#ifdef __ARM_NEON
#include <arm_neon.h>

#define ADD vaddq_u32
#define VAL vdupq_n_u32
#define word_t uint32x4_t
#define INPUT(k, set, ptr, offs, idx, var) ADD(var, VAL(k))
#define LOAD INPUT
#define LOAD4(set, ptr, offs, idx, var0, var1, var2, var3) { \
	uint32x4x2_t in01 = vzipq_u32( \
		vreinterpretq_u32_u8(vld1q_u8((uint8_t*)ptr[0+set*4] + offs + idx*4)), \
		vreinterpretq_u32_u8(vld1q_u8((uint8_t*)ptr[1+set*4] + offs + idx*4)) \
	); \
	uint32x4x2_t in23 = vzipq_u32( \
		vreinterpretq_u32_u8(vld1q_u8((uint8_t*)ptr[2+set*4] + offs + idx*4)), \
		vreinterpretq_u32_u8(vld1q_u8((uint8_t*)ptr[3+set*4] + offs + idx*4)) \
	); \
	var0 = vcombine_u32(vget_low_u32(in01.val[0]), vget_low_u32(in23.val[0])); \
	var1 = vcombine_u32(vget_high_u32(in01.val[0]), vget_high_u32(in23.val[0])); \
	var2 = vcombine_u32(vget_low_u32(in01.val[1]), vget_low_u32(in23.val[1])); \
	var3 = vcombine_u32(vget_high_u32(in01.val[1]), vget_high_u32(in23.val[1])); \
}

#define ROTATE(a, r) r==16 ? vreinterpretq_u32_u16(vrev32q_u16(vreinterpretq_u16_u32(a))) : vsliq_n_u32(vshrq_n_u32(a, 32-r), a, r)

#define md5mb_regions_neon 4
#define md5mb_max_regions_neon md5mb_regions_neon
#define md5mb_alignment_neon 16

#define F vbslq_u32
#define G(b,c,d) vbslq_u32(d, b, c)
#define H(b,c,d) veorq_u32(veorq_u32(d, c), b)
#define I(b,c,d) veorq_u32(vornq_u32(b, d), c)

/* the following (I function shortcut) often performs worse - is it because VBIT is more expensive than VORN?
#define F 1
#define G 2
#define H 3
#define I 4
#define ADDF(f,a,b,c,d) ( \
	f==G ? ADD(a, vbslq_u32(d, b, c)) : ( \
	f==I ? vsubq_u32(a, vbslq_u32(b, c, veorq_u32(c, d))) : ADD(a, f==F ? \
		vbslq_u32(b, c, d) : \
		veorq_u32(veorq_u32(d, c), b) \
	) \
	) \
)
#define IOFFSET -1
*/


#define _FN(f) f##_neon
#include "md5mb-base.h"
#define MD5X2
#include "md5mb-base.h"
#undef MD5X2
#undef _FN


#undef ROTATE
#undef ADD
#undef VAL
#undef word_t
#undef INPUT
#undef LOAD
#undef LOAD4

#undef F
#undef G
#undef H
#undef I
//#undef ADDF
//#undef IOFFSET

static HEDLEY_ALWAYS_INLINE void md5_extract_mb_neon(void* dst, void* state, int idx) {
	uint32x4_t* state_ = (uint32x4_t*)state + (idx & 4);
	// re-arrange to pairs
	uint32x4x2_t tmp1 = vzipq_u32(state_[0], state_[1]);
	uint32x4x2_t tmp2 = vzipq_u32(state_[2], state_[3]);
	
	idx &= 3;
	if(idx == 0)
		vst1q_u8((uint8_t*)dst, vreinterpretq_u8_u32(vcombine_u32(vget_low_u32(tmp1.val[0]), vget_low_u32(tmp2.val[0]))));
	if(idx == 1)
		vst1q_u8((uint8_t*)dst, vreinterpretq_u8_u32(vcombine_u32(vget_high_u32(tmp1.val[0]), vget_high_u32(tmp2.val[0]))));
	if(idx == 2)
		vst1q_u8((uint8_t*)dst, vreinterpretq_u8_u32(vcombine_u32(vget_low_u32(tmp1.val[1]), vget_low_u32(tmp2.val[1]))));
	if(idx == 3)
		vst1q_u8((uint8_t*)dst, vreinterpretq_u8_u32(vcombine_u32(vget_high_u32(tmp1.val[1]), vget_high_u32(tmp2.val[1]))));
}
static HEDLEY_ALWAYS_INLINE void md5_extract_all_mb_neon(void* dst, void* state, int group) {
	uint32x4_t* state_ = (uint32x4_t*)state + group*4;
	uint32x4x4_t t;
	t.val[0] = state_[0];
	t.val[1] = state_[1];
	t.val[2] = state_[2];
	t.val[3] = state_[3];
#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
	t.val[0] = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(t.val[0])));
	t.val[1] = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(t.val[1])));
	t.val[2] = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(t.val[2])));
	t.val[3] = vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(t.val[3])));
#endif
	vst4q_u32((uint32_t*)dst, t);
}

#endif