src/third_party/blake3/*.[hcS]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-This is a subset of https://github.com/BLAKE3-team/BLAKE3[BLAKE3] 0.3.6 with
+This is a subset of https://github.com/BLAKE3-team/BLAKE3[BLAKE3] 0.3.7 with
the following license:
-------------------------------------------------------------------------------
endif()
endfunction()
+add_source_if_enabled(sse2 "-msse2")
add_source_if_enabled(sse41 "-msse4.1")
add_source_if_enabled(avx2 "-mavx2")
add_source_if_enabled(avx512 "-mavx512f -mavx512vl")
memcpy(cv_words, self->input_cv, 32);
blake3_compress_in_place(cv_words, self->block, self->block_len,
self->counter, self->flags);
- memcpy(cv, cv_words, 32);
+ store_cv_words(cv, cv_words);
}
INLINE void output_root_bytes(const output_t *self, uint64_t seek, uint8_t *out,
hasher_init_base(self, key_words, KEYED_HASH);
}
-void blake3_hasher_init_derive_key(blake3_hasher *self, const char *context) {
+void blake3_hasher_init_derive_key_raw(blake3_hasher *self, const void *context,
+ size_t context_len) {
blake3_hasher context_hasher;
hasher_init_base(&context_hasher, IV, DERIVE_KEY_CONTEXT);
- blake3_hasher_update(&context_hasher, context, strlen(context));
+ blake3_hasher_update(&context_hasher, context, context_len);
uint8_t context_key[BLAKE3_KEY_LEN];
blake3_hasher_finalize(&context_hasher, context_key, BLAKE3_KEY_LEN);
uint32_t context_key_words[8];
hasher_init_base(self, context_key_words, DERIVE_KEY_MATERIAL);
}
+void blake3_hasher_init_derive_key(blake3_hasher *self, const char *context) {
+ blake3_hasher_init_derive_key_raw(self, context, strlen(context));
+}
+
// As described in hasher_push_cv() below, we do "lazy merging", delaying
// merges until right before the next CV is about to be added. This is
// different from the reference implementation. Another difference is that we
void blake3_hasher_init_keyed(blake3_hasher *self,
const uint8_t key[BLAKE3_KEY_LEN]);
void blake3_hasher_init_derive_key(blake3_hasher *self, const char *context);
+void blake3_hasher_init_derive_key_raw(blake3_hasher *self, const void *context,
+ size_t context_len);
void blake3_hasher_update(blake3_hasher *self, const void *input,
size_t input_len);
void blake3_hasher_finalize(const blake3_hasher *self, uint8_t *out,
+#if defined(__ELF__) && defined(__linux__)
+.section .note.GNU-stack,"",%progbits
+#endif
+
#if defined(__ELF__) && defined(__CET__) && defined(__has_include)
#if __has_include(<cet.h>)
#include <cet.h>
+#if defined(__ELF__) && defined(__linux__)
+.section .note.GNU-stack,"",%progbits
+#endif
+
#if defined(__ELF__) && defined(__CET__) && defined(__has_include)
#if __has_include(<cet.h>)
#include <cet.h>
return;
}
#endif
+#if !defined(BLAKE3_NO_SSE2)
+ if (features & SSE2) {
+ blake3_compress_in_place_sse2(cv, block, block_len, counter, flags);
+ return;
+ }
+#endif
#endif
blake3_compress_in_place_portable(cv, block, block_len, counter, flags);
}
return;
}
#endif
+#if !defined(BLAKE3_NO_SSE2)
+ if (features & SSE2) {
+ blake3_compress_xof_sse2(cv, block, block_len, counter, flags, out);
+ return;
+ }
+#endif
#endif
blake3_compress_xof_portable(cv, block, block_len, counter, flags, out);
}
return;
}
#endif
+#if !defined(BLAKE3_NO_SSE2)
+ if (features & SSE2) {
+ blake3_hash_many_sse2(inputs, num_inputs, blocks, key, counter,
+ increment_counter, flags, flags_start, flags_end,
+ out);
+ return;
+ }
+#endif
#endif
#if defined(BLAKE3_USE_NEON)
return 4;
}
#endif
+#if !defined(BLAKE3_NO_SSE2)
+ if (features & SSE2) {
+ return 4;
+ }
+#endif
#endif
#if defined(BLAKE3_USE_NEON)
return 4;
key_words[7] = load32(&key[7 * 4]);
}
+INLINE void store32(void *dst, uint32_t w) {
+ uint8_t *p = (uint8_t *)dst;
+ p[0] = (uint8_t)(w >> 0);
+ p[1] = (uint8_t)(w >> 8);
+ p[2] = (uint8_t)(w >> 16);
+ p[3] = (uint8_t)(w >> 24);
+}
+
+INLINE void store_cv_words(uint8_t bytes_out[32], uint32_t cv_words[8]) {
+ store32(&bytes_out[0 * 4], cv_words[0]);
+ store32(&bytes_out[1 * 4], cv_words[1]);
+ store32(&bytes_out[2 * 4], cv_words[2]);
+ store32(&bytes_out[3 * 4], cv_words[3]);
+ store32(&bytes_out[4 * 4], cv_words[4]);
+ store32(&bytes_out[5 * 4], cv_words[5]);
+ store32(&bytes_out[6 * 4], cv_words[6]);
+ store32(&bytes_out[7 * 4], cv_words[7]);
+}
+
void blake3_compress_in_place(uint32_t cv[8],
const uint8_t block[BLAKE3_BLOCK_LEN],
uint8_t block_len, uint64_t counter,
uint8_t flags_end, uint8_t *out);
#if defined(IS_X86)
+#if !defined(BLAKE3_NO_SSE2)
+void blake3_compress_in_place_sse2(uint32_t cv[8],
+ const uint8_t block[BLAKE3_BLOCK_LEN],
+ uint8_t block_len, uint64_t counter,
+ uint8_t flags);
+void blake3_compress_xof_sse2(const uint32_t cv[8],
+ const uint8_t block[BLAKE3_BLOCK_LEN],
+ uint8_t block_len, uint64_t counter,
+ uint8_t flags, uint8_t out[64]);
+void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs,
+ size_t blocks, const uint32_t key[8],
+ uint64_t counter, bool increment_counter,
+ uint8_t flags, uint8_t flags_start,
+ uint8_t flags_end, uint8_t *out);
+#endif
#if !defined(BLAKE3_NO_SSE41)
void blake3_compress_in_place_sse41(uint32_t cv[8],
const uint8_t block[BLAKE3_BLOCK_LEN],
#include "blake3_impl.h"
#include <string.h>
-INLINE void store32(void *dst, uint32_t w) {
- uint8_t *p = (uint8_t *)dst;
- p[0] = (uint8_t)(w >> 0);
- p[1] = (uint8_t)(w >> 8);
- p[2] = (uint8_t)(w >> 16);
- p[3] = (uint8_t)(w >> 24);
-}
-
INLINE uint32_t rotr32(uint32_t w, uint32_t c) {
return (w >> c) | (w << (32 - c));
}
blocks -= 1;
block_flags = flags;
}
- memcpy(out, cv, 32);
+ store_cv_words(out, cv);
}
void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs,
--- /dev/null
+#include "blake3_impl.h"
+
+#include <immintrin.h>
+
+#define DEGREE 4
+
+#define _mm_shuffle_ps2(a, b, c) \
+ (_mm_castps_si128( \
+ _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
+
+INLINE __m128i loadu(const uint8_t src[16]) {
+ return _mm_loadu_si128((const __m128i *)src);
+}
+
+INLINE void storeu(__m128i src, uint8_t dest[16]) {
+ _mm_storeu_si128((__m128i *)dest, src);
+}
+
+INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); }
+
+// Note that clang-format doesn't like the name "xor" for some reason.
+INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
+
+INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
+
+INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+ return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
+}
+
+INLINE __m128i rot16(__m128i x) {
+ return _mm_shufflehi_epi16(_mm_shufflelo_epi16(x, 0xB1), 0xB1);
+}
+
+INLINE __m128i rot12(__m128i x) {
+ return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
+}
+
+INLINE __m128i rot8(__m128i x) {
+ return xorv(_mm_srli_epi32(x, 8), _mm_slli_epi32(x, 32 - 8));
+}
+
+INLINE __m128i rot7(__m128i x) {
+ return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
+}
+
+INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+ __m128i m) {
+ *row0 = addv(addv(*row0, m), *row1);
+ *row3 = xorv(*row3, *row0);
+ *row3 = rot16(*row3);
+ *row2 = addv(*row2, *row3);
+ *row1 = xorv(*row1, *row2);
+ *row1 = rot12(*row1);
+}
+
+INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+ __m128i m) {
+ *row0 = addv(addv(*row0, m), *row1);
+ *row3 = xorv(*row3, *row0);
+ *row3 = rot8(*row3);
+ *row2 = addv(*row2, *row3);
+ *row1 = xorv(*row1, *row2);
+ *row1 = rot7(*row1);
+}
+
+// Note the optimization here of leaving row1 as the unrotated row, rather than
+// row0. All the message loads below are adjusted to compensate for this. See
+// discussion at https://github.com/sneves/blake2-avx2/pull/4
+INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+ *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
+ *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+ *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
+}
+
+INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+ *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
+ *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+ *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
+}
+
+INLINE __m128i blend_epi16(__m128i a, __m128i b, const int imm8) {
+ const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
+ __m128i mask = _mm_set1_epi16(imm8);
+ mask = _mm_and_si128(mask, bits);
+ mask = _mm_cmpeq_epi16(mask, bits);
+ return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
+}
+
+INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8],
+ const uint8_t block[BLAKE3_BLOCK_LEN],
+ uint8_t block_len, uint64_t counter, uint8_t flags) {
+ rows[0] = loadu((uint8_t *)&cv[0]);
+ rows[1] = loadu((uint8_t *)&cv[4]);
+ rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
+ rows[3] = set4(counter_low(counter), counter_high(counter),
+ (uint32_t)block_len, (uint32_t)flags);
+
+ __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
+ __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
+ __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
+ __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
+
+ __m128i t0, t1, t2, t3, tt;
+
+ // Round 1. The first round permutes the message words from the original
+ // input order, into the groups that get mixed in parallel.
+ t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); // 6 4 2 0
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+ t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 3, 1)); // 7 5 3 1
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+ diagonalize(&rows[0], &rows[2], &rows[3]);
+ t2 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10 8
+ t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3)); // 12 10 8 14
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+ t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11 9
+ t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3)); // 13 11 9 15
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+ undiagonalize(&rows[0], &rows[2], &rows[3]);
+ m0 = t0;
+ m1 = t1;
+ m2 = t2;
+ m3 = t3;
+
+ // Round 2. This round and all following rounds apply a fixed permutation
+ // to the message words from the round before.
+ t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+ t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+ t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+ tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+ t1 = blend_epi16(tt, t1, 0xCC);
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+ diagonalize(&rows[0], &rows[2], &rows[3]);
+ t2 = _mm_unpacklo_epi64(m3, m1);
+ tt = blend_epi16(t2, m2, 0xC0);
+ t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+ t3 = _mm_unpackhi_epi32(m1, m3);
+ tt = _mm_unpacklo_epi32(m2, t3);
+ t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+ undiagonalize(&rows[0], &rows[2], &rows[3]);
+ m0 = t0;
+ m1 = t1;
+ m2 = t2;
+ m3 = t3;
+
+ // Round 3
+ t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+ t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+ t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+ tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+ t1 = blend_epi16(tt, t1, 0xCC);
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+ diagonalize(&rows[0], &rows[2], &rows[3]);
+ t2 = _mm_unpacklo_epi64(m3, m1);
+ tt = blend_epi16(t2, m2, 0xC0);
+ t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+ t3 = _mm_unpackhi_epi32(m1, m3);
+ tt = _mm_unpacklo_epi32(m2, t3);
+ t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+ undiagonalize(&rows[0], &rows[2], &rows[3]);
+ m0 = t0;
+ m1 = t1;
+ m2 = t2;
+ m3 = t3;
+
+ // Round 4
+ t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+ t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+ t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+ tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+ t1 = blend_epi16(tt, t1, 0xCC);
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+ diagonalize(&rows[0], &rows[2], &rows[3]);
+ t2 = _mm_unpacklo_epi64(m3, m1);
+ tt = blend_epi16(t2, m2, 0xC0);
+ t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+ t3 = _mm_unpackhi_epi32(m1, m3);
+ tt = _mm_unpacklo_epi32(m2, t3);
+ t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+ undiagonalize(&rows[0], &rows[2], &rows[3]);
+ m0 = t0;
+ m1 = t1;
+ m2 = t2;
+ m3 = t3;
+
+ // Round 5
+ t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+ t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+ t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+ tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+ t1 = blend_epi16(tt, t1, 0xCC);
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+ diagonalize(&rows[0], &rows[2], &rows[3]);
+ t2 = _mm_unpacklo_epi64(m3, m1);
+ tt = blend_epi16(t2, m2, 0xC0);
+ t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+ t3 = _mm_unpackhi_epi32(m1, m3);
+ tt = _mm_unpacklo_epi32(m2, t3);
+ t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+ undiagonalize(&rows[0], &rows[2], &rows[3]);
+ m0 = t0;
+ m1 = t1;
+ m2 = t2;
+ m3 = t3;
+
+ // Round 6
+ t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+ t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+ t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+ tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+ t1 = blend_epi16(tt, t1, 0xCC);
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+ diagonalize(&rows[0], &rows[2], &rows[3]);
+ t2 = _mm_unpacklo_epi64(m3, m1);
+ tt = blend_epi16(t2, m2, 0xC0);
+ t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+ t3 = _mm_unpackhi_epi32(m1, m3);
+ tt = _mm_unpacklo_epi32(m2, t3);
+ t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+ undiagonalize(&rows[0], &rows[2], &rows[3]);
+ m0 = t0;
+ m1 = t1;
+ m2 = t2;
+ m3 = t3;
+
+ // Round 7
+ t0 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+ t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+ t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+ tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+ t1 = blend_epi16(tt, t1, 0xCC);
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+ diagonalize(&rows[0], &rows[2], &rows[3]);
+ t2 = _mm_unpacklo_epi64(m3, m1);
+ tt = blend_epi16(t2, m2, 0xC0);
+ t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+ g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+ t3 = _mm_unpackhi_epi32(m1, m3);
+ tt = _mm_unpacklo_epi32(m2, t3);
+ t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
+ g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+ undiagonalize(&rows[0], &rows[2], &rows[3]);
+}
+
+void blake3_compress_in_place_sse2(uint32_t cv[8],
+ const uint8_t block[BLAKE3_BLOCK_LEN],
+ uint8_t block_len, uint64_t counter,
+ uint8_t flags) {
+ __m128i rows[4];
+ compress_pre(rows, cv, block, block_len, counter, flags);
+ storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
+ storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
+}
+
+void blake3_compress_xof_sse2(const uint32_t cv[8],
+ const uint8_t block[BLAKE3_BLOCK_LEN],
+ uint8_t block_len, uint64_t counter,
+ uint8_t flags, uint8_t out[64]) {
+ __m128i rows[4];
+ compress_pre(rows, cv, block, block_len, counter, flags);
+ storeu(xorv(rows[0], rows[2]), &out[0]);
+ storeu(xorv(rows[1], rows[3]), &out[16]);
+ storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
+ storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
+}
+
+INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
+ v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
+ v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
+ v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
+ v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
+ v[0] = addv(v[0], v[4]);
+ v[1] = addv(v[1], v[5]);
+ v[2] = addv(v[2], v[6]);
+ v[3] = addv(v[3], v[7]);
+ v[12] = xorv(v[12], v[0]);
+ v[13] = xorv(v[13], v[1]);
+ v[14] = xorv(v[14], v[2]);
+ v[15] = xorv(v[15], v[3]);
+ v[12] = rot16(v[12]);
+ v[13] = rot16(v[13]);
+ v[14] = rot16(v[14]);
+ v[15] = rot16(v[15]);
+ v[8] = addv(v[8], v[12]);
+ v[9] = addv(v[9], v[13]);
+ v[10] = addv(v[10], v[14]);
+ v[11] = addv(v[11], v[15]);
+ v[4] = xorv(v[4], v[8]);
+ v[5] = xorv(v[5], v[9]);
+ v[6] = xorv(v[6], v[10]);
+ v[7] = xorv(v[7], v[11]);
+ v[4] = rot12(v[4]);
+ v[5] = rot12(v[5]);
+ v[6] = rot12(v[6]);
+ v[7] = rot12(v[7]);
+ v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
+ v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
+ v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
+ v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
+ v[0] = addv(v[0], v[4]);
+ v[1] = addv(v[1], v[5]);
+ v[2] = addv(v[2], v[6]);
+ v[3] = addv(v[3], v[7]);
+ v[12] = xorv(v[12], v[0]);
+ v[13] = xorv(v[13], v[1]);
+ v[14] = xorv(v[14], v[2]);
+ v[15] = xorv(v[15], v[3]);
+ v[12] = rot8(v[12]);
+ v[13] = rot8(v[13]);
+ v[14] = rot8(v[14]);
+ v[15] = rot8(v[15]);
+ v[8] = addv(v[8], v[12]);
+ v[9] = addv(v[9], v[13]);
+ v[10] = addv(v[10], v[14]);
+ v[11] = addv(v[11], v[15]);
+ v[4] = xorv(v[4], v[8]);
+ v[5] = xorv(v[5], v[9]);
+ v[6] = xorv(v[6], v[10]);
+ v[7] = xorv(v[7], v[11]);
+ v[4] = rot7(v[4]);
+ v[5] = rot7(v[5]);
+ v[6] = rot7(v[6]);
+ v[7] = rot7(v[7]);
+
+ v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
+ v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
+ v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
+ v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
+ v[0] = addv(v[0], v[5]);
+ v[1] = addv(v[1], v[6]);
+ v[2] = addv(v[2], v[7]);
+ v[3] = addv(v[3], v[4]);
+ v[15] = xorv(v[15], v[0]);
+ v[12] = xorv(v[12], v[1]);
+ v[13] = xorv(v[13], v[2]);
+ v[14] = xorv(v[14], v[3]);
+ v[15] = rot16(v[15]);
+ v[12] = rot16(v[12]);
+ v[13] = rot16(v[13]);
+ v[14] = rot16(v[14]);
+ v[10] = addv(v[10], v[15]);
+ v[11] = addv(v[11], v[12]);
+ v[8] = addv(v[8], v[13]);
+ v[9] = addv(v[9], v[14]);
+ v[5] = xorv(v[5], v[10]);
+ v[6] = xorv(v[6], v[11]);
+ v[7] = xorv(v[7], v[8]);
+ v[4] = xorv(v[4], v[9]);
+ v[5] = rot12(v[5]);
+ v[6] = rot12(v[6]);
+ v[7] = rot12(v[7]);
+ v[4] = rot12(v[4]);
+ v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
+ v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
+ v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
+ v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
+ v[0] = addv(v[0], v[5]);
+ v[1] = addv(v[1], v[6]);
+ v[2] = addv(v[2], v[7]);
+ v[3] = addv(v[3], v[4]);
+ v[15] = xorv(v[15], v[0]);
+ v[12] = xorv(v[12], v[1]);
+ v[13] = xorv(v[13], v[2]);
+ v[14] = xorv(v[14], v[3]);
+ v[15] = rot8(v[15]);
+ v[12] = rot8(v[12]);
+ v[13] = rot8(v[13]);
+ v[14] = rot8(v[14]);
+ v[10] = addv(v[10], v[15]);
+ v[11] = addv(v[11], v[12]);
+ v[8] = addv(v[8], v[13]);
+ v[9] = addv(v[9], v[14]);
+ v[5] = xorv(v[5], v[10]);
+ v[6] = xorv(v[6], v[11]);
+ v[7] = xorv(v[7], v[8]);
+ v[4] = xorv(v[4], v[9]);
+ v[5] = rot7(v[5]);
+ v[6] = rot7(v[6]);
+ v[7] = rot7(v[7]);
+ v[4] = rot7(v[4]);
+}
+
+INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
+ // Interleave 32-bit lates. The low unpack is lanes 00/11 and the high is
+ // 22/33. Note that this doesn't split the vector into two lanes, as the
+ // AVX2 counterparts do.
+ __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
+ __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
+ __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
+ __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
+
+ // Interleave 64-bit lanes.
+ __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
+ __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
+ __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
+ __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
+
+ vecs[0] = abcd_0;
+ vecs[1] = abcd_1;
+ vecs[2] = abcd_2;
+ vecs[3] = abcd_3;
+}
+
+INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
+ size_t block_offset, __m128i out[16]) {
+ out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
+ out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
+ out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
+ out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
+ out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
+ out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
+ out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
+ out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
+ out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
+ out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
+ out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
+ out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
+ out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
+ out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
+ out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
+ out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
+ for (size_t i = 0; i < 4; ++i) {
+ _mm_prefetch(&inputs[i][block_offset + 256], _MM_HINT_T0);
+ }
+ transpose_vecs(&out[0]);
+ transpose_vecs(&out[4]);
+ transpose_vecs(&out[8]);
+ transpose_vecs(&out[12]);
+}
+
+INLINE void load_counters(uint64_t counter, bool increment_counter,
+ __m128i *out_lo, __m128i *out_hi) {
+ const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
+ const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
+ const __m128i add1 = _mm_and_si128(mask, add0);
+ __m128i l = _mm_add_epi32(_mm_set1_epi32(counter), add1);
+ __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)),
+ _mm_xor_si128( l, _mm_set1_epi32(0x80000000)));
+ __m128i h = _mm_sub_epi32(_mm_set1_epi32(counter >> 32), carry);
+ *out_lo = l;
+ *out_hi = h;
+}
+
+void blake3_hash4_sse2(const uint8_t *const *inputs, size_t blocks,
+ const uint32_t key[8], uint64_t counter,
+ bool increment_counter, uint8_t flags,
+ uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
+ __m128i h_vecs[8] = {
+ set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
+ set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
+ };
+ __m128i counter_low_vec, counter_high_vec;
+ load_counters(counter, increment_counter, &counter_low_vec,
+ &counter_high_vec);
+ uint8_t block_flags = flags | flags_start;
+
+ for (size_t block = 0; block < blocks; block++) {
+ if (block + 1 == blocks) {
+ block_flags |= flags_end;
+ }
+ __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
+ __m128i block_flags_vec = set1(block_flags);
+ __m128i msg_vecs[16];
+ transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+ __m128i v[16] = {
+ h_vecs[0], h_vecs[1], h_vecs[2], h_vecs[3],
+ h_vecs[4], h_vecs[5], h_vecs[6], h_vecs[7],
+ set1(IV[0]), set1(IV[1]), set1(IV[2]), set1(IV[3]),
+ counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
+ };
+ round_fn(v, msg_vecs, 0);
+ round_fn(v, msg_vecs, 1);
+ round_fn(v, msg_vecs, 2);
+ round_fn(v, msg_vecs, 3);
+ round_fn(v, msg_vecs, 4);
+ round_fn(v, msg_vecs, 5);
+ round_fn(v, msg_vecs, 6);
+ h_vecs[0] = xorv(v[0], v[8]);
+ h_vecs[1] = xorv(v[1], v[9]);
+ h_vecs[2] = xorv(v[2], v[10]);
+ h_vecs[3] = xorv(v[3], v[11]);
+ h_vecs[4] = xorv(v[4], v[12]);
+ h_vecs[5] = xorv(v[5], v[13]);
+ h_vecs[6] = xorv(v[6], v[14]);
+ h_vecs[7] = xorv(v[7], v[15]);
+
+ block_flags = flags;
+ }
+
+ transpose_vecs(&h_vecs[0]);
+ transpose_vecs(&h_vecs[4]);
+ // The first four vecs now contain the first half of each output, and the
+ // second four vecs contain the second half of each output.
+ storeu(h_vecs[0], &out[0 * sizeof(__m128i)]);
+ storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
+ storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
+ storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
+ storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
+ storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
+ storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
+ storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
+}
+
+INLINE void hash_one_sse2(const uint8_t *input, size_t blocks,
+ const uint32_t key[8], uint64_t counter,
+ uint8_t flags, uint8_t flags_start,
+ uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
+ uint32_t cv[8];
+ memcpy(cv, key, BLAKE3_KEY_LEN);
+ uint8_t block_flags = flags | flags_start;
+ while (blocks > 0) {
+ if (blocks == 1) {
+ block_flags |= flags_end;
+ }
+ blake3_compress_in_place_sse2(cv, input, BLAKE3_BLOCK_LEN, counter,
+ block_flags);
+ input = &input[BLAKE3_BLOCK_LEN];
+ blocks -= 1;
+ block_flags = flags;
+ }
+ memcpy(out, cv, BLAKE3_OUT_LEN);
+}
+
+void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs,
+ size_t blocks, const uint32_t key[8],
+ uint64_t counter, bool increment_counter,
+ uint8_t flags, uint8_t flags_start,
+ uint8_t flags_end, uint8_t *out) {
+ while (num_inputs >= DEGREE) {
+ blake3_hash4_sse2(inputs, blocks, key, counter, increment_counter, flags,
+ flags_start, flags_end, out);
+ if (increment_counter) {
+ counter += DEGREE;
+ }
+ inputs += DEGREE;
+ num_inputs -= DEGREE;
+ out = &out[DEGREE * BLAKE3_OUT_LEN];
+ }
+ while (num_inputs > 0) {
+ hash_one_sse2(inputs[0], blocks, key, counter, flags, flags_start,
+ flags_end, out);
+ if (increment_counter) {
+ counter += 1;
+ }
+ inputs += 1;
+ num_inputs -= 1;
+ out = &out[BLAKE3_OUT_LEN];
+ }
+}
--- /dev/null
+#if defined(__ELF__) && defined(__linux__)
+.section .note.GNU-stack,"",%progbits
+#endif
+
+#if defined(__ELF__) && defined(__CET__) && defined(__has_include)
+#if __has_include(<cet.h>)
+#include <cet.h>
+#endif
+#endif
+
+#if !defined(_CET_ENDBR)
+#define _CET_ENDBR
+#endif
+
+.intel_syntax noprefix
+.global blake3_hash_many_sse2
+.global _blake3_hash_many_sse2
+.global blake3_compress_in_place_sse2
+.global _blake3_compress_in_place_sse2
+.global blake3_compress_xof_sse2
+.global _blake3_compress_xof_sse2
+#ifdef __APPLE__
+.text
+#else
+.section .text
+#endif
+ .p2align 6
+_blake3_hash_many_sse2:
+blake3_hash_many_sse2:
+ _CET_ENDBR
+ push r15
+ push r14
+ push r13
+ push r12
+ push rbx
+ push rbp
+ mov rbp, rsp
+ sub rsp, 360
+ and rsp, 0xFFFFFFFFFFFFFFC0
+ neg r9d
+ movd xmm0, r9d
+ pshufd xmm0, xmm0, 0x00
+ movdqa xmmword ptr [rsp+0x130], xmm0
+ movdqa xmm1, xmm0
+ pand xmm1, xmmword ptr [ADD0+rip]
+ pand xmm0, xmmword ptr [ADD1+rip]
+ movdqa xmmword ptr [rsp+0x150], xmm0
+ movd xmm0, r8d
+ pshufd xmm0, xmm0, 0x00
+ paddd xmm0, xmm1
+ movdqa xmmword ptr [rsp+0x110], xmm0
+ pxor xmm0, xmmword ptr [CMP_MSB_MASK+rip]
+ pxor xmm1, xmmword ptr [CMP_MSB_MASK+rip]
+ pcmpgtd xmm1, xmm0
+ shr r8, 32
+ movd xmm2, r8d
+ pshufd xmm2, xmm2, 0x00
+ psubd xmm2, xmm1
+ movdqa xmmword ptr [rsp+0x120], xmm2
+ mov rbx, qword ptr [rbp+0x50]
+ mov r15, rdx
+ shl r15, 6
+ movzx r13d, byte ptr [rbp+0x38]
+ movzx r12d, byte ptr [rbp+0x48]
+ cmp rsi, 4
+ jc 3f
+2:
+ movdqu xmm3, xmmword ptr [rcx]
+ pshufd xmm0, xmm3, 0x00
+ pshufd xmm1, xmm3, 0x55
+ pshufd xmm2, xmm3, 0xAA
+ pshufd xmm3, xmm3, 0xFF
+ movdqu xmm7, xmmword ptr [rcx+0x10]
+ pshufd xmm4, xmm7, 0x00
+ pshufd xmm5, xmm7, 0x55
+ pshufd xmm6, xmm7, 0xAA
+ pshufd xmm7, xmm7, 0xFF
+ mov r8, qword ptr [rdi]
+ mov r9, qword ptr [rdi+0x8]
+ mov r10, qword ptr [rdi+0x10]
+ mov r11, qword ptr [rdi+0x18]
+ movzx eax, byte ptr [rbp+0x40]
+ or eax, r13d
+ xor edx, edx
+9:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movdqu xmm8, xmmword ptr [r8+rdx-0x40]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x40]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x40]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x40]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp], xmm8
+ movdqa xmmword ptr [rsp+0x10], xmm9
+ movdqa xmmword ptr [rsp+0x20], xmm12
+ movdqa xmmword ptr [rsp+0x30], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-0x30]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x30]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x30]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x30]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+0x40], xmm8
+ movdqa xmmword ptr [rsp+0x50], xmm9
+ movdqa xmmword ptr [rsp+0x60], xmm12
+ movdqa xmmword ptr [rsp+0x70], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-0x20]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x20]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x20]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x20]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+0x80], xmm8
+ movdqa xmmword ptr [rsp+0x90], xmm9
+ movdqa xmmword ptr [rsp+0xA0], xmm12
+ movdqa xmmword ptr [rsp+0xB0], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-0x10]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x10]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x10]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x10]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+0xC0], xmm8
+ movdqa xmmword ptr [rsp+0xD0], xmm9
+ movdqa xmmword ptr [rsp+0xE0], xmm12
+ movdqa xmmword ptr [rsp+0xF0], xmm13
+ movdqa xmm9, xmmword ptr [BLAKE3_IV_1+rip]
+ movdqa xmm10, xmmword ptr [BLAKE3_IV_2+rip]
+ movdqa xmm11, xmmword ptr [BLAKE3_IV_3+rip]
+ movdqa xmm12, xmmword ptr [rsp+0x110]
+ movdqa xmm13, xmmword ptr [rsp+0x120]
+ movdqa xmm14, xmmword ptr [BLAKE3_BLOCK_LEN+rip]
+ movd xmm15, eax
+ pshufd xmm15, xmm15, 0x00
+ prefetcht0 [r8+rdx+0x80]
+ prefetcht0 [r9+rdx+0x80]
+ prefetcht0 [r10+rdx+0x80]
+ prefetcht0 [r11+rdx+0x80]
+ paddd xmm0, xmmword ptr [rsp]
+ paddd xmm1, xmmword ptr [rsp+0x20]
+ paddd xmm2, xmmword ptr [rsp+0x40]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [BLAKE3_IV_0+rip]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x10]
+ paddd xmm1, xmmword ptr [rsp+0x30]
+ paddd xmm2, xmmword ptr [rsp+0x50]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x80]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp+0xC0]
+ paddd xmm3, xmmword ptr [rsp+0xE0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x90]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0xD0]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x20]
+ paddd xmm1, xmmword ptr [rsp+0x30]
+ paddd xmm2, xmmword ptr [rsp+0x70]
+ paddd xmm3, xmmword ptr [rsp+0x40]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x60]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp]
+ paddd xmm3, xmmword ptr [rsp+0xD0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x10]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0x90]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xB0]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp+0xE0]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x30]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp+0xD0]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x40]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0x20]
+ paddd xmm3, xmmword ptr [rsp+0xE0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x60]
+ paddd xmm1, xmmword ptr [rsp+0x90]
+ paddd xmm2, xmmword ptr [rsp+0xB0]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x50]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+0xF0]
+ paddd xmm3, xmmword ptr [rsp+0x10]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xA0]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0xE0]
+ paddd xmm3, xmmword ptr [rsp+0xD0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x70]
+ paddd xmm1, xmmword ptr [rsp+0x90]
+ paddd xmm2, xmmword ptr [rsp+0x30]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x40]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0x50]
+ paddd xmm3, xmmword ptr [rsp+0x10]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp]
+ paddd xmm1, xmmword ptr [rsp+0x20]
+ paddd xmm2, xmmword ptr [rsp+0x80]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xC0]
+ paddd xmm1, xmmword ptr [rsp+0x90]
+ paddd xmm2, xmmword ptr [rsp+0xF0]
+ paddd xmm3, xmmword ptr [rsp+0xE0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xD0]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0xA0]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x70]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x20]
+ paddd xmm1, xmmword ptr [rsp+0x30]
+ paddd xmm2, xmmword ptr [rsp+0x10]
+ paddd xmm3, xmmword ptr [rsp+0x40]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x90]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0x80]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xE0]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp+0xC0]
+ paddd xmm3, xmmword ptr [rsp+0x10]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xD0]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+0x20]
+ paddd xmm3, xmmword ptr [rsp+0x40]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x30]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp+0x60]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xB0]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp+0x10]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xF0]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+0x90]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xE0]
+ paddd xmm1, xmmword ptr [rsp+0x20]
+ paddd xmm2, xmmword ptr [rsp+0x30]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xA0]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0x40]
+ paddd xmm3, xmmword ptr [rsp+0xD0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ pxor xmm0, xmm8
+ pxor xmm1, xmm9
+ pxor xmm2, xmm10
+ pxor xmm3, xmm11
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ pxor xmm4, xmm12
+ pxor xmm5, xmm13
+ pxor xmm6, xmm14
+ pxor xmm7, xmm15
+ mov eax, r13d
+ jne 9b
+ movdqa xmm9, xmm0
+ punpckldq xmm0, xmm1
+ punpckhdq xmm9, xmm1
+ movdqa xmm11, xmm2
+ punpckldq xmm2, xmm3
+ punpckhdq xmm11, xmm3
+ movdqa xmm1, xmm0
+ punpcklqdq xmm0, xmm2
+ punpckhqdq xmm1, xmm2
+ movdqa xmm3, xmm9
+ punpcklqdq xmm9, xmm11
+ punpckhqdq xmm3, xmm11
+ movdqu xmmword ptr [rbx], xmm0
+ movdqu xmmword ptr [rbx+0x20], xmm1
+ movdqu xmmword ptr [rbx+0x40], xmm9
+ movdqu xmmword ptr [rbx+0x60], xmm3
+ movdqa xmm9, xmm4
+ punpckldq xmm4, xmm5
+ punpckhdq xmm9, xmm5
+ movdqa xmm11, xmm6
+ punpckldq xmm6, xmm7
+ punpckhdq xmm11, xmm7
+ movdqa xmm5, xmm4
+ punpcklqdq xmm4, xmm6
+ punpckhqdq xmm5, xmm6
+ movdqa xmm7, xmm9
+ punpcklqdq xmm9, xmm11
+ punpckhqdq xmm7, xmm11
+ movdqu xmmword ptr [rbx+0x10], xmm4
+ movdqu xmmword ptr [rbx+0x30], xmm5
+ movdqu xmmword ptr [rbx+0x50], xmm9
+ movdqu xmmword ptr [rbx+0x70], xmm7
+ movdqa xmm1, xmmword ptr [rsp+0x110]
+ movdqa xmm0, xmm1
+ paddd xmm1, xmmword ptr [rsp+0x150]
+ movdqa xmmword ptr [rsp+0x110], xmm1
+ pxor xmm0, xmmword ptr [CMP_MSB_MASK+rip]
+ pxor xmm1, xmmword ptr [CMP_MSB_MASK+rip]
+ pcmpgtd xmm0, xmm1
+ movdqa xmm1, xmmword ptr [rsp+0x120]
+ psubd xmm1, xmm0
+ movdqa xmmword ptr [rsp+0x120], xmm1
+ add rbx, 128
+ add rdi, 32
+ sub rsi, 4
+ cmp rsi, 4
+ jnc 2b
+ test rsi, rsi
+ jnz 3f
+4:
+ mov rsp, rbp
+ pop rbp
+ pop rbx
+ pop r12
+ pop r13
+ pop r14
+ pop r15
+ ret
+.p2align 5
+3:
+ test esi, 0x2
+ je 3f
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+0x10]
+ movaps xmm8, xmm0
+ movaps xmm9, xmm1
+ movd xmm13, dword ptr [rsp+0x110]
+ movd xmm14, dword ptr [rsp+0x120]
+ punpckldq xmm13, xmm14
+ movaps xmmword ptr [rsp], xmm13
+ movd xmm14, dword ptr [rsp+0x114]
+ movd xmm13, dword ptr [rsp+0x124]
+ punpckldq xmm14, xmm13
+ movaps xmmword ptr [rsp+0x10], xmm14
+ mov r8, qword ptr [rdi]
+ mov r9, qword ptr [rdi+0x8]
+ movzx eax, byte ptr [rbp+0x40]
+ or eax, r13d
+ xor edx, edx
+2:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ movaps xmm10, xmm2
+ movups xmm4, xmmword ptr [r8+rdx-0x40]
+ movups xmm5, xmmword ptr [r8+rdx-0x30]
+ movaps xmm3, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm3, xmm5, 221
+ movaps xmm5, xmm3
+ movups xmm6, xmmword ptr [r8+rdx-0x20]
+ movups xmm7, xmmword ptr [r8+rdx-0x10]
+ movaps xmm3, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm3, xmm7, 221
+ pshufd xmm7, xmm3, 0x93
+ movups xmm12, xmmword ptr [r9+rdx-0x40]
+ movups xmm13, xmmword ptr [r9+rdx-0x30]
+ movaps xmm11, xmm12
+ shufps xmm12, xmm13, 136
+ shufps xmm11, xmm13, 221
+ movaps xmm13, xmm11
+ movups xmm14, xmmword ptr [r9+rdx-0x20]
+ movups xmm15, xmmword ptr [r9+rdx-0x10]
+ movaps xmm11, xmm14
+ shufps xmm14, xmm15, 136
+ pshufd xmm14, xmm14, 0x93
+ shufps xmm11, xmm15, 221
+ pshufd xmm15, xmm11, 0x93
+ shl rax, 0x20
+ or rax, 0x40
+ movd xmm3, rax
+ movdqa xmmword ptr [rsp+0x20], xmm3
+ movaps xmm3, xmmword ptr [rsp]
+ movaps xmm11, xmmword ptr [rsp+0x10]
+ punpcklqdq xmm3, xmmword ptr [rsp+0x20]
+ punpcklqdq xmm11, xmmword ptr [rsp+0x20]
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm8, xmm12
+ movaps xmmword ptr [rsp+0x20], xmm4
+ movaps xmmword ptr [rsp+0x30], xmm12
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ pshuflw xmm11, xmm11, 0xB1
+ pshufhw xmm11, xmm11, 0xB1
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 20
+ psrld xmm4, 12
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 20
+ psrld xmm4, 12
+ por xmm9, xmm4
+ paddd xmm0, xmm5
+ paddd xmm8, xmm13
+ movaps xmmword ptr [rsp+0x40], xmm5
+ movaps xmmword ptr [rsp+0x50], xmm13
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ movdqa xmm13, xmm3
+ psrld xmm3, 8
+ pslld xmm13, 24
+ pxor xmm3, xmm13
+ movdqa xmm13, xmm11
+ psrld xmm11, 8
+ pslld xmm13, 24
+ pxor xmm11, xmm13
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 25
+ psrld xmm4, 7
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 25
+ psrld xmm4, 7
+ por xmm9, xmm4
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm8, xmm8, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm11, xmm11, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ pshufd xmm10, xmm10, 0x39
+ paddd xmm0, xmm6
+ paddd xmm8, xmm14
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ pshuflw xmm11, xmm11, 0xB1
+ pshufhw xmm11, xmm11, 0xB1
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 20
+ psrld xmm4, 12
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 20
+ psrld xmm4, 12
+ por xmm9, xmm4
+ paddd xmm0, xmm7
+ paddd xmm8, xmm15
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ movdqa xmm13, xmm3
+ psrld xmm3, 8
+ pslld xmm13, 24
+ pxor xmm3, xmm13
+ movdqa xmm13, xmm11
+ psrld xmm11, 8
+ pslld xmm13, 24
+ pxor xmm11, xmm13
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 25
+ psrld xmm4, 7
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 25
+ psrld xmm4, 7
+ por xmm9, xmm4
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm8, xmm8, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm11, xmm11, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ pshufd xmm10, xmm10, 0x93
+ dec al
+ je 9f
+ movdqa xmm12, xmmword ptr [rsp+0x20]
+ movdqa xmm5, xmmword ptr [rsp+0x40]
+ pshufd xmm13, xmm12, 0x0F
+ shufps xmm12, xmm5, 214
+ pshufd xmm4, xmm12, 0x39
+ movdqa xmm12, xmm6
+ shufps xmm12, xmm7, 250
+ pand xmm13, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm12, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm13, xmm12
+ movdqa xmmword ptr [rsp+0x20], xmm13
+ movdqa xmm12, xmm7
+ punpcklqdq xmm12, xmm5
+ movdqa xmm13, xmm6
+ pand xmm12, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm13, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm12, xmm13
+ pshufd xmm12, xmm12, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmmword ptr [rsp+0x40], xmm12
+ movdqa xmm5, xmmword ptr [rsp+0x30]
+ movdqa xmm13, xmmword ptr [rsp+0x50]
+ pshufd xmm6, xmm5, 0x0F
+ shufps xmm5, xmm13, 214
+ pshufd xmm12, xmm5, 0x39
+ movdqa xmm5, xmm14
+ shufps xmm5, xmm15, 250
+ pand xmm6, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm5, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm6, xmm5
+ movdqa xmm5, xmm15
+ punpcklqdq xmm5, xmm13
+ movdqa xmmword ptr [rsp+0x30], xmm2
+ movdqa xmm2, xmm14
+ pand xmm5, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm2, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm5, xmm2
+ movdqa xmm2, xmmword ptr [rsp+0x30]
+ pshufd xmm5, xmm5, 0x78
+ punpckhdq xmm13, xmm15
+ punpckldq xmm14, xmm13
+ pshufd xmm15, xmm14, 0x1E
+ movdqa xmm13, xmm6
+ movdqa xmm14, xmm5
+ movdqa xmm5, xmmword ptr [rsp+0x20]
+ movdqa xmm6, xmmword ptr [rsp+0x40]
+ jmp 9b
+9:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ pxor xmm8, xmm10
+ pxor xmm9, xmm11
+ mov eax, r13d
+ cmp rdx, r15
+ jne 2b
+ movups xmmword ptr [rbx], xmm0
+ movups xmmword ptr [rbx+0x10], xmm1
+ movups xmmword ptr [rbx+0x20], xmm8
+ movups xmmword ptr [rbx+0x30], xmm9
+ mov eax, dword ptr [rsp+0x130]
+ neg eax
+ mov r10d, dword ptr [rsp+0x110+8*rax]
+ mov r11d, dword ptr [rsp+0x120+8*rax]
+ mov dword ptr [rsp+0x110], r10d
+ mov dword ptr [rsp+0x120], r11d
+ add rdi, 16
+ add rbx, 64
+ sub rsi, 2
+3:
+ test esi, 0x1
+ je 4b
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+0x10]
+ movd xmm13, dword ptr [rsp+0x110]
+ movd xmm14, dword ptr [rsp+0x120]
+ punpckldq xmm13, xmm14
+ mov r8, qword ptr [rdi]
+ movzx eax, byte ptr [rbp+0x40]
+ or eax, r13d
+ xor edx, edx
+2:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ shl rax, 32
+ or rax, 64
+ movd xmm12, rax
+ movdqa xmm3, xmm13
+ punpcklqdq xmm3, xmm12
+ movups xmm4, xmmword ptr [r8+rdx-0x40]
+ movups xmm5, xmmword ptr [r8+rdx-0x30]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [r8+rdx-0x20]
+ movups xmm7, xmmword ptr [r8+rdx-0x10]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 0x93
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ dec al
+ jz 9f
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0x0F
+ pshufd xmm4, xmm8, 0x39
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp 9b
+9:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ mov eax, r13d
+ cmp rdx, r15
+ jne 2b
+ movups xmmword ptr [rbx], xmm0
+ movups xmmword ptr [rbx+0x10], xmm1
+ jmp 4b
+
+.p2align 6
+blake3_compress_in_place_sse2:
+_blake3_compress_in_place_sse2:
+ _CET_ENDBR
+ movups xmm0, xmmword ptr [rdi]
+ movups xmm1, xmmword ptr [rdi+0x10]
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ shl r8, 32
+ add rdx, r8
+ movq xmm3, rcx
+ movq xmm4, rdx
+ punpcklqdq xmm3, xmm4
+ movups xmm4, xmmword ptr [rsi]
+ movups xmm5, xmmword ptr [rsi+0x10]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [rsi+0x20]
+ movups xmm7, xmmword ptr [rsi+0x30]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 0x93
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ dec al
+ jz 9f
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0x0F
+ pshufd xmm4, xmm8, 0x39
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp 9b
+9:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ movups xmmword ptr [rdi], xmm0
+ movups xmmword ptr [rdi+0x10], xmm1
+ ret
+
+.p2align 6
+blake3_compress_xof_sse2:
+_blake3_compress_xof_sse2:
+ _CET_ENDBR
+ movups xmm0, xmmword ptr [rdi]
+ movups xmm1, xmmword ptr [rdi+0x10]
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ movzx eax, r8b
+ movzx edx, dl
+ shl rax, 32
+ add rdx, rax
+ movq xmm3, rcx
+ movq xmm4, rdx
+ punpcklqdq xmm3, xmm4
+ movups xmm4, xmmword ptr [rsi]
+ movups xmm5, xmmword ptr [rsi+0x10]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [rsi+0x20]
+ movups xmm7, xmmword ptr [rsi+0x30]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 0x93
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ dec al
+ jz 9f
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0x0F
+ pshufd xmm4, xmm8, 0x39
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp 9b
+9:
+ movdqu xmm4, xmmword ptr [rdi]
+ movdqu xmm5, xmmword ptr [rdi+0x10]
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ pxor xmm2, xmm4
+ pxor xmm3, xmm5
+ movups xmmword ptr [r9], xmm0
+ movups xmmword ptr [r9+0x10], xmm1
+ movups xmmword ptr [r9+0x20], xmm2
+ movups xmmword ptr [r9+0x30], xmm3
+ ret
+
+
+#ifdef __APPLE__
+.static_data
+#else
+.section .rodata
+#endif
+.p2align 6
+BLAKE3_IV:
+ .long 0x6A09E667, 0xBB67AE85
+ .long 0x3C6EF372, 0xA54FF53A
+ADD0:
+ .long 0, 1, 2, 3
+ADD1:
+ .long 4, 4, 4, 4
+BLAKE3_IV_0:
+ .long 0x6A09E667, 0x6A09E667, 0x6A09E667, 0x6A09E667
+BLAKE3_IV_1:
+ .long 0xBB67AE85, 0xBB67AE85, 0xBB67AE85, 0xBB67AE85
+BLAKE3_IV_2:
+ .long 0x3C6EF372, 0x3C6EF372, 0x3C6EF372, 0x3C6EF372
+BLAKE3_IV_3:
+ .long 0xA54FF53A, 0xA54FF53A, 0xA54FF53A, 0xA54FF53A
+BLAKE3_BLOCK_LEN:
+ .long 64, 64, 64, 64
+CMP_MSB_MASK:
+ .long 0x80000000, 0x80000000, 0x80000000, 0x80000000
+PBLENDW_0x33_MASK:
+ .long 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF, 0x00000000
+PBLENDW_0xCC_MASK:
+ .long 0x00000000, 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF
+PBLENDW_0x3F_MASK:
+ .long 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000
+PBLENDW_0xC0_MASK:
+ .long 0x00000000, 0x00000000, 0x00000000, 0xFFFFFFFF
--- /dev/null
+.intel_syntax noprefix
+.global blake3_hash_many_sse2
+.global _blake3_hash_many_sse2
+.global blake3_compress_in_place_sse2
+.global _blake3_compress_in_place_sse2
+.global blake3_compress_xof_sse2
+.global _blake3_compress_xof_sse2
+.section .text
+ .p2align 6
+_blake3_hash_many_sse2:
+blake3_hash_many_sse2:
+ push r15
+ push r14
+ push r13
+ push r12
+ push rsi
+ push rdi
+ push rbx
+ push rbp
+ mov rbp, rsp
+ sub rsp, 528
+ and rsp, 0xFFFFFFFFFFFFFFC0
+ movdqa xmmword ptr [rsp+0x170], xmm6
+ movdqa xmmword ptr [rsp+0x180], xmm7
+ movdqa xmmword ptr [rsp+0x190], xmm8
+ movdqa xmmword ptr [rsp+0x1A0], xmm9
+ movdqa xmmword ptr [rsp+0x1B0], xmm10
+ movdqa xmmword ptr [rsp+0x1C0], xmm11
+ movdqa xmmword ptr [rsp+0x1D0], xmm12
+ movdqa xmmword ptr [rsp+0x1E0], xmm13
+ movdqa xmmword ptr [rsp+0x1F0], xmm14
+ movdqa xmmword ptr [rsp+0x200], xmm15
+ mov rdi, rcx
+ mov rsi, rdx
+ mov rdx, r8
+ mov rcx, r9
+ mov r8, qword ptr [rbp+0x68]
+ movzx r9, byte ptr [rbp+0x70]
+ neg r9d
+ movd xmm0, r9d
+ pshufd xmm0, xmm0, 0x00
+ movdqa xmmword ptr [rsp+0x130], xmm0
+ movdqa xmm1, xmm0
+ pand xmm1, xmmword ptr [ADD0+rip]
+ pand xmm0, xmmword ptr [ADD1+rip]
+ movdqa xmmword ptr [rsp+0x150], xmm0
+ movd xmm0, r8d
+ pshufd xmm0, xmm0, 0x00
+ paddd xmm0, xmm1
+ movdqa xmmword ptr [rsp+0x110], xmm0
+ pxor xmm0, xmmword ptr [CMP_MSB_MASK+rip]
+ pxor xmm1, xmmword ptr [CMP_MSB_MASK+rip]
+ pcmpgtd xmm1, xmm0
+ shr r8, 32
+ movd xmm2, r8d
+ pshufd xmm2, xmm2, 0x00
+ psubd xmm2, xmm1
+ movdqa xmmword ptr [rsp+0x120], xmm2
+ mov rbx, qword ptr [rbp+0x90]
+ mov r15, rdx
+ shl r15, 6
+ movzx r13d, byte ptr [rbp+0x78]
+ movzx r12d, byte ptr [rbp+0x88]
+ cmp rsi, 4
+ jc 3f
+2:
+ movdqu xmm3, xmmword ptr [rcx]
+ pshufd xmm0, xmm3, 0x00
+ pshufd xmm1, xmm3, 0x55
+ pshufd xmm2, xmm3, 0xAA
+ pshufd xmm3, xmm3, 0xFF
+ movdqu xmm7, xmmword ptr [rcx+0x10]
+ pshufd xmm4, xmm7, 0x00
+ pshufd xmm5, xmm7, 0x55
+ pshufd xmm6, xmm7, 0xAA
+ pshufd xmm7, xmm7, 0xFF
+ mov r8, qword ptr [rdi]
+ mov r9, qword ptr [rdi+0x8]
+ mov r10, qword ptr [rdi+0x10]
+ mov r11, qword ptr [rdi+0x18]
+ movzx eax, byte ptr [rbp+0x80]
+ or eax, r13d
+ xor edx, edx
+9:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movdqu xmm8, xmmword ptr [r8+rdx-0x40]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x40]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x40]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x40]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp], xmm8
+ movdqa xmmword ptr [rsp+0x10], xmm9
+ movdqa xmmword ptr [rsp+0x20], xmm12
+ movdqa xmmword ptr [rsp+0x30], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-0x30]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x30]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x30]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x30]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+0x40], xmm8
+ movdqa xmmword ptr [rsp+0x50], xmm9
+ movdqa xmmword ptr [rsp+0x60], xmm12
+ movdqa xmmword ptr [rsp+0x70], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-0x20]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x20]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x20]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x20]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+0x80], xmm8
+ movdqa xmmword ptr [rsp+0x90], xmm9
+ movdqa xmmword ptr [rsp+0xA0], xmm12
+ movdqa xmmword ptr [rsp+0xB0], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-0x10]
+ movdqu xmm9, xmmword ptr [r9+rdx-0x10]
+ movdqu xmm10, xmmword ptr [r10+rdx-0x10]
+ movdqu xmm11, xmmword ptr [r11+rdx-0x10]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+0xC0], xmm8
+ movdqa xmmword ptr [rsp+0xD0], xmm9
+ movdqa xmmword ptr [rsp+0xE0], xmm12
+ movdqa xmmword ptr [rsp+0xF0], xmm13
+ movdqa xmm9, xmmword ptr [BLAKE3_IV_1+rip]
+ movdqa xmm10, xmmword ptr [BLAKE3_IV_2+rip]
+ movdqa xmm11, xmmword ptr [BLAKE3_IV_3+rip]
+ movdqa xmm12, xmmword ptr [rsp+0x110]
+ movdqa xmm13, xmmword ptr [rsp+0x120]
+ movdqa xmm14, xmmword ptr [BLAKE3_BLOCK_LEN+rip]
+ movd xmm15, eax
+ pshufd xmm15, xmm15, 0x00
+ prefetcht0 [r8+rdx+0x80]
+ prefetcht0 [r9+rdx+0x80]
+ prefetcht0 [r10+rdx+0x80]
+ prefetcht0 [r11+rdx+0x80]
+ paddd xmm0, xmmword ptr [rsp]
+ paddd xmm1, xmmword ptr [rsp+0x20]
+ paddd xmm2, xmmword ptr [rsp+0x40]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [BLAKE3_IV_0+rip]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x10]
+ paddd xmm1, xmmword ptr [rsp+0x30]
+ paddd xmm2, xmmword ptr [rsp+0x50]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x80]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp+0xC0]
+ paddd xmm3, xmmword ptr [rsp+0xE0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x90]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0xD0]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x20]
+ paddd xmm1, xmmword ptr [rsp+0x30]
+ paddd xmm2, xmmword ptr [rsp+0x70]
+ paddd xmm3, xmmword ptr [rsp+0x40]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x60]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp]
+ paddd xmm3, xmmword ptr [rsp+0xD0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x10]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0x90]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xB0]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp+0xE0]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x30]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp+0xD0]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x40]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0x20]
+ paddd xmm3, xmmword ptr [rsp+0xE0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x60]
+ paddd xmm1, xmmword ptr [rsp+0x90]
+ paddd xmm2, xmmword ptr [rsp+0xB0]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x50]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+0xF0]
+ paddd xmm3, xmmword ptr [rsp+0x10]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xA0]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0xE0]
+ paddd xmm3, xmmword ptr [rsp+0xD0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x70]
+ paddd xmm1, xmmword ptr [rsp+0x90]
+ paddd xmm2, xmmword ptr [rsp+0x30]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x40]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0x50]
+ paddd xmm3, xmmword ptr [rsp+0x10]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp]
+ paddd xmm1, xmmword ptr [rsp+0x20]
+ paddd xmm2, xmmword ptr [rsp+0x80]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xC0]
+ paddd xmm1, xmmword ptr [rsp+0x90]
+ paddd xmm2, xmmword ptr [rsp+0xF0]
+ paddd xmm3, xmmword ptr [rsp+0xE0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xD0]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0xA0]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x70]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x20]
+ paddd xmm1, xmmword ptr [rsp+0x30]
+ paddd xmm2, xmmword ptr [rsp+0x10]
+ paddd xmm3, xmmword ptr [rsp+0x40]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x90]
+ paddd xmm1, xmmword ptr [rsp+0xB0]
+ paddd xmm2, xmmword ptr [rsp+0x80]
+ paddd xmm3, xmmword ptr [rsp+0xF0]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xE0]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp+0xC0]
+ paddd xmm3, xmmword ptr [rsp+0x10]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xD0]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+0x20]
+ paddd xmm3, xmmword ptr [rsp+0x40]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0x30]
+ paddd xmm1, xmmword ptr [rsp+0xA0]
+ paddd xmm2, xmmword ptr [rsp+0x60]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xB0]
+ paddd xmm1, xmmword ptr [rsp+0x50]
+ paddd xmm2, xmmword ptr [rsp+0x10]
+ paddd xmm3, xmmword ptr [rsp+0x80]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xF0]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+0x90]
+ paddd xmm3, xmmword ptr [rsp+0x60]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xE0]
+ paddd xmm1, xmmword ptr [rsp+0x20]
+ paddd xmm2, xmmword ptr [rsp+0x30]
+ paddd xmm3, xmmword ptr [rsp+0x70]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0xB1
+ pshufhw xmm15, xmm15, 0xB1
+ pshuflw xmm12, xmm12, 0xB1
+ pshufhw xmm12, xmm12, 0xB1
+ pshuflw xmm13, xmm13, 0xB1
+ pshufhw xmm13, xmm13, 0xB1
+ pshuflw xmm14, xmm14, 0xB1
+ pshufhw xmm14, xmm14, 0xB1
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+0x100], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0xA0]
+ paddd xmm1, xmmword ptr [rsp+0xC0]
+ paddd xmm2, xmmword ptr [rsp+0x40]
+ paddd xmm3, xmmword ptr [rsp+0xD0]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+0x100]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ pxor xmm0, xmm8
+ pxor xmm1, xmm9
+ pxor xmm2, xmm10
+ pxor xmm3, xmm11
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ pxor xmm4, xmm12
+ pxor xmm5, xmm13
+ pxor xmm6, xmm14
+ pxor xmm7, xmm15
+ mov eax, r13d
+ jne 9b
+ movdqa xmm9, xmm0
+ punpckldq xmm0, xmm1
+ punpckhdq xmm9, xmm1
+ movdqa xmm11, xmm2
+ punpckldq xmm2, xmm3
+ punpckhdq xmm11, xmm3
+ movdqa xmm1, xmm0
+ punpcklqdq xmm0, xmm2
+ punpckhqdq xmm1, xmm2
+ movdqa xmm3, xmm9
+ punpcklqdq xmm9, xmm11
+ punpckhqdq xmm3, xmm11
+ movdqu xmmword ptr [rbx], xmm0
+ movdqu xmmword ptr [rbx+0x20], xmm1
+ movdqu xmmword ptr [rbx+0x40], xmm9
+ movdqu xmmword ptr [rbx+0x60], xmm3
+ movdqa xmm9, xmm4
+ punpckldq xmm4, xmm5
+ punpckhdq xmm9, xmm5
+ movdqa xmm11, xmm6
+ punpckldq xmm6, xmm7
+ punpckhdq xmm11, xmm7
+ movdqa xmm5, xmm4
+ punpcklqdq xmm4, xmm6
+ punpckhqdq xmm5, xmm6
+ movdqa xmm7, xmm9
+ punpcklqdq xmm9, xmm11
+ punpckhqdq xmm7, xmm11
+ movdqu xmmword ptr [rbx+0x10], xmm4
+ movdqu xmmword ptr [rbx+0x30], xmm5
+ movdqu xmmword ptr [rbx+0x50], xmm9
+ movdqu xmmword ptr [rbx+0x70], xmm7
+ movdqa xmm1, xmmword ptr [rsp+0x110]
+ movdqa xmm0, xmm1
+ paddd xmm1, xmmword ptr [rsp+0x150]
+ movdqa xmmword ptr [rsp+0x110], xmm1
+ pxor xmm0, xmmword ptr [CMP_MSB_MASK+rip]
+ pxor xmm1, xmmword ptr [CMP_MSB_MASK+rip]
+ pcmpgtd xmm0, xmm1
+ movdqa xmm1, xmmword ptr [rsp+0x120]
+ psubd xmm1, xmm0
+ movdqa xmmword ptr [rsp+0x120], xmm1
+ add rbx, 128
+ add rdi, 32
+ sub rsi, 4
+ cmp rsi, 4
+ jnc 2b
+ test rsi, rsi
+ jne 3f
+4:
+ movdqa xmm6, xmmword ptr [rsp+0x170]
+ movdqa xmm7, xmmword ptr [rsp+0x180]
+ movdqa xmm8, xmmword ptr [rsp+0x190]
+ movdqa xmm9, xmmword ptr [rsp+0x1A0]
+ movdqa xmm10, xmmword ptr [rsp+0x1B0]
+ movdqa xmm11, xmmword ptr [rsp+0x1C0]
+ movdqa xmm12, xmmword ptr [rsp+0x1D0]
+ movdqa xmm13, xmmword ptr [rsp+0x1E0]
+ movdqa xmm14, xmmword ptr [rsp+0x1F0]
+ movdqa xmm15, xmmword ptr [rsp+0x200]
+ mov rsp, rbp
+ pop rbp
+ pop rbx
+ pop rdi
+ pop rsi
+ pop r12
+ pop r13
+ pop r14
+ pop r15
+ ret
+.p2align 5
+3:
+ test esi, 0x2
+ je 3f
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+0x10]
+ movaps xmm8, xmm0
+ movaps xmm9, xmm1
+ movd xmm13, dword ptr [rsp+0x110]
+ movd xmm14, dword ptr [rsp+0x120]
+ punpckldq xmm13, xmm14
+ movaps xmmword ptr [rsp], xmm13
+ movd xmm14, dword ptr [rsp+0x114]
+ movd xmm13, dword ptr [rsp+0x124]
+ punpckldq xmm14, xmm13
+ movaps xmmword ptr [rsp+0x10], xmm14
+ mov r8, qword ptr [rdi]
+ mov r9, qword ptr [rdi+0x8]
+ movzx eax, byte ptr [rbp+0x80]
+ or eax, r13d
+ xor edx, edx
+2:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ movaps xmm10, xmm2
+ movups xmm4, xmmword ptr [r8+rdx-0x40]
+ movups xmm5, xmmword ptr [r8+rdx-0x30]
+ movaps xmm3, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm3, xmm5, 221
+ movaps xmm5, xmm3
+ movups xmm6, xmmword ptr [r8+rdx-0x20]
+ movups xmm7, xmmword ptr [r8+rdx-0x10]
+ movaps xmm3, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm3, xmm7, 221
+ pshufd xmm7, xmm3, 0x93
+ movups xmm12, xmmword ptr [r9+rdx-0x40]
+ movups xmm13, xmmword ptr [r9+rdx-0x30]
+ movaps xmm11, xmm12
+ shufps xmm12, xmm13, 136
+ shufps xmm11, xmm13, 221
+ movaps xmm13, xmm11
+ movups xmm14, xmmword ptr [r9+rdx-0x20]
+ movups xmm15, xmmword ptr [r9+rdx-0x10]
+ movaps xmm11, xmm14
+ shufps xmm14, xmm15, 136
+ pshufd xmm14, xmm14, 0x93
+ shufps xmm11, xmm15, 221
+ pshufd xmm15, xmm11, 0x93
+ shl rax, 0x20
+ or rax, 0x40
+ movd xmm3, rax
+ movdqa xmmword ptr [rsp+0x20], xmm3
+ movaps xmm3, xmmword ptr [rsp]
+ movaps xmm11, xmmword ptr [rsp+0x10]
+ punpcklqdq xmm3, xmmword ptr [rsp+0x20]
+ punpcklqdq xmm11, xmmword ptr [rsp+0x20]
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm8, xmm12
+ movaps xmmword ptr [rsp+0x20], xmm4
+ movaps xmmword ptr [rsp+0x30], xmm12
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ pshuflw xmm11, xmm11, 0xB1
+ pshufhw xmm11, xmm11, 0xB1
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 20
+ psrld xmm4, 12
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 20
+ psrld xmm4, 12
+ por xmm9, xmm4
+ paddd xmm0, xmm5
+ paddd xmm8, xmm13
+ movaps xmmword ptr [rsp+0x40], xmm5
+ movaps xmmword ptr [rsp+0x50], xmm13
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ movdqa xmm13, xmm3
+ psrld xmm3, 8
+ pslld xmm13, 24
+ pxor xmm3, xmm13
+ movdqa xmm13, xmm11
+ psrld xmm11, 8
+ pslld xmm13, 24
+ pxor xmm11, xmm13
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 25
+ psrld xmm4, 7
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 25
+ psrld xmm4, 7
+ por xmm9, xmm4
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm8, xmm8, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm11, xmm11, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ pshufd xmm10, xmm10, 0x39
+ paddd xmm0, xmm6
+ paddd xmm8, xmm14
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ pshuflw xmm11, xmm11, 0xB1
+ pshufhw xmm11, xmm11, 0xB1
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 20
+ psrld xmm4, 12
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 20
+ psrld xmm4, 12
+ por xmm9, xmm4
+ paddd xmm0, xmm7
+ paddd xmm8, xmm15
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ movdqa xmm13, xmm3
+ psrld xmm3, 8
+ pslld xmm13, 24
+ pxor xmm3, xmm13
+ movdqa xmm13, xmm11
+ psrld xmm11, 8
+ pslld xmm13, 24
+ pxor xmm11, xmm13
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 25
+ psrld xmm4, 7
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 25
+ psrld xmm4, 7
+ por xmm9, xmm4
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm8, xmm8, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm11, xmm11, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ pshufd xmm10, xmm10, 0x93
+ dec al
+ je 9f
+ movdqa xmm12, xmmword ptr [rsp+0x20]
+ movdqa xmm5, xmmword ptr [rsp+0x40]
+ pshufd xmm13, xmm12, 0x0F
+ shufps xmm12, xmm5, 214
+ pshufd xmm4, xmm12, 0x39
+ movdqa xmm12, xmm6
+ shufps xmm12, xmm7, 250
+ pand xmm13, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm12, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm13, xmm12
+ movdqa xmmword ptr [rsp+0x20], xmm13
+ movdqa xmm12, xmm7
+ punpcklqdq xmm12, xmm5
+ movdqa xmm13, xmm6
+ pand xmm12, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm13, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm12, xmm13
+ pshufd xmm12, xmm12, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmmword ptr [rsp+0x40], xmm12
+ movdqa xmm5, xmmword ptr [rsp+0x30]
+ movdqa xmm13, xmmword ptr [rsp+0x50]
+ pshufd xmm6, xmm5, 0x0F
+ shufps xmm5, xmm13, 214
+ pshufd xmm12, xmm5, 0x39
+ movdqa xmm5, xmm14
+ shufps xmm5, xmm15, 250
+ pand xmm6, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm5, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm6, xmm5
+ movdqa xmm5, xmm15
+ punpcklqdq xmm5, xmm13
+ movdqa xmmword ptr [rsp+0x30], xmm2
+ movdqa xmm2, xmm14
+ pand xmm5, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm2, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm5, xmm2
+ movdqa xmm2, xmmword ptr [rsp+0x30]
+ pshufd xmm5, xmm5, 0x78
+ punpckhdq xmm13, xmm15
+ punpckldq xmm14, xmm13
+ pshufd xmm15, xmm14, 0x1E
+ movdqa xmm13, xmm6
+ movdqa xmm14, xmm5
+ movdqa xmm5, xmmword ptr [rsp+0x20]
+ movdqa xmm6, xmmword ptr [rsp+0x40]
+ jmp 9b
+9:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ pxor xmm8, xmm10
+ pxor xmm9, xmm11
+ mov eax, r13d
+ cmp rdx, r15
+ jne 2b
+ movups xmmword ptr [rbx], xmm0
+ movups xmmword ptr [rbx+0x10], xmm1
+ movups xmmword ptr [rbx+0x20], xmm8
+ movups xmmword ptr [rbx+0x30], xmm9
+ mov eax, dword ptr [rsp+0x130]
+ neg eax
+ mov r10d, dword ptr [rsp+0x110+8*rax]
+ mov r11d, dword ptr [rsp+0x120+8*rax]
+ mov dword ptr [rsp+0x110], r10d
+ mov dword ptr [rsp+0x120], r11d
+ add rdi, 16
+ add rbx, 64
+ sub rsi, 2
+3:
+ test esi, 0x1
+ je 4b
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+0x10]
+ movd xmm13, dword ptr [rsp+0x110]
+ movd xmm14, dword ptr [rsp+0x120]
+ punpckldq xmm13, xmm14
+ mov r8, qword ptr [rdi]
+ movzx eax, byte ptr [rbp+0x80]
+ or eax, r13d
+ xor edx, edx
+2:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ shl rax, 32
+ or rax, 64
+ movd xmm12, rax
+ movdqa xmm3, xmm13
+ punpcklqdq xmm3, xmm12
+ movups xmm4, xmmword ptr [r8+rdx-0x40]
+ movups xmm5, xmmword ptr [r8+rdx-0x30]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [r8+rdx-0x20]
+ movups xmm7, xmmword ptr [r8+rdx-0x10]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 0x93
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ dec al
+ jz 9f
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0x0F
+ pshufd xmm4, xmm8, 0x39
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp 9b
+9:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ mov eax, r13d
+ cmp rdx, r15
+ jne 2b
+ movups xmmword ptr [rbx], xmm0
+ movups xmmword ptr [rbx+0x10], xmm1
+ jmp 4b
+
+.p2align 6
+blake3_compress_in_place_sse2:
+_blake3_compress_in_place_sse2:
+ sub rsp, 120
+ movdqa xmmword ptr [rsp], xmm6
+ movdqa xmmword ptr [rsp+0x10], xmm7
+ movdqa xmmword ptr [rsp+0x20], xmm8
+ movdqa xmmword ptr [rsp+0x30], xmm9
+ movdqa xmmword ptr [rsp+0x40], xmm11
+ movdqa xmmword ptr [rsp+0x50], xmm14
+ movdqa xmmword ptr [rsp+0x60], xmm15
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+0x10]
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ movzx eax, byte ptr [rsp+0xA0]
+ movzx r8d, r8b
+ shl rax, 32
+ add r8, rax
+ movq xmm3, r9
+ movq xmm4, r8
+ punpcklqdq xmm3, xmm4
+ movups xmm4, xmmword ptr [rdx]
+ movups xmm5, xmmword ptr [rdx+0x10]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [rdx+0x20]
+ movups xmm7, xmmword ptr [rdx+0x30]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 0x93
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ dec al
+ jz 9f
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0x0F
+ pshufd xmm4, xmm8, 0x39
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp 9b
+9:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ movups xmmword ptr [rcx], xmm0
+ movups xmmword ptr [rcx+0x10], xmm1
+ movdqa xmm6, xmmword ptr [rsp]
+ movdqa xmm7, xmmword ptr [rsp+0x10]
+ movdqa xmm8, xmmword ptr [rsp+0x20]
+ movdqa xmm9, xmmword ptr [rsp+0x30]
+ movdqa xmm11, xmmword ptr [rsp+0x40]
+ movdqa xmm14, xmmword ptr [rsp+0x50]
+ movdqa xmm15, xmmword ptr [rsp+0x60]
+ add rsp, 120
+ ret
+
+
+.p2align 6
+_blake3_compress_xof_sse2:
+blake3_compress_xof_sse2:
+ sub rsp, 120
+ movdqa xmmword ptr [rsp], xmm6
+ movdqa xmmword ptr [rsp+0x10], xmm7
+ movdqa xmmword ptr [rsp+0x20], xmm8
+ movdqa xmmword ptr [rsp+0x30], xmm9
+ movdqa xmmword ptr [rsp+0x40], xmm11
+ movdqa xmmword ptr [rsp+0x50], xmm14
+ movdqa xmmword ptr [rsp+0x60], xmm15
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+0x10]
+ movaps xmm2, xmmword ptr [BLAKE3_IV+rip]
+ movzx eax, byte ptr [rsp+0xA0]
+ movzx r8d, r8b
+ mov r10, qword ptr [rsp+0xA8]
+ shl rax, 32
+ add r8, rax
+ movq xmm3, r9
+ movq xmm4, r8
+ punpcklqdq xmm3, xmm4
+ movups xmm4, xmmword ptr [rdx]
+ movups xmm5, xmmword ptr [rdx+0x10]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [rdx+0x20]
+ movups xmm7, xmmword ptr [rdx+0x30]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 0x93
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 0x93
+ mov al, 7
+9:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x93
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x39
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0xB1
+ pshufhw xmm3, xmm3, 0xB1
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 0x39
+ pshufd xmm3, xmm3, 0x4E
+ pshufd xmm2, xmm2, 0x93
+ dec al
+ jz 9f
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0x0F
+ pshufd xmm4, xmm8, 0x39
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK+rip]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK+rip]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK+rip]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK+rip]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 0x78
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 0x1E
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp 9b
+9:
+ movdqu xmm4, xmmword ptr [rcx]
+ movdqu xmm5, xmmword ptr [rcx+0x10]
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ pxor xmm2, xmm4
+ pxor xmm3, xmm5
+ movups xmmword ptr [r10], xmm0
+ movups xmmword ptr [r10+0x10], xmm1
+ movups xmmword ptr [r10+0x20], xmm2
+ movups xmmword ptr [r10+0x30], xmm3
+ movdqa xmm6, xmmword ptr [rsp]
+ movdqa xmm7, xmmword ptr [rsp+0x10]
+ movdqa xmm8, xmmword ptr [rsp+0x20]
+ movdqa xmm9, xmmword ptr [rsp+0x30]
+ movdqa xmm11, xmmword ptr [rsp+0x40]
+ movdqa xmm14, xmmword ptr [rsp+0x50]
+ movdqa xmm15, xmmword ptr [rsp+0x60]
+ add rsp, 120
+ ret
+
+
+.section .rodata
+.p2align 6
+BLAKE3_IV:
+ .long 0x6A09E667, 0xBB67AE85
+ .long 0x3C6EF372, 0xA54FF53A
+ADD0:
+ .long 0, 1, 2, 3
+ADD1:
+ .long 4, 4, 4, 4
+BLAKE3_IV_0:
+ .long 0x6A09E667, 0x6A09E667, 0x6A09E667, 0x6A09E667
+BLAKE3_IV_1:
+ .long 0xBB67AE85, 0xBB67AE85, 0xBB67AE85, 0xBB67AE85
+BLAKE3_IV_2:
+ .long 0x3C6EF372, 0x3C6EF372, 0x3C6EF372, 0x3C6EF372
+BLAKE3_IV_3:
+ .long 0xA54FF53A, 0xA54FF53A, 0xA54FF53A, 0xA54FF53A
+BLAKE3_BLOCK_LEN:
+ .long 64, 64, 64, 64
+CMP_MSB_MASK:
+ .long 0x80000000, 0x80000000, 0x80000000, 0x80000000
+PBLENDW_0x33_MASK:
+ .long 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF, 0x00000000
+PBLENDW_0xCC_MASK:
+ .long 0x00000000, 0xFFFFFFFF, 0x00000000, 0xFFFFFFFF
+PBLENDW_0x3F_MASK:
+ .long 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000
+PBLENDW_0xC0_MASK:
+ .long 0x00000000, 0x00000000, 0x00000000, 0xFFFFFFFF
--- /dev/null
+public _blake3_hash_many_sse2
+public blake3_hash_many_sse2
+public blake3_compress_in_place_sse2
+public _blake3_compress_in_place_sse2
+public blake3_compress_xof_sse2
+public _blake3_compress_xof_sse2
+
+_TEXT SEGMENT ALIGN(16) 'CODE'
+
+ALIGN 16
+blake3_hash_many_sse2 PROC
+_blake3_hash_many_sse2 PROC
+ push r15
+ push r14
+ push r13
+ push r12
+ push rsi
+ push rdi
+ push rbx
+ push rbp
+ mov rbp, rsp
+ sub rsp, 528
+ and rsp, 0FFFFFFFFFFFFFFC0H
+ movdqa xmmword ptr [rsp+170H], xmm6
+ movdqa xmmword ptr [rsp+180H], xmm7
+ movdqa xmmword ptr [rsp+190H], xmm8
+ movdqa xmmword ptr [rsp+1A0H], xmm9
+ movdqa xmmword ptr [rsp+1B0H], xmm10
+ movdqa xmmword ptr [rsp+1C0H], xmm11
+ movdqa xmmword ptr [rsp+1D0H], xmm12
+ movdqa xmmword ptr [rsp+1E0H], xmm13
+ movdqa xmmword ptr [rsp+1F0H], xmm14
+ movdqa xmmword ptr [rsp+200H], xmm15
+ mov rdi, rcx
+ mov rsi, rdx
+ mov rdx, r8
+ mov rcx, r9
+ mov r8, qword ptr [rbp+68H]
+ movzx r9, byte ptr [rbp+70H]
+ neg r9d
+ movd xmm0, r9d
+ pshufd xmm0, xmm0, 00H
+ movdqa xmmword ptr [rsp+130H], xmm0
+ movdqa xmm1, xmm0
+ pand xmm1, xmmword ptr [ADD0]
+ pand xmm0, xmmword ptr [ADD1]
+ movdqa xmmword ptr [rsp+150H], xmm0
+ movd xmm0, r8d
+ pshufd xmm0, xmm0, 00H
+ paddd xmm0, xmm1
+ movdqa xmmword ptr [rsp+110H], xmm0
+ pxor xmm0, xmmword ptr [CMP_MSB_MASK]
+ pxor xmm1, xmmword ptr [CMP_MSB_MASK]
+ pcmpgtd xmm1, xmm0
+ shr r8, 32
+ movd xmm2, r8d
+ pshufd xmm2, xmm2, 00H
+ psubd xmm2, xmm1
+ movdqa xmmword ptr [rsp+120H], xmm2
+ mov rbx, qword ptr [rbp+90H]
+ mov r15, rdx
+ shl r15, 6
+ movzx r13d, byte ptr [rbp+78H]
+ movzx r12d, byte ptr [rbp+88H]
+ cmp rsi, 4
+ jc final3blocks
+outerloop4:
+ movdqu xmm3, xmmword ptr [rcx]
+ pshufd xmm0, xmm3, 00H
+ pshufd xmm1, xmm3, 55H
+ pshufd xmm2, xmm3, 0AAH
+ pshufd xmm3, xmm3, 0FFH
+ movdqu xmm7, xmmword ptr [rcx+10H]
+ pshufd xmm4, xmm7, 00H
+ pshufd xmm5, xmm7, 55H
+ pshufd xmm6, xmm7, 0AAH
+ pshufd xmm7, xmm7, 0FFH
+ mov r8, qword ptr [rdi]
+ mov r9, qword ptr [rdi+8H]
+ mov r10, qword ptr [rdi+10H]
+ mov r11, qword ptr [rdi+18H]
+ movzx eax, byte ptr [rbp+80H]
+ or eax, r13d
+ xor edx, edx
+innerloop4:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movdqu xmm8, xmmword ptr [r8+rdx-40H]
+ movdqu xmm9, xmmword ptr [r9+rdx-40H]
+ movdqu xmm10, xmmword ptr [r10+rdx-40H]
+ movdqu xmm11, xmmword ptr [r11+rdx-40H]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp], xmm8
+ movdqa xmmword ptr [rsp+10H], xmm9
+ movdqa xmmword ptr [rsp+20H], xmm12
+ movdqa xmmword ptr [rsp+30H], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-30H]
+ movdqu xmm9, xmmword ptr [r9+rdx-30H]
+ movdqu xmm10, xmmword ptr [r10+rdx-30H]
+ movdqu xmm11, xmmword ptr [r11+rdx-30H]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+40H], xmm8
+ movdqa xmmword ptr [rsp+50H], xmm9
+ movdqa xmmword ptr [rsp+60H], xmm12
+ movdqa xmmword ptr [rsp+70H], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-20H]
+ movdqu xmm9, xmmword ptr [r9+rdx-20H]
+ movdqu xmm10, xmmword ptr [r10+rdx-20H]
+ movdqu xmm11, xmmword ptr [r11+rdx-20H]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+80H], xmm8
+ movdqa xmmword ptr [rsp+90H], xmm9
+ movdqa xmmword ptr [rsp+0A0H], xmm12
+ movdqa xmmword ptr [rsp+0B0H], xmm13
+ movdqu xmm8, xmmword ptr [r8+rdx-10H]
+ movdqu xmm9, xmmword ptr [r9+rdx-10H]
+ movdqu xmm10, xmmword ptr [r10+rdx-10H]
+ movdqu xmm11, xmmword ptr [r11+rdx-10H]
+ movdqa xmm12, xmm8
+ punpckldq xmm8, xmm9
+ punpckhdq xmm12, xmm9
+ movdqa xmm14, xmm10
+ punpckldq xmm10, xmm11
+ punpckhdq xmm14, xmm11
+ movdqa xmm9, xmm8
+ punpcklqdq xmm8, xmm10
+ punpckhqdq xmm9, xmm10
+ movdqa xmm13, xmm12
+ punpcklqdq xmm12, xmm14
+ punpckhqdq xmm13, xmm14
+ movdqa xmmword ptr [rsp+0C0H], xmm8
+ movdqa xmmword ptr [rsp+0D0H], xmm9
+ movdqa xmmword ptr [rsp+0E0H], xmm12
+ movdqa xmmword ptr [rsp+0F0H], xmm13
+ movdqa xmm9, xmmword ptr [BLAKE3_IV_1]
+ movdqa xmm10, xmmword ptr [BLAKE3_IV_2]
+ movdqa xmm11, xmmword ptr [BLAKE3_IV_3]
+ movdqa xmm12, xmmword ptr [rsp+110H]
+ movdqa xmm13, xmmword ptr [rsp+120H]
+ movdqa xmm14, xmmword ptr [BLAKE3_BLOCK_LEN]
+ movd xmm15, eax
+ pshufd xmm15, xmm15, 00H
+ prefetcht0 byte ptr [r8+rdx+80H]
+ prefetcht0 byte ptr [r9+rdx+80H]
+ prefetcht0 byte ptr [r10+rdx+80H]
+ prefetcht0 byte ptr [r11+rdx+80H]
+ paddd xmm0, xmmword ptr [rsp]
+ paddd xmm1, xmmword ptr [rsp+20H]
+ paddd xmm2, xmmword ptr [rsp+40H]
+ paddd xmm3, xmmword ptr [rsp+60H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ movdqa xmm8, xmmword ptr [BLAKE3_IV_0]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+10H]
+ paddd xmm1, xmmword ptr [rsp+30H]
+ paddd xmm2, xmmword ptr [rsp+50H]
+ paddd xmm3, xmmword ptr [rsp+70H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+80H]
+ paddd xmm1, xmmword ptr [rsp+0A0H]
+ paddd xmm2, xmmword ptr [rsp+0C0H]
+ paddd xmm3, xmmword ptr [rsp+0E0H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+90H]
+ paddd xmm1, xmmword ptr [rsp+0B0H]
+ paddd xmm2, xmmword ptr [rsp+0D0H]
+ paddd xmm3, xmmword ptr [rsp+0F0H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+20H]
+ paddd xmm1, xmmword ptr [rsp+30H]
+ paddd xmm2, xmmword ptr [rsp+70H]
+ paddd xmm3, xmmword ptr [rsp+40H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+60H]
+ paddd xmm1, xmmword ptr [rsp+0A0H]
+ paddd xmm2, xmmword ptr [rsp]
+ paddd xmm3, xmmword ptr [rsp+0D0H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+10H]
+ paddd xmm1, xmmword ptr [rsp+0C0H]
+ paddd xmm2, xmmword ptr [rsp+90H]
+ paddd xmm3, xmmword ptr [rsp+0F0H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0B0H]
+ paddd xmm1, xmmword ptr [rsp+50H]
+ paddd xmm2, xmmword ptr [rsp+0E0H]
+ paddd xmm3, xmmword ptr [rsp+80H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+30H]
+ paddd xmm1, xmmword ptr [rsp+0A0H]
+ paddd xmm2, xmmword ptr [rsp+0D0H]
+ paddd xmm3, xmmword ptr [rsp+70H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+40H]
+ paddd xmm1, xmmword ptr [rsp+0C0H]
+ paddd xmm2, xmmword ptr [rsp+20H]
+ paddd xmm3, xmmword ptr [rsp+0E0H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+60H]
+ paddd xmm1, xmmword ptr [rsp+90H]
+ paddd xmm2, xmmword ptr [rsp+0B0H]
+ paddd xmm3, xmmword ptr [rsp+80H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+50H]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+0F0H]
+ paddd xmm3, xmmword ptr [rsp+10H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0A0H]
+ paddd xmm1, xmmword ptr [rsp+0C0H]
+ paddd xmm2, xmmword ptr [rsp+0E0H]
+ paddd xmm3, xmmword ptr [rsp+0D0H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+70H]
+ paddd xmm1, xmmword ptr [rsp+90H]
+ paddd xmm2, xmmword ptr [rsp+30H]
+ paddd xmm3, xmmword ptr [rsp+0F0H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+40H]
+ paddd xmm1, xmmword ptr [rsp+0B0H]
+ paddd xmm2, xmmword ptr [rsp+50H]
+ paddd xmm3, xmmword ptr [rsp+10H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp]
+ paddd xmm1, xmmword ptr [rsp+20H]
+ paddd xmm2, xmmword ptr [rsp+80H]
+ paddd xmm3, xmmword ptr [rsp+60H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0C0H]
+ paddd xmm1, xmmword ptr [rsp+90H]
+ paddd xmm2, xmmword ptr [rsp+0F0H]
+ paddd xmm3, xmmword ptr [rsp+0E0H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0D0H]
+ paddd xmm1, xmmword ptr [rsp+0B0H]
+ paddd xmm2, xmmword ptr [rsp+0A0H]
+ paddd xmm3, xmmword ptr [rsp+80H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+70H]
+ paddd xmm1, xmmword ptr [rsp+50H]
+ paddd xmm2, xmmword ptr [rsp]
+ paddd xmm3, xmmword ptr [rsp+60H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+20H]
+ paddd xmm1, xmmword ptr [rsp+30H]
+ paddd xmm2, xmmword ptr [rsp+10H]
+ paddd xmm3, xmmword ptr [rsp+40H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+90H]
+ paddd xmm1, xmmword ptr [rsp+0B0H]
+ paddd xmm2, xmmword ptr [rsp+80H]
+ paddd xmm3, xmmword ptr [rsp+0F0H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0E0H]
+ paddd xmm1, xmmword ptr [rsp+50H]
+ paddd xmm2, xmmword ptr [rsp+0C0H]
+ paddd xmm3, xmmword ptr [rsp+10H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0D0H]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+20H]
+ paddd xmm3, xmmword ptr [rsp+40H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+30H]
+ paddd xmm1, xmmword ptr [rsp+0A0H]
+ paddd xmm2, xmmword ptr [rsp+60H]
+ paddd xmm3, xmmword ptr [rsp+70H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0B0H]
+ paddd xmm1, xmmword ptr [rsp+50H]
+ paddd xmm2, xmmword ptr [rsp+10H]
+ paddd xmm3, xmmword ptr [rsp+80H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0F0H]
+ paddd xmm1, xmmword ptr [rsp]
+ paddd xmm2, xmmword ptr [rsp+90H]
+ paddd xmm3, xmmword ptr [rsp+60H]
+ paddd xmm0, xmm4
+ paddd xmm1, xmm5
+ paddd xmm2, xmm6
+ paddd xmm3, xmm7
+ pxor xmm12, xmm0
+ pxor xmm13, xmm1
+ pxor xmm14, xmm2
+ pxor xmm15, xmm3
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm12
+ paddd xmm9, xmm13
+ paddd xmm10, xmm14
+ paddd xmm11, xmm15
+ pxor xmm4, xmm8
+ pxor xmm5, xmm9
+ pxor xmm6, xmm10
+ pxor xmm7, xmm11
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ paddd xmm0, xmmword ptr [rsp+0E0H]
+ paddd xmm1, xmmword ptr [rsp+20H]
+ paddd xmm2, xmmword ptr [rsp+30H]
+ paddd xmm3, xmmword ptr [rsp+70H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ pshuflw xmm15, xmm15, 0B1H
+ pshufhw xmm15, xmm15, 0B1H
+ pshuflw xmm12, xmm12, 0B1H
+ pshufhw xmm12, xmm12, 0B1H
+ pshuflw xmm13, xmm13, 0B1H
+ pshufhw xmm13, xmm13, 0B1H
+ pshuflw xmm14, xmm14, 0B1H
+ pshufhw xmm14, xmm14, 0B1H
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ movdqa xmmword ptr [rsp+100H], xmm8
+ movdqa xmm8, xmm5
+ psrld xmm8, 12
+ pslld xmm5, 20
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 12
+ pslld xmm6, 20
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 12
+ pslld xmm7, 20
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 12
+ pslld xmm4, 20
+ por xmm4, xmm8
+ paddd xmm0, xmmword ptr [rsp+0A0H]
+ paddd xmm1, xmmword ptr [rsp+0C0H]
+ paddd xmm2, xmmword ptr [rsp+40H]
+ paddd xmm3, xmmword ptr [rsp+0D0H]
+ paddd xmm0, xmm5
+ paddd xmm1, xmm6
+ paddd xmm2, xmm7
+ paddd xmm3, xmm4
+ pxor xmm15, xmm0
+ pxor xmm12, xmm1
+ pxor xmm13, xmm2
+ pxor xmm14, xmm3
+ movdqa xmm8, xmm15
+ psrld xmm15, 8
+ pslld xmm8, 24
+ pxor xmm15, xmm8
+ movdqa xmm8, xmm12
+ psrld xmm12, 8
+ pslld xmm8, 24
+ pxor xmm12, xmm8
+ movdqa xmm8, xmm13
+ psrld xmm13, 8
+ pslld xmm8, 24
+ pxor xmm13, xmm8
+ movdqa xmm8, xmm14
+ psrld xmm14, 8
+ pslld xmm8, 24
+ pxor xmm14, xmm8
+ paddd xmm10, xmm15
+ paddd xmm11, xmm12
+ movdqa xmm8, xmmword ptr [rsp+100H]
+ paddd xmm8, xmm13
+ paddd xmm9, xmm14
+ pxor xmm5, xmm10
+ pxor xmm6, xmm11
+ pxor xmm7, xmm8
+ pxor xmm4, xmm9
+ pxor xmm0, xmm8
+ pxor xmm1, xmm9
+ pxor xmm2, xmm10
+ pxor xmm3, xmm11
+ movdqa xmm8, xmm5
+ psrld xmm8, 7
+ pslld xmm5, 25
+ por xmm5, xmm8
+ movdqa xmm8, xmm6
+ psrld xmm8, 7
+ pslld xmm6, 25
+ por xmm6, xmm8
+ movdqa xmm8, xmm7
+ psrld xmm8, 7
+ pslld xmm7, 25
+ por xmm7, xmm8
+ movdqa xmm8, xmm4
+ psrld xmm8, 7
+ pslld xmm4, 25
+ por xmm4, xmm8
+ pxor xmm4, xmm12
+ pxor xmm5, xmm13
+ pxor xmm6, xmm14
+ pxor xmm7, xmm15
+ mov eax, r13d
+ jne innerloop4
+ movdqa xmm9, xmm0
+ punpckldq xmm0, xmm1
+ punpckhdq xmm9, xmm1
+ movdqa xmm11, xmm2
+ punpckldq xmm2, xmm3
+ punpckhdq xmm11, xmm3
+ movdqa xmm1, xmm0
+ punpcklqdq xmm0, xmm2
+ punpckhqdq xmm1, xmm2
+ movdqa xmm3, xmm9
+ punpcklqdq xmm9, xmm11
+ punpckhqdq xmm3, xmm11
+ movdqu xmmword ptr [rbx], xmm0
+ movdqu xmmword ptr [rbx+20H], xmm1
+ movdqu xmmword ptr [rbx+40H], xmm9
+ movdqu xmmword ptr [rbx+60H], xmm3
+ movdqa xmm9, xmm4
+ punpckldq xmm4, xmm5
+ punpckhdq xmm9, xmm5
+ movdqa xmm11, xmm6
+ punpckldq xmm6, xmm7
+ punpckhdq xmm11, xmm7
+ movdqa xmm5, xmm4
+ punpcklqdq xmm4, xmm6
+ punpckhqdq xmm5, xmm6
+ movdqa xmm7, xmm9
+ punpcklqdq xmm9, xmm11
+ punpckhqdq xmm7, xmm11
+ movdqu xmmword ptr [rbx+10H], xmm4
+ movdqu xmmword ptr [rbx+30H], xmm5
+ movdqu xmmword ptr [rbx+50H], xmm9
+ movdqu xmmword ptr [rbx+70H], xmm7
+ movdqa xmm1, xmmword ptr [rsp+110H]
+ movdqa xmm0, xmm1
+ paddd xmm1, xmmword ptr [rsp+150H]
+ movdqa xmmword ptr [rsp+110H], xmm1
+ pxor xmm0, xmmword ptr [CMP_MSB_MASK]
+ pxor xmm1, xmmword ptr [CMP_MSB_MASK]
+ pcmpgtd xmm0, xmm1
+ movdqa xmm1, xmmword ptr [rsp+120H]
+ psubd xmm1, xmm0
+ movdqa xmmword ptr [rsp+120H], xmm1
+ add rbx, 128
+ add rdi, 32
+ sub rsi, 4
+ cmp rsi, 4
+ jnc outerloop4
+ test rsi, rsi
+ jne final3blocks
+unwind:
+ movdqa xmm6, xmmword ptr [rsp+170H]
+ movdqa xmm7, xmmword ptr [rsp+180H]
+ movdqa xmm8, xmmword ptr [rsp+190H]
+ movdqa xmm9, xmmword ptr [rsp+1A0H]
+ movdqa xmm10, xmmword ptr [rsp+1B0H]
+ movdqa xmm11, xmmword ptr [rsp+1C0H]
+ movdqa xmm12, xmmword ptr [rsp+1D0H]
+ movdqa xmm13, xmmword ptr [rsp+1E0H]
+ movdqa xmm14, xmmword ptr [rsp+1F0H]
+ movdqa xmm15, xmmword ptr [rsp+200H]
+ mov rsp, rbp
+ pop rbp
+ pop rbx
+ pop rdi
+ pop rsi
+ pop r12
+ pop r13
+ pop r14
+ pop r15
+ ret
+ALIGN 16
+final3blocks:
+ test esi, 2H
+ je final1block
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+10H]
+ movaps xmm8, xmm0
+ movaps xmm9, xmm1
+ movd xmm13, dword ptr [rsp+110H]
+ movd xmm14, dword ptr [rsp+120H]
+ punpckldq xmm13, xmm14
+ movaps xmmword ptr [rsp], xmm13
+ movd xmm14, dword ptr [rsp+114H]
+ movd xmm13, dword ptr [rsp+124H]
+ punpckldq xmm14, xmm13
+ movaps xmmword ptr [rsp+10H], xmm14
+ mov r8, qword ptr [rdi]
+ mov r9, qword ptr [rdi+8H]
+ movzx eax, byte ptr [rbp+80H]
+ or eax, r13d
+ xor edx, edx
+innerloop2:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movaps xmm2, xmmword ptr [BLAKE3_IV]
+ movaps xmm10, xmm2
+ movups xmm4, xmmword ptr [r8+rdx-40H]
+ movups xmm5, xmmword ptr [r8+rdx-30H]
+ movaps xmm3, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm3, xmm5, 221
+ movaps xmm5, xmm3
+ movups xmm6, xmmword ptr [r8+rdx-20H]
+ movups xmm7, xmmword ptr [r8+rdx-10H]
+ movaps xmm3, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 93H
+ shufps xmm3, xmm7, 221
+ pshufd xmm7, xmm3, 93H
+ movups xmm12, xmmword ptr [r9+rdx-40H]
+ movups xmm13, xmmword ptr [r9+rdx-30H]
+ movaps xmm11, xmm12
+ shufps xmm12, xmm13, 136
+ shufps xmm11, xmm13, 221
+ movaps xmm13, xmm11
+ movups xmm14, xmmword ptr [r9+rdx-20H]
+ movups xmm15, xmmword ptr [r9+rdx-10H]
+ movaps xmm11, xmm14
+ shufps xmm14, xmm15, 136
+ pshufd xmm14, xmm14, 93H
+ shufps xmm11, xmm15, 221
+ pshufd xmm15, xmm11, 93H
+ shl rax, 20H
+ or rax, 40H
+ movd xmm3, rax
+ movdqa xmmword ptr [rsp+20H], xmm3
+ movaps xmm3, xmmword ptr [rsp]
+ movaps xmm11, xmmword ptr [rsp+10H]
+ punpcklqdq xmm3, xmmword ptr [rsp+20H]
+ punpcklqdq xmm11, xmmword ptr [rsp+20H]
+ mov al, 7
+roundloop2:
+ paddd xmm0, xmm4
+ paddd xmm8, xmm12
+ movaps xmmword ptr [rsp+20H], xmm4
+ movaps xmmword ptr [rsp+30H], xmm12
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ pshuflw xmm11, xmm11, 0B1H
+ pshufhw xmm11, xmm11, 0B1H
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 20
+ psrld xmm4, 12
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 20
+ psrld xmm4, 12
+ por xmm9, xmm4
+ paddd xmm0, xmm5
+ paddd xmm8, xmm13
+ movaps xmmword ptr [rsp+40H], xmm5
+ movaps xmmword ptr [rsp+50H], xmm13
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ movdqa xmm13, xmm3
+ psrld xmm3, 8
+ pslld xmm13, 24
+ pxor xmm3, xmm13
+ movdqa xmm13, xmm11
+ psrld xmm11, 8
+ pslld xmm13, 24
+ pxor xmm11, xmm13
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 25
+ psrld xmm4, 7
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 25
+ psrld xmm4, 7
+ por xmm9, xmm4
+ pshufd xmm0, xmm0, 93H
+ pshufd xmm8, xmm8, 93H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm11, xmm11, 4EH
+ pshufd xmm2, xmm2, 39H
+ pshufd xmm10, xmm10, 39H
+ paddd xmm0, xmm6
+ paddd xmm8, xmm14
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ pshuflw xmm11, xmm11, 0B1H
+ pshufhw xmm11, xmm11, 0B1H
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 20
+ psrld xmm4, 12
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 20
+ psrld xmm4, 12
+ por xmm9, xmm4
+ paddd xmm0, xmm7
+ paddd xmm8, xmm15
+ paddd xmm0, xmm1
+ paddd xmm8, xmm9
+ pxor xmm3, xmm0
+ pxor xmm11, xmm8
+ movdqa xmm13, xmm3
+ psrld xmm3, 8
+ pslld xmm13, 24
+ pxor xmm3, xmm13
+ movdqa xmm13, xmm11
+ psrld xmm11, 8
+ pslld xmm13, 24
+ pxor xmm11, xmm13
+ paddd xmm2, xmm3
+ paddd xmm10, xmm11
+ pxor xmm1, xmm2
+ pxor xmm9, xmm10
+ movdqa xmm4, xmm1
+ pslld xmm1, 25
+ psrld xmm4, 7
+ por xmm1, xmm4
+ movdqa xmm4, xmm9
+ pslld xmm9, 25
+ psrld xmm4, 7
+ por xmm9, xmm4
+ pshufd xmm0, xmm0, 39H
+ pshufd xmm8, xmm8, 39H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm11, xmm11, 4EH
+ pshufd xmm2, xmm2, 93H
+ pshufd xmm10, xmm10, 93H
+ dec al
+ je endroundloop2
+ movdqa xmm12, xmmword ptr [rsp+20H]
+ movdqa xmm5, xmmword ptr [rsp+40H]
+ pshufd xmm13, xmm12, 0FH
+ shufps xmm12, xmm5, 214
+ pshufd xmm4, xmm12, 39H
+ movdqa xmm12, xmm6
+ shufps xmm12, xmm7, 250
+ pand xmm13, xmmword ptr [PBLENDW_0x33_MASK]
+ pand xmm12, xmmword ptr [PBLENDW_0xCC_MASK]
+ por xmm13, xmm12
+ movdqa xmmword ptr [rsp+20H], xmm13
+ movdqa xmm12, xmm7
+ punpcklqdq xmm12, xmm5
+ movdqa xmm13, xmm6
+ pand xmm12, xmmword ptr [PBLENDW_0x3F_MASK]
+ pand xmm13, xmmword ptr [PBLENDW_0xC0_MASK]
+ por xmm12, xmm13
+ pshufd xmm12, xmm12, 78H
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 1EH
+ movdqa xmmword ptr [rsp+40H], xmm12
+ movdqa xmm5, xmmword ptr [rsp+30H]
+ movdqa xmm13, xmmword ptr [rsp+50H]
+ pshufd xmm6, xmm5, 0FH
+ shufps xmm5, xmm13, 214
+ pshufd xmm12, xmm5, 39H
+ movdqa xmm5, xmm14
+ shufps xmm5, xmm15, 250
+ pand xmm6, xmmword ptr [PBLENDW_0x33_MASK]
+ pand xmm5, xmmword ptr [PBLENDW_0xCC_MASK]
+ por xmm6, xmm5
+ movdqa xmm5, xmm15
+ punpcklqdq xmm5, xmm13
+ movdqa xmmword ptr [rsp+30H], xmm2
+ movdqa xmm2, xmm14
+ pand xmm5, xmmword ptr [PBLENDW_0x3F_MASK]
+ pand xmm2, xmmword ptr [PBLENDW_0xC0_MASK]
+ por xmm5, xmm2
+ movdqa xmm2, xmmword ptr [rsp+30H]
+ pshufd xmm5, xmm5, 78H
+ punpckhdq xmm13, xmm15
+ punpckldq xmm14, xmm13
+ pshufd xmm15, xmm14, 1EH
+ movdqa xmm13, xmm6
+ movdqa xmm14, xmm5
+ movdqa xmm5, xmmword ptr [rsp+20H]
+ movdqa xmm6, xmmword ptr [rsp+40H]
+ jmp roundloop2
+endroundloop2:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ pxor xmm8, xmm10
+ pxor xmm9, xmm11
+ mov eax, r13d
+ cmp rdx, r15
+ jne innerloop2
+ movups xmmword ptr [rbx], xmm0
+ movups xmmword ptr [rbx+10H], xmm1
+ movups xmmword ptr [rbx+20H], xmm8
+ movups xmmword ptr [rbx+30H], xmm9
+ mov eax, dword ptr [rsp+130H]
+ neg eax
+ mov r10d, dword ptr [rsp+110H+8*rax]
+ mov r11d, dword ptr [rsp+120H+8*rax]
+ mov dword ptr [rsp+110H], r10d
+ mov dword ptr [rsp+120H], r11d
+ add rdi, 16
+ add rbx, 64
+ sub rsi, 2
+final1block:
+ test esi, 1H
+ je unwind
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+10H]
+ movd xmm13, dword ptr [rsp+110H]
+ movd xmm14, dword ptr [rsp+120H]
+ punpckldq xmm13, xmm14
+ mov r8, qword ptr [rdi]
+ movzx eax, byte ptr [rbp+80H]
+ or eax, r13d
+ xor edx, edx
+innerloop1:
+ mov r14d, eax
+ or eax, r12d
+ add rdx, 64
+ cmp rdx, r15
+ cmovne eax, r14d
+ movaps xmm2, xmmword ptr [BLAKE3_IV]
+ shl rax, 32
+ or rax, 64
+ movd xmm12, rax
+ movdqa xmm3, xmm13
+ punpcklqdq xmm3, xmm12
+ movups xmm4, xmmword ptr [r8+rdx-40H]
+ movups xmm5, xmmword ptr [r8+rdx-30H]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [r8+rdx-20H]
+ movups xmm7, xmmword ptr [r8+rdx-10H]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 93H
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 93H
+ mov al, 7
+roundloop1:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 93H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm2, xmm2, 39H
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 39H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm2, xmm2, 93H
+ dec al
+ jz endroundloop1
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0FH
+ pshufd xmm4, xmm8, 39H
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 78H
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 1EH
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp roundloop1
+endroundloop1:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ mov eax, r13d
+ cmp rdx, r15
+ jne innerloop1
+ movups xmmword ptr [rbx], xmm0
+ movups xmmword ptr [rbx+10H], xmm1
+ jmp unwind
+_blake3_hash_many_sse2 ENDP
+blake3_hash_many_sse2 ENDP
+
+blake3_compress_in_place_sse2 PROC
+_blake3_compress_in_place_sse2 PROC
+ sub rsp, 120
+ movdqa xmmword ptr [rsp], xmm6
+ movdqa xmmword ptr [rsp+10H], xmm7
+ movdqa xmmword ptr [rsp+20H], xmm8
+ movdqa xmmword ptr [rsp+30H], xmm9
+ movdqa xmmword ptr [rsp+40H], xmm11
+ movdqa xmmword ptr [rsp+50H], xmm14
+ movdqa xmmword ptr [rsp+60H], xmm15
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+10H]
+ movaps xmm2, xmmword ptr [BLAKE3_IV]
+ movzx eax, byte ptr [rsp+0A0H]
+ movzx r8d, r8b
+ shl rax, 32
+ add r8, rax
+ movq xmm3, r9
+ movq xmm4, r8
+ punpcklqdq xmm3, xmm4
+ movups xmm4, xmmword ptr [rdx]
+ movups xmm5, xmmword ptr [rdx+10H]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [rdx+20H]
+ movups xmm7, xmmword ptr [rdx+30H]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 93H
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 93H
+ mov al, 7
+@@:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 93H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm2, xmm2, 39H
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 39H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm2, xmm2, 93H
+ dec al
+ jz @F
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0FH
+ pshufd xmm4, xmm8, 39H
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 78H
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 1EH
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp @B
+@@:
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ movups xmmword ptr [rcx], xmm0
+ movups xmmword ptr [rcx+10H], xmm1
+ movdqa xmm6, xmmword ptr [rsp]
+ movdqa xmm7, xmmword ptr [rsp+10H]
+ movdqa xmm8, xmmword ptr [rsp+20H]
+ movdqa xmm9, xmmword ptr [rsp+30H]
+ movdqa xmm11, xmmword ptr [rsp+40H]
+ movdqa xmm14, xmmword ptr [rsp+50H]
+ movdqa xmm15, xmmword ptr [rsp+60H]
+ add rsp, 120
+ ret
+_blake3_compress_in_place_sse2 ENDP
+blake3_compress_in_place_sse2 ENDP
+
+ALIGN 16
+blake3_compress_xof_sse2 PROC
+_blake3_compress_xof_sse2 PROC
+ sub rsp, 120
+ movdqa xmmword ptr [rsp], xmm6
+ movdqa xmmword ptr [rsp+10H], xmm7
+ movdqa xmmword ptr [rsp+20H], xmm8
+ movdqa xmmword ptr [rsp+30H], xmm9
+ movdqa xmmword ptr [rsp+40H], xmm11
+ movdqa xmmword ptr [rsp+50H], xmm14
+ movdqa xmmword ptr [rsp+60H], xmm15
+ movups xmm0, xmmword ptr [rcx]
+ movups xmm1, xmmword ptr [rcx+10H]
+ movaps xmm2, xmmword ptr [BLAKE3_IV]
+ movzx eax, byte ptr [rsp+0A0H]
+ movzx r8d, r8b
+ mov r10, qword ptr [rsp+0A8H]
+ shl rax, 32
+ add r8, rax
+ movq xmm3, r9
+ movq xmm4, r8
+ punpcklqdq xmm3, xmm4
+ movups xmm4, xmmword ptr [rdx]
+ movups xmm5, xmmword ptr [rdx+10H]
+ movaps xmm8, xmm4
+ shufps xmm4, xmm5, 136
+ shufps xmm8, xmm5, 221
+ movaps xmm5, xmm8
+ movups xmm6, xmmword ptr [rdx+20H]
+ movups xmm7, xmmword ptr [rdx+30H]
+ movaps xmm8, xmm6
+ shufps xmm6, xmm7, 136
+ pshufd xmm6, xmm6, 93H
+ shufps xmm8, xmm7, 221
+ pshufd xmm7, xmm8, 93H
+ mov al, 7
+@@:
+ paddd xmm0, xmm4
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm5
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 93H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm2, xmm2, 39H
+ paddd xmm0, xmm6
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ pshuflw xmm3, xmm3, 0B1H
+ pshufhw xmm3, xmm3, 0B1H
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 20
+ psrld xmm11, 12
+ por xmm1, xmm11
+ paddd xmm0, xmm7
+ paddd xmm0, xmm1
+ pxor xmm3, xmm0
+ movdqa xmm14, xmm3
+ psrld xmm3, 8
+ pslld xmm14, 24
+ pxor xmm3, xmm14
+ paddd xmm2, xmm3
+ pxor xmm1, xmm2
+ movdqa xmm11, xmm1
+ pslld xmm1, 25
+ psrld xmm11, 7
+ por xmm1, xmm11
+ pshufd xmm0, xmm0, 39H
+ pshufd xmm3, xmm3, 4EH
+ pshufd xmm2, xmm2, 93H
+ dec al
+ jz @F
+ movdqa xmm8, xmm4
+ shufps xmm8, xmm5, 214
+ pshufd xmm9, xmm4, 0FH
+ pshufd xmm4, xmm8, 39H
+ movdqa xmm8, xmm6
+ shufps xmm8, xmm7, 250
+ pand xmm9, xmmword ptr [PBLENDW_0x33_MASK]
+ pand xmm8, xmmword ptr [PBLENDW_0xCC_MASK]
+ por xmm9, xmm8
+ movdqa xmm8, xmm7
+ punpcklqdq xmm8, xmm5
+ movdqa xmm10, xmm6
+ pand xmm8, xmmword ptr [PBLENDW_0x3F_MASK]
+ pand xmm10, xmmword ptr [PBLENDW_0xC0_MASK]
+ por xmm8, xmm10
+ pshufd xmm8, xmm8, 78H
+ punpckhdq xmm5, xmm7
+ punpckldq xmm6, xmm5
+ pshufd xmm7, xmm6, 1EH
+ movdqa xmm5, xmm9
+ movdqa xmm6, xmm8
+ jmp @B
+@@:
+ movdqu xmm4, xmmword ptr [rcx]
+ movdqu xmm5, xmmword ptr [rcx+10H]
+ pxor xmm0, xmm2
+ pxor xmm1, xmm3
+ pxor xmm2, xmm4
+ pxor xmm3, xmm5
+ movups xmmword ptr [r10], xmm0
+ movups xmmword ptr [r10+10H], xmm1
+ movups xmmword ptr [r10+20H], xmm2
+ movups xmmword ptr [r10+30H], xmm3
+ movdqa xmm6, xmmword ptr [rsp]
+ movdqa xmm7, xmmword ptr [rsp+10H]
+ movdqa xmm8, xmmword ptr [rsp+20H]
+ movdqa xmm9, xmmword ptr [rsp+30H]
+ movdqa xmm11, xmmword ptr [rsp+40H]
+ movdqa xmm14, xmmword ptr [rsp+50H]
+ movdqa xmm15, xmmword ptr [rsp+60H]
+ add rsp, 120
+ ret
+_blake3_compress_xof_sse2 ENDP
+blake3_compress_xof_sse2 ENDP
+
+_TEXT ENDS
+
+
+_RDATA SEGMENT READONLY PAGE ALIAS(".rdata") 'CONST'
+ALIGN 64
+BLAKE3_IV:
+ dd 6A09E667H, 0BB67AE85H, 3C6EF372H, 0A54FF53AH
+
+ADD0:
+ dd 0, 1, 2, 3
+
+ADD1:
+ dd 4 dup (4)
+
+BLAKE3_IV_0:
+ dd 4 dup (6A09E667H)
+
+BLAKE3_IV_1:
+ dd 4 dup (0BB67AE85H)
+
+BLAKE3_IV_2:
+ dd 4 dup (3C6EF372H)
+
+BLAKE3_IV_3:
+ dd 4 dup (0A54FF53AH)
+
+BLAKE3_BLOCK_LEN:
+ dd 4 dup (64)
+
+CMP_MSB_MASK:
+ dd 8 dup(80000000H)
+
+PBLENDW_0x33_MASK:
+ dd 0FFFFFFFFH, 000000000H, 0FFFFFFFFH, 000000000H
+PBLENDW_0xCC_MASK:
+ dd 000000000H, 0FFFFFFFFH, 000000000H, 0FFFFFFFFH
+PBLENDW_0x3F_MASK:
+ dd 0FFFFFFFFH, 0FFFFFFFFH, 0FFFFFFFFH, 000000000H
+PBLENDW_0xC0_MASK:
+ dd 000000000H, 000000000H, 000000000H, 0FFFFFFFFH
+
+_RDATA ENDS
+END
+#if defined(__ELF__) && defined(__linux__)
+.section .note.GNU-stack,"",%progbits
+#endif
+
#if defined(__ELF__) && defined(__CET__) && defined(__has_include)
#if __has_include(<cet.h>)
#include <cet.h>