andre@0: /* This Source Code Form is subject to the terms of the Mozilla Public
andre@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
andre@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
andre@0: 
andre@0: /* This implementation of poly1305 is by Andrew Moon
andre@0:  * (https://github.com/floodyberry/poly1305-donna) and released as public
andre@0:  * domain. It implements SIMD vectorization based on the algorithm described in
andre@0:  * http://cr.yp.to/papers.html#neoncrypto. Unrolled to 2 powers, i.e. 64 byte
andre@0:  * block size. */
andre@0: 
andre@0: #include <emmintrin.h>
andre@0: #include <stdint.h>
andre@0: 
andre@0: #include "poly1305.h"
andre@0: 
andre@0: #define ALIGN(x) __attribute__((aligned(x)))
andre@0: #define INLINE inline
andre@0: #define U8TO64_LE(m) (*(uint64_t*)(m))
andre@0: #define U8TO32_LE(m) (*(uint32_t*)(m))
andre@0: #define U64TO8_LE(m,v) (*(uint64_t*)(m)) = v
andre@0: 
andre@0: typedef __m128i xmmi;
andre@0: typedef unsigned __int128 uint128_t;
andre@0: 
andre@0: static const uint32_t ALIGN(16) poly1305_x64_sse2_message_mask[4] = {(1 << 26) - 1, 0, (1 << 26) - 1, 0};
andre@0: static const uint32_t ALIGN(16) poly1305_x64_sse2_5[4] = {5, 0, 5, 0};
andre@0: static const uint32_t ALIGN(16) poly1305_x64_sse2_1shl128[4] = {(1 << 24), 0, (1 << 24), 0};
andre@0: 
andre@0: static uint128_t INLINE
andre@0: add128(uint128_t a, uint128_t b) {
andre@0: 	return a + b;
andre@0: }
andre@0: 
andre@0: static uint128_t INLINE
andre@0: add128_64(uint128_t a, uint64_t b) {
andre@0: 	return a + b;
andre@0: }
andre@0: 
andre@0: static uint128_t INLINE
andre@0: mul64x64_128(uint64_t a, uint64_t b) {
andre@0: 	return (uint128_t)a * b;
andre@0: }
andre@0: 
andre@0: static uint64_t INLINE
andre@0: lo128(uint128_t a) {
andre@0: 	return (uint64_t)a;
andre@0: }
andre@0: 
andre@0: static uint64_t INLINE
andre@0: shr128(uint128_t v, const int shift) {
andre@0: 	return (uint64_t)(v >> shift);
andre@0: }
andre@0: 
andre@0: static uint64_t INLINE
andre@0: shr128_pair(uint64_t hi, uint64_t lo, const int shift) {
andre@0: 	return (uint64_t)((((uint128_t)hi << 64) | lo) >> shift);
andre@0: }
andre@0: 
andre@0: typedef struct poly1305_power_t {
andre@0: 	union {
andre@0: 		xmmi v;
andre@0: 		uint64_t u[2];
andre@0: 		uint32_t d[4];
andre@0: 	} R20,R21,R22,R23,R24,S21,S22,S23,S24;
andre@0: } poly1305_power;
andre@0: 
andre@0: typedef struct poly1305_state_internal_t {
andre@0: 	poly1305_power P[2];     /* 288 bytes, top 32 bit halves unused = 144 bytes of free storage */
andre@0: 	union {
andre@0: 		xmmi H[5];           /*  80 bytes  */
andre@0: 		uint64_t HH[10];
andre@0: 	};
andre@0: 	/* uint64_t r0,r1,r2;       [24 bytes] */
andre@0: 	/* uint64_t pad0,pad1;      [16 bytes] */
andre@0: 	uint64_t started;        /*   8 bytes  */
andre@0: 	uint64_t leftover;       /*   8 bytes  */
andre@0: 	uint8_t buffer[64];      /*  64 bytes  */
andre@0: } poly1305_state_internal;   /* 448 bytes total + 63 bytes for alignment = 511 bytes raw */
andre@0: 
andre@0: static poly1305_state_internal INLINE
andre@0: *poly1305_aligned_state(poly1305_state *state) {
andre@0: 	return (poly1305_state_internal *)(((uint64_t)state + 63) & ~63);
andre@0: }
andre@0: 
andre@0: /* copy 0-63 bytes */
andre@0: static void INLINE
andre@0: poly1305_block_copy(uint8_t *dst, const uint8_t *src, size_t bytes) {
andre@0: 	size_t offset = src - dst;
andre@0: 	if (bytes & 32) {
andre@0: 		_mm_storeu_si128((xmmi *)(dst + 0), _mm_loadu_si128((xmmi *)(dst + offset + 0)));
andre@0: 		_mm_storeu_si128((xmmi *)(dst + 16), _mm_loadu_si128((xmmi *)(dst + offset + 16)));
andre@0: 		dst += 32;
andre@0: 	}
andre@0: 	if (bytes & 16) { _mm_storeu_si128((xmmi *)dst, _mm_loadu_si128((xmmi *)(dst + offset))); dst += 16; }
andre@0: 	if (bytes &  8) { *(uint64_t *)dst = *(uint64_t *)(dst + offset); dst += 8; }
andre@0: 	if (bytes &  4) { *(uint32_t *)dst = *(uint32_t *)(dst + offset); dst += 4; }
andre@0: 	if (bytes &  2) { *(uint16_t *)dst = *(uint16_t *)(dst + offset); dst += 2; }
andre@0: 	if (bytes &  1) { *( uint8_t *)dst = *( uint8_t *)(dst + offset);           }
andre@0: }
andre@0: 
andre@0: /* zero 0-15 bytes */
andre@0: static void INLINE
andre@0: poly1305_block_zero(uint8_t *dst, size_t bytes) {
andre@0: 	if (bytes &  8) { *(uint64_t *)dst = 0; dst += 8; }
andre@0: 	if (bytes &  4) { *(uint32_t *)dst = 0; dst += 4; }
andre@0: 	if (bytes &  2) { *(uint16_t *)dst = 0; dst += 2; }
andre@0: 	if (bytes &  1) { *( uint8_t *)dst = 0; }
andre@0: }
andre@0: 
andre@0: static size_t INLINE
andre@0: poly1305_min(size_t a, size_t b) {
andre@0: 	return (a < b) ? a : b;
andre@0: }
andre@0: 
andre@0: void
andre@0: Poly1305Init(poly1305_state *state, const unsigned char key[32]) {
andre@0: 	poly1305_state_internal *st = poly1305_aligned_state(state);
andre@0: 	poly1305_power *p;
andre@0: 	uint64_t r0,r1,r2;
andre@0: 	uint64_t t0,t1;
andre@0: 
andre@0: 	/* clamp key */
andre@0: 	t0 = U8TO64_LE(key + 0);
andre@0: 	t1 = U8TO64_LE(key + 8);
andre@0: 	r0 = t0 & 0xffc0fffffff; t0 >>= 44; t0 |= t1 << 20;
andre@0: 	r1 = t0 & 0xfffffc0ffff; t1 >>= 24;
andre@0: 	r2 = t1 & 0x00ffffffc0f;
andre@0: 
andre@0: 	/* store r in un-used space of st->P[1] */
andre@0: 	p = &st->P[1];
andre@0: 	p->R20.d[1] = (uint32_t)(r0      );
andre@0: 	p->R20.d[3] = (uint32_t)(r0 >> 32);
andre@0: 	p->R21.d[1] = (uint32_t)(r1      );
andre@0: 	p->R21.d[3] = (uint32_t)(r1 >> 32);
andre@0: 	p->R22.d[1] = (uint32_t)(r2      );
andre@0: 	p->R22.d[3] = (uint32_t)(r2 >> 32);
andre@0: 
andre@0: 	/* store pad */
andre@0: 	p->R23.d[1] = U8TO32_LE(key + 16);
andre@0: 	p->R23.d[3] = U8TO32_LE(key + 20);
andre@0: 	p->R24.d[1] = U8TO32_LE(key + 24);
andre@0: 	p->R24.d[3] = U8TO32_LE(key + 28);
andre@0: 
andre@0: 	/* H = 0 */
andre@0: 	st->H[0] = _mm_setzero_si128();
andre@0: 	st->H[1] = _mm_setzero_si128();
andre@0: 	st->H[2] = _mm_setzero_si128();
andre@0: 	st->H[3] = _mm_setzero_si128();
andre@0: 	st->H[4] = _mm_setzero_si128();
andre@0: 
andre@0: 	st->started = 0;
andre@0: 	st->leftover = 0;
andre@0: }
andre@0: 
andre@0: static void
andre@0: poly1305_first_block(poly1305_state_internal *st, const uint8_t *m) {
andre@0: 	const xmmi MMASK = _mm_load_si128((xmmi *)poly1305_x64_sse2_message_mask);
andre@0: 	const xmmi FIVE = _mm_load_si128((xmmi*)poly1305_x64_sse2_5);
andre@0: 	const xmmi HIBIT = _mm_load_si128((xmmi*)poly1305_x64_sse2_1shl128);
andre@0: 	xmmi T5,T6;
andre@0: 	poly1305_power *p;
andre@0: 	uint128_t d[3];
andre@0: 	uint64_t r0,r1,r2;
andre@0: 	uint64_t r20,r21,r22,s22;
andre@0: 	uint64_t pad0,pad1;
andre@0: 	uint64_t c;
andre@0: 	uint64_t i;
andre@0: 
andre@0: 	/* pull out stored info */
andre@0: 	p = &st->P[1];
andre@0: 
andre@0: 	r0   = ((uint64_t)p->R20.d[3] << 32) | (uint64_t)p->R20.d[1];
andre@0: 	r1   = ((uint64_t)p->R21.d[3] << 32) | (uint64_t)p->R21.d[1];
andre@0: 	r2   = ((uint64_t)p->R22.d[3] << 32) | (uint64_t)p->R22.d[1];
andre@0: 	pad0 = ((uint64_t)p->R23.d[3] << 32) | (uint64_t)p->R23.d[1];
andre@0: 	pad1 = ((uint64_t)p->R24.d[3] << 32) | (uint64_t)p->R24.d[1];
andre@0: 
andre@0: 	/* compute powers r^2,r^4 */
andre@0: 	r20 = r0;
andre@0: 	r21 = r1;
andre@0: 	r22 = r2;
andre@0: 	for (i = 0; i < 2; i++) {
andre@0: 		s22 = r22 * (5 << 2);
andre@0: 
andre@0: 		d[0] = add128(mul64x64_128(r20, r20), mul64x64_128(r21 * 2, s22));
andre@0: 		d[1] = add128(mul64x64_128(r22, s22), mul64x64_128(r20 * 2, r21));
andre@0: 		d[2] = add128(mul64x64_128(r21, r21), mul64x64_128(r22 * 2, r20));
andre@0: 
andre@0: 		                           r20 = lo128(d[0]) & 0xfffffffffff; c = shr128(d[0], 44);
andre@0: 		d[1] = add128_64(d[1], c); r21 = lo128(d[1]) & 0xfffffffffff; c = shr128(d[1], 44);
andre@0: 		d[2] = add128_64(d[2], c); r22 = lo128(d[2]) & 0x3ffffffffff; c = shr128(d[2], 42);
andre@0: 		r20 += c * 5; c = (r20 >> 44); r20 = r20 & 0xfffffffffff;
andre@0: 		r21 += c;
andre@0: 
andre@0: 		p->R20.v = _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)( r20                     ) & 0x3ffffff), _MM_SHUFFLE(1,0,1,0));
andre@0: 		p->R21.v = _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)((r20 >> 26) | (r21 << 18)) & 0x3ffffff), _MM_SHUFFLE(1,0,1,0));
andre@0: 		p->R22.v = _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)((r21 >> 8)               ) & 0x3ffffff), _MM_SHUFFLE(1,0,1,0));
andre@0: 		p->R23.v = _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)((r21 >> 34) | (r22 << 10)) & 0x3ffffff), _MM_SHUFFLE(1,0,1,0));
andre@0: 		p->R24.v = _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)((r22 >> 16)              )            ), _MM_SHUFFLE(1,0,1,0));
andre@0: 		p->S21.v = _mm_mul_epu32(p->R21.v, FIVE);
andre@0: 		p->S22.v = _mm_mul_epu32(p->R22.v, FIVE);
andre@0: 		p->S23.v = _mm_mul_epu32(p->R23.v, FIVE);
andre@0: 		p->S24.v = _mm_mul_epu32(p->R24.v, FIVE);
andre@0: 		p--;
andre@0: 	}
andre@0: 
andre@0: 	/* put saved info back */
andre@0: 	p = &st->P[1];
andre@0: 	p->R20.d[1] = (uint32_t)(r0        );
andre@0: 	p->R20.d[3] = (uint32_t)(r0   >> 32);
andre@0: 	p->R21.d[1] = (uint32_t)(r1        );
andre@0: 	p->R21.d[3] = (uint32_t)(r1   >> 32);
andre@0: 	p->R22.d[1] = (uint32_t)(r2        );
andre@0: 	p->R22.d[3] = (uint32_t)(r2   >> 32);
andre@0: 	p->R23.d[1] = (uint32_t)(pad0      );
andre@0: 	p->R23.d[3] = (uint32_t)(pad0 >> 32);
andre@0: 	p->R24.d[1] = (uint32_t)(pad1      );
andre@0: 	p->R24.d[3] = (uint32_t)(pad1 >> 32);
andre@0: 
andre@0: 	/* H = [Mx,My] */
andre@0: 	T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 0)), _mm_loadl_epi64((xmmi *)(m + 16)));
andre@0: 	T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 8)), _mm_loadl_epi64((xmmi *)(m + 24)));
andre@0: 	st->H[0] = _mm_and_si128(MMASK, T5);
andre@0: 	st->H[1] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 	T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
andre@0: 	st->H[2] = _mm_and_si128(MMASK, T5);
andre@0: 	st->H[3] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 	st->H[4] = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
andre@0: }
andre@0: 
andre@0: static void
andre@0: poly1305_blocks(poly1305_state_internal *st, const uint8_t *m, size_t bytes) {
andre@0: 	const xmmi MMASK = _mm_load_si128((xmmi *)poly1305_x64_sse2_message_mask);
andre@0: 	const xmmi FIVE = _mm_load_si128((xmmi*)poly1305_x64_sse2_5);
andre@0: 	const xmmi HIBIT = _mm_load_si128((xmmi*)poly1305_x64_sse2_1shl128);
andre@0: 
andre@0: 	poly1305_power *p;
andre@0: 	xmmi H0,H1,H2,H3,H4;
andre@0: 	xmmi T0,T1,T2,T3,T4,T5,T6;
andre@0: 	xmmi M0,M1,M2,M3,M4;
andre@0: 	xmmi C1,C2;
andre@0: 
andre@0: 	H0 = st->H[0];
andre@0: 	H1 = st->H[1];
andre@0: 	H2 = st->H[2];
andre@0: 	H3 = st->H[3];
andre@0: 	H4 = st->H[4];
andre@0: 
andre@0: 	while (bytes >= 64) {
andre@0: 		/* H *= [r^4,r^4] */
andre@0: 		p = &st->P[0];
andre@0: 		T0 = _mm_mul_epu32(H0, p->R20.v);
andre@0: 		T1 = _mm_mul_epu32(H0, p->R21.v);
andre@0: 		T2 = _mm_mul_epu32(H0, p->R22.v);
andre@0: 		T3 = _mm_mul_epu32(H0, p->R23.v);
andre@0: 		T4 = _mm_mul_epu32(H0, p->R24.v);
andre@0: 		T5 = _mm_mul_epu32(H1, p->S24.v); T6 = _mm_mul_epu32(H1, p->R20.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H2, p->S23.v); T6 = _mm_mul_epu32(H2, p->S24.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H3, p->S22.v); T6 = _mm_mul_epu32(H3, p->S23.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H4, p->S21.v); T6 = _mm_mul_epu32(H4, p->S22.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H1, p->R21.v); T6 = _mm_mul_epu32(H1, p->R22.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H2, p->R20.v); T6 = _mm_mul_epu32(H2, p->R21.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H3, p->S24.v); T6 = _mm_mul_epu32(H3, p->R20.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H4, p->S23.v); T6 = _mm_mul_epu32(H4, p->S24.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H1, p->R23.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(H2, p->R22.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(H3, p->R21.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(H4, p->R20.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 
andre@0: 		/* H += [Mx,My]*[r^2,r^2] */
andre@0: 		T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 0)), _mm_loadl_epi64((xmmi *)(m + 16)));
andre@0: 		T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 8)), _mm_loadl_epi64((xmmi *)(m + 24)));
andre@0: 		M0 = _mm_and_si128(MMASK, T5);
andre@0: 		M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 		T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
andre@0: 		M2 = _mm_and_si128(MMASK, T5);
andre@0: 		M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 		M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
andre@0: 
andre@0: 		p = &st->P[1];
andre@0: 		T5 = _mm_mul_epu32(M0, p->R20.v); T6 = _mm_mul_epu32(M0, p->R21.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(M1, p->S24.v); T6 = _mm_mul_epu32(M1, p->R20.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(M2, p->S23.v); T6 = _mm_mul_epu32(M2, p->S24.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(M3, p->S22.v); T6 = _mm_mul_epu32(M3, p->S23.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(M4, p->S21.v); T6 = _mm_mul_epu32(M4, p->S22.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(M0, p->R22.v); T6 = _mm_mul_epu32(M0, p->R23.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(M1, p->R21.v); T6 = _mm_mul_epu32(M1, p->R22.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(M2, p->R20.v); T6 = _mm_mul_epu32(M2, p->R21.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(M3, p->S24.v); T6 = _mm_mul_epu32(M3, p->R20.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(M4, p->S23.v); T6 = _mm_mul_epu32(M4, p->S24.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(M0, p->R24.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(M1, p->R23.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(M2, p->R22.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(M3, p->R21.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(M4, p->R20.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 
andre@0: 		/* H += [Mx,My] */
andre@0: 		T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 32)), _mm_loadl_epi64((xmmi *)(m + 48)));
andre@0: 		T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 40)), _mm_loadl_epi64((xmmi *)(m + 56)));
andre@0: 		M0 = _mm_and_si128(MMASK, T5);
andre@0: 		M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 		T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
andre@0: 		M2 = _mm_and_si128(MMASK, T5);
andre@0: 		M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 		M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
andre@0: 
andre@0: 		T0 = _mm_add_epi64(T0, M0);
andre@0: 		T1 = _mm_add_epi64(T1, M1);
andre@0: 		T2 = _mm_add_epi64(T2, M2);
andre@0: 		T3 = _mm_add_epi64(T3, M3);
andre@0: 		T4 = _mm_add_epi64(T4, M4);
andre@0: 
andre@0: 		/* reduce */
andre@0: 		C1 = _mm_srli_epi64(T0, 26); C2 = _mm_srli_epi64(T3, 26); T0 = _mm_and_si128(T0, MMASK); T3 = _mm_and_si128(T3, MMASK); T1 = _mm_add_epi64(T1, C1); T4 = _mm_add_epi64(T4, C2);
andre@0: 		C1 = _mm_srli_epi64(T1, 26); C2 = _mm_srli_epi64(T4, 26); T1 = _mm_and_si128(T1, MMASK); T4 = _mm_and_si128(T4, MMASK); T2 = _mm_add_epi64(T2, C1); T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
andre@0: 		C1 = _mm_srli_epi64(T2, 26); C2 = _mm_srli_epi64(T0, 26); T2 = _mm_and_si128(T2, MMASK); T0 = _mm_and_si128(T0, MMASK); T3 = _mm_add_epi64(T3, C1); T1 = _mm_add_epi64(T1, C2);
andre@0: 		C1 = _mm_srli_epi64(T3, 26);                              T3 = _mm_and_si128(T3, MMASK);                                T4 = _mm_add_epi64(T4, C1);
andre@0: 
andre@0: 		/* H = (H*[r^4,r^4] + [Mx,My]*[r^2,r^2] + [Mx,My]) */
andre@0: 		H0 = T0;
andre@0: 		H1 = T1;
andre@0: 		H2 = T2;
andre@0: 		H3 = T3;
andre@0: 		H4 = T4;
andre@0: 
andre@0: 		m += 64;
andre@0: 		bytes -= 64;
andre@0: 	}
andre@0: 
andre@0: 	st->H[0] = H0;
andre@0: 	st->H[1] = H1;
andre@0: 	st->H[2] = H2;
andre@0: 	st->H[3] = H3;
andre@0: 	st->H[4] = H4;
andre@0: }
andre@0: 
andre@0: static size_t
andre@0: poly1305_combine(poly1305_state_internal *st, const uint8_t *m, size_t bytes) {
andre@0: 	const xmmi MMASK = _mm_load_si128((xmmi *)poly1305_x64_sse2_message_mask);
andre@0: 	const xmmi HIBIT = _mm_load_si128((xmmi*)poly1305_x64_sse2_1shl128);
andre@0: 	const xmmi FIVE = _mm_load_si128((xmmi*)poly1305_x64_sse2_5);
andre@0: 
andre@0: 	poly1305_power *p;
andre@0: 	xmmi H0,H1,H2,H3,H4;
andre@0: 	xmmi M0,M1,M2,M3,M4;
andre@0: 	xmmi T0,T1,T2,T3,T4,T5,T6;
andre@0: 	xmmi C1,C2;
andre@0: 
andre@0: 	uint64_t r0,r1,r2;
andre@0: 	uint64_t t0,t1,t2,t3,t4;
andre@0: 	uint64_t c;
andre@0: 	size_t consumed = 0;
andre@0: 
andre@0: 	H0 = st->H[0];
andre@0: 	H1 = st->H[1];
andre@0: 	H2 = st->H[2];
andre@0: 	H3 = st->H[3];
andre@0: 	H4 = st->H[4];
andre@0: 
andre@0: 	/* p = [r^2,r^2] */
andre@0: 	p = &st->P[1];
andre@0: 
andre@0: 	if (bytes >= 32) {
andre@0: 		/* H *= [r^2,r^2] */
andre@0: 		T0 = _mm_mul_epu32(H0, p->R20.v);
andre@0: 		T1 = _mm_mul_epu32(H0, p->R21.v);
andre@0: 		T2 = _mm_mul_epu32(H0, p->R22.v);
andre@0: 		T3 = _mm_mul_epu32(H0, p->R23.v);
andre@0: 		T4 = _mm_mul_epu32(H0, p->R24.v);
andre@0: 		T5 = _mm_mul_epu32(H1, p->S24.v); T6 = _mm_mul_epu32(H1, p->R20.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H2, p->S23.v); T6 = _mm_mul_epu32(H2, p->S24.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H3, p->S22.v); T6 = _mm_mul_epu32(H3, p->S23.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H4, p->S21.v); T6 = _mm_mul_epu32(H4, p->S22.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 		T5 = _mm_mul_epu32(H1, p->R21.v); T6 = _mm_mul_epu32(H1, p->R22.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H2, p->R20.v); T6 = _mm_mul_epu32(H2, p->R21.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H3, p->S24.v); T6 = _mm_mul_epu32(H3, p->R20.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H4, p->S23.v); T6 = _mm_mul_epu32(H4, p->S24.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 		T5 = _mm_mul_epu32(H1, p->R23.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(H2, p->R22.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(H3, p->R21.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 		T5 = _mm_mul_epu32(H4, p->R20.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 
andre@0: 		/* H += [Mx,My] */
andre@0: 		T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 0)), _mm_loadl_epi64((xmmi *)(m + 16)));
andre@0: 		T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((xmmi *)(m + 8)), _mm_loadl_epi64((xmmi *)(m + 24)));
andre@0: 		M0 = _mm_and_si128(MMASK, T5);
andre@0: 		M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 		T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
andre@0: 		M2 = _mm_and_si128(MMASK, T5);
andre@0: 		M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
andre@0: 		M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
andre@0: 
andre@0: 		T0 = _mm_add_epi64(T0, M0);
andre@0: 		T1 = _mm_add_epi64(T1, M1);
andre@0: 		T2 = _mm_add_epi64(T2, M2);
andre@0: 		T3 = _mm_add_epi64(T3, M3);
andre@0: 		T4 = _mm_add_epi64(T4, M4);
andre@0: 
andre@0: 		/* reduce */
andre@0: 		C1 = _mm_srli_epi64(T0, 26); C2 = _mm_srli_epi64(T3, 26); T0 = _mm_and_si128(T0, MMASK); T3 = _mm_and_si128(T3, MMASK); T1 = _mm_add_epi64(T1, C1); T4 = _mm_add_epi64(T4, C2);
andre@0: 		C1 = _mm_srli_epi64(T1, 26); C2 = _mm_srli_epi64(T4, 26); T1 = _mm_and_si128(T1, MMASK); T4 = _mm_and_si128(T4, MMASK); T2 = _mm_add_epi64(T2, C1); T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
andre@0: 		C1 = _mm_srli_epi64(T2, 26); C2 = _mm_srli_epi64(T0, 26); T2 = _mm_and_si128(T2, MMASK); T0 = _mm_and_si128(T0, MMASK); T3 = _mm_add_epi64(T3, C1); T1 = _mm_add_epi64(T1, C2);
andre@0: 		C1 = _mm_srli_epi64(T3, 26);                              T3 = _mm_and_si128(T3, MMASK);                                T4 = _mm_add_epi64(T4, C1);
andre@0: 
andre@0: 		/* H = (H*[r^2,r^2] + [Mx,My]) */
andre@0: 		H0 = T0;
andre@0: 		H1 = T1;
andre@0: 		H2 = T2;
andre@0: 		H3 = T3;
andre@0: 		H4 = T4;
andre@0: 
andre@0: 		consumed = 32;
andre@0: 	}
andre@0: 
andre@0: 	/* finalize, H *= [r^2,r] */
andre@0: 	r0 = ((uint64_t)p->R20.d[3] << 32) | (uint64_t)p->R20.d[1];
andre@0: 	r1 = ((uint64_t)p->R21.d[3] << 32) | (uint64_t)p->R21.d[1];
andre@0: 	r2 = ((uint64_t)p->R22.d[3] << 32) | (uint64_t)p->R22.d[1];
andre@0: 
andre@0: 	p->R20.d[2] = (uint32_t)( r0                    ) & 0x3ffffff;
andre@0: 	p->R21.d[2] = (uint32_t)((r0 >> 26) | (r1 << 18)) & 0x3ffffff;
andre@0: 	p->R22.d[2] = (uint32_t)((r1 >> 8)              ) & 0x3ffffff;
andre@0: 	p->R23.d[2] = (uint32_t)((r1 >> 34) | (r2 << 10)) & 0x3ffffff;
andre@0: 	p->R24.d[2] = (uint32_t)((r2 >> 16)             )            ;
andre@0: 	p->S21.d[2] = p->R21.d[2] * 5;
andre@0: 	p->S22.d[2] = p->R22.d[2] * 5;
andre@0: 	p->S23.d[2] = p->R23.d[2] * 5;
andre@0: 	p->S24.d[2] = p->R24.d[2] * 5;
andre@0: 
andre@0: 	/* H *= [r^2,r] */
andre@0: 	T0 = _mm_mul_epu32(H0, p->R20.v);
andre@0: 	T1 = _mm_mul_epu32(H0, p->R21.v);
andre@0: 	T2 = _mm_mul_epu32(H0, p->R22.v);
andre@0: 	T3 = _mm_mul_epu32(H0, p->R23.v);
andre@0: 	T4 = _mm_mul_epu32(H0, p->R24.v);
andre@0: 	T5 = _mm_mul_epu32(H1, p->S24.v); T6 = _mm_mul_epu32(H1, p->R20.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 	T5 = _mm_mul_epu32(H2, p->S23.v); T6 = _mm_mul_epu32(H2, p->S24.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 	T5 = _mm_mul_epu32(H3, p->S22.v); T6 = _mm_mul_epu32(H3, p->S23.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 	T5 = _mm_mul_epu32(H4, p->S21.v); T6 = _mm_mul_epu32(H4, p->S22.v); T0 = _mm_add_epi64(T0, T5); T1 = _mm_add_epi64(T1, T6);
andre@0: 	T5 = _mm_mul_epu32(H1, p->R21.v); T6 = _mm_mul_epu32(H1, p->R22.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 	T5 = _mm_mul_epu32(H2, p->R20.v); T6 = _mm_mul_epu32(H2, p->R21.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 	T5 = _mm_mul_epu32(H3, p->S24.v); T6 = _mm_mul_epu32(H3, p->R20.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 	T5 = _mm_mul_epu32(H4, p->S23.v); T6 = _mm_mul_epu32(H4, p->S24.v); T2 = _mm_add_epi64(T2, T5); T3 = _mm_add_epi64(T3, T6);
andre@0: 	T5 = _mm_mul_epu32(H1, p->R23.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 	T5 = _mm_mul_epu32(H2, p->R22.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 	T5 = _mm_mul_epu32(H3, p->R21.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 	T5 = _mm_mul_epu32(H4, p->R20.v);                                   T4 = _mm_add_epi64(T4, T5);
andre@0: 
andre@0: 	C1 = _mm_srli_epi64(T0, 26); C2 = _mm_srli_epi64(T3, 26); T0 = _mm_and_si128(T0, MMASK); T3 = _mm_and_si128(T3, MMASK); T1 = _mm_add_epi64(T1, C1); T4 = _mm_add_epi64(T4, C2);
andre@0: 	C1 = _mm_srli_epi64(T1, 26); C2 = _mm_srli_epi64(T4, 26); T1 = _mm_and_si128(T1, MMASK); T4 = _mm_and_si128(T4, MMASK); T2 = _mm_add_epi64(T2, C1); T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
andre@0: 	C1 = _mm_srli_epi64(T2, 26); C2 = _mm_srli_epi64(T0, 26); T2 = _mm_and_si128(T2, MMASK); T0 = _mm_and_si128(T0, MMASK); T3 = _mm_add_epi64(T3, C1); T1 = _mm_add_epi64(T1, C2);
andre@0: 	C1 = _mm_srli_epi64(T3, 26);                              T3 = _mm_and_si128(T3, MMASK);                                T4 = _mm_add_epi64(T4, C1);
andre@0: 
andre@0: 	/* H = H[0]+H[1] */
andre@0: 	H0 = _mm_add_epi64(T0, _mm_srli_si128(T0, 8));
andre@0: 	H1 = _mm_add_epi64(T1, _mm_srli_si128(T1, 8));
andre@0: 	H2 = _mm_add_epi64(T2, _mm_srli_si128(T2, 8));
andre@0: 	H3 = _mm_add_epi64(T3, _mm_srli_si128(T3, 8));
andre@0: 	H4 = _mm_add_epi64(T4, _mm_srli_si128(T4, 8));
andre@0: 
andre@0: 	t0 = _mm_cvtsi128_si32(H0)    ; c = (t0 >> 26); t0 &= 0x3ffffff;
andre@0: 	t1 = _mm_cvtsi128_si32(H1) + c; c = (t1 >> 26); t1 &= 0x3ffffff;
andre@0: 	t2 = _mm_cvtsi128_si32(H2) + c; c = (t2 >> 26); t2 &= 0x3ffffff;
andre@0: 	t3 = _mm_cvtsi128_si32(H3) + c; c = (t3 >> 26); t3 &= 0x3ffffff;
andre@0: 	t4 = _mm_cvtsi128_si32(H4) + c; c = (t4 >> 26); t4 &= 0x3ffffff;
andre@0: 	t0 =              t0 + (c * 5); c = (t0 >> 26); t0 &= 0x3ffffff;
andre@0: 	t1 =              t1 + c;
andre@0: 
andre@0: 	st->HH[0] =  ((t0      ) | (t1 << 26)             ) & 0xfffffffffffull;
andre@0: 	st->HH[1] =  ((t1 >> 18) | (t2 <<  8) | (t3 << 34)) & 0xfffffffffffull;
andre@0: 	st->HH[2] =  ((t3 >> 10) | (t4 << 16)             ) & 0x3ffffffffffull;
andre@0: 
andre@0: 	return consumed;
andre@0: }
andre@0: 
andre@0: void
andre@0: Poly1305Update(poly1305_state *state, const unsigned char *m, size_t bytes) {
andre@0: 	poly1305_state_internal *st = poly1305_aligned_state(state);
andre@0: 	size_t want;
andre@0: 
andre@0: 	/* need at least 32 initial bytes to start the accelerated branch */
andre@0: 	if (!st->started) {
andre@0: 		if ((st->leftover == 0) && (bytes > 32)) {
andre@0: 			poly1305_first_block(st, m);
andre@0: 			m += 32;
andre@0: 			bytes -= 32;
andre@0: 		} else {
andre@0: 			want = poly1305_min(32 - st->leftover, bytes);
andre@0: 			poly1305_block_copy(st->buffer + st->leftover, m, want);
andre@0: 			bytes -= want;
andre@0: 			m += want;
andre@0: 			st->leftover += want;
andre@0: 			if ((st->leftover < 32) || (bytes == 0))
andre@0: 				return;
andre@0: 			poly1305_first_block(st, st->buffer);
andre@0: 			st->leftover = 0;
andre@0: 		}
andre@0: 		st->started = 1;
andre@0: 	}
andre@0: 
andre@0: 	/* handle leftover */
andre@0: 	if (st->leftover) {
andre@0: 		want = poly1305_min(64 - st->leftover, bytes);
andre@0: 		poly1305_block_copy(st->buffer + st->leftover, m, want);
andre@0: 		bytes -= want;
andre@0: 		m += want;
andre@0: 		st->leftover += want;
andre@0: 		if (st->leftover < 64)
andre@0: 			return;
andre@0: 		poly1305_blocks(st, st->buffer, 64);
andre@0: 		st->leftover = 0;
andre@0: 	}
andre@0: 
andre@0: 	/* process 64 byte blocks */
andre@0: 	if (bytes >= 64) {
andre@0: 		want = (bytes & ~63);
andre@0: 		poly1305_blocks(st, m, want);
andre@0: 		m += want;
andre@0: 		bytes -= want;
andre@0: 	}
andre@0: 
andre@0: 	if (bytes) {
andre@0: 		poly1305_block_copy(st->buffer + st->leftover, m, bytes);
andre@0: 		st->leftover += bytes;
andre@0: 	}
andre@0: }
andre@0: 
andre@0: void
andre@0: Poly1305Finish(poly1305_state *state, unsigned char mac[16]) {
andre@0: 	poly1305_state_internal *st = poly1305_aligned_state(state);
andre@0: 	size_t leftover = st->leftover;
andre@0: 	uint8_t *m = st->buffer;
andre@0: 	uint128_t d[3];
andre@0: 	uint64_t h0,h1,h2;
andre@0: 	uint64_t t0,t1;
andre@0: 	uint64_t g0,g1,g2,c,nc;
andre@0: 	uint64_t r0,r1,r2,s1,s2;
andre@0: 	poly1305_power *p;
andre@0: 
andre@0: 	if (st->started) {
andre@0: 		size_t consumed = poly1305_combine(st, m, leftover);
andre@0: 		leftover -= consumed;
andre@0: 		m += consumed;
andre@0: 	}
andre@0: 
andre@0: 	/* st->HH will either be 0 or have the combined result */
andre@0: 	h0 = st->HH[0];
andre@0: 	h1 = st->HH[1];
andre@0: 	h2 = st->HH[2];
andre@0: 
andre@0: 	p = &st->P[1];
andre@0: 	r0 = ((uint64_t)p->R20.d[3] << 32) | (uint64_t)p->R20.d[1];
andre@0: 	r1 = ((uint64_t)p->R21.d[3] << 32) | (uint64_t)p->R21.d[1];
andre@0: 	r2 = ((uint64_t)p->R22.d[3] << 32) | (uint64_t)p->R22.d[1];
andre@0: 	s1 = r1 * (5 << 2);
andre@0: 	s2 = r2 * (5 << 2);
andre@0: 
andre@0: 	if (leftover < 16)
andre@0: 		goto poly1305_donna_atmost15bytes;
andre@0: 
andre@0: poly1305_donna_atleast16bytes:
andre@0: 	t0 = U8TO64_LE(m + 0);
andre@0: 	t1 = U8TO64_LE(m + 8);
andre@0: 	h0 += t0 & 0xfffffffffff;
andre@0: 	t0 = shr128_pair(t1, t0, 44);
andre@0: 	h1 += t0 & 0xfffffffffff;
andre@0: 	h2 += (t1 >> 24) | ((uint64_t)1 << 40);
andre@0: 
andre@0: poly1305_donna_mul:
andre@0: 	d[0] = add128(add128(mul64x64_128(h0, r0), mul64x64_128(h1, s2)), mul64x64_128(h2, s1));
andre@0: 	d[1] = add128(add128(mul64x64_128(h0, r1), mul64x64_128(h1, r0)), mul64x64_128(h2, s2));
andre@0: 	d[2] = add128(add128(mul64x64_128(h0, r2), mul64x64_128(h1, r1)), mul64x64_128(h2, r0));
andre@0: 	                           h0 = lo128(d[0]) & 0xfffffffffff; c = shr128(d[0], 44);
andre@0: 	d[1] = add128_64(d[1], c); h1 = lo128(d[1]) & 0xfffffffffff; c = shr128(d[1], 44);
andre@0: 	d[2] = add128_64(d[2], c); h2 = lo128(d[2]) & 0x3ffffffffff; c = shr128(d[2], 42);
andre@0: 	h0   += c * 5;
andre@0: 
andre@0: 	m += 16;
andre@0: 	leftover -= 16;
andre@0: 	if (leftover >= 16) goto poly1305_donna_atleast16bytes;
andre@0: 
andre@0: 	/* final bytes */
andre@0: poly1305_donna_atmost15bytes:
andre@0: 	if (!leftover) goto poly1305_donna_finish;
andre@0: 
andre@0: 	m[leftover++] = 1;
andre@0: 	poly1305_block_zero(m + leftover, 16 - leftover);
andre@0: 	leftover = 16;
andre@0: 
andre@0: 	t0 = U8TO64_LE(m+0);
andre@0: 	t1 = U8TO64_LE(m+8);
andre@0: 	h0 += t0 & 0xfffffffffff; t0 = shr128_pair(t1, t0, 44);
andre@0: 	h1 += t0 & 0xfffffffffff;
andre@0: 	h2 += (t1 >> 24);
andre@0: 
andre@0: 	goto poly1305_donna_mul;
andre@0: 
andre@0: poly1305_donna_finish:
andre@0: 	             c = (h0 >> 44); h0 &= 0xfffffffffff;
andre@0: 	h1 += c;     c = (h1 >> 44); h1 &= 0xfffffffffff;
andre@0: 	h2 += c;     c = (h2 >> 42); h2 &= 0x3ffffffffff;
andre@0: 	h0 += c * 5;
andre@0: 
andre@0: 	g0 = h0 + 5; c = (g0 >> 44); g0 &= 0xfffffffffff;
andre@0: 	g1 = h1 + c; c = (g1 >> 44); g1 &= 0xfffffffffff;
andre@0: 	g2 = h2 + c - ((uint64_t)1 << 42);
andre@0: 
andre@0: 	c = (g2 >> 63) - 1;
andre@0: 	nc = ~c;
andre@0: 	h0 = (h0 & nc) | (g0 & c);
andre@0: 	h1 = (h1 & nc) | (g1 & c);
andre@0: 	h2 = (h2 & nc) | (g2 & c);
andre@0: 
andre@0: 	/* pad */
andre@0: 	t0 = ((uint64_t)p->R23.d[3] << 32) | (uint64_t)p->R23.d[1];
andre@0: 	t1 = ((uint64_t)p->R24.d[3] << 32) | (uint64_t)p->R24.d[1];
andre@0: 	h0 += (t0 & 0xfffffffffff)    ; c = (h0 >> 44); h0 &= 0xfffffffffff; t0 = shr128_pair(t1, t0, 44);
andre@0: 	h1 += (t0 & 0xfffffffffff) + c; c = (h1 >> 44); h1 &= 0xfffffffffff; t1 = (t1 >> 24);
andre@0: 	h2 += (t1                ) + c;
andre@0: 
andre@0: 	U64TO8_LE(mac + 0, ((h0      ) | (h1 << 44)));
andre@0: 	U64TO8_LE(mac + 8, ((h1 >> 20) | (h2 << 24)));
andre@0: }