define vtst_u8(a,b) __simd_vtst_u8(a,b)
#define vtst_s8(a,b) __simd_vtst_s8(a,b)
#define vtst_p8(a,b) __simd_vtst_p8(a,b)
#define vtst_u16(a,b) __simd_vtst_u16(a,b)
#define vtst_s16(a,b) __simd_vtst_s16(a,b)
#define vtst_p16(a,b) __simd_vtst_p16(a,b)
#define vtst_u32(a,b) __simd_vtst_u32(a,b)
#define vtst_s32(a,b) __simd_vtst_s32(a,b)
#define vtstq_u8(a,b) __simd_vtstq_u8(a,b)
#define vtstq_s8(a,b) __simd_vtstq_s8(a,b)
#define vtstq_p8(a,b) __simd_vtstq_p8(a,b)
#define vtstq_u16(a,b) __simd_vtstq_u16(a,b)
#define vtstq_s16(a,b) __simd_vtstq_s16(a,b)
#define vtstq_p16(a,b) __simd_vtstq_p16(a,b)
#define vtstq_u32(a,b) __simd_vtstq_u32(a,b)
#define vtstq_s32(a,b) __simd_vtstq_s32(a,b)
#define vtstq_u64(a,b) __simd_vtstq_u64(a,b)
#define vtstq_s64(a,b) __simd_vtstq_s64(a,b)
#define vclt_s8(a,b) __simd_vclt_s8(a,b)
#define vclt_s16(a,b) __simd_vclt_s16(a,b)
#define vclt_s32(a,b) __simd_vclt_s32(a,b)
#define vcltq_s8(a,b) __simd_vcltq_s8(a,b)
#define vcltq_s16(a,b) __simd_vcltq_s16(a,b)
#define vcltq_s32(a,b) __simd_vcltq_s32(a,b)
#define vcltq_s64(a,b) __simd_vcltq_s64(a,b)
#define vcle_s8(a,b) __simd_vcle_s8(a,b)
#define vcle_s16(a,b) __simd_vcle_s16(a,b)
#define vcle_s32(a,b) __simd_vcle_s32(a,b)
#define vcleq_s8(a,b) __simd_vcleq_s8(a,b)
#define vcleq_s16(a,b) __simd_vcleq_s16(a,b)
#define vcleq_s32(a,b) __simd_vcleq_s32(a,b)
#define vcleq_s64(a,b) __simd_vcleq_s64(a,b)
#define vclt_u8(a,b) __simd_vclt_u8(a,b)
#define vclt_u16(a,b) __simd_vclt_u16(a,b)
#define vclt_u32(a,b) __simd_vclt_u32(a,b)
#define vcltq_u8(a,b) __simd_vcltq_u8(a,b)
#define vcltq_u16(a,b) __simd_vcltq_u16(a,b)
#define vcltq_u32(a,b) __simd_vcltq_u32(a,b)
#define vcltq_u64(a,b) __simd_vcltq_u64(a,b)
#define vcle_u8(a,b) __simd_vcle_u8(a,b)
#define vcle_u16(a,b) __simd_vcle_u16(a,b)
#define vcle_u32(a,b) __simd_vcle_u32(a,b)
#define vcleq_u8(a,b) __simd_vcleq_u8(a,b)
#define vcleq_u16(a,b) __simd_vcleq_u16(a,b)
#define vcleq_u32(a,b) __simd_vcleq_u32(a,b)
#define vcleq_u64(a,b) __simd_vcleq_u64(a,b)
#define vmax_s8(a,b) __simd_vmax_s8(a,b)
#define vmax_s16(a,b) __simd_vmax_s16(a,b)
#define vmax_s32(a,b) __simd_vmax_s32(a,b)
#define vmaxq_s8(a,b) __simd_vmaxq_s8(a,b)
#define vmaxq_s16(a,b) __simd_vmaxq_s16(a,b)
#define vmaxq_s32(a,b) __simd_vmaxq_s32(a,b)
#define vmax_u8(a,b) __simd_vmax_u8(a,b)
#define vmax_u16(a,b) __simd_vmax_u16(a,b)
#define vmax_u32(a,b) __simd_vmax_u32(a,b)
#define vmaxq_u8(a,b) __simd_vmaxq_u8(a,b)
#define vmaxq_u16(a,b) __simd_vmaxq_u16(a,b)
#define vmaxq_u32(a,b) __simd_vmaxq_u32(a,b)
#define vmin_s8(a,b) __simd_vmin_s8(a,b)
#define vmin_s16(a,b) __simd_vmin_s16(a,b)
#define vmin_s32(a,b) __simd_vmin_s32(a,b)
#define vminq_s8(a,b) __simd_vminq_s8(a,b)
#define vminq_s16(a,b) __simd_vminq_s16(a,b)
#define vminq_s32(a,b) __simd_vminq_s32(a,b)
#define vmin_u8(a,b) __simd_vmin_u8(a,b)
#define vmin_u16(a,b) __simd_vmin_u16(a,b)
#define vmin_u32(a,b) __simd_vmin_u32(a,b)
#define vminq_u8(a,b) __simd_vminq_u8(a,b)
#define vminq_u16(a,b) __simd_vminq_u16(a,b)
#define vminq_u32(a,b) __simd_vminq_u32(a,b)
#define vpmax_s8(a,b) __simd_vpmax_s8(a,b)
#define vpmax_s16(a,b) __simd_vpmax_s16(a,b)
#define vpmax_s32(a,b) __simd_vpmax_s32(a,b)
#define vpmaxq_s8(a,b) __simd_vpmaxq_s8(a,b)
#define vpmaxq_s16(a,b) __simd_vpmaxq_s16(a,b)
#define vpmaxq_s32(a,b) __simd_vpmaxq_s32(a,b)
#define vpmax_u8(a,b) __simd_vpmax_u8(a,b)
#define vpmax_u16(a,b) __simd_vpmax_u16(a,b)
#define vpmax_u32(a,b) __simd_vpmax_u32(a,b)
#define vpmaxq_u8(a,b) __simd_vpmaxq_u8(a,b)
#define vpmaxq_u16(a,b) __simd_vpmaxq_u16(a,b)
#define vpmaxq_u32(a,b) __simd_vpmaxq_u32(a,b)
#define vpmin_s8(a,b) __simd_vpmin_s8(a,b)
#define vpmin_s16(a,b) __simd_vpmin_s16(a,b)
#define vpmin_s32(a,b) __simd_vpmin_s32(a,b)
#define vpminq_s8(a,b) __simd_vpminq_s8(a,b)
#define vpminq_s16(a,b) __simd_vpminq_s16(a,b)
#define vpminq_s32(a,b) __simd_vpminq_s32(a,b)
#define vpmin_u8(a,b) __simd_vpmin_u8(a,b)
#define vpmin_u16(a,b) __simd_vpmin_u16(a,b)
#define vpmin_u32(a,b) __simd_vpmin_u32(a,b)
#define vpminq_u8(a,b) __simd_vpminq_u8(a,b)
#define vpminq_u16(a,b) __simd_vpminq_u16(a,b)
#define vpminq_u32(a,b) __simd_vpminq_u32(a,b)
#define vaba_s8(a,b,c) __simd_vaba_s8(a,b,c)
#define vaba_s16(a,b,c) __simd_vaba_s16(a,b,c)
#define vaba_s32(a,b,c) __simd_vaba_s32(a,b,c)
#define vabaq_s8(a,b,c) __simd_vabaq_s8(a,b,c)
#define vabaq_s16(a,b,c) __simd_vabaq_s16(a,b,c)
#define vabaq_s32(a,b,c) __simd_vabaq_s32(a,b,c)
#define vaba_u8(a,b,c) __simd_vaba_u8(a,b,c)
#define vaba_u16(a,b,c) __simd_vaba_u16(a,b,c)
#define vaba_u32(a,b,c) __simd_vaba_u32(a,b,c)
#define vabaq_u8(a,b,c) __simd_vabaq_u8(a,b,c)
#define vabaq_u16(a,b,c) __simd_vabaq_u16(a,b,c)
#define vabaq_u32(a,b,c) __simd_vabaq_u32(a,b,c)
#define vabd_s8(a,b) __simd_vabd_s8(a,b)
#define vabd_s16(a,b) __simd_vabd_s16(a,b)
#define vabd_s32(a,b) __simd_vabd_s32(a,b)
#define vabdq_s8(a,b) __simd_vabdq_s8(a,b)
#define vabdq_s16(a,b) __simd_vabdq_s16(a,b)
#define vabdq_s32(a,b) __simd_vabdq_s32(a,b)
#define vabd_u8(a,b) __simd_vabd_u8(a,b)
#define vabd_u16(a,b) __simd_vabd_u16(a,b)
#define vabd_u32(a,b) __simd_vabd_u32(a,b)
#define vabdq_u8(a,b) __simd_vabdq_u8(a,b)
#define vabdq_u16(a,b) __simd_vabdq_u16(a,b)
#define vabdq_u32(a,b) __simd_vabdq_u32(a,b)
#define vreinterpret_u8_u8(a) __simd_vreinterpret_u8_u8(a)
#define vreinterpret_u8_u16(a) __simd_vreinterpret_u8_u16(a)
#define vreinterpret_u8_u32(a) __simd_vreinterpret_u8_u32(a)
#define vreinterpret_u8_u64(a) __simd_vreinterpret_u8_u64(a)
#define vreinterpret_u8_s8(a) __simd_vreinterpret_u8_s8(a)
#define vreinterpret_u8_s16(a) __simd_vreinterpret_u8_s16(a)
#define vreinterpret_u8_s32(a) __simd_vreinterpret_u8_s32(a)
#define vreinterpret_u8_s64(a) __simd_vreinterpret_u8_s64(a)
#define vreinterpret_u8_p8(a) __simd_vreinterpret_u8_p8(a)
#define vreinterpret_u8_p16(a) __simd_vreinterpret_u8_p16(a)
#define vreinterpret_u8_f32(a) __simd_vreinterpret_u8_f32(a)
#define vreinterpret_u16_u8(a) __simd_vreinterpret_u16_u8(a)
#define vreinterpret_u16_u16(a) __simd_vreinterpret_u16_u16(a)
#define vreinterpret_u16_u32(a) __simd_vreinterpret_u16_u32(a)
#define vreinterpret_u16_u64(a) __simd_vreinterpret_u16_u64(a)
#define vreinterpret_u16_s8(a) __simd_vreinterpret_u16_s8(a)
#define vreinterpret_u16_s16(a) __simd_vreinterpret_u16_s16(a)
#define vreinterpret_u16_s32(a) __simd_vreinterpret_u16_s32(a)
#define vreinterpret_u16_s64(a) __simd_vreinterpret_u16_s64(a)
#define vreinterpret_u16_p8(a) __simd_vreinterpret_u16_p8(a)
#define vreinterpret_u16_p16(a) __simd_vreinterpret_u16_p16(a)
#define vreinterpret_u16_f32(a) __simd_vreinterpret_u16_f32(a)
#define vreinterpret_u32_u8(a) __simd_vreinterpret_u32_u8(a)
#define vreinterpret_u32_u16(a) __simd_vreinterpret_u32_u16(a)
#define vreinterpret_u32_u32(a) __simd_vreinterpret_u32_u32(a)
#define vreinterpret_u32_u64(a) __simd_vreinterpret_u32_u64(a)
#define vreinterpret_u32_s8(a) __simd_vreinterpret_u32_s8(a)
#define vreinterpret_u32_s16(a) __simd_vreinterpret_u32_s16(a)
#define vreinterpret_u32_s32(a) __simd_vreinterpret_u32_s32(a)
#define vreinterpret_u32_s64(a) __simd_vreinterpret_u32_s64(a)
#define vreinterpret_u32_p8(a) __simd_vreinterpret_u32_p8(a)
#define vreinterpret_u32_p16(a) __simd_vreinterpret_u32_p16(a)
#define vreinterpret_u32_f32(a) __simd_vreinterpret_u32_f32(a)
#define vreinterpret_u64_u8(a) __simd_vreinterpret_u64_u8(a)
#define vreinterpret_u64_u16(a) __simd_vreinterpret_u64_u16(a)
#define vreinterpret_u64_u32(a) __simd_vreinterpret_u64_u32(a)
#define vreinterpret_u64_u64(a) __simd_vreinterpret_u64_u64(a)
#define vreinterpret_u64_s8(a) __simd_vreinterpret_u64_s8(a)
#define vreinterpret_u64_s16(a) __simd_vreinterpret_u64_s16(a)
#define vreinterpret_u64_s32(a) __simd_vreinterpret_u64_s32(a)
#define vreinterpret_u64_s64(a) __simd_vreinterpret_u64_s64(a)
#define vreinterpret_u64_p8(a) __simd_vreinterpret_u64_p8(a)
#define vreinterpret_u64_p16(a) __simd_vreinterpret_u64_p16(a)
#define vreinterpret_u64_f32(a) __simd_vreinterpret_u64_f32(a)
#define vreinterpret_s8_u8(a) __simd_vreinterpret_s8_u8(a)
#define vreinterpret_s8_u16(a) __simd_vreinterpret_s8_u16(a)
#define vreinterpret_s8_u32(a) __simd_vreinterpret_s8_u32(a)
#define vreinterpret_s8_u64(a) __simd_vreinterpret_s8_u64(a)
#define vreinterpret_s8_s8(a) __simd_vreinterpret_s8_s8(a)
#define vreinterpret_s8_s16(a) __simd_vreinterpret_s8_s16(a)
#define vreinterpret_s8_s32(a) __simd_vreinterpret_s8_s32(a)
#define vreinterpret_s8_s64(a) __simd_vreinterpret_s8_s64(a)
#define vreinterpret_s8_p8(a) __simd_vreinterpret_s8_p8(a)
#define vreinterpret_s8_p16(a) __simd_vreinterpret_s8_p16(a)
#define vreinterpret_s8_f32(a) __simd_vreinterpret_s8_f32(a)
#define vreinterpret_s16_u8(a) __simd_vreinterpret_s16_u8(a)
#define vreinterpret_s16_u16(a) __simd_vreinterpret_s16_u16(a)
#define vreinterpret_s16_u32(a) __simd_vreinterpret_s16_u32(a)
#define vreinterpret_s16_u64(a) __simd_vreinterpret_s16_u64(a)
#define vreinterpret_s16_s8(a) __simd_vreinterpret_s16_s8(a)
#define vreinterpret_s16_s16(a) __simd_vreinterpret_s16_s16(a)
#define vreinterpret_s16_s32(a) __simd_vreinterpret_s16_s32(a)
#define vreinterpret_s16_s64(a) __simd_vreinterpret_s16_s64(a)
#define vreinterpret_s16_p8(a) __simd_vreinterpret_s16_p8(a)
#define vreinterpret_s16_p16(a) __simd_vreinterpret_s16_p16(a)
#define vreinterpret_s16_f32(a) __simd_vreinterpret_s16_f32(a)
#define vreinterpret_s32_u8(a) __simd_vreinterpret_s32_u8(a)
#define vreinterpret_s32_u16(a) __simd_vreinterpret_s32_u16(a)
#define vreinterpret_s32_u32(a) __simd_vreinterpret_s32_u32(a)
#define vreinterpret_s32_u64(a) __simd_vreinterpret_s32_u64(a)
#define vreinterpret_s32_s8(a) __simd_vreinterpret_s32_s8(a)
#define vreinterpret_s32_s16(a) __simd_vreinterpret_s32_s16(a)
#define vreinterpret_s32_s32(a) __simd_vreinterpret_s32_s32(a)
#define vreinterpret_s32_s64(a) __simd_vreinterpret_s32_s64(a)
#define vreinterpret_s32_p8(a) __simd_vreinterpret_s32_p8(a)
#define vreinterpret_s32_p16(a) __simd_vreinterpret_s32_p16(a)
#define vreinterpret_s32_f32(a) __simd_vreinterpret_s32_f32(a)
#define vreinterpret_s64_u8(a) __simd_vreinterpret_s64_u8(a)
#define vreinterpret_s64_u16(a) __simd_vreinterpret_s64_u16(a)
#define vreinterpret_s64_u32(a) __simd_vreinterpret_s64_u32(a)
#define vreinterpret_s64_u64(a) __simd_vreinterpret_s64_u64(a)
#define vreinterpret_s64_s8(a) __simd_vreinterpret_s64_s8(a)
#define vreinterpret_s64_s16(a) __simd_vreinterpret_s64_s16(a)
#define vreinterpret_s64_s32(a) __simd_vreinterpret_s64_s32(a)
#define vreinterpret_s64_s64(a) __simd_vreinterpret_s64_s64(a)
#define vreinterpret_s64_p8(a) __simd_vreinterpret_s64_p8(a)
#define vreinterpret_s64_p16(a) __simd_vreinterpret_s64_p16(a)
#define vreinterpret_s64_f32(a) __simd_vreinterpret_s64_f32(a)
#define vreinterpret_p8_u8(a) __simd_vreinterpret_p8_u8(a)
#define vreinterpret_p8_u16(a) __simd_vreinterpret_p8_u16(a)
#define vreinterpret_p8_u32(a) __simd_vreinterpret_p8_u32(a)
#define vreinterpret_p8_u64(a) __simd_vreinterpret_p8_u64(a)
#define vreinterpret_p8_s8(a) __simd_vreinterpret_p8_s8(a)
#define vreinterpret_p8_s16(a) __simd_vreinterpret_p8_s16(a)
#define vreinterpret_p8_s32(a) __simd_vreinterpret_p8_s32(a)
#define vreinterpret_p8_s64(a) __simd_vreinterpret_p8_s64(a)
#define vreinterpret_p8_p8(a) __simd_vreinterpret_p8_p8(a)
#define vreinterpret_p8_p16(a) __simd_vreinterpret_p8_p16(a)
#define vreinterpret_p8_f32(a) __simd_vreinterpret_p8_f32(a)
#define vreinterpret_p16_u8(a) __simd_vreinterpret_p16_u8(a)
#define vreinterpret_p16_u16(a) __simd_vreinterpret_p16_u16(a)
#define vreinterpret_p16_u32(a) __simd_vreinterpret_p16_u32(a)
#define vreinterpret_p16_u64(a) __simd_vreinterpret_p16_u64(a)
#define vreinterpret_p16_s8(a) __simd_vreinterpret_p16_s8(a)
#define vreinterpret_p16_s16(a) __simd_vreinterpret_p16_s16(a)
#define vreinterpret_p16_s32(a) __simd_vreinterpret_p16_s32(a)
#define vreinterpret_p16_s64(a) __simd_vreinterpret_p16_s64(a)
#define vreinterpret_p16_p8(a) __simd_vreinterpret_p16_p8(a)
#define vreinterpret_p16_p16(a) __simd_vreinterpret_p16_p16(a)
#define vreinterpret_p16_f32(a) __simd_vreinterpret_p16_f32(a)
#define vreinterpret_f32_u8(a) __simd_vreinterpret_f32_u8(a)
#define vreinterpret_f32_u16(a) __simd_vreinterpret_f32_u16(a)
#define vreinterpret_f32_u32(a) __simd_vreinterpret_f32_u32(a)
#define vreinterpret_f32_u64(a) __simd_vreinterpret_f32_u64(a)
#define vreinterpret_f32_s8(a) __simd_vreinterpret_f32_s8(a)
#define vreinterpret_f32_s16(a) __simd_vreinterpret_f32_s16(a)
#define vreinterpret_f32_s32(a) __simd_vreinterpret_f32_s32(a)
#define vreinterpret_f32_s64(a) __simd_vreinterpret_f32_s64(a)
#define vreinterpret_f32_p8(a) __simd_vreinterpret_f32_p8(a)
#define vreinterpret_f32_p16(a) __simd_vreinterpret_f32_p16(a)
#define vreinterpret_f32_f32(a) __simd_vreinterpret_f32_f32(a)
#define vreinterpretq_u8_u8(a) __simd_vreinterpretq_u8_u8(a)
#define vreinterpretq_u8_u16(a) __simd_vreinterpretq_u8_u16(a)
#define vreinterpretq_u8_u32(a) __simd_vreinterpretq_u8_u32(a)
#define vreinterpretq_u8_u64(a) __simd_vreinterpretq_u8_u64(a)
#define vreinterpretq_u8_s8(a) __simd_vreinterpretq_u8_s8(a)
#define vreinterpretq_u8_s16(a) __simd_vreinterpretq_u8_s16(a)
#define vreinterpretq_u8_s32(a) __simd_vreinterpretq_u8_s32(a)
#define vreinterpretq_u8_s64(a) __simd_vreinterpretq_u8_s64(a)
#define vreinterpretq_u8_p8(a) __simd_vreinterpretq_u8_p8(a)
#define vreinterpretq_u8_p16(a) __simd_vreinterpretq_u8_p16(a)
#define vreinterpretq_u8_f32(a) __simd_vreinterpretq_u8_f32(a)
#define vreinterpretq_u8_f64(a) __simd_vreinterpretq_u8_f64(a)
#define vreinterpretq_u16_u8(a) __simd_vreinterpretq_u16_u8(a)
#define vreinterpretq_u16_u16(a) __simd_vreinterpretq_u16_u16(a)
#define vreinterpretq_u16_u32(a) __simd_vreinterpretq_u16_u32(a)
#define vreinterpretq_u16_u64(a) __simd_vreinterpretq_u16_u64(a)
#define vreinterpretq_u16_s8(a) __simd_vreinterpretq_u16_s8(a)
#define vreinterpretq_u16_s16(a) __simd_vreinterpretq_u16_s16(a)
#define vreinterpretq_u16_s32(a) __simd_vreinterpretq_u16_s32(a)
#define vreinterpretq_u16_s64(a) __simd_vreinterpretq_u16_s64(a)
#define vreinterpretq_u16_p8(a) __simd_vreinterpretq_u16_p8(a)
#define vreinterpretq_u16_p16(a) __simd_vreinterpretq_u16_p16(a)
#define vreinterpretq_u16_f32(a) __simd_vreinterpretq_u16_f32(a)
#define vreinterpretq_u16_f64(a) __simd_vreinterpretq_u16_f64(a)
#define vreinterpretq_u32_u8(a) __simd_vreinterpretq_u32_u8(a)
#define vreinterpretq_u32_u16(a) __simd_vreinterpretq_u32_u16(a)
#define vreinterpretq_u32_u32(a) __simd_vreinterpretq_u32_u32(a)
#define vreinterpretq_u32_u64(a) __simd_vreinterpretq_u32_u64(a)
#define vreinterpretq_u32_s8(a) __simd_vreinterpretq_u32_s8(a)
#define vreinterpretq_u32_s16(a) __simd_vreinterpretq_u32_s16(a)
#define vreinterpretq_u32_s32(a) __simd_vreinterpretq_u32_s32(a)
#define vreinterpretq_u32_s64(a) __simd_vreinterpretq_u32_s64(a)
#define vreinterpretq_u32_p8(a) __simd_vreinterpretq_u32_p8(a)
#define vreinterpretq_u32_p16(a) __simd_vreinterpretq_u32_p16(a)
#define vreinterpretq_u32_f32(a) __simd_vreinterpretq_u32_f32(a)
#define vreinterpretq_u32_f64(a) __simd_vreinterpretq_u32_f64(a)
#define vreinterpretq_u64_u8(a) __simd_vreinterpretq_u64_u8(a)
#define vreinterpretq_u64_u16(a) __simd_vreinterpretq_u64_u16(a)
#define vreinterpretq_u64_u32(a) __simd_vreinterpretq_u64_u32(a)
#define vreinterpretq_u64_u64(a) __simd_vreinterpretq_u64_u64(a)
#define vreinterpretq_u64_s8(a) __simd_vreinterpretq_u64_s8(a)
#define vreinterpretq_u64_s16(a) __simd_vreinterpretq_u64_s16(a)
#define vreinterpretq_u64_s32(a) __simd_vreinterpretq_u64_s32(a)
#define vreinterpretq_u64_s64(a) __simd_vreinterpretq_u64_s64(a)
#define vreinterpretq_u64_p8(a) __simd_vreinterpretq_u64_p8(a)
#define vreinterpretq_u64_p16(a) __simd_vreinterpretq_u64_p16(a)
#define vreinterpretq_u64_f32(a) __simd_vreinterpretq_u64_f32(a)
#define vreinterpretq_u64_f64(a) __simd_vreinterpretq_u64_f64(a)
#define vreinterpretq_s8_u8(a) __simd_vreinterpretq_s8_u8(a)
#define vreinterpretq_s8_u16(a) __simd_vreinterpretq_s8_u16(a)
#define vreinterpretq_s8_u32(a) __simd_vreinterpretq_s8_u32(a)
#define vreinterpretq_s8_u64(a) __simd_vreinterpretq_s8_u64(a)
#define vreinterpretq_s8_s8(a) __simd_vreinterpretq_s8_s8(a)
#define vreinterpretq_s8_s16(a) __simd_vreinterpretq_s8_s16(a)
#define vreinterpretq_s8_s32(a) __simd_vreinterpretq_s8_s32(a)
#define vreinterpretq_s8_s64(a) __simd_vreinterpretq_s8_s64(a)
#define vreinterpretq_s8_p8(a) __simd_vreinterpretq_s8_p8(a)
#define vreinterpretq_s8_p16(a) __simd_vreinterpretq_s8_p16(a)
#define vreinterpretq_s8_f32(a) __simd_vreinterpretq_s8_f32(a)
#define vreinterpretq_s8_f64(a) __simd_vreinterpretq_s8_f64(a)
#define vreinterpretq_s16_u8(a) __simd_vreinterpretq_s16_u8(a)
#define vreinterpretq_s16_u16(a) __simd_vreinterpretq_s16_u16(a)
#define vreinterpretq_s16_u32(a) __simd_vreinterpretq_s16_u32(a)
#define vreinterpretq_s16_u64(a) __simd_vreinterpretq_s16_u64(a)
#define vreinterpretq_s16_s8(a) __simd_vreinterpretq_s16_s8(a)
#define vreinterpretq_s16_s16(a) __simd_vreinterpretq_s16_s16(a)
#define vreinterpretq_s16_s32(a) __simd_vreinterpretq_s16_s32(a)
#define vreinterpretq_s16_s64(a) __simd_vreinterpretq_s16_s64(a)
#define vreinterpretq_s16_p8(a) __simd_vreinterpretq_s16_p8(a)
#define vreinterpretq_s16_p16(a) __simd_vreinterpretq_s16_p16(a)
#define vreinterpretq_s16_f32(a) __simd_vreinterpretq_s16_f32(a)
#define vreinterpretq_s16_f64(a) __simd_vreinterpretq_s16_f64(a)
#define vreinterpretq_s32_u8(a) __simd_vreinterpretq_s32_u8(a)
#define vreinterpretq_s32_u16(a) __simd_vreinterpretq_s32_u16(a)
#define vreinterpretq_s32_u32(a) __simd_vreinterpretq_s32_u32(a)
#define vreinterpretq_s32_u64(a) __simd_vreinterpretq_s32_u64(a)
#define vreinterpretq_s32_s8(a) __simd_vreinterpretq_s32_s8(a)
#define vreinterpretq_s32_s16(a) __simd_vreinterpretq_s32_s16(a)
#define vreinterpretq_s32_s32(a) __simd_vreinterpretq_s32_s32(a)
#define vreinterpretq_s32_s64(a) __simd_vreinterpretq_s32_s64(a)
#define vreinterpretq_s32_p8(a) __simd_vreinterpretq_s32_p8(a)
#define vreinterpretq_s32_p16(a) __simd_vreinterpretq_s32_p16(a)
#define vreinterpretq_s32_f32(a) __simd_vreinterpretq_s32_f32(a)
#define vreinterpretq_s32_f64(a) __simd_vreinterpretq_s32_f64(a)
#define vreinterpretq_s64_u8(a) __simd_vreinterpretq_s64_u8(a)
#define vreinterpretq_s64_u16(a) __simd_vreinterpretq_s64_u16(a)
#define vreinterpretq_s64_u32(a) __simd_vreinterpretq_s64_u32(a)
#define vreinterpretq_s64_u64(a) __simd_vreinterpretq_s64_u64(a)
#define vreinterpretq_s64_s8(a) __simd_vreinterpretq_s64_s8(a)
#define vreinterpretq_s64_s16(a) __simd_vreinterpretq_s64_s16(a)
#define vreinterpretq_s64_s32(a) __simd_vreinterpretq_s64_s32(a)
#define vreinterpretq_s64_s64(a) __simd_vreinterpretq_s64_s64(a)
#define vreinterpretq_s64_p8(a) __simd_vreinterpretq_s64_p8(a)
#define vreinterpretq_s64_p16(a) __simd_vreinterpretq_s64_p16(a)
#define vreinterpretq_s64_f32(a) __simd_vreinterpretq_s64_f32(a)
#define vreinterpretq_s64_f64(a) __simd_vreinterpretq_s64_f64(a)
#define vreinterpretq_p8_u8(a) __simd_vreinterpretq_p8_u8(a)
#define vreinterpretq_p8_u16(a) __simd_vreinterpretq_p8_u16(a)
#define vreinterpretq_p8_u32(a) __simd_vreinterpretq_p8_u32(a)
#define vreinterpretq_p8_u64(a) __simd_vreinterpretq_p8_u64(a)
#define vreinterpretq_p8_s8(a) __simd_vreinterpretq_p8_s8(a)
#define vreinterpretq_p8_s16(a) __simd_vreinterpretq_p8_s16(a)
#define vreinterpretq_p8_s32(a) __simd_vreinterpretq_p8_s32(a)
#define vreinterpretq_p8_s64(a) __simd_vreinterpretq_p8_s64(a)
#define vreinterpretq_p8_p8(a) __simd_vreinterpretq_p8_p8(a)
#define vreinterpretq_p8_p16(a) __simd_vreinterpretq_p8_p16(a)
#define vreinterpretq_p8_f32(a) __simd_vreinterpretq_p8_f32(a)
#define vreinterpretq_p8_f64(a) __simd_vreinterpretq_p8_f64(a)
#define vreinterpretq_p16_u8(a) __simd_vreinterpretq_p16_u8(a)
#define vreinterpretq_p16_u16(a) __simd_vreinterpretq_p16_u16(a)
#define vreinterpretq_p16_u32(a) __simd_vreinterpretq_p16_u32(a)
#define vreinterpretq_p16_u64(a) __simd_vreinterpretq_p16_u64(a)
#define vreinterpretq_p16_s8(a) __simd_vreinterpretq_p16_s8(a)
#define vreinterpretq_p16_s16(a) __simd_vreinterpretq_p16_s16(a)
#define vreinterpretq_p16_s32(a) __simd_vreinterpretq_p16_s32(a)
#define vreinterpretq_p16_s64(a) __simd_vreinterpretq_p16_s64(a)
#define vreinterpretq_p16_p8(a) __simd_vreinterpretq_p16_p8(a)
#define vreinterpretq_p16_p16(a) __simd_vreinterpretq_p16_p16(a)
#define vreinterpretq_p16_f32(a) __simd_vreinterpretq_p16_f32(a)
#define vreinterpretq_p16_f64(a) __simd_vreinterpretq_p16_f64(a)
#define vreinterpretq_f32_u8(a) __simd_vreinterpretq_f32_u8(a)
#define vreinterpretq_f32_u16(a) __simd_vreinterpretq_f32_u16(a)
#define vreinterpretq_f32_u32(a) __simd_vreinterpretq_f32_u32(a)
#define vreinterpretq_f32_u64(a) __simd_vreinterpretq_f32_u64(a)
#define vreinterpretq_f32_s8(a) __simd_vreinterpretq_f32_s8(a)
#define vreinterpretq_f32_s16(a) __simd_vreinterpretq_f32_s16(a)
#define vreinterpretq_f32_s32(a) __simd_vreinterpretq_f32_s32(a)
#define vreinterpretq_f32_s64(a) __simd_vreinterpretq_f32_s64(a)
#define vreinterpretq_f32_p8(a) __simd_vreinterpretq_f32_p8(a)
#define vreinterpretq_f32_p16(a) __simd_vreinterpretq_f32_p16(a)
#define vreinterpretq_f32_f32(a) __simd_vreinterpretq_f32_f32(a)
#define vreinterpretq_f32_f64(a) __simd_vreinterpretq_f32_f64(a)
#define vreinterpretq_f64_u8(a) __simd_vreinterpretq_f64_u8(a)
#define vreinterpretq_f64_u16(a) __simd_vreinterpretq_f64_u16(a)
#define vreinterpretq_f64_u32(a) __simd_vreinterpretq_f64_u32(a)
#define vreinterpretq_f64_u64(a) __simd_vreinterpretq_f64_u64(a)
#define vreinterpretq_f64_s8(a) __simd_vreinterpretq_f64_s8(a)
#define vreinterpretq_f64_s16(a) __simd_vreinterpretq_f64_s16(a)
#define vreinterpretq_f64_s32(a) __simd_vreinterpretq_f64_s32(a)
#define vreinterpretq_f64_s64(a) __simd_vreinterpretq_f64_s64(a)
#define vreinterpretq_f64_p8(a) __simd_vreinterpretq_f64_p8(a)
#define vreinterpretq_f64_p16(a) __simd_vreinterpretq_f64_p16(a)
#define vreinterpretq_f64_f32(a) __simd_vreinterpretq_f64_f32(a)
#define vreinterpretq_f64_f64(a) __simd_vreinterpretq_f64_f64(a)
#define vand_u8(a,b) __simd_vand_u8(a,b)
#define vand_u16(a,b) __simd_vand_u16(a,b)
#define vand_u32(a,b) __simd_vand_u32(a,b)
#define vand_u64(a,b) __simd_vand_u64(a,b)
#define vand_s8(a,b) __simd_vand_s8(a,b)
#define vand_s16(a,b) __simd_vand_s16(a,b)
#define vand_s32(a,b) __simd_vand_s32(a,b)
#define vand_s64(a,b) __simd_vand_s64(a,b)
#define vandq_u8(a,b) __simd_vandq_u8(a,b)
#define vandq_u16(a,b) __simd_vandq_u16(a,b)
#define vandq_u32(a,b) __simd_vandq_u32(a,b)
#define vandq_u64(a,b) __simd_vandq_u64(a,b)
#define vandq_s8(a,b) __simd_vandq_s8(a,b)
#define vandq_s16(a,b) __simd_vandq_s16(a,b)
#define vandq_s32(a,b) __simd_vandq_s32(a,b)
#define vandq_s64(a,b) __simd_vandq_s64(a,b)
#define vbic_u8(a,b) __simd_vbic_u8(a,b)
#define vbic_u16(a,b) __simd_vbic_u16(a,b)
#define vbic_u32(a,b) __simd_vbic_u32(a,b)
#define vbic_u64(a,b) __simd_vbic_u64(a,b)
#define vbic_s8(a,b) __simd_vbic_s8(a,b)
#define vbic_s16(a,b) __simd_vbic_s16(a,b)
#define vbic_s32(a,b) __simd_vbic_s32(a,b)
#define vbic_s64(a,b) __simd_vbic_s64(a,b)
#define vbicq_u8(a,b) __simd_vbicq_u8(a,b)
#define vbicq_u16(a,b) __simd_vbicq_u16(a,b)
#define vbicq_u32(a,b) __simd_vbicq_u32(a,b)
#define vbicq_u64(a,b) __simd_vbicq_u64(a,b)
#define vbicq_s8(a,b) __simd_vbicq_s8(a,b)
#define vbicq_s16(a,b) __simd_vbicq_s16(a,b)
#define vbicq_s32(a,b) __simd_vbicq_s32(a,b)
#define vbicq_s64(a,b) __simd_vbicq_s64(a,b)
#define vorr_u8(a,b) __simd_vorr_u8(a,b)
#define vorr_u16(a,b) __simd_vorr_u16(a,b)
#define vorr_u32(a,b) __simd_vorr_u32(a,b)
#define vorr_u64(a,b) __simd_vorr_u64(a,b)
#define vorr_s8(a,b) __simd_vorr_s8(a,b)
#define vorr_s16(a,b) __simd_vorr_s16(a,b)
#define vorr_s32(a,b) __simd_vorr_s32(a,b)
#define vorr_s64(a,b) __simd_vorr_s64(a,b)
#define vorrq_u8(a,b) __simd_vorrq_u8(a,b)
#define vorrq_u16(a,b) __simd_vorrq_u16(a,b)
#define vorrq_u32(a,b) __simd_vorrq_u32(a,b)
#define vorrq_u64(a,b) __simd_vorrq_u64(a,b)
#define vorrq_s8(a,b) __simd_vorrq_s8(a,b)
#define vorrq_s16(a,b) __simd_vorrq_s16(a,b)
#define vorrq_s32(a,b) __simd_vorrq_s32(a,b)
#define vorrq_s64(a,b) __simd_vorrq_s64(a,b)
#define vorn_u8(a,b) __simd_vorn_u8(a,b)
#define vorn_u16(a,b) __simd_vorn_u16(a,b)
#define vorn_u32(a,b) __simd_vorn_u32(a,b)
#define vorn_u64(a,b) __simd_vorn_u64(a,b)
#define vorn_s8(a,b) __simd_vorn_s8(a,b)
#define vorn_s16(a,b) __simd_vorn_s16(a,b)
#define vorn_s32(a,b) __simd_vorn_s32(a,b)
#define vorn_s64(a,b) __simd_vorn_s64(a,b)
#define vornq_u8(a,b) __simd_vornq_u8(a,b)
#define vornq_u16(a,b) __simd_vornq_u16(a,b)
#define vornq_u32(a,b) __simd_vornq_u32(a,b)
#define vornq_u64(a,b) __simd_vornq_u64(a,b)
#define vornq_s8(a,b) __simd_vornq_s8(a,b)
#define vornq_s16(a,b) __simd_vornq_s16(a,b)
#define vornq_s32(a,b) __simd_vornq_s32(a,b)
#define vornq_s64(a,b) __simd_vornq_s64(a,b)
#define veor_u8(a,b) __simd_veor_u8(a,b)
#define veor_u16(a,b) __simd_veor_u16(a,b)
#define veor_u32(a,b) __simd_veor_u32(a,b)
#define veor_u64(a,b) __simd_veor_u64(a,b)
#define veor_s8(a,b) __simd_veor_s8(a,b)
#define veor_s16(a,b) __simd_veor_s16(a,b)
#define veor_s32(a,b) __simd_veor_s32(a,b)
#define veor_s64(a,b) __simd_veor_s64(a,b)
#define veorq_u8(a,b) __simd_veorq_u8(a,b)
#define veorq_u16(a,b) __simd_veorq_u16(a,b)
#define veorq_u32(a,b) __simd_veorq_u32(a,b)
#define veorq_u64(a,b) __simd_veorq_u64(a,b)
#define veorq_s8(a,b) __simd_veorq_s8(a,b)
#define veorq_s16(a,b) __simd_veorq_s16(a,b)
#define veorq_s32(a,b) __simd_veorq_s32(a,b)
#define veorq_s64(a,b) __simd_veorq_s64(a,b)
#define vbsl_u8(a,b,c) __simd_vbsl_u8(a,b,c)
#define vbsl_u16(a,b,c) __simd_vbsl_u16(a,b,c)
#define vbsl_u32(a,b,c) __simd_vbsl_u32(a,b,c)
#define vbsl_u64(a,b,c) __simd_vbsl_u64(a,b,c)
#define vbsl_s8(a,b,c) __simd_vbsl_s8(a,b,c)
#define vbsl_s16(a,b,c) __simd_vbsl_s16(a,b,c)
#define vbsl_s32(a,b,c) __simd_vbsl_s32(a,b,c)
#define vbsl_s64(a,b,c) __simd_vbsl_s64(a,b,c)
#define vbsl_f32(a,b,c) __simd_vbsl_f32(a,b,c)
#define vbsl_p8(a,b,c) __simd_vbsl_p8(a,b,c)
#define vbsl_p16(a,b,c) __simd_vbsl_p16(a,b,c)
#define vbslq_u8(a,b,c) __simd_vbslq_u8(a,b,c)
#define vbslq_u16(a,b,c) __simd_vbslq_u16(a,b,c)
#define vbslq_u32(a,b,c) __simd_vbslq_u32(a,b,c)
#define vbslq_u64(a,b,c) __simd_vbslq_u64(a,b,c)
#define vbslq_s8(a,b,c) __simd_vbslq_s8(a,b,c)
#define vbslq_s16(a,b,c) __simd_vbslq_s16(a,b,c)
#define vbslq_s32(a,b,c) __simd_vbslq_s32(a,b,c)
#define vbslq_s64(a,b,c) __simd_vbslq_s64(a,b,c)
#define vbslq_f32(a,b,c) __simd_vbslq_f32(a,b,c)
#define vbslq_f64(a,b,c) __simd_vbslq_f64(a,b,c)
#define vbslq_p8(a,b,c) __simd_vbslq_p8(a,b,c)
#define vbslq_p16(a,b,c) __simd_vbslq_p16(a,b,c)
#define vqaddb_s8(a,b) __simd_vqaddb_s8(a,b)
#define vqaddh_s16(a,b) __simd_vqaddh_s16(a,b)
#define vqadds_s32(a,b) __simd_vqadds_s32(a,b)
#define vqadd_s64(a,b) __simd_vqadd_s64(a,b)
#define vqaddd_s64(a,b) __simd_vqaddd_s64(a,b)
#define vqaddb_u8(a,b) __simd_vqaddb_u8(a,b)
#define vqaddh_u16(a,b) __simd_vqaddh_u16(a,b)
#define vqadds_u32(a,b) __simd_vqadds_u32(a,b)
#define vqadd_u64(a,b) __simd_vqadd_u64(a,b)
#define vqaddd_u64(a,b) __simd_vqaddd_u64(a,b)
#define vqsubb_s8(a,b) __simd_vqsubb_s8(a,b)
#define vqsubh_s16(a,b) __simd_vqsubh_s16(a,b)
#define vqsubs_s32(a,b) __simd_vqsubs_s32(a,b)
#define vqsub_s64(a,b) __simd_vqsub_s64(a,b)
#define vqsubd_s64(a,b) __simd_vqsubd_s64(a,b)
#define vqsubb_u8(a,b) __simd_vqsubb_u8(a,b)
#define vqsubh_u16(a,b) __simd_vqsubh_u16(a,b)
#define vqsubs_u32(a,b) __simd_vqsubs_u32(a,b)
#define vqsub_u64(a,b) __simd_vqsub_u64(a,b)
#define vqsubd_u64(a,b) __simd_vqsubd_u64(a,b)
#define vadd_u64(a,b) __simd_vadd_u64(a,b)
#define vadd_s64(a,b) __simd_vadd_s64(a,b)
#define vaddd_u64(a,b) __simd_vaddd_u64(a,b)
#define vaddd_s64(a,b) __simd_vaddd_s64(a,b)
#define vsub_u64(a,b) __simd_vsub_u64(a,b)
#define vsub_s64(a,b) __simd_vsub_s64(a,b)
#define vsubd_u64(a,b) __simd_vsubd_u64(a,b)
#define vsubd_s64(a,b) __simd_vsubd_s64(a,b)
#define vqdmulhh_s16(a,b) __simd_vqdmulhh_s16(a,b)
#define vqdmulhs_s32(a,b) __simd_vqdmulhs_s32(a,b)
#define vqrdmulhh_s16(a,b) __simd_vqrdmulhh_s16(a,b)
#define vqrdmulhs_s32(a,b) __simd_vqrdmulhs_s32(a,b)
#define vshl_s64(a,b) __simd_vshl_s64(a,b)
#define vshld_s64(a,b) __simd_vshld_s64(a,b)
#define vshl_u64(a,b) __simd_vshl_u64(a,b)
#define vshld_u64(a,b) __simd_vshld_u64(a,b)
#define vrshl_s64(a,b) __simd_vrshl_s64(a,b)
#define vrshld_s64(a,b) __simd_vrshld_s64(a,b)
#define vrshl_u64(a,b) __simd_vrshl_u64(a,b)
#define vrshld_u64(a,b) __simd_vrshld_u64(a,b)
#define vqshlb_s8(a,b) __simd_vqshlb_s8(a,b)
#define vqshlh_s16(a,b) __simd_vqshlh_s16(a,b)
#define vqshls_s32(a,b) __simd_vqshls_s32(a,b)
#define vqshl_s64(a,b) __simd_vqshl_s64(a,b)
#define vqshld_s64(a,b) __simd_vqshld_s64(a,b)
#define vqshlb_u8(a,b) __simd_vqshlb_u8(a,b)
#define vqshlh_u16(a,b) __simd_vqshlh_u16(a,b)
#define vqshls_u32(a,b) __simd_vqshls_u32(a,b)
#define vqshl_u64(a,b) __simd_vqshl_u64(a,b)
#define vqshld_u64(a,b) __simd_vqshld_u64(a,b)
#define vqrshlb_s8(a,b) __simd_vqrshlb_s8(a,b)
#define vqrshlh_s16(a,b) __simd_vqrshlh_s16(a,b)
#define vqrshls_s32(a,b) __simd_vqrshls_s32(a,b)
#define vqrshl_s64(a,b) __simd_vqrshl_s64(a,b)
#define vqrshld_s64(a,b) __simd_vqrshld_s64(a,b)
#define vqrshlb_u8(a,b) __simd_vqrshlb_u8(a,b)
#define vqrshlh_u16(a,b) __simd_vqrshlh_u16(a,b)
#define vqrshls_u32(a,b) __simd_vqrshls_u32(a,b)
#define vqrshl_u64(a,b) __simd_vqrshl_u64(a,b)
#define vqrshld_u64(a,b) __simd_vqrshld_u64(a,b)
#define vceq_u64(a,b) __simd_vceq_u64(a,b)
#define vceq_s64(a,b) __simd_vceq_s64(a,b)
#define vceqd_u64(a,b) __simd_vceqd_u64(a,b)
#define vceqd_s64(a,b) __simd_vceqd_s64(a,b)
#define vcge_s64(a,b) __simd_vcge_s64(a,b)
#define vcged_s64(a,b) __simd_vcged_s64(a,b)
#define vcgt_s64(a,b) __simd_vcgt_s64(a,b)
#define vcgtd_s64(a,b) __simd_vcgtd_s64(a,b)
#define vcge_u64(a,b) __simd_vcge_u64(a,b)
#define vcged_u64(a,b) __simd_vcged_u64(a,b)
#define vcgt_u64(a,b) __simd_vcgt_u64(a,b)
#define vcgtd_u64(a,b) __simd_vcgtd_u64(a,b)
#define vtst_u64(a,b) __simd_vtst_u64(a,b)
#define vtst_s64(a,b) __simd_vtst_s64(a,b)
#define vtstd_u64(a,b) __simd_vtstd_u64(a,b)
#define vtstd_s64(a,b) __simd_vtstd_s64(a,b)
#define vcle_s64(a,b) __simd_vcle_s64(a,b)
#define vcled_s64(a,b) __simd_vcled_s64(a,b)
#define vclt_s64(a,b) __simd_vclt_s64(a,b)
#define vcltd_s64(a,b) __simd_vcltd_s64(a,b)
#define vcle_u64(a,b) __simd_vcle_u64(a,b)
#define vcled_u64(a,b) __simd_vcled_u64(a,b)
#define vclt_u64(a,b) __simd_vclt_u64(a,b)
#define vcltd_u64(a,b) __simd_vcltd_u64(a,b)
#define vadd_f32(a,b) __simd_vadd_f32(a,b)
#define vaddq_f32(a,b) __simd_vaddq_f32(a,b)
#define vaddq_f64(a,b) __simd_vaddq_f64(a,b)
#define vpadd_f32(a,b) __simd_vpadd_f32(a,b)
#define vpaddq_f32(a,b) __simd_vpaddq_f32(a,b)
#define vpaddq_f64(a,b) __simd_vpaddq_f64(a,b)
#define vsub_f32(a,b) __simd_vsub_f32(a,b)
#define vsubq_f32(a,b) __simd_vsubq_f32(a,b)
#define vsubq_f64(a,b) __simd_vsubq_f64(a,b)
#define vmul_f32(a,b) __simd_vmul_f32(a,b)
#define vmulq_f32(a,b) __simd_vmulq_f32(a,b)
#define vmulq_f64(a,b) __simd_vmulq_f64(a,b)
#define vmulx_f32(a,b) __simd_vmulx_f32(a,b)
#define vmulxq_f32(a,b) __simd_vmulxq_f32(a,b)
#define vmulxq_f64(a,b) __simd_vmulxq_f64(a,b)
#define vfma_f32(a,b,c) __simd_vfma_f32(a,b,c)
#define vfmaq_f32(a,b,c) __simd_vfmaq_f32(a,b,c)
#define vfmaq_f64(a,b,c) __simd_vfmaq_f64(a,b,c)
#define vfms_f32(a,b,c) __simd_vfms_f32(a,b,c)
#define vfmsq_f32(a,b,c) __simd_vfmsq_f32(a,b,c)
#define vfmsq_f64(a,b,c) __simd_vfmsq_f64(a,b,c)
#define vdiv_f32(a,b) __simd_vdiv_f32(a,b)
#define vdivq_f32(a,b) __simd_vdivq_f32(a,b)
#define vdivq_f64(a,b) __simd_vdivq_f64(a,b)
#define vceq_f32(a,b) __simd_vceq_f32(a,b)
#define vceqq_f32(a,b) __simd_vceqq_f32(a,b)
#define vceqq_f64(a,b) __simd_vceqq_f64(a,b)
#define vcge_f32(a,b) __simd_vcge_f32(a,b)
#define vcgeq_f32(a,b) __simd_vcgeq_f32(a,b)
#define vcgeq_f64(a,b) __simd_vcgeq_f64(a,b)
#define vcgt_f32(a,b) __simd_vcgt_f32(a,b)
#define vcgtq_f32(a,b) __simd_vcgtq_f32(a,b)
#define vcgtq_f64(a,b) __simd_vcgtq_f64(a,b)
#define vcage_f32(a,b) __simd_vcage_f32(a,b)
#define vcageq_f32(a,b) __simd_vcageq_f32(a,b)
#define vcageq_f64(a,b) __simd_vcageq_f64(a,b)
#define vcagt_f32(a,b) __simd_vcagt_f32(a,b)
#define vcagtq_f32(a,b) __simd_vcagtq_f32(a,b)
#define vcagtq_f64(a,b) __simd_vcagtq_f64(a,b)
#define vcle_f32(a,b) __simd_vcle_f32(a,b)
#define vcleq_f32(a,b) __simd_vcleq_f32(a,b)
#define vcleq_f64(a,b) __simd_vcleq_f64(a,b)
#define vclt_f32(a,b) __simd_vclt_f32(a,b)
#define vcltq_f32(a,b) __simd_vcltq_f32(a,b)
#define vcltq_f64(a,b) __simd_vcltq_f64(a,b)
#define vcale_f32(a,b) __simd_vcale_f32(a,b)
#define vcaleq_f32(a,b) __simd_vcaleq_f32(a,b)
#define vcaleq_f64(a,b) __simd_vcaleq_f64(a,b)
#define vcalt_f32(a,b) __simd_vcalt_f32(a,b)
#define vcaltq_f32(a,b) __simd_vcaltq_f32(a,b)
#define vcaltq_f64(a,b) __simd_vcaltq_f64(a,b)
#define vrecps_f32(a,b) __simd_vrecps_f32(a,b)
#define vrecpsq_f32(a,b) __simd_vrecpsq_f32(a,b)
#define vrecpsq_f64(a,b) __simd_vrecpsq_f64(a,b)
#define vrsqrts_f32(a,b) __simd_vrsqrts_f32(a,b)
#define vrsqrtsq_f32(a,b) __simd_vrsqrtsq_f32(a,b)
#define vrsqrtsq_f64(a,b) __simd_vrsqrtsq_f64(a,b)
#define vabd_f32(a,b) __simd_vabd_f32(a,b)
#define vabdq_f32(a,b) __simd_vabdq_f32(a,b)
#define vabdq_f64(a,b) __simd_vabdq_f64(a,b)
#define vmax_f32(a,b) __simd_vmax_f32(a,b)
#define vmaxq_f32(a,b) __simd_vmaxq_f32(a,b)
#define vmaxq_f64(a,b) __simd_vmaxq_f64(a,b)
#define vmaxnm_f32(a,b) __simd_vmaxnm_f32(a,b)
#define vmaxnmq_f32(a,b) __simd_vmaxnmq_f32(a,b)
#define vmaxnmq_f64(a,b) __simd_vmaxnmq_f64(a,b)
#define vmin_f32(a,b) __simd_vmin_f32(a,b)
#define vminq_f32(a,b) __simd_vminq_f32(a,b)
#define vminq_f64(a,b) __simd_vminq_f64(a,b)
#define vminnm_f32(a,b) __simd_vminnm_f32(a,b)
#define vminnmq_f32(a,b) __simd_vminnmq_f32(a,b)
#define vminnmq_f64(a,b) __simd_vminnmq_f64(a,b)
#define vpmax_f32(a,b) __simd_vpmax_f32(a,b)
#define vpmaxq_f32(a,b) __simd_vpmaxq_f32(a,b)
#define vpmaxq_f64(a,b) __simd_vpmaxq_f64(a,b)
#define vpmaxnm_f32(a,b) __simd_vpmaxnm_f32(a,b)
#define vpmaxnmq_f32(a,b) __simd_vpmaxnmq_f32(a,b)
#define vpmaxnmq_f64(a,b) __simd_vpmaxnmq_f64(a,b)
#define vpmin_f32(a,b) __simd_vpmin_f32(a,b)
#define vpminq_f32(a,b) __simd_vpminq_f32(a,b)
#define vpminq_f64(a,b) __simd_vpminq_f64(a,b)
#define vpminnm_f32(a,b) __simd_vpminnm_f32(a,b)
#define vpminnmq_f32(a,b) __simd_vpminnmq_f32(a,b)
#define vpminnmq_f64(a,b) __simd_vpminnmq_f64(a,b)
#define vmulxs_f32(a,b) __simd_vmulxs_f32(a,b)
#define vmulxd_f64(a,b) __simd_vmulxd_f64(a,b)
#define vabds_f32(a,b) __simd_vabds_f32(a,b)
#define vabdd_f64(a,b) __simd_vabdd_f64(a,b)
#define vrecpss_f32(a,b) __simd_vrecpss_f32(a,b)
#define vrecpsd_f64(a,b) __simd_vrecpsd_f64(a,b)
#define vrsqrtss_f32(a,b) __simd_vrsqrtss_f32(a,b)
#define vrsqrtsd_f64(a,b) __simd_vrsqrtsd_f64(a,b)
#define vceqs_f32(a,b) __simd_vceqs_f32(a,b)
#define vceqd_f64(a,b) __simd_vceqd_f64(a,b)
#define vcges_f32(a,b) __simd_vcges_f32(a,b)
#define vcged_f64(a,b) __simd_vcged_f64(a,b)
#define vcgts_f32(a,b) __simd_vcgts_f32(a,b)
#define vcgtd_f64(a,b) __simd_vcgtd_f64(a,b)
#define vcages_f32(a,b) __simd_vcages_f32(a,b)
#define vcaged_f64(a,b) __simd_vcaged_f64(a,b)
#define vcagts_f32(a,b) __simd_vcagts_f32(a,b)
#define vcagtd_f64(a,b) __simd_vcagtd_f64(a,b)
#define vcles_f32(a,b) __simd_vcles_f32(a,b)
#define vcled_f64(a,b) __simd_vcled_f64(a,b)
#define vclts_f32(a,b) __simd_vclts_f32(a,b)
#define vcltd_f64(a,b) __simd_vcltd_f64(a,b)
#define vcales_f32(a,b) __simd_vcales_f32(a,b)
#define vcaled_f64(a,b) __simd_vcaled_f64(a,b)
#define vcalts_f32(a,b) __simd_vcalts_f32(a,b)
#define vcaltd_f64(a,b) __simd_vcaltd_f64(a,b)
#define vaddl_s8(a,b) __simd_vaddl_s8(a,b)
#define vaddl_s16(a,b) __simd_vaddl_s16(a,b)
#define vaddl_s32(a,b) __simd_vaddl_s32(a,b)
#define vaddl_high_s8(a,b) __simd_vaddl_high_s8(a,b)
#define vaddl_high_s16(a,b) __simd_vaddl_high_s16(a,b)
#define vaddl_high_s32(a,b) __simd_vaddl_high_s32(a,b)
#define vaddl_u8(a,b) __simd_vaddl_u8(a,b)
#define vaddl_u16(a,b) __simd_vaddl_u16(a,b)
#define vaddl_u32(a,b) __simd_vaddl_u32(a,b)
#define vaddl_high_u8(a,b) __simd_vaddl_high_u8(a,b)
#define vaddl_high_u16(a,b) __simd_vaddl_high_u16(a,b)
#define vaddl_high_u32(a,b) __simd_vaddl_high_u32(a,b)
#define vaddw_s8(a,b) __simd_vaddw_s8(a,b)
#define vaddw_s16(a,b) __simd_vaddw_s16(a,b)
#define vaddw_s32(a,b) __simd_vaddw_s32(a,b)
#define vaddw_high_s8(a,b) __simd_vaddw_high_s8(a,b)
#define vaddw_high_s16(a,b) __simd_vaddw_high_s16(a,b)
#define vaddw_high_s32(a,b) __simd_vaddw_high_s32(a,b)
#define vaddw_u8(a,b) __simd_vaddw_u8(a,b)
#define vaddw_u16(a,b) __simd_vaddw_u16(a,b)
#define vaddw_u32(a,b) __simd_vaddw_u32(a,b)
#define vaddw_high_u8(a,b) __simd_vaddw_high_u8(a,b)
#define vaddw_high_u16(a,b) __simd_vaddw_high_u16(a,b)
#define vaddw_high_u32(a,b) __simd_vaddw_high_u32(a,b)
#define vsubl_s8(a,b) __simd_vsubl_s8(a,b)
#define vsubl_s16(a,b) __simd_vsubl_s16(a,b)
#define vsubl_s32(a,b) __simd_vsubl_s32(a,b)
#define vsubl_high_s8(a,b) __simd_vsubl_high_s8(a,b)
#define vsubl_high_s16(a,b) __simd_vsubl_high_s16(a,b)
#define vsubl_high_s32(a,b) __simd_vsubl_high_s32(a,b)
#define vsubl_u8(a,b) __simd_vsubl_u8(a,b)
#define vsubl_u16(a,b) __simd_vsubl_u16(a,b)
#define vsubl_u32(a,b) __simd_vsubl_u32(a,b)
#define vsubl_high_u8(a,b) __simd_vsubl_high_u8(a,b)
#define vsubl_high_u16(a,b) __simd_vsubl_high_u16(a,b)
#define vsubl_high_u32(a,b) __simd_vsubl_high_u32(a,b)
#define vsubw_s8(a,b) __simd_vsubw_s8(a,b)
#define vsubw_s16(a,b) __simd_vsubw_s16(a,b)
#define vsubw_s32(a,b) __simd_vsubw_s32(a,b)
#define vsubw_high_s8(a,b) __simd_vsubw_high_s8(a,b)
#define vsubw_high_s16(a,b) __simd_vsubw_high_s16(a,b)
#define vsubw_high_s32(a,b) __simd_vsubw_high_s32(a,b)
#define vsubw_u8(a,b) __simd_vsubw_u8(a,b)
#define vsubw_u16(a,b) __simd_vsubw_u16(a,b)
#define vsubw_u32(a,b) __simd_vsubw_u32(a,b)
#define vsubw_high_u8(a,b) __simd_vsubw_high_u8(a,b)
#define vsubw_high_u16(a,b) __simd_vsubw_high_u16(a,b)
#define vsubw_high_u32(a,b) __simd_vsubw_high_u32(a,b)
#define vaddhn_u16(a,b) __simd_vaddhn_u16(a,b)
#define vaddhn_s16(a,b) __simd_vaddhn_s16(a,b)
#define vaddhn_u32(a,b) __simd_vaddhn_u32(a,b)
#define vaddhn_s32(a,b) __simd_vaddhn_s32(a,b)
#define vaddhn_u64(a,b) __simd_vaddhn_u64(a,b)
#define vaddhn_s64(a,b) __simd_vaddhn_s64(a,b)
#define vaddhn_high_u16(a,b,c) __simd_vaddhn_high_u16(a,b,c)
#define vaddhn_high_s16(a,b,c) __simd_vaddhn_high_s16(a,b,c)
#define vaddhn_high_u32(a,b,c) __simd_vaddhn_high_u32(a,b,c)
#define vaddhn_high_s32(a,b,c) __simd_vaddhn_high_s32(a,b,c)
#define vaddhn_high_u64(a,b,c) __simd_vaddhn_high_u64(a,b,c)
#define vaddhn_high_s64(a,b,c) __simd_vaddhn_high_s64(a,b,c)
#define vraddhn_u16(a,b) __simd_vraddhn_u16(a,b)
#define vraddhn_s16(a,b) __simd_vraddhn_s16(a,b)
#define vraddhn_u32(a,b) __simd_vraddhn_u32(a,b)
#define vraddhn_s32(a,b) __simd_vraddhn_s32(a,b)
#define vraddhn_u64(a,b) __simd_vraddhn_u64(a,b)
#define vraddhn_s64(a,b) __simd_vraddhn_s64(a,b)
#define vraddhn_high_u16(a,b,c) __simd_vraddhn_high_u16(a,b,c)
#define vraddhn_high_s16(a,b,c) __simd_vraddhn_high_s16(a,b,c)
#define vraddhn_high_u32(a,b,c) __simd_vraddhn_high_u32(a,b,c)
#define vraddhn_high_s32(a,b,c) __simd_vraddhn_high_s32(a,b,c)
#define vraddhn_high_u64(a,b,c) __simd_vraddhn_high_u64(a,b,c)
#define vraddhn_high_s64(a,b,c) __simd_vraddhn_high_s64(a,b,c)
#define vsubhn_u16(a,b) __simd_vsubhn_u16(a,b)
#define vsubhn_s16(a,b) __simd_vsubhn_s16(a,b)
#define vsubhn_u32(a,b) __simd_vsubhn_u32(a,b)
#define vsubhn_s32(a,b) __simd_vsubhn_s32(a,b)
#define vsubhn_u64(a,b) __simd_vsubhn_u64(a,b)
#define vsubhn_s64(a,b) __simd_vsubhn_s64(a,b)
#define vsubhn_high_u16(a,b,c) __simd_vsubhn_high_u16(a,b,c)
#define vsubhn_high_s16(a,b,c) __simd_vsubhn_high_s16(a,b,c)
#define vsubhn_high_u32(a,b,c) __simd_vsubhn_high_u32(a,b,c)
#define vsubhn_high_s32(a,b,c) __simd_vsubhn_high_s32(a,b,c)
#define vsubhn_high_u64(a,b,c) __simd_vsubhn_high_u64(a,b,c)
#define vsubhn_high_s64(a,b,c) __simd_vsubhn_high_s64(a,b,c)
#define vrsubhn_u16(a,b) __simd_vrsubhn_u16(a,b)
#define vrsubhn_s16(a,b) __simd_vrsubhn_s16(a,b)
#define vrsubhn_u32(a,b) __simd_vrsubhn_u32(a,b)
#define vrsubhn_s32(a,b) __simd_vrsubhn_s32(a,b)
#define vrsubhn_u64(a,b) __simd_vrsubhn_u64(a,b)
#define vrsubhn_s64(a,b) __simd_vrsubhn_s64(a,b)
#define vrsubhn_high_u16(a,b,c) __simd_vrsubhn_high_u16(a,b,c)
#define vrsubhn_high_s16(a,b,c) __simd_vrsubhn_high_s16(a,b,c)
#define vrsubhn_high_u32(a,b,c) __simd_vrsubhn_high_u32(a,b,c)
#define vrsubhn_high_s32(a,b,c) __simd_vrsubhn_high_s32(a,b,c)
#define vrsubhn_high_u64(a,b,c) __simd_vrsubhn_high_u64(a,b,c)
#define vrsubhn_high_s64(a,b,c) __simd_vrsubhn_high_s64(a,b,c)
#define vabal_s8(a,b,c) __simd_vabal_s8(a,b,c)
#define vabal_s16(a,b,c) __simd_vabal_s16(a,b,c)
#define vabal_s32(a,b,c) __simd_vabal_s32(a,b,c)
#define vabal_high_s8(a,b,c) __simd_vabal_high_s8(a,b,c)
#define vabal_high_s16(a,b,c) __simd_vabal_high_s16(a,b,c)
#define vabal_high_s32(a,b,c) __simd_vabal_high_s32(a,b,c)
#define vabal_u8(a,b,c) __simd_vabal_u8(a,b,c)
#define vabal_u16(a,b,c) __simd_vabal_u16(a,b,c)
#define vabal_u32(a,b,c) __simd_vabal_u32(a,b,c)
#define vabal_high_u8(a,b,c) __simd_vabal_high_u8(a,b,c)
#define vabal_high_u16(a,b,c) __simd_vabal_high_u16(a,b,c)
#define vabal_high_u32(a,b,c) __simd_vabal_high_u32(a,b,c)
#define vabdl_s8(a,b) __simd_vabdl_s8(a,b)
#define vabdl_s16(a,b) __simd_vabdl_s16(a,b)
#define vabdl_s32(a,b) __simd_vabdl_s32(a,b)
#define vabdl_high_s8(a,b) __simd_vabdl_high_s8(a,b)
#define vabdl_high_s16(a,b) __simd_vabdl_high_s16(a,b)
#define vabdl_high_s32(a,b) __simd_vabdl_high_s32(a,b)
#define vabdl_u8(a,b) __simd_vabdl_u8(a,b)
#define vabdl_u16(a,b) __simd_vabdl_u16(a,b)
#define vabdl_u32(a,b) __simd_vabdl_u32(a,b)
#define vabdl_high_u8(a,b) __simd_vabdl_high_u8(a,b)
#define vabdl_high_u16(a,b) __simd_vabdl_high_u16(a,b)
#define vabdl_high_u32(a,b) __simd_vabdl_high_u32(a,b)
#define vmlal_s8(a,b,c) __simd_vmlal_s8(a,b,c)
#define vmlal_s16(a,b,c) __simd_vmlal_s16(a,b,c)
#define vmlal_s32(a,b,c) __simd_vmlal_s32(a,b,c)
#define vmlal_high_s8(a,b,c) __simd_vmlal_high_s8(a,b,c)
#define vmlal_high_s16(a,b,c) __simd_vmlal_high_s16(a,b,c)
#define vmlal_high_s32(a,b,c) __simd_vmlal_high_s32(a,b,c)
#define vmlal_u8(a,b,c) __simd_vmlal_u8(a,b,c)
#define vmlal_u16(a,b,c) __simd_vmlal_u16(a,b,c)
#define vmlal_u32(a,b,c) __simd_vmlal_u32(a,b,c)
#define vmlal_high_u8(a,b,c) __simd_vmlal_high_u8(a,b,c)
#define vmlal_high_u16(a,b,c) __simd_vmlal_high_u16(a,b,c)
#define vmlal_high_u32(a,b,c) __simd_vmlal_high_u32(a,b,c)
#define vmlsl_s8(a,b,c) __simd_vmlsl_s8(a,b,c)
#define vmlsl_s16(a,b,c) __simd_vmlsl_s16(a,b,c)
#define vmlsl_s32(a,b,c) __simd_vmlsl_s32(a,b,c)
#define vmlsl_high_s8(a,b,c) __simd_vmlsl_high_s8(a,b,c)
#define vmlsl_high_s16(a,b,c) __simd_vmlsl_high_s16(a,b,c)
#define vmlsl_high_s32(a,b,c) __simd_vmlsl_high_s32(a,b,c)
#define vmlsl_u8(a,b,c) __simd_vmlsl_u8(a,b,c)
#define vmlsl_u16(a,b,c) __simd_vmlsl_u16(a,b,c)
#define vmlsl_u32(a,b,c) __simd_vmlsl_u32(a,b,c)
#define vmlsl_high_u8(a,b,c) __simd_vmlsl_high_u8(a,b,c)
#define vmlsl_high_u16(a,b,c) __simd_vmlsl_high_u16(a,b,c)
#define vmlsl_high_u32(a,b,c) __simd_vmlsl_high_u32(a,b,c)
#define vmull_s8(a,b) __simd_vmull_s8(a,b)
#define vmull_s16(a,b) __simd_vmull_s16(a,b)
#define vmull_s32(a,b) __simd_vmull_s32(a,b)
#define vmull_high_s8(a,b) __simd_vmull_high_s8(a,b)
#define vmull_high_s16(a,b) __simd_vmull_high_s16(a,b)
#define vmull_high_s32(a,b) __simd_vmull_high_s32(a,b)
#define vmull_u8(a,b) __simd_vmull_u8(a,b)
#define vmull_u16(a,b) __simd_vmull_u16(a,b)
#define vmull_u32(a,b) __simd_vmull_u32(a,b)
#define vmull_high_u8(a,b) __simd_vmull_high_u8(a,b)
#define vmull_high_u16(a,b) __simd_vmull_high_u16(a,b)
#define vmull_high_u32(a,b) __simd_vmull_high_u32(a,b)
#define vqdmlal_s16(a,b,c) __simd_vqdmlal_s16(a,b,c)
#define vqdmlal_s32(a,b,c) __simd_vqdmlal_s32(a,b,c)
#define vqdmlal_high_s16(a,b,c) __simd_vqdmlal_high_s16(a,b,c)
#define vqdmlal_high_s32(a,b,c) __simd_vqdmlal_high_s32(a,b,c)
#define vqdmlsl_s16(a,b,c) __simd_vqdmlsl_s16(a,b,c)
#define vqdmlsl_s32(a,b,c) __simd_vqdmlsl_s32(a,b,c)
#define vqdmlsl_high_s16(a,b,c) __simd_vqdmlsl_high_s16(a,b,c)
#define vqdmlsl_high_s32(a,b,c) __simd_vqdmlsl_high_s32(a,b,c)
#define vqdmull_s16(a,b) __simd_vqdmull_s16(a,b)
#define vqdmull_s32(a,b) __simd_vqdmull_s32(a,b)
#define vqdmull_high_s16(a,b) __simd_vqdmull_high_s16(a,b)
#define vqdmull_high_s32(a,b) __simd_vqdmull_high_s32(a,b)
#define vmull_p8(a,b) __simd_vmull_p8(a,b)
#define vmull_high_p8(a,b) __simd_vmull_high_p8(a,b)
#define vqdmlalh_s16(a,b,c) __simd_vqdmlalh_s16(a,b,c)
#define vqdmlals_s32(a,b,c) __simd_vqdmlals_s32(a,b,c)
#define vqdmlslh_s16(a,b,c) __simd_vqdmlslh_s16(a,b,c)
#define vqdmlsls_s32(a,b,c) __simd_vqdmlsls_s32(a,b,c)
#define vqdmullh_s16(a,b) __simd_vqdmullh_s16(a,b)
#define vqdmulls_s32(a,b) __simd_vqdmulls_s32(a,b)
#define vmlal_lane_s16(a,b,c,d) __simd_vmlal_lane_s16(a,b,c,d)
#define vmlal_laneq_s16(a,b,c,d) __simd_vmlal_laneq_s16(a,b,c,d)
#define vmlal_n_s16(a,b,c) __simd_vmlal_n_s16(a,b,c)
#define vmlal_lane_s32(a,b,c,d) __simd_vmlal_lane_s32(a,b,c,d)
#define vmlal_laneq_s32(a,b,c,d) __simd_vmlal_laneq_s32(a,b,c,d)
#define vmlal_n_s32(a,b,c) __simd_vmlal_n_s32(a,b,c)
#define vmlal_high_lane_s16(a,b,c,d) __simd_vmlal_high_lane_s16(a,b,c,d)
#define vmlal_high_laneq_s16(a,b,c,d) __simd_vmlal_high_laneq_s16(a,b,c,d)
#define vmlal_high_n_s16(a,b,c) __simd_vmlal_high_n_s16(a,b,c)
#define vmlal_high_lane_s32(a,b,c,d) __simd_vmlal_high_lane_s32(a,b,c,d)
#define vmlal_high_laneq_s32(a,b,c,d) __simd_vmlal_high_laneq_s32(a,b,c,d)
#define vmlal_high_n_s32(a,b,c) __simd_vmlal_high_n_s32(a,b,c)
#define vmlal_lane_u16(a,b,c,d) __simd_vmlal_lane_u16(a,b,c,d)
#define vmlal_laneq_u16(a,b,c,d) __simd_vmlal_laneq_u16(a,b,c,d)
#define vmlal_n_u16(a,b,c) __simd_vmlal_n_u16(a,b,c)
#define vmlal_lane_u32(a,b,c,d) __simd_vmlal_lane_u32(a,b,c,d)
#define vmlal_laneq_u32(a,b,c,d) __simd_vmlal_laneq_u32(a,b,c,d)
#define vmlal_n_u32(a,b,c) __simd_vmlal_n_u32(a,b,c)
#define vmlal_high_lane_u16(a,b,c,d) __simd_vmlal_high_lane_u16(a,b,c,d)
#define vmlal_high_laneq_u16(a,b,c,d) __simd_vmlal_high_laneq_u16(a,b,c,d)
#define vmlal_high_n_u16(a,b,c) __simd_vmlal_high_n_u16(a,b,c)
#define vmlal_high_lane_u32(a,b,c,d) __simd_vmlal_high_lane_u32(a,b,c,d)
#define vmlal_high_laneq_u32(a,b,c,d) __simd_vmlal_high_laneq_u32(a,b,c,d)
#define vmlal_high_n_u32(a,b,c) __simd_vmlal_high_n_u32(a,b,c)
#define vqdmlal_lane_s16(a,b,c,d) __simd_vqdmlal_lane_s16(a,b,c,d)
#define vqdmlal_laneq_s16(a,b,c,d) __simd_vqdmlal_laneq_s16(a,b,c,d)
#define vqdmlal_n_s16(a,b,c) __simd_vqdmlal_n_s16(a,b,c)
#define vqdmlal_lane_s32(a,b,c,d) __simd_vqdmlal_lane_s32(a,b,c,d)
#define vqdmlal_laneq_s32(a,b,c,d) __simd_vqdmlal_laneq_s32(a,b,c,d)
#define vqdmlal_n_s32(a,b,c) __simd_vqdmlal_n_s32(a,b,c)
#define vqdmlal_high_lane_s16(a,b,c,d) __simd_vqdmlal_high_lane_s16(a,b,c,d)
#define vqdmlal_high_laneq_s16(a,b,c,d) __simd_vqdmlal_high_laneq_s16(a,b,c,d)
#define vqdmlal_high_n_s16(a,b,c) __simd_vqdmlal_high_n_s16(a,b,c)
#define vqdmlal_high_lane_s32(a,b,c,d) __simd_vqdmlal_high_lane_s32(a,b,c,d)
#define vqdmlal_high_laneq_s32(a,b,c,d) __simd_vqdmlal_high_laneq_s32(a,b,c,d)
#define vqdmlal_high_n_s32(a,b,c) __simd_vqdmlal_high_n_s32(a,b,c)
#define vmlsl_lane_s16(a,b,c,d) __simd_vmlsl_lane_s16(a,b,c,d)
#define vmlsl_laneq_s16(a,b,c,d) __simd_vmlsl_laneq_s16(a,b,c,d)
#define vmlsl_n_s16(a,b,c) __simd_vmlsl_n_s16(a,b,c)
#define vmlsl_lane_s32(a,b,c,d) __simd_vmlsl_lane_s32(a,b,c,d)
#define vmlsl_laneq_s32(a,b,c,d) __simd_vmlsl_laneq_s32(a,b,c,d)
#define vmlsl_n_s32(a,b,c) __simd_vmlsl_n_s32(a,b,c)
#define vmlsl_high_lane_s16(a,b,c,d) __simd_vmlsl_high_lane_s16(a,b,c,d)
#define vmlsl_high_laneq_s16(a,b,c,d) __simd_vmlsl_high_laneq_s16(a,b,c,d)
#define vmlsl_high_n_s16(a,b,c) __simd_vmlsl_high_n_s16(a,b,c)
#define vmlsl_high_lane_s32(a,b,c,d) __simd_vmlsl_high_lane_s32(a,b,c,d)
#define vmlsl_high_laneq_s32(a,b,c,d) __simd_vmlsl_high_laneq_s32(a,b,c,d)
#define vmlsl_high_n_s32(a,b,c) __simd_vmlsl_high_n_s32(a,b,c)
#define vmlsl_lane_u16(a,b,c,d) __simd_vmlsl_lane_u16(a,b,c,d)
#define vmlsl_laneq_u16(a,b,c,d) __simd_vmlsl_laneq_u16(a,b,c,d)
#define vmlsl_n_u16(a,b,c) __simd_vmlsl_n_u16(a,b,c)
#define vmlsl_lane_u32(a,b,c,d) __simd_vmlsl_lane_u32(a,b,c,d)
#define vmlsl_laneq_u32(a,b,c,d) __simd_vmlsl_laneq_u32(a,b,c,d)
#define vmlsl_n_u32(a,b,c) __simd_vmlsl_n_u32(a,b,c)
#define vmlsl_high_lane_u16(a,b,c,d) __simd_vmlsl_high_lane_u16(a,b,c,d)
#define vmlsl_high_laneq_u16(a,b,c,d) __simd_vmlsl_high_laneq_u16(a,b,c,d)
#define vmlsl_high_n_u16(a,b,c) __simd_vmlsl_high_n_u16(a,b,c)
#define vmlsl_high_lane_u32(a,b,c,d) __simd_vmlsl_high_lane_u32(a,b,c,d)
#define vmlsl_high_laneq_u32(a,b,c,d) __simd_vmlsl_high_laneq_u32(a,b,c,d)
#define vmlsl_high_n_u32(a,b,c) __simd_vmlsl_high_n_u32(a,b,c)
#define vqdmlsl_lane_s16(a,b,c,d) __simd_vqdmlsl_lane_s16(a,b,c,d)
#define vqdmlsl_laneq_s16(a,b,c,d) __simd_vqdmlsl_laneq_s16(a,b,c,d)
#define vqdmlsl_n_s16(a,b,c) __simd_vqdmlsl_n_s16(a,b,c)
#define vqdmlsl_lane_s32(a,b,c,d) __simd_vqdmlsl_lane_s32(a,b,c,d)
#define vqdmlsl_laneq_s32(a,b,c,d) __simd_vqdmlsl_laneq_s32(a,b,c,d)
#define vqdmlsl_n_s32(a,b,c) __simd_vqdmlsl_n_s32(a,b,c)
#define vqdmlsl_high_lane_s16(a,b,c,d) __simd_vqdmlsl_high_lane_s16(a,b,c,d)
#define vqdmlsl_high_laneq_s16(a,b,c,d) __simd_vqdmlsl_high_laneq_s16(a,b,c,d)
#define vqdmlsl_high_n_s16(a,b,c) __simd_vqdmlsl_high_n_s16(a,b,c)
#define vqdmlsl_high_lane_s32(a,b,c,d) __simd_vqdmlsl_high_lane_s32(a,b,c,d)
#define vqdmlsl_high_laneq_s32(a,b,c,d) __simd_vqdmlsl_high_laneq_s32(a,b,c,d)
#define vqdmlsl_high_n_s32(a,b,c) __simd_vqdmlsl_high_n_s32(a,b,c)
#define vmull_lane_s16(a,b,c) __simd_vmull_lane_s16(a,b,c)
#define vmull_laneq_s16(a,b,c) __simd_vmull_laneq_s16(a,b,c)
#define vmull_n_s16(a,b) __simd_vmull_n_s16(a,b)
#define vmull_lane_s32(a,b,c) __simd_vmull_lane_s32(a,b,c)
#define vmull_laneq_s32(a,b,c) __simd_vmull_laneq_s32(a,b,c)
#define vmull_n_s32(a,b) __simd_vmull_n_s32(a,b)
#define vmull_high_lane_s16(a,b,c) __simd_vmull_high_lane_s16(a,b,c)
#define vmull_high_laneq_s16(a,b,c) __simd_vmull_high_laneq_s16(a,b,c)
#define vmull_high_n_s16(a,b) __simd_vmull_high_n_s16(a,b)
#define vmull_high_lane_s32(a,b,c) __simd_vmull_high_lane_s32(a,b,c)
#define vmull_high_laneq_s32(a,b,c) __simd_vmull_high_laneq_s32(a,b,c)
#define vmull_high_n_s32(a,b) __simd_vmull_high_n_s32(a,b)
#define vmull_lane_u16(a,b,c) __simd_vmull_lane_u16(a,b,c)
#define vmull_laneq_u16(a,b,c) __simd_vmull_laneq_u16(a,b,c)
#define vmull_n_u16(a,b) __simd_vmull_n_u16(a,b)
#define vmull_lane_u32(a,b,c) __simd_vmull_lane_u32(a,b,c)
#define vmull_laneq_u32(a,b,c) __simd_vmull_laneq_u32(a,b,c)
#define vmull_n_u32(a,b) __simd_vmull_n_u32(a,b)
#define vmull_high_lane_u16(a,b,c) __simd_vmull_high_lane_u16(a,b,c)
#define vmull_high_laneq_u16(a,b,c) __simd_vmull_high_laneq_u16(a,b,c)
#define vmull_high_n_u16(a,b) __simd_vmull_high_n_u16(a,b)
#define vmull_high_lane_u32(a,b,c) __simd_vmull_high_lane_u32(a,b,c)
#define vmull_high_laneq_u32(a,b,c) __simd_vmull_high_laneq_u32(a,b,c)
#define vmull_high_n_u32(a,b) __simd_vmull_high_n_u32(a,b)
#define vqdmull_lane_s16(a,b,c) __simd_vqdmull_lane_s16(a,b,c)
#define vqdmull_laneq_s16(a,b,c) __simd_vqdmull_laneq_s16(a,b,c)
#define vqdmull_n_s16(a,b) __simd_vqdmull_n_s16(a,b)
#define vqdmull_lane_s32(a,b,c) __simd_vqdmull_lane_s32(a,b,c)
#define vqdmull_laneq_s32(a,b,c) __simd_vqdmull_laneq_s32(a,b,c)
#define vqdmull_n_s32(a,b) __simd_vqdmull_n_s32(a,b)
#define vqdmull_high_lane_s16(a,b,c) __simd_vqdmull_high_lane_s16(a,b,c)
#define vqdmull_high_laneq_s16(a,b,c) __simd_vqdmull_high_laneq_s16(a,b,c)
#define vqdmull_high_n_s16(a,b) __simd_vqdmull_high_n_s16(a,b)
#define vqdmull_high_lane_s32(a,b,c) __simd_vqdmull_high_lane_s32(a,b,c)
#define vqdmull_high_laneq_s32(a,b,c) __simd_vqdmull_high_laneq_s32(a,b,c)
#define vqdmull_high_n_s32(a,b) __simd_vqdmull_high_n_s32(a,b)
#define vmul_lane_u16(a,b,c) __simd_vmul_lane_u16(a,b,c)
#define vmul_lane_s16(a,b,c) __simd_vmul_lane_s16(a,b,c)
#define vmul_laneq_u16(a,b,c) __simd_vmul_laneq_u16(a,b,c)
#define vmul_laneq_s16(a,b,c) __simd_vmul_laneq_s16(a,b,c)
#define vmul_n_u16(a,b) __simd_vmul_n_u16(a,b)
#define vmul_n_s16(a,b) __simd_vmul_n_s16(a,b)
#define vmul_lane_u32(a,b,c) __simd_vmul_lane_u32(a,b,c)
#define vmul_lane_s32(a,b,c) __simd_vmul_lane_s32(a,b,c)
#define vmul_laneq_u32(a,b,c) __simd_vmul_laneq_u32(a,b,c)
#define vmul_laneq_s32(a,b,c) __simd_vmul_laneq_s32(a,b,c)
#define vmul_n_u32(a,b) __simd_vmul_n_u32(a,b)
#define vmul_n_s32(a,b) __simd_vmul_n_s32(a,b)
#define vmulq_lane_u16(a,b,c) __simd_vmulq_lane_u16(a,b,c)
#define vmulq_lane_s16(a,b,c) __simd_vmulq_lane_s16(a,b,c)
#define vmulq_laneq_u16(a,b,c) __simd_vmulq_laneq_u16(a,b,c)
#define vmulq_laneq_s16(a,b,c) __simd_vmulq_laneq_s16(a,b,c)
#define vmulq_n_u16(a,b) __simd_vmulq_n_u16(a,b)
#define vmulq_n_s16(a,b) __simd_vmulq_n_s16(a,b)
#define vmulq_lane_u32(a,b,c) __simd_vmulq_lane_u32(a,b,c)
#define vmulq_lane_s32(a,b,c) __simd_vmulq_lane_s32(a,b,c)
#define vmulq_laneq_u32(a,b,c) __simd_vmulq_laneq_u32(a,b,c)
#define vmulq_laneq_s32(a,b,c) __simd_vmulq_laneq_s32(a,b,c)
#define vmulq_n_u32(a,b) __simd_vmulq_n_u32(a,b)
#define vmulq_n_s32(a,b) __simd_vmulq_n_s32(a,b)
#define vmla_lane_u16(a,b,c,d) __simd_vmla_lane_u16(a,b,c,d)
#define vmla_lane_s16(a,b,c,d) __simd_vmla_lane_s16(a,b,c,d)
#define vmla_laneq_u16(a,b,c,d) __simd_vmla_laneq_u16(a,b,c,d)
#define vmla_laneq_s16(a,b,c,d) __simd_vmla_laneq_s16(a,b,c,d)
#define vmla_n_u16(a,b,c) __simd_vmla_n_u16(a,b,c)
#define vmla_n_s16(a,b,c) __simd_vmla_n_s16(a,b,c)
#define vmla_lane_u32(a,b,c,d) __simd_vmla_lane_u32(a,b,c,d)
#define vmla_lane_s32(a,b,c,d) __simd_vmla_lane_s32(a,b,c,d)
#define vmla_laneq_u32(a,b,c,d) __simd_vmla_laneq_u32(a,b,c,d)
#define vmla_laneq_s32(a,b,c,d) __simd_vmla_laneq_s32(a,b,c,d)
#define vmla_n_u32(a,b,c) __simd_vmla_n_u32(a,b,c)
#define vmla_n_s32(a,b,c) __simd_vmla_n_s32(a,b,c)
#define vmlaq_lane_u16(a,b,c,d) __simd_vmlaq_lane_u16(a,b,c,d)
#define vmlaq_lane_s16(a,b,c,d) __simd_vmlaq_lane_s16(a,b,c,d)
#define vmlaq_laneq_u16(a,b,c,d) __simd_vmlaq_laneq_u16(a,b,c,d)
#define vmlaq_laneq_s16(a,b,c,d) __simd_vmlaq_laneq_s16(a,b,c,d)
#define vmlaq_n_u16(a,b,c) __simd_vmlaq_n_u16(a,b,c)
#define vmlaq_n_s16(a,b,c) __simd_vmlaq_n_s16(a,b,c)
#define vmlaq_lane_u32(a,b,c,d) __simd_vmlaq_lane_u32(a,b,c,d)
#define vmlaq_lane_s32(a,b,c,d) __simd_vmlaq_lane_s32(a,b,c,d)
#define vmlaq_laneq_u32(a,b,c,d) __simd_vmlaq_laneq_u32(a,b,c,d)
#define vmlaq_laneq_s32(a,b,c,d) __simd_vmlaq_laneq_s32(a,b,c,d)
#define vmlaq_n_u32(a,b,c) __simd_vmlaq_n_u32(a,b,c)
#define vmlaq_n_s32(a,b,c) __simd_vmlaq_n_s32(a,b,c)
#define vmls_lane_u16(a,b,c,d) __simd_vmls_lane_u16(a,b,c,d)
#define vmls_lane_s16(a,b,c,d) __simd_vmls_lane_s16(a,b,c,d)
#define vmls_laneq_u16(a,b,c,d) __simd_vmls_laneq_u16(a,b,c,d)
#define vmls_laneq_s16(a,b,c,d) __simd_vmls_laneq_s16(a,b,c,d)
#define vmls_n_u16(a,b,c) __simd_vmls_n_u16(a,b,c)
#define vmls_n_s16(a,b,c) __simd_vmls_n_s16(a,b,c)
#define vmls_lane_u32(a,b,c,d) __simd_vmls_lane_u32(a,b,c,d)
#define vmls_lane_s32(a,b,c,d) __simd_vmls_lane_s32(a,b,c,d)
#define vmls_laneq_u32(a,b,c,d) __simd_vmls_laneq_u32(a,b,c,d)
#define vmls_laneq_s32(a,b,c,d) __simd_vmls_laneq_s32(a,b,c,d)
#define vmls_n_u32(a,b,c) __simd_vmls_n_u32(a,b,c)
#define vmls_n_s32(a,b,c) __simd_vmls_n_s32(a,b,c)
#define vmlsq_lane_u16(a,b,c,d) __simd_vmlsq_lane_u16(a,b,c,d)
#define vmlsq_lane_s16(a,b,c,d) __simd_vmlsq_lane_s16(a,b,c,d)
#define vmlsq_laneq_u16(a,b,c,d) __simd_vmlsq_laneq_u16(a,b,c,d)
#define vmlsq_laneq_s16(a,b,c,d) __simd_vmlsq_laneq_s16(a,b,c,d)
#define vmlsq_n_u16(a,b,c) __simd_vmlsq_n_u16(a,b,c)
#define vmlsq_n_s16(a,b,c) __simd_vmlsq_n_s16(a,b,c)
#define vmlsq_lane_u32(a,b,c,d) __simd_vmlsq_lane_u32(a,b,c,d)
#define vmlsq_lane_s32(a,b,c,d) __simd_vmlsq_lane_s32(a,b,c,d)
#define vmlsq_laneq_u32(a,b,c,d) __simd_vmlsq_laneq_u32(a,b,c,d)
#define vmlsq_laneq_s32(a,b,c,d) __simd_vmlsq_laneq_s32(a,b,c,d)
#define vmlsq_n_u32(a,b,c) __simd_vmlsq_n_u32(a,b,c)
#define vmlsq_n_s32(a,b,c) __simd_vmlsq_n_s32(a,b,c)
#define vqdmulh_lane_s16(a,b,c) __simd_vqdmulh_lane_s16(a,b,c)
#define vqdmulh_laneq_s16(a,b,c) __simd_vqdmulh_laneq_s16(a,b,c)
#define vqdmulh_n_s16(a,b) __simd_vqdmulh_n_s16(a,b)
#define vqdmulh_lane_s32(a,b,c) __simd_vqdmulh_lane_s32(a,b,c)
#define vqdmulh_laneq_s32(a,b,c) __simd_vqdmulh_laneq_s32(a,b,c)
#define vqdmulh_n_s32(a,b) __simd_vqdmulh_n_s32(a,b)
#define vqdmulhq_lane_s16(a,b,c) __simd_vqdmulhq_lane_s16(a,b,c)
#define vqdmulhq_laneq_s16(a,b,c) __simd_vqdmulhq_laneq_s16(a,b,c)
#define vqdmulhq_n_s16(a,b) __simd_vqdmulhq_n_s16(a,b)
#define vqdmulhq_lane_s32(a,b,c) __simd_vqdmulhq_lane_s32(a,b,c)
#define vqdmulhq_laneq_s32(a,b,c) __simd_vqdmulhq_laneq_s32(a,b,c)
#define vqdmulhq_n_s32(a,b) __simd_vqdmulhq_n_s32(a,b)
#define vqrdmulh_lane_s16(a,b,c) __simd_vqrdmulh_lane_s16(a,b,c)
#define vqrdmulh_laneq_s16(a,b,c) __simd_vqrdmulh_laneq_s16(a,b,c)
#define vqrdmulh_n_s16(a,b) __simd_vqrdmulh_n_s16(a,b)
#define vqrdmulh_lane_s32(a,b,c) __simd_vqrdmulh_lane_s32(a,b,c)
#define vqrdmulh_laneq_s32(a,b,c) __simd_vqrdmulh_laneq_s32(a,b,c)
#define vqrdmulh_n_s32(a,b) __simd_vqrdmulh_n_s32(a,b)
#define vqrdmulhq_lane_s16(a,b,c) __simd_vqrdmulhq_lane_s16(a,b,c)
#define vqrdmulhq_laneq_s16(a,b,c) __simd_vqrdmulhq_laneq_s16(a,b,c)
#define vqrdmulhq_n_s16(a,b) __simd_vqrdmulhq_n_s16(a,b)
#define vqrdmulhq_lane_s32(a,b,c) __simd_vqrdmulhq_lane_s32(a,b,c)
#define vqrdmulhq_laneq_s32(a,b,c) __simd_vqrdmulhq_laneq_s32(a,b,c)
#define vqrdmulhq_n_s32(a,b) __simd_vqrdmulhq_n_s32(a,b)
#define vmul_lane_f32(a,b,c) __simd_vmul_lane_f32(a,b,c)
#define vmul_laneq_f32(a,b,c) __simd_vmul_laneq_f32(a,b,c)
#define vmul_n_f32(a,b) __simd_vmul_n_f32(a,b)
#define vmulq_lane_f32(a,b,c) __simd_vmulq_lane_f32(a,b,c)
#define vmulq_laneq_f32(a,b,c) __simd_vmulq_laneq_f32(a,b,c)
#define vmulq_n_f32(a,b) __simd_vmulq_n_f32(a,b)
#define vmulq_laneq_f64(a,b,c) __simd_vmulq_laneq_f64(a,b,c)
#define vmulq_n_f64(a,b) __simd_vmulq_n_f64(a,b)
#define vfma_lane_f32(a,b,c,d) __simd_vfma_lane_f32(a,b,c,d)
#define vfma_laneq_f32(a,b,c,d) __simd_vfma_laneq_f32(a,b,c,d)
#define vfma_n_f32(a,b,c) __simd_vfma_n_f32(a,b,c)
#define vfmaq_lane_f32(a,b,c,d) __simd_vfmaq_lane_f32(a,b,c,d)
#define vfmaq_laneq_f32(a,b,c,d) __simd_vfmaq_laneq_f32(a,b,c,d)
#define vfmaq_n_f32(a,b,c) __simd_vfmaq_n_f32(a,b,c)
#define vfmaq_laneq_f64(a,b,c,d) __simd_vfmaq_laneq_f64(a,b,c,d)
#define vfmaq_n_f64(a,b,c) __simd_vfmaq_n_f64(a,b,c)
#define vfms_lane_f32(a,b,c,d) __simd_vfms_lane_f32(a,b,c,d)
#define vfms_laneq_f32(a,b,c,d) __simd_vfms_laneq_f32(a,b,c,d)
#define vfms_n_f32(a,b,c) __simd_vfms_n_f32(a,b,c)
#define vfmsq_lane_f32(a,b,c,d) __simd_vfmsq_lane_f32(a,b,c,d)
#define vfmsq_laneq_f32(a,b,c,d) __simd_vfmsq_laneq_f32(a,b,c,d)
#define vfmsq_n_f32(a,b,c) __simd_vfmsq_n_f32(a,b,c)
#define vfmsq_laneq_f64(a,b,c,d) __simd_vfmsq_laneq_f64(a,b,c,d)
#define vfmsq_n_f64(a,b,c) __simd_vfmsq_n_f64(a,b,c)
#define vmulx_lane_f32(a,b,c) __simd_vmulx_lane_f32(a,b,c)
#define vmulx_laneq_f32(a,b,c) __simd_vmulx_laneq_f32(a,b,c)
#define vmulx_n_f32(a,b) __simd_vmulx_n_f32(a,b)
#define vmulxq_lane_f32(a,b,c) __simd_vmulxq_lane_f32(a,b,c)
#define vmulxq_laneq_f32(a,b,c) __simd_vmulxq_laneq_f32(a,b,c)
#define vmulxq_n_f32(a,b) __simd_vmulxq_n_f32(a,b)
#define vmulxq_laneq_f64(a,b,c) __simd_vmulxq_laneq_f64(a,b,c)
#define vmulxq_n_f64(a,b) __simd_vmulxq_n_f64(a,b)
#define vqdmlalh_lane_s16(a,b,c,d) __simd_vqdmlalh_lane_s16(a,b,c,d)
#define vqdmlalh_laneq_s16(a,b,c,d) __simd_vqdmlalh_laneq_s16(a,b,c,d)
#define vqdmlals_lane_s32(a,b,c,d) __simd_vqdmlals_lane_s32(a,b,c,d)
#define vqdmlals_laneq_s32(a,b,c,d) __simd_vqdmlals_laneq_s32(a,b,c,d)
#define vqdmlslh_lane_s16(a,b,c,d) __simd_vqdmlslh_lane_s16(a,b,c,d)
#define vqdmlslh_laneq_s16(a,b,c,d) __simd_vqdmlslh_laneq_s16(a,b,c,d)
#define vqdmlsls_lane_s32(a,b,c,d) __simd_vqdmlsls_lane_s32(a,b,c,d)
#define vqdmlsls_laneq_s32(a,b,c,d) __simd_vqdmlsls_laneq_s32(a,b,c,d)
#define vqdmullh_lane_s16(a,b,c) __simd_vqdmullh_lane_s16(a,b,c)
#define vqdmullh_laneq_s16(a,b,c) __simd_vqdmullh_laneq_s16(a,b,c)
#define vqdmulls_lane_s32(a,b,c) __simd_vqdmulls_lane_s32(a,b,c)
#define vqdmulls_laneq_s32(a,b,c) __simd_vqdmulls_laneq_s32(a,b,c)
#define vqdmulhh_lane_s16(a,b,c) __simd_vqdmulhh_lane_s16(a,b,c)
#define vqdmulhh_laneq_s16(a,b,c) __simd_vqdmulhh_laneq_s16(a,b,c)
#define vqdmulhs_lane_s32(a,b,c) __simd_vqdmulhs_lane_s32(a,b,c)
#define vqdmulhs_laneq_s32(a,b,c) __simd_vqdmulhs_laneq_s32(a,b,c)
#define vqrdmulhh_lane_s16(a,b,c) __simd_vqrdmulhh_lane_s16(a,b,c)
#define vqrdmulhh_laneq_s16(a,b,c) __simd_vqrdmulhh_laneq_s16(a,b,c)
#define vqrdmulhs_lane_s32(a,b,c) __simd_vqrdmulhs_lane_s32(a,b,c)
#define vqrdmulhs_laneq_s32(a,b,c) __simd_vqrdmulhs_laneq_s32(a,b,c)
#define vfmas_lane_f32(a,b,c,d) __simd_vfmas_lane_f32(a,b,c,d)
#define vfmas_laneq_f32(a,b,c,d) __simd_vfmas_laneq_f32(a,b,c,d)
#define vfmad_laneq_f64(a,b,c,d) __simd_vfmad_laneq_f64(a,b,c,d)
#define vfmss_lane_f32(a,b,c,d) __simd_vfmss_lane_f32(a,b,c,d)
#define vfmss_laneq_f32(a,b,c,d) __simd_vfmss_laneq_f32(a,b,c,d)
#define vfmsd_laneq_f64(a,b,c,d) __simd_vfmsd_laneq_f64(a,b,c,d)
#define vmuls_lane_f32(a,b,c) __simd_vmuls_lane_f32(a,b,c)
#define vmuls_laneq_f32(a,b,c) __simd_vmuls_laneq_f32(a,b,c)
#define vmuld_laneq_f64(a,b,c) __simd_vmuld_laneq_f64(a,b,c)
#define vmulxs_lane_f32(a,b,c) __simd_vmulxs_lane_f32(a,b,c)
#define vmulxs_laneq_f32(a,b,c) __simd_vmulxs_laneq_f32(a,b,c)
#define vmulxd_laneq_f64(a,b,c) __simd_vmulxd_laneq_f64(a,b,c)
#define vrev64_u8(a) __simd_vrev64_u8(a)
#define vrev64_s8(a) __simd_vrev64_s8(a)
#define vrev64_p8(a) __simd_vrev64_p8(a)
#define vrev64_u16(a) __simd_vrev64_u16(a)
#define vrev64_s16(a) __simd_vrev64_s16(a)
#define vrev64_p16(a) __simd_vrev64_p16(a)
#define vrev64_u32(a) __simd_vrev64_u32(a)
#define vrev64_s32(a) __simd_vrev64_s32(a)
#define vrev64_f32(a) __simd_vrev64_f32(a)
#define vrev64q_u8(a) __simd_vrev64q_u8(a)
#define vrev64q_s8(a) __simd_vrev64q_s8(a)
#define vrev64q_p8(a) __simd_vrev64q_p8(a)
#define vrev64q_u16(a) __simd_vrev64q_u16(a)
#define vrev64q_s16(a) __simd_vrev64q_s16(a)
#define vrev64q_p16(a) __simd_vrev64q_p16(a)
#define vrev64q_u32(a) __simd_vrev64q_u32(a)
#define vrev64q_s32(a) __simd_vrev64q_s32(a)
#define vrev64q_f32(a) __simd_vrev64q_f32(a)
#define vrev32_u8(a) __simd_vrev32_u8(a)
#define vrev32_s8(a) __simd_vrev32_s8(a)
#define vrev32_p8(a) __simd_vrev32_p8(a)
#define vrev32_u16(a) __simd_vrev32_u16(a)
#define vrev32_s16(a) __simd_vrev32_s16(a)
#define vrev32_p16(a) __simd_vrev32_p16(a)
#define vrev32q_u8(a) __simd_vrev32q_u8(a)
#define vrev32q_s8(a) __simd_vrev32q_s8(a)
#define vrev32q_p8(a) __simd_vrev32q_p8(a)
#define vrev32q_u16(a) __simd_vrev32q_u16(a)
#define vrev32q_s16(a) __simd_vrev32q_s16(a)
#define vrev32q_p16(a) __simd_vrev32q_p16(a)
#define vrev16_u8(a) __simd_vrev16_u8(a)
#define vrev16_s8(a) __simd_vrev16_s8(a)
#define vrev16_p8(a) __simd_vrev16_p8(a)
#define vrev16q_u8(a) __simd_vrev16q_u8(a)
#define vrev16q_s8(a) __simd_vrev16q_s8(a)
#define vrev16q_p8(a) __simd_vrev16q_p8(a)
#define vpaddl_s8(a) __simd_vpaddl_s8(a)
#define vpaddl_s16(a) __simd_vpaddl_s16(a)
#define vpaddl_s32(a) __simd_vpaddl_s32(a)
#define vpaddlq_s8(a) __simd_vpaddlq_s8(a)
#define vpaddlq_s16(a) __simd_vpaddlq_s16(a)
#define vpaddlq_s32(a) __simd_vpaddlq_s32(a)
#define vpaddl_u8(a) __simd_vpaddl_u8(a)
#define vpaddl_u16(a) __simd_vpaddl_u16(a)
#define vpaddl_u32(a) __simd_vpaddl_u32(a)
#define vpaddlq_u8(a) __simd_vpaddlq_u8(a)
#define vpaddlq_u16(a) __simd_vpaddlq_u16(a)
#define vpaddlq_u32(a) __simd_vpaddlq_u32(a)
#define vpadal_s8(a,b) __simd_vpadal_s8(a,b)
#define vpadal_s16(a,b) __simd_vpadal_s16(a,b)
#define vpadal_s32(a,b) __simd_vpadal_s32(a,b)
#define vpadalq_s8(a,b) __simd_vpadalq_s8(a,b)
#define vpadalq_s16(a,b) __simd_vpadalq_s16(a,b)
#define vpadalq_s32(a,b) __simd_vpadalq_s32(a,b)
#define vpadal_u8(a,b) __simd_vpadal_u8(a,b)
#define vpadal_u16(a,b) __simd_vpadal_u16(a,b)
#define vpadal_u32(a,b) __simd_vpadal_u32(a,b)
#define vpadalq_u8(a,b) __simd_vpadalq_u8(a,b)
#define vpadalq_u16(a,b) __simd_vpadalq_u16(a,b)
#define vpadalq_u32(a,b) __simd_vpadalq_u32(a,b)
#define vuqadd_s8(a,b) __simd_vuqadd_s8(a,b)
#define vuqadd_s16(a,b) __simd_vuqadd_s16(a,b)
#define vuqadd_s32(a,b) __simd_vuqadd_s32(a,b)
#define vuqaddq_s8(a,b) __simd_vuqaddq_s8(a,b)
#define vuqaddq_s16(a,b) __simd_vuqaddq_s16(a,b)
#define vuqaddq_s32(a,b) __simd_vuqaddq_s32(a,b)
#define vuqaddq_s64(a,b) __simd_vuqaddq_s64(a,b)
#define vsqadd_u8(a,b) __simd_vsqadd_u8(a,b)
#define vsqadd_u16(a,b) __simd_vsqadd_u16(a,b)
#define vsqadd_u32(a,b) __simd_vsqadd_u32(a,b)
#define vsqaddq_u8(a,b) __simd_vsqaddq_u8(a,b)
#define vsqaddq_u16(a,b) __simd_vsqaddq_u16(a,b)
#define vsqaddq_u32(a,b) __simd_vsqaddq_u32(a,b)
#define vsqaddq_u64(a,b) __simd_vsqaddq_u64(a,b)
#define vcls_s8(a) __simd_vcls_s8(a)
#define vcls_s16(a) __simd_vcls_s16(a)
#define vcls_s32(a) __simd_vcls_s32(a)
#define vclsq_s8(a) __simd_vclsq_s8(a)
#define vclsq_s16(a) __simd_vclsq_s16(a)
#define vclsq_s32(a) __simd_vclsq_s32(a)
#define vclz_u8(a) __simd_vclz_u8(a)
#define vclz_s8(a) __simd_vclz_s8(a)
#define vclz_u16(a) __simd_vclz_u16(a)
#define vclz_s16(a) __simd_vclz_s16(a)
#define vclz_u32(a) __simd_vclz_u32(a)
#define vclz_s32(a) __simd_vclz_s32(a)
#define vclzq_u8(a) __simd_vclzq_u8(a)
#define vclzq_s8(a) __simd_vclzq_s8(a)
#define vclzq_u16(a) __simd_vclzq_u16(a)
#define vclzq_s16(a) __simd_vclzq_s16(a)
#define vclzq_u32(a) __simd_vclzq_u32(a)
#define vclzq_s32(a) __simd_vclzq_s32(a)
#define vcnt_u8(a) __simd_vcnt_u8(a)
#define vcnt_s8(a) __simd_vcnt_s8(a)
#define vcnt_p8(a) __simd_vcnt_p8(a)
#define vcntq_u8(a) __simd_vcntq_u8(a)
#define vcntq_s8(a) __simd_vcntq_s8(a)
#define vcntq_p8(a) __simd_vcntq_p8(a)
#define vqabs_s8(a) __simd_vqabs_s8(a)
#define vqabs_s16(a) __simd_vqabs_s16(a)
#define vqabs_s32(a) __simd_vqabs_s32(a)
#define vqabsq_s8(a) __simd_vqabsq_s8(a)
#define vqabsq_s16(a) __simd_vqabsq_s16(a)
#define vqabsq_s32(a) __simd_vqabsq_s32(a)
#define vqabsq_s64(a) __simd_vqabsq_s64(a)
#define vqneg_s8(a) __simd_vqneg_s8(a)
#define vqneg_s16(a) __simd_vqneg_s16(a)
#define vqneg_s32(a) __simd_vqneg_s32(a)
#define vqnegq_s8(a) __simd_vqnegq_s8(a)
#define vqnegq_s16(a) __simd_vqnegq_s16(a)
#define vqnegq_s32(a) __simd_vqnegq_s32(a)
#define vqnegq_s64(a) __simd_vqnegq_s64(a)
#define vabs_s8(a) __simd_vabs_s8(a)
#define vabs_s16(a) __simd_vabs_s16(a)
#define vabs_s32(a) __simd_vabs_s32(a)
#define vabsq_s8(a) __simd_vabsq_s8(a)
#define vabsq_s16(a) __simd_vabsq_s16(a)
#define vabsq_s32(a) __simd_vabsq_s32(a)
#define vabsq_s64(a) __simd_vabsq_s64(a)
#define vneg_s8(a) __simd_vneg_s8(a)
#define vneg_s16(a) __simd_vneg_s16(a)
#define vneg_s32(a) __simd_vneg_s32(a)
#define vnegq_s8(a) __simd_vnegq_s8(a)
#define vnegq_s16(a) __simd_vnegq_s16(a)
#define vnegq_s32(a) __simd_vnegq_s32(a)
#define vnegq_s64(a) __simd_vnegq_s64(a)
#define vmovn_u16(a) __simd_vmovn_u16(a)
#define vmovn_s16(a) __simd_vmovn_s16(a)
#define vmovn_u32(a) __simd_vmovn_u32(a)
#define vmovn_s32(a) __simd_vmovn_s32(a)
#define vmovn_u64(a) __simd_vmovn_u64(a)
#define vmovn_s64(a) __simd_vmovn_s64(a)
#define vmovn_high_u16(a,b) __simd_vmovn_high_u16(a,b)
#define vmovn_high_s16(a,b) __simd_vmovn_high_s16(a,b)
#define vmovn_high_u32(a,b) __simd_vmovn_high_u32(a,b)
#define vmovn_high_s32(a,b) __simd_vmovn_high_s32(a,b)
#define vmovn_high_u64(a,b) __simd_vmovn_high_u64(a,b)
#define vmovn_high_s64(a,b) __simd_vmovn_high_s64(a,b)
#define vqmovn_s16(a) __simd_vqmovn_s16(a)
#define vqmovn_s32(a) __simd_vqmovn_s32(a)
#define vqmovn_s64(a) __simd_vqmovn_s64(a)
#define vqmovn_high_s16(a,b) __simd_vqmovn_high_s16(a,b)
#define vqmovn_high_s32(a,b) __simd_vqmovn_high_s32(a,b)
#define vqmovn_high_s64(a,b) __simd_vqmovn_high_s64(a,b)
#define vqmovn_u16(a) __simd_vqmovn_u16(a)
#define vqmovn_u32(a) __simd_vqmovn_u32(a)
#define vqmovn_u64(a) __simd_vqmovn_u64(a)
#define vqmovn_high_u16(a,b) __simd_vqmovn_high_u16(a,b)
#define vqmovn_high_u32(a,b) __simd_vqmovn_high_u32(a,b)
#define vqmovn_high_u64(a,b) __simd_vqmovn_high_u64(a,b)
#define vqmovun_s16(a) __simd_vqmovun_s16(a)
#define vqmovun_s32(a) __simd_vqmovun_s32(a)
#define vqmovun_s64(a) __simd_vqmovun_s64(a)
#define vqmovun_high_s16(a,b) __simd_vqmovun_high_s16(a,b)
#define vqmovun_high_s32(a,b) __simd_vqmovun_high_s32(a,b)
#define vqmovun_high_s64(a,b) __simd_vqmovun_high_s64(a,b)
#define vmvn_u8(a) __simd_vmvn_u8(a)
#define vmvn_u16(a) __simd_vmvn_u16(a)
#define vmvn_u32(a) __simd_vmvn_u32(a)
#define vmvn_u64(a) __simd_vmvn_u64(a)
#define vmvn_s8(a) __simd_vmvn_s8(a)
#define vmvn_s16(a) __simd_vmvn_s16(a)
#define vmvn_s32(a) __simd_vmvn_s32(a)
#define vmvn_s64(a) __simd_vmvn_s64(a)
#define vmvn_p8(a) __simd_vmvn_p8(a)
#define vmvn_p16(a) __simd_vmvn_p16(a)
#define vmvnq_u8(a) __simd_vmvnq_u8(a)
#define vmvnq_u16(a) __simd_vmvnq_u16(a)
#define vmvnq_u32(a) __simd_vmvnq_u32(a)
#define vmvnq_u64(a) __simd_vmvnq_u64(a)
#define vmvnq_s8(a) __simd_vmvnq_s8(a)
#define vmvnq_s16(a) __simd_vmvnq_s16(a)
#define vmvnq_s32(a) __simd_vmvnq_s32(a)
#define vmvnq_s64(a) __simd_vmvnq_s64(a)
#define vmvnq_p8(a) __simd_vmvnq_p8(a)
#define vmvnq_p16(a) __simd_vmvnq_p16(a)
#define vrbit_u8(a) __simd_vrbit_u8(a)
#define vrbit_s8(a) __simd_vrbit_s8(a)
#define vrbit_p8(a) __simd_vrbit_p8(a)
#define vrbitq_u8(a) __simd_vrbitq_u8(a)
#define vrbitq_s8(a) __simd_vrbitq_s8(a)
#define vrbitq_p8(a) __simd_vrbitq_p8(a)
#define vceqz_u8(a) __simd_vceqz_u8(a)
#define vceqz_s8(a) __simd_vceqz_s8(a)
#define vceqz_p8(a) __simd_vceqz_p8(a)
#define vceqz_u16(a) __simd_vceqz_u16(a)
#define vceqz_s16(a) __simd_vceqz_s16(a)
#define vceqz_p16(a) __simd_vceqz_p16(a)
#define vceqz_u32(a) __simd_vceqz_u32(a)
#define vceqz_s32(a) __simd_vceqz_s32(a)
#define vceqzq_u8(a) __simd_vceqzq_u8(a)
#define vceqzq_s8(a) __simd_vceqzq_s8(a)
#define vceqzq_p8(a) __simd_vceqzq_p8(a)
#define vceqzq_u16(a) __simd_vceqzq_u16(a)
#define vceqzq_s16(a) __simd_vceqzq_s16(a)
#define vceqzq_p16(a) __simd_vceqzq_p16(a)
#define vceqzq_u32(a) __simd_vceqzq_u32(a)
#define vceqzq_s32(a) __simd_vceqzq_s32(a)
#define vceqzq_u64(a) __simd_vceqzq_u64(a)
#define vceqzq_s64(a) __simd_vceqzq_s64(a)
#define vcgtz_s8(a) __simd_vcgtz_s8(a)
#define vcgtz_s16(a) __simd_vcgtz_s16(a)
#define vcgtz_s32(a) __simd_vcgtz_s32(a)
#define vcgtzq_s8(a) __simd_vcgtzq_s8(a)
#define vcgtzq_s16(a) __simd_vcgtzq_s16(a)
#define vcgtzq_s32(a) __simd_vcgtzq_s32(a)
#define vcgtzq_s64(a) __simd_vcgtzq_s64(a)
#define vcltz_s8(a) __simd_vcltz_s8(a)
#define vcltz_s16(a) __simd_vcltz_s16(a)
#define vcltz_s32(a) __simd_vcltz_s32(a)
#define vcltzq_s8(a) __simd_vcltzq_s8(a)
#define vcltzq_s16(a) __simd_vcltzq_s16(a)
#define vcltzq_s32(a) __simd_vcltzq_s32(a)
#define vcltzq_s64(a) __simd_vcltzq_s64(a)
#define vcgez_s8(a) __simd_vcgez_s8(a)
#define vcgez_s16(a) __simd_vcgez_s16(a)
#define vcgez_s32(a) __simd_vcgez_s32(a)
#define vcgezq_s8(a) __simd_vcgezq_s8(a)
#define vcgezq_s16(a) __simd_vcgezq_s16(a)
#define vcgezq_s32(a) __simd_vcgezq_s32(a)
#define vcgezq_s64(a) __simd_vcgezq_s64(a)
#define vclez_s8(a) __simd_vclez_s8(a)
#define vclez_s16(a) __simd_vclez_s16(a)
#define vclez_s32(a) __simd_vclez_s32(a)
#define vclezq_s8(a) __simd_vclezq_s8(a)
#define vclezq_s16(a) __simd_vclezq_s16(a)
#define vclezq_s32(a) __simd_vclezq_s32(a)
#define vclezq_s64(a) __simd_vclezq_s64(a)
#define vrsqrte_u32(a) __simd_vrsqrte_u32(a)
#define vrsqrteq_u32(a) __simd_vrsqrteq_u32(a)
#define vrecpe_u32(a) __simd_vrecpe_u32(a)
#define vrecpeq_u32(a) __simd_vrecpeq_u32(a)
#define vcvt_f32_f64(a) __simd_vcvt_f32_f64(a)
#define vcvt_high_f32_f64(a,b) __simd_vcvt_high_f32_f64(a,b)
#define vcvt_f64_f32(a) __simd_vcvt_f64_f32(a)
#define vcvt_high_f64_f32(a) __simd_vcvt_high_f64_f32(a)
#define vcvtx_f32_f64(a) __simd_vcvtx_f32_f64(a)
#define vcvtx_high_f32_f64(a,b) __simd_vcvtx_high_f32_f64(a,b)
#define vrndn_f32(a) __simd_vrndn_f32(a)
#define vrndnq_f32(a) __simd_vrndnq_f32(a)
#define vrndnq_f64(a) __simd_vrndnq_f64(a)
#define vrndm_f32(a) __simd_vrndm_f32(a)
#define vrndmq_f32(a) __simd_vrndmq_f32(a)
#define vrndmq_f64(a) __simd_vrndmq_f64(a)
#define vrndp_f32(a) __simd_vrndp_f32(a)
#define vrndpq_f32(a) __simd_vrndpq_f32(a)
#define vrndpq_f64(a) __simd_vrndpq_f64(a)
#define vrnd_f32(a) __simd_vrnd_f32(a)
#define vrndq_f32(a) __simd_vrndq_f32(a)
#define vrndq_f64(a) __simd_vrndq_f64(a)
#define vrnda_f32(a) __simd_vrnda_f32(a)
#define vrndaq_f32(a) __simd_vrndaq_f32(a)
#define vrndaq_f64(a) __simd_vrndaq_f64(a)
#define vrndx_f32(a) __simd_vrndx_f32(a)
#define vrndxq_f32(a) __simd_vrndxq_f32(a)
#define vrndxq_f64(a) __simd_vrndxq_f64(a)
#define vrndi_f32(a) __simd_vrndi_f32(a)
#define vrndiq_f32(a) __simd_vrndiq_f32(a)
#define vrndiq_f64(a) __simd_vrndiq_f64(a)
#define vcvtn_s32_f32(a) __simd_vcvtn_s32_f32(a)
#define vcvtnq_s32_f32(a) __simd_vcvtnq_s32_f32(a)
#define vcvtnq_s64_f64(a) __simd_vcvtnq_s64_f64(a)
#define vcvtn_u32_f32(a) __simd_vcvtn_u32_f32(a)
#define vcvtnq_u32_f32(a) __simd_vcvtnq_u32_f32(a)
#define vcvtnq_u64_f64(a) __simd_vcvtnq_u64_f64(a)
#define vcvtm_s32_f32(a) __simd_vcvtm_s32_f32(a)
#define vcvtmq_s32_f32(a) __simd_vcvtmq_s32_f32(a)
#define vcvtmq_s64_f64(a) __simd_vcvtmq_s64_f64(a)
#define vcvtm_u32_f32(a) __simd_vcvtm_u32_f32(a)
#define vcvtmq_u32_f32(a) __simd_vcvtmq_u32_f32(a)
#define vcvtmq_u64_f64(a) __simd_vcvtmq_u64_f64(a)
#define vcvta_s32_f32(a) __simd_vcvta_s32_f32(a)
#define vcvtaq_s32_f32(a) __simd_vcvtaq_s32_f32(a)
#define vcvtaq_s64_f64(a) __simd_vcvtaq_s64_f64(a)
#define vcvta_u32_f32(a) __simd_vcvta_u32_f32(a)
#define vcvtaq_u32_f32(a) __simd_vcvtaq_u32_f32(a)
#define vcvtaq_u64_f64(a) __simd_vcvtaq_u64_f64(a)
#define vcvtp_s32_f32(a) __simd_vcvtp_s32_f32(a)
#define vcvtpq_s32_f32(a) __simd_vcvtpq_s32_f32(a)
#define vcvtpq_s64_f64(a) __simd_vcvtpq_s64_f64(a)
#define vcvtp_u32_f32(a) __simd_vcvtp_u32_f32(a)
#define vcvtpq_u32_f32(a) __simd_vcvtpq_u32_f32(a)
#define vcvtpq_u64_f64(a) __simd_vcvtpq_u64_f64(a)
#define vcvt_s32_f32(a) __simd_vcvt_s32_f32(a)
#define vcvtq_s32_f32(a) __simd_vcvtq_s32_f32(a)
#define vcvtq_s64_f64(a) __simd_vcvtq_s64_f64(a)
#define vcvt_u32_f32(a) __simd_vcvt_u32_f32(a)
#define vcvtq_u32_f32(a) __simd_vcvtq_u32_f32(a)
#define vcvtq_u64_f64(a) __simd_vcvtq_u64_f64(a)
#define vcvt_f32_s32(a) __simd_vcvt_f32_s32(a)
#define vcvtq_f32_s32(a) __simd_vcvtq_f32_s32(a)
#define vcvtq_f64_s64(a) __simd_vcvtq_f64_s64(a)
#define vcvt_f32_u32(a) __simd_vcvt_f32_u32(a)
#define vcvtq_f32_u32(a) __simd_vcvtq_f32_u32(a)
#define vcvtq_f64_u64(a) __simd_vcvtq_f64_u64(a)
#define vceqz_f32(a) __simd_vceqz_f32(a)
#define vceqzq_f32(a) __simd_vceqzq_f32(a)
#define vceqzq_f64(a) __simd_vceqzq_f64(a)
#define vcgtz_f32(a) __simd_vcgtz_f32(a)
#define vcgtzq_f32(a) __simd_vcgtzq_f32(a)
#define vcgtzq_f64(a) __simd_vcgtzq_f64(a)
#define vcltz_f32(a) __simd_vcltz_f32(a)
#define vcltzq_f32(a) __simd_vcltzq_f32(a)
#define vcltzq_f64(a) __simd_vcltzq_f64(a)
#define vcgez_f32(a) __simd_vcgez_f32(a)
#define vcgezq_f32(a) __simd_vcgezq_f32(a)
#define vcgezq_f64(a) __simd_vcgezq_f64(a)
#define vclez_f32(a) __simd_vclez_f32(a)
#define vclezq_f32(a) __simd_vclezq_f32(a)
#define vclezq_f64(a) __simd_vclezq_f64(a)
#define vabs_f32(a) __simd_vabs_f32(a)
#define vabsq_f32(a) __simd_vabsq_f32(a)
#define vabsq_f64(a) __simd_vabsq_f64(a)
#define vneg_f32(a) __simd_vneg_f32(a)
#define vnegq_f32(a) __simd_vnegq_f32(a)
#define vnegq_f64(a) __simd_vnegq_f64(a)
#define vrecpe_f32(a) __simd_vrecpe_f32(a)
#define vrecpeq_f32(a) __simd_vrecpeq_f32(a)
#define vrecpeq_f64(a) __simd_vrecpeq_f64(a)
#define vrsqrte_f32(a) __simd_vrsqrte_f32(a)
#define vrsqrteq_f32(a) __simd_vrsqrteq_f32(a)
#define vrsqrteq_f64(a) __simd_vrsqrteq_f64(a)
#define vsqrt_f32(a) __simd_vsqrt_f32(a)
#define vsqrtq_f32(a) __simd_vsqrtq_f32(a)
#define vsqrtq_f64(a) __simd_vsqrtq_f64(a)
#define vuqaddb_s8(a,b) __simd_vuqaddb_s8(a,b)
#define vuqaddh_s16(a,b) __simd_vuqaddh_s16(a,b)
#define vuqadds_s32(a,b) __simd_vuqadds_s32(a,b)
#define vuqadd_s64(a,b) __simd_vuqadd_s64(a,b)
#define vuqaddd_s64(a,b) __simd_vuqaddd_s64(a,b)
#define vsqaddb_u8(a,b) __simd_vsqaddb_u8(a,b)
#define vsqaddh_u16(a,b) __simd_vsqaddh_u16(a,b)
#define vsqadds_u32(a,b) __simd_vsqadds_u32(a,b)
#define vsqadd_u64(a,b) __simd_vsqadd_u64(a,b)
#define vsqaddd_u64(a,b) __simd_vsqaddd_u64(a,b)
#define vqabsb_s8(a) __simd_vqabsb_s8(a)
#define vqabsh_s16(a) __simd_vqabsh_s16(a)
#define vqabss_s32(a) __simd_vqabss_s32(a)
#define vqabs_s64(a) __simd_vqabs_s64(a)
#define vqabsd_s64(a) __simd_vqabsd_s64(a)
#define vqnegb_s8(a) __simd_vqnegb_s8(a)
#define vqnegh_s16(a) __simd_vqnegh_s16(a)
#define vqnegs_s32(a) __simd_vqnegs_s32(a)
#define vqneg_s64(a) __simd_vqneg_s64(a)
#define vqnegd_s64(a) __simd_vqnegd_s64(a)
#define vabs_s64(a) __simd_vabs_s64(a)
#define vabsd_s64(a) __simd_vabsd_s64(a)
#define vneg_s64(a) __simd_vneg_s64(a)
#define vnegd_s64(a) __simd_vnegd_s64(a)
#define vceqz_u64(a) __simd_vceqz_u64(a)
#define vceqz_s64(a) __simd_vceqz_s64(a)
#define vceqzd_u64(a) __simd_vceqzd_u64(a)
#define vceqzd_s64(a) __simd_vceqzd_s64(a)
#define vcgtz_s64(a) __simd_vcgtz_s64(a)
#define vcgtzd_s64(a) __simd_vcgtzd_s64(a)
#define vcltz_s64(a) __simd_vcltz_s64(a)
#define vcltzd_s64(a) __simd_vcltzd_s64(a)
#define vcgez_s64(a) __simd_vcgez_s64(a)
#define vcgezd_s64(a) __simd_vcgezd_s64(a)
#define vclez_s64(a) __simd_vclez_s64(a)
#define vclezd_s64(a) __simd_vclezd_s64(a)
#define vqmovnh_s16(a) __simd_vqmovnh_s16(a)
#define vqmovns_s32(a) __simd_vqmovns_s32(a)
#define vqmovnd_s64(a) __simd_vqmovnd_s64(a)
#define vqmovnh_u16(a) __simd_vqmovnh_u16(a)
#define vqmovns_u32(a) __simd_vqmovns_u32(a)
#define vqmovnd_u64(a) __simd_vqmovnd_u64(a)
#define vqmovunh_s16(a) __simd_vqmovunh_s16(a)
#define vqmovuns_s32(a) __simd_vqmovuns_s32(a)
#define vqmovund_s64(a) __simd_vqmovund_s64(a)
#define vcvtns_s32_f32(a) __simd_vcvtns_s32_f32(a)
#define vcvtnd_s64_f64(a) __simd_vcvtnd_s64_f64(a)
#define vcvtns_u32_f32(a) __simd_vcvtns_u32_f32(a)
#define vcvtnd_u64_f64(a) __simd_vcvtnd_u64_f64(a)
#define vcvtms_s32_f32(a) __simd_vcvtms_s32_f32(a)
#define vcvtmd_s64_f64(a) __simd_vcvtmd_s64_f64(a)
#define vcvtms_u32_f32(a) __simd_vcvtms_u32_f32(a)
#define vcvtmd_u64_f64(a) __simd_vcvtmd_u64_f64(a)
#define vcvtas_s32_f32(a) __simd_vcvtas_s32_f32(a)
#define vcvtad_s64_f64(a) __simd_vcvtad_s64_f64(a)
#define vcvtas_u32_f32(a) __simd_vcvtas_u32_f32(a)
#define vcvtad_u64_f64(a) __simd_vcvtad_u64_f64(a)
#define vcvtps_s32_f32(a) __simd_vcvtps_s32_f32(a)
#define vcvtpd_s64_f64(a) __simd_vcvtpd_s64_f64(a)
#define vcvtps_u32_f32(a) __simd_vcvtps_u32_f32(a)
#define vcvtpd_u64_f64(a) __simd_vcvtpd_u64_f64(a)
#define vcvts_s32_f32(a) __simd_vcvts_s32_f32(a)
#define vcvtd_s64_f64(a) __simd_vcvtd_s64_f64(a)
#define vcvts_u32_f32(a) __simd_vcvts_u32_f32(a)
#define vcvtd_u64_f64(a) __simd_vcvtd_u64_f64(a)
#define vcvts_f32_s32(a) __simd_vcvts_f32_s32(a)
#define vcvtd_f64_s64(a) __simd_vcvtd_f64_s64(a)
#define vcvts_f32_u32(a) __simd_vcvts_f32_u32(a)
#define vcvtd_f64_u64(a) __simd_vcvtd_f64_u64(a)
#define vcvtxd_f32_f64(a) __simd_vcvtxd_f32_f64(a)
#define vrecpes_f32(a) __simd_vrecpes_f32(a)
#define vrecped_f64(a) __simd_vrecped_f64(a)
#define vrecpxs_f32(a) __simd_vrecpxs_f32(a)
#define vrecpxd_f64(a) __simd_vrecpxd_f64(a)
#define vrsqrtes_f32(a) __simd_vrsqrtes_f32(a)
#define vrsqrted_f64(a) __simd_vrsqrted_f64(a)
#define vceqzs_f32(a) __simd_vceqzs_f32(a)
#define vceqzd_f64(a) __simd_vceqzd_f64(a)
#define vcgtzs_f32(a) __simd_vcgtzs_f32(a)
#define vcgtzd_f64(a) __simd_vcgtzd_f64(a)
#define vcltzs_f32(a) __simd_vcltzs_f32(a)
#define vcltzd_f64(a) __simd_vcltzd_f64(a)
#define vcgezs_f32(a) __simd_vcgezs_f32(a)
#define vcgezd_f64(a) __simd_vcgezd_f64(a)
#define vclezs_f32(a) __simd_vclezs_f32(a)
#define vclezd_f64(a) __simd_vclezd_f64(a)
#define vshr_n_s8(a,b) __simd_vshr_n_s8(a,b)
#define vshr_n_s16(a,b) __simd_vshr_n_s16(a,b)
#define vshr_n_s32(a,b) __simd_vshr_n_s32(a,b)
#define vshrq_n_s8(a,b) __simd_vshrq_n_s8(a,b)
#define vshrq_n_s16(a,b) __simd_vshrq_n_s16(a,b)
#define vshrq_n_s32(a,b) __simd_vshrq_n_s32(a,b)
#define vshrq_n_s64(a,b) __simd_vshrq_n_s64(a,b)
#define vshr_n_u8(a,b) __simd_vshr_n_u8(a,b)
#define vshr_n_u16(a,b) __simd_vshr_n_u16(a,b)
#define vshr_n_u32(a,b) __simd_vshr_n_u32(a,b)
#define vshrq_n_u8(a,b) __simd_vshrq_n_u8(a,b)
#define vshrq_n_u16(a,b) __simd_vshrq_n_u16(a,b)
#define vshrq_n_u32(a,b) __simd_vshrq_n_u32(a,b)
#define vshrq_n_u64(a,b) __simd_vshrq_n_u64(a,b)
#define vrshr_n_s8(a,b) __simd_vrshr_n_s8(a,b)
#define vrshr_n_s16(a,b) __simd_vrshr_n_s16(a,b)
#define vrshr_n_s32(a,b) __simd_vrshr_n_s32(a,b)
#define vrshrq_n_s8(a,b) __simd_vrshrq_n_s8(a,b)
#define vrshrq_n_s16(a,b) __simd_vrshrq_n_s16(a,b)
#define vrshrq_n_s32(a,b) __simd_vrshrq_n_s32(a,b)
#define vrshrq_n_s64(a,b) __simd_vrshrq_n_s64(a,b)
#define vrshr_n_u8(a,b) __simd_vrshr_n_u8(a,b)
#define vrshr_n_u16(a,b) __simd_vrshr_n_u16(a,b)
#define vrshr_n_u32(a,b) __simd_vrshr_n_u32(a,b)
#define vrshrq_n_u8(a,b) __simd_vrshrq_n_u8(a,b)
#define vrshrq_n_u16(a,b) __simd_vrshrq_n_u16(a,b)
#define vrshrq_n_u32(a,b) __simd_vrshrq_n_u32(a,b)
#define vrshrq_n_u64(a,b) __simd_vrshrq_n_u64(a,b)
#define vsra_n_s8(a,b,c) __simd_vsra_n_s8(a,b,c)
#define vsra_n_s16(a,b,c) __simd_vsra_n_s16(a,b,c)
#define vsra_n_s32(a,b,c) __simd_vsra_n_s32(a,b,c)
#define vsraq_n_s8(a,b,c) __simd_vsraq_n_s8(a,b,c)
#define vsraq_n_s16(a,b,c) __simd_vsraq_n_s16(a,b,c)
#define vsraq_n_s32(a,b,c) __simd_vsraq_n_s32(a,b,c)
#define vsraq_n_s64(a,b,c) __simd_vsraq_n_s64(a,b,c)
#define vsra_n_u8(a,b,c) __simd_vsra_n_u8(a,b,c)
#define vsra_n_u16(a,b,c) __simd_vsra_n_u16(a,b,c)
#define vsra_n_u32(a,b,c) __simd_vsra_n_u32(a,b,c)
#define vsraq_n_u8(a,b,c) __simd_vsraq_n_u8(a,b,c)
#define vsraq_n_u16(a,b,c) __simd_vsraq_n_u16(a,b,c)
#define vsraq_n_u32(a,b,c) __simd_vsraq_n_u32(a,b,c)
#define vsraq_n_u64(a,b,c) __simd_vsraq_n_u64(a,b,c)
#define vrsra_n_s8(a,b,c) __simd_vrsra_n_s8(a,b,c)
#define vrsra_n_s16(a,b,c) __simd_vrsra_n_s16(a,b,c)
#define vrsra_n_s32(a,b,c) __simd_vrsra_n_s32(a,b,c)
#define vrsraq_n_s8(a,b,c) __simd_vrsraq_n_s8(a,b,c)
#define vrsraq_n_s16(a,b,c) __simd_vrsraq_n_s16(a,b,c)
#define vrsraq_n_s32(a,b,c) __simd_vrsraq_n_s32(a,b,c)
#define vrsraq_n_s64(a,b,c) __simd_vrsraq_n_s64(a,b,c)
#define vrsra_n_u8(a,b,c) __simd_vrsra_n_u8(a,b,c)
#define vrsra_n_u16(a,b,c) __simd_vrsra_n_u16(a,b,c)
#define vrsra_n_u32(a,b,c) __simd_vrsra_n_u32(a,b,c)
#define vrsraq_n_u8(a,b,c) __simd_vrsraq_n_u8(a,b,c)
#define vrsraq_n_u16(a,b,c) __simd_vrsraq_n_u16(a,b,c)
#define vrsraq_n_u32(a,b,c) __simd_vrsraq_n_u32(a,b,c)
#define vrsraq_n_u64(a,b,c) __simd_vrsraq_n_u64(a,b,c)
#define vshl_n_u8(a,b) __simd_vshl_n_u8(a,b)
#define vshl_n_s8(a,b) __simd_vshl_n_s8(a,b)
#define vshl_n_u16(a,b) __simd_vshl_n_u16(a,b)
#define vshl_n_s16(a,b) __simd_vshl_n_s16(a,b)
#define vshl_n_u32(a,b) __simd_vshl_n_u32(a,b)
#define vshl_n_s32(a,b) __simd_vshl_n_s32(a,b)
#define vshlq_n_u8(a,b) __simd_vshlq_n_u8(a,b)
#define vshlq_n_s8(a,b) __simd_vshlq_n_s8(a,b)
#define vshlq_n_u16(a,b) __simd_vshlq_n_u16(a,b)
#define vshlq_n_s16(a,b) __simd_vshlq_n_s16(a,b)
#define vshlq_n_u32(a,b) __simd_vshlq_n_u32(a,b)
#define vshlq_n_s32(a,b) __simd_vshlq_n_s32(a,b)
#define vshlq_n_u64(a,b) __simd_vshlq_n_u64(a,b)
#define vshlq_n_s64(a,b) __simd_vshlq_n_s64(a,b)
#define vqshl_n_s8(a,b) __simd_vqshl_n_s8(a,b)
#define vqshl_n_s16(a,b) __simd_vqshl_n_s16(a,b)
#define vqshl_n_s32(a,b) __simd_vqshl_n_s32(a,b)
#define vqshlq_n_s8(a,b) __simd_vqshlq_n_s8(a,b)
#define vqshlq_n_s16(a,b) __simd_vqshlq_n_s16(a,b)
#define vqshlq_n_s32(a,b) __simd_vqshlq_n_s32(a,b)
#define vqshlq_n_s64(a,b) __simd_vqshlq_n_s64(a,b)
#define vqshl_n_u8(a,b) __simd_vqshl_n_u8(a,b)
#define vqshl_n_u16(a,b) __simd_vqshl_n_u16(a,b)
#define vqshl_n_u32(a,b) __simd_vqshl_n_u32(a,b)
#define vqshlq_n_u8(a,b) __simd_vqshlq_n_u8(a,b)
#define vqshlq_n_u16(a,b) __simd_vqshlq_n_u16(a,b)
#define vqshlq_n_u32(a,b) __simd_vqshlq_n_u32(a,b)
#define vqshlq_n_u64(a,b) __simd_vqshlq_n_u64(a,b)
#define vqshlu_n_s8(a,b) __simd_vqshlu_n_s8(a,b)
#define vqshlu_n_s16(a,b) __simd_vqshlu_n_s16(a,b)
#define vqshlu_n_s32(a,b) __simd_vqshlu_n_s32(a,b)
#define vqshluq_n_s8(a,b) __simd_vqshluq_n_s8(a,b)
#define vqshluq_n_s16(a,b) __simd_vqshluq_n_s16(a,b)
#define vqshluq_n_s32(a,b) __simd_vqshluq_n_s32(a,b)
#define vqshluq_n_s64(a,b) __simd_vqshluq_n_s64(a,b)
#define vsri_n_u8(a,b,c) __simd_vsri_n_u8(a,b,c)
#define vsri_n_s8(a,b,c) __simd_vsri_n_s8(a,b,c)
#define vsri_n_p8(a,b,c) __simd_vsri_n_p8(a,b,c)
#define vsri_n_u16(a,b,c) __simd_vsri_n_u16(a,b,c)
#define vsri_n_s16(a,b,c) __simd_vsri_n_s16(a,b,c)
#define vsri_n_p16(a,b,c) __simd_vsri_n_p16(a,b,c)
#define vsri_n_u32(a,b,c) __simd_vsri_n_u32(a,b,c)
#define vsri_n_s32(a,b,c) __simd_vsri_n_s32(a,b,c)
#define vsriq_n_u8(a,b,c) __simd_vsriq_n_u8(a,b,c)
#define vsriq_n_s8(a,b,c) __simd_vsriq_n_s8(a,b,c)
#define vsriq_n_p8(a,b,c) __simd_vsriq_n_p8(a,b,c)
#define vsriq_n_u16(a,b,c) __simd_vsriq_n_u16(a,b,c)
#define vsriq_n_s16(a,b,c) __simd_vsriq_n_s16(a,b,c)
#define vsriq_n_p16(a,b,c) __simd_vsriq_n_p16(a,b,c)
#define vsriq_n_u32(a,b,c) __simd_vsriq_n_u32(a,b,c)
#define vsriq_n_s32(a,b,c) __simd_vsriq_n_s32(a,b,c)
#define vsriq_n_u64(a,b,c) __simd_vsriq_n_u64(a,b,c)
#define vsriq_n_s64(a,b,c) __simd_vsriq_n_s64(a,b,c)
#define vsli_n_u8(a,b,c) __simd_vsli_n_u8(a,b,c)
#define vsli_n_s8(a,b,c) __simd_vsli_n_s8(a,b,c)
#define vsli_n_p8(a,b,c) __simd_vsli_n_p8(a,b,c)
#define vsli_n_u16(a,b,c) __simd_vsli_n_u16(a,b,c)
#define vsli_n_s16(a,b,c) __simd_vsli_n_s16(a,b,c)
#define vsli_n_p16(a,b,c) __simd_vsli_n_p16(a,b,c)
#define vsli_n_u32(a,b,c) __simd_vsli_n_u32(a,b,c)
#define vsli_n_s32(a,b,c) __simd_vsli_n_s32(a,b,c)
#define vsliq_n_u8(a,b,c) __simd_vsliq_n_u8(a,b,c)
#define vsliq_n_s8(a,b,c) __simd_vsliq_n_s8(a,b,c)
#define vsliq_n_p8(a,b,c) __simd_vsliq_n_p8(a,b,c)
#define vsliq_n_u16(a,b,c) __simd_vsliq_n_u16(a,b,c)
#define vsliq_n_s16(a,b,c) __simd_vsliq_n_s16(a,b,c)
#define vsliq_n_p16(a,b,c) __simd_vsliq_n_p16(a,b,c)
#define vsliq_n_u32(a,b,c) __simd_vsliq_n_u32(a,b,c)
#define vsliq_n_s32(a,b,c) __simd_vsliq_n_s32(a,b,c)
#define vsliq_n_u64(a,b,c) __simd_vsliq_n_u64(a,b,c)
#define vsliq_n_s64(a,b,c) __simd_vsliq_n_s64(a,b,c)
#define vshrn_n_u16(a,b) __simd_vshrn_n_u16(a,b)
#define vshrn_n_s16(a,b) __simd_vshrn_n_s16(a,b)
#define vshrn_n_u32(a,b) __simd_vshrn_n_u32(a,b)
#define vshrn_n_s32(a,b) __simd_vshrn_n_s32(a,b)
#define vshrn_n_u64(a,b) __simd_vshrn_n_u64(a,b)
#define vshrn_n_s64(a,b) __simd_vshrn_n_s64(a,b)
#define vshrn_high_n_u16(a,b,c) __simd_vshrn_high_n_u16(a,b,c)
#define vshrn_high_n_s16(a,b,c) __simd_vshrn_high_n_s16(a,b,c)
#define vshrn_high_n_u32(a,b,c) __simd_vshrn_high_n_u32(a,b,c)
#define vshrn_high_n_s32(a,b,c) __simd_vshrn_high_n_s32(a,b,c)
#define vshrn_high_n_u64(a,b,c) __simd_vshrn_high_n_u64(a,b,c)
#define vshrn_high_n_s64(a,b,c) __simd_vshrn_high_n_s64(a,b,c)
#define vrshrn_n_u16(a,b) __simd_vrshrn_n_u16(a,b)
#define vrshrn_n_s16(a,b) __simd_vrshrn_n_s16(a,b)
#define vrshrn_n_u32(a,b) __simd_vrshrn_n_u32(a,b)
#define vrshrn_n_s32(a,b) __simd_vrshrn_n_s32(a,b)
#define vrshrn_n_u64(a,b) __simd_vrshrn_n_u64(a,b)
#define vrshrn_n_s64(a,b) __simd_vrshrn_n_s64(a,b)
#define vrshrn_high_n_u16(a,b,c) __simd_vrshrn_high_n_u16(a,b,c)
#define vrshrn_high_n_s16(a,b,c) __simd_vrshrn_high_n_s16(a,b,c)
#define vrshrn_high_n_u32(a,b,c) __simd_vrshrn_high_n_u32(a,b,c)
#define vrshrn_high_n_s32(a,b,c) __simd_vrshrn_high_n_s32(a,b,c)
#define vrshrn_high_n_u64(a,b,c) __simd_vrshrn_high_n_u64(a,b,c)
#define vrshrn_high_n_s64(a,b,c) __simd_vrshrn_high_n_s64(a,b,c)
#define vqshrn_n_s16(a,b) __simd_vqshrn_n_s16(a,b)
#define vqshrn_n_s32(a,b) __simd_vqshrn_n_s32(a,b)
#define vqshrn_n_s64(a,b) __simd_vqshrn_n_s64(a,b)
#define vqshrn_high_n_s16(a,b,c) __simd_vqshrn_high_n_s16(a,b,c)
#define vqshrn_high_n_s32(a,b,c) __simd_vqshrn_high_n_s32(a,b,c)
#define vqshrn_high_n_s64(a,b,c) __simd_vqshrn_high_n_s64(a,b,c)
#define vqshrn_n_u16(a,b) __simd_vqshrn_n_u16(a,b)
#define vqshrn_n_u32(a,b) __simd_vqshrn_n_u32(a,b)
#define vqshrn_n_u64(a,b) __simd_vqshrn_n_u64(a,b)
#define vqshrn_high_n_u16(a,b,c) __simd_vqshrn_high_n_u16(a,b,c)
#define vqshrn_high_n_u32(a,b,c) __simd_vqshrn_high_n_u32(a,b,c)
#define vqshrn_high_n_u64(a,b,c) __simd_vqshrn_high_n_u64(a,b,c)
#define vqrshrn_n_s16(a,b) __simd_vqrshrn_n_s16(a,b)
#define vqrshrn_n_s32(a,b) __simd_vqrshrn_n_s32(a,b)
#define vqrshrn_n_s64(a,b) __simd_vqrshrn_n_s64(a,b)
#define vqrshrn_high_n_s16(a,b,c) __simd_vqrshrn_high_n_s16(a,b,c)
#define vqrshrn_high_n_s32(a,b,c) __simd_vqrshrn_high_n_s32(a,b,c)
#define vqrshrn_high_n_s64(a,b,c) __simd_vqrshrn_high_n_s64(a,b,c)
#define vqrshrn_n_u16(a,b) __simd_vqrshrn_n_u16(a,b)
#define vqrshrn_n_u32(a,b) __simd_vqrshrn_n_u32(a,b)
#define vqrshrn_n_u64(a,b) __simd_vqrshrn_n_u64(a,b)
#define vqrshrn_high_n_u16(a,b,c) __simd_vqrshrn_high_n_u16(a,b,c)
#define vqrshrn_high_n_u32(a,b,c) __simd_vqrshrn_high_n_u32(a,b,c)
#define vqrshrn_high_n_u64(a,b,c) __simd_vqrshrn_high_n_u64(a,b,c)
#define vqshrun_n_s16(a,b) __simd_vqshrun_n_s16(a,b)
#define vqshrun_n_s32(a,b) __simd_vqshrun_n_s32(a,b)
#define vqshrun_n_s64(a,b) __simd_vqshrun_n_s64(a,b)
#define vqshrun_high_n_s16(a,b,c) __simd_vqshrun_high_n_s16(a,b,c)
#define vqshrun_high_n_s32(a,b,c) __simd_vqshrun_high_n_s32(a,b,c)
#define vqshrun_high_n_s64(a,b,c) __simd_vqshrun_high_n_s64(a,b,c)
#define vqrshrun_n_s16(a,b) __simd_vqrshrun_n_s16(a,b)
#define vqrshrun_n_s32(a,b) __simd_vqrshrun_n_s32(a,b)
#define vqrshrun_n_s64(a,b) __simd_vqrshrun_n_s64(a,b)
#define vqrshrun_high_n_s16(a,b,c) __simd_vqrshrun_high_n_s16(a,b,c)
#define vqrshrun_high_n_s32(a,b,c) __simd_vqrshrun_high_n_s32(a,b,c)
#define vqrshrun_high_n_s64(a,b,c) __simd_vqrshrun_high_n_s64(a,b,c)
#define vmovl_s8(a) __simd_vmovl_s8(a)
#define vmovl_s16(a) __simd_vmovl_s16(a)
#define vmovl_s32(a) __simd_vmovl_s32(a)
#define vmovl_high_s8(a) __simd_vmovl_high_s8(a)
#define vmovl_high_s16(a) __simd_vmovl_high_s16(a)
#define vmovl_high_s32(a) __simd_vmovl_high_s32(a)
#define vmovl_u8(a) __simd_vmovl_u8(a)
#define vmovl_u16(a) __simd_vmovl_u16(a)
#define vmovl_u32(a) __simd_vmovl_u32(a)
#define vmovl_high_u8(a) __simd_vmovl_high_u8(a)
#define vmovl_high_u16(a) __simd_vmovl_high_u16(a)
#define vmovl_high_u32(a) __simd_vmovl_high_u32(a)
#define vshll_n_s8(a,b) __simd_vshll_n_s8(a,b)
#define vshll_n_s16(a,b) __simd_vshll_n_s16(a,b)
#define vshll_n_s32(a,b) __simd_vshll_n_s32(a,b)
#define vshll_high_n_s8(a,b) __simd_vshll_high_n_s8(a,b)
#define vshll_high_n_s16(a,b) __simd_vshll_high_n_s16(a,b)
#define vshll_high_n_s32(a,b) __simd_vshll_high_n_s32(a,b)
#define vshll_n_u8(a,b) __simd_vshll_n_u8(a,b)
#define vshll_n_u16(a,b) __simd_vshll_n_u16(a,b)
#define vshll_n_u32(a,b) __simd_vshll_n_u32(a,b)
#define vshll_high_n_u8(a,b) __simd_vshll_high_n_u8(a,b)
#define vshll_high_n_u16(a,b) __simd_vshll_high_n_u16(a,b)
#define vshll_high_n_u32(a,b) __simd_vshll_high_n_u32(a,b)
#define vcvt_n_f32_s32(a,b) __simd_vcvt_n_f32_s32(a,b)
#define vcvtq_n_f32_s32(a,b) __simd_vcvtq_n_f32_s32(a,b)
#define vcvtq_n_f64_s64(a,b) __simd_vcvtq_n_f64_s64(a,b)
#define vcvt_n_f32_u32(a,b) __simd_vcvt_n_f32_u32(a,b)
#define vcvtq_n_f32_u32(a,b) __simd_vcvtq_n_f32_u32(a,b)
#define vcvtq_n_f64_u64(a,b) __simd_vcvtq_n_f64_u64(a,b)
#define vcvt_n_s32_f32(a,b) __simd_vcvt_n_s32_f32(a,b)
#define vcvtq_n_s32_f32(a,b) __simd_vcvtq_n_s32_f32(a,b)
#define vcvtq_n_s64_f64(a,b) __simd_vcvtq_n_s64_f64(a,b)
#define vcvt_n_u32_f32(a,b) __simd_vcvt_n_u32_f32(a,b)
#define vcvtq_n_u32_f32(a,b) __simd_vcvtq_n_u32_f32(a,b)
#define vcvtq_n_u64_f64(a,b) __simd_vcvtq_n_u64_f64(a,b)
#define vshr_n_s64(a,b) __simd_vshr_n_s64(a,b)
#define vshrd_n_s64(a,b) __simd_vshrd_n_s64(a,b)
#define vshr_n_u64(a,b) __simd_vshr_n_u64(a,b)
#define vshrd_n_u64(a,b) __simd_vshrd_n_u64(a,b)
#define vsra_n_s64(a,b,c) __simd_vsra_n_s64(a,b,c)
#define vsrad_n_s64(a,b,c) __simd_vsrad_n_s64(a,b,c)
#define vsra_n_u64(a,b,c) __simd_vsra_n_u64(a,b,c)
#define vsrad_n_u64(a,b,c) __simd_vsrad_n_u64(a,b,c)
#define vrshr_n_s64(a,b) __simd_vrshr_n_s64(a,b)
#define vrshrd_n_s64(a,b) __simd_vrshrd_n_s64(a,b)
#define vrshr_n_u64(a,b) __simd_vrshr_n_u64(a,b)
#define vrshrd_n_u64(a,b) __simd_vrshrd_n_u64(a,b)
#define vrsra_n_s64(a,b,c) __simd_vrsra_n_s64(a,b,c)
#define vrsrad_n_s64(a,b,c) __simd_vrsrad_n_s64(a,b,c)
#define vrsra_n_u64(a,b,c) __simd_vrsra_n_u64(a,b,c)
#define vrsrad_n_u64(a,b,c) __simd_vrsrad_n_u64(a,b,c)
#define vshl_n_u64(a,b) __simd_vshl_n_u64(a,b)
#define vshl_n_s64(a,b) __simd_vshl_n_s64(a,b)
#define vshld_n_u64(a,b) __simd_vshld_n_u64(a,b)
#define vshld_n_s64(a,b) __simd_vshld_n_s64(a,b)
#define vqshlb_n_s8(a,b) __simd_vqshlb_n_s8(a,b)
#define vqshlh_n_s16(a,b) __simd_vqshlh_n_s16(a,b)
#define vqshls_n_s32(a,b) __simd_vqshls_n_s32(a,b)
#define vqshl_n_s64(a,b) __simd_vqshl_n_s64(a,b)
#define vqshld_n_s64(a,b) __simd_vqshld_n_s64(a,b)
#define vqshlb_n_u8(a,b) __simd_vqshlb_n_u8(a,b)
#define vqshlh_n_u16(a,b) __simd_vqshlh_n_u16(a,b)
#define vqshls_n_u32(a,b) __simd_vqshls_n_u32(a,b)
#define vqshl_n_u64(a,b) __simd_vqshl_n_u64(a,b)
#define vqshld_n_u64(a,b) __simd_vqshld_n_u64(a,b)
#define vqshlub_n_s8(a,b) __simd_vqshlub_n_s8(a,b)
#define vqshluh_n_s16(a,b) __simd_vqshluh_n_s16(a,b)
#define vqshlus_n_s32(a,b) __simd_vqshlus_n_s32(a,b)
#define vqshlu_n_s64(a,b) __simd_vqshlu_n_s64(a,b)
#define vqshlud_n_s64(a,b) __simd_vqshlud_n_s64(a,b)
#define vsri_n_u64(a,b,c) __simd_vsri_n_u64(a,b,c)
#define vsri_n_s64(a,b,c) __simd_vsri_n_s64(a,b,c)
#define vsrid_n_u64(a,b,c) __simd_vsrid_n_u64(a,b,c)
#define vsrid_n_s64(a,b,c) __simd_vsrid_n_s64(a,b,c)
#define vsli_n_u64(a,b,c) __simd_vsli_n_u64(a,b,c)
#define vsli_n_s64(a,b,c) __simd_vsli_n_s64(a,b,c)
#define vslid_n_u64(a,b,c) __simd_vslid_n_u64(a,b,c)
#define vslid_n_s64(a,b,c) __simd_vslid_n_s64(a,b,c)
#define vqshrnh_n_s16(a,b) __simd_vqshrnh_n_s16(a,b)
#define vqshrns_n_s32(a,b) __simd_vqshrns_n_s32(a,b)
#define vqshrnd_n_s64(a,b) __simd_vqshrnd_n_s64(a,b)
#define vqshrnh_n_u16(a,b) __simd_vqshrnh_n_u16(a,b)
#define vqshrns_n_u32(a,b) __simd_vqshrns_n_u32(a,b)
#define vqshrnd_n_u64(a,b) __simd_vqshrnd_n_u64(a,b)
#define vqrshrnh_n_s16(a,b) __simd_vqrshrnh_n_s16(a,b)
#define vqrshrns_n_s32(a,b) __simd_vqrshrns_n_s32(a,b)
#define vqrshrnd_n_s64(a,b) __simd_vqrshrnd_n_s64(a,b)
#define vqrshrnh_n_u16(a,b) __simd_vqrshrnh_n_u16(a,b)
#define vqrshrns_n_u32(a,b) __simd_vqrshrns_n_u32(a,b)
#define vqrshrnd_n_u64(a,b) __simd_vqrshrnd_n_u64(a,b)
#define vqshrunh_n_s16(a,b) __simd_vqshrunh_n_s16(a,b)
#define vqshruns_n_s32(a,b) __simd_vqshruns_n_s32(a,b)
#define vqshrund_n_s64(a,b) __simd_vqshrund_n_s64(a,b)
#define vqrshrunh_n_s16(a,b) __simd_vqrshrunh_n_s16(a,b)
#define vqrshruns_n_s32(a,b) __simd_vqrshruns_n_s32(a,b)
#define vqrshrund_n_s64(a,b) __simd_vqrshrund_n_s64(a,b)
#define vcvts_n_f32_s32(a,b) __simd_vcvts_n_f32_s32(a,b)
#define vcvtd_n_f64_s64(a,b) __simd_vcvtd_n_f64_s64(a,b)
#define vcvts_n_f32_u32(a,b) __simd_vcvts_n_f32_u32(a,b)
#define vcvtd_n_f64_u64(a,b) __simd_vcvtd_n_f64_u64(a,b)
#define vcvts_n_s32_f32(a,b) __simd_vcvts_n_s32_f32(a,b)
#define vcvtd_n_s64_f64(a,b) __simd_vcvtd_n_s64_f64(a,b)
#define vcvts_n_u32_f32(a,b) __simd_vcvts_n_u32_f32(a,b)
#define vcvtd_n_u64_f64(a,b) __simd_vcvtd_n_u64_f64(a,b)
#define vpadd_u64(a) __simd_vpadd_u64(a)
#define vpadd_s64(a) __simd_vpadd_s64(a)
#define vpaddd_u64(a) __simd_vpaddd_u64(a)
#define vpaddd_s64(a) __simd_vpaddd_s64(a)
#define vpadds_f32(a) __simd_vpadds_f32(a)
#define vpaddd_f64(a) __simd_vpaddd_f64(a)
#define vpmaxs_f32(a) __simd_vpmaxs_f32(a)
#define vpmaxd_f64(a) __simd_vpmaxd_f64(a)
#define vpmaxnms_f32(a) __simd_vpmaxnms_f32(a)
#define vpmaxnmd_f64(a) __simd_vpmaxnmd_f64(a)
#define vpmins_f32(a) __simd_vpmins_f32(a)
#define vpmind_f64(a) __simd_vpmind_f64(a)
#define vpminnms_f32(a) __simd_vpminnms_f32(a)
#define vpminnmd_f64(a) __simd_vpminnmd_f64(a)
#define vget_lane_f32(a,b) __simd_vget_lane_f32(a,b)
#define vgetq_lane_f32(a,b) __simd_vgetq_lane_f32(a,b)
#define vgetq_lane_f64(a,b) __simd_vgetq_lane_f64(a,b)
#define vdup_lane_u8(a,b) __simd_vdup_lane_u8(a,b)
#define vdup_lane_s8(a,b) __simd_vdup_lane_s8(a,b)
#define vdup_lane_p8(a,b) __simd_vdup_lane_p8(a,b)
#define vdup_laneq_u8(a,b) __simd_vdup_laneq_u8(a,b)
#define vdup_laneq_s8(a,b) __simd_vdup_laneq_s8(a,b)
#define vdup_laneq_p8(a,b) __simd_vdup_laneq_p8(a,b)
#define vdup_n_u8(a) __simd_vdup_n_u8(a)
#define vdup_n_s8(a) __simd_vdup_n_s8(a)
#define vdup_n_p8(a) __simd_vdup_n_p8(a)
#define vdupq_lane_u8(a,b) __simd_vdupq_lane_u8(a,b)
#define vdupq_lane_s8(a,b) __simd_vdupq_lane_s8(a,b)
#define vdupq_lane_p8(a,b) __simd_vdupq_lane_p8(a,b)
#define vdupq_laneq_u8(a,b) __simd_vdupq_laneq_u8(a,b)
#define vdupq_laneq_s8(a,b) __simd_vdupq_laneq_s8(a,b)
#define vdupq_laneq_p8(a,b) __simd_vdupq_laneq_p8(a,b)
#define vdupq_n_u8(a) __simd_vdupq_n_u8(a)
#define vdupq_n_s8(a) __simd_vdupq_n_s8(a)
#define vdupq_n_p8(a) __simd_vdupq_n_p8(a)
#define vdup_lane_u16(a,b) __simd_vdup_lane_u16(a,b)
#define vdup_lane_s16(a,b) __simd_vdup_lane_s16(a,b)
#define vdup_lane_p16(a,b) __simd_vdup_lane_p16(a,b)
#define vdup_laneq_u16(a,b) __simd_vdup_laneq_u16(a,b)
#define vdup_laneq_s16(a,b) __simd_vdup_laneq_s16(a,b)
#define vdup_laneq_p16(a,b) __simd_vdup_laneq_p16(a,b)
#define vdup_n_u16(a) __simd_vdup_n_u16(a)
#define vdup_n_s16(a) __simd_vdup_n_s16(a)
#define vdup_n_p16(a) __simd_vdup_n_p16(a)
#define vdupq_lane_u16(a,b) __simd_vdupq_lane_u16(a,b)
#define vdupq_lane_s16(a,b) __simd_vdupq_lane_s16(a,b)
#define vdupq_lane_p16(a,b) __simd_vdupq_lane_p16(a,b)
#define vdupq_laneq_u16(a,b) __simd_vdupq_laneq_u16(a,b)
#define vdupq_laneq_s16(a,b) __simd_vdupq_laneq_s16(a,b)
#define vdupq_laneq_p16(a,b) __simd_vdupq_laneq_p16(a,b)
#define vdupq_n_u16(a) __simd_vdupq_n_u16(a)
#define vdupq_n_s16(a) __simd_vdupq_n_s16(a)
#define vdupq_n_p16(a) __simd_vdupq_n_p16(a)
#define vdup_lane_u32(a,b) __simd_vdup_lane_u32(a,b)
#define vdup_lane_s32(a,b) __simd_vdup_lane_s32(a,b)
#define vdup_lane_f32(a,b) __simd_vdup_lane_f32(a,b)
#define vdup_laneq_u32(a,b) __simd_vdup_laneq_u32(a,b)
#define vdup_laneq_s32(a,b) __simd_vdup_laneq_s32(a,b)
#define vdup_laneq_f32(a,b) __simd_vdup_laneq_f32(a,b)
#define vdup_n_u32(a) __simd_vdup_n_u32(a)
#define vdup_n_s32(a) __simd_vdup_n_s32(a)
#define vdup_n_f32(a) __simd_vdup_n_f32(a)
#define vdupq_lane_u32(a,b) __simd_vdupq_lane_u32(a,b)
#define vdupq_lane_s32(a,b) __simd_vdupq_lane_s32(a,b)
#define vdupq_lane_f32(a,b) __simd_vdupq_lane_f32(a,b)
#define vdupq_laneq_u32(a,b) __simd_vdupq_laneq_u32(a,b)
#define vdupq_laneq_s32(a,b) __simd_vdupq_laneq_s32(a,b)
#define vdupq_laneq_f32(a,b) __simd_vdupq_laneq_f32(a,b)
#define vdupq_n_u32(a) __simd_vdupq_n_u32(a)
#define vdupq_n_s32(a) __simd_vdupq_n_s32(a)
#define vdupq_n_f32(a) __simd_vdupq_n_f32(a)
#define vdupq_lane_u64(a,b) __simd_vdupq_lane_u64(a,b)
#define vdupq_lane_s64(a,b) __simd_vdupq_lane_s64(a,b)
#define vdupq_laneq_u64(a,b) __simd_vdupq_laneq_u64(a,b)
#define vdupq_laneq_s64(a,b) __simd_vdupq_laneq_s64(a,b)
#define vdupq_laneq_f64(a,b) __simd_vdupq_laneq_f64(a,b)
#define vdupq_n_u64(a) __simd_vdupq_n_u64(a)
#define vdupq_n_s64(a) __simd_vdupq_n_s64(a)
#define vdupq_n_f64(a) __simd_vdupq_n_f64(a)
#define vget_lane_u32(a,b) __simd_vget_lane_u32(a,b)
#define vgetq_lane_u32(a,b) __simd_vgetq_lane_u32(a,b)
#define vget_lane_u8(a,b) __simd_vget_lane_u8(a,b)
#define vget_lane_p8(a,b) __simd_vget_lane_p8(a,b)
#define vgetq_lane_u8(a,b) __simd_vgetq_lane_u8(a,b)
#define vgetq_lane_p8(a,b) __simd_vgetq_lane_p8(a,b)
#define vget_lane_u16(a,b) __simd_vget_lane_u16(a,b)
#define vget_lane_p16(a,b) __simd_vget_lane_p16(a,b)
#define vgetq_lane_u16(a,b) __simd_vgetq_lane_u16(a,b)
#define vgetq_lane_p16(a,b) __simd_vgetq_lane_p16(a,b)
#define vget_lane_s8(a,b) __simd_vget_lane_s8(a,b)
#define vgetq_lane_s8(a,b) __simd_vgetq_lane_s8(a,b)
#define vget_lane_s16(a,b) __simd_vget_lane_s16(a,b)
#define vgetq_lane_s16(a,b) __simd_vgetq_lane_s16(a,b)
#define vgetq_lane_s32(a,b) __simd_vgetq_lane_s32(a,b)
#define vget_lane_s32(a,b) __simd_vget_lane_s32(a,b)
#define vget_lane_u64(a,b) __simd_vget_lane_u64(a,b)
#define vget_lane_s64(a,b) __simd_vget_lane_s64(a,b)
#define vgetq_lane_u64(a,b) __simd_vgetq_lane_u64(a,b)
#define vgetq_lane_s64(a,b) __simd_vgetq_lane_s64(a,b)
#define vset_lane_u8(a,b,c) __simd_vset_lane_u8(a,b,c)
#define vset_lane_s8(a,b,c) __simd_vset_lane_s8(a,b,c)
#define vset_lane_p8(a,b,c) __simd_vset_lane_p8(a,b,c)
#define vsetq_lane_u8(a,b,c) __simd_vsetq_lane_u8(a,b,c)
#define vsetq_lane_s8(a,b,c) __simd_vsetq_lane_s8(a,b,c)
#define vsetq_lane_p8(a,b,c) __simd_vsetq_lane_p8(a,b,c)
#define vset_lane_u16(a,b,c) __simd_vset_lane_u16(a,b,c)
#define vset_lane_s16(a,b,c) __simd_vset_lane_s16(a,b,c)
#define vset_lane_p16(a,b,c) __simd_vset_lane_p16(a,b,c)
#define vsetq_lane_u16(a,b,c) __simd_vsetq_lane_u16(a,b,c)
#define vsetq_lane_s16(a,b,c) __simd_vsetq_lane_s16(a,b,c)
#define vsetq_lane_p16(a,b,c) __simd_vsetq_lane_p16(a,b,c)
#define vset_lane_u32(a,b,c) __simd_vset_lane_u32(a,b,c)
#define vset_lane_s32(a,b,c) __simd_vset_lane_s32(a,b,c)
#define vsetq_lane_u32(a,b,c) __simd_vsetq_lane_u32(a,b,c)
#define vsetq_lane_s32(a,b,c) __simd_vsetq_lane_s32(a,b,c)
#define vset_lane_u64(a,b,c) __simd_vset_lane_u64(a,b,c)
#define vset_lane_s64(a,b,c) __simd_vset_lane_s64(a,b,c)
#define vsetq_lane_u64(a,b,c) __simd_vsetq_lane_u64(a,b,c)
#define vsetq_lane_s64(a,b,c) __simd_vsetq_lane_s64(a,b,c)
#define vcreate_u8(a) __simd_vcreate_u8(a)
#define vcreate_u16(a) __simd_vcreate_u16(a)
#define vcreate_u32(a) __simd_vcreate_u32(a)
#define vcreate_u64(a) __simd_vcreate_u64(a)
#define vcreate_s8(a) __simd_vcreate_s8(a)
#define vcreate_s16(a) __simd_vcreate_s16(a)
#define vcreate_s32(a) __simd_vcreate_s32(a)
#define vcreate_s64(a) __simd_vcreate_s64(a)
#define vcreate_p8(a) __simd_vcreate_p8(a)
#define vcreate_p16(a) __simd_vcreate_p16(a)
#define vcreate_f32(a) __simd_vcreate_f32(a)
#define vcopy_lane_u8(a,b,c,d) __simd_vcopy_lane_u8(a,b,c,d)
#define vcopy_lane_s8(a,b,c,d) __simd_vcopy_lane_s8(a,b,c,d)
#define vcopy_lane_p8(a,b,c,d) __simd_vcopy_lane_p8(a,b,c,d)
#define vcopy_laneq_u8(a,b,c,d) __simd_vcopy_laneq_u8(a,b,c,d)
#define vcopy_laneq_s8(a,b,c,d) __simd_vcopy_laneq_s8(a,b,c,d)
#define vcopy_laneq_p8(a,b,c,d) __simd_vcopy_laneq_p8(a,b,c,d)
#define vcopyq_lane_u8(a,b,c,d) __simd_vcopyq_lane_u8(a,b,c,d)
#define vcopyq_lane_s8(a,b,c,d) __simd_vcopyq_lane_s8(a,b,c,d)
#define vcopyq_lane_p8(a,b,c,d) __simd_vcopyq_lane_p8(a,b,c,d)
#define vcopyq_laneq_u8(a,b,c,d) __simd_vcopyq_laneq_u8(a,b,c,d)
#define vcopyq_laneq_s8(a,b,c,d) __simd_vcopyq_laneq_s8(a,b,c,d)
#define vcopyq_laneq_p8(a,b,c,d) __simd_vcopyq_laneq_p8(a,b,c,d)
#define vcopy_lane_u16(a,b,c,d) __simd_vcopy_lane_u16(a,b,c,d)
#define vcopy_lane_s16(a,b,c,d) __simd_vcopy_lane_s16(a,b,c,d)
#define vcopy_lane_p16(a,b,c,d) __simd_vcopy_lane_p16(a,b,c,d)
#define vcopy_laneq_u16(a,b,c,d) __simd_vcopy_laneq_u16(a,b,c,d)
#define vcopy_laneq_s16(a,b,c,d) __simd_vcopy_laneq_s16(a,b,c,d)
#define vcopy_laneq_p16(a,b,c,d) __simd_vcopy_laneq_p16(a,b,c,d)
#define vcopyq_lane_u16(a,b,c,d) __simd_vcopyq_lane_u16(a,b,c,d)
#define vcopyq_lane_s16(a,b,c,d) __simd_vcopyq_lane_s16(a,b,c,d)
#define vcopyq_lane_p16(a,b,c,d) __simd_vcopyq_lane_p16(a,b,c,d)
#define vcopyq_laneq_u16(a,b,c,d) __simd_vcopyq_laneq_u16(a,b,c,d)
#define vcopyq_laneq_s16(a,b,c,d) __simd_vcopyq_laneq_s16(a,b,c,d)
#define vcopyq_laneq_p16(a,b,c,d) __simd_vcopyq_laneq_p16(a,b,c,d)
#define vcopy_lane_u32(a,b,c,d) __simd_vcopy_lane_u32(a,b,c,d)
#define vcopy_lane_s32(a,b,c,d) __simd_vcopy_lane_s32(a,b,c,d)
#define vcopy_lane_f32(a,b,c,d) __simd_vcopy_lane_f32(a,b,c,d)
#define vcopy_laneq_u32(a,b,c,d) __simd_vcopy_laneq_u32(a,b,c,d)
#define vcopy_laneq_s32(a,b,c,d) __simd_vcopy_laneq_s32(a,b,c,d)
#define vcopy_laneq_f32(a,b,c,d) __simd_vcopy_laneq_f32(a,b,c,d)
#define vcopyq_lane_u32(a,b,c,d) __simd_vcopyq_lane_u32(a,b,c,d)
#define vcopyq_lane_s32(a,b,c,d) __simd_vcopyq_lane_s32(a,b,c,d)
#define vcopyq_lane_f32(a,b,c,d) __simd_vcopyq_lane_f32(a,b,c,d)
#define vcopyq_laneq_u32(a,b,c,d) __simd_vcopyq_laneq_u32(a,b,c,d)
#define vcopyq_laneq_s32(a,b,c,d) __simd_vcopyq_laneq_s32(a,b,c,d)
#define vcopyq_laneq_f32(a,b,c,d) __simd_vcopyq_laneq_f32(a,b,c,d)
#define vcopy_lane_u64(a,b,c,d) __simd_vcopy_lane_u64(a,b,c,d)
#define vcopy_lane_s64(a,b,c,d) __simd_vcopy_lane_s64(a,b,c,d)
#define vcopy_laneq_u64(a,b,c,d) __simd_vcopy_laneq_u64(a,b,c,d)
#define vcopy_laneq_s64(a,b,c,d) __simd_vcopy_laneq_s64(a,b,c,d)
#define vcopyq_lane_u64(a,b,c,d) __simd_vcopyq_lane_u64(a,b,c,d)
#define vcopyq_lane_s64(a,b,c,d) __simd_vcopyq_lane_s64(a,b,c,d)
#define vcopyq_laneq_u64(a,b,c,d) __simd_vcopyq_laneq_u64(a,b,c,d)
#define vcopyq_laneq_s64(a,b,c,d) __simd_vcopyq_laneq_s64(a,b,c,d)
#define vcopyq_laneq_f64(a,b,c,d) __simd_vcopyq_laneq_f64(a,b,c,d)
#define vset_lane_f32(a,b,c) __simd_vset_lane_f32(a,b,c)
#define vsetq_lane_f32(a,b,c) __simd_vsetq_lane_f32(a,b,c)
#define vsetq_lane_f64(a,b,c) __simd_vsetq_lane_f64(a,b,c)
#define vaddlv_s8(a) __simd_vaddlv_s8(a)
#define vaddlv_s16(a) __simd_vaddlv_s16(a)
#define vaddlvq_s8(a) __simd_vaddlvq_s8(a)
#define vaddlvq_s16(a) __simd_vaddlvq_s16(a)
#define vaddlvq_s32(a) __simd_vaddlvq_s32(a)
#define vaddlv_u8(a) __simd_vaddlv_u8(a)
#define vaddlv_u16(a) __simd_vaddlv_u16(a)
#define vaddlvq_u8(a) __simd_vaddlvq_u8(a)
#define vaddlvq_u16(a) __simd_vaddlvq_u16(a)
#define vaddlvq_u32(a) __simd_vaddlvq_u32(a)
#define vaddv_u8(a) __simd_vaddv_u8(a)
#define vaddv_s8(a) __simd_vaddv_s8(a)
#define vaddv_u16(a) __simd_vaddv_u16(a)
#define vaddv_s16(a) __simd_vaddv_s16(a)
#define vaddvq_u8(a) __simd_vaddvq_u8(a)
#define vaddvq_s8(a) __simd_vaddvq_s8(a)
#define vaddvq_u16(a) __simd_vaddvq_u16(a)
#define vaddvq_s16(a) __simd_vaddvq_s16(a)
#define vaddvq_u32(a) __simd_vaddvq_u32(a)
#define vaddvq_s32(a) __simd_vaddvq_s32(a)
#define vmaxv_s8(a) __simd_vmaxv_s8(a)
#define vmaxv_s16(a) __simd_vmaxv_s16(a)
#define vmaxvq_s8(a) __simd_vmaxvq_s8(a)
#define vmaxvq_s16(a) __simd_vmaxvq_s16(a)
#define vmaxvq_s32(a) __simd_vmaxvq_s32(a)
#define vmaxv_u8(a) __simd_vmaxv_u8(a)
#define vmaxv_u16(a) __simd_vmaxv_u16(a)
#define vmaxvq_u8(a) __simd_vmaxvq_u8(a)
#define vmaxvq_u16(a) __simd_vmaxvq_u16(a)
#define vmaxvq_u32(a) __simd_vmaxvq_u32(a)
#define vminv_s8(a) __simd_vminv_s8(a)
#define vminv_s16(a) __simd_vminv_s16(a)
#define vminvq_s8(a) __simd_vminvq_s8(a)
#define vminvq_s16(a) __simd_vminvq_s16(a)
#define vminvq_s32(a) __simd_vminvq_s32(a)
#define vminv_u8(a) __simd_vminv_u8(a)
#define vminv_u16(a) __simd_vminv_u16(a)
#define vminvq_u8(a) __simd_vminvq_u8(a)
#define vminvq_u16(a) __simd_vminvq_u16(a)
#define vminvq_u32(a) __simd_vminvq_u32(a)
#define vmaxvq_f32(a) __simd_vmaxvq_f32(a)
#define vmaxnmvq_f32(a) __simd_vmaxnmvq_f32(a)
#define vminvq_f32(a) __simd_vminvq_f32(a)
#define vminnmvq_f32(a) __simd_vminnmvq_f32(a)
#define vext_u8(a,b,c) __simd_vext_u8(a,b,c)
#define vext_s8(a,b,c) __simd_vext_s8(a,b,c)
#define vext_p8(a,b,c) __simd_vext_p8(a,b,c)
#define vextq_u8(a,b,c) __simd_vextq_u8(a,b,c)
#define vextq_s8(a,b,c) __simd_vextq_s8(a,b,c)
#define vextq_p8(a,b,c) __simd_vextq_p8(a,b,c)
#define vext_u16(a,b,c) __simd_vext_u16(a,b,c)
#define vext_s16(a,b,c) __simd_vext_s16(a,b,c)
#define vext_p16(a,b,c) __simd_vext_p16(a,b,c)
#define vext_u32(a,b,c) __simd_vext_u32(a,b,c)
#define vext_s32(a,b,c) __simd_vext_s32(a,b,c)
#define vext_f32(a,b,c) __simd_vext_f32(a,b,c)
#define vext_u64(a,b,c) __simd_vext_u64(a,b,c)
#define vext_s64(a,b,c) __simd_vext_s64(a,b,c)
#define vextd_u64(a,b,c) __simd_vextd_u64(a,b,c)
#define vextd_s64(a,b,c) __simd_vextd_s64(a,b,c)
#define vextq_u16(a,b,c) __simd_vextq_u16(a,b,c)
#define vextq_s16(a,b,c) __simd_vextq_s16(a,b,c)
#define vextq_p16(a,b,c) __simd_vextq_p16(a,b,c)
#define vextq_u32(a,b,c) __simd_vextq_u32(a,b,c)
#define vextq_s32(a,b,c) __simd_vextq_s32(a,b,c)
#define vextq_f32(a,b,c) __simd_vextq_f32(a,b,c)
#define vextq_u64(a,b,c) __simd_vextq_u64(a,b,c)
#define vextq_s64(a,b,c) __simd_vextq_s64(a,b,c)
#define vextq_f64(a,b,c) __simd_vextq_f64(a,b,c)
#define vtbl1_u8(a,b) __simd_vtbl1_u8(a,b)
#define vtbl1_s8(a,b) __simd_vtbl1_s8(a,b)
#define vtbl1_p8(a,b) __simd_vtbl1_p8(a,b)
#define vtbl2_u8(a,b) __simd_vtbl2_u8(a,b)
#define vtbl2_s8(a,b) __simd_vtbl2_s8(a,b)
#define vtbl2_p8(a,b) __simd_vtbl2_p8(a,b)
#define vqtbl1_u8(a,b) __simd_vqtbl1_u8(a,b)
#define vqtbl1_s8(a,b) __simd_vqtbl1_s8(a,b)
#define vqtbl1_p8(a,b) __simd_vqtbl1_p8(a,b)
#define vtbl3_u8(a,b) __simd_vtbl3_u8(a,b)
#define vtbl3_s8(a,b) __simd_vtbl3_s8(a,b)
#define vtbl3_p8(a,b) __simd_vtbl3_p8(a,b)
#define vtbl4_u8(a,b) __simd_vtbl4_u8(a,b)
#define vtbl4_s8(a,b) __simd_vtbl4_s8(a,b)
#define vtbl4_p8(a,b) __simd_vtbl4_p8(a,b)
#define vqtbl2_u8(a,b) __simd_vqtbl2_u8(a,b)
#define vqtbl2_s8(a,b) __simd_vqtbl2_s8(a,b)
#define vqtbl2_p8(a,b) __simd_vqtbl2_p8(a,b)
#define vqtbl3_u8(a,b) __simd_vqtbl3_u8(a,b)
#define vqtbl3_s8(a,b) __simd_vqtbl3_s8(a,b)
#define vqtbl3_p8(a,b) __simd_vqtbl3_p8(a,b)
#define vqtbl4_u8(a,b) __simd_vqtbl4_u8(a,b)
#define vqtbl4_s8(a,b) __simd_vqtbl4_s8(a,b)
#define vqtbl4_p8(a,b) __simd_vqtbl4_p8(a,b)
#define vqtbl1q_u8(a,b) __simd_vqtbl1q_u8(a,b)
#define vqtbl1q_s8(a,b) __simd_vqtbl1q_s8(a,b)
#define vqtbl1q_p8(a,b) __simd_vqtbl1q_p8(a,b)
#define vqtbl2q_u8(a,b) __simd_vqtbl2q_u8(a,b)
#define vqtbl2q_s8(a,b) __simd_vqtbl2q_s8(a,b)
#define vqtbl2q_p8(a,b) __simd_vqtbl2q_p8(a,b)
#define vqtbl3q_u8(a,b) __simd_vqtbl3q_u8(a,b)
#define vqtbl3q_s8(a,b) __simd_vqtbl3q_s8(a,b)
#define vqtbl3q_p8(a,b) __simd_vqtbl3q_p8(a,b)
#define vqtbl4q_u8(a,b) __simd_vqtbl4q_u8(a,b)
#define vqtbl4q_s8(a,b) __simd_vqtbl4q_s8(a,b)
#define vqtbl4q_p8(a,b) __simd_vqtbl4q_p8(a,b)
#define vtbx1_u8(a,b,c) __simd_vtbx1_u8(a,b,c)
#define vtbx1_s8(a,b,c) __simd_vtbx1_s8(a,b,c)
#define vtbx1_p8(a,b,c) __simd_vtbx1_p8(a,b,c)
#define vtbx2_u8(a,b,c) __simd_vtbx2_u8(a,b,c)
#define vtbx2_s8(a,b,c) __simd_vtbx2_s8(a,b,c)
#define vtbx2_p8(a,b,c) __simd_vtbx2_p8(a,b,c)
#define vqtbx1_u8(a,b,c) __simd_vqtbx1_u8(a,b,c)
#define vqtbx1_s8(a,b,c) __simd_vqtbx1_s8(a,b,c)
#define vqtbx1_p8(a,b,c) __simd_vqtbx1_p8(a,b,c)
#define vtbx3_u8(a,b,c) __simd_vtbx3_u8(a,b,c)
#define vtbx3_s8(a,b,c) __simd_vtbx3_s8(a,b,c)
#define vtbx3_p8(a,b,c) __simd_vtbx3_p8(a,b,c)
#define vtbx4_u8(a,b,c) __simd_vtbx4_u8(a,b,c)
#define vtbx4_s8(a,b,c) __simd_vtbx4_s8(a,b,c)
#define vtbx4_p8(a,b,c) __simd_vtbx4_p8(a,b,c)
#define vqtbx2_u8(a,b,c) __simd_vqtbx2_u8(a,b,c)
#define vqtbx2_s8(a,b,c) __simd_vqtbx2_s8(a,b,c)
#define vqtbx2_p8(a,b,c) __simd_vqtbx2_p8(a,b,c)
#define vqtbx3_u8(a,b,c) __simd_vqtbx3_u8(a,b,c)
#define vqtbx3_s8(a,b,c) __simd_vqtbx3_s8(a,b,c)
#define vqtbx3_p8(a,b,c) __simd_vqtbx3_p8(a,b,c)
#define vqtbx4_u8(a,b,c) __simd_vqtbx4_u8(a,b,c)
#define vqtbx4_s8(a,b,c) __simd_vqtbx4_s8(a,b,c)
#define vqtbx4_p8(a,b,c) __simd_vqtbx4_p8(a,b,c)
#define vqtbx1q_u8(a,b,c) __simd_vqtbx1q_u8(a,b,c)
#define vqtbx1q_s8(a,b,c) __simd_vqtbx1q_s8(a,b,c)
#define vqtbx1q_p8(a,b,c) __simd_vqtbx1q_p8(a,b,c)
#define vqtbx2q_u8(a,b,c) __simd_vqtbx2q_u8(a,b,c)
#define vqtbx2q_s8(a,b,c) __simd_vqtbx2q_s8(a,b,c)
#define vqtbx2q_p8(a,b,c) __simd_vqtbx2q_p8(a,b,c)
#define vqtbx3q_u8(a,b,c) __simd_vqtbx3q_u8(a,b,c)
#define vqtbx3q_s8(a,b,c) __simd_vqtbx3q_s8(a,b,c)
#define vqtbx3q_p8(a,b,c) __simd_vqtbx3q_p8(a,b,c)
#define vqtbx4q_u8(a,b,c) __simd_vqtbx4q_u8(a,b,c)
#define vqtbx4q_s8(a,b,c) __simd_vqtbx4q_s8(a,b,c)
#define vqtbx4q_p8(a,b,c) __simd_vqtbx4q_p8(a,b,c)
#define vcombine_u8(a,b) __simd_vcombine_u8(a,b)
#define vcombine_s8(a,b) __simd_vcombine_s8(a,b)
#define vcombine_p8(a,b) __simd_vcombine_p8(a,b)
#define vcombine_u16(a,b) __simd_vcombine_u16(a,b)
#define vcombine_s16(a,b) __simd_vcombine_s16(a,b)
#define vcombine_p16(a,b) __simd_vcombine_p16(a,b)
#define vcombine_u32(a,b) __simd_vcombine_u32(a,b)
#define vcombine_s32(a,b) __simd_vcombine_s32(a,b)
#define vcombine_f32(a,b) __simd_vcombine_f32(a,b)
#define vcombine_u64(a,b) __simd_vcombine_u64(a,b)
#define vcombine_s64(a,b) __simd_vcombine_s64(a,b)
#define vcombine_f64(a,b) __simd_vcombine_f64(a,b)
#define vget_high_u8(a) __simd_vget_high_u8(a)
#define vget_high_s8(a) __simd_vget_high_s8(a)
#define vget_high_p8(a) __simd_vget_high_p8(a)
#define vget_high_u16(a) __simd_vget_high_u16(a)
#define vget_high_s16(a) __simd_vget_high_s16(a)
#define vget_high_p16(a) __simd_vget_high_p16(a)
#define vget_high_u32(a) __simd_vget_high_u32(a)
#define vget_high_s32(a) __simd_vget_high_s32(a)
#define vget_high_f32(a) __simd_vget_high_f32(a)
#define vget_high_u64(a) __simd_vget_high_u64(a)
#define vget_high_s64(a) __simd_vget_high_s64(a)
#define vget_high_f64(a) __simd_vget_high_f64(a)
#define vget_low_u8(a) __simd_vget_low_u8(a)
#define vget_low_s8(a) __simd_vget_low_s8(a)
#define vget_low_p8(a) __simd_vget_low_p8(a)
#define vget_low_u16(a) __simd_vget_low_u16(a)
#define vget_low_s16(a) __simd_vget_low_s16(a)
#define vget_low_p16(a) __simd_vget_low_p16(a)
#define vget_low_u32(a) __simd_vget_low_u32(a)
#define vget_low_s32(a) __simd_vget_low_s32(a)
#define vget_low_f32(a) __simd_vget_low_f32(a)
#define vget_low_u64(a) __simd_vget_low_u64(a)
#define vget_low_s64(a) __simd_vget_low_s64(a)
#define vget_low_f64(a) __simd_vget_low_f64(a)
#define vtrn1_u8(a,b) __simd_vtrn1_u8(a,b)
#define vtrn1_s8(a,b) __simd_vtrn1_s8(a,b)
#define vtrn1_p8(a,b) __simd_vtrn1_p8(a,b)
#define vtrn1_u16(a,b) __simd_vtrn1_u16(a,b)
#define vtrn1_s16(a,b) __simd_vtrn1_s16(a,b)
#define vtrn1_p16(a,b) __simd_vtrn1_p16(a,b)
#define vtrn1_u32(a,b) __simd_vtrn1_u32(a,b)
#define vtrn1_s32(a,b) __simd_vtrn1_s32(a,b)
#define vtrn1_f32(a,b) __simd_vtrn1_f32(a,b)
#define vtrn1q_u8(a,b) __simd_vtrn1q_u8(a,b)
#define vtrn1q_s8(a,b) __simd_vtrn1q_s8(a,b)
#define vtrn1q_p8(a,b) __simd_vtrn1q_p8(a,b)
#define vtrn1q_u16(a,b) __simd_vtrn1q_u16(a,b)
#define vtrn1q_s16(a,b) __simd_vtrn1q_s16(a,b)
#define vtrn1q_p16(a,b) __simd_vtrn1q_p16(a,b)
#define vtrn1q_u32(a,b) __simd_vtrn1q_u32(a,b)
#define vtrn1q_s32(a,b) __simd_vtrn1q_s32(a,b)
#define vtrn1q_f32(a,b) __simd_vtrn1q_f32(a,b)
#define vtrn1q_u64(a,b) __simd_vtrn1q_u64(a,b)
#define vtrn1q_s64(a,b) __simd_vtrn1q_s64(a,b)
#define vtrn1q_f64(a,b) __simd_vtrn1q_f64(a,b)
#define vtrn2_u8(a,b) __simd_vtrn2_u8(a,b)
#define vtrn2_s8(a,b) __simd_vtrn2_s8(a,b)
#define vtrn2_p8(a,b) __simd_vtrn2_p8(a,b)
#define vtrn2_u16(a,b) __simd_vtrn2_u16(a,b)
#define vtrn2_s16(a,b) __simd_vtrn2_s16(a,b)
#define vtrn2_p16(a,b) __simd_vtrn2_p16(a,b)
#define vtrn2_u32(a,b) __simd_vtrn2_u32(a,b)
#define vtrn2_s32(a,b) __simd_vtrn2_s32(a,b)
#define vtrn2_f32(a,b) __simd_vtrn2_f32(a,b)
#define vtrn2q_u8(a,b) __simd_vtrn2q_u8(a,b)
#define vtrn2q_s8(a,b) __simd_vtrn2q_s8(a,b)
#define vtrn2q_p8(a,b) __simd_vtrn2q_p8(a,b)
#define vtrn2q_u16(a,b) __simd_vtrn2q_u16(a,b)
#define vtrn2q_s16(a,b) __simd_vtrn2q_s16(a,b)
#define vtrn2q_p16(a,b) __simd_vtrn2q_p16(a,b)
#define vtrn2q_u32(a,b) __simd_vtrn2q_u32(a,b)
#define vtrn2q_s32(a,b) __simd_vtrn2q_s32(a,b)
#define vtrn2q_f32(a,b) __simd_vtrn2q_f32(a,b)
#define vtrn2q_u64(a,b) __simd_vtrn2q_u64(a,b)
#define vtrn2q_s64(a,b) __simd_vtrn2q_s64(a,b)
#define vtrn2q_f64(a,b) __simd_vtrn2q_f64(a,b)
#define vuzp1_u8(a,b) __simd_vuzp1_u8(a,b)
#define vuzp1_s8(a,b) __simd_vuzp1_s8(a,b)
#define vuzp1_p8(a,b) __simd_vuzp1_p8(a,b)
#define vuzp1_u16(a,b) __simd_vuzp1_u16(a,b)
#define vuzp1_s16(a,b) __simd_vuzp1_s16(a,b)
#define vuzp1_p16(a,b) __simd_vuzp1_p16(a,b)
#define vuzp1_u32(a,b) __simd_vuzp1_u32(a,b)
#define vuzp1_s32(a,b) __simd_vuzp1_s32(a,b)
#define vuzp1_f32(a,b) __simd_vuzp1_f32(a,b)
#define vuzp1q_u8(a,b) __simd_vuzp1q_u8(a,b)
#define vuzp1q_s8(a,b) __simd_vuzp1q_s8(a,b)
#define vuzp1q_p8(a,b) __simd_vuzp1q_p8(a,b)
#define vuzp1q_u16(a,b) __simd_vuzp1q_u16(a,b)
#define vuzp1q_s16(a,b) __simd_vuzp1q_s16(a,b)
#define vuzp1q_p16(a,b) __simd_vuzp1q_p16(a,b)
#define vuzp1q_u32(a,b) __simd_vuzp1q_u32(a,b)
#define vuzp1q_s32(a,b) __simd_vuzp1q_s32(a,b)
#define vuzp1q_f32(a,b) __simd_vuzp1q_f32(a,b)
#define vuzp1q_u64(a,b) __simd_vuzp1q_u64(a,b)
#define vuzp1q_s64(a,b) __simd_vuzp1q_s64(a,b)
#define vuzp1q_f64(a,b) __simd_vuzp1q_f64(a,b)
#define vuzp2_u8(a,b) __simd_vuzp2_u8(a,b)
#define vuzp2_s8(a,b) __simd_vuzp2_s8(a,b)
#define vuzp2_p8(a,b) __simd_vuzp2_p8(a,b)
#define vuzp2_u16(a,b) __simd_vuzp2_u16(a,b)
#define vuzp2_s16(a,b) __simd_vuzp2_s16(a,b)
#define vuzp2_p16(a,b) __simd_vuzp2_p16(a,b)
#define vuzp2_u32(a,b) __simd_vuzp2_u32(a,b)
#define vuzp2_s32(a,b) __simd_vuzp2_s32(a,b)
#define vuzp2_f32(a,b) __simd_vuzp2_f32(a,b)
#define vuzp2q_u8(a,b) __simd_vuzp2q_u8(a,b)
#define vuzp2q_s8(a,b) __simd_vuzp2q_s8(a,b)
#define vuzp2q_p8(a,b) __simd_vuzp2q_p8(a,b)
#define vuzp2q_u16(a,b) __simd_vuzp2q_u16(a,b)
#define vuzp2q_s16(a,b) __simd_vuzp2q_s16(a,b)
#define vuzp2q_p16(a,b) __simd_vuzp2q_p16(a,b)
#define vuzp2q_u32(a,b) __simd_vuzp2q_u32(a,b)
#define vuzp2q_s32(a,b) __simd_vuzp2q_s32(a,b)
#define vuzp2q_f32(a,b) __simd_vuzp2q_f32(a,b)
#define vuzp2q_u64(a,b) __simd_vuzp2q_u64(a,b)
#define vuzp2q_s64(a,b) __simd_vuzp2q_s64(a,b)
#define vuzp2q_f64(a,b) __simd_vuzp2q_f64(a,b)
#define vzip1_u8(a,b) __simd_vzip1_u8(a,b)
#define vzip1_s8(a,b) __simd_vzip1_s8(a,b)
#define vzip1_p8(a,b) __simd_vzip1_p8(a,b)
#define vzip1_u16(a,b) __simd_vzip1_u16(a,b)
#define vzip1_s16(a,b) __simd_vzip1_s16(a,b)
#define vzip1_p16(a,b) __simd_vzip1_p16(a,b)
#define vzip1_u32(a,b) __simd_vzip1_u32(a,b)
#define vzip1_s32(a,b) __simd_vzip1_s32(a,b)
#define vzip1_f32(a,b) __simd_vzip1_f32(a,b)
#define vzip1q_u8(a,b) __simd_vzip1q_u8(a,b)
#define vzip1q_s8(a,b) __simd_vzip1q_s8(a,b)
#define vzip1q_p8(a,b) __simd_vzip1q_p8(a,b)
#define vzip1q_u16(a,b) __simd_vzip1q_u16(a,b)
#define vzip1q_s16(a,b) __simd_vzip1q_s16(a,b)
#define vzip1q_p16(a,b) __simd_vzip1q_p16(a,b)
#define vzip1q_u32(a,b) __simd_vzip1q_u32(a,b)
#define vzip1q_s32(a,b) __simd_vzip1q_s32(a,b)
#define vzip1q_f32(a,b) __simd_vzip1q_f32(a,b)
#define vzip1q_u64(a,b) __simd_vzip1q_u64(a,b)
#define vzip1q_s64(a,b) __simd_vzip1q_s64(a,b)
#define vzip1q_f64(a,b) __simd_vzip1q_f64(a,b)
#define vzip2_u8(a,b) __simd_vzip2_u8(a,b)
#define vzip2_s8(a,b) __simd_vzip2_s8(a,b)
#define vzip2_p8(a,b) __simd_vzip2_p8(a,b)
#define vzip2_u16(a,b) __simd_vzip2_u16(a,b)
#define vzip2_s16(a,b) __simd_vzip2_s16(a,b)
#define vzip2_p16(a,b) __simd_vzip2_p16(a,b)
#define vzip2_u32(a,b) __simd_vzip2_u32(a,b)
#define vzip2_s32(a,b) __simd_vzip2_s32(a,b)
#define vzip2_f32(a,b) __simd_vzip2_f32(a,b)
#define vzip2q_u8(a,b) __simd_vzip2q_u8(a,b)
#define vzip2q_s8(a,b) __simd_vzip2q_s8(a,b)
#define vzip2q_p8(a,b) __simd_vzip2q_p8(a,b)
#define vzip2q_u16(a,b) __simd_vzip2q_u16(a,b)
#define vzip2q_s16(a,b) __simd_vzip2q_s16(a,b)
#define vzip2q_p16(a,b) __simd_vzip2q_p16(a,b)
#define vzip2q_u32(a,b) __simd_vzip2q_u32(a,b)
#define vzip2q_s32(a,b) __simd_vzip2q_s32(a,b)
#define vzip2q_f32(a,b) __simd_vzip2q_f32(a,b)
#define vzip2q_u64(a,b) __simd_vzip2q_u64(a,b)
#define vzip2q_s64(a,b) __simd_vzip2q_s64(a,b)
#define vzip2q_f64(a,b) __simd_vzip2q_f64(a,b)
#define vtrn_u8(a,b) __simd_vtrn_u8(a,b)
#define vtrn_s8(a,b) __simd_vtrn_s8(a,b)
#define vtrn_p8(a,b) __simd_vtrn_p8(a,b)
#define vtrn_u16(a,b) __simd_vtrn_u16(a,b)
#define vtrn_s16(a,b) __simd_vtrn_s16(a,b)
#define vtrn_p16(a,b) __simd_vtrn_p16(a,b)
#define vtrn_u32(a,b) __simd_vtrn_u32(a,b)
#define vtrn_s32(a,b) __simd_vtrn_s32(a,b)
#define vtrn_f32(a,b) __simd_vtrn_f32(a,b)
#define vtrnq_u8(a,b) __simd_vtrnq_u8(a,b)
#define vtrnq_s8(a,b) __simd_vtrnq_s8(a,b)
#define vtrnq_p8(a,b) __simd_vtrnq_p8(a,b)
#define vtrnq_u16(a,b) __simd_vtrnq_u16(a,b)
#define vtrnq_s16(a,b) __simd_vtrnq_s16(a,b)
#define vtrnq_p16(a,b) __simd_vtrnq_p16(a,b)
#define vtrnq_u32(a,b) __simd_vtrnq_u32(a,b)
#define vtrnq_s32(a,b) __simd_vtrnq_s32(a,b)
#define vtrnq_f32(a,b) __simd_vtrnq_f32(a,b)
#define vtrnq_u64(a,b) __simd_vtrnq_u64(a,b)
#define vtrnq_s64(a,b) __simd_vtrnq_s64(a,b)
#define vtrnq_f64(a,b) __simd_vtrnq_f64(a,b)
#define vzip_u8(a,b) __simd_vzip_u8(a,b)
#define vzip_s8(a,b) __simd_vzip_s8(a,b)
#define vzip_p8(a,b) __simd_vzip_p8(a,b)
#define vzip_u16(a,b) __simd_vzip_u16(a,b)
#define vzip_s16(a,b) __simd_vzip_s16(a,b)
#define vzip_p16(a,b) __simd_vzip_p16(a,b)
#define vzip_u32(a,b) __simd_vzip_u32(a,b)
#define vzip_s32(a,b) __simd_vzip_s32(a,b)
#define vzip_f32(a,b) __simd_vzip_f32(a,b)
#define vzipq_u8(a,b) __simd_vzipq_u8(a,b)
#define vzipq_s8(a,b) __simd_vzipq_s8(a,b)
#define vzipq_p8(a,b) __simd_vzipq_p8(a,b)
#define vzipq_u16(a,b) __simd_vzipq_u16(a,b)
#define vzipq_s16(a,b) __simd_vzipq_s16(a,b)
#define vzipq_p16(a,b) __simd_vzipq_p16(a,b)
#define vzipq_u32(a,b) __simd_vzipq_u32(a,b)
#define vzipq_s32(a,b) __simd_vzipq_s32(a,b)
#define vzipq_f32(a,b) __simd_vzipq_f32(a,b)
#define vzipq_u64(a,b) __simd_vzipq_u64(a,b)
#define vzipq_s64(a,b) __simd_vzipq_s64(a,b)
#define vzipq_f64(a,b) __simd_vzipq_f64(a,b)
#define vuzp_u8(a,b) __simd_vuzp_u8(a,b)
#define vuzp_s8(a,b) __simd_vuzp_s8(a,b)
#define vuzp_p8(a,b) __simd_vuzp_p8(a,b)
#define vuzp_u16(a,b) __simd_vuzp_u16(a,b)
#define vuzp_s16(a,b) __simd_vuzp_s16(a,b)
#define vuzp_p16(a,b) __simd_vuzp_p16(a,b)
#define vuzp_u32(a,b) __simd_vuzp_u32(a,b)
#define vuzp_s32(a,b) __simd_vuzp_s32(a,b)
#define vuzp_f32(a,b) __simd_vuzp_f32(a,b)
#define vuzpq_u8(a,b) __simd_vuzpq_u8(a,b)
#define vuzpq_s8(a,b) __simd_vuzpq_s8(a,b)
#define vuzpq_p8(a,b) __simd_vuzpq_p8(a,b)
#define vuzpq_u16(a,b) __simd_vuzpq_u16(a,b)
#define vuzpq_s16(a,b) __simd_vuzpq_s16(a,b)
#define vuzpq_p16(a,b) __simd_vuzpq_p16(a,b)
#define vuzpq_u32(a,b) __simd_vuzpq_u32(a,b)
#define vuzpq_s32(a,b) __simd_vuzpq_s32(a,b)
#define vuzpq_f32(a,b) __simd_vuzpq_f32(a,b)
#define vuzpq_u64(a,b) __simd_vuzpq_u64(a,b)
#define vuzpq_s64(a,b) __simd_vuzpq_s64(a,b)
#define vuzpq_f64(a,b) __simd_vuzpq_f64(a,b)
#define vqrdmlah_s16(a,b,c) __simd_vqrdmlah_s16(a,b,c)
#define vqrdmlah_s32(a,b,c) __simd_vqrdmlah_s32(a,b,c)
#define vqrdmlahq_s16(a,b,c) __simd_vqrdmlahq_s16(a,b,c)
#define vqrdmlahq_s32(a,b,c) __simd_vqrdmlahq_s32(a,b,c)
#define vqrdmlahh_s16(a,b,c) __simd_vqrdmlahh_s16(a,b,c)
#define vqrdmlahs_s32(a,b,c) __simd_vqrdmlahs_s32(a,b,c)
#define vqrdmlah_lane_s16(a,b,c,d) __simd_vqrdmlah_lane_s16(a,b,c,d)
#define vqrdmlah_laneq_s16(a,b,c,d) __simd_vqrdmlah_laneq_s16(a,b,c,d)
#define vqrdmlah_n_s16(a,b,c) __simd_vqrdmlah_n_s16(a,b,c)
#define vqrdmlah_lane_s32(a,b,c,d) __simd_vqrdmlah_lane_s32(a,b,c,d)
#define vqrdmlah_laneq_s32(a,b,c,d) __simd_vqrdmlah_laneq_s32(a,b,c,d)
#define vqrdmlah_n_s32(a,b,c) __simd_vqrdmlah_n_s32(a,b,c)
#define vqrdmlahq_lane_s16(a,b,c,d) __simd_vqrdmlahq_lane_s16(a,b,c,d)
#define vqrdmlahq_laneq_s16(a,b,c,d) __simd_vqrdmlahq_laneq_s16(a,b,c,d)
#define vqrdmlahq_n_s16(a,b,c) __simd_vqrdmlahq_n_s16(a,b,c)
#define vqrdmlahq_lane_s32(a,b,c,d) __simd_vqrdmlahq_lane_s32(a,b,c,d)
#define vqrdmlahq_laneq_s32(a,b,c,d) __simd_vqrdmlahq_laneq_s32(a,b,c,d)
#define vqrdmlahq_n_s32(a,b,c) __simd_vqrdmlahq_n_s32(a,b,c)
#define vqrdmlahh_lane_s16(a,b,c,d) __simd_vqrdmlahh_lane_s16(a,b,c,d)
#define vqrdmlahh_laneq_s16(a,b,c,d) __simd_vqrdmlahh_laneq_s16(a,b,c,d)
#define vqrdmlahs_lane_s32(a,b,c,d) __simd_vqrdmlahs_lane_s32(a,b,c,d)
#define vqrdmlahs_laneq_s32(a,b,c,d) __simd_vqrdmlahs_laneq_s32(a,b,c,d)
#define vqrdmlsh_s16(a,b,c) __simd_vqrdmlsh_s16(a,b,c)
#define vqrdmlsh_s32(a,b,c) __simd_vqrdmlsh_s32(a,b,c)
#define vqrdmlshq_s16(a,b,c) __simd_vqrdmlshq_s16(a,b,c)
#define vqrdmlshq_s32(a,b,c) __simd_vqrdmlshq_s32(a,b,c)
#define vqrdmlshh_s16(a,b,c) __simd_vqrdmlshh_s16(a,b,c)
#define vqrdmlshs_s32(a,b,c) __simd_vqrdmlshs_s32(a,b,c)
#define vqrdmlsh_lane_s16(a,b,c,d) __simd_vqrdmlsh_lane_s16(a,b,c,d)
#define vqrdmlsh_laneq_s16(a,b,c,d) __simd_vqrdmlsh_laneq_s16(a,b,c,d)
#define vqrdmlsh_n_s16(a,b,c) __simd_vqrdmlsh_n_s16(a,b,c)
#define vqrdmlsh_lane_s32(a,b,c,d) __simd_vqrdmlsh_lane_s32(a,b,c,d)
#define vqrdmlsh_laneq_s32(a,b,c,d) __simd_vqrdmlsh_laneq_s32(a,b,c,d)
#define vqrdmlsh_n_s32(a,b,c) __simd_vqrdmlsh_n_s32(a,b,c)
#define vqrdmlshq_lane_s16(a,b,c,d) __simd_vqrdmlshq_lane_s16(a,b,c,d)
#define vqrdmlshq_laneq_s16(a,b,c,d) __simd_vqrdmlshq_laneq_s16(a,b,c,d)
#define vqrdmlshq_n_s16(a,b,c) __simd_vqrdmlshq_n_s16(a,b,c)
#define vqrdmlshq_lane_s32(a,b,c,d) __simd_vqrdmlshq_lane_s32(a,b,c,d)
#define vqrdmlshq_laneq_s32(a,b,c,d) __simd_vqrdmlshq_laneq_s32(a,b,c,d)
#define vqrdmlshq_n_s32(a,b,c) __simd_vqrdmlshq_n_s32(a,b,c)
#define vqrdmlshh_lane_s16(a,b,c,d) __simd_vqrdmlshh_lane_s16(a,b,c,d)
#define vqrdmlshh_laneq_s16(a,b,c,d) __simd_vqrdmlshh_laneq_s16(a,b,c,d)
#define vqrdmlshs_lane_s32(a,b,c,d) __simd_vqrdmlshs_lane_s32(a,b,c,d)
#define vqrdmlshs_laneq_s32(a,b,c,d) __simd_vqrdmlshs_laneq_s32(a,b,c,d)


/* These legacy ARM32 NEON instructions have no ARM64 equivalent. */
#define vmla_f32(a,b,c) vadd_f32(a, vmul_f32(b, c))
#define vmlaq_f32(a,b,c) vaddq_f32(a, vmulq_f32(b, c))
#define vmlaq_f64(a,b,c) vaddq_f64(a, vmulq_f64(b, c))
#define vmls_f32(a,b,c) vsub_f32(a, vmul_f32(b, c))
#define vmlsq_f32(a,b,c) vsubq_f32(a, vmulq_f32(b, c))
#define vmlsq_f64(a,b,c) vsubq_f64(a, vmulq_f64(b, c))
#define vmla_lane_f32(a,b,c,d) vadd_f32(a, vmul_lane_f32(b, c, d))
#define vmla_laneq_f32(a,b,c,d) vadd_f32(a, vmul_laneq_f32(b, c, d))
#define vmlaq_lane_f32(a,b,c,d) vaddq_f32(a, vmulq_lane_f32(b, c, d))
#define vmlaq_laneq_f32(a,b,c,d) vaddq_f32(a, vmulq_laneq_f32(b, c, d))
#define vmls_lane_f32(a,b,c,d) vsub_f32(a, vmul_lane_f32(b, c, d))
#define vmls_laneq_f32(a,b,c,d) vsub_f32(a, vmul_laneq_f32(b, c, d))
#define vmlsq_lane_f32(a,b,c,d) vsubq_f32(a, vmulq_lane_f32(b, c, d))
#define vmlsq_laneq_f32(a,b,c,d) vsubq_f32(a, vmulq_laneq_f32(b, c, d))
#define vmla_n_f32(a,b,c) vadd_f32(a, vmul_n_f32(b, c))
#define vmlaq_n_f32(a,b,c) vaddq_f32(a, vmulq_n_f32(b, c))
#define vmls_n_f32(a,b,c) vsub_f32(a, vmul_n_f32(b, c))
#define vmlsq_n_f32(a,b,c) vsubq_f32(a, vmulq_n_f32(b, c))

/* Single 64x1 operations may not correspond to unique neon instructions */
#define vdup_n_u64(a) vcreate_u64(a)
#define vdup_n_s64(a) vcreate_s64(a)
#define vdup_lane_u64(a,b) vdup_n_u64(vget_lane_u64(a, b))
#define vdup_lane_s64(a,b) vdup_n_s64(vget_lane_s64(a, b))
#define vdup_laneq_u64(a,b) vdup_n_u64(vgetq_lane_u64(a, b))
#define vdup_laneq_s64(a,b) vdup_n_s64(vgetq_lane_s64(a, b))

#define vget_lane_f64(a,b) (a)
#define vdup_n_f64(a) (a)
#define vdup_lane_f64(a,b) (a)
#define vdupq_lane_f64(a,b) vdupq_n_f64(a)

#define vld2_s64 vld1_s64_x2
#define vld2_u64 vld1_u64_x2
#define vld2_f64 vld1_f64_x2
#define vld3_s64 vld1_s64_x3
#define vld3_u64 vld1_u64_x3
#define vld3_f64 vld1_f64_x3
#define vld4_s64 vld1_s64_x4
#define vld4_u64 vld1_u64_x4
#define vld4_f64 vld1_f64_x4
#define vst2_s64 vst1_s64_x2
#define vst2_u64 vst1_u64_x2
#define vst2_f64 vst1_f64_x2
#define vst3_s64 vst1_s64_x3
#define vst3_u64 vst1_u64_x3
#define vst3_f64 vst1_f64_x3
#define vst4_s64 vst1_s64_x4
#define vst4_u64 vst1_u64_x4
#define vst4_f64 vst1_f64_x4

/* Scalar vdup instructions essentially the same as vget. */
#define vdupb_lane_s8 vget_lane_s8
#define vdupb_lane_u8 vget_lane_u8
#define vdupb_lane_p8 vget_lane_p8
#define vduph_lane_s16 vget_lane_s16
#define vduph_lane_u16 vget_lane_u16
#define vduph_lane_p16 vget_lane_p16
#define vdups_lane_s32 vget_lane_s32
#define vdups_lane_u32 vget_lane_u32
#define vdups_lane_f32 vget_lane_f32
#define vdupd_lane_s64 vget_lane_s64
#define vdupd_lane_u64 vget_lane_u64
#define vdupd_lane_f64 vget_lane_f64
#define vdupb_laneq_s8 vgetq_lane_s8
#define vdupb_laneq_u8 vgetq_lane_u8
#define vdupb_laneq_p8 vgetq_lane_p8
#define vduph_laneq_s16 vgetq_lane_s16
#define vduph_laneq_u16 vgetq_lane_u16
#define vduph_laneq_p16 vgetq_lane_p16
#define vdups_laneq_s32 vgetq_lane_s32
#define vdups_laneq_u32 vgetq_lane_u32
#define vdups_laneq_f32 vgetq_lane_f32
#define vdupd_laneq_s64 vgetq_lane_s64
#define vdupd_laneq_u64 vgetq_lane_u64
#define vdupd_laneq_f64 vgetq_lane_f64

/* mov_n essentially the same as dup_n. */
#define vmov_n_s8 vdup_n_s8
#define vmovq_n_s8 vdupq_n_s8
#define vmov_n_s16 vdup_n_s16
#define vmovq_n_s16 vdupq_n_s16
#define vmov_n_s32 vdup_n_s32
#define vmovq_n_s32 vdupq_n_s32
#define vmov_n_s64 vdup_n_s64
#define vmovq_n_s64 vdupq_n_s64
#define vmov_n_u8 vdup_n_u8
#define vmovq_n_u8 vdupq_n_u8
#define vmov_n_u16 vdup_n_u16
#define vmovq_n_u16 vdupq_n_u16
#define vmov_n_u32 vdup_n_u32
#define vmovq_n_u32 vdupq_n_u32
#define vmov_n_u64 vdup_n_u64
#define vmovq_n_u64 vdupq_n_u64
#define vmov_n_p8 vdup_n_p8
#define vmovq_n_p8 vdupq_n_p8
#define vmov_n_p16 vdup_n_p16
#define vmovq_n_p16 vdupq_n_p16
#define vmov_n_f32 vdup_n_f32
#define vmovq_n_f32 vdupq_n_f32
#define vmov_n_f64 vdup_n_f64
#define vmovq_n_f64 vdupq_n_f64

/* Exceptions to naming convention. */
#define vpmaxqd_f64 vpmaxd_f64
#define vpmaxnmqd_f64 vpmaxnmd_f64
#define vpminqd_f64 vpmind_f64
#define vpminnmqd_f64 vpminnmd_f64

/* Degenerate cases of across vector operations are really just pairs. */
#define vmaxv_s32(a) vget_lane_s32(vpmax_s32(a,vcreate_s32(0)),0)
#define vmaxv_u32(a) vget_lane_u32(vpmax_u32(a,vcreate_u32(0)),0)
#define vmaxv_f32(a) vget_lane_f32(vpmax_f32(a,vcreate_f32(0)),0)
#define vmaxnmv_f32(a) vpmaxnms_f32(a)
#define vmaxvq_f64(a) vpmaxqd_f64(a)
#define vmaxnmvq_f64(a) vpmaxnmqd_f64(a)
#define vminv_s32(a) vget_lane_s32(vpmin_s32(a,vcreate_s32(0)),0)
#define vminv_u32(a) vget_lane_u32(vpmin_u32(a,vcreate_u32(0)),0)
#define vminv_f32(a) vget_lane_f32(vpmin_f32(a,vcreate_f32(0)),0)
#define vminnmv_f32(a) vpminnms_f32(a)
#define vminvq_f64(a) vpminqd_f64(a)
#define vminnmvq_f64(a) vpminnmqd_f64(a)
#define vaddv_s32(a) vget_lane_s32(vpadd_s32(a,vcreate_s32(0)),0)
#define vaddv_u32(a) vget_lane_u32(vpadd_u32(a,vcreate_u32(0)),0)
#define vaddv_f32(a) vpadds_f32(a)
#define vaddvq_s64(a) vpaddd_s64(a)
#define vaddvq_u64(a) vpaddd_u64(a)
#define vaddvq_f32(a) vpadds_f32(vget_low_f32(vpaddq_f32(a,vdupq_n_f32(0.0))))
#define vaddlv_s32(a) vget_lane_s64(vpaddl_s32(a),0)
#define vaddlv_u32(a) vget_lane_u64(vpaddl_u32(a),0)
#define vaddvq_f64(a) vpaddd_f64(a)

#define vaesdq_u8(a,b) __simd_vaesdq_u8(a,b)
#define vaeseq_u8(a,b) __simd_vaeseq_u8(a,b)
#define vaesimcq_u8(a) __simd_vaesimcq_u8(a)
#define vaesmcq_u8(a) __simd_vaesmcq_u8(a)
#define vsha1cq_u32(a,b,c) __simd_vsha1cq_u32(a,b,c)
#define vsha1h_u32(a) __simd_vsha1h_u32(a)
#define vsha1mq_u32(a,b,c) __simd_vsha1mq_u32(a,b,c)
#define vsha1pq_u32(a,b,c) __simd_vsha1pq_u32(a,b,c)
#define vsha1su0q_u32(a,b,c) __simd_vsha1su0q_u32(a,b,c)
#define vsha1su1q_u32(a,b) __simd_vsha1su1q_u32(a,b)
#define vsha256hq_u32(a,b,c) __simd_vsha256hq_u32(a,b,c)
#define vsha256h2q_u32(a,b,c) __simd_vsha256h2q_u32(a,b,c)
#define vsha256su0q_u32(a,b) __simd_vsha256su0q_u32(a,b)
#define vsha256su1q_u32(a,b,c) __simd_vsha256su1q_u32(a,b,c)
  
#ifdef __cplusplus
}
#endif

#endif /* __LLONG_BIT */

#ifdef __ghs__
#pragma ghs endnowarning
#pragma ghs endnomisra
#endif

#endif