v4fmaddps(zmm1, zmm8, ptr [rdx + 64]);
v4fmaddss(xmm15, xmm8, ptr [rax + 64]);
v4fnmaddps(zmm5 | k5, zmm2, ptr [rcx + 0x80]);
v4fnmaddss(xmm31, xmm2, ptr [rsp + 0x80]);
vp4dpwssd(zmm23 | k7 | T_z, zmm1, ptr [rax + 64]);
vp4dpwssds(zmm10 | k4, zmm3, ptr [rsp + rax * 4 + 64]);
vaesdec(xmm20, xmm30, ptr [rcx + 64]);
vaesdec(ymm1, ymm2, ptr [rcx + 64]);
vaesdec(zmm1, zmm2, ptr [rcx + 64]);
vaesdeclast(xmm20, xmm30, ptr [rax + 64]);
vaesdeclast(ymm20, ymm30, ptr [rax + 64]);
vaesdeclast(zmm20, zmm30, ptr [rax + 64]);
vaesenc(xmm20, xmm30, ptr [rcx + 64]);
vaesenc(ymm1, ymm2, ptr [rcx + 64]);
vaesenc(zmm1, zmm2, ptr [rcx + 64]);
vaesenclast(xmm20, xmm30, ptr [rax + 64]);
vaesenclast(ymm20, ymm30, ptr [rax + 64]);
vaesenclast(zmm20, zmm30, ptr [rax + 64]);
vpclmulqdq(xmm2, xmm3, ptr [rax + 64], 3);
vpclmulqdq(ymm2, ymm3, ptr [rax + 64], 3);
vpclmulqdq(zmm2, zmm3, ptr [rax + 64], 3);
vpclmulqdq(xmm20, xmm3, ptr [rax + 64], 3);
vpclmulqdq(ymm20, ymm3, ptr [rax + 64], 3);
vpclmulqdq(zmm20, zmm3, ptr [rax + 64], 3);
vpcompressb(ptr[rax + 64], xmm1);
vpcompressb(xmm30 | k5, xmm1);
vpcompressb(ptr[rax + 64], ymm1);
vpcompressb(ymm30 | k3 |T_z, ymm1);
vpcompressb(ptr[rax + 64], zmm1);
vpcompressb(zmm30 | k2 |T_z, zmm1);
vpcompressw(ptr[rax + 64], xmm1);
vpcompressw(xmm30 | k5, xmm1);
vpcompressw(ptr[rax + 64], ymm1);
vpcompressw(ymm30 | k3 |T_z, ymm1);
vpcompressw(ptr[rax + 64], zmm1);
vpcompressw(zmm30 | k2 |T_z, zmm1);
vpshldw(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
vpshldw(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
vpshldw(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
vpshldd(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
vpshldd(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
vpshldd(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
vpshldq(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
vpshldq(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
vpshldq(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
vpshldvw(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
vpshldvw(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
vpshldvw(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
vpshldvd(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
vpshldvd(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
vpshldvd(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
vpshldvq(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
vpshldvq(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
vpshldvq(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
vpshrdw(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
vpshrdw(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
vpshrdw(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
vpshrdd(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
vpshrdd(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
vpshrdd(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
vpshrdq(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
vpshrdq(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
vpshrdq(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
vpshrdvw(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
vpshrdvw(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
vpshrdvw(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
vpshrdvd(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
vpshrdvd(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
vpshrdvd(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
vpshrdvq(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
vpshrdvq(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
vpshrdvq(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
vpshrdd(xmm5|k3|T_z, xmm2, ptr_b [rax + 0x40], 5);
vpshrdd(ymm5|k3|T_z, ymm2, ptr_b [rax + 0x40], 5);
vpshrdd(zmm5|k3|T_z, zmm2, ptr_b [rax + 0x40], 5);
vpshrdq(xmm5|k3|T_z, xmm2, ptr_b [rax + 0x40], 5);
vpshrdq(ymm5|k3|T_z, ymm2, ptr_b [rax + 0x40], 5);
vpshrdq(zmm5|k3|T_z, zmm2, ptr_b [rax + 0x40], 5);
vpshrdvd(xmm5|k3|T_z, xmm2, ptr_b [rax + 0x40]);
vpshrdvd(ymm5|k3|T_z, ymm2, ptr_b [rax + 0x40]);
vpshrdvd(zmm5|k3|T_z, zmm2, ptr_b [rax + 0x40]);
vpshrdvq(xmm5|k3|T_z, xmm2, ptr_b [rax + 0x40]);
vpshrdvq(ymm5|k3|T_z, ymm2, ptr_b [rax + 0x40]);
vpshrdvq(zmm5|k3|T_z, zmm2, ptr_b [rax + 0x40]);
vpopcntb(xmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntb(ymm5|k3|T_z, ptr [rax + 0x40]);
vpopcntb(zmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntw(xmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntw(ymm5|k3|T_z, ptr [rax + 0x40]);
vpopcntw(zmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntd(xmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntd(ymm5|k3|T_z, ptr [rax + 0x40]);
vpopcntd(zmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntd(xmm5|k3|T_z, ptr_b [rax + 0x40]);
vpopcntd(ymm5|k3|T_z, ptr_b [rax + 0x40]);
vpopcntd(zmm5|k3|T_z, ptr_b [rax + 0x40]);
vpopcntq(xmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntq(ymm5|k3|T_z, ptr [rax + 0x40]);
vpopcntq(zmm5|k3|T_z, ptr [rax + 0x40]);
vpopcntq(xmm5|k3|T_z, ptr_b [rax + 0x40]);
vpopcntq(ymm5|k3|T_z, ptr_b [rax + 0x40]);
vpopcntq(zmm5|k3|T_z, ptr_b [rax + 0x40]);
vpdpbusd(xmm5|k3|T_z, xmm20, ptr [rax + 0x40]);
vpdpbusd(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpbusd(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpbusd(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpbusd(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpbusd(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
vpdpbusds(xmm5|k3|T_z, xmm20, ptr [rax + 0x40]);
vpdpbusds(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpbusds(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpbusds(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpbusds(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpbusds(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
vpdpwssd(xmm5|k3|T_z, xmm20, ptr [rax + 0x40]);
vpdpwssd(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpwssd(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpwssd(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpwssd(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpwssd(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
vpdpwssds(xmm5|k3|T_z, xmm20, ptr [rax + 0x40]);
vpdpwssds(ymm5|k3|T_z, ymm20, ptr [rax + 0x40]);
vpdpwssds(zmm5|k3|T_z, zmm20, ptr [rax + 0x40]);
vpdpwssds(xmm5|k3|T_z, xmm20, ptr_b [rax + 0x40]);
vpdpwssds(ymm5|k3|T_z, ymm20, ptr_b [rax + 0x40]);
vpdpwssds(zmm5|k3|T_z, zmm20, ptr_b [rax + 0x40]);
vpexpandb(xmm5|k3|T_z, xmm30);
vpexpandb(ymm5|k3|T_z, ymm30);
vpexpandb(zmm5|k3|T_z, zmm30);
vpexpandb(xmm5|k3|T_z, ptr [rax + 0x40]);
vpexpandb(ymm5|k3|T_z, ptr [rax + 0x40]);
vpexpandb(zmm5|k3|T_z, ptr [rax + 0x40]);
vpexpandw(xmm5|k3|T_z, xmm30);
vpexpandw(ymm5|k3|T_z, ymm30);
vpexpandw(zmm5|k3|T_z, zmm30);
vpexpandw(xmm5|k3|T_z, ptr [rax + 0x40]);
vpexpandw(ymm5|k3|T_z, ptr [rax + 0x40]);
vpexpandw(zmm5|k3|T_z, ptr [rax + 0x40]);
vpshufbitqmb(k1|k2, xmm2, ptr [rax + 0x40]);
vpshufbitqmb(k1|k2, ymm2, ptr [rax + 0x40]);
vpshufbitqmb(k1|k2, zmm2, ptr [rax + 0x40]);
gf2p8affineinvqb(xmm1, xmm2, 3);
gf2p8affineinvqb(xmm1, ptr [rax + 0x40], 3);
vgf2p8affineinvqb(xmm1, xmm5, xmm2, 3);
vgf2p8affineinvqb(ymm1, ymm5, ymm2, 3);
vgf2p8affineinvqb(xmm1, xmm5, ptr [rax + 0x40], 3);
vgf2p8affineinvqb(ymm1, ymm5, ptr [rax + 0x40], 3);
vgf2p8affineinvqb(xmm30, xmm31, xmm4, 5);
vgf2p8affineinvqb(ymm30, ymm31, ymm4, 5);
vgf2p8affineinvqb(zmm30, zmm31, zmm4, 5);
vgf2p8affineinvqb(xmm30|k1|T_z, xmm5, ptr [rax + 0x40], 5);
vgf2p8affineinvqb(ymm30|k1|T_z, ymm5, ptr [rax + 0x40], 5);
vgf2p8affineinvqb(zmm30|k1|T_z, zmm5, ptr [rax + 0x40], 5);
vgf2p8affineinvqb(xmm30|k1|T_z, xmm5, ptr_b [rax + 0x40], 5);
vgf2p8affineinvqb(ymm30|k1|T_z, ymm5, ptr_b [rax + 0x40], 5);
vgf2p8affineinvqb(zmm30|k1|T_z, zmm5, ptr_b [rax + 0x40], 5);
gf2p8affineqb(xmm1, xmm2, 3);
gf2p8affineqb(xmm1, ptr [rax + 0x40], 3);
vgf2p8affineqb(xmm1, xmm5, xmm2, 3);
vgf2p8affineqb(ymm1, ymm5, ymm2, 3);
vgf2p8affineqb(xmm1, xmm5, ptr [rax + 0x40], 3);
vgf2p8affineqb(ymm1, ymm5, ptr [rax + 0x40], 3);
vgf2p8affineqb(xmm30, xmm31, xmm4, 5);
vgf2p8affineqb(ymm30, ymm31, ymm4, 5);
vgf2p8affineqb(zmm30, zmm31, zmm4, 5);
vgf2p8affineqb(xmm30|k1|T_z, xmm5, ptr [rax + 0x40], 5);
vgf2p8affineqb(ymm30|k1|T_z, ymm5, ptr [rax + 0x40], 5);
vgf2p8affineqb(zmm30|k1|T_z, zmm5, ptr [rax + 0x40], 5);
vgf2p8affineqb(xmm30|k1|T_z, xmm5, ptr_b [rax + 0x40], 5);
vgf2p8affineqb(ymm30|k1|T_z, ymm5, ptr_b [rax + 0x40], 5);
vgf2p8affineqb(zmm30|k1|T_z, zmm5, ptr_b [rax + 0x40], 5);
gf2p8mulb(xmm1, xmm2);
gf2p8mulb(xmm1, ptr [rax + 0x40]);
vgf2p8mulb(xmm1, xmm5, xmm2);
vgf2p8mulb(ymm1, ymm5, ymm2);
vgf2p8mulb(xmm1, xmm5, ptr [rax + 0x40]);
vgf2p8mulb(ymm1, ymm5, ptr [rax + 0x40]);
vgf2p8mulb(xmm30, xmm31, xmm4);
vgf2p8mulb(ymm30, ymm31, ymm4);
vgf2p8mulb(zmm30, zmm31, zmm4);
vgf2p8mulb(xmm30|k1|T_z, xmm5, ptr [rax + 0x40]);
vgf2p8mulb(ymm30|k1|T_z, ymm5, ptr [rax + 0x40]);
vgf2p8mulb(zmm30|k1|T_z, zmm5, ptr [rax + 0x40]);
vcvtne2ps2bf16(xmm0 | k1, xmm1, ptr [rax + 64]);
vcvtne2ps2bf16(ymm0 | k1 | T_z, ymm0, ptr [rax + 64]);
vcvtne2ps2bf16(zmm0 | k1, zmm1, ptr [rax + 64]);
vcvtneps2bf16(xmm0, xword [rax + 64]);
vcvtneps2bf16(xmm0 | k1, yword [rax + 64]);
vcvtneps2bf16(ymm0 | k1, zword [rax + 64]);
vcvtneps2bf16(ymm0 | k1, ptr [rax + 64]);
vdpbf16ps(xmm0 | k1, xmm1, ptr [rax + 64]);
vdpbf16ps(ymm0 | k1, ymm1, ptr [rax + 64]);
vdpbf16ps(zmm0 | k1, zmm1, ptr [rax + 64]);
ldtilecfg(ptr[rax + rcx * 4 + 64]);
sttilecfg(ptr[rsp + rax * 8 + 128]);
tileloadd(tmm3, ptr[rdi + rdx * 2 + 8]);
tileloaddt1(tmm4, ptr[r8 + r9 + 32]);
tilerelease();
tilestored(ptr[r10 + r11 * 2 + 32], tmm2);
tilezero(tmm7);
tdpbssd(tmm1, tmm2, tmm3);
tdpbsud(tmm2, tmm3, tmm4);
tdpbusd(tmm3, tmm4, tmm5);
tdpbuud(tmm4, tmm5, tmm6);
tdpbf16ps(tmm5, tmm6, tmm7);
tileloadd(tmm1, ptr[r8+r8]);
tileloadd(tmm1, ptr[rax+rcx*4]);
tileloadd(tmm1, ptr[r8+r9*1+0x40]);
vaddph(zmm0, zmm1, ptr[rax+64]);
vaddph(ymm0, ymm1, ptr[rax+64]);
vaddph(xmm0, xmm1, ptr[rax+64]);
vaddph(zmm0, zmm1, ptr_b[rax+64]);
vaddph(ymm0, ymm1, ptr_b[rax+64]);
vaddph(xmm0, xmm1, ptr_b[rax+64]);
vaddsh(xmm0, xmm15, ptr[rax+64]);
vaddsh(xmm0|k5|T_z|T_rd_sae, xmm15, xmm3);
vcmpph(k1, xm15, ptr[rax+64], 1);
vcmpph(k2, ym15, ptr[rax+64], 2);
vcmpph(k3, zm15, ptr[rax+64], 3);
vcmpph(k1, xm15, ptr_b[rax+64], 1);
vcmpph(k2, ym15, ptr_b[rax+64], 2);
vcmpph(k3, zm15, ptr_b[rax+64], 3);
vcmpsh(k1, xm15, ptr[rax+64], 1);
vcmpsh(k3|k5, xmm1, xmm25|T_sae, 4);
vcomish(xmm1, ptr[rax+64]);
vcomish(xmm1|T_sae, xmm15);
vucomish(xmm1, ptr [rax+0x40]);
vucomish(xmm1|T_sae, xmm15);
vfmaddsub213ph(xmm1, xmm2, ptr [rax+0x40]);
vfmaddsub213ph(xmm1, xmm2, ptr_b [rax+0x40]);
vfmaddsub213ph(xmm1|k3, xmm2, xmm5);
vfmaddsub213ph(ymm1, ymm2, ptr [rax+0x40]);
vfmaddsub213ph(ymm1, ymm2, ptr_b[rax+0x40]);
vfmaddsub213ph(ymm1|k3, ymm2, ymm5);
vfmaddsub213ph(zmm1, zmm2, ptr [rax+0x40]);
vfmaddsub213ph(zmm1, zmm2, ptr_b [rax+0x40]);
vfmaddsub213ph(zmm1|T_ru_sae, zmm2, zmm5);
vfmsubadd132ph(xmm1, xmm2, ptr [rax+0x40]);
vfmsubadd132ph(xmm1, xmm2, ptr_b [rax+0x40]);
vfmsubadd132ph(ymm1, ymm2, ptr [rax+0x40]);
vfmsubadd132ph(ymm1, ymm2, ptr_b [rax+0x40]);
vfmsubadd132ph(zmm1, zmm2, ptr [rax+0x40]);
vfmsubadd132ph(zmm1, zmm2, ptr_b [rax+0x40]);
vfmsubadd132ph(zmm1|T_ru_sae, zmm2, zmm5);
vfmadd132ph(xmm1, xmm2, ptr [rax+0x40]);
vfmadd132ph(xmm1, xmm2, ptr_b [rax+0x40]);
vfmadd132ph(ymm1, ymm2, ptr [rax+0x40]);
vfmadd132ph(ymm1, ymm2, ptr_b [rax+0x40]);
vfmadd132ph(zmm1, zmm2, ptr [rax+0x40]);
vfmadd132ph(zmm1, zmm2, ptr_b [rax+0x40]);
vfmadd132ph(zmm1|T_rd_sae, zmm2, zmm5);
vfmsub231ph(xmm1, xmm2, ptr [rax+0x40]);
vfmsub231ph(xmm1, xmm2, ptr_b [rax+0x40]);
vfmsub231ph(ymm1, ymm2, ptr [rax+0x40]);
vfmsub231ph(ymm1, ymm2, ptr_b [rax+0x40]);
vfmsub231ph(zmm1, zmm2, ptr [rax+0x40]);
vfmsub231ph(zmm1, zmm2, ptr_b [rax+0x40]);
vfmsub231ph(zmm1|T_rd_sae, zmm2, zmm5);
vfnmsub231ph(xmm1, xmm2, ptr [rax+0x40]);
vfnmsub231ph(ymm1, ymm2, ptr_b [rax+0x40]);
vfnmsub231ph(zmm1, zmm2, ptr_b [rax+0x40]);
vfnmsub231ph(zmm1|T_rd_sae, zmm2, zmm5);
vfmadd132sh(xmm1|k1|T_z|T_rd_sae, xmm2, xmm3);
vfmadd132sh(xmm1, xmm2, ptr [rax+0x40]);
vfnmadd132sh(xmm1|k1|T_z|T_rd_sae, xmm2, xmm3);
vfnmadd132sh(xmm1, xmm2, ptr [rax+0x40]);
vfmsub132sh(xmm1|k1|T_z|T_rd_sae, xmm2, xmm3);
vfmsub132sh(xmm1, xmm2, ptr [rax+0x40]);
vfnmsub132sh(xmm1|k1|T_z|T_rd_sae, xmm2, xmm3);
vfnmsub132sh(xmm1, xmm2, ptr [rax+0x40]);
vfcmaddcph(xmm1|k1|T_z, xmm2, ptr [rax+0x40]);
vfcmaddcph(ymm1|k1|T_z, ymm2, ptr [rax+0x40]);
vfcmaddcph(zmm1|k1, zmm2, ptr [rax+0x40]);
vfcmaddcph(zmm1|k1|T_rd_sae, zmm2, zmm5);
vfcmaddcph(xmm1|k1|T_z, xmm2, ptr_b [rax+0x40]);
vfcmaddcph(ymm1|k1|T_z, ymm2, ptr_b [rax+0x40]);
vfcmaddcph(zmm1|k1|T_z, zmm2, ptr_b [rax+0x40]);
vfmaddcph(xm1, xm2, ptr[rax+0x40]);
vfmaddcph(ym1|k1|T_z, ym2, ptr_b[rax+0x40]);
vfmaddcph(zm1, zm2, ptr_b[rax+0x40]);
vfcmulcph(xmm1, xmm2, ptr [rax+0x40]);
vfcmulcph(ymm1|k1|T_z, ymm2, ptr_b [rax+0x40]);
vfcmulcph(zmm1, zmm2, ptr_b [rax+0x40]);
vfmulcph(xmm1, xmm2, ptr [rax+0x40]);
vfmulcph(ymm1|k1|T_z, ymm2, ptr_b [rax+0x40]);
vfmulcph(zmm1, zmm2, ptr_b [rax+0x40]);
vrcpph(xmm1, ptr [rax+0x40]);
vrcpph(xmm1, ptr_b [rax+0x40]);
vrcpph(ymm1, ptr [rax+0x40]);
vrcpph(ymm1, ptr_b [rax+0x40]);
vrcpph(zmm1, ptr [rax+0x40]);
vrcpph(zmm1, ptr_b [rax+0x40]);
vrcpsh(xmm1, xmm3, ptr [rax+0x40]);
vrsqrtph(xmm1, ptr [rax+0x40]);
vrsqrtph(xmm1, ptr_b [rax+0x40]);
vrsqrtph(ymm2, ptr [rax+0x40]);
vrsqrtph(ymm2, ptr_b [rax+0x40]);
vrsqrtph(zmm2, ptr [rax+0x40]);
vrsqrtph(zmm2, ptr_b [rax+0x40]);
vrsqrtsh(xmm1|k5|T_z, xmm7, ptr [rax+0x40]);
vsqrtph(xmm1|k4|T_z, ptr [rax+0x40]);
vsqrtph(xmm1|k4|T_z, ptr_b [rax+0x40]);
vsqrtph(ymm1|k4|T_z, ptr_b [rax+0x40]);
vsqrtph(zmm1|k4|T_z, ptr [rax+0x40]);
vsqrtph(zmm1|k4|T_z, ptr_b [rax+0x40]);
vsqrtsh(xmm1|k4|T_z, xmm5, ptr [rax+0x40]);
vsqrtsh(xmm1|k4|T_z|T_rd_sae, xmm5, xmm7);
vscalefph(xmm1, xmm5, ptr [rax+0x40]);
vscalefph(xmm1, xmm5, ptr_b [rax+0x40]);
vscalefph(ymm1, ymm5, ptr [rax+0x40]);
vscalefph(ymm1, ymm5, ptr_b [rax+0x40]);
vscalefph(zmm1, zmm5, ptr [rax+0x40]);
vscalefph(zmm1, zmm5, ptr_b [rax+0x40]);
vscalefph(zmm1|k1|T_z|T_rd_sae, zmm5, zmm7);
vscalefsh(xmm1, xmm5, ptr [rax+0x40]);
vscalefsh(xmm1|k1|T_z|T_rd_sae, xmm5, xmm7);
vreduceph(xmm1, ptr [rax+0x40], 0x1);
vreduceph(xmm1, ptr_b [rax+0x40], 0x2);
vreduceph(ymm1, ptr [rax+0x40], 0x3);
vreduceph(ymm1, ptr_b [rax+0x40], 0x4);
vreduceph(zmm1, ptr [rax+0x40], 0x5);
vreduceph(zmm1, ptr_b [rax+0x40], 0x6);
vreduceph(zmm1|k1|T_z|T_sae, zmm5, 0x7);
vreducesh(xmm1, xmm3, ptr [rax+0x40], 0x1);
vreducesh(xmm1|k1|T_z|T_sae, xmm5, xmm4, 0x2);
vrndscaleph(xmm1, ptr [rax+0x40], 0x1);
vrndscaleph(xmm1, ptr_b [rax+0x40], 0x2);
vrndscaleph(ymm1, ptr [rax+0x40], 0x3);
vrndscaleph(ymm1, ptr_b [rax+0x40], 0x4);
vrndscaleph(zmm1, ptr [rax+0x40], 0x5);
vrndscaleph(zmm1, ptr_b [rax+0x40], 0x6);
vrndscaleph(zmm1|k1|T_z|T_sae, zmm5, 0x7);
vrndscalesh(xmm1, xmm3, ptr [rax+0x40], 0x1);
vrndscalesh(xmm1|k1|T_z|T_sae, xmm5, xmm4, 0x2);
vfpclassph(k1, xword [rax+0x40], 0x1);
vfpclassph(k1, xword_b[rax+0x40], 0x2);
vfpclassph(k1, yword [rax+0x40], 0x3);
vfpclassph(k1, yword_b[rax+0x40], 0x4);
vfpclassph(k1, zword [rax+0x40], 0x5);
vfpclassph(k1, zword_b[rax+0x40], 0x6);
vfpclasssh(k1|k2, xmm3, 0x5);
vfpclasssh(k1|k2, ptr [rax+0x40], 0x5);
vgetexpph(xmm1, ptr [rax+0x40]);
vgetexpph(ymm1, ptr_b [rax+0x40]);
vgetexpph(zmm1, ptr [rax+0x40]);
vgetexpph(zmm1|k1|T_z|T_sae, zmm5);
vgetexpsh(xmm1, xmm5, ptr [rax+0x40]);
vgetexpsh(xmm1|k1|T_z|T_sae, xmm3, xmm5);
vgetmantph(xmm1, ptr [rax+0x40], 0x1);
vgetmantph(ymm1, ptr_b [rax+0x40], 0x2);
vgetmantph(zmm1, ptr [rax+0x40], 0x3);
vgetmantph(zmm1|k1|T_z|T_sae, zmm5, 0x4);
vgetmantsh(xmm1, xmm5, ptr [rax+0x40], 0x5);
vgetmantsh(xmm1|k1|T_z|T_sae, xmm3, xmm5, 0x6);
vmovsh(xmm1|k1|T_z, ptr [rax+0x40]);
vmovsh(ptr [rax+0x40]|k1, xmm1);
vmovsh(xmm1|k2|T_z, xmm3, xmm5);
vmovw(xmm1, r13d);
vmovw(xmm3, ptr [rax+0x40]);
vmovw(r9d, xmm1);
vmovw(ptr [rax+0x40], xmm7);
vcvtsd2sh(xmm1|k1|T_z|T_rd_sae, xmm2, xmm3);
vcvtsd2sh(xmm1, xmm2, ptr [rax+0x40]);
vcvtsh2sd(xmm1|k1|T_z|T_sae, xmm2, xmm3);
vcvtsh2sd(xmm1, xmm2, ptr [rax+0x40]);
vcvtsh2ss(xmm1|k1|T_z|T_sae, xmm2, xmm3);
vcvtsh2ss(xmm1, xmm2, ptr [rax+0x40]);
vcvtss2sh(xmm1|k1|T_z|T_rd_sae, xmm2, xmm3);
vcvtss2sh(xmm1, xmm2, ptr [rax+0x40]);
vcvtsh2si(edx|T_rd_sae, xmm1);
vcvtsh2si(edx, ptr [rax+0x40]);
vcvtsh2si(rdx|T_rd_sae, xmm1);
vcvtsh2si(r8, ptr [rax+0x40]);
vcvtph2dq(xmm1, xmm5);
vcvtph2dq(xmm1, ptr [rax+0x40]);
vcvtph2dq(xmm1, ptr_b [rax+0x40]);
vcvtph2dq(ymm1|k2|T_z, xmm5);
vcvtph2dq(ymm1, ptr [rax+0x40]);
vcvtph2dq(ymm1, ptr_b [rax+0x40]);
vcvtph2dq(zmm1|k5|T_z|T_rd_sae, ymm3);
vcvtph2dq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvtph2dq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvtph2psx(xmm1, xmm5);
vcvtph2psx(xmm1, ptr [rax+0x40]);
vcvtph2psx(xmm1, ptr_b [rax+0x40]);
vcvtph2psx(ymm1|k2|T_z, xmm5);
vcvtph2psx(ymm1, ptr [rax+0x40]);
vcvtph2psx(ymm1, ptr_b [rax+0x40]);
vcvtph2psx(zmm1|k5|T_z|T_sae, ymm3);
vcvtph2psx(zmm1|k5|T_z, ptr [rax+0x40]);
vcvtph2psx(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvtph2udq(xmm1, xmm5);
vcvtph2udq(xmm1, ptr [rax+0x40]);
vcvtph2udq(xmm1, ptr_b [rax+0x40]);
vcvtph2udq(ymm1|k2|T_z, xmm5);
vcvtph2udq(ymm1, ptr [rax+0x40]);
vcvtph2udq(ymm1, ptr_b [rax+0x40]);
vcvtph2udq(zmm1|k5|T_z|T_rd_sae, ymm3);
vcvtph2udq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvtph2udq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvttph2dq(xmm1, xmm5);
vcvttph2dq(xmm1, ptr [rax+0x40]);
vcvttph2dq(xmm1, ptr_b [rax+0x40]);
vcvttph2dq(ymm1|k2|T_z, xmm5);
vcvttph2dq(ymm1, ptr [rax+0x40]);
vcvttph2dq(ymm1, ptr_b [rax+0x40]);
vcvttph2dq(zmm1|k5|T_z|T_sae, ymm3);
vcvttph2dq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvttph2dq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvttph2udq(xmm1, xmm5);
vcvttph2udq(xmm1, ptr [rax+0x40]);
vcvttph2udq(xmm1, ptr_b [rax+0x40]);
vcvttph2udq(ymm1|k2|T_z, xmm5);
vcvttph2udq(ymm1, ptr [rax+0x40]);
vcvttph2udq(ymm1, ptr_b [rax+0x40]);
vcvttph2udq(zmm1|k5|T_z|T_sae, ymm3);
vcvttph2udq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvttph2udq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvtph2pd(xmm1, xmm5);
vcvtph2pd(xmm1, ptr [rax+0x40]);
vcvtph2pd(xmm1, ptr_b [rax+0x40]);
vcvtph2pd(ymm1|k2|T_z, xmm5);
vcvtph2pd(ymm1, ptr [rax+0x40]);
vcvtph2pd(ymm1, ptr_b [rax+0x40]);
vcvtph2pd(zmm1|k5|T_z|T_sae, xmm3);
vcvtph2pd(zmm1|k5|T_z, ptr [rax+0x40]);
vcvtph2pd(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvtph2qq(xmm1, xmm5);
vcvtph2qq(xmm1, ptr [rax+0x40]);
vcvtph2qq(xmm1, ptr_b [rax+0x40]);
vcvtph2qq(ymm1|k2|T_z, xmm5);
vcvtph2qq(ymm1, ptr [rax+0x40]);
vcvtph2qq(ymm1, ptr_b [rax+0x40]);
vcvtph2qq(zmm1|k5|T_z|T_rd_sae, xmm3);
vcvtph2qq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvtph2qq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvtph2uqq(xmm1, xmm5);
vcvtph2uqq(xmm1, ptr [rax+0x40]);
vcvtph2uqq(xmm1, ptr_b [rax+0x40]);
vcvtph2uqq(ymm1|k2|T_z, xmm5);
vcvtph2uqq(ymm1, ptr [rax+0x40]);
vcvtph2uqq(ymm1, ptr_b [rax+0x40]);
vcvtph2uqq(zmm1|k5|T_z|T_rd_sae, xmm3);
vcvtph2uqq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvtph2uqq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvttph2uqq(xmm1, xmm5);
vcvttph2uqq(xmm1, ptr [rax+0x40]);
vcvttph2uqq(xmm1, ptr_b [rax+0x40]);
vcvttph2uqq(ymm1|k2|T_z, xmm5);
vcvttph2uqq(ymm1, ptr [rax+0x40]);
vcvttph2uqq(ymm1, ptr_b [rax+0x40]);
vcvttph2uqq(zmm1|k5|T_z|T_sae, xmm3);
vcvttph2uqq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvttph2uqq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvtdq2ph(xmm1, xmm5);
vcvtdq2ph(xmm1, xword [rax+0x40]);
vcvtdq2ph(xmm1, xword_b [rax+0x40]);
vcvtdq2ph(xmm1, yword [rax+0x40]);
vcvtdq2ph(xmm1, yword_b [rax+0x40]);
vcvtdq2ph(ymm1|k2|T_z|T_rd_sae, zmm5);
vcvtdq2ph(ymm1, ptr [rax+0x40]);
vcvtdq2ph(ymm1, ptr_b [rax+0x40]);
vcvtps2phx(xmm1, xmm5);
vcvtps2phx(xmm1, xword [rax+0x40]);
vcvtps2phx(xmm1, xword_b [rax+0x40]);
vcvtps2phx(xmm1, yword [rax+0x40]);
vcvtps2phx(xmm1, yword_b [rax+0x40]);
vcvtps2phx(ymm1|k2|T_z|T_rd_sae, zmm5);
vcvtps2phx(ymm1, ptr [rax+0x40]);
vcvtps2phx(ymm1, ptr_b [rax+0x40]);
vcvtudq2ph(xmm1, xmm5);
vcvtudq2ph(xmm1, xword [rax+0x40]);
vcvtudq2ph(xmm1, xword_b [rax+0x40]);
vcvtudq2ph(xmm1, yword [rax+0x40]);
vcvtudq2ph(xmm1, yword_b [rax+0x40]);
vcvtudq2ph(ymm1|k2|T_z|T_rd_sae, zmm5);
vcvtudq2ph(ymm1, ptr [rax+0x40]);
vcvtudq2ph(ymm1, ptr_b [rax+0x40]);
vcvtpd2ph(xmm1, xmm5);
vcvtpd2ph(xmm1, ymm5);
vcvtpd2ph(xmm1|k2|T_z|T_rd_sae, zmm5);
vcvtpd2ph(xmm1, xword [rax+0x40]);
vcvtpd2ph(xmm1, xword_b [rax+0x40]);
vcvtpd2ph(xmm1, yword [rax+0x40]);
vcvtpd2ph(xmm1, yword_b [rax+0x40]);
vcvtpd2ph(xmm1, zword [rax+0x40]);
vcvtpd2ph(xmm1, zword_b [rax+0x40]);
vcvtqq2ph(xmm1, xmm5);
vcvtqq2ph(xmm1, ymm5);
vcvtqq2ph(xmm1|k2|T_z|T_rd_sae, zmm5);
vcvtqq2ph(xmm1, xword [rax+0x40]);
vcvtqq2ph(xmm1, xword_b [rax+0x40]);
vcvtqq2ph(xmm1, yword [rax+0x40]);
vcvtqq2ph(xmm1, yword_b [rax+0x40]);
vcvtqq2ph(xmm1, zword [rax+0x40]);
vcvtqq2ph(xmm1, zword_b [rax+0x40]);
vcvtuqq2ph(xmm1, xmm5);
vcvtuqq2ph(xmm1, ymm5);
vcvtuqq2ph(xmm1|k2|T_z|T_rd_sae, zmm5);
vcvtuqq2ph(xmm1, xword [rax+0x40]);
vcvtuqq2ph(xmm1, xword_b [rax+0x40]);
vcvtuqq2ph(xmm1, yword [rax+0x40]);
vcvtuqq2ph(xmm1, yword_b [rax+0x40]);
vcvtuqq2ph(xmm1, zword [rax+0x40]);
vcvtuqq2ph(xmm1, zword_b [rax+0x40]);
vcvtph2uw(xmm1, xmm5);
vcvtph2uw(xmm1, ptr [rax+0x40]);
vcvtph2uw(xmm1, ptr_b [rax+0x40]);
vcvtph2uw(ymm1, ptr [rax+0x40]);
vcvtph2uw(ymm1, ptr_b [rax+0x40]);
vcvtph2uw(zmm1|k2|T_z|T_rd_sae, zmm5);
vcvtph2uw(zmm1, ptr [rax+0x40]);
vcvtph2uw(zmm1, ptr_b [rax+0x40]);
vcvtph2w(xmm1, xmm5);
vcvtph2w(xmm1, ptr [rax+0x40]);
vcvtph2w(xmm1, ptr_b [rax+0x40]);
vcvtph2w(ymm1, ptr [rax+0x40]);
vcvtph2w(ymm1, ptr_b [rax+0x40]);
vcvtph2w(zmm1|k2|T_z|T_rd_sae, zmm5);
vcvtph2w(zmm1, ptr [rax+0x40]);
vcvtph2w(zmm1, ptr_b [rax+0x40]);
vcvttph2uw(xmm1, xmm5);
vcvttph2uw(xmm1, ptr [rax+0x40]);
vcvttph2uw(xmm1, ptr_b [rax+0x40]);
vcvttph2uw(ymm1, ptr [rax+0x40]);
vcvttph2uw(ymm1, ptr_b [rax+0x40]);
vcvttph2uw(zmm1|k2|T_z|T_sae, zmm5);
vcvttph2uw(zmm1, ptr [rax+0x40]);
vcvttph2uw(zmm1, ptr_b [rax+0x40]);
vcvttph2w(xmm1, xmm5);
vcvttph2w(xmm1, ptr [rax+0x40]);
vcvttph2w(xmm1, ptr_b [rax+0x40]);
vcvttph2w(ymm1, ptr [rax+0x40]);
vcvttph2w(ymm1, ptr_b [rax+0x40]);
vcvttph2w(zmm1|k2|T_z|T_sae, zmm5);
vcvttph2w(zmm1, ptr [rax+0x40]);
vcvttph2w(zmm1, ptr_b [rax+0x40]);
vcvtuw2ph(xmm1, xmm5);
vcvtuw2ph(xmm1, ptr [rax+0x40]);
vcvtuw2ph(xmm1, ptr_b [rax+0x40]);
vcvtuw2ph(ymm1, ptr [rax+0x40]);
vcvtuw2ph(ymm1, ptr_b [rax+0x40]);
vcvtuw2ph(zmm1|k2|T_z|T_rd_sae, zmm5);
vcvtuw2ph(zmm1, ptr [rax+0x40]);
vcvtuw2ph(zmm1, ptr_b [rax+0x40]);
vcvtw2ph(xmm1, xmm5);
vcvtw2ph(xmm1, ptr [rax+0x40]);
vcvtw2ph(xmm1, ptr_b [rax+0x40]);
vcvtw2ph(ymm1, ptr [rax+0x40]);
vcvtw2ph(ymm1, ptr_b [rax+0x40]);
vcvtw2ph(zmm1|k2|T_z|T_rd_sae, zmm5);
vcvtw2ph(zmm1, ptr [rax+0x40]);
vcvtw2ph(zmm1, ptr_b [rax+0x40]);
vcvtps2ph(xmm1, xmm2, 0x1);
vcvtps2ph(ptr [rax+0x40], xmm2, 0x2);
vcvtps2ph(xmm1, ymm2, 0x3);
vcvtps2ph(ptr [rax+0x40], ymm2, 0x4);
vcvtps2ph(xmm1|k1|T_z, xmm2, 0x5);
vcvtps2ph(ptr [rax+0x40]|k1, xmm3, 0x6);
vcvtps2ph(xmm1|k2, ymm4, 0x7);
vcvtps2ph(ptr [rax+0x40]|k2, ymm5, 0x8);
vcvtps2ph(ymm1|k2|T_sae, zmm5, 0x9);
vcvtps2ph(ptr [rax+0x40]|k5, zmm4, 0xa);
vcvtsh2usi(ecx|T_rd_sae, xmm1);
vcvtsh2usi(eax, ptr [rax+0x40]);
vcvtsh2usi(r9|T_rd_sae, xmm1);
vcvtsh2usi(r13, ptr [rax+0x40]);
vcvttsh2si(ecx|T_sae, xmm1);
vcvttsh2si(eax, ptr [rax+0x40]);
vcvttsh2si(r9|T_sae, xmm1);
vcvttsh2si(r13, ptr [rax+0x40]);
vcvttsh2usi(ecx|T_sae, xmm1);
vcvttsh2usi(eax, ptr [rax+0x40]);
vcvttsh2usi(r9|T_sae, xmm1);
vcvttsh2usi(r13, ptr [rax+0x40]);
vcvttph2qq(xmm1, xmm5);
vcvttph2qq(xmm1, ptr [rax+0x40]);
vcvttph2qq(xmm1, ptr_b [rax+0x40]);
vcvttph2qq(ymm1|k2|T_z, xmm5);
vcvttph2qq(ymm1, ptr [rax+0x40]);
vcvttph2qq(ymm1, ptr_b [rax+0x40]);
vcvttph2qq(zmm1|k5|T_z|T_sae, xmm3);
vcvttph2qq(zmm1|k5|T_z, ptr [rax+0x40]);
vcvttph2qq(zmm1|k5|T_z, ptr_b [rax+0x40]);
vcvtsi2sh(xmm1|T_rd_sae, xmm2, eax);
vcvtsi2sh(xmm1, xmm2, dword [rax+0x40]);
vcvtsi2sh(xmm1|T_rd_sae, xmm2, r9);
vcvtsi2sh(xmm1, xmm2, qword [rax+0x40]);
vcvtusi2sh(xmm1|T_rd_sae, xmm2, eax);
vcvtusi2sh(xmm1, xmm2, dword [rax+0x40]);
vcvtusi2sh(xmm1|T_rd_sae, xmm2, r9);
vcvtusi2sh(xmm1, xmm2, qword [rax+0x40]);
aadd(ptr[rax], ecx);
aadd(ptr[eax], ecx);
aadd(ptr[rax], r10);
aand(ptr[rax], ecx);
aand(ptr[eax], ecx);
aand(ptr[rax], r10);
aor(ptr[rax], ecx);
aor(ptr[eax], ecx);
aor(ptr[rax], r10);
axor(ptr[rax], ecx);
axor(ptr[eax], ecx);
axor(ptr[rax], r10);
cmpbexadd(ptr[rax+r10*4], rcx, rdx);
cmpbxadd(ptr[rax+r10*4], rcx, rdx);
cmplexadd(ptr[rax+r10*4], rcx, rdx);
cmplxadd(ptr[rax+r10*4], rcx, rdx);
cmpnbexadd(ptr[rax+r10*4], rcx, rdx);
cmpnbxadd(ptr[rax+r10*4], rcx, rdx);
cmpnlexadd(ptr[rax+r10*4], rcx, rdx);
cmpnlxadd(ptr[rax+r10*4], rcx, rdx);
cmpnoxadd(ptr[rax+r10*4], rcx, rdx);
cmpnpxadd(ptr[rax+r10*4], rcx, rdx);
cmpnsxadd(ptr[rax+r10*4], rcx, rdx);
cmpnzxadd(ptr[rax+r10*4], rcx, rdx);
cmpoxadd(ptr[rax+r10*4], rcx, rdx);
cmppxadd(ptr[rax+r10*4], rcx, rdx);
cmpsxadd(ptr[rax+r10*4], rcx, rdx);
cmpzxadd(ptr[rax+r10*4], rcx, rdx);
vsha512msg1(ymm3, xmm5);
vsha512msg2(ymm9, ymm10);
vsha512rnds2(ymm1, ymm3, xmm2);
vsm3msg1(xmm1, xmm2, xmm3);
vsm3msg1(xmm1, xmm2, ptr [rax]);
vsm3msg2(xmm5, xmm7, xmm3);
vsm3msg2(xmm5, xmm6, ptr [rax]);
vsm3rnds2(xmm5, xmm7, xmm3, 0x12);
vsm3rnds2(xmm5, xmm7, ptr [rcx], 0x34);
vsm4key4(xmm1, xmm2, xmm3);
vsm4key4(xmm1, xmm2, ptr [rdx]);
vsm4rnds4(xmm1, xmm2, xmm3);
vsm4rnds4(xmm5, xmm6, ptr [rcx+rax*4]);
vpdpbssd(xmm1, xmm2, xmm3);
vpdpbssd(ymm1, ymm2, ptr [rax]);
vpdpbssds(xmm1, xmm2, xmm3);
vpdpbssds(ymm1, ymm2, ptr [rax]);
vpdpbsud(xmm1, xmm2, xmm3);
vpdpbsud(ymm1, ymm2, ptr [rax]);
vpdpbsuds(xmm1, xmm2, xmm3);
vpdpbsuds(ymm1, ymm2, ptr [rax]);
vpdpbuud(xmm1, xmm2, xmm3);
vpdpbuud(ymm1, ymm2, ptr [rax]);
vpdpbuuds(xmm1, xmm2, xmm3);
vpdpbuuds(ymm1, ymm2, ptr [rax]);
vpdpwsud(xmm1, xmm2, xmm3);
vpdpwsud(ymm1, ymm2, ptr [rax]);
vpdpwsuds(xmm1, xmm2, xmm3);
vpdpwsuds(ymm1, ymm2, ptr [rax]);
vpdpwusd(xmm1, xmm2, xmm3);
vpdpwusd(ymm1, ymm2, ptr [rax]);
vpdpwusds(xmm1, xmm2, xmm3);
vpdpwusds(ymm1, ymm2, ptr [rax]);
vpdpwuud(xmm1, xmm2, xmm3);
vpdpwuud(ymm1, ymm2, ptr [rax]);
vpdpwuuds(xmm1, xmm2, xmm3);
vpdpwuuds(ymm1, ymm2, ptr [rax]);