#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)))
#define __NEONALIGN256 __attribute__((aligned(4)))
#else
#define __NEONALIGN64 __attribute__((aligned(8)))
#if (__ghs_alignment < 16)
#define __NEONALIGN128 __attribute__((aligned(8)))
#define __NEONALIGN256 __attribute__((aligned(8)))
#else
#define __NEONALIGN128 __attribute__((aligned(16)))
#if (__ghs_alignment < 32)
#define __NEONALIGN256 __attribute__((aligned(16)))
#else
#define __NEONALIGN256 __attribute__((aligned(32)))
#endif
#endif
#endif


typedef float float32_t;
typedef uint8_t poly8_t;
typedef uint16_t poly16_t;
typedef uint16_t float16_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_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_uint16_t float16x4_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_poly8_t poly8x16_t;
typedef struct __simd128_poly16_t poly16x8_t;

struct __simd_array2_int8x8_t { struct __simd64_int8_t val[2]; } __NEONALIGN128;
struct __simd_array3_int8x8_t { struct __simd64_int8_t val[3]; } __NEONALIGN64;
struct __simd_array4_int8x8_t { struct __simd64_int8_t val[4]; } __NEONALIGN128;
struct __simd_array2_int8x16_t { struct __simd128_int8_t val[2]; } __NEONALIGN256;
struct __simd_array3_int8x16_t { struct __simd128_int8_t val[3]; } __NEONALIGN128;
struct __simd_array4_int8x16_t { struct __simd128_int8_t val[4]; } __NEONALIGN256;
struct __simd_array2_int16x4_t { struct __simd64_int16_t val[2]; } __NEONALIGN128;
struct __simd_array3_int16x4_t { struct __simd64_int16_t val[3]; } __NEONALIGN64;
struct __simd_array4_int16x4_t { struct __simd64_int16_t val[4]; } __NEONALIGN128;
struct __simd_array2_int16x8_t { struct __simd128_int16_t val[2]; } __NEONALIGN256;
struct __simd_array3_int16x8_t { struct __simd128_int16_t val[3]; } __NEONALIGN128;
struct __simd_array4_int16x8_t { struct __simd128_int16_t val[4]; } __NEONALIGN256;
struct __simd_array2_int32x2_t { struct __simd64_int32_t val[2]; } __NEONALIGN128;
struct __simd_array3_int32x2_t { struct __simd64_int32_t val[3]; } __NEONALIGN64;
struct __simd_array4_int32x2_t { struct __simd64_int32_t val[4]; } __NEONALIGN128;
struct __simd_array2_int32x4_t { struct __simd128_int32_t val[2]; } __NEONALIGN256;
struct __simd_array3_int32x4_t { struct __simd128_int32_t val[3]; } __NEONALIGN128;
struct __simd_array4_int32x4_t { struct __simd128_int32_t val[4]; } __NEONALIGN256;
struct __simd_array2_int64x1_t { struct __simd64_int64_t val[2]; } __NEONALIGN128;
struct __simd_array3_int64x1_t { struct __simd64_int64_t val[3]; } __NEONALIGN64;
struct __simd_array4_int64x1_t { struct __simd64_int64_t val[4]; } __NEONALIGN128;
struct __simd_array2_int64x2_t { struct __simd128_int64_t val[2]; } __NEONALIGN256;
struct __simd_array3_int64x2_t { struct __simd128_int64_t val[3]; } __NEONALIGN128;
struct __simd_array4_int64x2_t { struct __simd128_int64_t val[4]; } __NEONALIGN256;
struct __simd_array2_uint8x8_t { struct __simd64_uint8_t val[2]; } __NEONALIGN128;
struct __simd_array3_uint8x8_t { struct __simd64_uint8_t val[3]; } __NEONALIGN64;
struct __simd_array4_uint8x8_t { struct __simd64_uint8_t val[4]; } __NEONALIGN128;
struct __simd_array2_uint8x16_t { struct __simd128_uint8_t val[2]; } __NEONALIGN256;
struct __simd_array3_uint8x16_t { struct __simd128_uint8_t val[3]; } __NEONALIGN128;
struct __simd_array4_uint8x16_t { struct __simd128_uint8_t val[4]; } __NEONALIGN256;
struct __simd_array2_uint16x4_t { struct __simd64_uint16_t val[2]; } __NEONALIGN128;
struct __simd_array3_uint16x4_t { struct __simd64_uint16_t val[3]; } __NEONALIGN64;
struct __simd_array4_uint16x4_t { struct __simd64_uint16_t val[4]; } __NEONALIGN128;
struct __simd_array2_uint16x8_t { struct __simd128_uint16_t val[2]; } __NEONALIGN256;
struct __simd_array3_uint16x8_t { struct __simd128_uint16_t val[3]; } __NEONALIGN128;
struct __simd_array4_uint16x8_t { struct __simd128_uint16_t val[4]; } __NEONALIGN256;
struct __simd_array2_uint32x2_t { struct __simd64_uint32_t val[2]; } __NEONALIGN128;
struct __simd_array3_uint32x2_t { struct __simd64_uint32_t val[3]; } __NEONALIGN64;
struct __simd_array4_uint32x2_t { struct __simd64_uint32_t val[4]; } __NEONALIGN128;
struct __simd_array2_uint32x4_t { struct __simd128_uint32_t val[2]; } __NEONALIGN256;
struct __simd_array3_uint32x4_t { struct __simd128_uint32_t val[3]; } __NEONALIGN128;
struct __simd_array4_uint32x4_t { struct __simd128_uint32_t val[4]; } __NEONALIGN256;
struct __simd_array2_uint64x1_t { struct __simd64_uint64_t val[2]; } __NEONALIGN128;
struct __simd_array3_uint64x1_t { struct __simd64_uint64_t val[3]; } __NEONALIGN64;
struct __simd_array4_uint64x1_t { struct __simd64_uint64_t val[4]; } __NEONALIGN128;
struct __simd_array2_uint64x2_t { struct __simd128_uint64_t val[2]; } __NEONALIGN256;
struct __simd_array3_uint64x2_t { struct __simd128_uint64_t val[3]; } __NEONALIGN128;
struct __simd_array4_uint64x2_t { struct __simd128_uint64_t val[4]; } __NEONALIGN256;
struct __simd_array2_float32x2_t { struct __simd64_float32_t val[2]; } __NEONALIGN128;
struct __simd_array3_float32x2_t { struct __simd64_float32_t val[3]; } __NEONALIGN64;
struct __simd_array4_float32x2_t { struct __simd64_float32_t val[4]; } __NEONALIGN128;
struct __simd_array2_float32x4_t { struct __simd128_float32_t val[2]; } __NEONALIGN256;
struct __simd_array3_float32x4_t { struct __simd128_float32_t val[3]; } __NEONALIGN128;
struct __simd_array4_float32x4_t { struct __simd128_float32_t val[4]; } __NEONALIGN256;
struct __simd_array2_poly8x8_t { struct __simd64_poly8_t val[2]; } __NEONALIGN128;
struct __simd_array3_poly8x8_t { struct __simd64_poly8_t val[3]; } __NEONALIGN64;
struct __simd_array4_poly8x8_t { struct __simd64_poly8_t val[4]; } __NEONALIGN128;
struct __simd_array2_poly8x16_t { struct __simd128_poly8_t val[2]; } __NEONALIGN256;
struct __simd_array3_poly8x16_t { struct __simd128_poly8_t val[3]; } __NEONALIGN128;
struct __simd_array4_poly8x16_t { struct __simd128_poly8_t val[4]; } __NEONALIGN256;
struct __simd_array2_poly16x4_t { struct __simd64_poly16_t val[2]; } __NEONALIGN128;
struct __simd_array3_poly16x4_t { struct __simd64_poly16_t val[3]; } __NEONALIGN64;
struct __simd_array4_poly16x4_t { struct __simd64_poly16_t val[4]; } __NEONALIGN128;
struct __simd_array2_poly16x8_t { struct __simd128_poly16_t val[2]; } __NEONALIGN256;
struct __simd_array3_poly16x8_t { struct __simd128_poly16_t val[3]; } __NEONALIGN128;
struct __simd_array4_poly16x8_t { struct __simd128_poly16_t val[4]; } __NEONALIGN256;

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_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_SCOPE[sizeof(::__simd64_int8_t)];
#endif


int8x8_t __simd_vhadd_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vhaddq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vhadd_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vhaddq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vhadd_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vhaddq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vhadd_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vhaddq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vhadd_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vhaddq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vhadd_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vhaddq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vhsub_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vhsubq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vhsub_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vhsubq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vhsub_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vhsubq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vhsub_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vhsubq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vhsub_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vhsubq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vhsub_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vhsubq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vqadd_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vqaddq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vqadd_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vqaddq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vqadd_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vqaddq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vqadd_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vqaddq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vqadd_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqaddq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vqadd_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vqaddq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int64x1_t __simd_vqadd_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
int64x2_t __simd_vqaddq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint64x1_t __simd_vqadd_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
uint64x2_t __simd_vqaddq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int8x8_t __simd_vrhadd_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vrhaddq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vrhadd_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vrhaddq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vrhadd_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vrhaddq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vrhadd_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vrhaddq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vrhadd_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vrhaddq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vrhadd_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vrhaddq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vqsub_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vqsubq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vqsub_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vqsubq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vqsub_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vqsubq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vqsub_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vqsubq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vqsub_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqsubq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vqsub_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vqsubq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int64x1_t __simd_vqsub_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
int64x2_t __simd_vqsubq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint64x1_t __simd_vqsub_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
uint64x2_t __simd_vqsubq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vcgt_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcgtq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vcgt_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcgtq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vcgt_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcgtq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vcgt_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcgtq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vcgt_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgtq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcgt_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgtq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vcge_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcgeq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vcge_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcgeq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vcge_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcgeq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vcge_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcgeq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vcge_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgeq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcge_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgeq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vmax_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vmaxq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vmax_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vmaxq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vmax_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vmaxq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vmax_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vmaxq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vmax_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vmaxq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vmax_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vmaxq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vmin_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vminq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vmin_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vminq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vmin_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vminq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vmin_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vminq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vmin_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vminq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vmin_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vminq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vabd_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vabdq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vabd_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vabdq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vabd_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vabdq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vabd_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vabdq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vabd_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vabdq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vabd_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vabdq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vaba_s8(int8x8_t Dd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vabaq_s8(int8x16_t Qd, int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vaba_u8(uint8x8_t Dd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vabaq_u8(uint8x16_t Qd, uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vaba_s16(int16x4_t Dd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vabaq_s16(int16x8_t Qd, int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vaba_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vabaq_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vaba_s32(int32x2_t Dd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vabaq_s32(int32x4_t Qd, int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vaba_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vabaq_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vadd_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vadd_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vaddq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vaddq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vadd_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vadd_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vaddq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vaddq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vadd_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vadd_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vaddq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vaddq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint64x1_t __simd_vadd_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vadd_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint64x2_t __simd_vaddq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vaddq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vsub_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vsub_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vsubq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vsubq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vsub_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vsub_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vsubq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vsubq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vsub_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vsub_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vsubq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vsubq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint64x1_t __simd_vsub_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vsub_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint64x2_t __simd_vsubq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vsubq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vtst_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtst_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vtstq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
uint8x16_t __simd_vtstq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vtst_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x4_t __simd_vtst_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vtstq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
uint16x8_t __simd_vtstq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vtst_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vtst_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vtstq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
uint32x4_t __simd_vtstq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vtst_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vtstq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vceq_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vceq_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vceqq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
uint8x16_t __simd_vceqq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vceq_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
uint16x4_t __simd_vceq_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vceqq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
uint16x8_t __simd_vceqq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vceq_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vceq_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vceqq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
uint32x4_t __simd_vceqq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vceq_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vceqq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vmla_u8(uint8x8_t Dd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vmla_s8(int8x8_t Dd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vmlaq_u8(uint8x16_t Qd, uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vmlaq_s8(int8x16_t Qd, int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vmla_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vmla_s16(int16x4_t Dd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vmlaq_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vmlaq_s16(int16x8_t Qd, int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vmla_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vmla_s32(int32x2_t Dd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vmlaq_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vmlaq_s32(int32x4_t Qd, int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vmls_u8(uint8x8_t Dd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vmls_s8(int8x8_t Dd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vmlsq_u8(uint8x16_t Qd, uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vmlsq_s8(int8x16_t Qd, int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vmls_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vmls_s16(int16x4_t Dd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vmlsq_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vmlsq_s16(int16x8_t Qd, int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vmls_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vmls_s32(int32x2_t Dd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vmlsq_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vmlsq_s32(int32x4_t Qd, int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vmul_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vmul_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vmulq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vmulq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vmul_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vmul_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vmulq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vmulq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vmul_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vmul_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vmulq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vmulq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
poly8x8_t __simd_vmul_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
poly8x16_t __simd_vmulq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vqdmulh_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vqdmulhq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vqdmulh_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqdmulhq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vqrdmulh_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vqrdmulhq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vqrdmulh_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqrdmulhq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vpmax_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vpmax_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int16x4_t __simd_vpmax_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x4_t __simd_vpmax_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int32x2_t __simd_vpmax_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vpmax_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int8x8_t __simd_vpmin_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vpmin_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int16x4_t __simd_vpmin_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint16x4_t __simd_vpmin_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int32x2_t __simd_vpmin_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vpmin_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
uint8x8_t __simd_vpadd_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vpadd_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vpadd_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vpadd_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vpadd_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vpadd_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int8x8_t __simd_vshl_s8(int8x8_t Dm, int8x8_t Dn) __NEONSAFE;
int8x16_t __simd_vshlq_s8(int8x16_t Qm, int8x16_t Qn) __NEONSAFE;
uint8x8_t __simd_vshl_u8(uint8x8_t Dm, int8x8_t Dn) __NEONSAFE;
uint8x16_t __simd_vshlq_u8(uint8x16_t Qm, int8x16_t Qn) __NEONSAFE;
int16x4_t __simd_vshl_s16(int16x4_t Dm, int16x4_t Dn) __NEONSAFE;
int16x8_t __simd_vshlq_s16(int16x8_t Qm, int16x8_t Qn) __NEONSAFE;
uint16x4_t __simd_vshl_u16(uint16x4_t Dm, int16x4_t Dn) __NEONSAFE;
uint16x8_t __simd_vshlq_u16(uint16x8_t Qm, int16x8_t Qn) __NEONSAFE;
int32x2_t __simd_vshl_s32(int32x2_t Dm, int32x2_t Dn) __NEONSAFE;
int32x4_t __simd_vshlq_s32(int32x4_t Qm, int32x4_t Qn) __NEONSAFE;
uint32x2_t __simd_vshl_u32(uint32x2_t Dm, int32x2_t Dn) __NEONSAFE;
uint32x4_t __simd_vshlq_u32(uint32x4_t Qm, int32x4_t Qn) __NEONSAFE;
int64x1_t __simd_vshl_s64(int64x1_t Dm, int64x1_t Dn) __NEONSAFE;
int64x2_t __simd_vshlq_s64(int64x2_t Qm, int64x2_t Qn) __NEONSAFE;
uint64x1_t __simd_vshl_u64(uint64x1_t Dm, int64x1_t Dn) __NEONSAFE;
uint64x2_t __simd_vshlq_u64(uint64x2_t Qm, int64x2_t Qn) __NEONSAFE;
int8x8_t __simd_vqshl_s8(int8x8_t Dm, int8x8_t Dn) __NEONSAFE;
int8x16_t __simd_vqshlq_s8(int8x16_t Qm, int8x16_t Qn) __NEONSAFE;
uint8x8_t __simd_vqshl_u8(uint8x8_t Dm, int8x8_t Dn) __NEONSAFE;
uint8x16_t __simd_vqshlq_u8(uint8x16_t Qm, int8x16_t Qn) __NEONSAFE;
int16x4_t __simd_vqshl_s16(int16x4_t Dm, int16x4_t Dn) __NEONSAFE;
int16x8_t __simd_vqshlq_s16(int16x8_t Qm, int16x8_t Qn) __NEONSAFE;
uint16x4_t __simd_vqshl_u16(uint16x4_t Dm, int16x4_t Dn) __NEONSAFE;
uint16x8_t __simd_vqshlq_u16(uint16x8_t Qm, int16x8_t Qn) __NEONSAFE;
int32x2_t __simd_vqshl_s32(int32x2_t Dm, int32x2_t Dn) __NEONSAFE;
int32x4_t __simd_vqshlq_s32(int32x4_t Qm, int32x4_t Qn) __NEONSAFE;
uint32x2_t __simd_vqshl_u32(uint32x2_t Dm, int32x2_t Dn) __NEONSAFE;
uint32x4_t __simd_vqshlq_u32(uint32x4_t Qm, int32x4_t Qn) __NEONSAFE;
int64x1_t __simd_vqshl_s64(int64x1_t Dm, int64x1_t Dn) __NEONSAFE;
int64x2_t __simd_vqshlq_s64(int64x2_t Qm, int64x2_t Qn) __NEONSAFE;
uint64x1_t __simd_vqshl_u64(uint64x1_t Dm, int64x1_t Dn) __NEONSAFE;
uint64x2_t __simd_vqshlq_u64(uint64x2_t Qm, int64x2_t Qn) __NEONSAFE;
int8x8_t __simd_vrshl_s8(int8x8_t Dm, int8x8_t Dn) __NEONSAFE;
int8x16_t __simd_vrshlq_s8(int8x16_t Qm, int8x16_t Qn) __NEONSAFE;
uint8x8_t __simd_vrshl_u8(uint8x8_t Dm, int8x8_t Dn) __NEONSAFE;
uint8x16_t __simd_vrshlq_u8(uint8x16_t Qm, int8x16_t Qn) __NEONSAFE;
int16x4_t __simd_vrshl_s16(int16x4_t Dm, int16x4_t Dn) __NEONSAFE;
int16x8_t __simd_vrshlq_s16(int16x8_t Qm, int16x8_t Qn) __NEONSAFE;
uint16x4_t __simd_vrshl_u16(uint16x4_t Dm, int16x4_t Dn) __NEONSAFE;
uint16x8_t __simd_vrshlq_u16(uint16x8_t Qm, int16x8_t Qn) __NEONSAFE;
int32x2_t __simd_vrshl_s32(int32x2_t Dm, int32x2_t Dn) __NEONSAFE;
int32x4_t __simd_vrshlq_s32(int32x4_t Qm, int32x4_t Qn) __NEONSAFE;
uint32x2_t __simd_vrshl_u32(uint32x2_t Dm, int32x2_t Dn) __NEONSAFE;
uint32x4_t __simd_vrshlq_u32(uint32x4_t Qm, int32x4_t Qn) __NEONSAFE;
int64x1_t __simd_vrshl_s64(int64x1_t Dm, int64x1_t Dn) __NEONSAFE;
int64x2_t __simd_vrshlq_s64(int64x2_t Qm, int64x2_t Qn) __NEONSAFE;
uint64x1_t __simd_vrshl_u64(uint64x1_t Dm, int64x1_t Dn) __NEONSAFE;
uint64x2_t __simd_vrshlq_u64(uint64x2_t Qm, int64x2_t Qn) __NEONSAFE;
int8x8_t __simd_vqrshl_s8(int8x8_t Dm, int8x8_t Dn) __NEONSAFE;
int8x16_t __simd_vqrshlq_s8(int8x16_t Qm, int8x16_t Qn) __NEONSAFE;
uint8x8_t __simd_vqrshl_u8(uint8x8_t Dm, int8x8_t Dn) __NEONSAFE;
uint8x16_t __simd_vqrshlq_u8(uint8x16_t Qm, int8x16_t Qn) __NEONSAFE;
int16x4_t __simd_vqrshl_s16(int16x4_t Dm, int16x4_t Dn) __NEONSAFE;
int16x8_t __simd_vqrshlq_s16(int16x8_t Qm, int16x8_t Qn) __NEONSAFE;
uint16x4_t __simd_vqrshl_u16(uint16x4_t Dm, int16x4_t Dn) __NEONSAFE;
uint16x8_t __simd_vqrshlq_u16(uint16x8_t Qm, int16x8_t Qn) __NEONSAFE;
int32x2_t __simd_vqrshl_s32(int32x2_t Dm, int32x2_t Dn) __NEONSAFE;
int32x4_t __simd_vqrshlq_s32(int32x4_t Qm, int32x4_t Qn) __NEONSAFE;
uint32x2_t __simd_vqrshl_u32(uint32x2_t Dm, int32x2_t Dn) __NEONSAFE;
uint32x4_t __simd_vqrshlq_u32(uint32x4_t Qm, int32x4_t Qn) __NEONSAFE;
int64x1_t __simd_vqrshl_s64(int64x1_t Dm, int64x1_t Dn) __NEONSAFE;
int64x2_t __simd_vqrshlq_s64(int64x2_t Qm, int64x2_t Qn) __NEONSAFE;
uint64x1_t __simd_vqrshl_u64(uint64x1_t Dm, int64x1_t Dn) __NEONSAFE;
uint64x2_t __simd_vqrshlq_u64(uint64x2_t Qm, int64x2_t Qn) __NEONSAFE;
uint8x8_t __simd_vand_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vand_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vand_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vand_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vand_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vand_p16(poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vand_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vand_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vand_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vand_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vand_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vandq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vandq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vandq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vandq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vandq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vandq_p16(poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vandq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vandq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vandq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vandq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vandq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vbic_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vbic_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vbic_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vbic_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vbic_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vbic_p16(poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vbic_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vbic_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vbic_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vbic_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vbic_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vbicq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vbicq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vbicq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vbicq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vbicq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vbicq_p16(poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vbicq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vbicq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vbicq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vbicq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vbicq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vorr_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vorr_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vorr_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vorr_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vorr_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vorr_p16(poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vorr_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vorr_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vorr_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vorr_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vorr_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vorrq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vorrq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vorrq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vorrq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vorrq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vorrq_p16(poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vorrq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vorrq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vorrq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vorrq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vorrq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vorn_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vorn_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vorn_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vorn_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vorn_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vorn_p16(poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vorn_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vorn_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vorn_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vorn_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vorn_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vornq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vornq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vornq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vornq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vornq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vornq_p16(poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vornq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vornq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vornq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vornq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vornq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_veor_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_veor_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_veor_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_veor_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_veor_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_veor_p16(poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_veor_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_veor_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_veor_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_veor_u64(uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_veor_s64(int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_veorq_u8(uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_veorq_s8(int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_veorq_p8(poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_veorq_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_veorq_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_veorq_p16(poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_veorq_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_veorq_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_veorq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_veorq_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_veorq_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vbsl_u8(uint8x8_t Dd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vbsl_s8(uint8x8_t Dd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vbsl_p8(uint8x8_t Dd, poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vbsl_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vbsl_s16(uint16x4_t Dd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vbsl_p16(uint16x4_t Dd, poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vbsl_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vbsl_s32(uint32x2_t Dd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vbsl_f32(uint32x2_t Dd, float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vbsl_u64(uint64x1_t Dd, uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vbsl_s64(uint64x1_t Dd, int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vbslq_u8(uint8x16_t Qd, uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vbslq_s8(uint8x16_t Qd, int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vbslq_p8(uint8x16_t Qd, poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vbslq_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vbslq_s16(uint16x8_t Qd, int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vbslq_p16(uint16x8_t Qd, poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vbslq_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vbslq_s32(uint32x4_t Qd, int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vbslq_f32(uint32x4_t Qd, float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vbslq_u64(uint64x2_t Qd, uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vbslq_s64(uint64x2_t Qd, int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vbit_u8(uint8x8_t Dd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vbit_s8(int8x8_t Dd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vbit_p8(poly8x8_t Dd, poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vbit_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vbit_s16(int16x4_t Dd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vbit_p16(poly16x4_t Dd, poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vbit_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vbit_s32(int32x2_t Dd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vbit_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vbit_u64(uint64x1_t Dd, uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vbit_s64(int64x1_t Dd, int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vbitq_u8(uint8x16_t Qd, uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vbitq_s8(int8x16_t Qd, int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vbitq_p8(poly8x16_t Qd, poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vbitq_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vbitq_s16(int16x8_t Qd, int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vbitq_p16(poly16x8_t Qd, poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vbitq_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vbitq_s32(int32x4_t Qd, int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vbitq_f32(float32x4_t Qd, float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vbitq_u64(uint64x2_t Qd, uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vbitq_s64(int64x2_t Qd, int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vbif_u8(uint8x8_t Dd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vbif_s8(int8x8_t Dd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vbif_p8(poly8x8_t Dd, poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vbif_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vbif_s16(int16x4_t Dd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vbif_p16(poly16x4_t Dd, poly16x4_t Dn, poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vbif_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vbif_s32(int32x2_t Dd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vbif_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vbif_u64(uint64x1_t Dd, uint64x1_t Dn, uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vbif_s64(int64x1_t Dd, int64x1_t Dn, int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vbifq_u8(uint8x16_t Qd, uint8x16_t Qn, uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vbifq_s8(int8x16_t Qd, int8x16_t Qn, int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vbifq_p8(poly8x16_t Qd, poly8x16_t Qn, poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vbifq_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vbifq_s16(int16x8_t Qd, int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vbifq_p16(poly16x8_t Qd, poly16x8_t Qn, poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vbifq_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vbifq_s32(int32x4_t Qd, int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vbifq_f32(float32x4_t Qd, float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vbifq_u64(uint64x2_t Qd, uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vbifq_s64(int64x2_t Qd, int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
float32x2_t __simd_vfma_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vfmaq_f32(float32x4_t Qd, float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vfms_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vfmsq_f32(float32x4_t Qd, float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vadd_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vaddq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vsub_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vsubq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vabd_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vabdq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vmla_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vmlaq_f32(float32x4_t Qd, float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vmls_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vmlsq_f32(float32x4_t Qd, float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vmul_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vmulq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vceq_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vceqq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcge_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgeq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcgt_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgtq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vacge_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vacgeq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vacgt_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vacgtq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vmax_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vmaxq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vmin_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vminq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrecps_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrecpsq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrsqrts_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrsqrtsq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vpadd_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vpmax_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vpmin_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vaddl_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vaddl_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vaddl_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vaddl_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vaddl_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vaddl_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vsubl_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vsubl_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vsubl_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vsubl_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vsubl_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vsubl_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vabal_s8(int16x8_t Qd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vabal_u8(uint16x8_t Qd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vabal_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vabal_u16(uint32x4_t Qd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vabal_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vabal_u32(uint64x2_t Qd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vabdl_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vabdl_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vabdl_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vabdl_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vabdl_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vabdl_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vmlal_s8(int16x8_t Qd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vmlal_u8(uint16x8_t Qd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vmlal_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vmlal_u16(uint32x4_t Qd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vmlal_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vmlal_u32(uint64x2_t Qd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqdmlal_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vqdmlal_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vmlsl_s8(int16x8_t Qd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vmlsl_u8(uint16x8_t Qd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vmlsl_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vmlsl_u16(uint32x4_t Qd, uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vmlsl_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vmlsl_u32(uint64x2_t Qd, uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqdmlsl_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vqdmlsl_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vmull_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vmull_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vmull_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vmull_u16(uint16x4_t Dn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vmull_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vmull_u32(uint32x2_t Dn, uint32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqdmull_s16(int16x4_t Dn, int16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vqdmull_s32(int32x2_t Dn, int32x2_t Dm) __NEONSAFE;
poly16x8_t __simd_vmull_p8(poly8x8_t Dn, poly8x8_t Dm) __NEONSAFE;
int16x8_t __simd_vaddw_s8(int16x8_t Qn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vaddw_u8(uint16x8_t Qn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vaddw_s16(int32x4_t Qn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vaddw_u16(uint32x4_t Qn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vaddw_s32(int64x2_t Qn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vaddw_u32(uint64x2_t Qn, uint32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vsubw_s8(int16x8_t Qn, int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vsubw_u8(uint16x8_t Qn, uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vsubw_s16(int32x4_t Qn, int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vsubw_u16(uint32x4_t Qn, uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vsubw_s32(int64x2_t Qn, int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vsubw_u32(uint64x2_t Qn, uint32x2_t Dm) __NEONSAFE;
uint8x8_t __simd_vaddhn_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int8x8_t __simd_vaddhn_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vaddhn_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vaddhn_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vaddhn_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int32x2_t __simd_vaddhn_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vraddhn_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int8x8_t __simd_vraddhn_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vraddhn_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vraddhn_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vraddhn_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int32x2_t __simd_vraddhn_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vsubhn_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int8x8_t __simd_vsubhn_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vsubhn_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vsubhn_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vsubhn_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int32x2_t __simd_vsubhn_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vrsubhn_u16(uint16x8_t Qn, uint16x8_t Qm) __NEONSAFE;
int8x8_t __simd_vrsubhn_s16(int16x8_t Qn, int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vrsubhn_u32(uint32x4_t Qn, uint32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vrsubhn_s32(int32x4_t Qn, int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vrsubhn_u64(uint64x2_t Qn, uint64x2_t Qm) __NEONSAFE;
int32x2_t __simd_vrsubhn_s64(int64x2_t Qn, int64x2_t Qm) __NEONSAFE;
uint16x4_t __simd_vmla_lane_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x4_t __simd_vmla_n_u16(uint16x4_t Dd, uint16x4_t Dn, uint16_t Rm) __NEONSAFE;
int16x4_t __simd_vmla_lane_s16(int16x4_t Dd, int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x4_t __simd_vmla_n_s16(int16x4_t Dd, int16x4_t Dn, int16_t Rm) __NEONSAFE;
uint16x8_t __simd_vmlaq_lane_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x8_t __simd_vmlaq_n_u16(uint16x8_t Qd, uint16x8_t Qn, uint16_t Rm) __NEONSAFE;
int16x8_t __simd_vmlaq_lane_s16(int16x8_t Qd, int16x8_t Qn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x8_t __simd_vmlaq_n_s16(int16x8_t Qd, int16x8_t Qn, int16_t Rm) __NEONSAFE;
uint32x2_t __simd_vmla_lane_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x2_t __simd_vmla_n_u32(uint32x2_t Dd, uint32x2_t Dn, uint32_t Rm) __NEONSAFE;
int32x2_t __simd_vmla_lane_s32(int32x2_t Dd, int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x2_t __simd_vmla_n_s32(int32x2_t Dd, int32x2_t Dn, int32_t Rm) __NEONSAFE;
uint32x4_t __simd_vmlaq_lane_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vmlaq_n_u32(uint32x4_t Qd, uint32x4_t Qn, uint32_t Rm) __NEONSAFE;
int32x4_t __simd_vmlaq_lane_s32(int32x4_t Qd, int32x4_t Qn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vmlaq_n_s32(int32x4_t Qd, int32x4_t Qn, int32_t Rm) __NEONSAFE;
float32x2_t __simd_vmla_lane_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x2_t __simd_vmla_n_f32(float32x2_t Dd, float32x2_t Dn, float32_t Rm) __NEONSAFE;
float32x4_t __simd_vmlaq_lane_f32(float32x4_t Qd, float32x4_t Qn, float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x4_t __simd_vmlaq_n_f32(float32x4_t Qd, float32x4_t Qn, float32_t Rm) __NEONSAFE;
uint16x4_t __simd_vmls_lane_u16(uint16x4_t Dd, uint16x4_t Dn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x4_t __simd_vmls_n_u16(uint16x4_t Dd, uint16x4_t Dn, uint16_t Rm) __NEONSAFE;
int16x4_t __simd_vmls_lane_s16(int16x4_t Dd, int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x4_t __simd_vmls_n_s16(int16x4_t Dd, int16x4_t Dn, int16_t Rm) __NEONSAFE;
uint16x8_t __simd_vmlsq_lane_u16(uint16x8_t Qd, uint16x8_t Qn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x8_t __simd_vmlsq_n_u16(uint16x8_t Qd, uint16x8_t Qn, uint16_t Rm) __NEONSAFE;
int16x8_t __simd_vmlsq_lane_s16(int16x8_t Qd, int16x8_t Qn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x8_t __simd_vmlsq_n_s16(int16x8_t Qd, int16x8_t Qn, int16_t Rm) __NEONSAFE;
uint32x2_t __simd_vmls_lane_u32(uint32x2_t Dd, uint32x2_t Dn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x2_t __simd_vmls_n_u32(uint32x2_t Dd, uint32x2_t Dn, uint32_t Rm) __NEONSAFE;
int32x2_t __simd_vmls_lane_s32(int32x2_t Dd, int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x2_t __simd_vmls_n_s32(int32x2_t Dd, int32x2_t Dn, int32_t Rm) __NEONSAFE;
uint32x4_t __simd_vmlsq_lane_u32(uint32x4_t Qd, uint32x4_t Qn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vmlsq_n_u32(uint32x4_t Qd, uint32x4_t Qn, uint32_t Rm) __NEONSAFE;
int32x4_t __simd_vmlsq_lane_s32(int32x4_t Qd, int32x4_t Qn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vmlsq_n_s32(int32x4_t Qd, int32x4_t Qn, int32_t Rm) __NEONSAFE;
float32x2_t __simd_vmls_lane_f32(float32x2_t Dd, float32x2_t Dn, float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x2_t __simd_vmls_n_f32(float32x2_t Dd, float32x2_t Dn, float32_t Rm) __NEONSAFE;
float32x4_t __simd_vmlsq_lane_f32(float32x4_t Qd, float32x4_t Qn, float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x4_t __simd_vmlsq_n_f32(float32x4_t Qd, float32x4_t Qn, float32_t Rm) __NEONSAFE;
uint16x4_t __simd_vmul_lane_u16(uint16x4_t Dn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x4_t __simd_vmul_n_u16(uint16x4_t Dn, uint16_t Rm) __NEONSAFE;
int16x4_t __simd_vmul_lane_s16(int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x4_t __simd_vmul_n_s16(int16x4_t Dn, int16_t Rm) __NEONSAFE;
uint16x8_t __simd_vmulq_lane_u16(uint16x8_t Qn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x8_t __simd_vmulq_n_u16(uint16x8_t Qn, uint16_t Rm) __NEONSAFE;
int16x8_t __simd_vmulq_lane_s16(int16x8_t Qn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x8_t __simd_vmulq_n_s16(int16x8_t Qn, int16_t Rm) __NEONSAFE;
uint32x2_t __simd_vmul_lane_u32(uint32x2_t Dn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x2_t __simd_vmul_n_u32(uint32x2_t Dn, uint32_t Rm) __NEONSAFE;
int32x2_t __simd_vmul_lane_s32(int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x2_t __simd_vmul_n_s32(int32x2_t Dn, int32_t Rm) __NEONSAFE;
uint32x4_t __simd_vmulq_lane_u32(uint32x4_t Qn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vmulq_n_u32(uint32x4_t Qn, uint32_t Rm) __NEONSAFE;
int32x4_t __simd_vmulq_lane_s32(int32x4_t Qn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vmulq_n_s32(int32x4_t Qn, int32_t Rm) __NEONSAFE;
float32x2_t __simd_vmul_lane_f32(float32x2_t Dn, float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x2_t __simd_vmul_n_f32(float32x2_t Dn, float32_t Rm) __NEONSAFE;
float32x4_t __simd_vmulq_lane_f32(float32x4_t Qn, float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x4_t __simd_vmulq_n_f32(float32x4_t Qn, float32_t Rm) __NEONSAFE;
int16x4_t __simd_vqdmulh_lane_s16(int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x4_t __simd_vqdmulh_n_s16(int16x4_t Dn, int16_t Rm) __NEONSAFE;
int16x8_t __simd_vqdmulhq_lane_s16(int16x8_t Qn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x8_t __simd_vqdmulhq_n_s16(int16x8_t Qn, int16_t Rm) __NEONSAFE;
int32x2_t __simd_vqdmulh_lane_s32(int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x2_t __simd_vqdmulh_n_s32(int32x2_t Dn, int32_t Rm) __NEONSAFE;
int32x4_t __simd_vqdmulhq_lane_s32(int32x4_t Qn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vqdmulhq_n_s32(int32x4_t Qn, int32_t Rm) __NEONSAFE;
int16x4_t __simd_vqrdmulh_lane_s16(int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x4_t __simd_vqrdmulh_n_s16(int16x4_t Dn, int16_t Rm) __NEONSAFE;
int16x8_t __simd_vqrdmulhq_lane_s16(int16x8_t Qn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x8_t __simd_vqrdmulhq_n_s16(int16x8_t Qn, int16_t Rm) __NEONSAFE;
int32x2_t __simd_vqrdmulh_lane_s32(int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x2_t __simd_vqrdmulh_n_s32(int32x2_t Dn, int32_t Rm) __NEONSAFE;
int32x4_t __simd_vqrdmulhq_lane_s32(int32x4_t Qn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vqrdmulhq_n_s32(int32x4_t Qn, int32_t Rm) __NEONSAFE;
int32x4_t __simd_vmlal_lane_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vmlal_n_s16(int32x4_t Qd, int16x4_t Dn, int16_t Rm) __NEONSAFE;
uint32x4_t __simd_vmlal_lane_u16(uint32x4_t Qd, uint16x4_t Dn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vmlal_n_u16(uint32x4_t Qd, uint16x4_t Dn, uint16_t Rm) __NEONSAFE;
int64x2_t __simd_vmlal_lane_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vmlal_n_s32(int64x2_t Qd, int32x2_t Dn, int32_t Rm) __NEONSAFE;
uint64x2_t __simd_vmlal_lane_u32(uint64x2_t Qd, uint32x2_t Dn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint64x2_t __simd_vmlal_n_u32(uint64x2_t Qd, uint32x2_t Dn, uint32_t Rm) __NEONSAFE;
int32x4_t __simd_vqdmlal_lane_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vqdmlal_n_s16(int32x4_t Qd, int16x4_t Dn, int16_t Rm) __NEONSAFE;
int64x2_t __simd_vqdmlal_lane_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vqdmlal_n_s32(int64x2_t Qd, int32x2_t Dn, int32_t Rm) __NEONSAFE;
int32x4_t __simd_vmlsl_lane_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vmlsl_n_s16(int32x4_t Qd, int16x4_t Dn, int16_t Rm) __NEONSAFE;
uint32x4_t __simd_vmlsl_lane_u16(uint32x4_t Qd, uint16x4_t Dn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vmlsl_n_u16(uint32x4_t Qd, uint16x4_t Dn, uint16_t Rm) __NEONSAFE;
int64x2_t __simd_vmlsl_lane_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vmlsl_n_s32(int64x2_t Qd, int32x2_t Dn, int32_t Rm) __NEONSAFE;
uint64x2_t __simd_vmlsl_lane_u32(uint64x2_t Qd, uint32x2_t Dn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint64x2_t __simd_vmlsl_n_u32(uint64x2_t Qd, uint32x2_t Dn, uint32_t Rm) __NEONSAFE;
int32x4_t __simd_vqdmlsl_lane_s16(int32x4_t Qd, int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vqdmlsl_n_s16(int32x4_t Qd, int16x4_t Dn, int16_t Rm) __NEONSAFE;
int64x2_t __simd_vqdmlsl_lane_s32(int64x2_t Qd, int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vqdmlsl_n_s32(int64x2_t Qd, int32x2_t Dn, int32_t Rm) __NEONSAFE;
int32x4_t __simd_vmull_lane_s16(int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vmull_n_s16(int16x4_t Dn, int16_t Rm) __NEONSAFE;
uint32x4_t __simd_vmull_lane_u16(uint16x4_t Dn, uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vmull_n_u16(uint16x4_t Dn, uint16_t Rm) __NEONSAFE;
int64x2_t __simd_vmull_lane_s32(int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vmull_n_s32(int32x2_t Dn, int32_t Rm) __NEONSAFE;
uint64x2_t __simd_vmull_lane_u32(uint32x2_t Dn, uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint64x2_t __simd_vmull_n_u32(uint32x2_t Dn, uint32_t Rm) __NEONSAFE;
int32x4_t __simd_vqdmull_lane_s16(int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vqdmull_n_s16(int16x4_t Dn, int16_t Rm) __NEONSAFE;
int64x2_t __simd_vqdmull_lane_s32(int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vqdmull_n_s32(int32x2_t Dn, int32_t Rm) __NEONSAFE;
int8x8_t __simd_vshr_n_s8(int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vshrq_n_s8(int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vshr_n_u8(uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vshrq_n_u8(uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vshr_n_s16(int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vshrq_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vshr_n_u16(uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vshrq_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vshr_n_s32(int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vshrq_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vshr_n_u32(uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vshrq_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vshr_n_s64(int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vshrq_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vshr_n_u64(uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vshrq_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vsra_n_s8(int8x8_t Dd, int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vsraq_n_s8(int8x16_t Qd, int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vsra_n_u8(uint8x8_t Dd, uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vsraq_n_u8(uint8x16_t Qd, uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vsra_n_s16(int16x4_t Dd, int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vsraq_n_s16(int16x8_t Qd, int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vsra_n_u16(uint16x4_t Dd, uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vsraq_n_u16(uint16x8_t Qd, uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vsra_n_s32(int32x2_t Dd, int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vsraq_n_s32(int32x4_t Qd, int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vsra_n_u32(uint32x2_t Dd, uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vsraq_n_u32(uint32x4_t Qd, uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vsra_n_s64(int64x1_t Dd, int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vsraq_n_s64(int64x2_t Qd, int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vsra_n_u64(uint64x1_t Dd, uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vsraq_n_u64(uint64x2_t Qd, uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vrshr_n_s8(int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vrshrq_n_s8(int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vrshr_n_u8(uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vrshrq_n_u8(uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vrshr_n_s16(int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vrshrq_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vrshr_n_u16(uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vrshrq_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vrshr_n_s32(int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vrshrq_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vrshr_n_u32(uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vrshrq_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vrshr_n_s64(int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vrshrq_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vrshr_n_u64(uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vrshrq_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vrsra_n_s8(int8x8_t Dd, int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vrsraq_n_s8(int8x16_t Qd, int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vrsra_n_u8(uint8x8_t Dd, uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vrsraq_n_u8(uint8x16_t Qd, uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vrsra_n_s16(int16x4_t Dd, int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vrsraq_n_s16(int16x8_t Qd, int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vrsra_n_u16(uint16x4_t Dd, uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vrsraq_n_u16(uint16x8_t Qd, uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vrsra_n_s32(int32x2_t Dd, int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vrsraq_n_s32(int32x4_t Qd, int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vrsra_n_u32(uint32x2_t Dd, uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vrsraq_n_u32(uint32x4_t Qd, uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vrsra_n_s64(int64x1_t Dd, int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vrsraq_n_s64(int64x2_t Qd, int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vrsra_n_u64(uint64x1_t Dd, uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vrsraq_n_u64(uint64x2_t Qd, uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vsri_n_u8(uint8x8_t Dd, uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vsri_n_s8(int8x8_t Dd, int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly8x8_t __simd_vsri_n_p8(poly8x8_t Dd, poly8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vsriq_n_u8(uint8x16_t Qd, uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vsriq_n_s8(int8x16_t Qd, int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
poly8x16_t __simd_vsriq_n_p8(poly8x16_t Qd, poly8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vsri_n_u16(uint16x4_t Dd, uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vsri_n_s16(int16x4_t Dd, int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly16x4_t __simd_vsri_n_p16(poly16x4_t Dd, poly16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vsriq_n_u16(uint16x8_t Qd, uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vsriq_n_s16(int16x8_t Qd, int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
poly16x8_t __simd_vsriq_n_p16(poly16x8_t Qd, poly16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vsri_n_u32(uint32x2_t Dd, uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vsri_n_s32(int32x2_t Dd, int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
float32x2_t __simd_vsri_n_f32(float32x2_t Dd, float32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vsriq_n_u32(uint32x4_t Qd, uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vsriq_n_s32(int32x4_t Qd, int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
float32x4_t __simd_vsriq_n_f32(float32x4_t Qd, float32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vsri_n_u64(uint64x1_t Dd, uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vsri_n_s64(int64x1_t Dd, int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vsriq_n_u64(uint64x2_t Qd, uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vsriq_n_s64(int64x2_t Qd, int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vshl_n_u8(uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vshl_n_s8(int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vshlq_n_u8(uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vshlq_n_s8(int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vshl_n_u16(uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vshl_n_s16(int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vshlq_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vshlq_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vshl_n_u32(uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vshl_n_s32(int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vshlq_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vshlq_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vshl_n_u64(uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vshl_n_s64(int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vshlq_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vshlq_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vsli_n_u8(uint8x8_t Dd, uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vsli_n_s8(int8x8_t Dd, int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly8x8_t __simd_vsli_n_p8(poly8x8_t Dd, poly8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vsliq_n_u8(uint8x16_t Qd, uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vsliq_n_s8(int8x16_t Qd, int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
poly8x16_t __simd_vsliq_n_p8(poly8x16_t Qd, poly8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vsli_n_u16(uint16x4_t Dd, uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vsli_n_s16(int16x4_t Dd, int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly16x4_t __simd_vsli_n_p16(poly16x4_t Dd, poly16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vsliq_n_u16(uint16x8_t Qd, uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vsliq_n_s16(int16x8_t Qd, int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
poly16x8_t __simd_vsliq_n_p16(poly16x8_t Qd, poly16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vsli_n_u32(uint32x2_t Dd, uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vsli_n_s32(int32x2_t Dd, int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
float32x2_t __simd_vsli_n_f32(float32x2_t Dd, float32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vsliq_n_u32(uint32x4_t Qd, uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vsliq_n_s32(int32x4_t Qd, int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
float32x4_t __simd_vsliq_n_f32(float32x4_t Qd, float32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vsli_n_u64(uint64x1_t Dd, uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vsli_n_s64(int64x1_t Dd, int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vsliq_n_u64(uint64x2_t Qd, uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vsliq_n_s64(int64x2_t Qd, int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vqshlu_n_s8(int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vqshluq_n_s8(int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vqshlu_n_s16(int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vqshluq_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vqshlu_n_s32(int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vqshluq_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vqshlu_n_s64(int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vqshluq_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vqshl_n_s8(int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vqshlq_n_s8(int8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vqshl_n_u8(uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vqshlq_n_u8(uint8x16_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vqshl_n_s16(int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vqshlq_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vqshl_n_u16(uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vqshlq_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vqshl_n_s32(int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vqshlq_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vqshl_n_u32(uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vqshlq_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vqshl_n_s64(int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vqshlq_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vqshl_n_u64(uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vqshlq_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vmovl_s8(int8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vmovl_u8(uint8x8_t Dm) __NEONSAFE;
int32x4_t __simd_vmovl_s16(int16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vmovl_u16(uint16x4_t Dm) __NEONSAFE;
int64x2_t __simd_vmovl_s32(int32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vmovl_u32(uint32x2_t Dm) __NEONSAFE;
int16x8_t __simd_vshll_n_s8(int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vshll_n_u8(uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vshll_n_s16(int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vshll_n_u16(uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vshll_n_s32(int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vshll_n_u32(uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vshrn_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vshrn_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vshrn_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vshrn_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vshrn_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vshrn_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vrshrn_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vrshrn_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vrshrn_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vrshrn_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vrshrn_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vrshrn_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vqshrun_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vqshrun_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vqshrun_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vqrshrun_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vqrshrun_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vqrshrun_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vqshrn_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vqshrn_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vqshrn_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vqshrn_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vqshrn_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vqshrn_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int8x8_t __simd_vqrshrn_n_s16(int16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vqrshrn_n_u16(uint16x8_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vqrshrn_n_s32(int32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vqrshrn_n_u32(uint32x4_t Qm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vqrshrn_n_s64(int64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vqrshrn_n_u64(uint64x2_t Qm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vrev64_u8(uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vrev64_s8(int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vrev64_p8(poly8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vrev64q_u8(uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vrev64q_s8(int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vrev64q_p8(poly8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vrev64_u16(uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vrev64_s16(int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vrev64_p16(poly16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vrev64q_u16(uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vrev64q_s16(int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vrev64q_p16(poly16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vrev64_u32(uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vrev64_s32(int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vrev64_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vrev64q_u32(uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vrev64q_s32(int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vrev64q_f32(float32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vrev32_u8(uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vrev32_s8(int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vrev32_p8(poly8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vrev32q_u8(uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vrev32q_s8(int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vrev32q_p8(poly8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vrev32_u16(uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vrev32_s16(int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vrev32_p16(poly16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vrev32q_u16(uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vrev32q_s16(int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vrev32q_p16(poly16x8_t Qm) __NEONSAFE;
uint8x8_t __simd_vrev16_u8(uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vrev16_s8(int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vrev16_p8(poly8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vrev16q_u8(uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vrev16q_s8(int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vrev16q_p8(poly8x16_t Qm) __NEONSAFE;
int8x8_t __simd_vcls_s8(int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vclsq_s8(int8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vcls_s16(int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vclsq_s16(int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vcls_s32(int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vclsq_s32(int32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vclz_u8(uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vclz_s8(int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vclzq_u8(uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vclzq_s8(int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vclz_u16(uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vclz_s16(int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vclzq_u16(uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vclzq_s16(int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vclz_u32(uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vclz_s32(int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vclzq_u32(uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vclzq_s32(int32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vcnt_u8(uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vcnt_s8(int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vcnt_p8(poly8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcntq_u8(uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vcntq_s8(int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vcntq_p8(poly8x16_t Qm) __NEONSAFE;
uint8x8_t __simd_vmvn_u8(uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vmvn_s8(int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vmvn_p8(poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vmvn_u16(uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vmvn_s16(int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vmvn_p16(poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vmvn_u32(uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vmvn_s32(int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vmvn_f32(float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vmvn_u64(uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vmvn_s64(int64x1_t Dm) __NEONSAFE;
uint8x16_t __simd_vmvnq_u8(uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vmvnq_s8(int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vmvnq_p8(poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vmvnq_u16(uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vmvnq_s16(int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vmvnq_p16(poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vmvnq_u32(uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vmvnq_s32(int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vmvnq_f32(float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vmvnq_u64(uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vmvnq_s64(int64x2_t Qm) __NEONSAFE;
int8x8_t __simd_vqabs_s8(int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vqabsq_s8(int8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vqabs_s16(int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vqabsq_s16(int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vqabs_s32(int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqabsq_s32(int32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vqneg_s8(int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vqnegq_s8(int8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vqneg_s16(int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vqnegq_s16(int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vqneg_s32(int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vqnegq_s32(int32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vabs_s8(int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vabsq_s8(int8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vabs_s16(int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vabsq_s16(int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vabs_s32(int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vabsq_s32(int32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vabs_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vabsq_f32(float32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vneg_s8(int8x8_t Dm) __NEONSAFE;
int8x16_t __simd_vnegq_s8(int8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vneg_s16(int16x4_t Dm) __NEONSAFE;
int16x8_t __simd_vnegq_s16(int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vneg_s32(int32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vnegq_s32(int32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vneg_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vnegq_f32(float32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vpaddl_s8(int8x8_t Dm) __NEONSAFE;
int16x8_t __simd_vpaddlq_s8(int8x16_t Qm) __NEONSAFE;
int32x2_t __simd_vpaddl_s16(int16x4_t Dm) __NEONSAFE;
int32x4_t __simd_vpaddlq_s16(int16x8_t Qm) __NEONSAFE;
int64x1_t __simd_vpaddl_s32(int32x2_t Dm) __NEONSAFE;
int64x2_t __simd_vpaddlq_s32(int32x4_t Qm) __NEONSAFE;
uint16x4_t __simd_vpaddl_u8(uint8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vpaddlq_u8(uint8x16_t Qm) __NEONSAFE;
uint32x2_t __simd_vpaddl_u16(uint16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vpaddlq_u16(uint16x8_t Qm) __NEONSAFE;
uint64x1_t __simd_vpaddl_u32(uint32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vpaddlq_u32(uint32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vpadal_s8(int16x4_t Dd, int8x8_t Dm) __NEONSAFE;
int16x8_t __simd_vpadalq_s8(int16x8_t Qd, int8x16_t Qm) __NEONSAFE;
int32x2_t __simd_vpadal_s16(int32x2_t Dd, int16x4_t Dm) __NEONSAFE;
int32x4_t __simd_vpadalq_s16(int32x4_t Qd, int16x8_t Qm) __NEONSAFE;
int64x1_t __simd_vpadal_s32(int64x1_t Dd, int32x2_t Dm) __NEONSAFE;
int64x2_t __simd_vpadalq_s32(int64x2_t Qd, int32x4_t Qm) __NEONSAFE;
uint16x4_t __simd_vpadal_u8(uint16x4_t Dd, uint8x8_t Dm) __NEONSAFE;
uint16x8_t __simd_vpadalq_u8(uint16x8_t Qd, uint8x16_t Qm) __NEONSAFE;
uint32x2_t __simd_vpadal_u16(uint32x2_t Dd, uint16x4_t Dm) __NEONSAFE;
uint32x4_t __simd_vpadalq_u16(uint32x4_t Qd, uint16x8_t Qm) __NEONSAFE;
uint64x1_t __simd_vpadal_u32(uint64x1_t Dd, uint32x2_t Dm) __NEONSAFE;
uint64x2_t __simd_vpadalq_u32(uint64x2_t Qd, uint32x4_t Qm) __NEONSAFE;
uint8x8x2_t __simd_vtrn_u8(uint8x8_t Dd, uint8x8_t Dm) __NEONSAFE;
int8x8x2_t __simd_vtrn_s8(int8x8_t Dd, int8x8_t Dm) __NEONSAFE;
poly8x8x2_t __simd_vtrn_p8(poly8x8_t Dd, poly8x8_t Dm) __NEONSAFE;
uint8x16x2_t __simd_vtrnq_u8(uint8x16_t Qd, uint8x16_t Qm) __NEONSAFE;
int8x16x2_t __simd_vtrnq_s8(int8x16_t Qd, int8x16_t Qm) __NEONSAFE;
poly8x16x2_t __simd_vtrnq_p8(poly8x16_t Qd, poly8x16_t Qm) __NEONSAFE;
uint16x4x2_t __simd_vtrn_u16(uint16x4_t Dd, uint16x4_t Dm) __NEONSAFE;
int16x4x2_t __simd_vtrn_s16(int16x4_t Dd, int16x4_t Dm) __NEONSAFE;
poly16x4x2_t __simd_vtrn_p16(poly16x4_t Dd, poly16x4_t Dm) __NEONSAFE;
uint16x8x2_t __simd_vtrnq_u16(uint16x8_t Qd, uint16x8_t Qm) __NEONSAFE;
int16x8x2_t __simd_vtrnq_s16(int16x8_t Qd, int16x8_t Qm) __NEONSAFE;
poly16x8x2_t __simd_vtrnq_p16(poly16x8_t Qd, poly16x8_t Qm) __NEONSAFE;
uint32x2x2_t __simd_vtrn_u32(uint32x2_t Dd, uint32x2_t Dm) __NEONSAFE;
int32x2x2_t __simd_vtrn_s32(int32x2_t Dd, int32x2_t Dm) __NEONSAFE;
float32x2x2_t __simd_vtrn_f32(float32x2_t Dd, float32x2_t Dm) __NEONSAFE;
uint32x4x2_t __simd_vtrnq_u32(uint32x4_t Qd, uint32x4_t Qm) __NEONSAFE;
int32x4x2_t __simd_vtrnq_s32(int32x4_t Qd, int32x4_t Qm) __NEONSAFE;
float32x4x2_t __simd_vtrnq_f32(float32x4_t Qd, float32x4_t Qm) __NEONSAFE;
uint8x8x2_t __simd_vuzp_u8(uint8x8_t Dd, uint8x8_t Dm) __NEONSAFE;
int8x8x2_t __simd_vuzp_s8(int8x8_t Dd, int8x8_t Dm) __NEONSAFE;
poly8x8x2_t __simd_vuzp_p8(poly8x8_t Dd, poly8x8_t Dm) __NEONSAFE;
uint8x16x2_t __simd_vuzpq_u8(uint8x16_t Qd, uint8x16_t Qm) __NEONSAFE;
int8x16x2_t __simd_vuzpq_s8(int8x16_t Qd, int8x16_t Qm) __NEONSAFE;
poly8x16x2_t __simd_vuzpq_p8(poly8x16_t Qd, poly8x16_t Qm) __NEONSAFE;
uint16x4x2_t __simd_vuzp_u16(uint16x4_t Dd, uint16x4_t Dm) __NEONSAFE;
int16x4x2_t __simd_vuzp_s16(int16x4_t Dd, int16x4_t Dm) __NEONSAFE;
poly16x4x2_t __simd_vuzp_p16(poly16x4_t Dd, poly16x4_t Dm) __NEONSAFE;
uint16x8x2_t __simd_vuzpq_u16(uint16x8_t Qd, uint16x8_t Qm) __NEONSAFE;
int16x8x2_t __simd_vuzpq_s16(int16x8_t Qd, int16x8_t Qm) __NEONSAFE;
poly16x8x2_t __simd_vuzpq_p16(poly16x8_t Qd, poly16x8_t Qm) __NEONSAFE;
uint8x8x2_t __simd_vzip_u8(uint8x8_t Dd, uint8x8_t Dm) __NEONSAFE;
int8x8x2_t __simd_vzip_s8(int8x8_t Dd, int8x8_t Dm) __NEONSAFE;
poly8x8x2_t __simd_vzip_p8(poly8x8_t Dd, poly8x8_t Dm) __NEONSAFE;
uint8x16x2_t __simd_vzipq_u8(uint8x16_t Qd, uint8x16_t Qm) __NEONSAFE;
int8x16x2_t __simd_vzipq_s8(int8x16_t Qd, int8x16_t Qm) __NEONSAFE;
poly8x16x2_t __simd_vzipq_p8(poly8x16_t Qd, poly8x16_t Qm) __NEONSAFE;
uint16x4x2_t __simd_vzip_u16(uint16x4_t Dd, uint16x4_t Dm) __NEONSAFE;
int16x4x2_t __simd_vzip_s16(int16x4_t Dd, int16x4_t Dm) __NEONSAFE;
poly16x4x2_t __simd_vzip_p16(poly16x4_t Dd, poly16x4_t Dm) __NEONSAFE;
uint16x8x2_t __simd_vzipq_u16(uint16x8_t Qd, uint16x8_t Qm) __NEONSAFE;
int16x8x2_t __simd_vzipq_s16(int16x8_t Qd, int16x8_t Qm) __NEONSAFE;
poly16x8x2_t __simd_vzipq_p16(poly16x8_t Qd, poly16x8_t Qm) __NEONSAFE;
uint32x4x2_t __simd_vuzpq_u32(uint32x4_t Qd, uint32x4_t Qm) __NEONSAFE;
int32x4x2_t __simd_vuzpq_s32(int32x4_t Qd, int32x4_t Qm) __NEONSAFE;
float32x4x2_t __simd_vuzpq_f32(float32x4_t Qd, float32x4_t Qm) __NEONSAFE;
uint32x2x2_t __simd_vuzp_u32(uint32x2_t Dd, uint32x2_t Dm) __NEONSAFE;
int32x2x2_t __simd_vuzp_s32(int32x2_t Dd, int32x2_t Dm) __NEONSAFE;
float32x2x2_t __simd_vuzp_f32(float32x2_t Dd, float32x2_t Dm) __NEONSAFE;
uint32x4x2_t __simd_vzipq_u32(uint32x4_t Qd, uint32x4_t Qm) __NEONSAFE;
int32x4x2_t __simd_vzipq_s32(int32x4_t Qd, int32x4_t Qm) __NEONSAFE;
float32x4x2_t __simd_vzipq_f32(float32x4_t Qd, float32x4_t Qm) __NEONSAFE;
uint32x2x2_t __simd_vzip_u32(uint32x2_t Dd, uint32x2_t Dm) __NEONSAFE;
int32x2x2_t __simd_vzip_s32(int32x2_t Dd, int32x2_t Dm) __NEONSAFE;
float32x2x2_t __simd_vzip_f32(float32x2_t Dd, float32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vrecpe_u32(uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vrecpeq_u32(uint32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrecpe_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrecpeq_f32(float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vrsqrte_u32(uint32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vrsqrteq_u32(uint32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrsqrte_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrsqrteq_f32(float32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vmovn_u16(uint16x8_t Qm) __NEONSAFE;
int8x8_t __simd_vmovn_s16(int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vmovn_u32(uint32x4_t Qm) __NEONSAFE;
int16x4_t __simd_vmovn_s32(int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vmovn_u64(uint64x2_t Qm) __NEONSAFE;
int32x2_t __simd_vmovn_s64(int64x2_t Qm) __NEONSAFE;
int8x8_t __simd_vqmovn_s16(int16x8_t Qm) __NEONSAFE;
int16x4_t __simd_vqmovn_s32(int32x4_t Qm) __NEONSAFE;
int32x2_t __simd_vqmovn_s64(int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vqmovn_u16(uint16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vqmovn_u32(uint32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vqmovn_u64(uint64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vqmovun_s16(int16x8_t Qm) __NEONSAFE;
uint16x4_t __simd_vqmovun_s32(int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vqmovun_s64(int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vcgt_n_s8(int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcgtq_n_s8(int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vcgt_n_s16(int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcgtq_n_s16(int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vcgt_n_s32(int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgtq_n_s32(int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcgt_n_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgtq_n_f32(float32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vcge_n_s8(int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcgeq_n_s8(int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vcge_n_s16(int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcgeq_n_s16(int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vcge_n_s32(int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgeq_n_s32(int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcge_n_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcgeq_n_f32(float32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vceq_n_u8(uint8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vceq_n_s8(int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vceqq_n_u8(uint8x16_t Qm) __NEONSAFE;
uint8x16_t __simd_vceqq_n_s8(int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vceq_n_u16(uint16x4_t Dm) __NEONSAFE;
uint16x4_t __simd_vceq_n_s16(int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vceqq_n_u16(uint16x8_t Qm) __NEONSAFE;
uint16x8_t __simd_vceqq_n_s16(int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vceq_n_u32(uint32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vceq_n_s32(int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vceqq_n_u32(uint32x4_t Qm) __NEONSAFE;
uint32x4_t __simd_vceqq_n_s32(int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vceq_n_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vceqq_n_f32(float32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vcle_n_s8(int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcleq_n_s8(int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vcle_n_s16(int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcleq_n_s16(int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vcle_n_s32(int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcleq_n_s32(int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcle_n_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcleq_n_f32(float32x4_t Qm) __NEONSAFE;
uint8x8_t __simd_vclt_n_s8(int8x8_t Dm) __NEONSAFE;
uint8x16_t __simd_vcltq_n_s8(int8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vclt_n_s16(int16x4_t Dm) __NEONSAFE;
uint16x8_t __simd_vcltq_n_s16(int16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vclt_n_s32(int32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcltq_n_s32(int32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vclt_n_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcltq_n_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrinta_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrintaq_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrintm_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrintmq_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrintn_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrintnq_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrintp_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrintpq_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrintx_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrintxq_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vrintz_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vrintzq_f32(float32x4_t Qm) __NEONSAFE;
int32x2_t __simd_vcvta_s32_f32(float32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vcvtaq_s32_f32(float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcvta_u32_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcvtaq_u32_f32(float32x4_t Qm) __NEONSAFE;
int32x2_t __simd_vcvtm_s32_f32(float32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vcvtmq_s32_f32(float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcvtm_u32_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcvtmq_u32_f32(float32x4_t Qm) __NEONSAFE;
int32x2_t __simd_vcvtn_s32_f32(float32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vcvtnq_s32_f32(float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcvtn_u32_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcvtnq_u32_f32(float32x4_t Qm) __NEONSAFE;
int32x2_t __simd_vcvtp_s32_f32(float32x2_t Dm) __NEONSAFE;
int32x4_t __simd_vcvtpq_s32_f32(float32x4_t Qm) __NEONSAFE;
uint32x2_t __simd_vcvtp_u32_f32(float32x2_t Dm) __NEONSAFE;
uint32x4_t __simd_vcvtpq_u32_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vmaxnm_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vmaxnmq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vminnm_f32(float32x2_t Dn, float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vminnmq_f32(float32x4_t Qn, float32x4_t Qm) __NEONSAFE;

int8x8_t __simd_vdup_n_s8(int8_t Rt) __NEONSAFE;
int16x4_t __simd_vdup_n_s16(int16_t Rt) __NEONSAFE;
int32x2_t __simd_vdup_n_s32(int32_t Rt) __NEONSAFE;
int64x1_t __simd_vdup_n_s64(int64_t Rt) __NEONSAFE;
uint8x8_t __simd_vdup_n_u8(uint8_t Rt) __NEONSAFE;
uint16x4_t __simd_vdup_n_u16(uint16_t Rt) __NEONSAFE;
uint32x2_t __simd_vdup_n_u32(uint32_t Rt) __NEONSAFE;
uint64x1_t __simd_vdup_n_u64(uint64_t Rt) __NEONSAFE;
poly8x8_t __simd_vdup_n_p8(poly8_t Rt) __NEONSAFE;
poly16x4_t __simd_vdup_n_p16(poly16_t Rt) __NEONSAFE;
float32x2_t __simd_vdup_n_f32(float32_t Rt) __NEONSAFE;
int8x16_t __simd_vdupq_n_s8(int8_t Rt) __NEONSAFE;
int16x8_t __simd_vdupq_n_s16(int16_t Rt) __NEONSAFE;
int32x4_t __simd_vdupq_n_s32(int32_t Rt) __NEONSAFE;
int64x2_t __simd_vdupq_n_s64(int64_t Rt) __NEONSAFE;
uint8x16_t __simd_vdupq_n_u8(uint8_t Rt) __NEONSAFE;
uint16x8_t __simd_vdupq_n_u16(uint16_t Rt) __NEONSAFE;
uint32x4_t __simd_vdupq_n_u32(uint32_t Rt) __NEONSAFE;
uint64x2_t __simd_vdupq_n_u64(uint64_t Rt) __NEONSAFE;
poly8x16_t __simd_vdupq_n_p8(poly8_t Rt) __NEONSAFE;
poly16x8_t __simd_vdupq_n_p16(poly16_t Rt) __NEONSAFE;
float32x4_t __simd_vdupq_n_f32(float32_t Rt) __NEONSAFE;
int8x8_t __simd_vdup_lane_s8(int8x8_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x4_t __simd_vdup_lane_s16(int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x2_t __simd_vdup_lane_s32(int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x1_t __simd_vdup_lane_s64(int64x1_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint8x8_t __simd_vdup_lane_u8(uint8x8_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x4_t __simd_vdup_lane_u16(uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x2_t __simd_vdup_lane_u32(uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint64x1_t __simd_vdup_lane_u64(uint64x1_t Dm, __ghs_c_int__ lane) __NEONSAFE;
poly8x8_t __simd_vdup_lane_p8(poly8x8_t Dm, __ghs_c_int__ lane) __NEONSAFE;
poly16x4_t __simd_vdup_lane_p16(poly16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x2_t __simd_vdup_lane_f32(float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int8x16_t __simd_vdupq_lane_s8(int8x8_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int16x8_t __simd_vdupq_lane_s16(int16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vdupq_lane_s32(int32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vdupq_lane_s64(int64x1_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint8x16_t __simd_vdupq_lane_u8(uint8x8_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint16x8_t __simd_vdupq_lane_u16(uint16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vdupq_lane_u32(uint32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;
uint64x2_t __simd_vdupq_lane_u64(uint64x1_t Dm, __ghs_c_int__ lane) __NEONSAFE;
poly8x16_t __simd_vdupq_lane_p8(poly8x8_t Dm, __ghs_c_int__ lane) __NEONSAFE;
poly16x8_t __simd_vdupq_lane_p16(poly16x4_t Dm, __ghs_c_int__ lane) __NEONSAFE;
float32x4_t __simd_vdupq_lane_f32(float32x2_t Dm, __ghs_c_int__ lane) __NEONSAFE;

float32x2_t __simd_vcvt_f32_u32(uint32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vcvt_f32_s32(int32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vcvt_u32_f32(float32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vcvt_s32_f32(float32x2_t Dm) __NEONSAFE;
float32x4_t __simd_vcvtq_f32_u32(uint32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vcvtq_f32_s32(int32x4_t Qm) __NEONSAFE;
uint32x4_t __simd_vcvtq_u32_f32(float32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vcvtq_s32_f32(float32x4_t Qm) __NEONSAFE;
float32x2_t __simd_vcvt_n_f32_u32(uint32x2_t Dm, __ghs_c_int__ fbits) __NEONSAFE;
float32x2_t __simd_vcvt_n_f32_s32(int32x2_t Dm, __ghs_c_int__ fbits) __NEONSAFE;
uint32x2_t __simd_vcvt_n_u32_f32(float32x2_t Dm, __ghs_c_int__ fbits) __NEONSAFE;
int32x2_t __simd_vcvt_n_s32_f32(float32x2_t Dm, __ghs_c_int__ fbits) __NEONSAFE;
float32x4_t __simd_vcvtq_n_f32_u32(uint32x4_t Qm, __ghs_c_int__ fbits) __NEONSAFE;
float32x4_t __simd_vcvtq_n_f32_s32(int32x4_t Qm, __ghs_c_int__ fbits) __NEONSAFE;
uint32x4_t __simd_vcvtq_n_u32_f32(float32x4_t Qm, __ghs_c_int__ fbits) __NEONSAFE;
int32x4_t __simd_vcvtq_n_s32_f32(float32x4_t Qm, __ghs_c_int__ fbits) __NEONSAFE;

float32x4_t __simd_vcvt_f32_f16(float16x4_t Dm) __NEONSAFE;
float16x4_t __simd_vcvt_f16_f32(float32x4_t Qm) __NEONSAFE;

int8x8_t __simd_vext_s8(int8x8_t Dn, int8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x4_t __simd_vext_s16(int16x4_t Dn, int16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x2_t __simd_vext_s32(int32x2_t Dn, int32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x1_t __simd_vext_s64(int64x1_t Dn, int64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x8_t __simd_vext_u8(uint8x8_t Dn, uint8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x4_t __simd_vext_u16(uint16x4_t Dn, uint16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x2_t __simd_vext_u32(uint32x2_t Dn, uint32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x1_t __simd_vext_u64(uint64x1_t Dn, uint64x1_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly8x8_t __simd_vext_p8(poly8x8_t Dn, poly8x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly16x4_t __simd_vext_p16(poly16x4_t Dn, poly16x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
float32x2_t __simd_vext_f32(float32x2_t Dn, float32x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int8x16_t __simd_vextq_s8(int8x16_t Dn, int8x16_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int16x8_t __simd_vextq_s16(int16x8_t Dn, int16x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int32x4_t __simd_vextq_s32(int32x4_t Dn, int32x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
int64x2_t __simd_vextq_s64(int64x2_t Dn, int64x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint8x16_t __simd_vextq_u8(uint8x16_t Dn, uint8x16_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint16x8_t __simd_vextq_u16(uint16x8_t Dn, uint16x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint32x4_t __simd_vextq_u32(uint32x4_t Dn, uint32x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;
uint64x2_t __simd_vextq_u64(uint64x2_t Dn, uint64x2_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly8x16_t __simd_vextq_p8(poly8x16_t Dn, poly8x16_t Dm, __ghs_c_int__ amt) __NEONSAFE;
poly16x8_t __simd_vextq_p16(poly16x8_t Dn, poly16x8_t Dm, __ghs_c_int__ amt) __NEONSAFE;
float32x4_t __simd_vextq_f32(float32x4_t Dn, float32x4_t Dm, __ghs_c_int__ amt) __NEONSAFE;

int8x8_t __simd_vtbl1_s8(int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbl1_u8(uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbl1_p8(poly8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vtbl2_s8(int8x8x2_t Qn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbl2_u8(uint8x8x2_t Qn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbl2_p8(poly8x8x2_t Qn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vtbl3_s8(int8x8x3_t Vn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbl3_u8(uint8x8x3_t Vn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbl3_p8(poly8x8x3_t Vn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vtbl4_s8(int8x8x4_t Vn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbl4_u8(uint8x8x4_t Vn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbl4_p8(poly8x8x4_t Vn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vtbx1_s8(int8x8_t Dd, int8x8_t Dn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbx1_u8(uint8x8_t Dd, uint8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbx1_p8(poly8x8_t Dd, poly8x8_t Dn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vtbx2_s8(int8x8_t Dd, int8x8x2_t Qn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbx2_u8(uint8x8_t Dd, uint8x8x2_t Qn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbx2_p8(poly8x8_t Dd, poly8x8x2_t Qn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vtbx3_s8(int8x8_t Dd, int8x8x3_t Vn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbx3_u8(uint8x8_t Dd, uint8x8x3_t Vn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbx3_p8(poly8x8_t Dd, poly8x8x3_t Vn, uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vtbx4_s8(int8x8_t Dd, int8x8x4_t Vn, int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vtbx4_u8(uint8x8_t Dd, uint8x8x4_t Vn, uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vtbx4_p8(poly8x8_t Dd, poly8x8x4_t Vn, uint8x8_t Dm) __NEONSAFE;

int8x8_t __simd_vld1_s8(const int8_t * Rn) __NEONPURE;
int16x4_t __simd_vld1_s16(const int16_t * Rn) __NEONPURE;
int32x2_t __simd_vld1_s32(const int32_t * Rn) __NEONPURE;
int64x1_t __simd_vld1_s64(const int64_t * Rn) __NEONPURE;
uint8x8_t __simd_vld1_u8(const uint8_t * Rn) __NEONPURE;
uint16x4_t __simd_vld1_u16(const uint16_t * Rn) __NEONPURE;
uint32x2_t __simd_vld1_u32(const uint32_t * Rn) __NEONPURE;
uint64x1_t __simd_vld1_u64(const uint64_t * Rn) __NEONPURE;
poly8x8_t __simd_vld1_p8(const poly8_t * Rn) __NEONPURE;
poly16x4_t __simd_vld1_p16(const poly16_t * Rn) __NEONPURE;
float32x2_t __simd_vld1_f32(const float32_t * Rn) __NEONPURE;
int8x16_t __simd_vld1q_s8(const int8_t * Rn) __NEONPURE;
int16x8_t __simd_vld1q_s16(const int16_t * Rn) __NEONPURE;
int32x4_t __simd_vld1q_s32(const int32_t * Rn) __NEONPURE;
int64x2_t __simd_vld1q_s64(const int64_t * Rn) __NEONPURE;
uint8x16_t __simd_vld1q_u8(const uint8_t * Rn) __NEONPURE;
uint16x8_t __simd_vld1q_u16(const uint16_t * Rn) __NEONPURE;
uint32x4_t __simd_vld1q_u32(const uint32_t * Rn) __NEONPURE;
uint64x2_t __simd_vld1q_u64(const uint64_t * Rn) __NEONPURE;
poly8x16_t __simd_vld1q_p8(const poly8_t * Rn) __NEONPURE;
poly16x8_t __simd_vld1q_p16(const poly16_t * Rn) __NEONPURE;
float32x4_t __simd_vld1q_f32(const float32_t * Rn) __NEONPURE;
int8x8_t __simd_vld1_lane_s8(const int8_t * Rn, int8x8_t Dd, __ghs_c_int__ lane) __NEONPURE;
int16x4_t __simd_vld1_lane_s16(const int16_t * Rn, int16x4_t Dd, __ghs_c_int__ lane) __NEONPURE;
int32x2_t __simd_vld1_lane_s32(const int32_t * Rn, int32x2_t Dd, __ghs_c_int__ lane) __NEONPURE;
int64x1_t __simd_vld1_lane_s64(const int64_t * Rn, int64x1_t Dd, __ghs_c_int__ lane) __NEONPURE;
uint8x8_t __simd_vld1_lane_u8(const uint8_t * Rn, uint8x8_t Dd, __ghs_c_int__ lane) __NEONPURE;
uint16x4_t __simd_vld1_lane_u16(const uint16_t * Rn, uint16x4_t Dd, __ghs_c_int__ lane) __NEONPURE;
uint32x2_t __simd_vld1_lane_u32(const uint32_t * Rn, uint32x2_t Dd, __ghs_c_int__ lane) __NEONPURE;
uint64x1_t __simd_vld1_lane_u64(const uint64_t * Rn, uint64x1_t Dd, __ghs_c_int__ lane) __NEONPURE;
poly8x8_t __simd_vld1_lane_p8(const poly8_t * Rn, poly8x8_t Dd, __ghs_c_int__ lane) __NEONPURE;
poly16x4_t __simd_vld1_lane_p16(const poly16_t * Rn, poly16x4_t Dd, __ghs_c_int__ lane) __NEONPURE;
float32x2_t __simd_vld1_lane_f32(const float32_t * Rn, float32x2_t Dd, __ghs_c_int__ lane) __NEONPURE;
int8x16_t __simd_vld1q_lane_s8(const int8_t * Rn, int8x16_t Qd, __ghs_c_int__ lane) __NEONPURE;
int16x8_t __simd_vld1q_lane_s16(const int16_t * Rn, int16x8_t Qd, __ghs_c_int__ lane) __NEONPURE;
int32x4_t __simd_vld1q_lane_s32(const int32_t * Rn, int32x4_t Qd, __ghs_c_int__ lane) __NEONPURE;
int64x2_t __simd_vld1q_lane_s64(const int64_t * Rn, int64x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
uint8x16_t __simd_vld1q_lane_u8(const uint8_t * Rn, uint8x16_t Qd, __ghs_c_int__ lane) __NEONPURE;
uint16x8_t __simd_vld1q_lane_u16(const uint16_t * Rn, uint16x8_t Qd, __ghs_c_int__ lane) __NEONPURE;
uint32x4_t __simd_vld1q_lane_u32(const uint32_t * Rn, uint32x4_t Qd, __ghs_c_int__ lane) __NEONPURE;
uint64x2_t __simd_vld1q_lane_u64(const uint64_t * Rn, uint64x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
poly8x16_t __simd_vld1q_lane_p8(const poly8_t * Rn, poly8x16_t Qd, __ghs_c_int__ lane) __NEONPURE;
poly16x8_t __simd_vld1q_lane_p16(const poly16_t * Rn, poly16x8_t Qd, __ghs_c_int__ lane) __NEONPURE;
float32x4_t __simd_vld1q_lane_f32(const float32_t * Rn, float32x4_t Qd, __ghs_c_int__ lane) __NEONPURE;
int8x8_t __simd_vld1_dup_s8(const int8_t * Rn) __NEONPURE;
int16x4_t __simd_vld1_dup_s16(const int16_t * Rn) __NEONPURE;
int32x2_t __simd_vld1_dup_s32(const int32_t * Rn) __NEONPURE;
int64x1_t __simd_vld1_dup_s64(const int64_t * Rn) __NEONPURE;
uint8x8_t __simd_vld1_dup_u8(const uint8_t * Rn) __NEONPURE;
uint16x4_t __simd_vld1_dup_u16(const uint16_t * Rn) __NEONPURE;
uint32x2_t __simd_vld1_dup_u32(const uint32_t * Rn) __NEONPURE;
uint64x1_t __simd_vld1_dup_u64(const uint64_t * Rn) __NEONPURE;
poly8x8_t __simd_vld1_dup_p8(const poly8_t * Rn) __NEONPURE;
poly16x4_t __simd_vld1_dup_p16(const poly16_t * Rn) __NEONPURE;
float32x2_t __simd_vld1_dup_f32(const float32_t * Rn) __NEONPURE;
int8x16_t __simd_vld1q_dup_s8(const int8_t * Rn) __NEONPURE;
int16x8_t __simd_vld1q_dup_s16(const int16_t * Rn) __NEONPURE;
int32x4_t __simd_vld1q_dup_s32(const int32_t * Rn) __NEONPURE;
int64x2_t __simd_vld1q_dup_s64(const int64_t * Rn) __NEONPURE;
uint8x16_t __simd_vld1q_dup_u8(const uint8_t * Rn) __NEONPURE;
uint16x8_t __simd_vld1q_dup_u16(const uint16_t * Rn) __NEONPURE;
uint32x4_t __simd_vld1q_dup_u32(const uint32_t * Rn) __NEONPURE;
uint64x2_t __simd_vld1q_dup_u64(const uint64_t * Rn) __NEONPURE;
poly8x16_t __simd_vld1q_dup_p8(const poly8_t * Rn) __NEONPURE;
poly16x8_t __simd_vld1q_dup_p16(const poly16_t * Rn) __NEONPURE;
float32x4_t __simd_vld1q_dup_f32(const float32_t * Rn) __NEONPURE;

int8x8x2_t __simd_vld2_s8(const int8_t * Rn) __NEONPURE;
int16x4x2_t __simd_vld2_s16(const int16_t * Rn) __NEONPURE;
int32x2x2_t __simd_vld2_s32(const int32_t * Rn) __NEONPURE;
int64x1x2_t __simd_vld2_s64(const int64_t * Rn) __NEONPURE;
uint8x8x2_t __simd_vld2_u8(const uint8_t * Rn) __NEONPURE;
uint16x4x2_t __simd_vld2_u16(const uint16_t * Rn) __NEONPURE;
uint32x2x2_t __simd_vld2_u32(const uint32_t * Rn) __NEONPURE;
uint64x1x2_t __simd_vld2_u64(const uint64_t * Rn) __NEONPURE;
poly8x8x2_t __simd_vld2_p8(const poly8_t * Rn) __NEONPURE;
poly16x4x2_t __simd_vld2_p16(const poly16_t * Rn) __NEONPURE;
float32x2x2_t __simd_vld2_f32(const float32_t * Rn) __NEONPURE;
int8x16x2_t __simd_vld2q_s8(const int8_t * Rn) __NEONPURE;
int16x8x2_t __simd_vld2q_s16(const int16_t * Rn) __NEONPURE;
int32x4x2_t __simd_vld2q_s32(const int32_t * Rn) __NEONPURE;
uint8x16x2_t __simd_vld2q_u8(const uint8_t * Rn) __NEONPURE;
uint16x8x2_t __simd_vld2q_u16(const uint16_t * Rn) __NEONPURE;
uint32x4x2_t __simd_vld2q_u32(const uint32_t * Rn) __NEONPURE;
poly8x16x2_t __simd_vld2q_p8(const poly8_t * Rn) __NEONPURE;
poly16x8x2_t __simd_vld2q_p16(const poly16_t * Rn) __NEONPURE;
float32x4x2_t __simd_vld2q_f32(const float32_t * Rn) __NEONPURE;
int8x8x2_t __simd_vld2_lane_s8(const int8_t * Rn, int8x8x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
int16x4x2_t __simd_vld2_lane_s16(const int16_t * Rn, int16x4x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
int32x2x2_t __simd_vld2_lane_s32(const int32_t * Rn, int32x2x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
uint8x8x2_t __simd_vld2_lane_u8(const uint8_t * Rn, uint8x8x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
uint16x4x2_t __simd_vld2_lane_u16(const uint16_t * Rn, uint16x4x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
uint32x2x2_t __simd_vld2_lane_u32(const uint32_t * Rn, uint32x2x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
poly8x8x2_t __simd_vld2_lane_p8(const poly8_t * Rn, poly8x8x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
poly16x4x2_t __simd_vld2_lane_p16(const poly16_t * Rn, poly16x4x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
float32x2x2_t __simd_vld2_lane_f32(const float32_t * Rn, float32x2x2_t Qd, __ghs_c_int__ lane) __NEONPURE;
int16x8x2_t __simd_vld2q_lane_s16(const int16_t * Rn, int16x8x2_t Vd, __ghs_c_int__ lane) __NEONPURE;
int32x4x2_t __simd_vld2q_lane_s32(const int32_t * Rn, int32x4x2_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint16x8x2_t __simd_vld2q_lane_u16(const uint16_t * Rn, uint16x8x2_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint32x4x2_t __simd_vld2q_lane_u32(const uint32_t * Rn, uint32x4x2_t Vd, __ghs_c_int__ lane) __NEONPURE;
poly16x8x2_t __simd_vld2q_lane_p16(const poly16_t * Rn, poly16x8x2_t Vd, __ghs_c_int__ lane) __NEONPURE;
float32x4x2_t __simd_vld2q_lane_f32(const float32_t * Rn, float32x4x2_t Vd, __ghs_c_int__ lane) __NEONPURE;
int8x8x2_t __simd_vld2_dup_s8(const int8_t * Rn) __NEONPURE;
int16x4x2_t __simd_vld2_dup_s16(const int16_t * Rn) __NEONPURE;
int32x2x2_t __simd_vld2_dup_s32(const int32_t * Rn) __NEONPURE;
int64x1x2_t __simd_vld2_dup_s64(const int64_t * Rn) __NEONPURE;
uint8x8x2_t __simd_vld2_dup_u8(const uint8_t * Rn) __NEONPURE;
uint16x4x2_t __simd_vld2_dup_u16(const uint16_t * Rn) __NEONPURE;
uint32x2x2_t __simd_vld2_dup_u32(const uint32_t * Rn) __NEONPURE;
uint64x1x2_t __simd_vld2_dup_u64(const uint64_t * Rn) __NEONPURE;
poly8x8x2_t __simd_vld2_dup_p8(const poly8_t * Rn) __NEONPURE;
poly16x4x2_t __simd_vld2_dup_p16(const poly16_t * Rn) __NEONPURE;
float32x2x2_t __simd_vld2_dup_f32(const float32_t * Rn) __NEONPURE;

int8x8x3_t __simd_vld3_s8(const int8_t * Rn) __NEONPURE;
int16x4x3_t __simd_vld3_s16(const int16_t * Rn) __NEONPURE;
int32x2x3_t __simd_vld3_s32(const int32_t * Rn) __NEONPURE;
int64x1x3_t __simd_vld3_s64(const int64_t * Rn) __NEONPURE;
uint8x8x3_t __simd_vld3_u8(const uint8_t * Rn) __NEONPURE;
uint16x4x3_t __simd_vld3_u16(const uint16_t * Rn) __NEONPURE;
uint32x2x3_t __simd_vld3_u32(const uint32_t * Rn) __NEONPURE;
uint64x1x3_t __simd_vld3_u64(const uint64_t * Rn) __NEONPURE;
poly8x8x3_t __simd_vld3_p8(const poly8_t * Rn) __NEONPURE;
poly16x4x3_t __simd_vld3_p16(const poly16_t * Rn) __NEONPURE;
float32x2x3_t __simd_vld3_f32(const float32_t * Rn) __NEONPURE;
int8x16x3_t __simd_vld3q_s8(const int8_t * Rn) __NEONPURE;
int16x8x3_t __simd_vld3q_s16(const int16_t * Rn) __NEONPURE;
int32x4x3_t __simd_vld3q_s32(const int32_t * Rn) __NEONPURE;
uint8x16x3_t __simd_vld3q_u8(const uint8_t * Rn) __NEONPURE;
uint16x8x3_t __simd_vld3q_u16(const uint16_t * Rn) __NEONPURE;
uint32x4x3_t __simd_vld3q_u32(const uint32_t * Rn) __NEONPURE;
poly8x16x3_t __simd_vld3q_p8(const poly8_t * Rn) __NEONPURE;
poly16x8x3_t __simd_vld3q_p16(const poly16_t * Rn) __NEONPURE;
float32x4x3_t __simd_vld3q_f32(const float32_t * Rn) __NEONPURE;
int8x8x3_t __simd_vld3_lane_s8(const int8_t * Rn, int8x8x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
int16x4x3_t __simd_vld3_lane_s16(const int16_t * Rn, int16x4x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
int32x2x3_t __simd_vld3_lane_s32(const int32_t * Rn, int32x2x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint8x8x3_t __simd_vld3_lane_u8(const uint8_t * Rn, uint8x8x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint16x4x3_t __simd_vld3_lane_u16(const uint16_t * Rn, uint16x4x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint32x2x3_t __simd_vld3_lane_u32(const uint32_t * Rn, uint32x2x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
poly8x8x3_t __simd_vld3_lane_p8(const poly8_t * Rn, poly8x8x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
poly16x4x3_t __simd_vld3_lane_p16(const poly16_t * Rn, poly16x4x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
float32x2x3_t __simd_vld3_lane_f32(const float32_t * Rn, float32x2x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
int16x8x3_t __simd_vld3q_lane_s16(const int16_t * Rn, int16x8x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
int32x4x3_t __simd_vld3q_lane_s32(const int32_t * Rn, int32x4x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint16x8x3_t __simd_vld3q_lane_u16(const uint16_t * Rn, uint16x8x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint32x4x3_t __simd_vld3q_lane_u32(const uint32_t * Rn, uint32x4x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
poly16x8x3_t __simd_vld3q_lane_p16(const poly16_t * Rn, poly16x8x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
float32x4x3_t __simd_vld3q_lane_f32(const float32_t * Rn, float32x4x3_t Vd, __ghs_c_int__ lane) __NEONPURE;
int8x8x3_t __simd_vld3_dup_s8(const int8_t * Rn) __NEONPURE;
int16x4x3_t __simd_vld3_dup_s16(const int16_t * Rn) __NEONPURE;
int32x2x3_t __simd_vld3_dup_s32(const int32_t * Rn) __NEONPURE;
int64x1x3_t __simd_vld3_dup_s64(const int64_t * Rn) __NEONPURE;
uint8x8x3_t __simd_vld3_dup_u8(const uint8_t * Rn) __NEONPURE;
uint16x4x3_t __simd_vld3_dup_u16(const uint16_t * Rn) __NEONPURE;
uint32x2x3_t __simd_vld3_dup_u32(const uint32_t * Rn) __NEONPURE;
uint64x1x3_t __simd_vld3_dup_u64(const uint64_t * Rn) __NEONPURE;
poly8x8x3_t __simd_vld3_dup_p8(const poly8_t * Rn) __NEONPURE;
poly16x4x3_t __simd_vld3_dup_p16(const poly16_t * Rn) __NEONPURE;
float32x2x3_t __simd_vld3_dup_f32(const float32_t * Rn) __NEONPURE;

int8x8x4_t __simd_vld4_s8(const int8_t * Rn) __NEONPURE;
int16x4x4_t __simd_vld4_s16(const int16_t * Rn) __NEONPURE;
int32x2x4_t __simd_vld4_s32(const int32_t * Rn) __NEONPURE;
int64x1x4_t __simd_vld4_s64(const int64_t * Rn) __NEONPURE;
uint8x8x4_t __simd_vld4_u8(const uint8_t * Rn) __NEONPURE;
uint16x4x4_t __simd_vld4_u16(const uint16_t * Rn) __NEONPURE;
uint32x2x4_t __simd_vld4_u32(const uint32_t * Rn) __NEONPURE;
uint64x1x4_t __simd_vld4_u64(const uint64_t * Rn) __NEONPURE;
poly8x8x4_t __simd_vld4_p8(const poly8_t * Rn) __NEONPURE;
poly16x4x4_t __simd_vld4_p16(const poly16_t * Rn) __NEONPURE;
float32x2x4_t __simd_vld4_f32(const float32_t * Rn) __NEONPURE;
int8x16x4_t __simd_vld4q_s8(const int8_t * Rn) __NEONPURE;
int16x8x4_t __simd_vld4q_s16(const int16_t * Rn) __NEONPURE;
int32x4x4_t __simd_vld4q_s32(const int32_t * Rn) __NEONPURE;
uint8x16x4_t __simd_vld4q_u8(const uint8_t * Rn) __NEONPURE;
uint16x8x4_t __simd_vld4q_u16(const uint16_t * Rn) __NEONPURE;
uint32x4x4_t __simd_vld4q_u32(const uint32_t * Rn) __NEONPURE;
poly8x16x4_t __simd_vld4q_p8(const poly8_t * Rn) __NEONPURE;
poly16x8x4_t __simd_vld4q_p16(const poly16_t * Rn) __NEONPURE;
float32x4x4_t __simd_vld4q_f32(const float32_t * Rn) __NEONPURE;
int8x8x4_t __simd_vld4_lane_s8(const int8_t * Rn, int8x8x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
int16x4x4_t __simd_vld4_lane_s16(const int16_t * Rn, int16x4x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
int32x2x4_t __simd_vld4_lane_s32(const int32_t * Rn, int32x2x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint8x8x4_t __simd_vld4_lane_u8(const uint8_t * Rn, uint8x8x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint16x4x4_t __simd_vld4_lane_u16(const uint16_t * Rn, uint16x4x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint32x2x4_t __simd_vld4_lane_u32(const uint32_t * Rn, uint32x2x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
poly8x8x4_t __simd_vld4_lane_p8(const poly8_t * Rn, poly8x8x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
poly16x4x4_t __simd_vld4_lane_p16(const poly16_t * Rn, poly16x4x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
float32x2x4_t __simd_vld4_lane_f32(const float32_t * Rn, float32x2x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
int16x8x4_t __simd_vld4q_lane_s16(const int16_t * Rn, int16x8x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
int32x4x4_t __simd_vld4q_lane_s32(const int32_t * Rn, int32x4x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint16x8x4_t __simd_vld4q_lane_u16(const uint16_t * Rn, uint16x8x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
uint32x4x4_t __simd_vld4q_lane_u32(const uint32_t * Rn, uint32x4x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
poly16x8x4_t __simd_vld4q_lane_p16(const poly16_t * Rn, poly16x8x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
float32x4x4_t __simd_vld4q_lane_f32(const float32_t * Rn, float32x4x4_t Vd, __ghs_c_int__ lane) __NEONPURE;
int8x8x4_t __simd_vld4_dup_s8(const int8_t * Rn) __NEONPURE;
int16x4x4_t __simd_vld4_dup_s16(const int16_t * Rn) __NEONPURE;
int32x2x4_t __simd_vld4_dup_s32(const int32_t * Rn) __NEONPURE;
int64x1x4_t __simd_vld4_dup_s64(const int64_t * Rn) __NEONPURE;
uint8x8x4_t __simd_vld4_dup_u8(const uint8_t * Rn) __NEONPURE;
uint16x4x4_t __simd_vld4_dup_u16(const uint16_t * Rn) __NEONPURE;
uint32x2x4_t __simd_vld4_dup_u32(const uint32_t * Rn) __NEONPURE;
uint64x1x4_t __simd_vld4_dup_u64(const uint64_t * Rn) __NEONPURE;
poly8x8x4_t __simd_vld4_dup_p8(const poly8_t * Rn) __NEONPURE;
poly16x4x4_t __simd_vld4_dup_p16(const poly16_t * Rn) __NEONPURE;
float32x2x4_t __simd_vld4_dup_f32(const float32_t * Rn) __NEONPURE;

void __simd_vst1_s8(int8_t * Rn, int8x8_t Dd);
void __simd_vst1_s16(int16_t * Rn, int16x4_t Dd);
void __simd_vst1_s32(int32_t * Rn, int32x2_t Dd);
void __simd_vst1_s64(int64_t * Rn, int64x1_t Dd);
void __simd_vst1_u8(uint8_t * Rn, uint8x8_t Dd);
void __simd_vst1_u16(uint16_t * Rn, uint16x4_t Dd);
void __simd_vst1_u32(uint32_t * Rn, uint32x2_t Dd);
void __simd_vst1_u64(uint64_t * Rn, uint64x1_t Dd);
void __simd_vst1_p8(poly8_t * Rn, poly8x8_t Dd);
void __simd_vst1_p16(poly16_t * Rn, poly16x4_t Dd);
void __simd_vst1_f32(float32_t * Rn, float32x2_t Dd);
void __simd_vst1q_s8(int8_t * Rn, int8x16_t Qd);
void __simd_vst1q_s16(int16_t * Rn, int16x8_t Qd);
void __simd_vst1q_s32(int32_t * Rn, int32x4_t Qd);
void __simd_vst1q_s64(int64_t * Rn, int64x2_t Qd);
void __simd_vst1q_u8(uint8_t * Rn, uint8x16_t Qd);
void __simd_vst1q_u16(uint16_t * Rn, uint16x8_t Qd);
void __simd_vst1q_u32(uint32_t * Rn, uint32x4_t Qd);
void __simd_vst1q_u64(uint64_t * Rn, uint64x2_t Qd);
void __simd_vst1q_p8(poly8_t * Rn, poly8x16_t Qd);
void __simd_vst1q_p16(poly16_t * Rn, poly16x8_t Qd);
void __simd_vst1q_f32(float32_t * Rn, float32x4_t Qd);
void __simd_vst1_lane_s8(int8_t * Rn, int8x8_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_s16(int16_t * Rn, int16x4_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_s32(int32_t * Rn, int32x2_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_s64(int64_t * Rn, int64x1_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_u8(uint8_t * Rn, uint8x8_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_u16(uint16_t * Rn, uint16x4_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_u32(uint32_t * Rn, uint32x2_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_u64(uint64_t * Rn, uint64x1_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_p8(poly8_t * Rn, poly8x8_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_p16(poly16_t * Rn, poly16x4_t Dd, __ghs_c_int__ lane);
void __simd_vst1_lane_f32(float32_t * Rn, float32x2_t Dd, __ghs_c_int__ lane);
void __simd_vst1q_lane_s8(int8_t * Rn, int8x16_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_s16(int16_t * Rn, int16x8_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_s32(int32_t * Rn, int32x4_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_s64(int64_t * Rn, int64x2_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_u8(uint8_t * Rn, uint8x16_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_u16(uint16_t * Rn, uint16x8_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_u32(uint32_t * Rn, uint32x4_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_u64(uint64_t * Rn, uint64x2_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_p8(poly8_t * Rn, poly8x16_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_p16(poly16_t * Rn, poly16x8_t Qd, __ghs_c_int__ lane);
void __simd_vst1q_lane_f32(float32_t * Rn, float32x4_t Qd, __ghs_c_int__ lane);

void __simd_vst2_s8(int8_t * Rn, int8x8x2_t Qd);
void __simd_vst2_s16(int16_t * Rn, int16x4x2_t Qd);
void __simd_vst2_s32(int32_t * Rn, int32x2x2_t Qd);
void __simd_vst2_s64(int64_t * Rn, int64x1x2_t Qd);
void __simd_vst2_u8(uint8_t * Rn, uint8x8x2_t Qd);
void __simd_vst2_u16(uint16_t * Rn, uint16x4x2_t Qd);
void __simd_vst2_u32(uint32_t * Rn, uint32x2x2_t Qd);
void __simd_vst2_u64(uint64_t * Rn, uint64x1x2_t Qd);
void __simd_vst2_p8(poly8_t * Rn, poly8x8x2_t Qd);
void __simd_vst2_p16(poly16_t * Rn, poly16x4x2_t Qd);
void __simd_vst2_f32(float32_t * Rn, float32x2x2_t Qd);
void __simd_vst2q_s8(int8_t * Rn, int8x16x2_t Vd);
void __simd_vst2q_s16(int16_t * Rn, int16x8x2_t Vd);
void __simd_vst2q_s32(int32_t * Rn, int32x4x2_t Vd);
void __simd_vst2q_u8(uint8_t * Rn, uint8x16x2_t Vd);
void __simd_vst2q_u16(uint16_t * Rn, uint16x8x2_t Vd);
void __simd_vst2q_u32(uint32_t * Rn, uint32x4x2_t Vd);
void __simd_vst2q_p8(poly8_t * Rn, poly8x16x2_t Vd);
void __simd_vst2q_p16(poly16_t * Rn, poly16x8x2_t Vd);
void __simd_vst2q_f32(float32_t * Rn, float32x4x2_t Vd);
void __simd_vst2_lane_s8(int8_t * Rn, int8x8x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_s16(int16_t * Rn, int16x4x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_s32(int32_t * Rn, int32x2x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_u8(uint8_t * Rn, uint8x8x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_u16(uint16_t * Rn, uint16x4x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_u32(uint32_t * Rn, uint32x2x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_p8(poly8_t * Rn, poly8x8x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_p16(poly16_t * Rn, poly16x4x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2_lane_f32(float32_t * Rn, float32x2x2_t Qd, __ghs_c_int__ lane);
void __simd_vst2q_lane_s16(int16_t * Rn, int16x8x2_t Vd, __ghs_c_int__ lane);
void __simd_vst2q_lane_s32(int32_t * Rn, int32x4x2_t Vd, __ghs_c_int__ lane);
void __simd_vst2q_lane_u16(uint16_t * Rn, uint16x8x2_t Vd, __ghs_c_int__ lane);
void __simd_vst2q_lane_u32(uint32_t * Rn, uint32x4x2_t Vd, __ghs_c_int__ lane);
void __simd_vst2q_lane_p16(poly16_t * Rn, poly16x8x2_t Vd, __ghs_c_int__ lane);
void __simd_vst2q_lane_f32(float32_t * Rn, float32x4x2_t Vd, __ghs_c_int__ lane);

void __simd_vst3_s8(int8_t * Rn, int8x8x3_t Qd);
void __simd_vst3_s16(int16_t * Rn, int16x4x3_t Qd);
void __simd_vst3_s32(int32_t * Rn, int32x2x3_t Qd);
void __simd_vst3_s64(int64_t * Rn, int64x1x3_t Qd);
void __simd_vst3_u8(uint8_t * Rn, uint8x8x3_t Qd);
void __simd_vst3_u16(uint16_t * Rn, uint16x4x3_t Qd);
void __simd_vst3_u32(uint32_t * Rn, uint32x2x3_t Qd);
void __simd_vst3_u64(uint64_t * Rn, uint64x1x3_t Qd);
void __simd_vst3_p8(poly8_t * Rn, poly8x8x3_t Qd);
void __simd_vst3_p16(poly16_t * Rn, poly16x4x3_t Qd);
void __simd_vst3_f32(float32_t * Rn, float32x2x3_t Qd);
void __simd_vst3q_s8(int8_t * Rn, int8x16x3_t Vd);
void __simd_vst3q_s16(int16_t * Rn, int16x8x3_t Vd);
void __simd_vst3q_s32(int32_t * Rn, int32x4x3_t Vd);
void __simd_vst3q_u8(uint8_t * Rn, uint8x16x3_t Vd);
void __simd_vst3q_u16(uint16_t * Rn, uint16x8x3_t Vd);
void __simd_vst3q_u32(uint32_t * Rn, uint32x4x3_t Vd);
void __simd_vst3q_p8(poly8_t * Rn, poly8x16x3_t Vd);
void __simd_vst3q_p16(poly16_t * Rn, poly16x8x3_t Vd);
void __simd_vst3q_f32(float32_t * Rn, float32x4x3_t Vd);
void __simd_vst3_lane_s8(int8_t * Rn, int8x8x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_s16(int16_t * Rn, int16x4x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_s32(int32_t * Rn, int32x2x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_u8(uint8_t * Rn, uint8x8x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_u16(uint16_t * Rn, uint16x4x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_u32(uint32_t * Rn, uint32x2x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_p8(poly8_t * Rn, poly8x8x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_p16(poly16_t * Rn, poly16x4x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3_lane_f32(float32_t * Rn, float32x2x3_t Qd, __ghs_c_int__ lane);
void __simd_vst3q_lane_s16(int16_t * Rn, int16x8x3_t Vd, __ghs_c_int__ lane);
void __simd_vst3q_lane_s32(int32_t * Rn, int32x4x3_t Vd, __ghs_c_int__ lane);
void __simd_vst3q_lane_u16(uint16_t * Rn, uint16x8x3_t Vd, __ghs_c_int__ lane);
void __simd_vst3q_lane_u32(uint32_t * Rn, uint32x4x3_t Vd, __ghs_c_int__ lane);
void __simd_vst3q_lane_p16(poly16_t * Rn, poly16x8x3_t Vd, __ghs_c_int__ lane);
void __simd_vst3q_lane_f32(float32_t * Rn, float32x4x3_t Vd, __ghs_c_int__ lane);

void __simd_vst4_s8(int8_t * Rn, int8x8x4_t Qd);
void __simd_vst4_s16(int16_t * Rn, int16x4x4_t Qd);
void __simd_vst4_s32(int32_t * Rn, int32x2x4_t Qd);
void __simd_vst4_s64(int64_t * Rn, int64x1x4_t Qd);
void __simd_vst4_u8(uint8_t * Rn, uint8x8x4_t Qd);
void __simd_vst4_u16(uint16_t * Rn, uint16x4x4_t Qd);
void __simd_vst4_u32(uint32_t * Rn, uint32x2x4_t Qd);
void __simd_vst4_u64(uint64_t * Rn, uint64x1x4_t Qd);
void __simd_vst4_p8(poly8_t * Rn, poly8x8x4_t Qd);
void __simd_vst4_p16(poly16_t * Rn, poly16x4x4_t Qd);
void __simd_vst4_f32(float32_t * Rn, float32x2x4_t Qd);
void __simd_vst4q_s8(int8_t * Rn, int8x16x4_t Vd);
void __simd_vst4q_s16(int16_t * Rn, int16x8x4_t Vd);
void __simd_vst4q_s32(int32_t * Rn, int32x4x4_t Vd);
void __simd_vst4q_u8(uint8_t * Rn, uint8x16x4_t Vd);
void __simd_vst4q_u16(uint16_t * Rn, uint16x8x4_t Vd);
void __simd_vst4q_u32(uint32_t * Rn, uint32x4x4_t Vd);
void __simd_vst4q_p8(poly8_t * Rn, poly8x16x4_t Vd);
void __simd_vst4q_p16(poly16_t * Rn, poly16x8x4_t Vd);
void __simd_vst4q_f32(float32_t * Rn, float32x4x4_t Vd);
void __simd_vst4_lane_s8(int8_t * Rn, int8x8x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_s16(int16_t * Rn, int16x4x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_s32(int32_t * Rn, int32x2x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_u8(uint8_t * Rn, uint8x8x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_u16(uint16_t * Rn, uint16x4x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_u32(uint32_t * Rn, uint32x2x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_p8(poly8_t * Rn, poly8x8x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_p16(poly16_t * Rn, poly16x4x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4_lane_f32(float32_t * Rn, float32x2x4_t Qd, __ghs_c_int__ lane);
void __simd_vst4q_lane_s16(int16_t * Rn, int16x8x4_t Vd, __ghs_c_int__ lane);
void __simd_vst4q_lane_s32(int32_t * Rn, int32x4x4_t Vd, __ghs_c_int__ lane);
void __simd_vst4q_lane_u16(uint16_t * Rn, uint16x8x4_t Vd, __ghs_c_int__ lane);
void __simd_vst4q_lane_u32(uint32_t * Rn, uint32x4x4_t Vd, __ghs_c_int__ lane);
void __simd_vst4q_lane_p16(poly16_t * Rn, poly16x8x4_t Vd, __ghs_c_int__ lane);
void __simd_vst4q_lane_f32(float32_t * Rn, float32x4x4_t Vd, __ghs_c_int__ lane);

int8x8_t __simd_vld2_stride_s8(const int8_t * Rn) __NEONPURE;
int16x4_t __simd_vld2_stride_s16(const int16_t * Rn) __NEONPURE;
int32x2_t __simd_vld2_stride_s32(const int32_t * Rn) __NEONPURE;
uint8x8_t __simd_vld2_stride_u8(const uint8_t * Rn) __NEONPURE;
uint16x4_t __simd_vld2_stride_u16(const uint16_t * Rn) __NEONPURE;
uint32x2_t __simd_vld2_stride_u32(const uint32_t * Rn) __NEONPURE;
float32x2_t __simd_vld2_stride_f32(const float32_t * Rn) __NEONPURE;
int8x16_t __simd_vld2q_stride_s8(const int8_t * Rn) __NEONPURE;
int16x8_t __simd_vld2q_stride_s16(const int16_t * Rn) __NEONPURE;
int32x4_t __simd_vld2q_stride_s32(const int32_t * Rn) __NEONPURE;
uint8x16_t __simd_vld2q_stride_u8(const uint8_t * Rn) __NEONPURE;
uint16x8_t __simd_vld2q_stride_u16(const uint16_t * Rn) __NEONPURE;
uint32x4_t __simd_vld2q_stride_u32(const uint32_t * Rn) __NEONPURE;
float32x4_t __simd_vld2q_stride_f32(const float32_t * Rn) __NEONPURE;

int8x8_t __simd_vld3_stride_s8(const int8_t * Rn) __NEONPURE;
int16x4_t __simd_vld3_stride_s16(const int16_t * Rn) __NEONPURE;
int32x2_t __simd_vld3_stride_s32(const int32_t * Rn) __NEONPURE;
uint8x8_t __simd_vld3_stride_u8(const uint8_t * Rn) __NEONPURE;
uint16x4_t __simd_vld3_stride_u16(const uint16_t * Rn) __NEONPURE;
uint32x2_t __simd_vld3_stride_u32(const uint32_t * Rn) __NEONPURE;
float32x2_t __simd_vld3_stride_f32(const float32_t * Rn) __NEONPURE;
int8x16_t __simd_vld3q_stride_s8(const int8_t * Rn) __NEONPURE;
int16x8_t __simd_vld3q_stride_s16(const int16_t * Rn) __NEONPURE;
int32x4_t __simd_vld3q_stride_s32(const int32_t * Rn) __NEONPURE;
uint8x16_t __simd_vld3q_stride_u8(const uint8_t * Rn) __NEONPURE;
uint16x8_t __simd_vld3q_stride_u16(const uint16_t * Rn) __NEONPURE;
uint32x4_t __simd_vld3q_stride_u32(const uint32_t * Rn) __NEONPURE;
float32x4_t __simd_vld3q_stride_f32(const float32_t * Rn) __NEONPURE;

int8x8_t __simd_vld4_stride_s8(const int8_t * Rn) __NEONPURE;
int16x4_t __simd_vld4_stride_s16(const int16_t * Rn) __NEONPURE;
int32x2_t __simd_vld4_stride_s32(const int32_t * Rn) __NEONPURE;
uint8x8_t __simd_vld4_stride_u8(const uint8_t * Rn) __NEONPURE;
uint16x4_t __simd_vld4_stride_u16(const uint16_t * Rn) __NEONPURE;
uint32x2_t __simd_vld4_stride_u32(const uint32_t * Rn) __NEONPURE;
float32x2_t __simd_vld4_stride_f32(const float32_t * Rn) __NEONPURE;
int8x16_t __simd_vld4q_stride_s8(const int8_t * Rn) __NEONPURE;
int16x8_t __simd_vld4q_stride_s16(const int16_t * Rn) __NEONPURE;
int32x4_t __simd_vld4q_stride_s32(const int32_t * Rn) __NEONPURE;
uint8x16_t __simd_vld4q_stride_u8(const uint8_t * Rn) __NEONPURE;
uint16x8_t __simd_vld4q_stride_u16(const uint16_t * Rn) __NEONPURE;
uint32x4_t __simd_vld4q_stride_u32(const uint32_t * Rn) __NEONPURE;
float32x4_t __simd_vld4q_stride_f32(const float32_t * Rn) __NEONPURE;

void __simd_vst2_stride_s8(int8_t * Rn, int8x8_t Qd);
void __simd_vst2_stride_s16(int16_t * Rn, int16x4_t Qd);
void __simd_vst2_stride_s32(int32_t * Rn, int32x2_t Qd);
void __simd_vst2_stride_u8(uint8_t * Rn, uint8x8_t Qd);
void __simd_vst2_stride_u16(uint16_t * Rn, uint16x4_t Qd);
void __simd_vst2_stride_u32(uint32_t * Rn, uint32x2_t Qd);
void __simd_vst2_stride_f32(float32_t * Rn, float32x2_t Qd);
void __simd_vst2q_stride_s8(int8_t * Rn, int8x16_t Vd);
void __simd_vst2q_stride_s16(int16_t * Rn, int16x8_t Vd);
void __simd_vst2q_stride_s32(int32_t * Rn, int32x4_t Vd);
void __simd_vst2q_stride_u8(uint8_t * Rn, uint8x16_t Vd);
void __simd_vst2q_stride_u16(uint16_t * Rn, uint16x8_t Vd);
void __simd_vst2q_stride_u32(uint32_t * Rn, uint32x4_t Vd);
void __simd_vst2q_stride_f32(float32_t * Rn, float32x4_t Vd);

void __simd_vst3_stride_s8(int8_t * Rn, int8x8_t Qd);
void __simd_vst3_stride_s16(int16_t * Rn, int16x4_t Qd);
void __simd_vst3_stride_s32(int32_t * Rn, int32x2_t Qd);
void __simd_vst3_stride_u8(uint8_t * Rn, uint8x8_t Qd);
void __simd_vst3_stride_u16(uint16_t * Rn, uint16x4_t Qd);
void __simd_vst3_stride_u32(uint32_t * Rn, uint32x2_t Qd);
void __simd_vst3_stride_f32(float32_t * Rn, float32x2_t Qd);
void __simd_vst3q_stride_s8(int8_t * Rn, int8x16_t Vd);
void __simd_vst3q_stride_s16(int16_t * Rn, int16x8_t Vd);
void __simd_vst3q_stride_s32(int32_t * Rn, int32x4_t Vd);
void __simd_vst3q_stride_u8(uint8_t * Rn, uint8x16_t Vd);
void __simd_vst3q_stride_u16(uint16_t * Rn, uint16x8_t Vd);
void __simd_vst3q_stride_u32(uint32_t * Rn, uint32x4_t Vd);
void __simd_vst3q_stride_f32(float32_t * Rn, float32x4_t Vd);

void __simd_vst4_stride_s8(int8_t * Rn, int8x8_t Qd);
void __simd_vst4_stride_s16(int16_t * Rn, int16x4_t Qd);
void __simd_vst4_stride_s32(int32_t * Rn, int32x2_t Qd);
void __simd_vst4_stride_u8(uint8_t * Rn, uint8x8_t Qd);
void __simd_vst4_stride_u16(uint16_t * Rn, uint16x4_t Qd);
void __simd_vst4_stride_u32(uint32_t * Rn, uint32x2_t Qd);
void __simd_vst4_stride_f32(float32_t * Rn, float32x2_t Qd);
void __simd_vst4q_stride_s8(int8_t * Rn, int8x16_t Vd);
void __simd_vst4q_stride_s16(int16_t * Rn, int16x8_t Vd);
void __simd_vst4q_stride_s32(int32_t * Rn, int32x4_t Vd);
void __simd_vst4q_stride_u8(uint8_t * Rn, uint8x16_t Vd);
void __simd_vst4q_stride_u16(uint16_t * Rn, uint16x8_t Vd);
void __simd_vst4q_stride_u32(uint32_t * Rn, uint32x4_t Vd);
void __simd_vst4q_stride_f32(float32_t * Rn, float32x4_t Vd);

int8_t __simd_vget_lane_s8(int8x8_t Dn, __ghs_c_int__ lane) __NEONSAFE;
int16_t __simd_vget_lane_s16(int16x4_t Dn, __ghs_c_int__ lane) __NEONSAFE;
int32_t __simd_vget_lane_s32(int32x2_t Dn, __ghs_c_int__ lane) __NEONSAFE;
int64_t __simd_vget_lane_s64(int64x1_t Dn, __ghs_c_int__ lane) __NEONSAFE;
uint8_t __simd_vget_lane_u8(uint8x8_t Dn, __ghs_c_int__ lane) __NEONSAFE;
uint16_t __simd_vget_lane_u16(uint16x4_t Dn, __ghs_c_int__ lane) __NEONSAFE;
uint32_t __simd_vget_lane_u32(uint32x2_t Dn, __ghs_c_int__ lane) __NEONSAFE;
uint64_t __simd_vget_lane_u64(uint64x1_t Dn, __ghs_c_int__ lane) __NEONSAFE;
poly8_t __simd_vget_lane_p8(poly8x8_t Dn, __ghs_c_int__ lane) __NEONSAFE;
poly16_t __simd_vget_lane_p16(poly16x4_t Dn, __ghs_c_int__ lane) __NEONSAFE;
float32_t __simd_vget_lane_f32(float32x2_t Dn, __ghs_c_int__ lane) __NEONSAFE;
int8_t __simd_vgetq_lane_s8(int8x16_t Qn, __ghs_c_int__ lane) __NEONSAFE;
int16_t __simd_vgetq_lane_s16(int16x8_t Qn, __ghs_c_int__ lane) __NEONSAFE;
int32_t __simd_vgetq_lane_s32(int32x4_t Qn, __ghs_c_int__ lane) __NEONSAFE;
int64_t __simd_vgetq_lane_s64(int64x2_t Qn, __ghs_c_int__ lane) __NEONSAFE;
uint8_t __simd_vgetq_lane_u8(uint8x16_t Qn, __ghs_c_int__ lane) __NEONSAFE;
uint16_t __simd_vgetq_lane_u16(uint16x8_t Qn, __ghs_c_int__ lane) __NEONSAFE;
uint32_t __simd_vgetq_lane_u32(uint32x4_t Qn, __ghs_c_int__ lane) __NEONSAFE;
uint64_t __simd_vgetq_lane_u64(uint64x2_t Qn, __ghs_c_int__ lane) __NEONSAFE;
poly8_t __simd_vgetq_lane_p8(poly8x16_t Qn, __ghs_c_int__ lane) __NEONSAFE;
poly16_t __simd_vgetq_lane_p16(poly16x8_t Qn, __ghs_c_int__ lane) __NEONSAFE;
float32_t __simd_vgetq_lane_f32(float32x4_t Qn, __ghs_c_int__ lane) __NEONSAFE;
int8x8_t __simd_vset_lane_s8(int8_t Rt, int8x8_t Dd, __ghs_c_int__ lane) __NEONSAFE;
int16x4_t __simd_vset_lane_s16(int16_t Rt, int16x4_t Dd, __ghs_c_int__ lane) __NEONSAFE;
int32x2_t __simd_vset_lane_s32(int32_t Rt, int32x2_t Dd, __ghs_c_int__ lane) __NEONSAFE;
int64x1_t __simd_vset_lane_s64(int64_t Rt, int64x1_t Dd, __ghs_c_int__ lane) __NEONSAFE;
uint8x8_t __simd_vset_lane_u8(uint8_t Rt, uint8x8_t Dd, __ghs_c_int__ lane) __NEONSAFE;
uint16x4_t __simd_vset_lane_u16(uint16_t Rt, uint16x4_t Dd, __ghs_c_int__ lane) __NEONSAFE;
uint32x2_t __simd_vset_lane_u32(uint32_t Rt, uint32x2_t Dd, __ghs_c_int__ lane) __NEONSAFE;
uint64x1_t __simd_vset_lane_u64(uint64_t Rt, uint64x1_t Dd, __ghs_c_int__ lane) __NEONSAFE;
poly8x8_t __simd_vset_lane_p8(poly8_t Rt, poly8x8_t Dd, __ghs_c_int__ lane) __NEONSAFE;
poly16x4_t __simd_vset_lane_p16(poly16_t Rt, poly16x4_t Dd, __ghs_c_int__ lane) __NEONSAFE;
float32x2_t __simd_vset_lane_f32(float32_t Rt, float32x2_t Dd, __ghs_c_int__ lane) __NEONSAFE;
int8x16_t __simd_vsetq_lane_s8(int8_t Rt, int8x16_t Qd, __ghs_c_int__ lane) __NEONSAFE;
int16x8_t __simd_vsetq_lane_s16(int16_t Rt, int16x8_t Qd, __ghs_c_int__ lane) __NEONSAFE;
int32x4_t __simd_vsetq_lane_s32(int32_t Rt, int32x4_t Qd, __ghs_c_int__ lane) __NEONSAFE;
int64x2_t __simd_vsetq_lane_s64(int64_t Rt, int64x2_t Qd, __ghs_c_int__ lane) __NEONSAFE;
uint8x16_t __simd_vsetq_lane_u8(uint8_t Rt, uint8x16_t Qd, __ghs_c_int__ lane) __NEONSAFE;
uint16x8_t __simd_vsetq_lane_u16(uint16_t Rt, uint16x8_t Qd, __ghs_c_int__ lane) __NEONSAFE;
uint32x4_t __simd_vsetq_lane_u32(uint32_t Rt, uint32x4_t Qd, __ghs_c_int__ lane) __NEONSAFE;
uint64x2_t __simd_vsetq_lane_u64(uint64_t Rt, uint64x2_t Qd, __ghs_c_int__ lane) __NEONSAFE;
poly8x16_t __simd_vsetq_lane_p8(poly8_t Rt, poly8x16_t Qd, __ghs_c_int__ lane) __NEONSAFE;
poly16x8_t __simd_vsetq_lane_p16(poly16_t Rt, poly16x8_t Qd, __ghs_c_int__ lane) __NEONSAFE;
float32x4_t __simd_vsetq_lane_f32(float32_t Rt, float32x4_t Qd, __ghs_c_int__ lane) __NEONSAFE;
int8x8_t __simd_vcreate_s8(uint64_t Rt) __NEONSAFE;
int16x4_t __simd_vcreate_s16(uint64_t Rt) __NEONSAFE;
int32x2_t __simd_vcreate_s32(uint64_t Rt) __NEONSAFE;
int64x1_t __simd_vcreate_s64(uint64_t Rt) __NEONSAFE;
uint8x8_t __simd_vcreate_u8(uint64_t Rt) __NEONSAFE;
uint16x4_t __simd_vcreate_u16(uint64_t Rt) __NEONSAFE;
uint32x2_t __simd_vcreate_u32(uint64_t Rt) __NEONSAFE;
uint64x1_t __simd_vcreate_u64(uint64_t Rt) __NEONSAFE;
poly8x8_t __simd_vcreate_p8(uint64_t Rt) __NEONSAFE;
poly16x4_t __simd_vcreate_p16(uint64_t Rt) __NEONSAFE;
float32x2_t __simd_vcreate_f32(uint64_t Rt) __NEONSAFE;
int8x16_t __simd_vcombine_s8(int8x8_t Da, int8x8_t Db) __NEONSAFE;
int16x8_t __simd_vcombine_s16(int16x4_t Da, int16x4_t Db) __NEONSAFE;
int32x4_t __simd_vcombine_s32(int32x2_t Da, int32x2_t Db) __NEONSAFE;
int64x2_t __simd_vcombine_s64(int64x1_t Da, int64x1_t Db) __NEONSAFE;
uint8x16_t __simd_vcombine_u8(uint8x8_t Da, uint8x8_t Db) __NEONSAFE;
uint16x8_t __simd_vcombine_u16(uint16x4_t Da, uint16x4_t Db) __NEONSAFE;
uint32x4_t __simd_vcombine_u32(uint32x2_t Da, uint32x2_t Db) __NEONSAFE;
uint64x2_t __simd_vcombine_u64(uint64x1_t Da, uint64x1_t Db) __NEONSAFE;
poly8x16_t __simd_vcombine_p8(poly8x8_t Da, poly8x8_t Db) __NEONSAFE;
poly16x8_t __simd_vcombine_p16(poly16x4_t Da, poly16x4_t Db) __NEONSAFE;
float32x4_t __simd_vcombine_f32(float32x2_t Da, float32x2_t Db) __NEONSAFE;
int8x8_t __simd_vget_high_s8(int8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vget_high_s16(int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vget_high_s32(int32x4_t Qm) __NEONSAFE;
int64x1_t __simd_vget_high_s64(int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vget_high_u8(uint8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vget_high_u16(uint16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vget_high_u32(uint32x4_t Qm) __NEONSAFE;
uint64x1_t __simd_vget_high_u64(uint64x2_t Qm) __NEONSAFE;
poly8x8_t __simd_vget_high_p8(poly8x16_t Qm) __NEONSAFE;
poly16x4_t __simd_vget_high_p16(poly16x8_t Qm) __NEONSAFE;
float32x2_t __simd_vget_high_f32(float32x4_t Qm) __NEONSAFE;
int8x8_t __simd_vget_low_s8(int8x16_t Qm) __NEONSAFE;
int16x4_t __simd_vget_low_s16(int16x8_t Qm) __NEONSAFE;
int32x2_t __simd_vget_low_s32(int32x4_t Qm) __NEONSAFE;
int64x1_t __simd_vget_low_s64(int64x2_t Qm) __NEONSAFE;
uint8x8_t __simd_vget_low_u8(uint8x16_t Qm) __NEONSAFE;
uint16x4_t __simd_vget_low_u16(uint16x8_t Qm) __NEONSAFE;
uint32x2_t __simd_vget_low_u32(uint32x4_t Qm) __NEONSAFE;
uint64x1_t __simd_vget_low_u64(uint64x2_t Qm) __NEONSAFE;
poly8x8_t __simd_vget_low_p8(poly8x16_t Qm) __NEONSAFE;
poly16x4_t __simd_vget_low_p16(poly16x8_t Qm) __NEONSAFE;
float32x2_t __simd_vget_low_f32(float32x4_t Qm) __NEONSAFE;

int8x8_t __simd_vreinterpret_s8_s16(int16x4_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_s32(int32x2_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_s64(int64x1_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u8(uint8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u16(uint16x4_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u32(uint32x2_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_u64(uint64x1_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_p8(poly8x8_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_p16(poly16x4_t Dm) __NEONSAFE;
int8x8_t __simd_vreinterpret_s8_f32(float32x2_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_s8(int8x8_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_s32(int32x2_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_s64(int64x1_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u8(uint8x8_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u16(uint16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u32(uint32x2_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_u64(uint64x1_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_p8(poly8x8_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_p16(poly16x4_t Dm) __NEONSAFE;
int16x4_t __simd_vreinterpret_s16_f32(float32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_s8(int8x8_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_s16(int16x4_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_s64(int64x1_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u8(uint8x8_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u16(uint16x4_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u32(uint32x2_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_u64(uint64x1_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_p8(poly8x8_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_p16(poly16x4_t Dm) __NEONSAFE;
int32x2_t __simd_vreinterpret_s32_f32(float32x2_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_s8(int8x8_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_s16(int16x4_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_s32(int32x2_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u8(uint8x8_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u16(uint16x4_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u32(uint32x2_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_u64(uint64x1_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_p8(poly8x8_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_p16(poly16x4_t Dm) __NEONSAFE;
int64x1_t __simd_vreinterpret_s64_f32(float32x2_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s8(int8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s16(int16x4_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s32(int32x2_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_s64(int64x1_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_u16(uint16x4_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_u32(uint32x2_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_u64(uint64x1_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_p8(poly8x8_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_p16(poly16x4_t Dm) __NEONSAFE;
uint8x8_t __simd_vreinterpret_u8_f32(float32x2_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s8(int8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s16(int16x4_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s32(int32x2_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_s64(int64x1_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_u8(uint8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_u32(uint32x2_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_u64(uint64x1_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_p8(poly8x8_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_p16(poly16x4_t Dm) __NEONSAFE;
uint16x4_t __simd_vreinterpret_u16_f32(float32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s8(int8x8_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s16(int16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s32(int32x2_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_s64(int64x1_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_u8(uint8x8_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_u16(uint16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_u64(uint64x1_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_p8(poly8x8_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_p16(poly16x4_t Dm) __NEONSAFE;
uint32x2_t __simd_vreinterpret_u32_f32(float32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s8(int8x8_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s16(int16x4_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s32(int32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_s64(int64x1_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_u8(uint8x8_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_u16(uint16x4_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_u32(uint32x2_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_p8(poly8x8_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_p16(poly16x4_t Dm) __NEONSAFE;
uint64x1_t __simd_vreinterpret_u64_f32(float32x2_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s8(int8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s16(int16x4_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s32(int32x2_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_s64(int64x1_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u8(uint8x8_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u16(uint16x4_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u32(uint32x2_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_u64(uint64x1_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_p16(poly16x4_t Dm) __NEONSAFE;
poly8x8_t __simd_vreinterpret_p8_f32(float32x2_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s8(int8x8_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s16(int16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s32(int32x2_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_s64(int64x1_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u8(uint8x8_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u16(uint16x4_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u32(uint32x2_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_u64(uint64x1_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_p8(poly8x8_t Dm) __NEONSAFE;
poly16x4_t __simd_vreinterpret_p16_f32(float32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s8(int8x8_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s16(int16x4_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s32(int32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_s64(int64x1_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u8(uint8x8_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u16(uint16x4_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u32(uint32x2_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_u64(uint64x1_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_p8(poly8x8_t Dm) __NEONSAFE;
float32x2_t __simd_vreinterpret_f32_p16(poly16x4_t Dm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_s16(int16x8_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_s32(int32x4_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_s64(int64x2_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u8(uint8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u16(uint16x8_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u32(uint32x4_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_u64(uint64x2_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_p8(poly8x16_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_p16(poly16x8_t Qm) __NEONSAFE;
int8x16_t __simd_vreinterpretq_s8_f32(float32x4_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_s8(int8x16_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_s32(int32x4_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_s64(int64x2_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u8(uint8x16_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u16(uint16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u32(uint32x4_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_u64(uint64x2_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_p8(poly8x16_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_p16(poly16x8_t Qm) __NEONSAFE;
int16x8_t __simd_vreinterpretq_s16_f32(float32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_s8(int8x16_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_s16(int16x8_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_s64(int64x2_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u8(uint8x16_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u16(uint16x8_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u32(uint32x4_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_u64(uint64x2_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_p8(poly8x16_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_p16(poly16x8_t Qm) __NEONSAFE;
int32x4_t __simd_vreinterpretq_s32_f32(float32x4_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_s8(int8x16_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_s16(int16x8_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_s32(int32x4_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u8(uint8x16_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u16(uint16x8_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u32(uint32x4_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_u64(uint64x2_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_p8(poly8x16_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_p16(poly16x8_t Qm) __NEONSAFE;
int64x2_t __simd_vreinterpretq_s64_f32(float32x4_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s8(int8x16_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s16(int16x8_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s32(int32x4_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_s64(int64x2_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_u16(uint16x8_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_u32(uint32x4_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_u64(uint64x2_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_p8(poly8x16_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_p16(poly16x8_t Qm) __NEONSAFE;
uint8x16_t __simd_vreinterpretq_u8_f32(float32x4_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s8(int8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s16(int16x8_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s32(int32x4_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_s64(int64x2_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_u8(uint8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_u32(uint32x4_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_u64(uint64x2_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_p8(poly8x16_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_p16(poly16x8_t Qm) __NEONSAFE;
uint16x8_t __simd_vreinterpretq_u16_f32(float32x4_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s8(int8x16_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s16(int16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s32(int32x4_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_s64(int64x2_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_u8(uint8x16_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_u16(uint16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_u64(uint64x2_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_p8(poly8x16_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_p16(poly16x8_t Qm) __NEONSAFE;
uint32x4_t __simd_vreinterpretq_u32_f32(float32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s8(int8x16_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s16(int16x8_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s32(int32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_s64(int64x2_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_u8(uint8x16_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_u16(uint16x8_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_u32(uint32x4_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_p8(poly8x16_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_p16(poly16x8_t Qm) __NEONSAFE;
uint64x2_t __simd_vreinterpretq_u64_f32(float32x4_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s8(int8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s16(int16x8_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s32(int32x4_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_s64(int64x2_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u8(uint8x16_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u16(uint16x8_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u32(uint32x4_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_u64(uint64x2_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_p16(poly16x8_t Qm) __NEONSAFE;
poly8x16_t __simd_vreinterpretq_p8_f32(float32x4_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s8(int8x16_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s16(int16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s32(int32x4_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_s64(int64x2_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u8(uint8x16_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u16(uint16x8_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u32(uint32x4_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_u64(uint64x2_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_p8(poly8x16_t Qm) __NEONSAFE;
poly16x8_t __simd_vreinterpretq_p16_f32(float32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s8(int8x16_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s16(int16x8_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s32(int32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_s64(int64x2_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u8(uint8x16_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u16(uint16x8_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u32(uint32x4_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_u64(uint64x2_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_p8(poly8x16_t Qm) __NEONSAFE;
float32x4_t __simd_vreinterpretq_f32_p16(poly16x8_t Qm) __NEONSAFE;


#define vhadd_s8(a,b) __simd_vhadd_s8(a,b)
#define vhaddq_s8(a,b) __simd_vhaddq_s8(a,b)
#define vhadd_u8(a,b) __simd_vhadd_u8(a,b)
#define vhaddq_u8(a,b) __simd_vhaddq_u8(a,b)
#define vhadd_s16(a,b) __simd_vhadd_s16(a,b)
#define vhaddq_s16(a,b) __simd_vhaddq_s16(a,b)
#define vhadd_u16(a,b) __simd_vhadd_u16(a,b)
#define vhaddq_u16(a,b) __simd_vhaddq_u16(a,b)
#define vhadd_s32(a,b) __simd_vhadd_s32(a,b)
#define vhaddq_s32(a,b) __simd_vhaddq_s32(a,b)
#define vhadd_u32(a,b) __simd_vhadd_u32(a,b)
#define vhaddq_u32(a,b) __simd_vhaddq_u32(a,b)
#define vhsub_s8(a,b) __simd_vhsub_s8(a,b)
#define vhsubq_s8(a,b) __simd_vhsubq_s8(a,b)
#define vhsub_u8(a,b) __simd_vhsub_u8(a,b)
#define vhsubq_u8(a,b) __simd_vhsubq_u8(a,b)
#define vhsub_s16(a,b) __simd_vhsub_s16(a,b)
#define vhsubq_s16(a,b) __simd_vhsubq_s16(a,b)
#define vhsub_u16(a,b) __simd_vhsub_u16(a,b)
#define vhsubq_u16(a,b) __simd_vhsubq_u16(a,b)
#define vhsub_s32(a,b) __simd_vhsub_s32(a,b)
#define vhsubq_s32(a,b) __simd_vhsubq_s32(a,b)
#define vhsub_u32(a,b) __simd_vhsub_u32(a,b)
#define vhsubq_u32(a,b) __simd_vhsubq_u32(a,b)
#define vqadd_s8(a,b) __simd_vqadd_s8(a,b)
#define vqaddq_s8(a,b) __simd_vqaddq_s8(a,b)
#define vqadd_u8(a,b) __simd_vqadd_u8(a,b)
#define vqaddq_u8(a,b) __simd_vqaddq_u8(a,b)
#define vqadd_s16(a,b) __simd_vqadd_s16(a,b)
#define vqaddq_s16(a,b) __simd_vqaddq_s16(a,b)
#define vqadd_u16(a,b) __simd_vqadd_u16(a,b)
#define vqaddq_u16(a,b) __simd_vqaddq_u16(a,b)
#define vqadd_s32(a,b) __simd_vqadd_s32(a,b)
#define vqaddq_s32(a,b) __simd_vqaddq_s32(a,b)
#define vqadd_u32(a,b) __simd_vqadd_u32(a,b)
#define vqaddq_u32(a,b) __simd_vqaddq_u32(a,b)
#define vqadd_s64(a,b) __simd_vqadd_s64(a,b)
#define vqaddq_s64(a,b) __simd_vqaddq_s64(a,b)
#define vqadd_u64(a,b) __simd_vqadd_u64(a,b)
#define vqaddq_u64(a,b) __simd_vqaddq_u64(a,b)
#define vrhadd_s8(a,b) __simd_vrhadd_s8(a,b)
#define vrhaddq_s8(a,b) __simd_vrhaddq_s8(a,b)
#define vrhadd_u8(a,b) __simd_vrhadd_u8(a,b)
#define vrhaddq_u8(a,b) __simd_vrhaddq_u8(a,b)
#define vrhadd_s16(a,b) __simd_vrhadd_s16(a,b)
#define vrhaddq_s16(a,b) __simd_vrhaddq_s16(a,b)
#define vrhadd_u16(a,b) __simd_vrhadd_u16(a,b)
#define vrhaddq_u16(a,b) __simd_vrhaddq_u16(a,b)
#define vrhadd_s32(a,b) __simd_vrhadd_s32(a,b)
#define vrhaddq_s32(a,b) __simd_vrhaddq_s32(a,b)
#define vrhadd_u32(a,b) __simd_vrhadd_u32(a,b)
#define vrhaddq_u32(a,b) __simd_vrhaddq_u32(a,b)
#define vqsub_s8(a,b) __simd_vqsub_s8(a,b)
#define vqsubq_s8(a,b) __simd_vqsubq_s8(a,b)
#define vqsub_u8(a,b) __simd_vqsub_u8(a,b)
#define vqsubq_u8(a,b) __simd_vqsubq_u8(a,b)
#define vqsub_s16(a,b) __simd_vqsub_s16(a,b)
#define vqsubq_s16(a,b) __simd_vqsubq_s16(a,b)
#define vqsub_u16(a,b) __simd_vqsub_u16(a,b)
#define vqsubq_u16(a,b) __simd_vqsubq_u16(a,b)
#define vqsub_s32(a,b) __simd_vqsub_s32(a,b)
#define vqsubq_s32(a,b) __simd_vqsubq_s32(a,b)
#define vqsub_u32(a,b) __simd_vqsub_u32(a,b)
#define vqsubq_u32(a,b) __simd_vqsubq_u32(a,b)
#define vqsub_s64(a,b) __simd_vqsub_s64(a,b)
#define vqsubq_s64(a,b) __simd_vqsubq_s64(a,b)
#define vqsub_u64(a,b) __simd_vqsub_u64(a,b)
#define vqsubq_u64(a,b) __simd_vqsubq_u64(a,b)
#define vcgt_s8(a,b) __simd_vcgt_s8(a,b)
#define vcgtq_s8(a,b) __simd_vcgtq_s8(a,b)
#define vcgt_u8(a,b) __simd_vcgt_u8(a,b)
#define vcgtq_u8(a,b) __simd_vcgtq_u8(a,b)
#define vcgt_s16(a,b) __simd_vcgt_s16(a,b)
#define vcgtq_s16(a,b) __simd_vcgtq_s16(a,b)
#define vcgt_u16(a,b) __simd_vcgt_u16(a,b)
#define vcgtq_u16(a,b) __simd_vcgtq_u16(a,b)
#define vcgt_s32(a,b) __simd_vcgt_s32(a,b)
#define vcgtq_s32(a,b) __simd_vcgtq_s32(a,b)
#define vcgt_u32(a,b) __simd_vcgt_u32(a,b)
#define vcgtq_u32(a,b) __simd_vcgtq_u32(a,b)
#define vcge_s8(a,b) __simd_vcge_s8(a,b)
#define vcgeq_s8(a,b) __simd_vcgeq_s8(a,b)
#define vcge_u8(a,b) __simd_vcge_u8(a,b)
#define vcgeq_u8(a,b) __simd_vcgeq_u8(a,b)
#define vcge_s16(a,b) __simd_vcge_s16(a,b)
#define vcgeq_s16(a,b) __simd_vcgeq_s16(a,b)
#define vcge_u16(a,b) __simd_vcge_u16(a,b)
#define vcgeq_u16(a,b) __simd_vcgeq_u16(a,b)
#define vcge_s32(a,b) __simd_vcge_s32(a,b)
#define vcgeq_s32(a,b) __simd_vcgeq_s32(a,b)
#define vcge_u32(a,b) __simd_vcge_u32(a,b)
#define vcgeq_u32(a,b) __simd_vcgeq_u32(a,b)
#define vcle_s8(a,b) __simd_vcge_s8(b,a)
#define vcleq_s8(a,b) __simd_vcgeq_s8(b,a)
#define vcle_u8(a,b) __simd_vcge_u8(b,a)
#define vcleq_u8(a,b) __simd_vcgeq_u8(b,a)
#define vcle_s16(a,b) __simd_vcge_s16(b,a)
#define vcleq_s16(a,b) __simd_vcgeq_s16(b,a)
#define vcle_u16(a,b) __simd_vcge_u16(b,a)
#define vcleq_u16(a,b) __simd_vcgeq_u16(b,a)
#define vcle_s32(a,b) __simd_vcge_s32(b,a)
#define vcleq_s32(a,b) __simd_vcgeq_s32(b,a)
#define vcle_u32(a,b) __simd_vcge_u32(b,a)
#define vcleq_u32(a,b) __simd_vcgeq_u32(b,a)
#define vclt_s8(a,b) __simd_vcgt_s8(b,a)
#define vcltq_s8(a,b) __simd_vcgtq_s8(b,a)
#define vclt_u8(a,b) __simd_vcgt_u8(b,a)
#define vcltq_u8(a,b) __simd_vcgtq_u8(b,a)
#define vclt_s16(a,b) __simd_vcgt_s16(b,a)
#define vcltq_s16(a,b) __simd_vcgtq_s16(b,a)
#define vclt_u16(a,b) __simd_vcgt_u16(b,a)
#define vcltq_u16(a,b) __simd_vcgtq_u16(b,a)
#define vclt_s32(a,b) __simd_vcgt_s32(b,a)
#define vcltq_s32(a,b) __simd_vcgtq_s32(b,a)
#define vclt_u32(a,b) __simd_vcgt_u32(b,a)
#define vcltq_u32(a,b) __simd_vcgtq_u32(b,a)
#define vmax_s8(a,b) __simd_vmax_s8(a,b)
#define vmaxq_s8(a,b) __simd_vmaxq_s8(a,b)
#define vmax_u8(a,b) __simd_vmax_u8(a,b)
#define vmaxq_u8(a,b) __simd_vmaxq_u8(a,b)
#define vmax_s16(a,b) __simd_vmax_s16(a,b)
#define vmaxq_s16(a,b) __simd_vmaxq_s16(a,b)
#define vmax_u16(a,b) __simd_vmax_u16(a,b)
#define vmaxq_u16(a,b) __simd_vmaxq_u16(a,b)
#define vmax_s32(a,b) __simd_vmax_s32(a,b)
#define vmaxq_s32(a,b) __simd_vmaxq_s32(a,b)
#define vmax_u32(a,b) __simd_vmax_u32(a,b)
#define vmaxq_u32(a,b) __simd_vmaxq_u32(a,b)
#define vmin_s8(a,b) __simd_vmin_s8(a,b)
#define vminq_s8(a,b) __simd_vminq_s8(a,b)
#define vmin_u8(a,b) __simd_vmin_u8(a,b)
#define vminq_u8(a,b) __simd_vminq_u8(a,b)
#define vmin_s16(a,b) __simd_vmin_s16(a,b)
#define vminq_s16(a,b) __simd_vminq_s16(a,b)
#define vmin_u16(a,b) __simd_vmin_u16(a,b)
#define vminq_u16(a,b) __simd_vminq_u16(a,b)
#define vmin_s32(a,b) __simd_vmin_s32(a,b)
#define vminq_s32(a,b) __simd_vminq_s32(a,b)
#define vmin_u32(a,b) __simd_vmin_u32(a,b)
#define vminq_u32(a,b) __simd_vminq_u32(a,b)
#define vabd_s8(a,b) __simd_vabd_s8(a,b)
#define vabdq_s8(a,b) __simd_vabdq_s8(a,b)
#define vabd_u8(a,b) __simd_vabd_u8(a,b)
#define vabdq_u8(a,b) __simd_vabdq_u8(a,b)
#define vabd_s16(a,b) __simd_vabd_s16(a,b)
#define vabdq_s16(a,b) __simd_vabdq_s16(a,b)
#define vabd_u16(a,b) __simd_vabd_u16(a,b)
#define vabdq_u16(a,b) __simd_vabdq_u16(a,b)
#define vabd_s32(a,b) __simd_vabd_s32(a,b)
#define vabdq_s32(a,b) __simd_vabdq_s32(a,b)
#define vabd_u32(a,b) __simd_vabd_u32(a,b)
#define vabdq_u32(a,b) __simd_vabdq_u32(a,b)
#define vaba_s8(a,b,c) __simd_vaba_s8(a,b,c)
#define vabaq_s8(a,b,c) __simd_vabaq_s8(a,b,c)
#define vaba_u8(a,b,c) __simd_vaba_u8(a,b,c)
#define vabaq_u8(a,b,c) __simd_vabaq_u8(a,b,c)
#define vaba_s16(a,b,c) __simd_vaba_s16(a,b,c)
#define vabaq_s16(a,b,c) __simd_vabaq_s16(a,b,c)
#define vaba_u16(a,b,c) __simd_vaba_u16(a,b,c)
#define vabaq_u16(a,b,c) __simd_vabaq_u16(a,b,c)
#define vaba_s32(a,b,c) __simd_vaba_s32(a,b,c)
#define vabaq_s32(a,b,c) __simd_vabaq_s32(a,b,c)
#define vaba_u32(a,b,c) __simd_vaba_u32(a,b,c)
#define vabaq_u32(a,b,c) __simd_vabaq_u32(a,b,c)
#define vadd_u8(a,b) __simd_vadd_u8(a,b)
#define vadd_s8(a,b) __simd_vadd_s8(a,b)
#define vaddq_u8(a,b) __simd_vaddq_u8(a,b)
#define vaddq_s8(a,b) __simd_vaddq_s8(a,b)
#define vadd_u16(a,b) __simd_vadd_u16(a,b)
#define vadd_s16(a,b) __simd_vadd_s16(a,b)
#define vaddq_u16(a,b) __simd_vaddq_u16(a,b)
#define vaddq_s16(a,b) __simd_vaddq_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_u32(a,b) __simd_vaddq_u32(a,b)
#define vaddq_s32(a,b) __simd_vaddq_s32(a,b)
#define vadd_u64(a,b) __simd_vadd_u64(a,b)
#define vadd_s64(a,b) __simd_vadd_s64(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 vsubq_u8(a,b) __simd_vsubq_u8(a,b)
#define vsubq_s8(a,b) __simd_vsubq_s8(a,b)
#define vsub_u16(a,b) __simd_vsub_u16(a,b)
#define vsub_s16(a,b) __simd_vsub_s16(a,b)
#define vsubq_u16(a,b) __simd_vsubq_u16(a,b)
#define vsubq_s16(a,b) __simd_vsubq_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_u32(a,b) __simd_vsubq_u32(a,b)
#define vsubq_s32(a,b) __simd_vsubq_s32(a,b)
#define vsub_u64(a,b) __simd_vsub_u64(a,b)
#define vsub_s64(a,b) __simd_vsub_s64(a,b)
#define vsubq_u64(a,b) __simd_vsubq_u64(a,b)
#define vsubq_s64(a,b) __simd_vsubq_s64(a,b)
#define vtst_u8(a,b) __simd_vtst_u8(a,b)
#define vtst_s8(a,b) __simd_vtst_s8(a,b)
#define vtstq_u8(a,b) __simd_vtstq_u8(a,b)
#define vtstq_s8(a,b) __simd_vtstq_s8(a,b)
#define vtst_u16(a,b) __simd_vtst_u16(a,b)
#define vtst_s16(a,b) __simd_vtst_s16(a,b)
#define vtstq_u16(a,b) __simd_vtstq_u16(a,b)
#define vtstq_s16(a,b) __simd_vtstq_s16(a,b)
#define vtst_u32(a,b) __simd_vtst_u32(a,b)
#define vtst_s32(a,b) __simd_vtst_s32(a,b)
#define vtstq_u32(a,b) __simd_vtstq_u32(a,b)
#define vtstq_s32(a,b) __simd_vtstq_s32(a,b)
#define vtst_p8(a,b) __simd_vtst_p8(a,b)
#define vtstq_p8(a,b) __simd_vtstq_p8(a,b)
#define vceq_u8(a,b) __simd_vceq_u8(a,b)
#define vceq_s8(a,b) __simd_vceq_s8(a,b)
#define vceqq_u8(a,b) __simd_vceqq_u8(a,b)
#define vceqq_s8(a,b) __simd_vceqq_s8(a,b)
#define vceq_u16(a,b) __simd_vceq_u16(a,b)
#define vceq_s16(a,b) __simd_vceq_s16(a,b)
#define vceqq_u16(a,b) __simd_vceqq_u16(a,b)
#define vceqq_s16(a,b) __simd_vceqq_s16(a,b)
#define vceq_u32(a,b) __simd_vceq_u32(a,b)
#define vceq_s32(a,b) __simd_vceq_s32(a,b)
#define vceqq_u32(a,b) __simd_vceqq_u32(a,b)
#define vceqq_s32(a,b) __simd_vceqq_s32(a,b)
#define vceq_p8(a,b) __simd_vceq_p8(a,b)
#define vceqq_p8(a,b) __simd_vceqq_p8(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 vmlaq_u8(a,b,c) __simd_vmlaq_u8(a,b,c)
#define vmlaq_s8(a,b,c) __simd_vmlaq_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 vmlaq_u16(a,b,c) __simd_vmlaq_u16(a,b,c)
#define vmlaq_s16(a,b,c) __simd_vmlaq_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_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 vmlsq_u8(a,b,c) __simd_vmlsq_u8(a,b,c)
#define vmlsq_s8(a,b,c) __simd_vmlsq_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 vmlsq_u16(a,b,c) __simd_vmlsq_u16(a,b,c)
#define vmlsq_s16(a,b,c) __simd_vmlsq_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_u32(a,b,c) __simd_vmlsq_u32(a,b,c)
#define vmlsq_s32(a,b,c) __simd_vmlsq_s32(a,b,c)
#define vmul_u8(a,b) __simd_vmul_u8(a,b)
#define vmul_s8(a,b) __simd_vmul_s8(a,b)
#define vmulq_u8(a,b) __simd_vmulq_u8(a,b)
#define vmulq_s8(a,b) __simd_vmulq_s8(a,b)
#define vmul_u16(a,b) __simd_vmul_u16(a,b)
#define vmul_s16(a,b) __simd_vmul_s16(a,b)
#define vmulq_u16(a,b) __simd_vmulq_u16(a,b)
#define vmulq_s16(a,b) __simd_vmulq_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_u32(a,b) __simd_vmulq_u32(a,b)
#define vmulq_s32(a,b) __simd_vmulq_s32(a,b)
#define vmul_p8(a,b) __simd_vmul_p8(a,b)
#define vmulq_p8(a,b) __simd_vmulq_p8(a,b)
#define vqdmulh_s16(a,b) __simd_vqdmulh_s16(a,b)
#define vqdmulhq_s16(a,b) __simd_vqdmulhq_s16(a,b)
#define vqdmulh_s32(a,b) __simd_vqdmulh_s32(a,b)
#define vqdmulhq_s32(a,b) __simd_vqdmulhq_s32(a,b)
#define vqrdmulh_s16(a,b) __simd_vqrdmulh_s16(a,b)
#define vqrdmulhq_s16(a,b) __simd_vqrdmulhq_s16(a,b)
#define vqrdmulh_s32(a,b) __simd_vqrdmulh_s32(a,b)
#define vqrdmulhq_s32(a,b) __simd_vqrdmulhq_s32(a,b)
#define vpmax_s8(a,b) __simd_vpmax_s8(a,b)
#define vpmax_u8(a,b) __simd_vpmax_u8(a,b)
#define vpmax_s16(a,b) __simd_vpmax_s16(a,b)
#define vpmax_u16(a,b) __simd_vpmax_u16(a,b)
#define vpmax_s32(a,b) __simd_vpmax_s32(a,b)
#define vpmax_u32(a,b) __simd_vpmax_u32(a,b)
#define vpmin_s8(a,b) __simd_vpmin_s8(a,b)
#define vpmin_u8(a,b) __simd_vpmin_u8(a,b)
#define vpmin_s16(a,b) __simd_vpmin_s16(a,b)
#define vpmin_u16(a,b) __simd_vpmin_u16(a,b)
#define vpmin_s32(a,b) __simd_vpmin_s32(a,b)
#define vpmin_u32(a,b) __simd_vpmin_u32(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 vshl_s8(a,b) __simd_vshl_s8(a,b)
#define vshlq_s8(a,b) __simd_vshlq_s8(a,b)
#define vshl_u8(a,b) __simd_vshl_u8(a,b)
#define vshlq_u8(a,b) __simd_vshlq_u8(a,b)
#define vshl_s16(a,b) __simd_vshl_s16(a,b)
#define vshlq_s16(a,b) __simd_vshlq_s16(a,b)
#define vshl_u16(a,b) __simd_vshl_u16(a,b)
#define vshlq_u16(a,b) __simd_vshlq_u16(a,b)
#define vshl_s32(a,b) __simd_vshl_s32(a,b)
#define vshlq_s32(a,b) __simd_vshlq_s32(a,b)
#define vshl_u32(a,b) __simd_vshl_u32(a,b)
#define vshlq_u32(a,b) __simd_vshlq_u32(a,b)
#define vshl_s64(a,b) __simd_vshl_s64(a,b)
#define vshlq_s64(a,b) __simd_vshlq_s64(a,b)
#define vshl_u64(a,b) __simd_vshl_u64(a,b)
#define vshlq_u64(a,b) __simd_vshlq_u64(a,b)
#define vqshl_s8(a,b) __simd_vqshl_s8(a,b)
#define vqshlq_s8(a,b) __simd_vqshlq_s8(a,b)
#define vqshl_u8(a,b) __simd_vqshl_u8(a,b)
#define vqshlq_u8(a,b) __simd_vqshlq_u8(a,b)
#define vqshl_s16(a,b) __simd_vqshl_s16(a,b)
#define vqshlq_s16(a,b) __simd_vqshlq_s16(a,b)
#define vqshl_u16(a,b) __simd_vqshl_u16(a,b)
#define vqshlq_u16(a,b) __simd_vqshlq_u16(a,b)
#define vqshl_s32(a,b) __simd_vqshl_s32(a,b)
#define vqshlq_s32(a,b) __simd_vqshlq_s32(a,b)
#define vqshl_u32(a,b) __simd_vqshl_u32(a,b)
#define vqshlq_u32(a,b) __simd_vqshlq_u32(a,b)
#define vqshl_s64(a,b) __simd_vqshl_s64(a,b)
#define vqshlq_s64(a,b) __simd_vqshlq_s64(a,b)
#define vqshl_u64(a,b) __simd_vqshl_u64(a,b)
#define vqshlq_u64(a,b) __simd_vqshlq_u64(a,b)
#define vrshl_s8(a,b) __simd_vrshl_s8(a,b)
#define vrshlq_s8(a,b) __simd_vrshlq_s8(a,b)
#define vrshl_u8(a,b) __simd_vrshl_u8(a,b)
#define vrshlq_u8(a,b) __simd_vrshlq_u8(a,b)
#define vrshl_s16(a,b) __simd_vrshl_s16(a,b)
#define vrshlq_s16(a,b) __simd_vrshlq_s16(a,b)
#define vrshl_u16(a,b) __simd_vrshl_u16(a,b)
#define vrshlq_u16(a,b) __simd_vrshlq_u16(a,b)
#define vrshl_s32(a,b) __simd_vrshl_s32(a,b)
#define vrshlq_s32(a,b) __simd_vrshlq_s32(a,b)
#define vrshl_u32(a,b) __simd_vrshl_u32(a,b)
#define vrshlq_u32(a,b) __simd_vrshlq_u32(a,b)
#define vrshl_s64(a,b) __simd_vrshl_s64(a,b)
#define vrshlq_s64(a,b) __simd_vrshlq_s64(a,b)
#define vrshl_u64(a,b) __simd_vrshl_u64(a,b)
#define vrshlq_u64(a,b) __simd_vrshlq_u64(a,b)
#define vqrshl_s8(a,b) __simd_vqrshl_s8(a,b)
#define vqrshlq_s8(a,b) __simd_vqrshlq_s8(a,b)
#define vqrshl_u8(a,b) __simd_vqrshl_u8(a,b)
#define vqrshlq_u8(a,b) __simd_vqrshlq_u8(a,b)
#define vqrshl_s16(a,b) __simd_vqrshl_s16(a,b)
#define vqrshlq_s16(a,b) __simd_vqrshlq_s16(a,b)
#define vqrshl_u16(a,b) __simd_vqrshl_u16(a,b)
#define vqrshlq_u16(a,b) __simd_vqrshlq_u16(a,b)
#define vqrshl_s32(a,b) __simd_vqrshl_s32(a,b)
#define vqrshlq_s32(a,b) __simd_vqrshlq_s32(a,b)
#define vqrshl_u32(a,b) __simd_vqrshl_u32(a,b)
#define vqrshlq_u32(a,b) __simd_vqrshlq_u32(a,b)
#define vqrshl_s64(a,b) __simd_vqrshl_s64(a,b)
#define vqrshlq_s64(a,b) __simd_vqrshlq_s64(a,b)
#define vqrshl_u64(a,b) __simd_vqrshl_u64(a,b)
#define vqrshlq_u64(a,b) __simd_vqrshlq_u64(a,b)
#define vand_u8(a,b) __simd_vand_u8(a,b)
#define vand_s8(a,b) __simd_vand_s8(a,b)
#define vand_p8(a,b) __simd_vand_p8(a,b)
#define vand_u16(a,b) __simd_vand_u16(a,b)
#define vand_s16(a,b) __simd_vand_s16(a,b)
#define vand_p16(a,b) __simd_vand_p16(a,b)
#define vand_u32(a,b) __simd_vand_u32(a,b)
#define vand_s32(a,b) __simd_vand_s32(a,b)
#define vand_f32(a,b) __simd_vand_f32(a,b)
#define vand_u64(a,b) __simd_vand_u64(a,b)
#define vand_s64(a,b) __simd_vand_s64(a,b)
#define vandq_u8(a,b) __simd_vandq_u8(a,b)
#define vandq_s8(a,b) __simd_vandq_s8(a,b)
#define vandq_p8(a,b) __simd_vandq_p8(a,b)
#define vandq_u16(a,b) __simd_vandq_u16(a,b)
#define vandq_s16(a,b) __simd_vandq_s16(a,b)
#define vandq_p16(a,b) __simd_vandq_p16(a,b)
#define vandq_u32(a,b) __simd_vandq_u32(a,b)
#define vandq_s32(a,b) __simd_vandq_s32(a,b)
#define vandq_f32(a,b) __simd_vandq_f32(a,b)
#define vandq_u64(a,b) __simd_vandq_u64(a,b)
#define vandq_s64(a,b) __simd_vandq_s64(a,b)
#define vbic_u8(a,b) __simd_vbic_u8(a,b)
#define vbic_s8(a,b) __simd_vbic_s8(a,b)
#define vbic_p8(a,b) __simd_vbic_p8(a,b)
#define vbic_u16(a,b) __simd_vbic_u16(a,b)
#define vbic_s16(a,b) __simd_vbic_s16(a,b)
#define vbic_p16(a,b) __simd_vbic_p16(a,b)
#define vbic_u32(a,b) __simd_vbic_u32(a,b)
#define vbic_s32(a,b) __simd_vbic_s32(a,b)
#define vbic_f32(a,b) __simd_vbic_f32(a,b)
#define vbic_u64(a,b) __simd_vbic_u64(a,b)
#define vbic_s64(a,b) __simd_vbic_s64(a,b)
#define vbicq_u8(a,b) __simd_vbicq_u8(a,b)
#define vbicq_s8(a,b) __simd_vbicq_s8(a,b)
#define vbicq_p8(a,b) __simd_vbicq_p8(a,b)
#define vbicq_u16(a,b) __simd_vbicq_u16(a,b)
#define vbicq_s16(a,b) __simd_vbicq_s16(a,b)
#define vbicq_p16(a,b) __simd_vbicq_p16(a,b)
#define vbicq_u32(a,b) __simd_vbicq_u32(a,b)
#define vbicq_s32(a,b) __simd_vbicq_s32(a,b)
#define vbicq_f32(a,b) __simd_vbicq_f32(a,b)
#define vbicq_u64(a,b) __simd_vbicq_u64(a,b)
#define vbicq_s64(a,b) __simd_vbicq_s64(a,b)
#define vorr_u8(a,b) __simd_vorr_u8(a,b)
#define vorr_s8(a,b) __simd_vorr_s8(a,b)
#define vorr_p8(a,b) __simd_vorr_p8(a,b)
#define vorr_u16(a,b) __simd_vorr_u16(a,b)
#define vorr_s16(a,b) __simd_vorr_s16(a,b)
#define vorr_p16(a,b) __simd_vorr_p16(a,b)
#define vorr_u32(a,b) __simd_vorr_u32(a,b)
#define vorr_s32(a,b) __simd_vorr_s32(a,b)
#define vorr_f32(a,b) __simd_vorr_f32(a,b)
#define vorr_u64(a,b) __simd_vorr_u64(a,b)
#define vorr_s64(a,b) __simd_vorr_s64(a,b)
#define vorrq_u8(a,b) __simd_vorrq_u8(a,b)
#define vorrq_s8(a,b) __simd_vorrq_s8(a,b)
#define vorrq_p8(a,b) __simd_vorrq_p8(a,b)
#define vorrq_u16(a,b) __simd_vorrq_u16(a,b)
#define vorrq_s16(a,b) __simd_vorrq_s16(a,b)
#define vorrq_p16(a,b) __simd_vorrq_p16(a,b)
#define vorrq_u32(a,b) __simd_vorrq_u32(a,b)
#define vorrq_s32(a,b) __simd_vorrq_s32(a,b)
#define vorrq_f32(a,b) __simd_vorrq_f32(a,b)
#define vorrq_u64(a,b) __simd_vorrq_u64(a,b)
#define vorrq_s64(a,b) __simd_vorrq_s64(a,b)
#define vorn_u8(a,b) __simd_vorn_u8(a,b)
#define vorn_s8(a,b) __simd_vorn_s8(a,b)
#define vorn_p8(a,b) __simd_vorn_p8(a,b)
#define vorn_u16(a,b) __simd_vorn_u16(a,b)
#define vorn_s16(a,b) __simd_vorn_s16(a,b)
#define vorn_p16(a,b) __simd_vorn_p16(a,b)
#define vorn_u32(a,b) __simd_vorn_u32(a,b)
#define vorn_s32(a,b) __simd_vorn_s32(a,b)
#define vorn_f32(a,b) __simd_vorn_f32(a,b)
#define vorn_u64(a,b) __simd_vorn_u64(a,b)
#define vorn_s64(a,b) __simd_vorn_s64(a,b)
#define vornq_u8(a,b) __simd_vornq_u8(a,b)
#define vornq_s8(a,b) __simd_vornq_s8(a,b)
#define vornq_p8(a,b) __simd_vornq_p8(a,b)
#define vornq_u16(a,b) __simd_vornq_u16(a,b)
#define vornq_s16(a,b) __simd_vornq_s16(a,b)
#define vornq_p16(a,b) __simd_vornq_p16(a,b)
#define vornq_u32(a,b) __simd_vornq_u32(a,b)
#define vornq_s32(a,b) __simd_vornq_s32(a,b)
#define vornq_f32(a,b) __simd_vornq_f32(a,b)
#define vornq_u64(a,b) __simd_vornq_u64(a,b)
#define vornq_s64(a,b) __simd_vornq_s64(a,b)
#define veor_u8(a,b) __simd_veor_u8(a,b)
#define veor_s8(a,b) __simd_veor_s8(a,b)
#define veor_p8(a,b) __simd_veor_p8(a,b)
#define veor_u16(a,b) __simd_veor_u16(a,b)
#define veor_s16(a,b) __simd_veor_s16(a,b)
#define veor_p16(a,b) __simd_veor_p16(a,b)
#define veor_u32(a,b) __simd_veor_u32(a,b)
#define veor_s32(a,b) __simd_veor_s32(a,b)
#define veor_f32(a,b) __simd_veor_f32(a,b)
#define veor_u64(a,b) __simd_veor_u64(a,b)
#define veor_s64(a,b) __simd_veor_s64(a,b)
#define veorq_u8(a,b) __simd_veorq_u8(a,b)
#define veorq_s8(a,b) __simd_veorq_s8(a,b)
#define veorq_p8(a,b) __simd_veorq_p8(a,b)
#define veorq_u16(a,b) __simd_veorq_u16(a,b)
#define veorq_s16(a,b) __simd_veorq_s16(a,b)
#define veorq_p16(a,b) __simd_veorq_p16(a,b)
#define veorq_u32(a,b) __simd_veorq_u32(a,b)
#define veorq_s32(a,b) __simd_veorq_s32(a,b)
#define veorq_f32(a,b) __simd_veorq_f32(a,b)
#define veorq_u64(a,b) __simd_veorq_u64(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_s8(a,b,c) __simd_vbsl_s8(a,b,c)
#define vbsl_p8(a,b,c) __simd_vbsl_p8(a,b,c)
#define vbsl_u16(a,b,c) __simd_vbsl_u16(a,b,c)
#define vbsl_s16(a,b,c) __simd_vbsl_s16(a,b,c)
#define vbsl_p16(a,b,c) __simd_vbsl_p16(a,b,c)
#define vbsl_u32(a,b,c) __simd_vbsl_u32(a,b,c)
#define vbsl_s32(a,b,c) __simd_vbsl_s32(a,b,c)
#define vbsl_f32(a,b,c) __simd_vbsl_f32(a,b,c)
#define vbsl_u64(a,b,c) __simd_vbsl_u64(a,b,c)
#define vbsl_s64(a,b,c) __simd_vbsl_s64(a,b,c)
#define vbslq_u8(a,b,c) __simd_vbslq_u8(a,b,c)
#define vbslq_s8(a,b,c) __simd_vbslq_s8(a,b,c)
#define vbslq_p8(a,b,c) __simd_vbslq_p8(a,b,c)
#define vbslq_u16(a,b,c) __simd_vbslq_u16(a,b,c)
#define vbslq_s16(a,b,c) __simd_vbslq_s16(a,b,c)
#define vbslq_p16(a,b,c) __simd_vbslq_p16(a,b,c)
#define vbslq_u32(a,b,c) __simd_vbslq_u32(a,b,c)
#define vbslq_s32(a,b,c) __simd_vbslq_s32(a,b,c)
#define vbslq_f32(a,b,c) __simd_vbslq_f32(a,b,c)
#define vbslq_u64(a,b,c) __simd_vbslq_u64(a,b,c)
#define vbslq_s64(a,b,c) __simd_vbslq_s64(a,b,c)
#define vbit_u8(a,b,c) __simd_vbit_u8(a,b,c)
#define vbit_s8(a,b,c) __simd_vbit_s8(a,b,c)
#define vbit_p8(a,b,c) __simd_vbit_p8(a,b,c)
#define vbit_u16(a,b,c) __simd_vbit_u16(a,b,c)
#define vbit_s16(a,b,c) __simd_vbit_s16(a,b,c)
#define vbit_p16(a,b,c) __simd_vbit_p16(a,b,c)
#define vbit_u32(a,b,c) __simd_vbit_u32(a,b,c)
#define vbit_s32(a,b,c) __simd_vbit_s32(a,b,c)
#define vbit_f32(a,b,c) __simd_vbit_f32(a,b,c)
#define vbit_u64(a,b,c) __simd_vbit_u64(a,b,c)
#define vbit_s64(a,b,c) __simd_vbit_s64(a,b,c)
#define vbitq_u8(a,b,c) __simd_vbitq_u8(a,b,c)
#define vbitq_s8(a,b,c) __simd_vbitq_s8(a,b,c)
#define vbitq_p8(a,b,c) __simd_vbitq_p8(a,b,c)
#define vbitq_u16(a,b,c) __simd_vbitq_u16(a,b,c)
#define vbitq_s16(a,b,c) __simd_vbitq_s16(a,b,c)
#define vbitq_p16(a,b,c) __simd_vbitq_p16(a,b,c)
#define vbitq_u32(a,b,c) __simd_vbitq_u32(a,b,c)
#define vbitq_s32(a,b,c) __simd_vbitq_s32(a,b,c)
#define vbitq_f32(a,b,c) __simd_vbitq_f32(a,b,c)
#define vbitq_u64(a,b,c) __simd_vbitq_u64(a,b,c)
#define vbitq_s64(a,b,c) __simd_vbitq_s64(a,b,c)
#define vbif_u8(a,b,c) __simd_vbif_u8(a,b,c)
#define vbif_s8(a,b,c) __simd_vbif_s8(a,b,c)
#define vbif_p8(a,b,c) __simd_vbif_p8(a,b,c)
#define vbif_u16(a,b,c) __simd_vbif_u16(a,b,c)
#define vbif_s16(a,b,c) __simd_vbif_s16(a,b,c)
#define vbif_p16(a,b,c) __simd_vbif_p16(a,b,c)
#define vbif_u32(a,b,c) __simd_vbif_u32(a,b,c)
#define vbif_s32(a,b,c) __simd_vbif_s32(a,b,c)
#define vbif_f32(a,b,c) __simd_vbif_f32(a,b,c)
#define vbif_u64(a,b,c) __simd_vbif_u64(a,b,c)
#define vbif_s64(a,b,c) __simd_vbif_s64(a,b,c)
#define vbifq_u8(a,b,c) __simd_vbifq_u8(a,b,c)
#define vbifq_s8(a,b,c) __simd_vbifq_s8(a,b,c)
#define vbifq_p8(a,b,c) __simd_vbifq_p8(a,b,c)
#define vbifq_u16(a,b,c) __simd_vbifq_u16(a,b,c)
#define vbifq_s16(a,b,c) __simd_vbifq_s16(a,b,c)
#define vbifq_p16(a,b,c) __simd_vbifq_p16(a,b,c)
#define vbifq_u32(a,b,c) __simd_vbifq_u32(a,b,c)
#define vbifq_s32(a,b,c) __simd_vbifq_s32(a,b,c)
#define vbifq_f32(a,b,c) __simd_vbifq_f32(a,b,c)
#define vbifq_u64(a,b,c) __simd_vbifq_u64(a,b,c)
#define vbifq_s64(a,b,c) __simd_vbifq_s64(a,b,c)
#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 vfms_f32(a,b,c) __simd_vfms_f32(a,b,c)
#define vfmsq_f32(a,b,c) __simd_vfmsq_f32(a,b,c)
#define vadd_f32(a,b) __simd_vadd_f32(a,b)
#define vaddq_f32(a,b) __simd_vaddq_f32(a,b)
#define vsub_f32(a,b) __simd_vsub_f32(a,b)
#define vsubq_f32(a,b) __simd_vsubq_f32(a,b)
#define vabd_f32(a,b) __simd_vabd_f32(a,b)
#define vabdq_f32(a,b) __simd_vabdq_f32(a,b)
#define vmla_f32(a,b,c) __simd_vmla_f32(a,b,c)
#define vmlaq_f32(a,b,c) __simd_vmlaq_f32(a,b,c)
#define vmls_f32(a,b,c) __simd_vmls_f32(a,b,c)
#define vmlsq_f32(a,b,c) __simd_vmlsq_f32(a,b,c)
#define vmul_f32(a,b) __simd_vmul_f32(a,b)
#define vmulq_f32(a,b) __simd_vmulq_f32(a,b)
#define vceq_f32(a,b) __simd_vceq_f32(a,b)
#define vceqq_f32(a,b) __simd_vceqq_f32(a,b)
#define vcge_f32(a,b) __simd_vcge_f32(a,b)
#define vcgeq_f32(a,b) __simd_vcgeq_f32(a,b)
#define vcgt_f32(a,b) __simd_vcgt_f32(a,b)
#define vcgtq_f32(a,b) __simd_vcgtq_f32(a,b)
#define vclt_f32(a,b) __simd_vcgt_f32(b,a)
#define vcltq_f32(a,b) __simd_vcgtq_f32(b,a)
#define vrinta_f32(a) __simd_vrinta_f32(a)
#define vrintaq_f32(a) __simd_vrintaq_f32(a)
#define vrintm_f32(a) __simd_vrintm_f32(a)
#define vrintmq_f32(a) __simd_vrintmq_f32(a)
#define vrintn_f32(a) __simd_vrintn_f32(a)
#define vrintnq_f32(a) __simd_vrintnq_f32(a)
#define vrintp_f32(a) __simd_vrintp_f32(a)
#define vrintpq_f32(a) __simd_vrintpq_f32(a)
#define vrintx_f32(a) __simd_vrintx_f32(a)
#define vrintxq_f32(a) __simd_vrintxq_f32(a)
#define vrintz_f32(a) __simd_vrintz_f32(a)
#define vrintzq_f32(a) __simd_vrintzq_f32(a)
#define vcvta_s32_f32(a) __simd_vcvta_s32_f32(a)
#define vcvtaq_s32_f32(a) __simd_vcvtaq_s32_f32(a)
#define vcvta_u32_f32(a) __simd_vcvta_u32_f32(a)
#define vcvtaq_u32_f32(a) __simd_vcvtaq_u32_f32(a)
#define vcvtm_s32_f32(a) __simd_vcvtm_s32_f32(a)
#define vcvtmq_s32_f32(a) __simd_vcvtmq_s32_f32(a)
#define vcvtm_u32_f32(a) __simd_vcvtm_u32_f32(a)
#define vcvtmq_u32_f32(a) __simd_vcvtmq_u32_f32(a)
#define vcvtn_s32_f32(a) __simd_vcvtn_s32_f32(a)
#define vcvtnq_s32_f32(a) __simd_vcvtnq_s32_f32(a)
#define vcvtn_u32_f32(a) __simd_vcvtn_u32_f32(a)
#define vcvtnq_u32_f32(a) __simd_vcvtnq_u32_f32(a)
#define vcvtp_s32_f32(a) __simd_vcvtp_s32_f32(a)
#define vcvtpq_s32_f32(a) __simd_vcvtpq_s32_f32(a)
#define vcvtp_u32_f32(a) __simd_vcvtp_u32_f32(a)
#define vcvtpq_u32_f32(a) __simd_vcvtpq_u32_f32(a)
#define vmaxnm_f32(a,b) __simd_vmaxnm_f32(a,b)
#define vmaxnmq_f32(a,b) __simd_vmaxnmq_f32(a,b)
#define vminnm_f32(a,b) __simd_vminnm_f32(a,b)
#define vminnmq_f32(a,b) __simd_vminnmq_f32(a,b)
#define vcle_f32(a,b) __simd_vcge_f32(b,a)
#define vcleq_f32(a,b) __simd_vcgeq_f32(b,a)
#define vacge_f32(a,b) __simd_vacge_f32(a,b)
#define vacgeq_f32(a,b) __simd_vacgeq_f32(a,b)
#define vacgt_f32(a,b) __simd_vacgt_f32(a,b)
#define vacgtq_f32(a,b) __simd_vacgtq_f32(a,b)
#define vaclt_f32(a,b) __simd_vacgt_f32(b,a)
#define vacltq_f32(a,b) __simd_vacgtq_f32(b,a)
#define vacle_f32(a,b) __simd_vacge_f32(b,a)
#define vacleq_f32(a,b) __simd_vacgeq_f32(b,a)
#define vcage_f32(a,b) __simd_vacge_f32(a,b)
#define vcageq_f32(a,b) __simd_vacgeq_f32(a,b)
#define vcagt_f32(a,b) __simd_vacgt_f32(a,b)
#define vcagtq_f32(a,b) __simd_vacgtq_f32(a,b)
#define vcalt_f32(a,b) __simd_vacgt_f32(b,a)
#define vcaltq_f32(a,b) __simd_vacgtq_f32(b,a)
#define vcale_f32(a,b) __simd_vacge_f32(b,a)
#define vcaleq_f32(a,b) __simd_vacgeq_f32(b,a)
#define vmax_f32(a,b) __simd_vmax_f32(a,b)
#define vmaxq_f32(a,b) __simd_vmaxq_f32(a,b)
#define vmin_f32(a,b) __simd_vmin_f32(a,b)
#define vminq_f32(a,b) __simd_vminq_f32(a,b)
#define vrecps_f32(a,b) __simd_vrecps_f32(a,b)
#define vrecpsq_f32(a,b) __simd_vrecpsq_f32(a,b)
#define vrsqrts_f32(a,b) __simd_vrsqrts_f32(a,b)
#define vrsqrtsq_f32(a,b) __simd_vrsqrtsq_f32(a,b)
#define vpadd_f32(a,b) __simd_vpadd_f32(a,b)
#define vpmax_f32(a,b) __simd_vpmax_f32(a,b)
#define vpmin_f32(a,b) __simd_vpmin_f32(a,b)
#define vaddl_s8(a,b) __simd_vaddl_s8(a,b)
#define vaddl_u8(a,b) __simd_vaddl_u8(a,b)
#define vaddl_s16(a,b) __simd_vaddl_s16(a,b)
#define vaddl_u16(a,b) __simd_vaddl_u16(a,b)
#define vaddl_s32(a,b) __simd_vaddl_s32(a,b)
#define vaddl_u32(a,b) __simd_vaddl_u32(a,b)
#define vsubl_s8(a,b) __simd_vsubl_s8(a,b)
#define vsubl_u8(a,b) __simd_vsubl_u8(a,b)
#define vsubl_s16(a,b) __simd_vsubl_s16(a,b)
#define vsubl_u16(a,b) __simd_vsubl_u16(a,b)
#define vsubl_s32(a,b) __simd_vsubl_s32(a,b)
#define vsubl_u32(a,b) __simd_vsubl_u32(a,b)
#define vabal_s8(a,b,c) __simd_vabal_s8(a,b,c)
#define vabal_u8(a,b,c) __simd_vabal_u8(a,b,c)
#define vabal_s16(a,b,c) __simd_vabal_s16(a,b,c)
#define vabal_u16(a,b,c) __simd_vabal_u16(a,b,c)
#define vabal_s32(a,b,c) __simd_vabal_s32(a,b,c)
#define vabal_u32(a,b,c) __simd_vabal_u32(a,b,c)
#define vabdl_s8(a,b) __simd_vabdl_s8(a,b)
#define vabdl_u8(a,b) __simd_vabdl_u8(a,b)
#define vabdl_s16(a,b) __simd_vabdl_s16(a,b)
#define vabdl_u16(a,b) __simd_vabdl_u16(a,b)
#define vabdl_s32(a,b) __simd_vabdl_s32(a,b)
#define vabdl_u32(a,b) __simd_vabdl_u32(a,b)
#define vmlal_s8(a,b,c) __simd_vmlal_s8(a,b,c)
#define vmlal_u8(a,b,c) __simd_vmlal_u8(a,b,c)
#define vmlal_s16(a,b,c) __simd_vmlal_s16(a,b,c)
#define vmlal_u16(a,b,c) __simd_vmlal_u16(a,b,c)
#define vmlal_s32(a,b,c) __simd_vmlal_s32(a,b,c)
#define vmlal_u32(a,b,c) __simd_vmlal_u32(a,b,c)
#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 vmlsl_s8(a,b,c) __simd_vmlsl_s8(a,b,c)
#define vmlsl_u8(a,b,c) __simd_vmlsl_u8(a,b,c)
#define vmlsl_s16(a,b,c) __simd_vmlsl_s16(a,b,c)
#define vmlsl_u16(a,b,c) __simd_vmlsl_u16(a,b,c)
#define vmlsl_s32(a,b,c) __simd_vmlsl_s32(a,b,c)
#define vmlsl_u32(a,b,c) __simd_vmlsl_u32(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 vmull_s8(a,b) __simd_vmull_s8(a,b)
#define vmull_u8(a,b) __simd_vmull_u8(a,b)
#define vmull_s16(a,b) __simd_vmull_s16(a,b)
#define vmull_u16(a,b) __simd_vmull_u16(a,b)
#define vmull_s32(a,b) __simd_vmull_s32(a,b)
#define vmull_u32(a,b) __simd_vmull_u32(a,b)
#define vqdmull_s16(a,b) __simd_vqdmull_s16(a,b)
#define vqdmull_s32(a,b) __simd_vqdmull_s32(a,b)
#define vmull_p8(a,b) __simd_vmull_p8(a,b)
#define vaddw_s8(a,b) __simd_vaddw_s8(a,b)
#define vaddw_u8(a,b) __simd_vaddw_u8(a,b)
#define vaddw_s16(a,b) __simd_vaddw_s16(a,b)
#define vaddw_u16(a,b) __simd_vaddw_u16(a,b)
#define vaddw_s32(a,b) __simd_vaddw_s32(a,b)
#define vaddw_u32(a,b) __simd_vaddw_u32(a,b)
#define vsubw_s8(a,b) __simd_vsubw_s8(a,b)
#define vsubw_u8(a,b) __simd_vsubw_u8(a,b)
#define vsubw_s16(a,b) __simd_vsubw_s16(a,b)
#define vsubw_u16(a,b) __simd_vsubw_u16(a,b)
#define vsubw_s32(a,b) __simd_vsubw_s32(a,b)
#define vsubw_u32(a,b) __simd_vsubw_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 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 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 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 vmla_lane_u16(a,b,c,d) __simd_vmla_lane_u16(a,b,c,d)
#define vmla_n_u16(a,b,c) __simd_vmla_n_u16(a,b,c)
#define vmla_lane_s16(a,b,c,d) __simd_vmla_lane_s16(a,b,c,d)
#define vmla_n_s16(a,b,c) __simd_vmla_n_s16(a,b,c)
#define vmlaq_lane_u16(a,b,c,d) __simd_vmlaq_lane_u16(a,b,c,d)
#define vmlaq_n_u16(a,b,c) __simd_vmlaq_n_u16(a,b,c)
#define vmlaq_lane_s16(a,b,c,d) __simd_vmlaq_lane_s16(a,b,c,d)
#define vmlaq_n_s16(a,b,c) __simd_vmlaq_n_s16(a,b,c)
#define vmla_lane_u32(a,b,c,d) __simd_vmla_lane_u32(a,b,c,d)
#define vmla_n_u32(a,b,c) __simd_vmla_n_u32(a,b,c)
#define vmla_lane_s32(a,b,c,d) __simd_vmla_lane_s32(a,b,c,d)
#define vmla_n_s32(a,b,c) __simd_vmla_n_s32(a,b,c)
#define vmlaq_lane_u32(a,b,c,d) __simd_vmlaq_lane_u32(a,b,c,d)
#define vmlaq_n_u32(a,b,c) __simd_vmlaq_n_u32(a,b,c)
#define vmlaq_lane_s32(a,b,c,d) __simd_vmlaq_lane_s32(a,b,c,d)
#define vmlaq_n_s32(a,b,c) __simd_vmlaq_n_s32(a,b,c)
#define vmla_lane_f32(a,b,c,d) __simd_vmla_lane_f32(a,b,c,d)
#define vmla_n_f32(a,b,c) __simd_vmla_n_f32(a,b,c)
#define vmlaq_lane_f32(a,b,c,d) __simd_vmlaq_lane_f32(a,b,c,d)
#define vmlaq_n_f32(a,b,c) __simd_vmlaq_n_f32(a,b,c)
#define vmls_lane_u16(a,b,c,d) __simd_vmls_lane_u16(a,b,c,d)
#define vmls_n_u16(a,b,c) __simd_vmls_n_u16(a,b,c)
#define vmls_lane_s16(a,b,c,d) __simd_vmls_lane_s16(a,b,c,d)
#define vmls_n_s16(a,b,c) __simd_vmls_n_s16(a,b,c)
#define vmlsq_lane_u16(a,b,c,d) __simd_vmlsq_lane_u16(a,b,c,d)
#define vmlsq_n_u16(a,b,c) __simd_vmlsq_n_u16(a,b,c)
#define vmlsq_lane_s16(a,b,c,d) __simd_vmlsq_lane_s16(a,b,c,d)
#define vmlsq_n_s16(a,b,c) __simd_vmlsq_n_s16(a,b,c)
#define vmls_lane_u32(a,b,c,d) __simd_vmls_lane_u32(a,b,c,d)
#define vmls_n_u32(a,b,c) __simd_vmls_n_u32(a,b,c)
#define vmls_lane_s32(a,b,c,d) __simd_vmls_lane_s32(a,b,c,d)
#define vmls_n_s32(a,b,c) __simd_vmls_n_s32(a,b,c)
#define vmlsq_lane_u32(a,b,c,d) __simd_vmlsq_lane_u32(a,b,c,d)
#define vmlsq_n_u32(a,b,c) __simd_vmlsq_n_u32(a,b,c)
#define vmlsq_lane_s32(a,b,c,d) __simd_vmlsq_lane_s32(a,b,c,d)
#define vmlsq_n_s32(a,b,c) __simd_vmlsq_n_s32(a,b,c)
#define vmls_lane_f32(a,b,c,d) __simd_vmls_lane_f32(a,b,c,d)
#define vmls_n_f32(a,b,c) __simd_vmls_n_f32(a,b,c)
#define vmlsq_lane_f32(a,b,c,d) __simd_vmlsq_lane_f32(a,b,c,d)
#define vmlsq_n_f32(a,b,c) __simd_vmlsq_n_f32(a,b,c)
#define vmul_lane_u16(a,b,c) __simd_vmul_lane_u16(a,b,c)
#define vmul_n_u16(a,b) __simd_vmul_n_u16(a,b)
#define vmul_lane_s16(a,b,c) __simd_vmul_lane_s16(a,b,c)
#define vmul_n_s16(a,b) __simd_vmul_n_s16(a,b)
#define vmulq_lane_u16(a,b,c) __simd_vmulq_lane_u16(a,b,c)
#define vmulq_n_u16(a,b) __simd_vmulq_n_u16(a,b)
#define vmulq_lane_s16(a,b,c) __simd_vmulq_lane_s16(a,b,c)
#define vmulq_n_s16(a,b) __simd_vmulq_n_s16(a,b)
#define vmul_lane_u32(a,b,c) __simd_vmul_lane_u32(a,b,c)
#define vmul_n_u32(a,b) __simd_vmul_n_u32(a,b)
#define vmul_lane_s32(a,b,c) __simd_vmul_lane_s32(a,b,c)
#define vmul_n_s32(a,b) __simd_vmul_n_s32(a,b)
#define vmulq_lane_u32(a,b,c) __simd_vmulq_lane_u32(a,b,c)
#define vmulq_n_u32(a,b) __simd_vmulq_n_u32(a,b)
#define vmulq_lane_s32(a,b,c) __simd_vmulq_lane_s32(a,b,c)
#define vmulq_n_s32(a,b) __simd_vmulq_n_s32(a,b)
#define vmul_lane_f32(a,b,c) __simd_vmul_lane_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_n_f32(a,b) __simd_vmulq_n_f32(a,b)
#define vqdmulh_lane_s16(a,b,c) __simd_vqdmulh_lane_s16(a,b,c)
#define vqdmulh_n_s16(a,b) __simd_vqdmulh_n_s16(a,b)
#define vqdmulhq_lane_s16(a,b,c) __simd_vqdmulhq_lane_s16(a,b,c)
#define vqdmulhq_n_s16(a,b) __simd_vqdmulhq_n_s16(a,b)
#define vqdmulh_lane_s32(a,b,c) __simd_vqdmulh_lane_s32(a,b,c)
#define vqdmulh_n_s32(a,b) __simd_vqdmulh_n_s32(a,b)
#define vqdmulhq_lane_s32(a,b,c) __simd_vqdmulhq_lane_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_n_s16(a,b) __simd_vqrdmulh_n_s16(a,b)
#define vqrdmulhq_lane_s16(a,b,c) __simd_vqrdmulhq_lane_s16(a,b,c)
#define vqrdmulhq_n_s16(a,b) __simd_vqrdmulhq_n_s16(a,b)
#define vqrdmulh_lane_s32(a,b,c) __simd_vqrdmulh_lane_s32(a,b,c)
#define vqrdmulh_n_s32(a,b) __simd_vqrdmulh_n_s32(a,b)
#define vqrdmulhq_lane_s32(a,b,c) __simd_vqrdmulhq_lane_s32(a,b,c)
#define vqrdmulhq_n_s32(a,b) __simd_vqrdmulhq_n_s32(a,b)
#define vmlal_lane_s16(a,b,c,d) __simd_vmlal_lane_s16(a,b,c,d)
#define vmlal_n_s16(a,b,c) __simd_vmlal_n_s16(a,b,c)
#define vmlal_lane_u16(a,b,c,d) __simd_vmlal_lane_u16(a,b,c,d)
#define vmlal_n_u16(a,b,c) __simd_vmlal_n_u16(a,b,c)
#define vmlal_lane_s32(a,b,c,d) __simd_vmlal_lane_s32(a,b,c,d)
#define vmlal_n_s32(a,b,c) __simd_vmlal_n_s32(a,b,c)
#define vmlal_lane_u32(a,b,c,d) __simd_vmlal_lane_u32(a,b,c,d)
#define vmlal_n_u32(a,b,c) __simd_vmlal_n_u32(a,b,c)
#define vqdmlal_lane_s16(a,b,c,d) __simd_vqdmlal_lane_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_n_s32(a,b,c) __simd_vqdmlal_n_s32(a,b,c)
#define vmlsl_lane_s16(a,b,c,d) __simd_vmlsl_lane_s16(a,b,c,d)
#define vmlsl_n_s16(a,b,c) __simd_vmlsl_n_s16(a,b,c)
#define vmlsl_lane_u16(a,b,c,d) __simd_vmlsl_lane_u16(a,b,c,d)
#define vmlsl_n_u16(a,b,c) __simd_vmlsl_n_u16(a,b,c)
#define vmlsl_lane_s32(a,b,c,d) __simd_vmlsl_lane_s32(a,b,c,d)
#define vmlsl_n_s32(a,b,c) __simd_vmlsl_n_s32(a,b,c)
#define vmlsl_lane_u32(a,b,c,d) __simd_vmlsl_lane_u32(a,b,c,d)
#define vmlsl_n_u32(a,b,c) __simd_vmlsl_n_u32(a,b,c)
#define vqdmlsl_lane_s16(a,b,c,d) __simd_vqdmlsl_lane_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_n_s32(a,b,c) __simd_vqdmlsl_n_s32(a,b,c)
#define vmull_lane_s16(a,b,c) __simd_vmull_lane_s16(a,b,c)
#define vmull_n_s16(a,b) __simd_vmull_n_s16(a,b)
#define vmull_lane_u16(a,b,c) __simd_vmull_lane_u16(a,b,c)
#define vmull_n_u16(a,b) __simd_vmull_n_u16(a,b)
#define vmull_lane_s32(a,b,c) __simd_vmull_lane_s32(a,b,c)
#define vmull_n_s32(a,b) __simd_vmull_n_s32(a,b)
#define vmull_lane_u32(a,b,c) __simd_vmull_lane_u32(a,b,c)
#define vmull_n_u32(a,b) __simd_vmull_n_u32(a,b)
#define vqdmull_lane_s16(a,b,c) __simd_vqdmull_lane_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_n_s32(a,b) __simd_vqdmull_n_s32(a,b)
#define vshr_n_s8(a,b) __simd_vshr_n_s8(a,b)
#define vshrq_n_s8(a,b) __simd_vshrq_n_s8(a,b)
#define vshr_n_u8(a,b) __simd_vshr_n_u8(a,b)
#define vshrq_n_u8(a,b) __simd_vshrq_n_u8(a,b)
#define vshr_n_s16(a,b) __simd_vshr_n_s16(a,b)
#define vshrq_n_s16(a,b) __simd_vshrq_n_s16(a,b)
#define vshr_n_u16(a,b) __simd_vshr_n_u16(a,b)
#define vshrq_n_u16(a,b) __simd_vshrq_n_u16(a,b)
#define vshr_n_s32(a,b) __simd_vshr_n_s32(a,b)
#define vshrq_n_s32(a,b) __simd_vshrq_n_s32(a,b)
#define vshr_n_u32(a,b) __simd_vshr_n_u32(a,b)
#define vshrq_n_u32(a,b) __simd_vshrq_n_u32(a,b)
#define vshr_n_s64(a,b) __simd_vshr_n_s64(a,b)
#define vshrq_n_s64(a,b) __simd_vshrq_n_s64(a,b)
#define vshr_n_u64(a,b) __simd_vshr_n_u64(a,b)
#define vshrq_n_u64(a,b) __simd_vshrq_n_u64(a,b)
#define vsra_n_s8(a,b,c) __simd_vsra_n_s8(a,b,c)
#define vsraq_n_s8(a,b,c) __simd_vsraq_n_s8(a,b,c)
#define vsra_n_u8(a,b,c) __simd_vsra_n_u8(a,b,c)
#define vsraq_n_u8(a,b,c) __simd_vsraq_n_u8(a,b,c)
#define vsra_n_s16(a,b,c) __simd_vsra_n_s16(a,b,c)
#define vsraq_n_s16(a,b,c) __simd_vsraq_n_s16(a,b,c)
#define vsra_n_u16(a,b,c) __simd_vsra_n_u16(a,b,c)
#define vsraq_n_u16(a,b,c) __simd_vsraq_n_u16(a,b,c)
#define vsra_n_s32(a,b,c) __simd_vsra_n_s32(a,b,c)
#define vsraq_n_s32(a,b,c) __simd_vsraq_n_s32(a,b,c)
#define vsra_n_u32(a,b,c) __simd_vsra_n_u32(a,b,c)
#define vsraq_n_u32(a,b,c) __simd_vsraq_n_u32(a,b,c)
#define vsra_n_s64(a,b,c) __simd_vsra_n_s64(a,b,c)
#define vsraq_n_s64(a,b,c) __simd_vsraq_n_s64(a,b,c)
#define vsra_n_u64(a,b,c) __simd_vsra_n_u64(a,b,c)
#define vsraq_n_u64(a,b,c) __simd_vsraq_n_u64(a,b,c)
#define vrshr_n_s8(a,b) __simd_vrshr_n_s8(a,b)
#define vrshrq_n_s8(a,b) __simd_vrshrq_n_s8(a,b)
#define vrshr_n_u8(a,b) __simd_vrshr_n_u8(a,b)
#define vrshrq_n_u8(a,b) __simd_vrshrq_n_u8(a,b)
#define vrshr_n_s16(a,b) __simd_vrshr_n_s16(a,b)
#define vrshrq_n_s16(a,b) __simd_vrshrq_n_s16(a,b)
#define vrshr_n_u16(a,b) __simd_vrshr_n_u16(a,b)
#define vrshrq_n_u16(a,b) __simd_vrshrq_n_u16(a,b)
#define vrshr_n_s32(a,b) __simd_vrshr_n_s32(a,b)
#define vrshrq_n_s32(a,b) __simd_vrshrq_n_s32(a,b)
#define vrshr_n_u32(a,b) __simd_vrshr_n_u32(a,b)
#define vrshrq_n_u32(a,b) __simd_vrshrq_n_u32(a,b)
#define vrshr_n_s64(a,b) __simd_vrshr_n_s64(a,b)
#define vrshrq_n_s64(a,b) __simd_vrshrq_n_s64(a,b)
#define vrshr_n_u64(a,b) __simd_vrshr_n_u64(a,b)
#define vrshrq_n_u64(a,b) __simd_vrshrq_n_u64(a,b)
#define vrsra_n_s8(a,b,c) __simd_vrsra_n_s8(a,b,c)
#define vrsraq_n_s8(a,b,c) __simd_vrsraq_n_s8(a,b,c)
#define vrsra_n_u8(a,b,c) __simd_vrsra_n_u8(a,b,c)
#define vrsraq_n_u8(a,b,c) __simd_vrsraq_n_u8(a,b,c)
#define vrsra_n_s16(a,b,c) __simd_vrsra_n_s16(a,b,c)
#define vrsraq_n_s16(a,b,c) __simd_vrsraq_n_s16(a,b,c)
#define vrsra_n_u16(a,b,c) __simd_vrsra_n_u16(a,b,c)
#define vrsraq_n_u16(a,b,c) __simd_vrsraq_n_u16(a,b,c)
#define vrsra_n_s32(a,b,c) __simd_vrsra_n_s32(a,b,c)
#define vrsraq_n_s32(a,b,c) __simd_vrsraq_n_s32(a,b,c)
#define vrsra_n_u32(a,b,c) __simd_vrsra_n_u32(a,b,c)
#define vrsraq_n_u32(a,b,c) __simd_vrsraq_n_u32(a,b,c)
#define vrsra_n_s64(a,b,c) __simd_vrsra_n_s64(a,b,c)
#define vrsraq_n_s64(a,b,c) __simd_vrsraq_n_s64(a,b,c)
#define vrsra_n_u64(a,b,c) __simd_vrsra_n_u64(a,b,c)
#define vrsraq_n_u64(a,b,c) __simd_vrsraq_n_u64(a,b,c)
#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 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 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 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 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 vsri_n_f32(a,b,c) __simd_vsri_n_f32(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_f32(a,b,c) __simd_vsriq_n_f32(a,b,c)
#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 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 vshl_n_u8(a,b) __simd_vshl_n_u8(a,b)
#define vshl_n_s8(a,b) __simd_vshl_n_s8(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 vshl_n_u16(a,b) __simd_vshl_n_u16(a,b)
#define vshl_n_s16(a,b) __simd_vshl_n_s16(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 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_u32(a,b) __simd_vshlq_n_u32(a,b)
#define vshlq_n_s32(a,b) __simd_vshlq_n_s32(a,b)
#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 vshlq_n_u64(a,b) __simd_vshlq_n_u64(a,b)
#define vshlq_n_s64(a,b) __simd_vshlq_n_s64(a,b)
#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 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 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 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 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 vsli_n_f32(a,b,c) __simd_vsli_n_f32(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_f32(a,b,c) __simd_vsliq_n_f32(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 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 vqshlu_n_s8(a,b) __simd_vqshlu_n_s8(a,b)
#define vqshluq_n_s8(a,b) __simd_vqshluq_n_s8(a,b)
#define vqshlu_n_s16(a,b) __simd_vqshlu_n_s16(a,b)
#define vqshluq_n_s16(a,b) __simd_vqshluq_n_s16(a,b)
#define vqshlu_n_s32(a,b) __simd_vqshlu_n_s32(a,b)
#define vqshluq_n_s32(a,b) __simd_vqshluq_n_s32(a,b)
#define vqshlu_n_s64(a,b) __simd_vqshlu_n_s64(a,b)
#define vqshluq_n_s64(a,b) __simd_vqshluq_n_s64(a,b)
#define vqshl_n_s8(a,b) __simd_vqshl_n_s8(a,b)
#define vqshlq_n_s8(a,b) __simd_vqshlq_n_s8(a,b)
#define vqshl_n_u8(a,b) __simd_vqshl_n_u8(a,b)
#define vqshlq_n_u8(a,b) __simd_vqshlq_n_u8(a,b)
#define vqshl_n_s16(a,b) __simd_vqshl_n_s16(a,b)
#define vqshlq_n_s16(a,b) __simd_vqshlq_n_s16(a,b)
#define vqshl_n_u16(a,b) __simd_vqshl_n_u16(a,b)
#define vqshlq_n_u16(a,b) __simd_vqshlq_n_u16(a,b)
#define vqshl_n_s32(a,b) __simd_vqshl_n_s32(a,b)
#define vqshlq_n_s32(a,b) __simd_vqshlq_n_s32(a,b)
#define vqshl_n_u32(a,b) __simd_vqshl_n_u32(a,b)
#define vqshlq_n_u32(a,b) __simd_vqshlq_n_u32(a,b)
#define vqshl_n_s64(a,b) __simd_vqshl_n_s64(a,b)
#define vqshlq_n_s64(a,b) __simd_vqshlq_n_s64(a,b)
#define vqshl_n_u64(a,b) __simd_vqshl_n_u64(a,b)
#define vqshlq_n_u64(a,b) __simd_vqshlq_n_u64(a,b)
#define vmovl_s8(a) __simd_vmovl_s8(a)
#define vmovl_u8(a) __simd_vmovl_u8(a)
#define vmovl_s16(a) __simd_vmovl_s16(a)
#define vmovl_u16(a) __simd_vmovl_u16(a)
#define vmovl_s32(a) __simd_vmovl_s32(a)
#define vmovl_u32(a) __simd_vmovl_u32(a)
#define vshll_n_s8(a,b) __simd_vshll_n_s8(a,b)
#define vshll_n_u8(a,b) __simd_vshll_n_u8(a,b)
#define vshll_n_s16(a,b) __simd_vshll_n_s16(a,b)
#define vshll_n_u16(a,b) __simd_vshll_n_u16(a,b)
#define vshll_n_s32(a,b) __simd_vshll_n_s32(a,b)
#define vshll_n_u32(a,b) __simd_vshll_n_u32(a,b)
#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 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 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 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 vqshrn_n_s16(a,b) __simd_vqshrn_n_s16(a,b)
#define vqshrn_n_u16(a,b) __simd_vqshrn_n_u16(a,b)
#define vqshrn_n_s32(a,b) __simd_vqshrn_n_s32(a,b)
#define vqshrn_n_u32(a,b) __simd_vqshrn_n_u32(a,b)
#define vqshrn_n_s64(a,b) __simd_vqshrn_n_s64(a,b)
#define vqshrn_n_u64(a,b) __simd_vqshrn_n_u64(a,b)
#define vqrshrn_n_s16(a,b) __simd_vqrshrn_n_s16(a,b)
#define vqrshrn_n_u16(a,b) __simd_vqrshrn_n_u16(a,b)
#define vqrshrn_n_s32(a,b) __simd_vqrshrn_n_s32(a,b)
#define vqrshrn_n_u32(a,b) __simd_vqrshrn_n_u32(a,b)
#define vqrshrn_n_s64(a,b) __simd_vqrshrn_n_s64(a,b)
#define vqrshrn_n_u64(a,b) __simd_vqrshrn_n_u64(a,b)
#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 vrev64q_u8(a) __simd_vrev64q_u8(a)
#define vrev64q_s8(a) __simd_vrev64q_s8(a)
#define vrev64q_p8(a) __simd_vrev64q_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 vrev64q_u16(a) __simd_vrev64q_u16(a)
#define vrev64q_s16(a) __simd_vrev64q_s16(a)
#define vrev64q_p16(a) __simd_vrev64q_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_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 vrev32q_u8(a) __simd_vrev32q_u8(a)
#define vrev32q_s8(a) __simd_vrev32q_s8(a)
#define vrev32q_p8(a) __simd_vrev32q_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_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 vpaddlq_s8(a) __simd_vpaddlq_s8(a)
#define vpaddl_s16(a) __simd_vpaddl_s16(a)
#define vpaddlq_s16(a) __simd_vpaddlq_s16(a)
#define vpaddl_s32(a) __simd_vpaddl_s32(a)
#define vpaddlq_s32(a) __simd_vpaddlq_s32(a)
#define vpaddl_u8(a) __simd_vpaddl_u8(a)
#define vpaddlq_u8(a) __simd_vpaddlq_u8(a)
#define vpaddl_u16(a) __simd_vpaddl_u16(a)
#define vpaddlq_u16(a) __simd_vpaddlq_u16(a)
#define vpaddl_u32(a) __simd_vpaddl_u32(a)
#define vpaddlq_u32(a) __simd_vpaddlq_u32(a)
#define vcls_s8(a) __simd_vcls_s8(a)
#define vclsq_s8(a) __simd_vclsq_s8(a)
#define vcls_s16(a) __simd_vcls_s16(a)
#define vclsq_s16(a) __simd_vclsq_s16(a)
#define vcls_s32(a) __simd_vcls_s32(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 vclzq_u8(a) __simd_vclzq_u8(a)
#define vclzq_s8(a) __simd_vclzq_s8(a)
#define vclz_u16(a) __simd_vclz_u16(a)
#define vclz_s16(a) __simd_vclz_s16(a)
#define vclzq_u16(a) __simd_vclzq_u16(a)
#define vclzq_s16(a) __simd_vclzq_s16(a)
#define vclz_u32(a) __simd_vclz_u32(a)
#define vclz_s32(a) __simd_vclz_s32(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 vmvn_u8(a) __simd_vmvn_u8(a)
#define vmvn_s8(a) __simd_vmvn_s8(a)
#define vmvn_p8(a) __simd_vmvn_p8(a)
#define vmvn_u16(a) __simd_vmvn_u16(a)
#define vmvn_s16(a) __simd_vmvn_s16(a)
#define vmvn_p16(a) __simd_vmvn_p16(a)
#define vmvn_u32(a) __simd_vmvn_u32(a)
#define vmvn_s32(a) __simd_vmvn_s32(a)
#define vmvn_f32(a) __simd_vmvn_f32(a)
#define vmvn_u64(a) __simd_vmvn_u64(a)
#define vmvn_s64(a) __simd_vmvn_s64(a)
#define vmvnq_u8(a) __simd_vmvnq_u8(a)
#define vmvnq_s8(a) __simd_vmvnq_s8(a)
#define vmvnq_p8(a) __simd_vmvnq_p8(a)
#define vmvnq_u16(a) __simd_vmvnq_u16(a)
#define vmvnq_s16(a) __simd_vmvnq_s16(a)
#define vmvnq_p16(a) __simd_vmvnq_p16(a)
#define vmvnq_u32(a) __simd_vmvnq_u32(a)
#define vmvnq_s32(a) __simd_vmvnq_s32(a)
#define vmvnq_f32(a) __simd_vmvnq_f32(a)
#define vmvnq_u64(a) __simd_vmvnq_u64(a)
#define vmvnq_s64(a) __simd_vmvnq_s64(a)
#define vpadal_s8(a,b) __simd_vpadal_s8(a,b)
#define vpadalq_s8(a,b) __simd_vpadalq_s8(a,b)
#define vpadal_s16(a,b) __simd_vpadal_s16(a,b)
#define vpadalq_s16(a,b) __simd_vpadalq_s16(a,b)
#define vpadal_s32(a,b) __simd_vpadal_s32(a,b)
#define vpadalq_s32(a,b) __simd_vpadalq_s32(a,b)
#define vpadal_u8(a,b) __simd_vpadal_u8(a,b)
#define vpadalq_u8(a,b) __simd_vpadalq_u8(a,b)
#define vpadal_u16(a,b) __simd_vpadal_u16(a,b)
#define vpadalq_u16(a,b) __simd_vpadalq_u16(a,b)
#define vpadal_u32(a,b) __simd_vpadal_u32(a,b)
#define vpadalq_u32(a,b) __simd_vpadalq_u32(a,b)
#define vqabs_s8(a) __simd_vqabs_s8(a)
#define vqabsq_s8(a) __simd_vqabsq_s8(a)
#define vqabs_s16(a) __simd_vqabs_s16(a)
#define vqabsq_s16(a) __simd_vqabsq_s16(a)
#define vqabs_s32(a) __simd_vqabs_s32(a)
#define vqabsq_s32(a) __simd_vqabsq_s32(a)
#define vqneg_s8(a) __simd_vqneg_s8(a)
#define vqnegq_s8(a) __simd_vqnegq_s8(a)
#define vqneg_s16(a) __simd_vqneg_s16(a)
#define vqnegq_s16(a) __simd_vqnegq_s16(a)
#define vqneg_s32(a) __simd_vqneg_s32(a)
#define vqnegq_s32(a) __simd_vqnegq_s32(a)
#define vabs_s8(a) __simd_vabs_s8(a)
#define vabsq_s8(a) __simd_vabsq_s8(a)
#define vabs_s16(a) __simd_vabs_s16(a)
#define vabsq_s16(a) __simd_vabsq_s16(a)
#define vabs_s32(a) __simd_vabs_s32(a)
#define vabsq_s32(a) __simd_vabsq_s32(a)
#define vabs_f32(a) __simd_vabs_f32(a)
#define vabsq_f32(a) __simd_vabsq_f32(a)
#define vneg_s8(a) __simd_vneg_s8(a)
#define vnegq_s8(a) __simd_vnegq_s8(a)
#define vneg_s16(a) __simd_vneg_s16(a)
#define vnegq_s16(a) __simd_vnegq_s16(a)
#define vneg_s32(a) __simd_vneg_s32(a)
#define vnegq_s32(a) __simd_vnegq_s32(a)
#define vneg_f32(a) __simd_vneg_f32(a)
#define vnegq_f32(a) __simd_vnegq_f32(a)
#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 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 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 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 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_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 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 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 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 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 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 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 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 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 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 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 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 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 vrecpe_u32(a) __simd_vrecpe_u32(a)
#define vrecpeq_u32(a) __simd_vrecpeq_u32(a)
#define vrecpe_f32(a) __simd_vrecpe_f32(a)
#define vrecpeq_f32(a) __simd_vrecpeq_f32(a)
#define vrsqrte_u32(a) __simd_vrsqrte_u32(a)
#define vrsqrteq_u32(a) __simd_vrsqrteq_u32(a)
#define vrsqrte_f32(a) __simd_vrsqrte_f32(a)
#define vrsqrteq_f32(a) __simd_vrsqrteq_f32(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 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_u16(a) __simd_vqmovn_u16(a)
#define vqmovn_u32(a) __simd_vqmovn_u32(a)
#define vqmovn_u64(a) __simd_vqmovn_u64(a)
#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 vcgt_n_s8(a) __simd_vcgt_n_s8(a)
#define vcgtq_n_s8(a) __simd_vcgtq_n_s8(a)
#define vcgt_n_s16(a) __simd_vcgt_n_s16(a)
#define vcgtq_n_s16(a) __simd_vcgtq_n_s16(a)
#define vcgt_n_s32(a) __simd_vcgt_n_s32(a)
#define vcgtq_n_s32(a) __simd_vcgtq_n_s32(a)
#define vcgt_n_f32(a) __simd_vcgt_n_f32(a)
#define vcgtq_n_f32(a) __simd_vcgtq_n_f32(a)
#define vcge_n_s8(a) __simd_vcge_n_s8(a)
#define vcgeq_n_s8(a) __simd_vcgeq_n_s8(a)
#define vcge_n_s16(a) __simd_vcge_n_s16(a)
#define vcgeq_n_s16(a) __simd_vcgeq_n_s16(a)
#define vcge_n_s32(a) __simd_vcge_n_s32(a)
#define vcgeq_n_s32(a) __simd_vcgeq_n_s32(a)
#define vcge_n_f32(a) __simd_vcge_n_f32(a)
#define vcgeq_n_f32(a) __simd_vcgeq_n_f32(a)
#define vceq_n_u8(a) __simd_vceq_n_u8(a)
#define vceq_n_s8(a) __simd_vceq_n_s8(a)
#define vceqq_n_u8(a) __simd_vceqq_n_u8(a)
#define vceqq_n_s8(a) __simd_vceqq_n_s8(a)
#define vceq_n_u16(a) __simd_vceq_n_u16(a)
#define vceq_n_s16(a) __simd_vceq_n_s16(a)
#define vceqq_n_u16(a) __simd_vceqq_n_u16(a)
#define vceqq_n_s16(a) __simd_vceqq_n_s16(a)
#define vceq_n_u32(a) __simd_vceq_n_u32(a)
#define vceq_n_s32(a) __simd_vceq_n_s32(a)
#define vceqq_n_u32(a) __simd_vceqq_n_u32(a)
#define vceqq_n_s32(a) __simd_vceqq_n_s32(a)
#define vceq_n_f32(a) __simd_vceq_n_f32(a)
#define vceqq_n_f32(a) __simd_vceqq_n_f32(a)
#define vcle_n_s8(a) __simd_vcle_n_s8(a)
#define vcleq_n_s8(a) __simd_vcleq_n_s8(a)
#define vcle_n_s16(a) __simd_vcle_n_s16(a)
#define vcleq_n_s16(a) __simd_vcleq_n_s16(a)
#define vcle_n_s32(a) __simd_vcle_n_s32(a)
#define vcleq_n_s32(a) __simd_vcleq_n_s32(a)
#define vcle_n_f32(a) __simd_vcle_n_f32(a)
#define vcleq_n_f32(a) __simd_vcleq_n_f32(a)
#define vclt_n_s8(a) __simd_vclt_n_s8(a)
#define vcltq_n_s8(a) __simd_vcltq_n_s8(a)
#define vclt_n_s16(a) __simd_vclt_n_s16(a)
#define vcltq_n_s16(a) __simd_vcltq_n_s16(a)
#define vclt_n_s32(a) __simd_vclt_n_s32(a)
#define vcltq_n_s32(a) __simd_vcltq_n_s32(a)
#define vclt_n_f32(a) __simd_vclt_n_f32(a)
#define vcltq_n_f32(a) __simd_vcltq_n_f32(a)

#define vdup_n_s8(a) __simd_vdup_n_s8(a)
#define vdup_n_s16(a) __simd_vdup_n_s16(a)
#define vdup_n_s32(a) __simd_vdup_n_s32(a)
#define vdup_n_s64(a) __simd_vdup_n_s64(a)
#define vdup_n_u8(a) __simd_vdup_n_u8(a)
#define vdup_n_u16(a) __simd_vdup_n_u16(a)
#define vdup_n_u32(a) __simd_vdup_n_u32(a)
#define vdup_n_u64(a) __simd_vdup_n_u64(a)
#define vdup_n_p8(a) __simd_vdup_n_p8(a)
#define vdup_n_p16(a) __simd_vdup_n_p16(a)
#define vdup_n_f32(a) __simd_vdup_n_f32(a)
#define vdupq_n_s8(a) __simd_vdupq_n_s8(a)
#define vdupq_n_s16(a) __simd_vdupq_n_s16(a)
#define vdupq_n_s32(a) __simd_vdupq_n_s32(a)
#define vdupq_n_s64(a) __simd_vdupq_n_s64(a)
#define vdupq_n_u8(a) __simd_vdupq_n_u8(a)
#define vdupq_n_u16(a) __simd_vdupq_n_u16(a)
#define vdupq_n_u32(a) __simd_vdupq_n_u32(a)
#define vdupq_n_u64(a) __simd_vdupq_n_u64(a)
#define vdupq_n_p8(a) __simd_vdupq_n_p8(a)
#define vdupq_n_p16(a) __simd_vdupq_n_p16(a)
#define vdupq_n_f32(a) __simd_vdupq_n_f32(a)
#define vmov_n_s8(a) __simd_vdup_n_s8(a)
#define vmov_n_s16(a) __simd_vdup_n_s16(a)
#define vmov_n_s32(a) __simd_vdup_n_s32(a)
#define vmov_n_s64(a) __simd_vdup_n_s64(a)
#define vmov_n_u8(a) __simd_vdup_n_u8(a)
#define vmov_n_u16(a) __simd_vdup_n_u16(a)
#define vmov_n_u32(a) __simd_vdup_n_u32(a)
#define vmov_n_u64(a) __simd_vdup_n_u64(a)
#define vmov_n_p8(a) __simd_vdup_n_p8(a)
#define vmov_n_p16(a) __simd_vdup_n_p16(a)
#define vmov_n_f32(a) __simd_vdup_n_f32(a)
#define vmovq_n_s8(a) __simd_vdupq_n_s8(a)
#define vmovq_n_s16(a) __simd_vdupq_n_s16(a)
#define vmovq_n_s32(a) __simd_vdupq_n_s32(a)
#define vmovq_n_s64(a) __simd_vdupq_n_s64(a)
#define vmovq_n_u8(a) __simd_vdupq_n_u8(a)
#define vmovq_n_u16(a) __simd_vdupq_n_u16(a)
#define vmovq_n_u32(a) __simd_vdupq_n_u32(a)
#define vmovq_n_u64(a) __simd_vdupq_n_u64(a)
#define vmovq_n_p8(a) __simd_vdupq_n_p8(a)
#define vmovq_n_p16(a) __simd_vdupq_n_p16(a)
#define vmovq_n_f32(a) __simd_vdupq_n_f32(a)
#define vdup_lane_s8(a,b) __simd_vdup_lane_s8(a,b)
#define vdup_lane_s16(a,b) __simd_vdup_lane_s16(a,b)
#define vdup_lane_s32(a,b) __simd_vdup_lane_s32(a,b)
#define vdup_lane_s64(a,b) __simd_vdup_lane_s64(a,b)
#define vdup_lane_u8(a,b) __simd_vdup_lane_u8(a,b)
#define vdup_lane_u16(a,b) __simd_vdup_lane_u16(a,b)
#define vdup_lane_u32(a,b) __simd_vdup_lane_u32(a,b)
#define vdup_lane_u64(a,b) __simd_vdup_lane_u64(a,b)
#define vdup_lane_p8(a,b) __simd_vdup_lane_p8(a,b)
#define vdup_lane_p16(a,b) __simd_vdup_lane_p16(a,b)
#define vdup_lane_f32(a,b) __simd_vdup_lane_f32(a,b)
#define vdupq_lane_s8(a,b) __simd_vdupq_lane_s8(a,b)
#define vdupq_lane_s16(a,b) __simd_vdupq_lane_s16(a,b)
#define vdupq_lane_s32(a,b) __simd_vdupq_lane_s32(a,b)
#define vdupq_lane_s64(a,b) __simd_vdupq_lane_s64(a,b)
#define vdupq_lane_u8(a,b) __simd_vdupq_lane_u8(a,b)
#define vdupq_lane_u16(a,b) __simd_vdupq_lane_u16(a,b)
#define vdupq_lane_u32(a,b) __simd_vdupq_lane_u32(a,b)
#define vdupq_lane_u64(a,b) __simd_vdupq_lane_u64(a,b)
#define vdupq_lane_p8(a,b) __simd_vdupq_lane_p8(a,b)
#define vdupq_lane_p16(a,b) __simd_vdupq_lane_p16(a,b)
#define vdupq_lane_f32(a,b) __simd_vdupq_lane_f32(a,b)
#define vcvt_f32_s32(a) __simd_vcvt_f32_s32(a)
#define vcvt_f32_u32(a) __simd_vcvt_f32_u32(a)
#define vcvt_s32_f32(a) __simd_vcvt_s32_f32(a)
#define vcvt_u32_f32(a) __simd_vcvt_u32_f32(a)
#define vcvtq_f32_s32(a) __simd_vcvtq_f32_s32(a)
#define vcvtq_f32_u32(a) __simd_vcvtq_f32_u32(a)
#define vcvtq_s32_f32(a) __simd_vcvtq_s32_f32(a)
#define vcvtq_u32_f32(a) __simd_vcvtq_u32_f32(a)
#define vcvt_n_f32_s32(a,b) __simd_vcvt_n_f32_s32(a,b)
#define vcvt_n_f32_u32(a,b) __simd_vcvt_n_f32_u32(a,b)
#define vcvt_n_s32_f32(a,b) __simd_vcvt_n_s32_f32(a,b)
#define vcvt_n_u32_f32(a,b) __simd_vcvt_n_u32_f32(a,b)
#define vcvtq_n_f32_s32(a,b) __simd_vcvtq_n_f32_s32(a,b)
#define vcvtq_n_f32_u32(a,b) __simd_vcvtq_n_f32_u32(a,b)
#define vcvtq_n_s32_f32(a,b) __simd_vcvtq_n_s32_f32(a,b)
#define vcvtq_n_u32_f32(a,b) __simd_vcvtq_n_u32_f32(a,b)
#define vcvt_f32_f16(a) __simd_vcvt_f32_f16(a)
#define vcvt_f16_f32(a) __simd_vcvt_f16_f32(a)
#define vext_s8(a,b,c) __simd_vext_s8(a,b,c)
#define vext_s16(a,b,c) __simd_vext_s16(a,b,c)
#define vext_s32(a,b,c) __simd_vext_s32(a,b,c)
#define vext_s64(a,b,c) __simd_vext_s64(a,b,c)
#define vext_u8(a,b,c) __simd_vext_u8(a,b,c)
#define vext_u16(a,b,c) __simd_vext_u16(a,b,c)
#define vext_u32(a,b,c) __simd_vext_u32(a,b,c)
#define vext_u64(a,b,c) __simd_vext_u64(a,b,c)
#define vext_p8(a,b,c) __simd_vext_p8(a,b,c)
#define vext_p16(a,b,c) __simd_vext_p16(a,b,c)
#define vext_f32(a,b,c) __simd_vext_f32(a,b,c)
#define vextq_s8(a,b,c) __simd_vextq_s8(a,b,c)
#define vextq_s16(a,b,c) __simd_vextq_s16(a,b,c)
#define vextq_s32(a,b,c) __simd_vextq_s32(a,b,c)
#define vextq_s64(a,b,c) __simd_vextq_s64(a,b,c)
#define vextq_u8(a,b,c) __simd_vextq_u8(a,b,c)
#define vextq_u16(a,b,c) __simd_vextq_u16(a,b,c)
#define vextq_u32(a,b,c) __simd_vextq_u32(a,b,c)
#define vextq_u64(a,b,c) __simd_vextq_u64(a,b,c)
#define vextq_p8(a,b,c) __simd_vextq_p8(a,b,c)
#define vextq_p16(a,b,c) __simd_vextq_p16(a,b,c)
#define vextq_f32(a,b,c) __simd_vextq_f32(a,b,c)
#define vtbl1_s8(a,b) __simd_vtbl1_s8(a,b)
#define vtbl1_u8(a,b) __simd_vtbl1_u8(a,b)
#define vtbl1_p8(a,b) __simd_vtbl1_p8(a,b)
#define vtbl2_s8(a,b) __simd_vtbl2_s8(a,b)
#define vtbl2_u8(a,b) __simd_vtbl2_u8(a,b)
#define vtbl2_p8(a,b) __simd_vtbl2_p8(a,b)
#define vtbl3_s8(a,b) __simd_vtbl3_s8(a,b)
#define vtbl3_u8(a,b) __simd_vtbl3_u8(a,b)
#define vtbl3_p8(a,b) __simd_vtbl3_p8(a,b)
#define vtbl4_s8(a,b) __simd_vtbl4_s8(a,b)
#define vtbl4_u8(a,b) __simd_vtbl4_u8(a,b)
#define vtbl4_p8(a,b) __simd_vtbl4_p8(a,b)
#define vtbx1_s8(a,b,c) __simd_vtbx1_s8(a,b,c)
#define vtbx1_u8(a,b,c) __simd_vtbx1_u8(a,b,c)
#define vtbx1_p8(a,b,c) __simd_vtbx1_p8(a,b,c)
#define vtbx2_s8(a,b,c) __simd_vtbx2_s8(a,b,c)
#define vtbx2_u8(a,b,c) __simd_vtbx2_u8(a,b,c)
#define vtbx2_p8(a,b,c) __simd_vtbx2_p8(a,b,c)
#define vtbx3_s8(a,b,c) __simd_vtbx3_s8(a,b,c)
#define vtbx3_u8(a,b,c) __simd_vtbx3_u8(a,b,c)
#define vtbx3_p8(a,b,c) __simd_vtbx3_p8(a,b,c)
#define vtbx4_s8(a,b,c) __simd_vtbx4_s8(a,b,c)
#define vtbx4_u8(a,b,c) __simd_vtbx4_u8(a,b,c)
#define vtbx4_p8(a,b,c) __simd_vtbx4_p8(a,b,c)

#define vld1_s8(a) __simd_vld1_s8(a)
#define vld1_s16(a) __simd_vld1_s16(a)
#define vld1_s32(a) __simd_vld1_s32(a)
#define vld1_s64(a) __simd_vld1_s64(a)
#define vld1_u8(a) __simd_vld1_u8(a)
#define vld1_u16(a) __simd_vld1_u16(a)
#define vld1_u32(a) __simd_vld1_u32(a)
#define vld1_u64(a) __simd_vld1_u64(a)
#define vld1_p8(a) __simd_vld1_p8(a)
#define vld1_p16(a) __simd_vld1_p16(a)
#define vld1_f32(a) __simd_vld1_f32(a)
#define vld1q_s8(a) __simd_vld1q_s8(a)
#define vld1q_s16(a) __simd_vld1q_s16(a)
#define vld1q_s32(a) __simd_vld1q_s32(a)
#define vld1q_s64(a) __simd_vld1q_s64(a)
#define vld1q_u8(a) __simd_vld1q_u8(a)
#define vld1q_u16(a) __simd_vld1q_u16(a)
#define vld1q_u32(a) __simd_vld1q_u32(a)
#define vld1q_u64(a) __simd_vld1q_u64(a)
#define vld1q_p8(a) __simd_vld1q_p8(a)
#define vld1q_p16(a) __simd_vld1q_p16(a)
#define vld1q_f32(a) __simd_vld1q_f32(a)
#define vld1_lane_s8(a,b,c) __simd_vld1_lane_s8(a,b,c)
#define vld1_lane_s16(a,b,c) __simd_vld1_lane_s16(a,b,c)
#define vld1_lane_s32(a,b,c) __simd_vld1_lane_s32(a,b,c)
#define vld1_lane_s64(a,b,c) __simd_vld1_lane_s64(a,b,c)
#define vld1_lane_u8(a,b,c) __simd_vld1_lane_u8(a,b,c)
#define vld1_lane_u16(a,b,c) __simd_vld1_lane_u16(a,b,c)
#define vld1_lane_u32(a,b,c) __simd_vld1_lane_u32(a,b,c)
#define vld1_lane_u64(a,b,c) __simd_vld1_lane_u64(a,b,c)
#define vld1_lane_p8(a,b,c) __simd_vld1_lane_p8(a,b,c)
#define vld1_lane_p16(a,b,c) __simd_vld1_lane_p16(a,b,c)
#define vld1_lane_f32(a,b,c) __simd_vld1_lane_f32(a,b,c)
#define vld1q_lane_s8(a,b,c) __simd_vld1q_lane_s8(a,b,c)
#define vld1q_lane_s16(a,b,c) __simd_vld1q_lane_s16(a,b,c)
#define vld1q_lane_s32(a,b,c) __simd_vld1q_lane_s32(a,b,c)
#define vld1q_lane_s64(a,b,c) __simd_vld1q_lane_s64(a,b,c)
#define vld1q_lane_u8(a,b,c) __simd_vld1q_lane_u8(a,b,c)
#define vld1q_lane_u16(a,b,c) __simd_vld1q_lane_u16(a,b,c)
#define vld1q_lane_u32(a,b,c) __simd_vld1q_lane_u32(a,b,c)
#define vld1q_lane_u64(a,b,c) __simd_vld1q_lane_u64(a,b,c)
#define vld1q_lane_p8(a,b,c) __simd_vld1q_lane_p8(a,b,c)
#define vld1q_lane_p16(a,b,c) __simd_vld1q_lane_p16(a,b,c)
#define vld1q_lane_f32(a,b,c) __simd_vld1q_lane_f32(a,b,c)
#define vld1_dup_s8(a) __simd_vld1_dup_s8(a)
#define vld1_dup_s16(a) __simd_vld1_dup_s16(a)
#define vld1_dup_s32(a) __simd_vld1_dup_s32(a)
#define vld1_dup_s64(a) __simd_vld1_dup_s64(a)
#define vld1_dup_u8(a) __simd_vld1_dup_u8(a)
#define vld1_dup_u16(a) __simd_vld1_dup_u16(a)
#define vld1_dup_u32(a) __simd_vld1_dup_u32(a)
#define vld1_dup_u64(a) __simd_vld1_dup_u64(a)
#define vld1_dup_p8(a) __simd_vld1_dup_p8(a)
#define vld1_dup_p16(a) __simd_vld1_dup_p16(a)
#define vld1_dup_f32(a) __simd_vld1_dup_f32(a)
#define vld1q_dup_s8(a) __simd_vld1q_dup_s8(a)
#define vld1q_dup_s16(a) __simd_vld1q_dup_s16(a)
#define vld1q_dup_s32(a) __simd_vld1q_dup_s32(a)
#define vld1q_dup_s64(a) __simd_vld1q_dup_s64(a)
#define vld1q_dup_u8(a) __simd_vld1q_dup_u8(a)
#define vld1q_dup_u16(a) __simd_vld1q_dup_u16(a)
#define vld1q_dup_u32(a) __simd_vld1q_dup_u32(a)
#define vld1q_dup_u64(a) __simd_vld1q_dup_u64(a)
#define vld1q_dup_p8(a) __simd_vld1q_dup_p8(a)
#define vld1q_dup_p16(a) __simd_vld1q_dup_p16(a)
#define vld1q_dup_f32(a) __simd_vld1q_dup_f32(a)
#define vld2_s8(a) __simd_vld2_s8(a)
#define vld2_s16(a) __simd_vld2_s16(a)
#define vld2_s32(a) __simd_vld2_s32(a)
#define vld2_s64(a) __simd_vld2_s64(a)
#define vld2_u8(a) __simd_vld2_u8(a)
#define vld2_u16(a) __simd_vld2_u16(a)
#define vld2_u32(a) __simd_vld2_u32(a)
#define vld2_u64(a) __simd_vld2_u64(a)
#define vld2_p8(a) __simd_vld2_p8(a)
#define vld2_p16(a) __simd_vld2_p16(a)
#define vld2_f32(a) __simd_vld2_f32(a)
#define vld2q_s8(a) __simd_vld2q_s8(a)
#define vld2q_s16(a) __simd_vld2q_s16(a)
#define vld2q_s32(a) __simd_vld2q_s32(a)
#define vld2q_u8(a) __simd_vld2q_u8(a)
#define vld2q_u16(a) __simd_vld2q_u16(a)
#define vld2q_u32(a) __simd_vld2q_u32(a)
#define vld2q_p8(a) __simd_vld2q_p8(a)
#define vld2q_p16(a) __simd_vld2q_p16(a)
#define vld2q_f32(a) __simd_vld2q_f32(a)
#define vld2_lane_s8(a,b,c) __simd_vld2_lane_s8(a,b,c)
#define vld2_lane_s16(a,b,c) __simd_vld2_lane_s16(a,b,c)
#define vld2_lane_s32(a,b,c) __simd_vld2_lane_s32(a,b,c)
#define vld2_lane_u8(a,b,c) __simd_vld2_lane_u8(a,b,c)
#define vld2_lane_u16(a,b,c) __simd_vld2_lane_u16(a,b,c)
#define vld2_lane_u32(a,b,c) __simd_vld2_lane_u32(a,b,c)
#define vld2_lane_p8(a,b,c) __simd_vld2_lane_p8(a,b,c)
#define vld2_lane_p16(a,b,c) __simd_vld2_lane_p16(a,b,c)
#define vld2_lane_f32(a,b,c) __simd_vld2_lane_f32(a,b,c)
#define vld2q_lane_s16(a,b,c) __simd_vld2q_lane_s16(a,b,c)
#define vld2q_lane_s32(a,b,c) __simd_vld2q_lane_s32(a,b,c)
#define vld2q_lane_u16(a,b,c) __simd_vld2q_lane_u16(a,b,c)
#define vld2q_lane_u32(a,b,c) __simd_vld2q_lane_u32(a,b,c)
#define vld2q_lane_p16(a,b,c) __simd_vld2q_lane_p16(a,b,c)
#define vld2q_lane_f32(a,b,c) __simd_vld2q_lane_f32(a,b,c)
#define vld2_dup_s8(a) __simd_vld2_dup_s8(a)
#define vld2_dup_s16(a) __simd_vld2_dup_s16(a)
#define vld2_dup_s32(a) __simd_vld2_dup_s32(a)
#define vld2_dup_s64(a) __simd_vld2_dup_s64(a)
#define vld2_dup_u8(a) __simd_vld2_dup_u8(a)
#define vld2_dup_u16(a) __simd_vld2_dup_u16(a)
#define vld2_dup_u32(a) __simd_vld2_dup_u32(a)
#define vld2_dup_u64(a) __simd_vld2_dup_u64(a)
#define vld2_dup_p8(a) __simd_vld2_dup_p8(a)
#define vld2_dup_p16(a) __simd_vld2_dup_p16(a)
#define vld2_dup_f32(a) __simd_vld2_dup_f32(a)
#define vld3_s8(a) __simd_vld3_s8(a)
#define vld3_s16(a) __simd_vld3_s16(a)
#define vld3_s32(a) __simd_vld3_s32(a)
#define vld3_s64(a) __simd_vld3_s64(a)
#define vld3_u8(a) __simd_vld3_u8(a)
#define vld3_u16(a) __simd_vld3_u16(a)
#define vld3_u32(a) __simd_vld3_u32(a)
#define vld3_u64(a) __simd_vld3_u64(a)
#define vld3_p8(a) __simd_vld3_p8(a)
#define vld3_p16(a) __simd_vld3_p16(a)
#define vld3_f32(a) __simd_vld3_f32(a)
#define vld3q_s8(a) __simd_vld3q_s8(a)
#define vld3q_s16(a) __simd_vld3q_s16(a)
#define vld3q_s32(a) __simd_vld3q_s32(a)
#define vld3q_u8(a) __simd_vld3q_u8(a)
#define vld3q_u16(a) __simd_vld3q_u16(a)
#define vld3q_u32(a) __simd_vld3q_u32(a)
#define vld3q_p8(a) __simd_vld3q_p8(a)
#define vld3q_p16(a) __simd_vld3q_p16(a)
#define vld3q_f32(a) __simd_vld3q_f32(a)
#define vld3_lane_s8(a,b,c) __simd_vld3_lane_s8(a,b,c)
#define vld3_lane_s16(a,b,c) __simd_vld3_lane_s16(a,b,c)
#define vld3_lane_s32(a,b,c) __simd_vld3_lane_s32(a,b,c)
#define vld3_lane_u8(a,b,c) __simd_vld3_lane_u8(a,b,c)
#define vld3_lane_u16(a,b,c) __simd_vld3_lane_u16(a,b,c)
#define vld3_lane_u32(a,b,c) __simd_vld3_lane_u32(a,b,c)
#define vld3_lane_p8(a,b,c) __simd_vld3_lane_p8(a,b,c)
#define vld3_lane_p16(a,b,c) __simd_vld3_lane_p16(a,b,c)
#define vld3_lane_f32(a,b,c) __simd_vld3_lane_f32(a,b,c)
#define vld3q_lane_s16(a,b,c) __simd_vld3q_lane_s16(a,b,c)
#define vld3q_lane_s32(a,b,c) __simd_vld3q_lane_s32(a,b,c)
#define vld3q_lane_u16(a,b,c) __simd_vld3q_lane_u16(a,b,c)
#define vld3q_lane_u32(a,b,c) __simd_vld3q_lane_u32(a,b,c)
#define vld3q_lane_p16(a,b,c) __simd_vld3q_lane_p16(a,b,c)
#define vld3q_lane_f32(a,b,c) __simd_vld3q_lane_f32(a,b,c)
#define vld3_dup_s8(a) __simd_vld3_dup_s8(a)
#define vld3_dup_s16(a) __simd_vld3_dup_s16(a)
#define vld3_dup_s32(a) __simd_vld3_dup_s32(a)
#define vld3_dup_s64(a) __simd_vld3_dup_s64(a)
#define vld3_dup_u8(a) __simd_vld3_dup_u8(a)
#define vld3_dup_u16(a) __simd_vld3_dup_u16(a)
#define vld3_dup_u32(a) __simd_vld3_dup_u32(a)
#define vld3_dup_u64(a) __simd_vld3_dup_u64(a)
#define vld3_dup_p8(a) __simd_vld3_dup_p8(a)
#define vld3_dup_p16(a) __simd_vld3_dup_p16(a)
#define vld3_dup_f32(a) __simd_vld3_dup_f32(a)
#define vld4_s8(a) __simd_vld4_s8(a)
#define vld4_s16(a) __simd_vld4_s16(a)
#define vld4_s32(a) __simd_vld4_s32(a)
#define vld4_s64(a) __simd_vld4_s64(a)
#define vld4_u8(a) __simd_vld4_u8(a)
#define vld4_u16(a) __simd_vld4_u16(a)
#define vld4_u32(a) __simd_vld4_u32(a)
#define vld4_u64(a) __simd_vld4_u64(a)
#define vld4_p8(a) __simd_vld4_p8(a)
#define vld4_p16(a) __simd_vld4_p16(a)
#define vld4_f32(a) __simd_vld4_f32(a)
#define vld4q_s8(a) __simd_vld4q_s8(a)
#define vld4q_s16(a) __simd_vld4q_s16(a)
#define vld4q_s32(a) __simd_vld4q_s32(a)
#define vld4q_u8(a) __simd_vld4q_u8(a)
#define vld4q_u16(a) __simd_vld4q_u16(a)
#define vld4q_u32(a) __simd_vld4q_u32(a)
#define vld4q_p8(a) __simd_vld4q_p8(a)
#define vld4q_p16(a) __simd_vld4q_p16(a)
#define vld4q_f32(a) __simd_vld4q_f32(a)
#define vld4_lane_s8(a,b,c) __simd_vld4_lane_s8(a,b,c)
#define vld4_lane_s16(a,b,c) __simd_vld4_lane_s16(a,b,c)
#define vld4_lane_s32(a,b,c) __simd_vld4_lane_s32(a,b,c)
#define vld4_lane_u8(a,b,c) __simd_vld4_lane_u8(a,b,c)
#define vld4_lane_u16(a,b,c) __simd_vld4_lane_u16(a,b,c)
#define vld4_lane_u32(a,b,c) __simd_vld4_lane_u32(a,b,c)
#define vld4_lane_p8(a,b,c) __simd_vld4_lane_p8(a,b,c)
#define vld4_lane_p16(a,b,c) __simd_vld4_lane_p16(a,b,c)
#define vld4_lane_f32(a,b,c) __simd_vld4_lane_f32(a,b,c)
#define vld4q_lane_s16(a,b,c) __simd_vld4q_lane_s16(a,b,c)
#define vld4q_lane_s32(a,b,c) __simd_vld4q_lane_s32(a,b,c)
#define vld4q_lane_u16(a,b,c) __simd_vld4q_lane_u16(a,b,c)
#define vld4q_lane_u32(a,b,c) __simd_vld4q_lane_u32(a,b,c)
#define vld4q_lane_p16(a,b,c) __simd_vld4q_lane_p16(a,b,c)
#define vld4q_lane_f32(a,b,c) __simd_vld4q_lane_f32(a,b,c)
#define vld4_dup_s8(a) __simd_vld4_dup_s8(a)
#define vld4_dup_s16(a) __simd_vld4_dup_s16(a)
#define vld4_dup_s32(a) __simd_vld4_dup_s32(a)
#define vld4_dup_s64(a) __simd_vld4_dup_s64(a)
#define vld4_dup_u8(a) __simd_vld4_dup_u8(a)
#define vld4_dup_u16(a) __simd_vld4_dup_u16(a)
#define vld4_dup_u32(a) __simd_vld4_dup_u32(a)
#define vld4_dup_u64(a) __simd_vld4_dup_u64(a)
#define vld4_dup_p8(a) __simd_vld4_dup_p8(a)
#define vld4_dup_p16(a) __simd_vld4_dup_p16(a)
#define vld4_dup_f32(a) __simd_vld4_dup_f32(a)

#define vst1_s8(a,b) __simd_vst1_s8(a,b)
#define vst1_s16(a,b) __simd_vst1_s16(a,b)
#define vst1_s32(a,b) __simd_vst1_s32(a,b)
#define vst1_s64(a,b) __simd_vst1_s64(a,b)
#define vst1_u8(a,b) __simd_vst1_u8(a,b)
#define vst1_u16(a,b) __simd_vst1_u16(a,b)
#define vst1_u32(a,b) __simd_vst1_u32(a,b)
#define vst1_u64(a,b) __simd_vst1_u64(a,b)
#define vst1_p8(a,b) __simd_vst1_p8(a,b)
#define vst1_p16(a,b) __simd_vst1_p16(a,b)
#define vst1_f32(a,b) __simd_vst1_f32(a,b)
#define vst1q_s8(a,b) __simd_vst1q_s8(a,b)
#define vst1q_s16(a,b) __simd_vst1q_s16(a,b)
#define vst1q_s32(a,b) __simd_vst1q_s32(a,b)
#define vst1q_s64(a,b) __simd_vst1q_s64(a,b)
#define vst1q_u8(a,b) __simd_vst1q_u8(a,b)
#define vst1q_u16(a,b) __simd_vst1q_u16(a,b)
#define vst1q_u32(a,b) __simd_vst1q_u32(a,b)
#define vst1q_u64(a,b) __simd_vst1q_u64(a,b)
#define vst1q_p8(a,b) __simd_vst1q_p8(a,b)
#define vst1q_p16(a,b) __simd_vst1q_p16(a,b)
#define vst1q_f32(a,b) __simd_vst1q_f32(a,b)
#define vst1_lane_s8(a,b,c) __simd_vst1_lane_s8(a,b,c)
#define vst1_lane_s16(a,b,c) __simd_vst1_lane_s16(a,b,c)
#define vst1_lane_s32(a,b,c) __simd_vst1_lane_s32(a,b,c)
#define vst1_lane_s64(a,b,c) __simd_vst1_lane_s64(a,b,c)
#define vst1_lane_u8(a,b,c) __simd_vst1_lane_u8(a,b,c)
#define vst1_lane_u16(a,b,c) __simd_vst1_lane_u16(a,b,c)
#define vst1_lane_u32(a,b,c) __simd_vst1_lane_u32(a,b,c)
#define vst1_lane_u64(a,b,c) __simd_vst1_lane_u64(a,b,c)
#define vst1_lane_p8(a,b,c) __simd_vst1_lane_p8(a,b,c)
#define vst1_lane_p16(a,b,c) __simd_vst1_lane_p16(a,b,c)
#define vst1_lane_f32(a,b,c) __simd_vst1_lane_f32(a,b,c)
#define vst1q_lane_s8(a,b,c) __simd_vst1q_lane_s8(a,b,c)
#define vst1q_lane_s16(a,b,c) __simd_vst1q_lane_s16(a,b,c)
#define vst1q_lane_s32(a,b,c) __simd_vst1q_lane_s32(a,b,c)
#define vst1q_lane_s64(a,b,c) __simd_vst1q_lane_s64(a,b,c)
#define vst1q_lane_u8(a,b,c) __simd_vst1q_lane_u8(a,b,c)
#define vst1q_lane_u16(a,b,c) __simd_vst1q_lane_u16(a,b,c)
#define vst1q_lane_u32(a,b,c) __simd_vst1q_lane_u32(a,b,c)
#define vst1q_lane_u64(a,b,c) __simd_vst1q_lane_u64(a,b,c)
#define vst1q_lane_p8(a,b,c) __simd_vst1q_lane_p8(a,b,c)
#define vst1q_lane_p16(a,b,c) __simd_vst1q_lane_p16(a,b,c)
#define vst1q_lane_f32(a,b,c) __simd_vst1q_lane_f32(a,b,c)
#define vst2_s8(a,b) __simd_vst2_s8(a,b)
#define vst2_s16(a,b) __simd_vst2_s16(a,b)
#define vst2_s32(a,b) __simd_vst2_s32(a,b)
#define vst2_s64(a,b) __simd_vst2_s64(a,b)
#define vst2_u8(a,b) __simd_vst2_u8(a,b)
#define vst2_u16(a,b) __simd_vst2_u16(a,b)
#define vst2_u32(a,b) __simd_vst2_u32(a,b)
#define vst2_u64(a,b) __simd_vst2_u64(a,b)
#define vst2_p8(a,b) __simd_vst2_p8(a,b)
#define vst2_p16(a,b) __simd_vst2_p16(a,b)
#define vst2_f32(a,b) __simd_vst2_f32(a,b)
#define vst2q_s8(a,b) __simd_vst2q_s8(a,b)
#define vst2q_s16(a,b) __simd_vst2q_s16(a,b)
#define vst2q_s32(a,b) __simd_vst2q_s32(a,b)
#define vst2q_u8(a,b) __simd_vst2q_u8(a,b)
#define vst2q_u16(a,b) __simd_vst2q_u16(a,b)
#define vst2q_u32(a,b) __simd_vst2q_u32(a,b)
#define vst2q_p8(a,b) __simd_vst2q_p8(a,b)
#define vst2q_p16(a,b) __simd_vst2q_p16(a,b)
#define vst2q_f32(a,b) __simd_vst2q_f32(a,b)
#define vst2_lane_s8(a,b,c) __simd_vst2_lane_s8(a,b,c)
#define vst2_lane_s16(a,b,c) __simd_vst2_lane_s16(a,b,c)
#define vst2_lane_s32(a,b,c) __simd_vst2_lane_s32(a,b,c)
#define vst2_lane_u8(a,b,c) __simd_vst2_lane_u8(a,b,c)
#define vst2_lane_u16(a,b,c) __simd_vst2_lane_u16(a,b,c)
#define vst2_lane_u32(a,b,c) __simd_vst2_lane_u32(a,b,c)
#define vst2_lane_p8(a,b,c) __simd_vst2_lane_p8(a,b,c)
#define vst2_lane_p16(a,b,c) __simd_vst2_lane_p16(a,b,c)
#define vst2_lane_f32(a,b,c) __simd_vst2_lane_f32(a,b,c)
#define vst2q_lane_s16(a,b,c) __simd_vst2q_lane_s16(a,b,c)
#define vst2q_lane_s32(a,b,c) __simd_vst2q_lane_s32(a,b,c)
#define vst2q_lane_u16(a,b,c) __simd_vst2q_lane_u16(a,b,c)
#define vst2q_lane_u32(a,b,c) __simd_vst2q_lane_u32(a,b,c)
#define vst2q_lane_p16(a,b,c) __simd_vst2q_lane_p16(a,b,c)
#define vst2q_lane_f32(a,b,c) __simd_vst2q_lane_f32(a,b,c)
#define vst3_s8(a,b) __simd_vst3_s8(a,b)
#define vst3_s16(a,b) __simd_vst3_s16(a,b)
#define vst3_s32(a,b) __simd_vst3_s32(a,b)
#define vst3_s64(a,b) __simd_vst3_s64(a,b)
#define vst3_u8(a,b) __simd_vst3_u8(a,b)
#define vst3_u16(a,b) __simd_vst3_u16(a,b)
#define vst3_u32(a,b) __simd_vst3_u32(a,b)
#define vst3_u64(a,b) __simd_vst3_u64(a,b)
#define vst3_p8(a,b) __simd_vst3_p8(a,b)
#define vst3_p16(a,b) __simd_vst3_p16(a,b)
#define vst3_f32(a,b) __simd_vst3_f32(a,b)
#define vst3q_s8(a,b) __simd_vst3q_s8(a,b)
#define vst3q_s16(a,b) __simd_vst3q_s16(a,b)
#define vst3q_s32(a,b) __simd_vst3q_s32(a,b)
#define vst3q_u8(a,b) __simd_vst3q_u8(a,b)
#define vst3q_u16(a,b) __simd_vst3q_u16(a,b)
#define vst3q_u32(a,b) __simd_vst3q_u32(a,b)
#define vst3q_p8(a,b) __simd_vst3q_p8(a,b)
#define vst3q_p16(a,b) __simd_vst3q_p16(a,b)
#define vst3q_f32(a,b) __simd_vst3q_f32(a,b)
#define vst3_lane_s8(a,b,c) __simd_vst3_lane_s8(a,b,c)
#define vst3_lane_s16(a,b,c) __simd_vst3_lane_s16(a,b,c)
#define vst3_lane_s32(a,b,c) __simd_vst3_lane_s32(a,b,c)
#define vst3_lane_u8(a,b,c) __simd_vst3_lane_u8(a,b,c)
#define vst3_lane_u16(a,b,c) __simd_vst3_lane_u16(a,b,c)
#define vst3_lane_u32(a,b,c) __simd_vst3_lane_u32(a,b,c)
#define vst3_lane_p8(a,b,c) __simd_vst3_lane_p8(a,b,c)
#define vst3_lane_p16(a,b,c) __simd_vst3_lane_p16(a,b,c)
#define vst3_lane_f32(a,b,c) __simd_vst3_lane_f32(a,b,c)
#define vst3q_lane_s16(a,b,c) __simd_vst3q_lane_s16(a,b,c)
#define vst3q_lane_s32(a,b,c) __simd_vst3q_lane_s32(a,b,c)
#define vst3q_lane_u16(a,b,c) __simd_vst3q_lane_u16(a,b,c)
#define vst3q_lane_u32(a,b,c) __simd_vst3q_lane_u32(a,b,c)
#define vst3q_lane_p16(a,b,c) __simd_vst3q_lane_p16(a,b,c)
#define vst3q_lane_f32(a,b,c) __simd_vst3q_lane_f32(a,b,c)
#define vst4_s8(a,b) __simd_vst4_s8(a,b)
#define vst4_s16(a,b) __simd_vst4_s16(a,b)
#define vst4_s32(a,b) __simd_vst4_s32(a,b)
#define vst4_s64(a,b) __simd_vst4_s64(a,b)
#define vst4_u8(a,b) __simd_vst4_u8(a,b)
#define vst4_u16(a,b) __simd_vst4_u16(a,b)
#define vst4_u32(a,b) __simd_vst4_u32(a,b)
#define vst4_u64(a,b) __simd_vst4_u64(a,b)
#define vst4_p8(a,b) __simd_vst4_p8(a,b)
#define vst4_p16(a,b) __simd_vst4_p16(a,b)
#define vst4_f32(a,b) __simd_vst4_f32(a,b)
#define vst4q_s8(a,b) __simd_vst4q_s8(a,b)
#define vst4q_s16(a,b) __simd_vst4q_s16(a,b)
#define vst4q_s32(a,b) __simd_vst4q_s32(a,b)
#define vst4q_u8(a,b) __simd_vst4q_u8(a,b)
#define vst4q_u16(a,b) __simd_vst4q_u16(a,b)
#define vst4q_u32(a,b) __simd_vst4q_u32(a,b)
#define vst4q_p8(a,b) __simd_vst4q_p8(a,b)
#define vst4q_p16(a,b) __simd_vst4q_p16(a,b)
#define vst4q_f32(a,b) __simd_vst4q_f32(a,b)
#define vst4_lane_s8(a,b,c) __simd_vst4_lane_s8(a,b,c)
#define vst4_lane_s16(a,b,c) __simd_vst4_lane_s16(a,b,c)
#define vst4_lane_s32(a,b,c) __simd_vst4_lane_s32(a,b,c)
#define vst4_lane_u8(a,b,c) __simd_vst4_lane_u8(a,b,c)
#define vst4_lane_u16(a,b,c) __simd_vst4_lane_u16(a,b,c)
#define vst4_lane_u32(a,b,c) __simd_vst4_lane_u32(a,b,c)
#define vst4_lane_p8(a,b,c) __simd_vst4_lane_p8(a,b,c)
#define vst4_lane_p16(a,b,c) __simd_vst4_lane_p16(a,b,c)
#define vst4_lane_f32(a,b,c) __simd_vst4_lane_f32(a,b,c)
#define vst4q_lane_s16(a,b,c) __simd_vst4q_lane_s16(a,b,c)
#define vst4q_lane_s32(a,b,c) __simd_vst4q_lane_s32(a,b,c)
#define vst4q_lane_u16(a,b,c) __simd_vst4q_lane_u16(a,b,c)
#define vst4q_lane_u32(a,b,c) __simd_vst4q_lane_u32(a,b,c)
#define vst4q_lane_p16(a,b,c) __simd_vst4q_lane_p16(a,b,c)
#define vst4q_lane_f32(a,b,c) __simd_vst4q_lane_f32(a,b,c)

#define vget_lane_s8(a,b) __simd_vget_lane_s8(a,b)
#define vget_lane_s16(a,b) __simd_vget_lane_s16(a,b)
#define vget_lane_s32(a,b) __simd_vget_lane_s32(a,b)
#define vget_lane_s64(a,b) __simd_vget_lane_s64(a,b)
#define vget_lane_u8(a,b) __simd_vget_lane_u8(a,b)
#define vget_lane_u16(a,b) __simd_vget_lane_u16(a,b)
#define vget_lane_u32(a,b) __simd_vget_lane_u32(a,b)
#define vget_lane_u64(a,b) __simd_vget_lane_u64(a,b)
#define vget_lane_p8(a,b) __simd_vget_lane_p8(a,b)
#define vget_lane_p16(a,b) __simd_vget_lane_p16(a,b)
#define vget_lane_f32(a,b) __simd_vget_lane_f32(a,b)
#define vgetq_lane_s8(a,b) __simd_vgetq_lane_s8(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 vgetq_lane_s64(a,b) __simd_vgetq_lane_s64(a,b)
#define vgetq_lane_u8(a,b) __simd_vgetq_lane_u8(a,b)
#define vgetq_lane_u16(a,b) __simd_vgetq_lane_u16(a,b)
#define vgetq_lane_u32(a,b) __simd_vgetq_lane_u32(a,b)
#define vgetq_lane_u64(a,b) __simd_vgetq_lane_u64(a,b)
#define vgetq_lane_p8(a,b) __simd_vgetq_lane_p8(a,b)
#define vgetq_lane_p16(a,b) __simd_vgetq_lane_p16(a,b)
#define vgetq_lane_f32(a,b) __simd_vgetq_lane_f32(a,b)
#define vset_lane_s8(a,b,c) __simd_vset_lane_s8(a,b,c)
#define vset_lane_s16(a,b,c) __simd_vset_lane_s16(a,b,c)
#define vset_lane_s32(a,b,c) __simd_vset_lane_s32(a,b,c)
#define vset_lane_s64(a,b,c) __simd_vset_lane_s64(a,b,c)
#define vset_lane_u8(a,b,c) __simd_vset_lane_u8(a,b,c)
#define vset_lane_u16(a,b,c) __simd_vset_lane_u16(a,b,c)
#define vset_lane_u32(a,b,c) __simd_vset_lane_u32(a,b,c)
#define vset_lane_u64(a,b,c) __simd_vset_lane_u64(a,b,c)
#define vset_lane_p8(a,b,c) __simd_vset_lane_p8(a,b,c)
#define vset_lane_p16(a,b,c) __simd_vset_lane_p16(a,b,c)
#define vset_lane_f32(a,b,c) __simd_vset_lane_f32(a,b,c)
#define vsetq_lane_s8(a,b,c) __simd_vsetq_lane_s8(a,b,c)
#define vsetq_lane_s16(a,b,c) __simd_vsetq_lane_s16(a,b,c)
#define vsetq_lane_s32(a,b,c) __simd_vsetq_lane_s32(a,b,c)
#define vsetq_lane_s64(a,b,c) __simd_vsetq_lane_s64(a,b,c)
#define vsetq_lane_u8(a,b,c) __simd_vsetq_lane_u8(a,b,c)
#define vsetq_lane_u16(a,b,c) __simd_vsetq_lane_u16(a,b,c)
#define vsetq_lane_u32(a,b,c) __simd_vsetq_lane_u32(a,b,c)
#define vsetq_lane_u64(a,b,c) __simd_vsetq_lane_u64(a,b,c)
#define vsetq_lane_p8(a,b,c) __simd_vsetq_lane_p8(a,b,c)
#define vsetq_lane_p16(a,b,c) __simd_vsetq_lane_p16(a,b,c)
#define vsetq_lane_f32(a,b,c) __simd_vsetq_lane_f32(a,b,c)
#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_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_p8(a) __simd_vcreate_p8(a)
#define vcreate_p16(a) __simd_vcreate_p16(a)
#define vcreate_f32(a) __simd_vcreate_f32(a)
#define vcombine_s8(a,b) __simd_vcombine_s8(a,b)
#define vcombine_s16(a,b) __simd_vcombine_s16(a,b)
#define vcombine_s32(a,b) __simd_vcombine_s32(a,b)
#define vcombine_s64(a,b) __simd_vcombine_s64(a,b)
#define vcombine_u8(a,b) __simd_vcombine_u8(a,b)
#define vcombine_u16(a,b) __simd_vcombine_u16(a,b)
#define vcombine_u32(a,b) __simd_vcombine_u32(a,b)
#define vcombine_u64(a,b) __simd_vcombine_u64(a,b)
#define vcombine_p8(a,b) __simd_vcombine_p8(a,b)
#define vcombine_p16(a,b) __simd_vcombine_p16(a,b)
#define vcombine_f32(a,b) __simd_vcombine_f32(a,b)
#define vget_high_s8(a) __simd_vget_high_s8(a)
#define vget_high_s16(a) __simd_vget_high_s16(a)
#define vget_high_s32(a) __simd_vget_high_s32(a)
#define vget_high_s64(a) __simd_vget_high_s64(a)
#define vget_high_u8(a) __simd_vget_high_u8(a)
#define vget_high_u16(a) __simd_vget_high_u16(a)
#define vget_high_u32(a) __simd_vget_high_u32(a)
#define vget_high_u64(a) __simd_vget_high_u64(a)
#define vget_high_p8(a) __simd_vget_high_p8(a)
#define vget_high_p16(a) __simd_vget_high_p16(a)
#define vget_high_f32(a) __simd_vget_high_f32(a)
#define vget_low_s8(a) __simd_vget_low_s8(a)
#define vget_low_s16(a) __simd_vget_low_s16(a)
#define vget_low_s32(a) __simd_vget_low_s32(a)
#define vget_low_s64(a) __simd_vget_low_s64(a)
#define vget_low_u8(a) __simd_vget_low_u8(a)
#define vget_low_u16(a) __simd_vget_low_u16(a)
#define vget_low_u32(a) __simd_vget_low_u32(a)
#define vget_low_u64(a) __simd_vget_low_u64(a)
#define vget_low_p8(a) __simd_vget_low_p8(a)
#define vget_low_p16(a) __simd_vget_low_p16(a)
#define vget_low_f32(a) __simd_vget_low_f32(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_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_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_s8(a) __simd_vreinterpret_s16_s8(a)
#define vreinterpret_s16_s32(a) __simd_vreinterpret_s16_s32(a)
#define vreinterpret_s16_s64(a) __simd_vreinterpret_s16_s64(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_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_s8(a) __simd_vreinterpret_s32_s8(a)
#define vreinterpret_s32_s16(a) __simd_vreinterpret_s32_s16(a)
#define vreinterpret_s32_s64(a) __simd_vreinterpret_s32_s64(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_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_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_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_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_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_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_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_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_u8(a) __simd_vreinterpret_u16_u8(a)
#define vreinterpret_u16_u32(a) __simd_vreinterpret_u16_u32(a)
#define vreinterpret_u16_u64(a) __simd_vreinterpret_u16_u64(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_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_u8(a) __simd_vreinterpret_u32_u8(a)
#define vreinterpret_u32_u16(a) __simd_vreinterpret_u32_u16(a)
#define vreinterpret_u32_u64(a) __simd_vreinterpret_u32_u64(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_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_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_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_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_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_p16(a) __simd_vreinterpret_p8_p16(a)
#define vreinterpret_p8_f32(a) __simd_vreinterpret_p8_f32(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_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_p8(a) __simd_vreinterpret_p16_p8(a)
#define vreinterpret_p16_f32(a) __simd_vreinterpret_p16_f32(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_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_p8(a) __simd_vreinterpret_f32_p8(a)
#define vreinterpret_f32_p16(a) __simd_vreinterpret_f32_p16(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_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_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_s16_s8(a) __simd_vreinterpretq_s16_s8(a)
#define vreinterpretq_s16_s32(a) __simd_vreinterpretq_s16_s32(a)
#define vreinterpretq_s16_s64(a) __simd_vreinterpretq_s16_s64(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_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_s32_s8(a) __simd_vreinterpretq_s32_s8(a)
#define vreinterpretq_s32_s16(a) __simd_vreinterpretq_s32_s16(a)
#define vreinterpretq_s32_s64(a) __simd_vreinterpretq_s32_s64(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_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_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_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_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_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_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_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_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_u8(a) __simd_vreinterpretq_u16_u8(a)
#define vreinterpretq_u16_u32(a) __simd_vreinterpretq_u16_u32(a)
#define vreinterpretq_u16_u64(a) __simd_vreinterpretq_u16_u64(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_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_u8(a) __simd_vreinterpretq_u32_u8(a)
#define vreinterpretq_u32_u16(a) __simd_vreinterpretq_u32_u16(a)
#define vreinterpretq_u32_u64(a) __simd_vreinterpretq_u32_u64(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_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_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_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_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_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_p16(a) __simd_vreinterpretq_p8_p16(a)
#define vreinterpretq_p8_f32(a) __simd_vreinterpretq_p8_f32(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_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_p8(a) __simd_vreinterpretq_p16_p8(a)
#define vreinterpretq_p16_f32(a) __simd_vreinterpretq_p16_f32(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_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_p8(a) __simd_vreinterpretq_f32_p8(a)
#define vreinterpretq_f32_p16(a) __simd_vreinterpretq_f32_p16(a)


#ifdef __cplusplus
}
#endif

#endif /* __LLONG_BIT */

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

#endif