rbnacl-libsodium 1.0.10 → 1.0.11
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGES.md +6 -2
- data/lib/rbnacl/libsodium/version.rb +1 -1
- data/vendor/libsodium/ChangeLog +24 -0
- data/vendor/libsodium/LICENSE +2 -0
- data/vendor/libsodium/Makefile.in +6 -1
- data/vendor/libsodium/aclocal.m4 +1 -0
- data/vendor/libsodium/autom4te.cache/output.1 +1896 -1136
- data/vendor/libsodium/autom4te.cache/output.7 +19837 -0
- data/vendor/libsodium/autom4te.cache/requests +1271 -1031
- data/vendor/libsodium/autom4te.cache/traces.1 +532 -491
- data/vendor/libsodium/autom4te.cache/traces.7 +3614 -0
- data/vendor/libsodium/builds/msvc/version.h +2 -2
- data/vendor/libsodium/builds/msvc/vs2010/libsodium/libsodium.vcxproj +22 -1
- data/vendor/libsodium/builds/msvc/vs2010/libsodium/libsodium.vcxproj.filters +23 -20
- data/vendor/libsodium/builds/msvc/vs2012/libsodium/libsodium.vcxproj +13 -2
- data/vendor/libsodium/builds/msvc/vs2012/libsodium/libsodium.vcxproj.filters +12 -9
- data/vendor/libsodium/builds/msvc/vs2013/libsodium/libsodium.vcxproj +12 -1
- data/vendor/libsodium/builds/msvc/vs2013/libsodium/libsodium.vcxproj.filters +3 -0
- data/vendor/libsodium/builds/msvc/vs2015/libsodium/libsodium.vcxproj +2 -1
- data/vendor/libsodium/builds/msvc/vs2015/libsodium/libsodium.vcxproj.filters +14 -11
- data/vendor/libsodium/configure +1887 -1127
- data/vendor/libsodium/configure.ac +44 -11
- data/vendor/libsodium/dist-build/Makefile.am +4 -2
- data/vendor/libsodium/dist-build/Makefile.in +9 -2
- data/vendor/libsodium/dist-build/android-armv8-a.sh +1 -1
- data/vendor/libsodium/dist-build/android-build.sh +40 -3
- data/vendor/libsodium/dist-build/android-mips64.sh +1 -1
- data/vendor/libsodium/dist-build/android-x86_64.sh +4 -0
- data/vendor/libsodium/dist-build/emscripten-symbols.def +25 -25
- data/vendor/libsodium/dist-build/emscripten.sh +69 -35
- data/vendor/libsodium/dist-build/msys2-win32.sh +8 -2
- data/vendor/libsodium/dist-build/msys2-win64.sh +8 -2
- data/vendor/libsodium/dist-build/{nativeclient.sh → nativeclient-pnacl.sh} +3 -4
- data/vendor/libsodium/dist-build/nativeclient-x86.sh +19 -0
- data/vendor/libsodium/dist-build/nativeclient-x86_64.sh +19 -0
- data/vendor/libsodium/logo.png +0 -0
- data/vendor/libsodium/m4/ax_pthread.m4 +485 -0
- data/vendor/libsodium/msvc-scripts/Makefile.in +5 -0
- data/vendor/libsodium/msvc-scripts/process.bat +2 -2
- data/vendor/libsodium/packaging/nuget/package.bat +3 -3
- data/vendor/libsodium/src/Makefile.in +5 -0
- data/vendor/libsodium/src/libsodium/Makefile.am +2 -1
- data/vendor/libsodium/src/libsodium/Makefile.in +13 -8
- data/vendor/libsodium/src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c +70 -87
- data/vendor/libsodium/src/libsodium/crypto_aead/chacha20poly1305/sodium/aead_chacha20poly1305.c +6 -0
- data/vendor/libsodium/src/libsodium/crypto_generichash/blake2/ref/blake2b-compress-avx2.c +1 -1
- data/vendor/libsodium/src/libsodium/crypto_generichash/blake2/ref/blake2b-ref.c +9 -1
- data/vendor/libsodium/src/libsodium/crypto_hash/sha256/cp/hash_sha256.c +11 -4
- data/vendor/libsodium/src/libsodium/crypto_hash/sha512/cp/hash_sha512.c +14 -8
- data/vendor/libsodium/src/libsodium/crypto_pwhash/argon2/argon2-encoding.c +0 -1
- data/vendor/libsodium/src/libsodium/crypto_pwhash/argon2/argon2.c +10 -6
- data/vendor/libsodium/src/libsodium/crypto_pwhash/scryptsalsa208sha256/sse/pwhash_scryptsalsa208sha256_sse.c +2 -1
- data/vendor/libsodium/src/libsodium/crypto_scalarmult/curve25519/donna_c64/curve25519_donna_c64.c +1 -1
- data/vendor/libsodium/src/libsodium/crypto_scalarmult/curve25519/ref10/x25519_ref10.c +11 -11
- data/vendor/libsodium/src/libsodium/crypto_scalarmult/curve25519/ref10/x25519_ref10.h +2 -2
- data/vendor/libsodium/src/libsodium/crypto_scalarmult/curve25519/sandy2x/fe51_nsquare.S +3 -0
- data/vendor/libsodium/src/libsodium/crypto_scalarmult/curve25519/sandy2x/fe51_pack.S +3 -0
- data/vendor/libsodium/src/libsodium/crypto_scalarmult/curve25519/sandy2x/ladder.S +830 -826
- data/vendor/libsodium/src/libsodium/crypto_scalarmult/curve25519/sandy2x/ladder_base.S +732 -728
- data/vendor/libsodium/src/libsodium/crypto_secretbox/crypto_secretbox_easy.c +3 -0
- data/vendor/libsodium/src/libsodium/crypto_stream/aes128ctr/portable/int128.h +1 -1
- data/vendor/libsodium/src/libsodium/crypto_stream/salsa20/amd64_xmm6/stream_salsa20_amd64_xmm6.S +8 -0
- data/vendor/libsodium/src/libsodium/include/Makefile.in +5 -0
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_aead_aes256gcm.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_aead_chacha20poly1305.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_auth.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_auth_hmacsha256.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_auth_hmacsha512.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_auth_hmacsha512256.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_box.h +3 -3
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_box_curve25519xsalsa20poly1305.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_generichash.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_generichash_blake2b.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_hash.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_hash_sha256.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_hash_sha512.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_onetimeauth.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_onetimeauth_poly1305.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_pwhash.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_pwhash_argon2i.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_pwhash_scryptsalsa208sha256.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_secretbox.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_secretbox_xsalsa20poly1305.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_shorthash.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_shorthash_siphash24.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_sign.h +3 -3
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_sign_ed25519.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_sign_edwards25519sha512batch.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_stream.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_stream_aes128ctr.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_stream_chacha20.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_stream_salsa20.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_stream_salsa2012.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_stream_salsa208.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/crypto_stream_xsalsa20.h +1 -1
- data/vendor/libsodium/src/libsodium/include/sodium/private/curve25519_ref10.h +2 -2
- data/vendor/libsodium/src/libsodium/include/sodium/private/mutex.h +7 -0
- data/vendor/libsodium/src/libsodium/include/sodium/randombytes.h +1 -1
- data/vendor/libsodium/src/libsodium/randombytes/nativeclient/randombytes_nativeclient.c +15 -4
- data/vendor/libsodium/src/libsodium/randombytes/randombytes.c +25 -15
- data/vendor/libsodium/src/libsodium/randombytes/salsa20/randombytes_salsa20_random.c +72 -15
- data/vendor/libsodium/src/libsodium/randombytes/sysrandom/randombytes_sysrandom.c +43 -4
- data/vendor/libsodium/src/libsodium/sodium/core.c +131 -2
- data/vendor/libsodium/src/libsodium/sodium/runtime.c +5 -2
- data/vendor/libsodium/src/libsodium/sodium/utils.c +16 -8
- data/vendor/libsodium/test/Makefile.in +5 -0
- data/vendor/libsodium/test/default/Makefile.in +5 -0
- data/vendor/libsodium/test/default/generichash2.c +8 -2
- data/vendor/libsodium/test/default/nacl-test-wrapper.sh +19 -17
- data/vendor/libsodium/test/default/randombytes.c +4 -2
- data/vendor/libsodium/test/quirks/quirks.h +23 -2
- metadata +11 -15
- data/vendor/libsodium/README +0 -1
- data/vendor/libsodium/dist-build/android-x86-64.sh +0 -4
- data/vendor/libsodium/examples/Makefile +0 -21
- data/vendor/libsodium/examples/auth.c +0 -68
- data/vendor/libsodium/examples/box.c +0 -135
- data/vendor/libsodium/examples/box_detached.c +0 -134
- data/vendor/libsodium/examples/generichash.c +0 -80
- data/vendor/libsodium/examples/generichash_stream.c +0 -58
- data/vendor/libsodium/examples/shorthash.c +0 -58
- data/vendor/libsodium/examples/sign.c +0 -78
- data/vendor/libsodium/examples/utils.h +0 -110
@@ -210,6 +210,7 @@ crypto_hash_sha256_update(crypto_hash_sha256_state *state,
|
|
210
210
|
const unsigned char *in,
|
211
211
|
unsigned long long inlen)
|
212
212
|
{
|
213
|
+
unsigned long long i;
|
213
214
|
uint32_t r;
|
214
215
|
|
215
216
|
if (inlen <= 0U) {
|
@@ -219,10 +220,14 @@ crypto_hash_sha256_update(crypto_hash_sha256_state *state,
|
|
219
220
|
state->count += (uint64_t)(inlen) << 3;
|
220
221
|
|
221
222
|
if (inlen < 64 - r) {
|
222
|
-
|
223
|
+
for (i = 0; i < inlen; i++) {
|
224
|
+
state->buf[r + i] = in[i];
|
225
|
+
}
|
223
226
|
return 0;
|
224
227
|
}
|
225
|
-
|
228
|
+
for (i = 0; i < 64 - r; i++) {
|
229
|
+
state->buf[r + i] = in[i];
|
230
|
+
}
|
226
231
|
SHA256_Transform(state->state, state->buf);
|
227
232
|
in += 64 - r;
|
228
233
|
inlen -= 64 - r;
|
@@ -232,8 +237,10 @@ crypto_hash_sha256_update(crypto_hash_sha256_state *state,
|
|
232
237
|
in += 64;
|
233
238
|
inlen -= 64;
|
234
239
|
}
|
235
|
-
|
236
|
-
|
240
|
+
inlen &= 63;
|
241
|
+
for (i = 0; i < inlen; i++) {
|
242
|
+
state->buf[i] = in[i];
|
243
|
+
}
|
237
244
|
return 0;
|
238
245
|
}
|
239
246
|
|
@@ -232,9 +232,9 @@ crypto_hash_sha512_update(crypto_hash_sha512_state *state,
|
|
232
232
|
const unsigned char *in,
|
233
233
|
unsigned long long inlen)
|
234
234
|
{
|
235
|
+
unsigned long long i;
|
235
236
|
uint64_t bitlen[2];
|
236
237
|
uint64_t r;
|
237
|
-
const unsigned char *src = in;
|
238
238
|
|
239
239
|
r = (state->count[1] >> 3) & 0x7f;
|
240
240
|
|
@@ -249,21 +249,27 @@ crypto_hash_sha512_update(crypto_hash_sha512_state *state,
|
|
249
249
|
state->count[0] += bitlen[0];
|
250
250
|
|
251
251
|
if (inlen < 128 - r) {
|
252
|
-
|
252
|
+
for (i = 0; i < inlen; i++) {
|
253
|
+
state->buf[r + i] = in[i];
|
254
|
+
}
|
253
255
|
return 0;
|
254
256
|
}
|
255
|
-
|
257
|
+
for (i = 0; i < 128 - r; i++) {
|
258
|
+
state->buf[r + i] = in[i];
|
259
|
+
}
|
256
260
|
SHA512_Transform(state->state, state->buf);
|
257
|
-
|
261
|
+
in += 128 - r;
|
258
262
|
inlen -= 128 - r;
|
259
263
|
|
260
264
|
while (inlen >= 128) {
|
261
|
-
SHA512_Transform(state->state,
|
262
|
-
|
265
|
+
SHA512_Transform(state->state, in);
|
266
|
+
in += 128;
|
263
267
|
inlen -= 128;
|
264
268
|
}
|
265
|
-
|
266
|
-
|
269
|
+
inlen &= 127;
|
270
|
+
for (i = 0; i < inlen; i++) {
|
271
|
+
state->buf[i] = in[i];
|
272
|
+
}
|
267
273
|
return 0;
|
268
274
|
}
|
269
275
|
|
@@ -17,7 +17,6 @@
|
|
17
17
|
#include <stdio.h>
|
18
18
|
#include <limits.h>
|
19
19
|
|
20
|
-
|
21
20
|
#include "utils.h"
|
22
21
|
|
23
22
|
#include "argon2.h"
|
@@ -179,18 +178,23 @@ int argon2_verify(const char *encoded, const void *pwd, const size_t pwdlen,
|
|
179
178
|
uint8_t *out;
|
180
179
|
int decode_result;
|
181
180
|
int ret;
|
182
|
-
|
181
|
+
size_t encoded_len;
|
183
182
|
|
184
183
|
memset(&ctx, 0, sizeof ctx);
|
185
184
|
|
185
|
+
ctx.pwd = NULL;
|
186
|
+
ctx.pwdlen = 0;
|
186
187
|
ctx.secret = NULL;
|
187
188
|
ctx.secretlen = 0;
|
188
189
|
|
189
190
|
/* max values, to be updated in decode_string */
|
190
|
-
encoded_len =
|
191
|
-
|
192
|
-
|
193
|
-
|
191
|
+
encoded_len = strlen(encoded);
|
192
|
+
if (encoded_len > UINT32_MAX) {
|
193
|
+
return ARGON2_DECODING_LENGTH_FAIL;
|
194
|
+
}
|
195
|
+
ctx.adlen = (uint32_t) encoded_len;
|
196
|
+
ctx.saltlen = (uint32_t) encoded_len;
|
197
|
+
ctx.outlen = (uint32_t) encoded_len;
|
194
198
|
|
195
199
|
ctx.ad = (uint8_t *) malloc(ctx.adlen);
|
196
200
|
ctx.salt = (uint8_t *) malloc(ctx.saltlen);
|
@@ -30,7 +30,7 @@
|
|
30
30
|
|
31
31
|
#if defined(HAVE_EMMINTRIN_H) || \
|
32
32
|
(defined(_MSC_VER) && (defined(_M_X64) || defined(_M_AMD64) || defined(_M_IX86)))
|
33
|
-
#
|
33
|
+
#ifdef __GNUC__
|
34
34
|
# pragma GCC target("sse2")
|
35
35
|
#endif
|
36
36
|
#include <emmintrin.h>
|
@@ -210,6 +210,7 @@ blockmix_salsa8_xor(const __m128i * Bin1, const __m128i * Bin2, __m128i * Bout,
|
|
210
210
|
/**
|
211
211
|
* integerify(B, r):
|
212
212
|
* Return the result of parsing B_{2r-1} as a little-endian integer.
|
213
|
+
* Note that B's layout is permuted compared to the generic implementation.
|
213
214
|
*/
|
214
215
|
static inline uint32_t
|
215
216
|
integerify(const void * B, size_t r)
|
@@ -127,17 +127,17 @@ fe_mul121666(fe h,const fe f)
|
|
127
127
|
int64_t carry8;
|
128
128
|
int64_t carry9;
|
129
129
|
|
130
|
-
carry9 = (h9 + ((int64_t) 1 << 24)) >> 25; h0 += carry9 * 19; h9 -= carry9 << 25;
|
131
|
-
carry1 = (h1 + ((int64_t) 1 << 24)) >> 25; h2 += carry1; h1 -= carry1 << 25;
|
132
|
-
carry3 = (h3 + ((int64_t) 1 << 24)) >> 25; h4 += carry3; h3 -= carry3 << 25;
|
133
|
-
carry5 = (h5 + ((int64_t) 1 << 24)) >> 25; h6 += carry5; h5 -= carry5 << 25;
|
134
|
-
carry7 = (h7 + ((int64_t) 1 << 24)) >> 25; h8 += carry7; h7 -= carry7 << 25;
|
135
|
-
|
136
|
-
carry0 = (h0 + ((int64_t) 1 << 25)) >> 26; h1 += carry0; h0 -= carry0 << 26;
|
137
|
-
carry2 = (h2 + ((int64_t) 1 << 25)) >> 26; h3 += carry2; h2 -= carry2 << 26;
|
138
|
-
carry4 = (h4 + ((int64_t) 1 << 25)) >> 26; h5 += carry4; h4 -= carry4 << 26;
|
139
|
-
carry6 = (h6 + ((int64_t) 1 << 25)) >> 26; h7 += carry6; h6 -= carry6 << 26;
|
140
|
-
carry8 = (h8 + ((int64_t) 1 << 25)) >> 26; h9 += carry8; h8 -= carry8 << 26;
|
130
|
+
carry9 = (h9 + ((int64_t) 1 << 24)) >> 25; h0 += carry9 * 19; h9 -= carry9 * ((int64_t) 1 << 25);
|
131
|
+
carry1 = (h1 + ((int64_t) 1 << 24)) >> 25; h2 += carry1; h1 -= carry1 * ((int64_t) 1 << 25);
|
132
|
+
carry3 = (h3 + ((int64_t) 1 << 24)) >> 25; h4 += carry3; h3 -= carry3 * ((int64_t) 1 << 25);
|
133
|
+
carry5 = (h5 + ((int64_t) 1 << 24)) >> 25; h6 += carry5; h5 -= carry5 * ((int64_t) 1 << 25);
|
134
|
+
carry7 = (h7 + ((int64_t) 1 << 24)) >> 25; h8 += carry7; h7 -= carry7 * ((int64_t) 1 << 25);
|
135
|
+
|
136
|
+
carry0 = (h0 + ((int64_t) 1 << 25)) >> 26; h1 += carry0; h0 -= carry0 * ((int64_t) 1 << 26);
|
137
|
+
carry2 = (h2 + ((int64_t) 1 << 25)) >> 26; h3 += carry2; h2 -= carry2 * ((int64_t) 1 << 26);
|
138
|
+
carry4 = (h4 + ((int64_t) 1 << 25)) >> 26; h5 += carry4; h4 -= carry4 * ((int64_t) 1 << 26);
|
139
|
+
carry6 = (h6 + ((int64_t) 1 << 25)) >> 26; h7 += carry6; h6 -= carry6 * ((int64_t) 1 << 26);
|
140
|
+
carry8 = (h8 + ((int64_t) 1 << 25)) >> 26; h9 += carry8; h8 -= carry8 * ((int64_t) 1 << 26);
|
141
141
|
|
142
142
|
h[0] = h0;
|
143
143
|
h[1] = h1;
|
@@ -41,6 +41,8 @@ movq %r9,16(%rdi)
|
|
41
41
|
movq %rax,24(%rdi)
|
42
42
|
movq %rsi,32(%rdi)
|
43
43
|
mov %rdx,%rsi
|
44
|
+
|
45
|
+
.p2align 4
|
44
46
|
._loop:
|
45
47
|
sub $1,%rsi
|
46
48
|
mov %rcx,%rax
|
@@ -154,6 +156,7 @@ and %rdx,%r10
|
|
154
156
|
movq %r10,32(%rdi)
|
155
157
|
cmp $0,%rsi
|
156
158
|
jne ._loop
|
159
|
+
|
157
160
|
movq %rcx,0(%rdi)
|
158
161
|
movq %r8,8(%rdi)
|
159
162
|
movq 0(%rsp),%r11
|
@@ -35,6 +35,8 @@ movq 32(%rsi),%rsi
|
|
35
35
|
movq REDMASK51(%rip),%rax
|
36
36
|
lea -18(%rax),%r10
|
37
37
|
mov $3,%r11
|
38
|
+
|
39
|
+
.p2align 4
|
38
40
|
._reduceloop:
|
39
41
|
mov %rdx,%r12
|
40
42
|
shr $51,%r12
|
@@ -59,6 +61,7 @@ imulq $19, %r12,%r12
|
|
59
61
|
add %r12,%rdx
|
60
62
|
sub $1,%r11
|
61
63
|
ja ._reduceloop
|
64
|
+
|
62
65
|
mov $1,%r12
|
63
66
|
cmp %r10,%rdx
|
64
67
|
cmovl %r11,%r12
|
@@ -25,77 +25,77 @@ movq %r11,1824(%rsp)
|
|
25
25
|
movq %r12,1832(%rsp)
|
26
26
|
movq %r13,1840(%rsp)
|
27
27
|
movq %r14,1848(%rsp)
|
28
|
-
|
29
|
-
|
30
|
-
|
31
|
-
|
32
|
-
|
33
|
-
|
34
|
-
|
35
|
-
|
36
|
-
|
37
|
-
|
38
|
-
|
39
|
-
|
40
|
-
|
41
|
-
|
42
|
-
|
43
|
-
|
44
|
-
|
45
|
-
|
46
|
-
|
47
|
-
|
48
|
-
|
49
|
-
|
50
|
-
|
51
|
-
|
52
|
-
|
53
|
-
|
54
|
-
|
55
|
-
|
56
|
-
|
57
|
-
|
58
|
-
|
59
|
-
|
60
|
-
|
61
|
-
|
62
|
-
|
63
|
-
|
64
|
-
|
65
|
-
|
66
|
-
|
67
|
-
|
68
|
-
|
69
|
-
|
70
|
-
|
71
|
-
|
72
|
-
|
73
|
-
|
74
|
-
|
75
|
-
|
76
|
-
|
77
|
-
|
78
|
-
|
79
|
-
|
80
|
-
|
81
|
-
|
82
|
-
|
83
|
-
|
84
|
-
|
85
|
-
|
86
|
-
|
87
|
-
|
88
|
-
|
89
|
-
|
90
|
-
|
91
|
-
|
92
|
-
|
93
|
-
|
94
|
-
|
95
|
-
|
96
|
-
|
97
|
-
|
98
|
-
|
28
|
+
vmovdqa v0_0(%rip),%xmm0
|
29
|
+
vmovdqa v1_0(%rip),%xmm1
|
30
|
+
vmovdqu 0(%rdi),%xmm2
|
31
|
+
vmovdqa %xmm2,0(%rsp)
|
32
|
+
vmovdqu 16(%rdi),%xmm2
|
33
|
+
vmovdqa %xmm2,16(%rsp)
|
34
|
+
vmovdqu 32(%rdi),%xmm2
|
35
|
+
vmovdqa %xmm2,32(%rsp)
|
36
|
+
vmovdqu 48(%rdi),%xmm2
|
37
|
+
vmovdqa %xmm2,48(%rsp)
|
38
|
+
vmovdqu 64(%rdi),%xmm2
|
39
|
+
vmovdqa %xmm2,64(%rsp)
|
40
|
+
vmovdqa %xmm1,80(%rsp)
|
41
|
+
vmovdqa %xmm0,96(%rsp)
|
42
|
+
vmovdqa %xmm0,112(%rsp)
|
43
|
+
vmovdqa %xmm0,128(%rsp)
|
44
|
+
vmovdqa %xmm0,144(%rsp)
|
45
|
+
vmovdqa %xmm1,%xmm0
|
46
|
+
vpxor %xmm1,%xmm1,%xmm1
|
47
|
+
vpxor %xmm2,%xmm2,%xmm2
|
48
|
+
vpxor %xmm3,%xmm3,%xmm3
|
49
|
+
vpxor %xmm4,%xmm4,%xmm4
|
50
|
+
vpxor %xmm5,%xmm5,%xmm5
|
51
|
+
vpxor %xmm6,%xmm6,%xmm6
|
52
|
+
vpxor %xmm7,%xmm7,%xmm7
|
53
|
+
vpxor %xmm8,%xmm8,%xmm8
|
54
|
+
vpxor %xmm9,%xmm9,%xmm9
|
55
|
+
vmovdqu 0(%rdi),%xmm10
|
56
|
+
vmovdqa %xmm10,160(%rsp)
|
57
|
+
vmovdqu 16(%rdi),%xmm10
|
58
|
+
vmovdqa %xmm10,176(%rsp)
|
59
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
60
|
+
vmovdqa %xmm10,192(%rsp)
|
61
|
+
vmovdqu 32(%rdi),%xmm10
|
62
|
+
vmovdqa %xmm10,208(%rsp)
|
63
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
64
|
+
vmovdqa %xmm10,224(%rsp)
|
65
|
+
vmovdqu 48(%rdi),%xmm10
|
66
|
+
vmovdqa %xmm10,240(%rsp)
|
67
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
68
|
+
vmovdqa %xmm10,256(%rsp)
|
69
|
+
vmovdqu 64(%rdi),%xmm10
|
70
|
+
vmovdqa %xmm10,272(%rsp)
|
71
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
72
|
+
vmovdqa %xmm10,288(%rsp)
|
73
|
+
vmovdqu 8(%rdi),%xmm10
|
74
|
+
vpmuludq v2_1(%rip),%xmm10,%xmm10
|
75
|
+
vmovdqa %xmm10,304(%rsp)
|
76
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
77
|
+
vmovdqa %xmm10,320(%rsp)
|
78
|
+
vmovdqu 24(%rdi),%xmm10
|
79
|
+
vpmuludq v2_1(%rip),%xmm10,%xmm10
|
80
|
+
vmovdqa %xmm10,336(%rsp)
|
81
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
82
|
+
vmovdqa %xmm10,352(%rsp)
|
83
|
+
vmovdqu 40(%rdi),%xmm10
|
84
|
+
vpmuludq v2_1(%rip),%xmm10,%xmm10
|
85
|
+
vmovdqa %xmm10,368(%rsp)
|
86
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
87
|
+
vmovdqa %xmm10,384(%rsp)
|
88
|
+
vmovdqu 56(%rdi),%xmm10
|
89
|
+
vpmuludq v2_1(%rip),%xmm10,%xmm10
|
90
|
+
vmovdqa %xmm10,400(%rsp)
|
91
|
+
vpmuludq v19_19(%rip),%xmm10,%xmm10
|
92
|
+
vmovdqa %xmm10,416(%rsp)
|
93
|
+
vmovdqu 0(%rdi),%xmm10
|
94
|
+
vmovdqu 64(%rdi),%xmm11
|
95
|
+
vblendps $12, %xmm11, %xmm10, %xmm10
|
96
|
+
vpshufd $2,%xmm10,%xmm10
|
97
|
+
vpmuludq v38_1(%rip),%xmm10,%xmm10
|
98
|
+
vmovdqa %xmm10,432(%rsp)
|
99
99
|
movq 0(%rsi),%rdx
|
100
100
|
movq 8(%rsi),%rcx
|
101
101
|
movq 16(%rsi),%r8
|
@@ -110,6 +110,8 @@ xorq 16(%rsi),%r8
|
|
110
110
|
xorq 24(%rsi),%r9
|
111
111
|
leaq 800(%rsp),%rsi
|
112
112
|
mov $64,%rax
|
113
|
+
|
114
|
+
.p2align 4
|
113
115
|
._ladder_small_loop:
|
114
116
|
mov %rdx,%r10
|
115
117
|
mov %rcx,%r11
|
@@ -136,1274 +138,1276 @@ sub $1,%rax
|
|
136
138
|
jne ._ladder_small_loop
|
137
139
|
mov $255,%rdx
|
138
140
|
add $760,%rsi
|
141
|
+
|
142
|
+
.p2align 4
|
139
143
|
._ladder_loop:
|
140
144
|
sub $1,%rdx
|
141
145
|
vbroadcastss 0(%rsi),%xmm10
|
142
146
|
sub $4,%rsi
|
143
|
-
|
144
|
-
|
147
|
+
vmovdqa 0(%rsp),%xmm11
|
148
|
+
vmovdqa 80(%rsp),%xmm12
|
145
149
|
vpxor %xmm11,%xmm0,%xmm13
|
146
|
-
|
147
|
-
|
148
|
-
|
150
|
+
vpand %xmm10,%xmm13,%xmm13
|
151
|
+
vpxor %xmm13,%xmm0,%xmm0
|
152
|
+
vpxor %xmm13,%xmm11,%xmm11
|
149
153
|
vpxor %xmm12,%xmm1,%xmm13
|
150
|
-
|
151
|
-
|
152
|
-
|
153
|
-
|
154
|
-
|
154
|
+
vpand %xmm10,%xmm13,%xmm13
|
155
|
+
vpxor %xmm13,%xmm1,%xmm1
|
156
|
+
vpxor %xmm13,%xmm12,%xmm12
|
157
|
+
vmovdqa 16(%rsp),%xmm13
|
158
|
+
vmovdqa 96(%rsp),%xmm14
|
155
159
|
vpxor %xmm13,%xmm2,%xmm15
|
156
|
-
|
157
|
-
|
158
|
-
|
160
|
+
vpand %xmm10,%xmm15,%xmm15
|
161
|
+
vpxor %xmm15,%xmm2,%xmm2
|
162
|
+
vpxor %xmm15,%xmm13,%xmm13
|
159
163
|
vpxor %xmm14,%xmm3,%xmm15
|
160
|
-
|
161
|
-
|
162
|
-
|
163
|
-
|
164
|
-
|
165
|
-
|
166
|
-
|
164
|
+
vpand %xmm10,%xmm15,%xmm15
|
165
|
+
vpxor %xmm15,%xmm3,%xmm3
|
166
|
+
vpxor %xmm15,%xmm14,%xmm14
|
167
|
+
vmovdqa %xmm13,0(%rsp)
|
168
|
+
vmovdqa %xmm14,16(%rsp)
|
169
|
+
vmovdqa 32(%rsp),%xmm13
|
170
|
+
vmovdqa 112(%rsp),%xmm14
|
167
171
|
vpxor %xmm13,%xmm4,%xmm15
|
168
|
-
|
169
|
-
|
170
|
-
|
172
|
+
vpand %xmm10,%xmm15,%xmm15
|
173
|
+
vpxor %xmm15,%xmm4,%xmm4
|
174
|
+
vpxor %xmm15,%xmm13,%xmm13
|
171
175
|
vpxor %xmm14,%xmm5,%xmm15
|
172
|
-
|
173
|
-
|
174
|
-
|
175
|
-
|
176
|
-
|
177
|
-
|
178
|
-
|
176
|
+
vpand %xmm10,%xmm15,%xmm15
|
177
|
+
vpxor %xmm15,%xmm5,%xmm5
|
178
|
+
vpxor %xmm15,%xmm14,%xmm14
|
179
|
+
vmovdqa %xmm13,32(%rsp)
|
180
|
+
vmovdqa %xmm14,80(%rsp)
|
181
|
+
vmovdqa 48(%rsp),%xmm13
|
182
|
+
vmovdqa 128(%rsp),%xmm14
|
179
183
|
vpxor %xmm13,%xmm6,%xmm15
|
180
|
-
|
181
|
-
|
182
|
-
|
184
|
+
vpand %xmm10,%xmm15,%xmm15
|
185
|
+
vpxor %xmm15,%xmm6,%xmm6
|
186
|
+
vpxor %xmm15,%xmm13,%xmm13
|
183
187
|
vpxor %xmm14,%xmm7,%xmm15
|
184
|
-
|
185
|
-
|
186
|
-
|
187
|
-
|
188
|
-
|
189
|
-
|
190
|
-
|
188
|
+
vpand %xmm10,%xmm15,%xmm15
|
189
|
+
vpxor %xmm15,%xmm7,%xmm7
|
190
|
+
vpxor %xmm15,%xmm14,%xmm14
|
191
|
+
vmovdqa %xmm13,48(%rsp)
|
192
|
+
vmovdqa %xmm14,96(%rsp)
|
193
|
+
vmovdqa 64(%rsp),%xmm13
|
194
|
+
vmovdqa 144(%rsp),%xmm14
|
191
195
|
vpxor %xmm13,%xmm8,%xmm15
|
192
|
-
|
193
|
-
|
194
|
-
|
196
|
+
vpand %xmm10,%xmm15,%xmm15
|
197
|
+
vpxor %xmm15,%xmm8,%xmm8
|
198
|
+
vpxor %xmm15,%xmm13,%xmm13
|
195
199
|
vpxor %xmm14,%xmm9,%xmm15
|
196
|
-
|
197
|
-
|
198
|
-
|
199
|
-
|
200
|
-
|
200
|
+
vpand %xmm10,%xmm15,%xmm15
|
201
|
+
vpxor %xmm15,%xmm9,%xmm9
|
202
|
+
vpxor %xmm15,%xmm14,%xmm14
|
203
|
+
vmovdqa %xmm13,64(%rsp)
|
204
|
+
vmovdqa %xmm14,112(%rsp)
|
201
205
|
vpaddq subc0(%rip),%xmm11,%xmm10
|
202
|
-
|
203
|
-
|
206
|
+
vpsubq %xmm12,%xmm10,%xmm10
|
207
|
+
vpaddq %xmm12,%xmm11,%xmm11
|
204
208
|
vpunpckhqdq %xmm10,%xmm11,%xmm12
|
205
209
|
vpunpcklqdq %xmm10,%xmm11,%xmm10
|
206
210
|
vpaddq %xmm1,%xmm0,%xmm11
|
207
|
-
|
208
|
-
|
211
|
+
vpaddq subc0(%rip),%xmm0,%xmm0
|
212
|
+
vpsubq %xmm1,%xmm0,%xmm0
|
209
213
|
vpunpckhqdq %xmm11,%xmm0,%xmm1
|
210
214
|
vpunpcklqdq %xmm11,%xmm0,%xmm0
|
211
215
|
vpmuludq %xmm0,%xmm10,%xmm11
|
212
216
|
vpmuludq %xmm1,%xmm10,%xmm13
|
213
|
-
|
214
|
-
|
217
|
+
vmovdqa %xmm1,128(%rsp)
|
218
|
+
vpaddq %xmm1,%xmm1,%xmm1
|
215
219
|
vpmuludq %xmm0,%xmm12,%xmm14
|
216
|
-
|
217
|
-
|
220
|
+
vmovdqa %xmm0,144(%rsp)
|
221
|
+
vpaddq %xmm14,%xmm13,%xmm13
|
218
222
|
vpmuludq %xmm1,%xmm12,%xmm0
|
219
|
-
|
223
|
+
vmovdqa %xmm1,448(%rsp)
|
220
224
|
vpaddq %xmm3,%xmm2,%xmm1
|
221
|
-
|
222
|
-
|
225
|
+
vpaddq subc2(%rip),%xmm2,%xmm2
|
226
|
+
vpsubq %xmm3,%xmm2,%xmm2
|
223
227
|
vpunpckhqdq %xmm1,%xmm2,%xmm3
|
224
228
|
vpunpcklqdq %xmm1,%xmm2,%xmm1
|
225
229
|
vpmuludq %xmm1,%xmm10,%xmm2
|
226
|
-
|
230
|
+
vpaddq %xmm2,%xmm0,%xmm0
|
227
231
|
vpmuludq %xmm3,%xmm10,%xmm2
|
228
|
-
|
229
|
-
|
232
|
+
vmovdqa %xmm3,464(%rsp)
|
233
|
+
vpaddq %xmm3,%xmm3,%xmm3
|
230
234
|
vpmuludq %xmm1,%xmm12,%xmm14
|
231
|
-
|
232
|
-
|
235
|
+
vmovdqa %xmm1,480(%rsp)
|
236
|
+
vpaddq %xmm14,%xmm2,%xmm2
|
233
237
|
vpmuludq %xmm3,%xmm12,%xmm1
|
234
|
-
|
238
|
+
vmovdqa %xmm3,496(%rsp)
|
235
239
|
vpaddq %xmm5,%xmm4,%xmm3
|
236
|
-
|
237
|
-
|
240
|
+
vpaddq subc2(%rip),%xmm4,%xmm4
|
241
|
+
vpsubq %xmm5,%xmm4,%xmm4
|
238
242
|
vpunpckhqdq %xmm3,%xmm4,%xmm5
|
239
243
|
vpunpcklqdq %xmm3,%xmm4,%xmm3
|
240
244
|
vpmuludq %xmm3,%xmm10,%xmm4
|
241
|
-
|
245
|
+
vpaddq %xmm4,%xmm1,%xmm1
|
242
246
|
vpmuludq %xmm5,%xmm10,%xmm4
|
243
|
-
|
244
|
-
|
247
|
+
vmovdqa %xmm5,512(%rsp)
|
248
|
+
vpaddq %xmm5,%xmm5,%xmm5
|
245
249
|
vpmuludq %xmm3,%xmm12,%xmm14
|
246
|
-
|
247
|
-
|
250
|
+
vmovdqa %xmm3,528(%rsp)
|
251
|
+
vpaddq %xmm14,%xmm4,%xmm4
|
248
252
|
vpaddq %xmm7,%xmm6,%xmm3
|
249
|
-
|
250
|
-
|
253
|
+
vpaddq subc2(%rip),%xmm6,%xmm6
|
254
|
+
vpsubq %xmm7,%xmm6,%xmm6
|
251
255
|
vpunpckhqdq %xmm3,%xmm6,%xmm7
|
252
256
|
vpunpcklqdq %xmm3,%xmm6,%xmm3
|
253
257
|
vpmuludq %xmm3,%xmm10,%xmm6
|
254
258
|
vpmuludq %xmm5,%xmm12,%xmm14
|
255
|
-
|
256
|
-
|
257
|
-
|
258
|
-
|
259
|
+
vmovdqa %xmm5,544(%rsp)
|
260
|
+
vpmuludq v19_19(%rip),%xmm5,%xmm5
|
261
|
+
vmovdqa %xmm5,560(%rsp)
|
262
|
+
vpaddq %xmm14,%xmm6,%xmm6
|
259
263
|
vpmuludq %xmm7,%xmm10,%xmm5
|
260
|
-
|
261
|
-
|
264
|
+
vmovdqa %xmm7,576(%rsp)
|
265
|
+
vpaddq %xmm7,%xmm7,%xmm7
|
262
266
|
vpmuludq %xmm3,%xmm12,%xmm14
|
263
|
-
|
264
|
-
|
265
|
-
|
266
|
-
|
267
|
+
vmovdqa %xmm3,592(%rsp)
|
268
|
+
vpaddq %xmm14,%xmm5,%xmm5
|
269
|
+
vpmuludq v19_19(%rip),%xmm3,%xmm3
|
270
|
+
vmovdqa %xmm3,608(%rsp)
|
267
271
|
vpaddq %xmm9,%xmm8,%xmm3
|
268
|
-
|
269
|
-
|
272
|
+
vpaddq subc2(%rip),%xmm8,%xmm8
|
273
|
+
vpsubq %xmm9,%xmm8,%xmm8
|
270
274
|
vpunpckhqdq %xmm3,%xmm8,%xmm9
|
271
275
|
vpunpcklqdq %xmm3,%xmm8,%xmm3
|
272
|
-
|
276
|
+
vmovdqa %xmm3,624(%rsp)
|
273
277
|
vpmuludq %xmm7,%xmm12,%xmm8
|
274
|
-
|
275
|
-
|
276
|
-
|
278
|
+
vmovdqa %xmm7,640(%rsp)
|
279
|
+
vpmuludq v19_19(%rip),%xmm7,%xmm7
|
280
|
+
vmovdqa %xmm7,656(%rsp)
|
277
281
|
vpmuludq %xmm3,%xmm10,%xmm7
|
278
|
-
|
282
|
+
vpaddq %xmm7,%xmm8,%xmm8
|
279
283
|
vpmuludq %xmm9,%xmm10,%xmm7
|
280
|
-
|
281
|
-
|
284
|
+
vmovdqa %xmm9,672(%rsp)
|
285
|
+
vpaddq %xmm9,%xmm9,%xmm9
|
282
286
|
vpmuludq %xmm3,%xmm12,%xmm10
|
283
|
-
|
284
|
-
|
285
|
-
|
286
|
-
|
287
|
+
vpaddq %xmm10,%xmm7,%xmm7
|
288
|
+
vpmuludq v19_19(%rip),%xmm3,%xmm3
|
289
|
+
vmovdqa %xmm3,688(%rsp)
|
290
|
+
vpmuludq v19_19(%rip),%xmm12,%xmm12
|
287
291
|
vpmuludq %xmm9,%xmm12,%xmm3
|
288
|
-
|
289
|
-
|
290
|
-
|
291
|
-
|
292
|
+
vmovdqa %xmm9,704(%rsp)
|
293
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
294
|
+
vmovdqa 0(%rsp),%xmm3
|
295
|
+
vmovdqa 16(%rsp),%xmm9
|
292
296
|
vpaddq subc2(%rip),%xmm3,%xmm10
|
293
|
-
|
294
|
-
|
297
|
+
vpsubq %xmm9,%xmm10,%xmm10
|
298
|
+
vpaddq %xmm9,%xmm3,%xmm3
|
295
299
|
vpunpckhqdq %xmm10,%xmm3,%xmm9
|
296
300
|
vpunpcklqdq %xmm10,%xmm3,%xmm3
|
297
301
|
vpmuludq 144(%rsp),%xmm3,%xmm10
|
298
|
-
|
302
|
+
vpaddq %xmm10,%xmm0,%xmm0
|
299
303
|
vpmuludq 128(%rsp),%xmm3,%xmm10
|
300
|
-
|
304
|
+
vpaddq %xmm10,%xmm2,%xmm2
|
301
305
|
vpmuludq 480(%rsp),%xmm3,%xmm10
|
302
|
-
|
306
|
+
vpaddq %xmm10,%xmm1,%xmm1
|
303
307
|
vpmuludq 464(%rsp),%xmm3,%xmm10
|
304
|
-
|
308
|
+
vpaddq %xmm10,%xmm4,%xmm4
|
305
309
|
vpmuludq 528(%rsp),%xmm3,%xmm10
|
306
|
-
|
310
|
+
vpaddq %xmm10,%xmm6,%xmm6
|
307
311
|
vpmuludq 512(%rsp),%xmm3,%xmm10
|
308
|
-
|
312
|
+
vpaddq %xmm10,%xmm5,%xmm5
|
309
313
|
vpmuludq 592(%rsp),%xmm3,%xmm10
|
310
|
-
|
314
|
+
vpaddq %xmm10,%xmm8,%xmm8
|
311
315
|
vpmuludq 576(%rsp),%xmm3,%xmm10
|
312
|
-
|
313
|
-
|
316
|
+
vpaddq %xmm10,%xmm7,%xmm7
|
317
|
+
vpmuludq v19_19(%rip),%xmm3,%xmm3
|
314
318
|
vpmuludq 624(%rsp),%xmm3,%xmm10
|
315
|
-
|
316
|
-
|
317
|
-
|
319
|
+
vpaddq %xmm10,%xmm11,%xmm11
|
320
|
+
vpmuludq 672(%rsp),%xmm3,%xmm3
|
321
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
318
322
|
vpmuludq 144(%rsp),%xmm9,%xmm3
|
319
|
-
|
323
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
320
324
|
vpmuludq 448(%rsp),%xmm9,%xmm3
|
321
|
-
|
325
|
+
vpaddq %xmm3,%xmm1,%xmm1
|
322
326
|
vpmuludq 480(%rsp),%xmm9,%xmm3
|
323
|
-
|
327
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
324
328
|
vpmuludq 496(%rsp),%xmm9,%xmm3
|
325
|
-
|
329
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
326
330
|
vpmuludq 528(%rsp),%xmm9,%xmm3
|
327
|
-
|
331
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
328
332
|
vpmuludq 544(%rsp),%xmm9,%xmm3
|
329
|
-
|
333
|
+
vpaddq %xmm3,%xmm8,%xmm8
|
330
334
|
vpmuludq 592(%rsp),%xmm9,%xmm3
|
331
|
-
|
332
|
-
|
335
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
336
|
+
vpmuludq v19_19(%rip),%xmm9,%xmm9
|
333
337
|
vpmuludq 640(%rsp),%xmm9,%xmm3
|
334
|
-
|
338
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
335
339
|
vpmuludq 624(%rsp),%xmm9,%xmm3
|
336
|
-
|
337
|
-
|
338
|
-
|
339
|
-
|
340
|
-
|
340
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
341
|
+
vpmuludq 704(%rsp),%xmm9,%xmm9
|
342
|
+
vpaddq %xmm9,%xmm0,%xmm0
|
343
|
+
vmovdqa 32(%rsp),%xmm3
|
344
|
+
vmovdqa 80(%rsp),%xmm9
|
341
345
|
vpaddq subc2(%rip),%xmm3,%xmm10
|
342
|
-
|
343
|
-
|
346
|
+
vpsubq %xmm9,%xmm10,%xmm10
|
347
|
+
vpaddq %xmm9,%xmm3,%xmm3
|
344
348
|
vpunpckhqdq %xmm10,%xmm3,%xmm9
|
345
349
|
vpunpcklqdq %xmm10,%xmm3,%xmm3
|
346
350
|
vpmuludq 144(%rsp),%xmm3,%xmm10
|
347
|
-
|
351
|
+
vpaddq %xmm10,%xmm1,%xmm1
|
348
352
|
vpmuludq 128(%rsp),%xmm3,%xmm10
|
349
|
-
|
353
|
+
vpaddq %xmm10,%xmm4,%xmm4
|
350
354
|
vpmuludq 480(%rsp),%xmm3,%xmm10
|
351
|
-
|
355
|
+
vpaddq %xmm10,%xmm6,%xmm6
|
352
356
|
vpmuludq 464(%rsp),%xmm3,%xmm10
|
353
|
-
|
357
|
+
vpaddq %xmm10,%xmm5,%xmm5
|
354
358
|
vpmuludq 528(%rsp),%xmm3,%xmm10
|
355
|
-
|
359
|
+
vpaddq %xmm10,%xmm8,%xmm8
|
356
360
|
vpmuludq 512(%rsp),%xmm3,%xmm10
|
357
|
-
|
358
|
-
|
361
|
+
vpaddq %xmm10,%xmm7,%xmm7
|
362
|
+
vpmuludq v19_19(%rip),%xmm3,%xmm3
|
359
363
|
vpmuludq 592(%rsp),%xmm3,%xmm10
|
360
|
-
|
364
|
+
vpaddq %xmm10,%xmm11,%xmm11
|
361
365
|
vpmuludq 576(%rsp),%xmm3,%xmm10
|
362
|
-
|
366
|
+
vpaddq %xmm10,%xmm13,%xmm13
|
363
367
|
vpmuludq 624(%rsp),%xmm3,%xmm10
|
364
|
-
|
365
|
-
|
366
|
-
|
368
|
+
vpaddq %xmm10,%xmm0,%xmm0
|
369
|
+
vpmuludq 672(%rsp),%xmm3,%xmm3
|
370
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
367
371
|
vpmuludq 144(%rsp),%xmm9,%xmm3
|
368
|
-
|
372
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
369
373
|
vpmuludq 448(%rsp),%xmm9,%xmm3
|
370
|
-
|
374
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
371
375
|
vpmuludq 480(%rsp),%xmm9,%xmm3
|
372
|
-
|
376
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
373
377
|
vpmuludq 496(%rsp),%xmm9,%xmm3
|
374
|
-
|
378
|
+
vpaddq %xmm3,%xmm8,%xmm8
|
375
379
|
vpmuludq 528(%rsp),%xmm9,%xmm3
|
376
|
-
|
377
|
-
|
380
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
381
|
+
vpmuludq v19_19(%rip),%xmm9,%xmm9
|
378
382
|
vpmuludq 544(%rsp),%xmm9,%xmm3
|
379
|
-
|
383
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
380
384
|
vpmuludq 592(%rsp),%xmm9,%xmm3
|
381
|
-
|
385
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
382
386
|
vpmuludq 640(%rsp),%xmm9,%xmm3
|
383
|
-
|
387
|
+
vpaddq %xmm3,%xmm0,%xmm0
|
384
388
|
vpmuludq 624(%rsp),%xmm9,%xmm3
|
385
|
-
|
386
|
-
|
387
|
-
|
388
|
-
|
389
|
-
|
389
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
390
|
+
vpmuludq 704(%rsp),%xmm9,%xmm9
|
391
|
+
vpaddq %xmm9,%xmm1,%xmm1
|
392
|
+
vmovdqa 48(%rsp),%xmm3
|
393
|
+
vmovdqa 96(%rsp),%xmm9
|
390
394
|
vpaddq subc2(%rip),%xmm3,%xmm10
|
391
|
-
|
392
|
-
|
395
|
+
vpsubq %xmm9,%xmm10,%xmm10
|
396
|
+
vpaddq %xmm9,%xmm3,%xmm3
|
393
397
|
vpunpckhqdq %xmm10,%xmm3,%xmm9
|
394
398
|
vpunpcklqdq %xmm10,%xmm3,%xmm3
|
395
399
|
vpmuludq 144(%rsp),%xmm3,%xmm10
|
396
|
-
|
400
|
+
vpaddq %xmm10,%xmm6,%xmm6
|
397
401
|
vpmuludq 128(%rsp),%xmm3,%xmm10
|
398
|
-
|
402
|
+
vpaddq %xmm10,%xmm5,%xmm5
|
399
403
|
vpmuludq 480(%rsp),%xmm3,%xmm10
|
400
|
-
|
404
|
+
vpaddq %xmm10,%xmm8,%xmm8
|
401
405
|
vpmuludq 464(%rsp),%xmm3,%xmm10
|
402
|
-
|
403
|
-
|
406
|
+
vpaddq %xmm10,%xmm7,%xmm7
|
407
|
+
vpmuludq v19_19(%rip),%xmm3,%xmm3
|
404
408
|
vpmuludq 528(%rsp),%xmm3,%xmm10
|
405
|
-
|
409
|
+
vpaddq %xmm10,%xmm11,%xmm11
|
406
410
|
vpmuludq 512(%rsp),%xmm3,%xmm10
|
407
|
-
|
411
|
+
vpaddq %xmm10,%xmm13,%xmm13
|
408
412
|
vpmuludq 592(%rsp),%xmm3,%xmm10
|
409
|
-
|
413
|
+
vpaddq %xmm10,%xmm0,%xmm0
|
410
414
|
vpmuludq 576(%rsp),%xmm3,%xmm10
|
411
|
-
|
415
|
+
vpaddq %xmm10,%xmm2,%xmm2
|
412
416
|
vpmuludq 624(%rsp),%xmm3,%xmm10
|
413
|
-
|
414
|
-
|
415
|
-
|
417
|
+
vpaddq %xmm10,%xmm1,%xmm1
|
418
|
+
vpmuludq 672(%rsp),%xmm3,%xmm3
|
419
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
416
420
|
vpmuludq 144(%rsp),%xmm9,%xmm3
|
417
|
-
|
421
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
418
422
|
vpmuludq 448(%rsp),%xmm9,%xmm3
|
419
|
-
|
423
|
+
vpaddq %xmm3,%xmm8,%xmm8
|
420
424
|
vpmuludq 480(%rsp),%xmm9,%xmm3
|
421
|
-
|
422
|
-
|
425
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
426
|
+
vpmuludq v19_19(%rip),%xmm9,%xmm9
|
423
427
|
vpmuludq 496(%rsp),%xmm9,%xmm3
|
424
|
-
|
428
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
425
429
|
vpmuludq 528(%rsp),%xmm9,%xmm3
|
426
|
-
|
430
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
427
431
|
vpmuludq 544(%rsp),%xmm9,%xmm3
|
428
|
-
|
432
|
+
vpaddq %xmm3,%xmm0,%xmm0
|
429
433
|
vpmuludq 592(%rsp),%xmm9,%xmm3
|
430
|
-
|
434
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
431
435
|
vpmuludq 640(%rsp),%xmm9,%xmm3
|
432
|
-
|
436
|
+
vpaddq %xmm3,%xmm1,%xmm1
|
433
437
|
vpmuludq 624(%rsp),%xmm9,%xmm3
|
434
|
-
|
435
|
-
|
436
|
-
|
437
|
-
|
438
|
-
|
438
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
439
|
+
vpmuludq 704(%rsp),%xmm9,%xmm9
|
440
|
+
vpaddq %xmm9,%xmm6,%xmm6
|
441
|
+
vmovdqa 64(%rsp),%xmm3
|
442
|
+
vmovdqa 112(%rsp),%xmm9
|
439
443
|
vpaddq subc2(%rip),%xmm3,%xmm10
|
440
|
-
|
441
|
-
|
444
|
+
vpsubq %xmm9,%xmm10,%xmm10
|
445
|
+
vpaddq %xmm9,%xmm3,%xmm3
|
442
446
|
vpunpckhqdq %xmm10,%xmm3,%xmm9
|
443
447
|
vpunpcklqdq %xmm10,%xmm3,%xmm3
|
444
448
|
vpmuludq 144(%rsp),%xmm3,%xmm10
|
445
|
-
|
449
|
+
vpaddq %xmm10,%xmm8,%xmm8
|
446
450
|
vpmuludq 128(%rsp),%xmm3,%xmm10
|
447
|
-
|
448
|
-
|
451
|
+
vpaddq %xmm10,%xmm7,%xmm7
|
452
|
+
vpmuludq v19_19(%rip),%xmm3,%xmm3
|
449
453
|
vpmuludq 480(%rsp),%xmm3,%xmm10
|
450
|
-
|
454
|
+
vpaddq %xmm10,%xmm11,%xmm11
|
451
455
|
vpmuludq 464(%rsp),%xmm3,%xmm10
|
452
|
-
|
456
|
+
vpaddq %xmm10,%xmm13,%xmm13
|
453
457
|
vpmuludq 528(%rsp),%xmm3,%xmm10
|
454
|
-
|
458
|
+
vpaddq %xmm10,%xmm0,%xmm0
|
455
459
|
vpmuludq 512(%rsp),%xmm3,%xmm10
|
456
|
-
|
460
|
+
vpaddq %xmm10,%xmm2,%xmm2
|
457
461
|
vpmuludq 592(%rsp),%xmm3,%xmm10
|
458
|
-
|
462
|
+
vpaddq %xmm10,%xmm1,%xmm1
|
459
463
|
vpmuludq 576(%rsp),%xmm3,%xmm10
|
460
|
-
|
464
|
+
vpaddq %xmm10,%xmm4,%xmm4
|
461
465
|
vpmuludq 624(%rsp),%xmm3,%xmm10
|
462
|
-
|
463
|
-
|
464
|
-
|
466
|
+
vpaddq %xmm10,%xmm6,%xmm6
|
467
|
+
vpmuludq 672(%rsp),%xmm3,%xmm3
|
468
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
465
469
|
vpmuludq 144(%rsp),%xmm9,%xmm3
|
466
|
-
|
467
|
-
|
470
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
471
|
+
vpmuludq v19_19(%rip),%xmm9,%xmm9
|
468
472
|
vpmuludq 448(%rsp),%xmm9,%xmm3
|
469
|
-
|
473
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
470
474
|
vpmuludq 480(%rsp),%xmm9,%xmm3
|
471
|
-
|
475
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
472
476
|
vpmuludq 496(%rsp),%xmm9,%xmm3
|
473
|
-
|
477
|
+
vpaddq %xmm3,%xmm0,%xmm0
|
474
478
|
vpmuludq 528(%rsp),%xmm9,%xmm3
|
475
|
-
|
479
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
476
480
|
vpmuludq 544(%rsp),%xmm9,%xmm3
|
477
|
-
|
481
|
+
vpaddq %xmm3,%xmm1,%xmm1
|
478
482
|
vpmuludq 592(%rsp),%xmm9,%xmm3
|
479
|
-
|
483
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
480
484
|
vpmuludq 640(%rsp),%xmm9,%xmm3
|
481
|
-
|
485
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
482
486
|
vpmuludq 624(%rsp),%xmm9,%xmm3
|
483
|
-
|
484
|
-
|
485
|
-
|
487
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
488
|
+
vpmuludq 704(%rsp),%xmm9,%xmm9
|
489
|
+
vpaddq %xmm9,%xmm8,%xmm8
|
486
490
|
vpsrlq $25,%xmm4,%xmm3
|
487
|
-
|
488
|
-
|
491
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
492
|
+
vpand m25(%rip),%xmm4,%xmm4
|
489
493
|
vpsrlq $26,%xmm11,%xmm3
|
490
|
-
|
491
|
-
|
494
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
495
|
+
vpand m26(%rip),%xmm11,%xmm11
|
492
496
|
vpsrlq $26,%xmm6,%xmm3
|
493
|
-
|
494
|
-
|
497
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
498
|
+
vpand m26(%rip),%xmm6,%xmm6
|
495
499
|
vpsrlq $25,%xmm13,%xmm3
|
496
|
-
|
497
|
-
|
500
|
+
vpaddq %xmm3,%xmm0,%xmm0
|
501
|
+
vpand m25(%rip),%xmm13,%xmm13
|
498
502
|
vpsrlq $25,%xmm5,%xmm3
|
499
|
-
|
500
|
-
|
503
|
+
vpaddq %xmm3,%xmm8,%xmm8
|
504
|
+
vpand m25(%rip),%xmm5,%xmm5
|
501
505
|
vpsrlq $26,%xmm0,%xmm3
|
502
|
-
|
503
|
-
|
506
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
507
|
+
vpand m26(%rip),%xmm0,%xmm0
|
504
508
|
vpsrlq $26,%xmm8,%xmm3
|
505
|
-
|
506
|
-
|
509
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
510
|
+
vpand m26(%rip),%xmm8,%xmm8
|
507
511
|
vpsrlq $25,%xmm2,%xmm3
|
508
|
-
|
509
|
-
|
512
|
+
vpaddq %xmm3,%xmm1,%xmm1
|
513
|
+
vpand m25(%rip),%xmm2,%xmm2
|
510
514
|
vpsrlq $25,%xmm7,%xmm3
|
511
515
|
vpsllq $4,%xmm3,%xmm9
|
512
|
-
|
513
|
-
|
514
|
-
|
515
|
-
|
516
|
-
|
516
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
517
|
+
vpsllq $1,%xmm3,%xmm3
|
518
|
+
vpaddq %xmm3,%xmm9,%xmm9
|
519
|
+
vpaddq %xmm9,%xmm11,%xmm11
|
520
|
+
vpand m25(%rip),%xmm7,%xmm7
|
517
521
|
vpsrlq $26,%xmm1,%xmm3
|
518
|
-
|
519
|
-
|
522
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
523
|
+
vpand m26(%rip),%xmm1,%xmm1
|
520
524
|
vpsrlq $26,%xmm11,%xmm3
|
521
|
-
|
522
|
-
|
525
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
526
|
+
vpand m26(%rip),%xmm11,%xmm11
|
523
527
|
vpsrlq $25,%xmm4,%xmm3
|
524
|
-
|
525
|
-
|
528
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
529
|
+
vpand m25(%rip),%xmm4,%xmm4
|
526
530
|
vpunpcklqdq %xmm13,%xmm11,%xmm3
|
527
531
|
vpunpckhqdq %xmm13,%xmm11,%xmm9
|
528
532
|
vpaddq subc0(%rip),%xmm9,%xmm10
|
529
|
-
|
530
|
-
|
533
|
+
vpsubq %xmm3,%xmm10,%xmm10
|
534
|
+
vpaddq %xmm9,%xmm3,%xmm3
|
531
535
|
vpunpckhqdq %xmm3,%xmm10,%xmm9
|
532
|
-
|
536
|
+
vpunpcklqdq %xmm3,%xmm10,%xmm10
|
533
537
|
vpmuludq %xmm10,%xmm10,%xmm3
|
534
|
-
|
538
|
+
vpaddq %xmm10,%xmm10,%xmm10
|
535
539
|
vpmuludq %xmm9,%xmm10,%xmm11
|
536
540
|
vpunpcklqdq %xmm2,%xmm0,%xmm12
|
537
541
|
vpunpckhqdq %xmm2,%xmm0,%xmm0
|
538
542
|
vpaddq subc2(%rip),%xmm0,%xmm2
|
539
|
-
|
540
|
-
|
543
|
+
vpsubq %xmm12,%xmm2,%xmm2
|
544
|
+
vpaddq %xmm0,%xmm12,%xmm12
|
541
545
|
vpunpckhqdq %xmm12,%xmm2,%xmm0
|
542
|
-
|
546
|
+
vpunpcklqdq %xmm12,%xmm2,%xmm2
|
543
547
|
vpmuludq %xmm2,%xmm10,%xmm12
|
544
548
|
vpaddq %xmm9,%xmm9,%xmm13
|
545
549
|
vpmuludq %xmm13,%xmm9,%xmm9
|
546
|
-
|
550
|
+
vpaddq %xmm9,%xmm12,%xmm12
|
547
551
|
vpmuludq %xmm0,%xmm10,%xmm9
|
548
552
|
vpmuludq %xmm2,%xmm13,%xmm14
|
549
|
-
|
553
|
+
vpaddq %xmm14,%xmm9,%xmm9
|
550
554
|
vpunpcklqdq %xmm4,%xmm1,%xmm14
|
551
555
|
vpunpckhqdq %xmm4,%xmm1,%xmm1
|
552
556
|
vpaddq subc2(%rip),%xmm1,%xmm4
|
553
|
-
|
554
|
-
|
557
|
+
vpsubq %xmm14,%xmm4,%xmm4
|
558
|
+
vpaddq %xmm1,%xmm14,%xmm14
|
555
559
|
vpunpckhqdq %xmm14,%xmm4,%xmm1
|
556
|
-
|
557
|
-
|
558
|
-
|
559
|
-
|
560
|
-
|
561
|
-
|
560
|
+
vpunpcklqdq %xmm14,%xmm4,%xmm4
|
561
|
+
vmovdqa %xmm1,0(%rsp)
|
562
|
+
vpaddq %xmm1,%xmm1,%xmm1
|
563
|
+
vmovdqa %xmm1,16(%rsp)
|
564
|
+
vpmuludq v19_19(%rip),%xmm1,%xmm1
|
565
|
+
vmovdqa %xmm1,32(%rsp)
|
562
566
|
vpmuludq %xmm4,%xmm10,%xmm1
|
563
567
|
vpmuludq %xmm2,%xmm2,%xmm14
|
564
|
-
|
568
|
+
vpaddq %xmm14,%xmm1,%xmm1
|
565
569
|
vpmuludq 0(%rsp),%xmm10,%xmm14
|
566
570
|
vpmuludq %xmm4,%xmm13,%xmm15
|
567
|
-
|
571
|
+
vpaddq %xmm15,%xmm14,%xmm14
|
568
572
|
vpunpcklqdq %xmm5,%xmm6,%xmm15
|
569
573
|
vpunpckhqdq %xmm5,%xmm6,%xmm5
|
570
574
|
vpaddq subc2(%rip),%xmm5,%xmm6
|
571
|
-
|
572
|
-
|
575
|
+
vpsubq %xmm15,%xmm6,%xmm6
|
576
|
+
vpaddq %xmm5,%xmm15,%xmm15
|
573
577
|
vpunpckhqdq %xmm15,%xmm6,%xmm5
|
574
|
-
|
575
|
-
|
576
|
-
|
577
|
-
|
578
|
-
|
579
|
-
|
580
|
-
|
578
|
+
vpunpcklqdq %xmm15,%xmm6,%xmm6
|
579
|
+
vmovdqa %xmm6,48(%rsp)
|
580
|
+
vpmuludq v19_19(%rip),%xmm6,%xmm6
|
581
|
+
vmovdqa %xmm6,64(%rsp)
|
582
|
+
vmovdqa %xmm5,80(%rsp)
|
583
|
+
vpmuludq v38_38(%rip),%xmm5,%xmm5
|
584
|
+
vmovdqa %xmm5,96(%rsp)
|
581
585
|
vpmuludq 48(%rsp),%xmm10,%xmm5
|
582
586
|
vpaddq %xmm0,%xmm0,%xmm6
|
583
587
|
vpmuludq %xmm6,%xmm0,%xmm0
|
584
|
-
|
588
|
+
vpaddq %xmm0,%xmm5,%xmm5
|
585
589
|
vpmuludq 80(%rsp),%xmm10,%xmm0
|
586
590
|
vpmuludq %xmm4,%xmm6,%xmm15
|
587
|
-
|
591
|
+
vpaddq %xmm15,%xmm0,%xmm0
|
588
592
|
vpmuludq %xmm6,%xmm13,%xmm15
|
589
|
-
|
593
|
+
vpaddq %xmm15,%xmm1,%xmm1
|
590
594
|
vpmuludq %xmm6,%xmm2,%xmm15
|
591
|
-
|
595
|
+
vpaddq %xmm15,%xmm14,%xmm14
|
592
596
|
vpunpcklqdq %xmm7,%xmm8,%xmm15
|
593
597
|
vpunpckhqdq %xmm7,%xmm8,%xmm7
|
594
598
|
vpaddq subc2(%rip),%xmm7,%xmm8
|
595
|
-
|
596
|
-
|
599
|
+
vpsubq %xmm15,%xmm8,%xmm8
|
600
|
+
vpaddq %xmm7,%xmm15,%xmm15
|
597
601
|
vpunpckhqdq %xmm15,%xmm8,%xmm7
|
598
|
-
|
599
|
-
|
600
|
-
|
601
|
-
|
602
|
+
vpunpcklqdq %xmm15,%xmm8,%xmm8
|
603
|
+
vmovdqa %xmm8,112(%rsp)
|
604
|
+
vpmuludq v19_19(%rip),%xmm8,%xmm8
|
605
|
+
vmovdqa %xmm8,448(%rsp)
|
602
606
|
vpmuludq 112(%rsp),%xmm10,%xmm8
|
603
607
|
vpmuludq %xmm7,%xmm10,%xmm10
|
604
608
|
vpmuludq v38_38(%rip),%xmm7,%xmm15
|
605
609
|
vpmuludq %xmm15,%xmm7,%xmm7
|
606
|
-
|
610
|
+
vpaddq %xmm7,%xmm8,%xmm8
|
607
611
|
vpmuludq %xmm15,%xmm13,%xmm7
|
608
|
-
|
612
|
+
vpaddq %xmm7,%xmm3,%xmm3
|
609
613
|
vpmuludq %xmm15,%xmm2,%xmm7
|
610
|
-
|
614
|
+
vpaddq %xmm7,%xmm11,%xmm11
|
611
615
|
vpmuludq 80(%rsp),%xmm13,%xmm7
|
612
|
-
|
613
|
-
|
616
|
+
vpaddq %xmm7,%xmm7,%xmm7
|
617
|
+
vpaddq %xmm7,%xmm8,%xmm8
|
614
618
|
vpmuludq 16(%rsp),%xmm13,%xmm7
|
615
|
-
|
619
|
+
vpaddq %xmm7,%xmm5,%xmm5
|
616
620
|
vpmuludq 48(%rsp),%xmm13,%xmm7
|
617
|
-
|
621
|
+
vpaddq %xmm7,%xmm0,%xmm0
|
618
622
|
vpmuludq 112(%rsp),%xmm13,%xmm7
|
619
|
-
|
623
|
+
vpaddq %xmm7,%xmm10,%xmm10
|
620
624
|
vpmuludq %xmm15,%xmm6,%xmm7
|
621
|
-
|
625
|
+
vpaddq %xmm7,%xmm12,%xmm12
|
622
626
|
vpmuludq %xmm15,%xmm4,%xmm7
|
623
|
-
|
627
|
+
vpaddq %xmm7,%xmm9,%xmm9
|
624
628
|
vpaddq %xmm2,%xmm2,%xmm2
|
625
629
|
vpmuludq %xmm4,%xmm2,%xmm7
|
626
|
-
|
630
|
+
vpaddq %xmm7,%xmm5,%xmm5
|
627
631
|
vpmuludq 448(%rsp),%xmm2,%xmm7
|
628
|
-
|
632
|
+
vpaddq %xmm7,%xmm3,%xmm3
|
629
633
|
vpmuludq 448(%rsp),%xmm6,%xmm7
|
630
|
-
|
634
|
+
vpaddq %xmm7,%xmm11,%xmm11
|
631
635
|
vpmuludq 0(%rsp),%xmm2,%xmm7
|
632
|
-
|
636
|
+
vpaddq %xmm7,%xmm0,%xmm0
|
633
637
|
vpmuludq 48(%rsp),%xmm2,%xmm7
|
634
|
-
|
638
|
+
vpaddq %xmm7,%xmm8,%xmm8
|
635
639
|
vpmuludq 80(%rsp),%xmm2,%xmm2
|
636
|
-
|
640
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
637
641
|
vpmuludq 96(%rsp),%xmm4,%xmm2
|
638
|
-
|
642
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
639
643
|
vpmuludq %xmm4,%xmm4,%xmm2
|
640
|
-
|
644
|
+
vpaddq %xmm2,%xmm8,%xmm8
|
641
645
|
vpaddq %xmm4,%xmm4,%xmm2
|
642
646
|
vpmuludq 448(%rsp),%xmm2,%xmm4
|
643
|
-
|
647
|
+
vpaddq %xmm4,%xmm12,%xmm12
|
644
648
|
vpmuludq 16(%rsp),%xmm15,%xmm4
|
645
|
-
|
649
|
+
vpaddq %xmm4,%xmm1,%xmm1
|
646
650
|
vpmuludq 48(%rsp),%xmm15,%xmm4
|
647
|
-
|
651
|
+
vpaddq %xmm4,%xmm14,%xmm14
|
648
652
|
vpmuludq 96(%rsp),%xmm6,%xmm4
|
649
|
-
|
650
|
-
|
651
|
-
|
652
|
-
|
653
|
+
vpaddq %xmm4,%xmm3,%xmm3
|
654
|
+
vmovdqa 16(%rsp),%xmm4
|
655
|
+
vpmuludq 448(%rsp),%xmm4,%xmm4
|
656
|
+
vpaddq %xmm4,%xmm9,%xmm9
|
653
657
|
vpmuludq 16(%rsp),%xmm6,%xmm4
|
654
|
-
|
658
|
+
vpaddq %xmm4,%xmm8,%xmm8
|
655
659
|
vpmuludq 48(%rsp),%xmm6,%xmm4
|
656
|
-
|
660
|
+
vpaddq %xmm4,%xmm10,%xmm10
|
657
661
|
vpmuludq 80(%rsp),%xmm15,%xmm4
|
658
|
-
|
659
|
-
|
662
|
+
vpaddq %xmm4,%xmm4,%xmm4
|
663
|
+
vpaddq %xmm4,%xmm5,%xmm5
|
660
664
|
vpmuludq 112(%rsp),%xmm15,%xmm4
|
661
|
-
|
662
|
-
|
663
|
-
|
664
|
-
|
665
|
-
|
666
|
-
|
667
|
-
|
668
|
-
|
669
|
-
|
665
|
+
vpaddq %xmm4,%xmm0,%xmm0
|
666
|
+
vmovdqa 48(%rsp),%xmm4
|
667
|
+
vpaddq %xmm4,%xmm4,%xmm4
|
668
|
+
vpmuludq 448(%rsp),%xmm4,%xmm4
|
669
|
+
vpaddq %xmm4,%xmm1,%xmm1
|
670
|
+
vmovdqa 80(%rsp),%xmm4
|
671
|
+
vpaddq %xmm4,%xmm4,%xmm4
|
672
|
+
vpmuludq 448(%rsp),%xmm4,%xmm4
|
673
|
+
vpaddq %xmm4,%xmm14,%xmm14
|
670
674
|
vpmuludq 64(%rsp),%xmm2,%xmm4
|
671
|
-
|
672
|
-
|
673
|
-
|
674
|
-
|
675
|
-
|
676
|
-
|
677
|
-
|
678
|
-
|
679
|
-
|
680
|
-
|
675
|
+
vpaddq %xmm4,%xmm3,%xmm3
|
676
|
+
vmovdqa 16(%rsp),%xmm4
|
677
|
+
vpmuludq 64(%rsp),%xmm4,%xmm4
|
678
|
+
vpaddq %xmm4,%xmm11,%xmm11
|
679
|
+
vmovdqa 16(%rsp),%xmm4
|
680
|
+
vpmuludq 96(%rsp),%xmm4,%xmm4
|
681
|
+
vpaddq %xmm4,%xmm12,%xmm12
|
682
|
+
vmovdqa 48(%rsp),%xmm4
|
683
|
+
vpmuludq 96(%rsp),%xmm4,%xmm4
|
684
|
+
vpaddq %xmm4,%xmm9,%xmm9
|
685
|
+
vpmuludq 0(%rsp),%xmm2,%xmm2
|
686
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
687
|
+
vmovdqa 32(%rsp),%xmm2
|
681
688
|
vpmuludq 0(%rsp),%xmm2,%xmm2
|
682
|
-
|
683
|
-
|
684
|
-
|
685
|
-
|
686
|
-
|
687
|
-
|
688
|
-
|
689
|
-
|
690
|
-
|
691
|
-
|
692
|
-
movdqa 448(%rsp),%xmm2
|
693
|
-
pmuludq 112(%rsp),%xmm2
|
694
|
-
paddq %xmm2,%xmm5
|
689
|
+
vpaddq %xmm2,%xmm3,%xmm3
|
690
|
+
vmovdqa 64(%rsp),%xmm2
|
691
|
+
vpmuludq 48(%rsp),%xmm2,%xmm2
|
692
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
693
|
+
vmovdqa 96(%rsp),%xmm2
|
694
|
+
vpmuludq 80(%rsp),%xmm2,%xmm2
|
695
|
+
vpaddq %xmm2,%xmm1,%xmm1
|
696
|
+
vmovdqa 448(%rsp),%xmm2
|
697
|
+
vpmuludq 112(%rsp),%xmm2,%xmm2
|
698
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
695
699
|
vpsrlq $26,%xmm3,%xmm2
|
696
|
-
|
697
|
-
|
700
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
701
|
+
vpand m26(%rip),%xmm3,%xmm3
|
698
702
|
vpsrlq $25,%xmm14,%xmm2
|
699
|
-
|
700
|
-
|
703
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
704
|
+
vpand m25(%rip),%xmm14,%xmm14
|
701
705
|
vpsrlq $25,%xmm11,%xmm2
|
702
|
-
|
703
|
-
|
706
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
707
|
+
vpand m25(%rip),%xmm11,%xmm11
|
704
708
|
vpsrlq $26,%xmm5,%xmm2
|
705
|
-
|
706
|
-
|
709
|
+
vpaddq %xmm2,%xmm0,%xmm0
|
710
|
+
vpand m26(%rip),%xmm5,%xmm5
|
707
711
|
vpsrlq $26,%xmm12,%xmm2
|
708
|
-
|
709
|
-
|
712
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
713
|
+
vpand m26(%rip),%xmm12,%xmm12
|
710
714
|
vpsrlq $25,%xmm0,%xmm2
|
711
|
-
|
712
|
-
|
715
|
+
vpaddq %xmm2,%xmm8,%xmm8
|
716
|
+
vpand m25(%rip),%xmm0,%xmm0
|
713
717
|
vpsrlq $25,%xmm9,%xmm2
|
714
|
-
|
715
|
-
|
718
|
+
vpaddq %xmm2,%xmm1,%xmm1
|
719
|
+
vpand m25(%rip),%xmm9,%xmm9
|
716
720
|
vpsrlq $26,%xmm8,%xmm2
|
717
|
-
|
718
|
-
|
721
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
722
|
+
vpand m26(%rip),%xmm8,%xmm8
|
719
723
|
vpsrlq $26,%xmm1,%xmm2
|
720
|
-
|
721
|
-
|
724
|
+
vpaddq %xmm2,%xmm14,%xmm14
|
725
|
+
vpand m26(%rip),%xmm1,%xmm1
|
722
726
|
vpsrlq $25,%xmm10,%xmm2
|
723
727
|
vpsllq $4,%xmm2,%xmm4
|
724
|
-
|
725
|
-
|
726
|
-
|
727
|
-
|
728
|
-
|
728
|
+
vpaddq %xmm2,%xmm3,%xmm3
|
729
|
+
vpsllq $1,%xmm2,%xmm2
|
730
|
+
vpaddq %xmm2,%xmm4,%xmm4
|
731
|
+
vpaddq %xmm4,%xmm3,%xmm3
|
732
|
+
vpand m25(%rip),%xmm10,%xmm10
|
729
733
|
vpsrlq $25,%xmm14,%xmm2
|
730
|
-
|
731
|
-
|
734
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
735
|
+
vpand m25(%rip),%xmm14,%xmm14
|
732
736
|
vpsrlq $26,%xmm3,%xmm2
|
733
|
-
|
734
|
-
|
737
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
738
|
+
vpand m26(%rip),%xmm3,%xmm3
|
735
739
|
vpunpckhqdq %xmm11,%xmm3,%xmm2
|
736
|
-
|
737
|
-
|
738
|
-
|
740
|
+
vmovdqa %xmm2,0(%rsp)
|
741
|
+
vpshufd $0,%xmm3,%xmm2
|
742
|
+
vpshufd $0,%xmm11,%xmm3
|
739
743
|
vpmuludq 160(%rsp),%xmm2,%xmm4
|
740
744
|
vpmuludq 432(%rsp),%xmm3,%xmm6
|
741
|
-
|
745
|
+
vpaddq %xmm6,%xmm4,%xmm4
|
742
746
|
vpmuludq 176(%rsp),%xmm2,%xmm6
|
743
747
|
vpmuludq 304(%rsp),%xmm3,%xmm7
|
744
|
-
|
748
|
+
vpaddq %xmm7,%xmm6,%xmm6
|
745
749
|
vpmuludq 208(%rsp),%xmm2,%xmm7
|
746
750
|
vpmuludq 336(%rsp),%xmm3,%xmm11
|
747
|
-
|
751
|
+
vpaddq %xmm11,%xmm7,%xmm7
|
748
752
|
vpmuludq 240(%rsp),%xmm2,%xmm11
|
749
753
|
vpmuludq 368(%rsp),%xmm3,%xmm13
|
750
|
-
|
754
|
+
vpaddq %xmm13,%xmm11,%xmm11
|
751
755
|
vpmuludq 272(%rsp),%xmm2,%xmm2
|
752
756
|
vpmuludq 400(%rsp),%xmm3,%xmm3
|
753
|
-
|
757
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
754
758
|
vpunpckhqdq %xmm9,%xmm12,%xmm3
|
755
|
-
|
756
|
-
|
757
|
-
|
759
|
+
vmovdqa %xmm3,16(%rsp)
|
760
|
+
vpshufd $0,%xmm12,%xmm3
|
761
|
+
vpshufd $0,%xmm9,%xmm9
|
758
762
|
vpmuludq 288(%rsp),%xmm3,%xmm12
|
759
|
-
|
763
|
+
vpaddq %xmm12,%xmm4,%xmm4
|
760
764
|
vpmuludq 416(%rsp),%xmm9,%xmm12
|
761
|
-
|
765
|
+
vpaddq %xmm12,%xmm4,%xmm4
|
762
766
|
vpmuludq 160(%rsp),%xmm3,%xmm12
|
763
|
-
|
767
|
+
vpaddq %xmm12,%xmm6,%xmm6
|
764
768
|
vpmuludq 432(%rsp),%xmm9,%xmm12
|
765
|
-
|
769
|
+
vpaddq %xmm12,%xmm6,%xmm6
|
766
770
|
vpmuludq 176(%rsp),%xmm3,%xmm12
|
767
|
-
|
771
|
+
vpaddq %xmm12,%xmm7,%xmm7
|
768
772
|
vpmuludq 304(%rsp),%xmm9,%xmm12
|
769
|
-
|
773
|
+
vpaddq %xmm12,%xmm7,%xmm7
|
770
774
|
vpmuludq 208(%rsp),%xmm3,%xmm12
|
771
|
-
|
775
|
+
vpaddq %xmm12,%xmm11,%xmm11
|
772
776
|
vpmuludq 336(%rsp),%xmm9,%xmm12
|
773
|
-
|
777
|
+
vpaddq %xmm12,%xmm11,%xmm11
|
774
778
|
vpmuludq 240(%rsp),%xmm3,%xmm3
|
775
|
-
|
779
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
776
780
|
vpmuludq 368(%rsp),%xmm9,%xmm3
|
777
|
-
|
781
|
+
vpaddq %xmm3,%xmm2,%xmm2
|
778
782
|
vpunpckhqdq %xmm14,%xmm1,%xmm3
|
779
|
-
|
780
|
-
|
781
|
-
|
783
|
+
vmovdqa %xmm3,32(%rsp)
|
784
|
+
vpshufd $0,%xmm1,%xmm1
|
785
|
+
vpshufd $0,%xmm14,%xmm3
|
782
786
|
vpmuludq 256(%rsp),%xmm1,%xmm9
|
783
|
-
|
787
|
+
vpaddq %xmm9,%xmm4,%xmm4
|
784
788
|
vpmuludq 384(%rsp),%xmm3,%xmm9
|
785
|
-
|
789
|
+
vpaddq %xmm9,%xmm4,%xmm4
|
786
790
|
vpmuludq 288(%rsp),%xmm1,%xmm9
|
787
|
-
|
791
|
+
vpaddq %xmm9,%xmm6,%xmm6
|
788
792
|
vpmuludq 416(%rsp),%xmm3,%xmm9
|
789
|
-
|
793
|
+
vpaddq %xmm9,%xmm6,%xmm6
|
790
794
|
vpmuludq 160(%rsp),%xmm1,%xmm9
|
791
|
-
|
795
|
+
vpaddq %xmm9,%xmm7,%xmm7
|
792
796
|
vpmuludq 432(%rsp),%xmm3,%xmm9
|
793
|
-
|
797
|
+
vpaddq %xmm9,%xmm7,%xmm7
|
794
798
|
vpmuludq 176(%rsp),%xmm1,%xmm9
|
795
|
-
|
799
|
+
vpaddq %xmm9,%xmm11,%xmm11
|
796
800
|
vpmuludq 304(%rsp),%xmm3,%xmm9
|
797
|
-
|
801
|
+
vpaddq %xmm9,%xmm11,%xmm11
|
798
802
|
vpmuludq 208(%rsp),%xmm1,%xmm1
|
799
|
-
|
803
|
+
vpaddq %xmm1,%xmm2,%xmm2
|
800
804
|
vpmuludq 336(%rsp),%xmm3,%xmm1
|
801
|
-
|
805
|
+
vpaddq %xmm1,%xmm2,%xmm2
|
802
806
|
vpunpckhqdq %xmm0,%xmm5,%xmm1
|
803
|
-
|
804
|
-
|
805
|
-
|
807
|
+
vmovdqa %xmm1,48(%rsp)
|
808
|
+
vpshufd $0,%xmm5,%xmm1
|
809
|
+
vpshufd $0,%xmm0,%xmm0
|
806
810
|
vpmuludq 224(%rsp),%xmm1,%xmm3
|
807
|
-
|
811
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
808
812
|
vpmuludq 352(%rsp),%xmm0,%xmm3
|
809
|
-
|
813
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
810
814
|
vpmuludq 256(%rsp),%xmm1,%xmm3
|
811
|
-
|
815
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
812
816
|
vpmuludq 384(%rsp),%xmm0,%xmm3
|
813
|
-
|
817
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
814
818
|
vpmuludq 288(%rsp),%xmm1,%xmm3
|
815
|
-
|
819
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
816
820
|
vpmuludq 416(%rsp),%xmm0,%xmm3
|
817
|
-
|
821
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
818
822
|
vpmuludq 160(%rsp),%xmm1,%xmm3
|
819
|
-
|
823
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
820
824
|
vpmuludq 432(%rsp),%xmm0,%xmm3
|
821
|
-
|
825
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
822
826
|
vpmuludq 176(%rsp),%xmm1,%xmm1
|
823
|
-
|
827
|
+
vpaddq %xmm1,%xmm2,%xmm2
|
824
828
|
vpmuludq 304(%rsp),%xmm0,%xmm0
|
825
|
-
|
829
|
+
vpaddq %xmm0,%xmm2,%xmm2
|
826
830
|
vpunpckhqdq %xmm10,%xmm8,%xmm0
|
827
|
-
|
828
|
-
|
829
|
-
|
831
|
+
vmovdqa %xmm0,64(%rsp)
|
832
|
+
vpshufd $0,%xmm8,%xmm0
|
833
|
+
vpshufd $0,%xmm10,%xmm1
|
830
834
|
vpmuludq 192(%rsp),%xmm0,%xmm3
|
831
|
-
|
835
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
832
836
|
vpmuludq 320(%rsp),%xmm1,%xmm3
|
833
|
-
|
837
|
+
vpaddq %xmm3,%xmm4,%xmm4
|
834
838
|
vpmuludq 224(%rsp),%xmm0,%xmm3
|
835
|
-
|
839
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
836
840
|
vpmuludq 352(%rsp),%xmm1,%xmm3
|
837
|
-
|
841
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
838
842
|
vpmuludq 256(%rsp),%xmm0,%xmm3
|
839
|
-
|
843
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
840
844
|
vpmuludq 384(%rsp),%xmm1,%xmm3
|
841
|
-
|
845
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
842
846
|
vpmuludq 288(%rsp),%xmm0,%xmm3
|
843
|
-
|
847
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
844
848
|
vpmuludq 416(%rsp),%xmm1,%xmm3
|
845
|
-
|
849
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
846
850
|
vpmuludq 160(%rsp),%xmm0,%xmm0
|
847
|
-
|
851
|
+
vpaddq %xmm0,%xmm2,%xmm2
|
848
852
|
vpmuludq 432(%rsp),%xmm1,%xmm0
|
849
|
-
|
850
|
-
|
851
|
-
|
852
|
-
|
853
|
-
|
854
|
-
|
855
|
-
|
853
|
+
vpaddq %xmm0,%xmm2,%xmm2
|
854
|
+
vmovdqa %xmm4,80(%rsp)
|
855
|
+
vmovdqa %xmm6,96(%rsp)
|
856
|
+
vmovdqa %xmm7,112(%rsp)
|
857
|
+
vmovdqa %xmm11,448(%rsp)
|
858
|
+
vmovdqa %xmm2,496(%rsp)
|
859
|
+
vmovdqa 144(%rsp),%xmm0
|
856
860
|
vpmuludq %xmm0,%xmm0,%xmm1
|
857
|
-
|
858
|
-
|
861
|
+
vpaddq %xmm0,%xmm0,%xmm0
|
862
|
+
vmovdqa 128(%rsp),%xmm2
|
859
863
|
vpmuludq %xmm2,%xmm0,%xmm3
|
860
|
-
|
864
|
+
vmovdqa 480(%rsp),%xmm4
|
861
865
|
vpmuludq %xmm4,%xmm0,%xmm5
|
862
|
-
|
866
|
+
vmovdqa 464(%rsp),%xmm6
|
863
867
|
vpmuludq %xmm6,%xmm0,%xmm7
|
864
|
-
|
868
|
+
vmovdqa 528(%rsp),%xmm8
|
865
869
|
vpmuludq %xmm8,%xmm0,%xmm9
|
866
870
|
vpmuludq 512(%rsp),%xmm0,%xmm10
|
867
871
|
vpmuludq 592(%rsp),%xmm0,%xmm11
|
868
872
|
vpmuludq 576(%rsp),%xmm0,%xmm12
|
869
873
|
vpmuludq 624(%rsp),%xmm0,%xmm13
|
870
|
-
|
874
|
+
vmovdqa 672(%rsp),%xmm14
|
871
875
|
vpmuludq %xmm14,%xmm0,%xmm0
|
872
876
|
vpmuludq v38_38(%rip),%xmm14,%xmm15
|
873
877
|
vpmuludq %xmm15,%xmm14,%xmm14
|
874
|
-
|
878
|
+
vpaddq %xmm14,%xmm13,%xmm13
|
875
879
|
vpaddq %xmm6,%xmm6,%xmm14
|
876
880
|
vpmuludq %xmm14,%xmm6,%xmm6
|
877
|
-
|
881
|
+
vpaddq %xmm6,%xmm11,%xmm11
|
878
882
|
vpaddq %xmm2,%xmm2,%xmm6
|
879
883
|
vpmuludq %xmm6,%xmm2,%xmm2
|
880
|
-
|
884
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
881
885
|
vpmuludq %xmm15,%xmm6,%xmm2
|
882
|
-
|
886
|
+
vpaddq %xmm2,%xmm1,%xmm1
|
883
887
|
vpmuludq %xmm15,%xmm4,%xmm2
|
884
|
-
|
888
|
+
vpaddq %xmm2,%xmm3,%xmm3
|
885
889
|
vpmuludq 544(%rsp),%xmm6,%xmm2
|
886
|
-
|
890
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
887
891
|
vpmuludq 592(%rsp),%xmm6,%xmm2
|
888
|
-
|
892
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
889
893
|
vpmuludq 640(%rsp),%xmm6,%xmm2
|
890
|
-
|
894
|
+
vpaddq %xmm2,%xmm13,%xmm13
|
891
895
|
vpmuludq 624(%rsp),%xmm6,%xmm2
|
892
|
-
|
896
|
+
vpaddq %xmm2,%xmm0,%xmm0
|
893
897
|
vpmuludq %xmm4,%xmm6,%xmm2
|
894
|
-
|
898
|
+
vpaddq %xmm2,%xmm7,%xmm7
|
895
899
|
vpmuludq %xmm14,%xmm6,%xmm2
|
896
|
-
|
900
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
897
901
|
vpmuludq %xmm8,%xmm6,%xmm2
|
898
|
-
|
902
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
899
903
|
vpmuludq %xmm15,%xmm14,%xmm2
|
900
|
-
|
904
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
901
905
|
vpmuludq %xmm15,%xmm8,%xmm2
|
902
|
-
|
906
|
+
vpaddq %xmm2,%xmm7,%xmm7
|
903
907
|
vpmuludq %xmm4,%xmm4,%xmm2
|
904
|
-
|
908
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
905
909
|
vpmuludq %xmm14,%xmm4,%xmm2
|
906
|
-
|
910
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
907
911
|
vpaddq %xmm4,%xmm4,%xmm2
|
908
912
|
vpmuludq %xmm8,%xmm2,%xmm4
|
909
|
-
|
913
|
+
vpaddq %xmm4,%xmm11,%xmm11
|
910
914
|
vpmuludq 688(%rsp),%xmm2,%xmm4
|
911
|
-
|
915
|
+
vpaddq %xmm4,%xmm1,%xmm1
|
912
916
|
vpmuludq 688(%rsp),%xmm14,%xmm4
|
913
|
-
|
917
|
+
vpaddq %xmm4,%xmm3,%xmm3
|
914
918
|
vpmuludq 512(%rsp),%xmm2,%xmm4
|
915
|
-
|
919
|
+
vpaddq %xmm4,%xmm12,%xmm12
|
916
920
|
vpmuludq 592(%rsp),%xmm2,%xmm4
|
917
|
-
|
921
|
+
vpaddq %xmm4,%xmm13,%xmm13
|
918
922
|
vpmuludq 576(%rsp),%xmm2,%xmm2
|
919
|
-
|
923
|
+
vpaddq %xmm2,%xmm0,%xmm0
|
920
924
|
vpmuludq 656(%rsp),%xmm8,%xmm2
|
921
|
-
|
925
|
+
vpaddq %xmm2,%xmm3,%xmm3
|
922
926
|
vpmuludq %xmm8,%xmm14,%xmm2
|
923
|
-
|
927
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
924
928
|
vpmuludq %xmm8,%xmm8,%xmm2
|
925
|
-
|
929
|
+
vpaddq %xmm2,%xmm13,%xmm13
|
926
930
|
vpaddq %xmm8,%xmm8,%xmm2
|
927
931
|
vpmuludq 688(%rsp),%xmm2,%xmm4
|
928
|
-
|
932
|
+
vpaddq %xmm4,%xmm5,%xmm5
|
929
933
|
vpmuludq 544(%rsp),%xmm15,%xmm4
|
930
|
-
|
934
|
+
vpaddq %xmm4,%xmm9,%xmm9
|
931
935
|
vpmuludq 592(%rsp),%xmm15,%xmm4
|
932
|
-
|
936
|
+
vpaddq %xmm4,%xmm10,%xmm10
|
933
937
|
vpmuludq 656(%rsp),%xmm14,%xmm4
|
934
|
-
|
935
|
-
|
936
|
-
|
937
|
-
|
938
|
+
vpaddq %xmm4,%xmm1,%xmm1
|
939
|
+
vmovdqa 544(%rsp),%xmm4
|
940
|
+
vpmuludq 688(%rsp),%xmm4,%xmm4
|
941
|
+
vpaddq %xmm4,%xmm7,%xmm7
|
938
942
|
vpmuludq 544(%rsp),%xmm14,%xmm4
|
939
|
-
|
943
|
+
vpaddq %xmm4,%xmm13,%xmm13
|
940
944
|
vpmuludq 592(%rsp),%xmm14,%xmm4
|
941
|
-
|
945
|
+
vpaddq %xmm4,%xmm0,%xmm0
|
942
946
|
vpmuludq 640(%rsp),%xmm15,%xmm4
|
943
|
-
|
947
|
+
vpaddq %xmm4,%xmm11,%xmm11
|
944
948
|
vpmuludq 624(%rsp),%xmm15,%xmm4
|
945
|
-
|
946
|
-
|
947
|
-
|
948
|
-
|
949
|
-
|
949
|
+
vpaddq %xmm4,%xmm12,%xmm12
|
950
|
+
vmovdqa 592(%rsp),%xmm4
|
951
|
+
vpaddq %xmm4,%xmm4,%xmm4
|
952
|
+
vpmuludq 688(%rsp),%xmm4,%xmm4
|
953
|
+
vpaddq %xmm4,%xmm9,%xmm9
|
950
954
|
vpmuludq 608(%rsp),%xmm2,%xmm4
|
951
|
-
|
952
|
-
|
953
|
-
|
954
|
-
|
955
|
-
|
956
|
-
|
957
|
-
|
958
|
-
|
959
|
-
|
960
|
-
|
961
|
-
|
962
|
-
|
963
|
-
|
955
|
+
vpaddq %xmm4,%xmm1,%xmm1
|
956
|
+
vmovdqa 544(%rsp),%xmm4
|
957
|
+
vpmuludq 608(%rsp),%xmm4,%xmm4
|
958
|
+
vpaddq %xmm4,%xmm3,%xmm3
|
959
|
+
vmovdqa 544(%rsp),%xmm4
|
960
|
+
vpmuludq 656(%rsp),%xmm4,%xmm4
|
961
|
+
vpaddq %xmm4,%xmm5,%xmm5
|
962
|
+
vmovdqa 592(%rsp),%xmm4
|
963
|
+
vpmuludq 656(%rsp),%xmm4,%xmm4
|
964
|
+
vpaddq %xmm4,%xmm7,%xmm7
|
965
|
+
vmovdqa 640(%rsp),%xmm4
|
966
|
+
vpmuludq 688(%rsp),%xmm4,%xmm4
|
967
|
+
vpaddq %xmm4,%xmm10,%xmm10
|
968
|
+
vpmuludq 512(%rsp),%xmm2,%xmm2
|
969
|
+
vpaddq %xmm2,%xmm0,%xmm0
|
970
|
+
vmovdqa 560(%rsp),%xmm2
|
964
971
|
vpmuludq 512(%rsp),%xmm2,%xmm2
|
965
|
-
|
966
|
-
|
967
|
-
|
968
|
-
|
969
|
-
|
970
|
-
|
971
|
-
|
972
|
-
|
973
|
-
|
974
|
-
|
975
|
-
movdqa 688(%rsp),%xmm2
|
976
|
-
pmuludq 624(%rsp),%xmm2
|
977
|
-
paddq %xmm2,%xmm11
|
972
|
+
vpaddq %xmm2,%xmm1,%xmm1
|
973
|
+
vmovdqa 608(%rsp),%xmm2
|
974
|
+
vpmuludq 592(%rsp),%xmm2,%xmm2
|
975
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
976
|
+
vmovdqa 656(%rsp),%xmm2
|
977
|
+
vpmuludq 576(%rsp),%xmm2,%xmm2
|
978
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
979
|
+
vmovdqa 688(%rsp),%xmm2
|
980
|
+
vpmuludq 624(%rsp),%xmm2,%xmm2
|
981
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
978
982
|
vpsrlq $26,%xmm1,%xmm2
|
979
|
-
|
980
|
-
|
983
|
+
vpaddq %xmm2,%xmm3,%xmm3
|
984
|
+
vpand m26(%rip),%xmm1,%xmm1
|
981
985
|
vpsrlq $25,%xmm10,%xmm2
|
982
|
-
|
983
|
-
|
986
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
987
|
+
vpand m25(%rip),%xmm10,%xmm10
|
984
988
|
vpsrlq $25,%xmm3,%xmm2
|
985
|
-
|
986
|
-
|
989
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
990
|
+
vpand m25(%rip),%xmm3,%xmm3
|
987
991
|
vpsrlq $26,%xmm11,%xmm2
|
988
|
-
|
989
|
-
|
992
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
993
|
+
vpand m26(%rip),%xmm11,%xmm11
|
990
994
|
vpsrlq $26,%xmm5,%xmm2
|
991
|
-
|
992
|
-
|
995
|
+
vpaddq %xmm2,%xmm7,%xmm7
|
996
|
+
vpand m26(%rip),%xmm5,%xmm5
|
993
997
|
vpsrlq $25,%xmm12,%xmm2
|
994
|
-
|
995
|
-
|
998
|
+
vpaddq %xmm2,%xmm13,%xmm13
|
999
|
+
vpand m25(%rip),%xmm12,%xmm12
|
996
1000
|
vpsrlq $25,%xmm7,%xmm2
|
997
|
-
|
998
|
-
|
1001
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
1002
|
+
vpand m25(%rip),%xmm7,%xmm7
|
999
1003
|
vpsrlq $26,%xmm13,%xmm2
|
1000
|
-
|
1001
|
-
|
1004
|
+
vpaddq %xmm2,%xmm0,%xmm0
|
1005
|
+
vpand m26(%rip),%xmm13,%xmm13
|
1002
1006
|
vpsrlq $26,%xmm9,%xmm2
|
1003
|
-
|
1004
|
-
|
1007
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
1008
|
+
vpand m26(%rip),%xmm9,%xmm9
|
1005
1009
|
vpsrlq $25,%xmm0,%xmm2
|
1006
1010
|
vpsllq $4,%xmm2,%xmm4
|
1007
|
-
|
1008
|
-
|
1009
|
-
|
1010
|
-
|
1011
|
-
|
1011
|
+
vpaddq %xmm2,%xmm1,%xmm1
|
1012
|
+
vpsllq $1,%xmm2,%xmm2
|
1013
|
+
vpaddq %xmm2,%xmm4,%xmm4
|
1014
|
+
vpaddq %xmm4,%xmm1,%xmm1
|
1015
|
+
vpand m25(%rip),%xmm0,%xmm0
|
1012
1016
|
vpsrlq $25,%xmm10,%xmm2
|
1013
|
-
|
1014
|
-
|
1017
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
1018
|
+
vpand m25(%rip),%xmm10,%xmm10
|
1015
1019
|
vpsrlq $26,%xmm1,%xmm2
|
1016
|
-
|
1017
|
-
|
1020
|
+
vpaddq %xmm2,%xmm3,%xmm3
|
1021
|
+
vpand m26(%rip),%xmm1,%xmm1
|
1018
1022
|
vpunpckhqdq %xmm3,%xmm1,%xmm2
|
1019
1023
|
vpunpcklqdq %xmm3,%xmm1,%xmm1
|
1020
|
-
|
1024
|
+
vmovdqa %xmm1,464(%rsp)
|
1021
1025
|
vpaddq subc0(%rip),%xmm2,%xmm3
|
1022
|
-
|
1026
|
+
vpsubq %xmm1,%xmm3,%xmm3
|
1023
1027
|
vpunpckhqdq %xmm3,%xmm2,%xmm1
|
1024
1028
|
vpunpcklqdq %xmm3,%xmm2,%xmm2
|
1025
|
-
|
1026
|
-
|
1027
|
-
|
1028
|
-
|
1029
|
-
|
1030
|
-
|
1029
|
+
vmovdqa %xmm2,480(%rsp)
|
1030
|
+
vmovdqa %xmm1,512(%rsp)
|
1031
|
+
vpsllq $1,%xmm1,%xmm1
|
1032
|
+
vmovdqa %xmm1,528(%rsp)
|
1033
|
+
vpmuludq v121666_121666(%rip),%xmm3,%xmm3
|
1034
|
+
vmovdqa 80(%rsp),%xmm1
|
1031
1035
|
vpunpcklqdq %xmm1,%xmm3,%xmm2
|
1032
1036
|
vpunpckhqdq %xmm1,%xmm3,%xmm1
|
1033
1037
|
vpunpckhqdq %xmm7,%xmm5,%xmm3
|
1034
1038
|
vpunpcklqdq %xmm7,%xmm5,%xmm4
|
1035
|
-
|
1039
|
+
vmovdqa %xmm4,544(%rsp)
|
1036
1040
|
vpaddq subc2(%rip),%xmm3,%xmm5
|
1037
|
-
|
1041
|
+
vpsubq %xmm4,%xmm5,%xmm5
|
1038
1042
|
vpunpckhqdq %xmm5,%xmm3,%xmm4
|
1039
1043
|
vpunpcklqdq %xmm5,%xmm3,%xmm3
|
1040
|
-
|
1041
|
-
|
1042
|
-
|
1043
|
-
|
1044
|
-
|
1045
|
-
|
1044
|
+
vmovdqa %xmm3,560(%rsp)
|
1045
|
+
vmovdqa %xmm4,576(%rsp)
|
1046
|
+
vpsllq $1,%xmm4,%xmm4
|
1047
|
+
vmovdqa %xmm4,592(%rsp)
|
1048
|
+
vpmuludq v121666_121666(%rip),%xmm5,%xmm5
|
1049
|
+
vmovdqa 96(%rsp),%xmm3
|
1046
1050
|
vpunpcklqdq %xmm3,%xmm5,%xmm4
|
1047
1051
|
vpunpckhqdq %xmm3,%xmm5,%xmm3
|
1048
1052
|
vpunpckhqdq %xmm10,%xmm9,%xmm5
|
1049
1053
|
vpunpcklqdq %xmm10,%xmm9,%xmm6
|
1050
|
-
|
1054
|
+
vmovdqa %xmm6,608(%rsp)
|
1051
1055
|
vpaddq subc2(%rip),%xmm5,%xmm7
|
1052
|
-
|
1056
|
+
vpsubq %xmm6,%xmm7,%xmm7
|
1053
1057
|
vpunpckhqdq %xmm7,%xmm5,%xmm6
|
1054
1058
|
vpunpcklqdq %xmm7,%xmm5,%xmm5
|
1055
|
-
|
1056
|
-
|
1057
|
-
|
1058
|
-
|
1059
|
-
|
1060
|
-
|
1059
|
+
vmovdqa %xmm5,624(%rsp)
|
1060
|
+
vmovdqa %xmm6,640(%rsp)
|
1061
|
+
vpsllq $1,%xmm6,%xmm6
|
1062
|
+
vmovdqa %xmm6,656(%rsp)
|
1063
|
+
vpmuludq v121666_121666(%rip),%xmm7,%xmm7
|
1064
|
+
vmovdqa 112(%rsp),%xmm5
|
1061
1065
|
vpunpcklqdq %xmm5,%xmm7,%xmm6
|
1062
1066
|
vpunpckhqdq %xmm5,%xmm7,%xmm5
|
1063
1067
|
vpunpckhqdq %xmm12,%xmm11,%xmm7
|
1064
1068
|
vpunpcklqdq %xmm12,%xmm11,%xmm8
|
1065
|
-
|
1069
|
+
vmovdqa %xmm8,672(%rsp)
|
1066
1070
|
vpaddq subc2(%rip),%xmm7,%xmm9
|
1067
|
-
|
1071
|
+
vpsubq %xmm8,%xmm9,%xmm9
|
1068
1072
|
vpunpckhqdq %xmm9,%xmm7,%xmm8
|
1069
1073
|
vpunpcklqdq %xmm9,%xmm7,%xmm7
|
1070
|
-
|
1071
|
-
|
1072
|
-
|
1073
|
-
|
1074
|
-
|
1075
|
-
|
1074
|
+
vmovdqa %xmm7,688(%rsp)
|
1075
|
+
vmovdqa %xmm8,704(%rsp)
|
1076
|
+
vpsllq $1,%xmm8,%xmm8
|
1077
|
+
vmovdqa %xmm8,720(%rsp)
|
1078
|
+
vpmuludq v121666_121666(%rip),%xmm9,%xmm9
|
1079
|
+
vmovdqa 448(%rsp),%xmm7
|
1076
1080
|
vpunpcklqdq %xmm7,%xmm9,%xmm8
|
1077
1081
|
vpunpckhqdq %xmm7,%xmm9,%xmm7
|
1078
1082
|
vpunpckhqdq %xmm0,%xmm13,%xmm9
|
1079
1083
|
vpunpcklqdq %xmm0,%xmm13,%xmm0
|
1080
|
-
|
1084
|
+
vmovdqa %xmm0,448(%rsp)
|
1081
1085
|
vpaddq subc2(%rip),%xmm9,%xmm10
|
1082
|
-
|
1086
|
+
vpsubq %xmm0,%xmm10,%xmm10
|
1083
1087
|
vpunpckhqdq %xmm10,%xmm9,%xmm0
|
1084
1088
|
vpunpcklqdq %xmm10,%xmm9,%xmm9
|
1085
|
-
|
1086
|
-
|
1087
|
-
|
1088
|
-
|
1089
|
-
|
1090
|
-
|
1089
|
+
vmovdqa %xmm9,736(%rsp)
|
1090
|
+
vmovdqa %xmm0,752(%rsp)
|
1091
|
+
vpsllq $1,%xmm0,%xmm0
|
1092
|
+
vmovdqa %xmm0,768(%rsp)
|
1093
|
+
vpmuludq v121666_121666(%rip),%xmm10,%xmm10
|
1094
|
+
vmovdqa 496(%rsp),%xmm0
|
1091
1095
|
vpunpcklqdq %xmm0,%xmm10,%xmm9
|
1092
1096
|
vpunpckhqdq %xmm0,%xmm10,%xmm0
|
1093
1097
|
vpsrlq $26,%xmm2,%xmm10
|
1094
|
-
|
1095
|
-
|
1098
|
+
vpaddq %xmm10,%xmm1,%xmm1
|
1099
|
+
vpand m26(%rip),%xmm2,%xmm2
|
1096
1100
|
vpsrlq $25,%xmm5,%xmm10
|
1097
|
-
|
1098
|
-
|
1101
|
+
vpaddq %xmm10,%xmm8,%xmm8
|
1102
|
+
vpand m25(%rip),%xmm5,%xmm5
|
1099
1103
|
vpsrlq $25,%xmm1,%xmm10
|
1100
|
-
|
1101
|
-
|
1104
|
+
vpaddq %xmm10,%xmm4,%xmm4
|
1105
|
+
vpand m25(%rip),%xmm1,%xmm1
|
1102
1106
|
vpsrlq $26,%xmm8,%xmm10
|
1103
|
-
|
1104
|
-
|
1107
|
+
vpaddq %xmm10,%xmm7,%xmm7
|
1108
|
+
vpand m26(%rip),%xmm8,%xmm8
|
1105
1109
|
vpsrlq $26,%xmm4,%xmm10
|
1106
|
-
|
1107
|
-
|
1110
|
+
vpaddq %xmm10,%xmm3,%xmm3
|
1111
|
+
vpand m26(%rip),%xmm4,%xmm4
|
1108
1112
|
vpsrlq $25,%xmm7,%xmm10
|
1109
|
-
|
1110
|
-
|
1113
|
+
vpaddq %xmm10,%xmm9,%xmm9
|
1114
|
+
vpand m25(%rip),%xmm7,%xmm7
|
1111
1115
|
vpsrlq $25,%xmm3,%xmm10
|
1112
|
-
|
1113
|
-
|
1116
|
+
vpaddq %xmm10,%xmm6,%xmm6
|
1117
|
+
vpand m25(%rip),%xmm3,%xmm3
|
1114
1118
|
vpsrlq $26,%xmm9,%xmm10
|
1115
|
-
|
1116
|
-
|
1119
|
+
vpaddq %xmm10,%xmm0,%xmm0
|
1120
|
+
vpand m26(%rip),%xmm9,%xmm9
|
1117
1121
|
vpsrlq $26,%xmm6,%xmm10
|
1118
|
-
|
1119
|
-
|
1122
|
+
vpaddq %xmm10,%xmm5,%xmm5
|
1123
|
+
vpand m26(%rip),%xmm6,%xmm6
|
1120
1124
|
vpsrlq $25,%xmm0,%xmm10
|
1121
1125
|
vpsllq $4,%xmm10,%xmm11
|
1122
|
-
|
1123
|
-
|
1124
|
-
|
1125
|
-
|
1126
|
-
|
1126
|
+
vpaddq %xmm10,%xmm2,%xmm2
|
1127
|
+
vpsllq $1,%xmm10,%xmm10
|
1128
|
+
vpaddq %xmm10,%xmm11,%xmm11
|
1129
|
+
vpaddq %xmm11,%xmm2,%xmm2
|
1130
|
+
vpand m25(%rip),%xmm0,%xmm0
|
1127
1131
|
vpsrlq $25,%xmm5,%xmm10
|
1128
|
-
|
1129
|
-
|
1132
|
+
vpaddq %xmm10,%xmm8,%xmm8
|
1133
|
+
vpand m25(%rip),%xmm5,%xmm5
|
1130
1134
|
vpsrlq $26,%xmm2,%xmm10
|
1131
|
-
|
1132
|
-
|
1135
|
+
vpaddq %xmm10,%xmm1,%xmm1
|
1136
|
+
vpand m26(%rip),%xmm2,%xmm2
|
1133
1137
|
vpunpckhqdq %xmm1,%xmm2,%xmm10
|
1134
|
-
|
1138
|
+
vmovdqa %xmm10,80(%rsp)
|
1135
1139
|
vpunpcklqdq %xmm1,%xmm2,%xmm1
|
1136
1140
|
vpunpckhqdq %xmm3,%xmm4,%xmm2
|
1137
|
-
|
1141
|
+
vmovdqa %xmm2,96(%rsp)
|
1138
1142
|
vpunpcklqdq %xmm3,%xmm4,%xmm2
|
1139
1143
|
vpunpckhqdq %xmm5,%xmm6,%xmm3
|
1140
|
-
|
1144
|
+
vmovdqa %xmm3,112(%rsp)
|
1141
1145
|
vpunpcklqdq %xmm5,%xmm6,%xmm3
|
1142
1146
|
vpunpckhqdq %xmm7,%xmm8,%xmm4
|
1143
|
-
|
1147
|
+
vmovdqa %xmm4,128(%rsp)
|
1144
1148
|
vpunpcklqdq %xmm7,%xmm8,%xmm4
|
1145
1149
|
vpunpckhqdq %xmm0,%xmm9,%xmm5
|
1146
|
-
|
1150
|
+
vmovdqa %xmm5,144(%rsp)
|
1147
1151
|
vpunpcklqdq %xmm0,%xmm9,%xmm0
|
1148
|
-
|
1149
|
-
|
1152
|
+
vmovdqa 464(%rsp),%xmm5
|
1153
|
+
vpaddq %xmm5,%xmm1,%xmm1
|
1150
1154
|
vpunpcklqdq %xmm1,%xmm5,%xmm6
|
1151
1155
|
vpunpckhqdq %xmm1,%xmm5,%xmm1
|
1152
1156
|
vpmuludq 512(%rsp),%xmm6,%xmm5
|
1153
1157
|
vpmuludq 480(%rsp),%xmm1,%xmm7
|
1154
|
-
|
1158
|
+
vpaddq %xmm7,%xmm5,%xmm5
|
1155
1159
|
vpmuludq 560(%rsp),%xmm6,%xmm7
|
1156
1160
|
vpmuludq 528(%rsp),%xmm1,%xmm8
|
1157
|
-
|
1161
|
+
vpaddq %xmm8,%xmm7,%xmm7
|
1158
1162
|
vpmuludq 576(%rsp),%xmm6,%xmm8
|
1159
1163
|
vpmuludq 560(%rsp),%xmm1,%xmm9
|
1160
|
-
|
1164
|
+
vpaddq %xmm9,%xmm8,%xmm8
|
1161
1165
|
vpmuludq 624(%rsp),%xmm6,%xmm9
|
1162
1166
|
vpmuludq 592(%rsp),%xmm1,%xmm10
|
1163
|
-
|
1167
|
+
vpaddq %xmm10,%xmm9,%xmm9
|
1164
1168
|
vpmuludq 640(%rsp),%xmm6,%xmm10
|
1165
1169
|
vpmuludq 624(%rsp),%xmm1,%xmm11
|
1166
|
-
|
1170
|
+
vpaddq %xmm11,%xmm10,%xmm10
|
1167
1171
|
vpmuludq 688(%rsp),%xmm6,%xmm11
|
1168
1172
|
vpmuludq 656(%rsp),%xmm1,%xmm12
|
1169
|
-
|
1173
|
+
vpaddq %xmm12,%xmm11,%xmm11
|
1170
1174
|
vpmuludq 704(%rsp),%xmm6,%xmm12
|
1171
1175
|
vpmuludq 688(%rsp),%xmm1,%xmm13
|
1172
|
-
|
1176
|
+
vpaddq %xmm13,%xmm12,%xmm12
|
1173
1177
|
vpmuludq 736(%rsp),%xmm6,%xmm13
|
1174
1178
|
vpmuludq 720(%rsp),%xmm1,%xmm14
|
1175
|
-
|
1179
|
+
vpaddq %xmm14,%xmm13,%xmm13
|
1176
1180
|
vpmuludq 752(%rsp),%xmm6,%xmm14
|
1177
1181
|
vpmuludq 736(%rsp),%xmm1,%xmm15
|
1178
|
-
|
1182
|
+
vpaddq %xmm15,%xmm14,%xmm14
|
1179
1183
|
vpmuludq 480(%rsp),%xmm6,%xmm6
|
1180
|
-
|
1184
|
+
vpmuludq v19_19(%rip),%xmm1,%xmm1
|
1181
1185
|
vpmuludq 768(%rsp),%xmm1,%xmm1
|
1182
|
-
|
1183
|
-
|
1184
|
-
|
1186
|
+
vpaddq %xmm1,%xmm6,%xmm6
|
1187
|
+
vmovdqa 544(%rsp),%xmm1
|
1188
|
+
vpaddq %xmm1,%xmm2,%xmm2
|
1185
1189
|
vpunpcklqdq %xmm2,%xmm1,%xmm15
|
1186
1190
|
vpunpckhqdq %xmm2,%xmm1,%xmm1
|
1187
1191
|
vpmuludq 480(%rsp),%xmm15,%xmm2
|
1188
|
-
|
1192
|
+
vpaddq %xmm2,%xmm7,%xmm7
|
1189
1193
|
vpmuludq 512(%rsp),%xmm15,%xmm2
|
1190
|
-
|
1194
|
+
vpaddq %xmm2,%xmm8,%xmm8
|
1191
1195
|
vpmuludq 560(%rsp),%xmm15,%xmm2
|
1192
|
-
|
1196
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
1193
1197
|
vpmuludq 576(%rsp),%xmm15,%xmm2
|
1194
|
-
|
1198
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
1195
1199
|
vpmuludq 624(%rsp),%xmm15,%xmm2
|
1196
|
-
|
1200
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
1197
1201
|
vpmuludq 640(%rsp),%xmm15,%xmm2
|
1198
|
-
|
1202
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
1199
1203
|
vpmuludq 688(%rsp),%xmm15,%xmm2
|
1200
|
-
|
1204
|
+
vpaddq %xmm2,%xmm13,%xmm13
|
1201
1205
|
vpmuludq 704(%rsp),%xmm15,%xmm2
|
1202
|
-
|
1203
|
-
|
1206
|
+
vpaddq %xmm2,%xmm14,%xmm14
|
1207
|
+
vpmuludq v19_19(%rip),%xmm15,%xmm15
|
1204
1208
|
vpmuludq 736(%rsp),%xmm15,%xmm2
|
1205
|
-
|
1206
|
-
|
1207
|
-
|
1209
|
+
vpaddq %xmm2,%xmm6,%xmm6
|
1210
|
+
vpmuludq 752(%rsp),%xmm15,%xmm15
|
1211
|
+
vpaddq %xmm15,%xmm5,%xmm5
|
1208
1212
|
vpmuludq 480(%rsp),%xmm1,%xmm2
|
1209
|
-
|
1213
|
+
vpaddq %xmm2,%xmm8,%xmm8
|
1210
1214
|
vpmuludq 528(%rsp),%xmm1,%xmm2
|
1211
|
-
|
1215
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
1212
1216
|
vpmuludq 560(%rsp),%xmm1,%xmm2
|
1213
|
-
|
1217
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
1214
1218
|
vpmuludq 592(%rsp),%xmm1,%xmm2
|
1215
|
-
|
1219
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
1216
1220
|
vpmuludq 624(%rsp),%xmm1,%xmm2
|
1217
|
-
|
1221
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
1218
1222
|
vpmuludq 656(%rsp),%xmm1,%xmm2
|
1219
|
-
|
1223
|
+
vpaddq %xmm2,%xmm13,%xmm13
|
1220
1224
|
vpmuludq 688(%rsp),%xmm1,%xmm2
|
1221
|
-
|
1222
|
-
|
1225
|
+
vpaddq %xmm2,%xmm14,%xmm14
|
1226
|
+
vpmuludq v19_19(%rip),%xmm1,%xmm1
|
1223
1227
|
vpmuludq 720(%rsp),%xmm1,%xmm2
|
1224
|
-
|
1228
|
+
vpaddq %xmm2,%xmm6,%xmm6
|
1225
1229
|
vpmuludq 736(%rsp),%xmm1,%xmm2
|
1226
|
-
|
1227
|
-
|
1228
|
-
|
1229
|
-
|
1230
|
-
|
1230
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
1231
|
+
vpmuludq 768(%rsp),%xmm1,%xmm1
|
1232
|
+
vpaddq %xmm1,%xmm7,%xmm7
|
1233
|
+
vmovdqa 608(%rsp),%xmm1
|
1234
|
+
vpaddq %xmm1,%xmm3,%xmm3
|
1231
1235
|
vpunpcklqdq %xmm3,%xmm1,%xmm2
|
1232
1236
|
vpunpckhqdq %xmm3,%xmm1,%xmm1
|
1233
1237
|
vpmuludq 480(%rsp),%xmm2,%xmm3
|
1234
|
-
|
1238
|
+
vpaddq %xmm3,%xmm9,%xmm9
|
1235
1239
|
vpmuludq 512(%rsp),%xmm2,%xmm3
|
1236
|
-
|
1240
|
+
vpaddq %xmm3,%xmm10,%xmm10
|
1237
1241
|
vpmuludq 560(%rsp),%xmm2,%xmm3
|
1238
|
-
|
1242
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
1239
1243
|
vpmuludq 576(%rsp),%xmm2,%xmm3
|
1240
|
-
|
1244
|
+
vpaddq %xmm3,%xmm12,%xmm12
|
1241
1245
|
vpmuludq 624(%rsp),%xmm2,%xmm3
|
1242
|
-
|
1246
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
1243
1247
|
vpmuludq 640(%rsp),%xmm2,%xmm3
|
1244
|
-
|
1245
|
-
|
1248
|
+
vpaddq %xmm3,%xmm14,%xmm14
|
1249
|
+
vpmuludq v19_19(%rip),%xmm2,%xmm2
|
1246
1250
|
vpmuludq 688(%rsp),%xmm2,%xmm3
|
1247
|
-
|
1251
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
1248
1252
|
vpmuludq 704(%rsp),%xmm2,%xmm3
|
1249
|
-
|
1253
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
1250
1254
|
vpmuludq 736(%rsp),%xmm2,%xmm3
|
1251
|
-
|
1252
|
-
|
1253
|
-
|
1255
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
1256
|
+
vpmuludq 752(%rsp),%xmm2,%xmm2
|
1257
|
+
vpaddq %xmm2,%xmm8,%xmm8
|
1254
1258
|
vpmuludq 480(%rsp),%xmm1,%xmm2
|
1255
|
-
|
1259
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
1256
1260
|
vpmuludq 528(%rsp),%xmm1,%xmm2
|
1257
|
-
|
1261
|
+
vpaddq %xmm2,%xmm11,%xmm11
|
1258
1262
|
vpmuludq 560(%rsp),%xmm1,%xmm2
|
1259
|
-
|
1263
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
1260
1264
|
vpmuludq 592(%rsp),%xmm1,%xmm2
|
1261
|
-
|
1265
|
+
vpaddq %xmm2,%xmm13,%xmm13
|
1262
1266
|
vpmuludq 624(%rsp),%xmm1,%xmm2
|
1263
|
-
|
1264
|
-
|
1267
|
+
vpaddq %xmm2,%xmm14,%xmm14
|
1268
|
+
vpmuludq v19_19(%rip),%xmm1,%xmm1
|
1265
1269
|
vpmuludq 656(%rsp),%xmm1,%xmm2
|
1266
|
-
|
1270
|
+
vpaddq %xmm2,%xmm6,%xmm6
|
1267
1271
|
vpmuludq 688(%rsp),%xmm1,%xmm2
|
1268
|
-
|
1272
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
1269
1273
|
vpmuludq 720(%rsp),%xmm1,%xmm2
|
1270
|
-
|
1274
|
+
vpaddq %xmm2,%xmm7,%xmm7
|
1271
1275
|
vpmuludq 736(%rsp),%xmm1,%xmm2
|
1272
|
-
|
1273
|
-
|
1274
|
-
|
1275
|
-
|
1276
|
-
|
1276
|
+
vpaddq %xmm2,%xmm8,%xmm8
|
1277
|
+
vpmuludq 768(%rsp),%xmm1,%xmm1
|
1278
|
+
vpaddq %xmm1,%xmm9,%xmm9
|
1279
|
+
vmovdqa 672(%rsp),%xmm1
|
1280
|
+
vpaddq %xmm1,%xmm4,%xmm4
|
1277
1281
|
vpunpcklqdq %xmm4,%xmm1,%xmm2
|
1278
1282
|
vpunpckhqdq %xmm4,%xmm1,%xmm1
|
1279
1283
|
vpmuludq 480(%rsp),%xmm2,%xmm3
|
1280
|
-
|
1284
|
+
vpaddq %xmm3,%xmm11,%xmm11
|
1281
1285
|
vpmuludq 512(%rsp),%xmm2,%xmm3
|
1282
|
-
|
1286
|
+
vpaddq %xmm3,%xmm12,%xmm12
|
1283
1287
|
vpmuludq 560(%rsp),%xmm2,%xmm3
|
1284
|
-
|
1288
|
+
vpaddq %xmm3,%xmm13,%xmm13
|
1285
1289
|
vpmuludq 576(%rsp),%xmm2,%xmm3
|
1286
|
-
|
1287
|
-
|
1290
|
+
vpaddq %xmm3,%xmm14,%xmm14
|
1291
|
+
vpmuludq v19_19(%rip),%xmm2,%xmm2
|
1288
1292
|
vpmuludq 624(%rsp),%xmm2,%xmm3
|
1289
|
-
|
1293
|
+
vpaddq %xmm3,%xmm6,%xmm6
|
1290
1294
|
vpmuludq 640(%rsp),%xmm2,%xmm3
|
1291
|
-
|
1295
|
+
vpaddq %xmm3,%xmm5,%xmm5
|
1292
1296
|
vpmuludq 688(%rsp),%xmm2,%xmm3
|
1293
|
-
|
1297
|
+
vpaddq %xmm3,%xmm7,%xmm7
|
1294
1298
|
vpmuludq 704(%rsp),%xmm2,%xmm3
|
1295
|
-
|
1299
|
+
vpaddq %xmm3,%xmm8,%xmm8
|
1296
1300
|
vpmuludq 736(%rsp),%xmm2,%xmm3
|
1297
|
-
|
1298
|
-
|
1299
|
-
|
1301
|
+
vpaddq %xmm3,%xmm9,%xmm9
|
1302
|
+
vpmuludq 752(%rsp),%xmm2,%xmm2
|
1303
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
1300
1304
|
vpmuludq 480(%rsp),%xmm1,%xmm2
|
1301
|
-
|
1305
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
1302
1306
|
vpmuludq 528(%rsp),%xmm1,%xmm2
|
1303
|
-
|
1307
|
+
vpaddq %xmm2,%xmm13,%xmm13
|
1304
1308
|
vpmuludq 560(%rsp),%xmm1,%xmm2
|
1305
|
-
|
1306
|
-
|
1309
|
+
vpaddq %xmm2,%xmm14,%xmm14
|
1310
|
+
vpmuludq v19_19(%rip),%xmm1,%xmm1
|
1307
1311
|
vpmuludq 592(%rsp),%xmm1,%xmm2
|
1308
|
-
|
1312
|
+
vpaddq %xmm2,%xmm6,%xmm6
|
1309
1313
|
vpmuludq 624(%rsp),%xmm1,%xmm2
|
1310
|
-
|
1314
|
+
vpaddq %xmm2,%xmm5,%xmm5
|
1311
1315
|
vpmuludq 656(%rsp),%xmm1,%xmm2
|
1312
|
-
|
1316
|
+
vpaddq %xmm2,%xmm7,%xmm7
|
1313
1317
|
vpmuludq 688(%rsp),%xmm1,%xmm2
|
1314
|
-
|
1318
|
+
vpaddq %xmm2,%xmm8,%xmm8
|
1315
1319
|
vpmuludq 720(%rsp),%xmm1,%xmm2
|
1316
|
-
|
1320
|
+
vpaddq %xmm2,%xmm9,%xmm9
|
1317
1321
|
vpmuludq 736(%rsp),%xmm1,%xmm2
|
1318
|
-
|
1319
|
-
|
1320
|
-
|
1321
|
-
|
1322
|
-
|
1322
|
+
vpaddq %xmm2,%xmm10,%xmm10
|
1323
|
+
vpmuludq 768(%rsp),%xmm1,%xmm1
|
1324
|
+
vpaddq %xmm1,%xmm11,%xmm11
|
1325
|
+
vmovdqa 448(%rsp),%xmm1
|
1326
|
+
vpaddq %xmm1,%xmm0,%xmm0
|
1323
1327
|
vpunpcklqdq %xmm0,%xmm1,%xmm2
|
1324
1328
|
vpunpckhqdq %xmm0,%xmm1,%xmm0
|
1325
1329
|
vpmuludq 480(%rsp),%xmm2,%xmm1
|
1326
|
-
|
1330
|
+
vpaddq %xmm1,%xmm13,%xmm13
|
1327
1331
|
vpmuludq 512(%rsp),%xmm2,%xmm1
|
1328
|
-
|
1329
|
-
|
1332
|
+
vpaddq %xmm1,%xmm14,%xmm14
|
1333
|
+
vpmuludq v19_19(%rip),%xmm2,%xmm2
|
1330
1334
|
vpmuludq 560(%rsp),%xmm2,%xmm1
|
1331
|
-
|
1335
|
+
vpaddq %xmm1,%xmm6,%xmm6
|
1332
1336
|
vpmuludq 576(%rsp),%xmm2,%xmm1
|
1333
|
-
|
1337
|
+
vpaddq %xmm1,%xmm5,%xmm5
|
1334
1338
|
vpmuludq 624(%rsp),%xmm2,%xmm1
|
1335
|
-
|
1339
|
+
vpaddq %xmm1,%xmm7,%xmm7
|
1336
1340
|
vpmuludq 640(%rsp),%xmm2,%xmm1
|
1337
|
-
|
1341
|
+
vpaddq %xmm1,%xmm8,%xmm8
|
1338
1342
|
vpmuludq 688(%rsp),%xmm2,%xmm1
|
1339
|
-
|
1343
|
+
vpaddq %xmm1,%xmm9,%xmm9
|
1340
1344
|
vpmuludq 704(%rsp),%xmm2,%xmm1
|
1341
|
-
|
1345
|
+
vpaddq %xmm1,%xmm10,%xmm10
|
1342
1346
|
vpmuludq 736(%rsp),%xmm2,%xmm1
|
1343
|
-
|
1344
|
-
|
1345
|
-
|
1347
|
+
vpaddq %xmm1,%xmm11,%xmm11
|
1348
|
+
vpmuludq 752(%rsp),%xmm2,%xmm2
|
1349
|
+
vpaddq %xmm2,%xmm12,%xmm12
|
1346
1350
|
vpmuludq 480(%rsp),%xmm0,%xmm1
|
1347
|
-
|
1348
|
-
|
1351
|
+
vpaddq %xmm1,%xmm14,%xmm14
|
1352
|
+
vpmuludq v19_19(%rip),%xmm0,%xmm0
|
1349
1353
|
vpmuludq 528(%rsp),%xmm0,%xmm1
|
1350
|
-
|
1354
|
+
vpaddq %xmm1,%xmm6,%xmm6
|
1351
1355
|
vpmuludq 560(%rsp),%xmm0,%xmm1
|
1352
|
-
|
1356
|
+
vpaddq %xmm1,%xmm5,%xmm5
|
1353
1357
|
vpmuludq 592(%rsp),%xmm0,%xmm1
|
1354
|
-
|
1358
|
+
vpaddq %xmm1,%xmm7,%xmm7
|
1355
1359
|
vpmuludq 624(%rsp),%xmm0,%xmm1
|
1356
|
-
|
1360
|
+
vpaddq %xmm1,%xmm8,%xmm8
|
1357
1361
|
vpmuludq 656(%rsp),%xmm0,%xmm1
|
1358
|
-
|
1362
|
+
vpaddq %xmm1,%xmm9,%xmm9
|
1359
1363
|
vpmuludq 688(%rsp),%xmm0,%xmm1
|
1360
|
-
|
1364
|
+
vpaddq %xmm1,%xmm10,%xmm10
|
1361
1365
|
vpmuludq 720(%rsp),%xmm0,%xmm1
|
1362
|
-
|
1366
|
+
vpaddq %xmm1,%xmm11,%xmm11
|
1363
1367
|
vpmuludq 736(%rsp),%xmm0,%xmm1
|
1364
|
-
|
1365
|
-
|
1366
|
-
|
1368
|
+
vpaddq %xmm1,%xmm12,%xmm12
|
1369
|
+
vpmuludq 768(%rsp),%xmm0,%xmm0
|
1370
|
+
vpaddq %xmm0,%xmm13,%xmm13
|
1367
1371
|
vpsrlq $26,%xmm6,%xmm0
|
1368
|
-
|
1369
|
-
|
1372
|
+
vpaddq %xmm0,%xmm5,%xmm5
|
1373
|
+
vpand m26(%rip),%xmm6,%xmm6
|
1370
1374
|
vpsrlq $25,%xmm10,%xmm0
|
1371
|
-
|
1372
|
-
|
1375
|
+
vpaddq %xmm0,%xmm11,%xmm11
|
1376
|
+
vpand m25(%rip),%xmm10,%xmm10
|
1373
1377
|
vpsrlq $25,%xmm5,%xmm0
|
1374
|
-
|
1375
|
-
|
1378
|
+
vpaddq %xmm0,%xmm7,%xmm7
|
1379
|
+
vpand m25(%rip),%xmm5,%xmm5
|
1376
1380
|
vpsrlq $26,%xmm11,%xmm0
|
1377
|
-
|
1378
|
-
|
1381
|
+
vpaddq %xmm0,%xmm12,%xmm12
|
1382
|
+
vpand m26(%rip),%xmm11,%xmm11
|
1379
1383
|
vpsrlq $26,%xmm7,%xmm0
|
1380
|
-
|
1381
|
-
|
1384
|
+
vpaddq %xmm0,%xmm8,%xmm8
|
1385
|
+
vpand m26(%rip),%xmm7,%xmm7
|
1382
1386
|
vpsrlq $25,%xmm12,%xmm0
|
1383
|
-
|
1384
|
-
|
1387
|
+
vpaddq %xmm0,%xmm13,%xmm13
|
1388
|
+
vpand m25(%rip),%xmm12,%xmm12
|
1385
1389
|
vpsrlq $25,%xmm8,%xmm0
|
1386
|
-
|
1387
|
-
|
1390
|
+
vpaddq %xmm0,%xmm9,%xmm9
|
1391
|
+
vpand m25(%rip),%xmm8,%xmm8
|
1388
1392
|
vpsrlq $26,%xmm13,%xmm0
|
1389
|
-
|
1390
|
-
|
1393
|
+
vpaddq %xmm0,%xmm14,%xmm14
|
1394
|
+
vpand m26(%rip),%xmm13,%xmm13
|
1391
1395
|
vpsrlq $26,%xmm9,%xmm0
|
1392
|
-
|
1393
|
-
|
1396
|
+
vpaddq %xmm0,%xmm10,%xmm10
|
1397
|
+
vpand m26(%rip),%xmm9,%xmm9
|
1394
1398
|
vpsrlq $25,%xmm14,%xmm0
|
1395
1399
|
vpsllq $4,%xmm0,%xmm1
|
1396
|
-
|
1397
|
-
|
1398
|
-
|
1399
|
-
|
1400
|
-
|
1400
|
+
vpaddq %xmm0,%xmm6,%xmm6
|
1401
|
+
vpsllq $1,%xmm0,%xmm0
|
1402
|
+
vpaddq %xmm0,%xmm1,%xmm1
|
1403
|
+
vpaddq %xmm1,%xmm6,%xmm6
|
1404
|
+
vpand m25(%rip),%xmm14,%xmm14
|
1401
1405
|
vpsrlq $25,%xmm10,%xmm0
|
1402
|
-
|
1403
|
-
|
1406
|
+
vpaddq %xmm0,%xmm11,%xmm11
|
1407
|
+
vpand m25(%rip),%xmm10,%xmm10
|
1404
1408
|
vpsrlq $26,%xmm6,%xmm0
|
1405
|
-
|
1406
|
-
|
1409
|
+
vpaddq %xmm0,%xmm5,%xmm5
|
1410
|
+
vpand m26(%rip),%xmm6,%xmm6
|
1407
1411
|
vpunpckhqdq %xmm5,%xmm6,%xmm1
|
1408
1412
|
vpunpcklqdq %xmm5,%xmm6,%xmm0
|
1409
1413
|
vpunpckhqdq %xmm8,%xmm7,%xmm3
|
@@ -1416,16 +1420,16 @@ vpunpckhqdq %xmm14,%xmm13,%xmm9
|
|
1416
1420
|
vpunpcklqdq %xmm14,%xmm13,%xmm8
|
1417
1421
|
cmp $0,%rdx
|
1418
1422
|
jne ._ladder_loop
|
1419
|
-
|
1420
|
-
|
1421
|
-
|
1422
|
-
|
1423
|
-
|
1424
|
-
|
1425
|
-
|
1426
|
-
|
1427
|
-
|
1428
|
-
|
1423
|
+
vmovdqu %xmm1,160(%rdi)
|
1424
|
+
vmovdqu %xmm0,80(%rdi)
|
1425
|
+
vmovdqu %xmm3,176(%rdi)
|
1426
|
+
vmovdqu %xmm2,96(%rdi)
|
1427
|
+
vmovdqu %xmm5,192(%rdi)
|
1428
|
+
vmovdqu %xmm4,112(%rdi)
|
1429
|
+
vmovdqu %xmm7,208(%rdi)
|
1430
|
+
vmovdqu %xmm6,128(%rdi)
|
1431
|
+
vmovdqu %xmm9,224(%rdi)
|
1432
|
+
vmovdqu %xmm8,144(%rdi)
|
1429
1433
|
movq 1824(%rsp),%r11
|
1430
1434
|
movq 1832(%rsp),%r12
|
1431
1435
|
movq 1840(%rsp),%r13
|