/dports/biology/fasta3/fasta-36.3.8/src/ |
H A D | global_sse2.c | 173 vTemp = _mm_slli_si128 (vF, 2); in global_sse2_word() 177 vTemp = _mm_slli_si128 (vF, 4); in global_sse2_word() 185 vTemp = _mm_slli_si128 (vF, 8); in global_sse2_word() 257 vFPrev = _mm_slli_si128 (vF, 2); in global_sse2_word() 264 vH = _mm_slli_si128 (vH, 2); in global_sse2_word() 273 vH = _mm_slli_si128 (vH, 2); in global_sse2_word() 391 vH = _mm_slli_si128 (vH, 1); in global_sse2_byte() 436 vTemp = _mm_slli_si128 (vF, 1); in global_sse2_byte() 442 vTemp = _mm_slli_si128 (vF, 2); in global_sse2_byte() 520 vH = _mm_slli_si128 (vH, 1); in global_sse2_byte() [all …]
|
H A D | glocal_sse2.c | 139 vH = _mm_slli_si128 (vH, 2); in glocal_sse2_word() 185 vTemp = _mm_slli_si128 (vF, 2); in glocal_sse2_word() 189 vTemp = _mm_slli_si128 (vF, 4); in glocal_sse2_word() 197 vTemp = _mm_slli_si128 (vF, 8); in glocal_sse2_word() 280 vFPrev = _mm_slli_si128 (vF, 2); in glocal_sse2_word() 289 vH = _mm_slli_si128 (vH, 2); in glocal_sse2_word() 302 vH = _mm_slli_si128 (vH, 2); in glocal_sse2_word() 427 vH = _mm_slli_si128 (vH, 1); in glocal_sse2_byte() 472 vTemp = _mm_slli_si128 (vF, 1); in glocal_sse2_byte() 569 vH = _mm_slli_si128 (vH, 1); in glocal_sse2_byte() [all …]
|
/dports/security/bearssl/bearssl-0.6/src/symcipher/ |
H A D | aes_x86ni.c | 53 k = _mm_xor_si128(k, _mm_slli_si128(k, 4)); in expand_step128() 67 t4 = _mm_slli_si128(*t1, 0x4); in expand_step192() 69 t4 = _mm_slli_si128(t4, 0x4); in expand_step192() 71 t4 = _mm_slli_si128(t4, 0x4); in expand_step192() 75 t4 = _mm_slli_si128(*t3, 0x4); in expand_step192() 87 t4 = _mm_slli_si128(*t1, 0x4); in expand_step256_1() 89 t4 = _mm_slli_si128(t4, 0x4); in expand_step256_1() 91 t4 = _mm_slli_si128(t4, 0x4); in expand_step256_1() 104 t4 = _mm_slli_si128(*t3, 0x4); in expand_step256_2() 106 t4 = _mm_slli_si128(t4, 0x4); in expand_step256_2() [all …]
|
/dports/games/retroarch/RetroArch-1.9.7/deps/bearssl-0.6/src/symcipher/ |
H A D | aes_x86ni.c | 53 k = _mm_xor_si128(k, _mm_slli_si128(k, 4)); in expand_step128() 67 t4 = _mm_slli_si128(*t1, 0x4); in expand_step192() 69 t4 = _mm_slli_si128(t4, 0x4); in expand_step192() 71 t4 = _mm_slli_si128(t4, 0x4); in expand_step192() 75 t4 = _mm_slli_si128(*t3, 0x4); in expand_step192() 87 t4 = _mm_slli_si128(*t1, 0x4); in expand_step256_1() 89 t4 = _mm_slli_si128(t4, 0x4); in expand_step256_1() 91 t4 = _mm_slli_si128(t4, 0x4); in expand_step256_1() 104 t4 = _mm_slli_si128(*t3, 0x4); in expand_step256_2() 106 t4 = _mm_slli_si128(t4, 0x4); in expand_step256_2() [all …]
|
/dports/net-im/libsignal-client/libsignal-client-0.9.6/cargo-crates/aes-0.7.4/src/ni/aes256/ |
H A D | expand.rs | 15 t4 = _mm_slli_si128(t1, 0x4); 17 t4 = _mm_slli_si128(t4, 0x4); 19 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 50 t4 = _mm_slli_si128(t1, 0x4); 52 t4 = _mm_slli_si128(t4, 0x4); 54 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/net-im/zkgroup/zkgroup-0.7.4/cargo-crates/aes-0.7.4/src/ni/aes256/ |
H A D | expand.rs | 15 t4 = _mm_slli_si128(t1, 0x4); 17 t4 = _mm_slli_si128(t4, 0x4); 19 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 50 t4 = _mm_slli_si128(t1, 0x4); 52 t4 = _mm_slli_si128(t4, 0x4); 54 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/security/arti/arti-9d0ede26801cdb182daa85c3eb5f0058dc178eb6/cargo-crates/aes-0.7.5/src/ni/aes256/ |
H A D | expand.rs | 15 t4 = _mm_slli_si128(t1, 0x4); 17 t4 = _mm_slli_si128(t4, 0x4); 19 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 50 t4 = _mm_slli_si128(t1, 0x4); 52 t4 = _mm_slli_si128(t4, 0x4); 54 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/devel/pijul/pijul-1.0.0.a55/cargo-crates/aes-0.7.5/src/ni/aes256/ |
H A D | expand.rs | 15 t4 = _mm_slli_si128(t1, 0x4); 17 t4 = _mm_slli_si128(t4, 0x4); 19 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 50 t4 = _mm_slli_si128(t1, 0x4); 52 t4 = _mm_slli_si128(t4, 0x4); 54 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/multimedia/librespot/librespot-0.1.3/cargo-crates/aesni-0.6.0/src/aes256/ |
H A D | expand.rs | 14 t4 = _mm_slli_si128(t1, 0x4); 16 t4 = _mm_slli_si128(t4, 0x4); 18 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 51 t4 = _mm_slli_si128(t1, 0x4); 53 t4 = _mm_slli_si128(t4, 0x4); 55 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/dns/doh-proxy/doh-proxy-0.9.2/cargo-crates/aesni-0.10.0/src/aes256/ |
H A D | expand.rs | 14 t4 = _mm_slli_si128(t1, 0x4); 16 t4 = _mm_slli_si128(t4, 0x4); 18 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 51 t4 = _mm_slli_si128(t1, 0x4); 53 t4 = _mm_slli_si128(t4, 0x4); 55 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/net-im/fractal/fractal-4.4.0/vendor/aesni/src/aes256/ |
H A D | expand.rs | 14 t4 = _mm_slli_si128(t1, 0x4); 16 t4 = _mm_slli_si128(t4, 0x4); 18 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 51 t4 = _mm_slli_si128(t1, 0x4); 53 t4 = _mm_slli_si128(t4, 0x4); 55 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/audio/ncspot/ncspot-0.9.3/cargo-crates/aesni-0.10.0/src/aes256/ |
H A D | expand.rs | 14 t4 = _mm_slli_si128(t1, 0x4); 16 t4 = _mm_slli_si128(t4, 0x4); 18 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 51 t4 = _mm_slli_si128(t1, 0x4); 53 t4 = _mm_slli_si128(t4, 0x4); 55 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/audio/spotifyd/spotifyd-0.3.0/cargo-crates/aesni-0.6.0/src/aes256/ |
H A D | expand.rs | 14 t4 = _mm_slli_si128(t1, 0x4); 16 t4 = _mm_slli_si128(t4, 0x4); 18 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 51 t4 = _mm_slli_si128(t1, 0x4); 53 t4 = _mm_slli_si128(t4, 0x4); 55 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/multimedia/neolink/neolink-cf54129/cargo-crates/aesni-0.10.0/src/aes256/ |
H A D | expand.rs | 14 t4 = _mm_slli_si128(t1, 0x4); 16 t4 = _mm_slli_si128(t4, 0x4); 18 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 51 t4 = _mm_slli_si128(t1, 0x4); 53 t4 = _mm_slli_si128(t4, 0x4); 55 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/devel/py-maturin/maturin-0.11.3/cargo-crates/aesni-0.10.0/src/aes256/ |
H A D | expand.rs | 14 t4 = _mm_slli_si128(t1, 0x4); 16 t4 = _mm_slli_si128(t4, 0x4); 18 t4 = _mm_slli_si128(t4, 0x4); 28 t4 = _mm_slli_si128(t3, 0x4); 30 t4 = _mm_slli_si128(t4, 0x4); 32 t4 = _mm_slli_si128(t4, 0x4); 51 t4 = _mm_slli_si128(t1, 0x4); 53 t4 = _mm_slli_si128(t4, 0x4); 55 t4 = _mm_slli_si128(t4, 0x4);
|
/dports/lang/spidermonkey60/firefox-60.9.0/security/nss/lib/freebl/ |
H A D | aes-x86.c | 16 tmp = _mm_xor_si128(k, _mm_slli_si128(k, 4)); \ 17 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 18 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 41 tmp2 = _mm_slli_si128(k0, 4); \ 43 tmp2 = _mm_slli_si128(tmp2, 4); \ 44 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \ 49 tmp2 = _mm_xor_si128(k1, _mm_slli_si128(k1, 4)); \ 86 tmp2 = _mm_slli_si128(k1x, 4); \ 88 tmp2 = _mm_slli_si128(tmp2, 4); \ 89 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \
|
/dports/games/libretro-tyrquake/tyrquake-e76477d/deps/libFLAC/ |
H A D | lpc_intrin_sse2.c | 95 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 97 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 117 xmm4 = _mm_slli_si128(xmm4, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 119 xmm3 = _mm_slli_si128(xmm3, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 138 xmm4 = _mm_slli_si128(xmm4, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 196 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 198 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 200 xmm9 = _mm_slli_si128(xmm1, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 202 xmm8 = _mm_slli_si128(xmm0, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 204 xmmB = _mm_slli_si128(xmm1, 6); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() [all …]
|
/dports/games/libretro-scummvm/scummvm-7b1e929/backends/platform/libretro/deps/libFLAC/ |
H A D | lpc_intrin_sse2.c | 95 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 97 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 117 xmm4 = _mm_slli_si128(xmm4, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 119 xmm3 = _mm_slli_si128(xmm3, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 138 xmm4 = _mm_slli_si128(xmm4, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 196 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 198 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 200 xmm9 = _mm_slli_si128(xmm1, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 202 xmm8 = _mm_slli_si128(xmm0, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 204 xmmB = _mm_slli_si128(xmm1, 6); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() [all …]
|
/dports/games/libretro-mame2003/mame2003-libretro-4358db4/src/libretro-deps/libFLAC/ |
H A D | lpc_intrin_sse2.c | 95 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 97 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 117 xmm4 = _mm_slli_si128(xmm4, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 119 xmm3 = _mm_slli_si128(xmm3, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 138 xmm4 = _mm_slli_si128(xmm4, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 196 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 198 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 200 xmm9 = _mm_slli_si128(xmm1, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 202 xmm8 = _mm_slli_si128(xmm0, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 204 xmmB = _mm_slli_si128(xmm1, 6); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() [all …]
|
/dports/games/libretro-mame2003_plus/mame2003-plus-libretro-17e9889/src/libretro-deps/libFLAC/ |
H A D | lpc_intrin_sse2.c | 95 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 97 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 117 xmm4 = _mm_slli_si128(xmm4, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 119 xmm3 = _mm_slli_si128(xmm3, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 138 xmm4 = _mm_slli_si128(xmm4, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 196 xmm7 = _mm_slli_si128(xmm1, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 198 xmm2 = _mm_slli_si128(xmm0, 2); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 200 xmm9 = _mm_slli_si128(xmm1, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 202 xmm8 = _mm_slli_si128(xmm0, 4); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() 204 xmmB = _mm_slli_si128(xmm1, 6); in FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2() [all …]
|
/dports/multimedia/handbrake/HandBrake-1.4.2/libhb/ |
H A D | nlmeans_x86.c | 70 ltmp = _mm_slli_si128(lldiff, 4); // Temp shift diff low left 4 bytes in build_integral_sse2() 72 ltmp = _mm_slli_si128(lldiff, 8); // Temp shift diff low left 8 bytes in build_integral_sse2() 78 htmp = _mm_slli_si128(lhdiff, 4); // Temp shift diff high left 4 bytes in build_integral_sse2() 80 htmp = _mm_slli_si128(lhdiff, 8); // Temp shift diff high left 8 bytes in build_integral_sse2() 96 l2tmp = _mm_slli_si128(hldiff, 4); // Temp shift diff low 4 bytes in build_integral_sse2() 98 l2tmp = _mm_slli_si128(hldiff, 8); // Temp shift diff low left 8 bytes in build_integral_sse2() 104 h2tmp = _mm_slli_si128(hhdiff, 4); // Temp shift diff high left 4 bytes in build_integral_sse2() 106 h2tmp = _mm_slli_si128(hhdiff, 8); // Temp shift diff high left 8 bytes in build_integral_sse2()
|
/dports/devel/vc/Vc-1.4.2/Vc/sse/ |
H A D | intrinsics.h | 433 c = _mm_slli_si128(_mm_setallone_si128(), 12); in blend_ps() 445 c = _mm_slli_si128(_mm_setallone_si128(), 8); in blend_ps() 451 c = _mm_slli_si128(_mm_setallone_si128(), 4); in blend_ps() 479 return _mm_unpackhi_epi64(_mm_slli_si128(b, 8), a); in blend_epi16() 490 c = _mm_slli_si128(_mm_setallone_si128(), 14); in blend_epi16() 493 c = _mm_slli_si128(_mm_setallone_si128(), 12); in blend_epi16() 496 c = _mm_slli_si128(_mm_setallone_si128(), 10); in blend_epi16() 499 c = _mm_slli_si128(_mm_setallone_si128(), 8); in blend_epi16() 502 c = _mm_slli_si128(_mm_setallone_si128(), 6); in blend_epi16() 505 c = _mm_slli_si128(_mm_setallone_si128(), 4); in blend_epi16() [all …]
|
/dports/www/firefox-esr/firefox-91.8.0/security/nss/lib/freebl/ |
H A D | aes-x86.c | 16 tmp = _mm_xor_si128(k, _mm_slli_si128(k, 4)); \ 17 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 18 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 41 tmp2 = _mm_slli_si128(k0, 4); \ 43 tmp2 = _mm_slli_si128(tmp2, 4); \ 44 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \ 49 tmp2 = _mm_xor_si128(k1, _mm_slli_si128(k1, 4)); \ 86 tmp2 = _mm_slli_si128(k1x, 4); \ 88 tmp2 = _mm_slli_si128(tmp2, 4); \ 89 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \
|
/dports/security/ca_root_nss/nss-3.71/nss/lib/freebl/ |
H A D | aes-x86.c | 16 tmp = _mm_xor_si128(k, _mm_slli_si128(k, 4)); \ 17 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 18 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 41 tmp2 = _mm_slli_si128(k0, 4); \ 43 tmp2 = _mm_slli_si128(tmp2, 4); \ 44 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \ 49 tmp2 = _mm_xor_si128(k1, _mm_slli_si128(k1, 4)); \ 86 tmp2 = _mm_slli_si128(k1x, 4); \ 88 tmp2 = _mm_slli_si128(tmp2, 4); \ 89 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \
|
/dports/www/firefox/firefox-99.0/security/nss/lib/freebl/ |
H A D | aes-x86.c | 16 tmp = _mm_xor_si128(k, _mm_slli_si128(k, 4)); \ 17 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 18 tmp = _mm_xor_si128(tmp, _mm_slli_si128(tmp, 4)); \ 41 tmp2 = _mm_slli_si128(k0, 4); \ 43 tmp2 = _mm_slli_si128(tmp2, 4); \ 44 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \ 49 tmp2 = _mm_xor_si128(k1, _mm_slli_si128(k1, 4)); \ 86 tmp2 = _mm_slli_si128(k1x, 4); \ 88 tmp2 = _mm_slli_si128(tmp2, 4); \ 89 tmp1 = _mm_xor_si128(_mm_xor_si128(tmp1, tmp2), _mm_slli_si128(tmp2, 4)); \
|