; FMUL_R e0, a2 mulpd xmm4, xmm10 ; IADD_RC r2, r5, 2673743102 lea r10, [r10+r13-1621224194] ; ISTORE L2[r2], r7 mov eax, r10d and eax, 262136 mov qword ptr [rsi+rax], r15 ; FNEG_R f2 xorps xmm2, xmm15 ; IMUL_9C r6, 3291464084 lea r14, [r14+r14*8-1003503212] ; FSUB_R f1, a0 subpd xmm1, xmm8 ; IXOR_M r5, L2[r3] mov eax, r11d and eax, 262136 xor r13, qword ptr [rsi+rax] ; FNEG_R f2 xorps xmm2, xmm15 ; FSUB_R f3, a0 subpd xmm3, xmm8 ; ISDIV_C r0, 1400272688 mov rax, 7072565507528518045 imul r8 xor eax, eax sar rdx, 29 sets al add rdx, rax add r8, rdx ; IMUL_M r3, L1[r7] mov eax, r15d and eax, 16376 imul r11, qword ptr [rsi+rax] ; ISWAP_R r2, r3 xchg r10, r11 ; IMULH_R r6, r0 mov rax, r14 mul r8 mov r14, rdx ; FMUL_R e0, a2 mulpd xmm4, xmm10 ; IADD_RC r3, r4, 4242706868 lea r11, [r11+r12-52260428] ; IADD_R r7, 3156349536 add r15, -1138617760 ; IXOR_M r2, L1[r6] mov eax, r14d and eax, 16376 xor r10, qword ptr [rsi+rax] ; FSUB_M f2, L1[r5] mov eax, r13d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] subpd xmm2, xmm12 ; IXOR_R r7, r1 xor r15, r9 ; COND_R r2, lt(r7, 4253348488) xor ecx, ecx cmp r15d, -41618808 setl cl add r10, rcx ; FMUL_R e3, a0 mulpd xmm7, xmm8 ; COND_R r4, sg(r1, 3333776931) xor ecx, ecx cmp r9d, -961190365 sets cl add r12, rcx ; FADD_R f2, a1 addpd xmm2, xmm9 ; FSUB_R f0, a3 subpd xmm0, xmm11 ; ISTORE L1[r6], r2 mov eax, r14d and eax, 16376 mov qword ptr [rsi+rax], r10 ; ISUB_R r6, r5 sub r14, r13 ; IADD_M r0, L1[r4] mov eax, r12d and eax, 16376 add r8, qword ptr [rsi+rax] ; ISTORE L1[r4], r3 mov eax, r12d and eax, 16376 mov qword ptr [rsi+rax], r11 ; COND_M r6, sg(L1[r6], 1048782623) xor ecx, ecx mov eax, r14d and eax, 16376 cmp dword ptr [rsi+rax], 1048782623 sets cl add r14, rcx ; FSQRT_R e0 sqrtpd xmm4, xmm4 ; INEG_R r2 neg r10 ; FSQRT_R e1 sqrtpd xmm5, xmm5 ; FMUL_R e1, a3 mulpd xmm5, xmm11 ; IMUL_R r7, r6 imul r15, r14 ; IMULH_R r0, r4 mov rax, r8 mul r12 mov r8, rdx ; IMUL_R r5, r3 imul r13, r11 ; FSQRT_R e2 sqrtpd xmm6, xmm6 ; FADD_R f3, a0 addpd xmm3, xmm8 ; IADD_R r3, r2 add r11, r10 ; FADD_R f1, a0 addpd xmm1, xmm8 ; FMUL_R e3, a2 mulpd xmm7, xmm10 ; FADD_M f0, L2[r5] mov eax, r13d and eax, 262136 cvtdq2pd xmm12, qword ptr [rsi+rax] addpd xmm0, xmm12 ; IMUL_R r5, r6 imul r13, r14 ; IADD_RC r1, r2, 3031682053 lea r9, [r9+r10-1263285243] ; ISUB_M r4, L1[r6] mov eax, r14d and eax, 16376 sub r12, qword ptr [rsi+rax] ; FSWAP_R e3 shufpd xmm7, xmm7, 1 ; IMUL_R r0, r7 imul r8, r15 ; IXOR_R r1, r6 xor r9, r14 ; IXOR_M r2, L1[r4] mov eax, r12d and eax, 16376 xor r10, qword ptr [rsi+rax] ; FSUB_R f3, a1 subpd xmm3, xmm9 ; ISTORE L1[r0], r5 mov eax, r8d and eax, 16376 mov qword ptr [rsi+rax], r13 ; FDIV_M e2, L2[r3] mov eax, r11d and eax, 262136 cvtdq2pd xmm12, qword ptr [rsi+rax] andps xmm12, xmm14 divpd xmm6, xmm12 maxpd xmm6, xmm13 ; ISWAP_R r2, r0 xchg r10, r8 ; IADD_R r7, r5 add r15, r13 ; FDIV_M e0, L1[r4] mov eax, r12d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] andps xmm12, xmm14 divpd xmm4, xmm12 maxpd xmm4, xmm13 ; FADD_R f3, a1 addpd xmm3, xmm9 ; FADD_R f0, a3 addpd xmm0, xmm11 ; IADD_R r2, r0 add r10, r8 ; ISTORE L1[r3], r6 mov eax, r11d and eax, 16376 mov qword ptr [rsi+rax], r14 ; IXOR_R r1, r7 xor r9, r15 ; ISUB_M r5, L2[r7] mov eax, r15d and eax, 262136 sub r13, qword ptr [rsi+rax] ; ISDIV_C r7, 266992378 mov rax, -9173520256920442565 imul r15 xor eax, eax add rdx, r15 sar rdx, 27 sets al add rdx, rax add r15, rdx ; FDIV_M e3, L1[r4] mov eax, r12d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] andps xmm12, xmm14 divpd xmm7, xmm12 maxpd xmm7, xmm13 ; IMUL_R r2, r0 imul r10, r8 ; FMUL_R e3, a2 mulpd xmm7, xmm10 ; IMUL_R r0, r6 imul r8, r14 ; ISTORE L1[r0], r7 mov eax, r8d and eax, 16376 mov qword ptr [rsi+rax], r15 ; FSUB_M f0, L2[r1] mov eax, r9d and eax, 262136 cvtdq2pd xmm12, qword ptr [rsi+rax] subpd xmm0, xmm12 ; FADD_R f3, a1 addpd xmm3, xmm9 ; IXOR_R r5, r4 xor r13, r12 ; ISTORE L2[r7], r2 mov eax, r15d and eax, 262136 mov qword ptr [rsi+rax], r10 ; FSWAP_R e2 shufpd xmm6, xmm6, 1 ; FADD_R f3, a2 addpd xmm3, xmm10 ; ISMULH_R r5, r0 mov rax, r13 imul r8 mov r13, rdx ; IADD_M r0, L1[r4] mov eax, r12d and eax, 16376 add r8, qword ptr [rsi+rax] ; COND_R r7, ge(r6, 2322068811) xor ecx, ecx cmp r14d, -1972898485 setge cl add r15, rcx ; FADD_R f2, a2 addpd xmm2, xmm10 ; IROR_R r7, r6 mov ecx, r14d ror r15, cl ; IADD_RC r2, r4, 4177509323 lea r10, [r10+r12-117457973] ; IMUL_R r0, 2794074228 imul r8, -1500893068 ; IADD_R r2, r3 add r10, r11 ; FSQRT_R e2 sqrtpd xmm6, xmm6 ; IROR_R r7, r4 mov ecx, r12d ror r15, cl ; IMUL_9C r4, 381194890 lea r12, [r12+r12*8+381194890] ; IADD_RC r3, r7, 1050899263 lea r11, [r11+r15+1050899263] ; IADD_R r2, r7 add r10, r15 ; FMUL_R e3, a0 mulpd xmm7, xmm8 ; IADD_RC r6, r6, 540663146 lea r14, [r14+r14+540663146] ; IROR_R r5, 58 ror r13, 58 ; FADD_R f2, a1 addpd xmm2, xmm9 ; FADD_R f2, a2 addpd xmm2, xmm10 ; FMUL_R e1, a2 mulpd xmm5, xmm10 ; FSWAP_R e1 shufpd xmm5, xmm5, 1 ; IADD_R r5, r3 add r13, r11 ; IADD_R r7, 2514699120 add r15, -1780268176 ; IADD_RC r7, r0, 2797210442 lea r15, [r15+r8-1497756854] ; ISTORE L2[r0], r7 mov eax, r8d and eax, 262136 mov qword ptr [rsi+rax], r15 ; ISMULH_R r2, r4 mov rax, r10 imul r12 mov r10, rdx ; FSUB_M f0, L1[r2] mov eax, r10d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] subpd xmm0, xmm12 ; ISMULH_R r2, r3 mov rax, r10 imul r11 mov r10, rdx ; IADD_R r0, r3 add r8, r11 ; ISUB_R r7, r2 sub r15, r10 ; FADD_R f2, a0 addpd xmm2, xmm8 ; FMUL_R e0, a2 mulpd xmm4, xmm10 ; FADD_M f2, L1[r3] mov eax, r11d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] addpd xmm2, xmm12 ; IMUL_R r1, r2 imul r9, r10 ; IMUL_M r7, L1[r5] mov eax, r13d and eax, 16376 imul r15, qword ptr [rsi+rax] ; IMUL_R r3, r2 imul r11, r10 ; IXOR_R r1, r0 xor r9, r8 ; FNEG_R f0 xorps xmm0, xmm15 ; IADD_RC r4, r4, 1456841848 lea r12, [r12+r12+1456841848] ; IXOR_R r3, r2 xor r11, r10 ; COND_R r0, of(r4, 1678513610) xor ecx, ecx cmp r12d, 1678513610 seto cl add r8, rcx ; ISMULH_R r4, r4 mov rax, r12 imul r12 mov r12, rdx ; IMUL_R r4, r1 imul r12, r9 ; FADD_R f1, a2 addpd xmm1, xmm10 ; FSUB_R f2, a0 subpd xmm2, xmm8 ; FMUL_R e1, a2 mulpd xmm5, xmm10 ; FSUB_R f0, a3 subpd xmm0, xmm11 ; IXOR_R r0, r7 xor r8, r15 ; ISTORE L2[r1], r4 mov eax, r9d and eax, 262136 mov qword ptr [rsi+rax], r12 ; IXOR_M r7, L1[r6] mov eax, r14d and eax, 16376 xor r15, qword ptr [rsi+rax] ; ISUB_R r2, r4 sub r10, r12 ; ISUB_M r4, L1[r6] mov eax, r14d and eax, 16376 sub r12, qword ptr [rsi+rax] ; FADD_R f2, a2 addpd xmm2, xmm10 ; FSUB_M f3, L2[r4] mov eax, r12d and eax, 262136 cvtdq2pd xmm12, qword ptr [rsi+rax] subpd xmm3, xmm12 ; IXOR_R r7, r2 xor r15, r10 ; IXOR_R r0, r5 xor r8, r13 ; FADD_R f1, a2 addpd xmm1, xmm10 ; FMUL_R e3, a2 mulpd xmm7, xmm10 ; FSWAP_R e3 shufpd xmm7, xmm7, 1 ; FSWAP_R f1 shufpd xmm1, xmm1, 1 ; COND_R r2, ge(r2, 4068636356) xor ecx, ecx cmp r10d, -226330940 setge cl add r10, rcx ; FMUL_R e2, a3 mulpd xmm6, xmm11 ; FSUB_M f2, L2[r1] mov eax, r9d and eax, 262136 cvtdq2pd xmm12, qword ptr [rsi+rax] subpd xmm2, xmm12 ; FADD_R f1, a0 addpd xmm1, xmm8 ; ISUB_R r7, r5 sub r15, r13 ; ISUB_M r0, L1[r1] mov eax, r9d and eax, 16376 sub r8, qword ptr [rsi+rax] ; FSUB_R f3, a1 subpd xmm3, xmm9 ; ISWAP_R r3, r5 xchg r11, r13 ; IADD_RC r5, r2, 795784298 lea r13, [r13+r10+795784298] ; IADD_RC r0, r4, 2244788743 lea r8, [r8+r12-2050178553] ; IMUL_9C r5, 1062534001 lea r13, [r13+r13*8+1062534001] ; FADD_R f0, a2 addpd xmm0, xmm10 ; FMUL_R e3, a1 mulpd xmm7, xmm9 ; IDIV_C r3, 1662492575 mov rax, 11914062610815620875 mul r11 shr rdx, 30 add r11, rdx ; IMUL_M r5, L1[r0] mov eax, r8d and eax, 16376 imul r13, qword ptr [rsi+rax] ; IDIV_C r4, 1963597892 mov rax, r12 shr rax, 2 mov rcx, 1260889558222626443 mul rcx shr rdx, 25 add r12, rdx ; IMUL_9C r7, 1820045218 lea r15, [r15+r15*8+1820045218] ; IMUL_M r0, L1[r3] mov eax, r11d and eax, 16376 imul r8, qword ptr [rsi+rax] ; IXOR_R r3, r7 xor r11, r15 ; ISMULH_R r4, r2 mov rax, r12 imul r10 mov r12, rdx ; ISWAP_R r3, r0 xchg r11, r8 ; IXOR_R r2, r0 xor r10, r8 ; IXOR_M r0, L2[r1] mov eax, r9d and eax, 262136 xor r8, qword ptr [rsi+rax] ; ISDIV_C r7, 3359520316 mov rax, 7859804860668271393 imul r15 xor eax, eax sub rdx, r15 sar rdx, 29 sets al add rdx, rax add r15, rdx ; IMUL_M r6, L1[r2] mov eax, r10d and eax, 16376 imul r14, qword ptr [rsi+rax] ; FNEG_R f3 xorps xmm3, xmm15 ; IADD_RC r4, r2, 1704868083 lea r12, [r12+r10+1704868083] ; FADD_R f2, a0 addpd xmm2, xmm8 ; ISTORE L1[r0], r0 mov eax, r8d and eax, 16376 mov qword ptr [rsi+rax], r8 ; FADD_M f0, L1[r7] mov eax, r15d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] addpd xmm0, xmm12 ; FMUL_R e0, a3 mulpd xmm4, xmm11 ; FSUB_R f3, a2 subpd xmm3, xmm10 ; IADD_RC r7, r7, 1302457878 lea r15, [r15+r15+1302457878] ; ISUB_R r1, 1330165941 sub r9, 1330165941 ; FNEG_R f1 xorps xmm1, xmm15 ; IROR_R r0, r4 mov ecx, r12d ror r8, cl ; FSUB_R f1, a0 subpd xmm1, xmm8 ; IROR_R r5, r6 mov ecx, r14d ror r13, cl ; COND_R r0, ab(r1, 3984033425) xor ecx, ecx cmp r9d, -310933871 seta cl add r8, rcx ; COND_R r4, ab(r7, 757929676) xor ecx, ecx cmp r15d, 757929676 seta cl add r12, rcx ; FMUL_R e0, a1 mulpd xmm4, xmm9 ; IMUL_R r1, r3 imul r9, r11 ; ISUB_R r3, r2 sub r11, r10 ; FSUB_R f3, a2 subpd xmm3, xmm10 ; FDIV_M e1, L1[r4] mov eax, r12d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] andps xmm12, xmm14 divpd xmm5, xmm12 maxpd xmm5, xmm13 ; FSWAP_R f1 shufpd xmm1, xmm1, 1 ; IADD_R r7, 2873779272 add r15, -1421188024 ; FSUB_M f3, L2[r2] mov eax, r10d and eax, 262136 cvtdq2pd xmm12, qword ptr [rsi+rax] subpd xmm3, xmm12 ; FSUB_R f2, a3 subpd xmm2, xmm11 ; FSUB_R f3, a1 subpd xmm3, xmm9 ; FMUL_R e1, a3 mulpd xmm5, xmm11 ; IADD_RC r2, r4, 3977135268 lea r10, [r10+r12-317832028] ; IMUL_M r4, L1[r5] mov eax, r13d and eax, 16376 imul r12, qword ptr [rsi+rax] ; FDIV_M e1, L1[r7] mov eax, r15d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] andps xmm12, xmm14 divpd xmm5, xmm12 maxpd xmm5, xmm13 ; IADD_R r5, r2 add r13, r10 ; ISUB_R r4, 401020510 sub r12, 401020510 ; IROR_R r3, r0 mov ecx, r8d ror r11, cl ; ISTORE L1[r7], r0 mov eax, r15d and eax, 16376 mov qword ptr [rsi+rax], r8 ; FSUB_R f2, a1 subpd xmm2, xmm9 ; FMUL_R e3, a1 mulpd xmm7, xmm9 ; IMUL_9C r3, 720965215 lea r11, [r11+r11*8+720965215] ; IMUL_9C r6, 74948046 lea r14, [r14+r14*8+74948046] ; ISTORE L1[r7], r3 mov eax, r15d and eax, 16376 mov qword ptr [rsi+rax], r11 ; IXOR_R r2, r6 xor r10, r14 ; FMUL_R e3, a1 mulpd xmm7, xmm9 ; ISUB_R r4, r1 sub r12, r9 ; ISUB_R r3, r0 sub r11, r8 ; ISWAP_R r7, r5 xchg r15, r13 ; IMUL_R r2, r6 imul r10, r14 ; COND_R r2, ge(r2, 2402809790) xor ecx, ecx cmp r10d, -1892157506 setge cl add r10, rcx ; FADD_R f1, a3 addpd xmm1, xmm11 ; IADD_R r7, r0 add r15, r8 ; IDIV_C r1, 624867857 mov rax, 15848983434401622933 mul r9 shr rdx, 29 add r9, rdx ; FADD_R f0, a1 addpd xmm0, xmm9 ; IADD_RC r5, r7, 3817376178 lea r13, [r13+r15-477591118] ; FSUB_R f0, a3 subpd xmm0, xmm11 ; ISUB_M r6, L1[r2] mov eax, r10d and eax, 16376 sub r14, qword ptr [rsi+rax] ; FMUL_R e3, a1 mulpd xmm7, xmm9 ; IADD_R r0, r4 add r8, r12 ; FSUB_R f3, a1 subpd xmm3, xmm9 ; FSUB_M f2, L1[r4] mov eax, r12d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] subpd xmm2, xmm12 ; ISDIV_C r2, 3898255608 mov rax, 5964731804029407733 imul r10 xor eax, eax sub rdx, r10 sar rdx, 28 sets al add rdx, rax add r10, rdx ; FNEG_R f2 xorps xmm2, xmm15 ; FSUB_R f3, a2 subpd xmm3, xmm10 ; FADD_R f1, a3 addpd xmm1, xmm11 ; IMUL_R r3, r2 imul r11, r10 ; FADD_M f0, L1[r3] mov eax, r11d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] addpd xmm0, xmm12 ; ISMULH_R r5, r2 mov rax, r13 imul r10 mov r13, rdx ; IMULH_R r6, r2 mov rax, r14 mul r10 mov r14, rdx ; FADD_M f3, L1[r3] mov eax, r11d and eax, 16376 cvtdq2pd xmm12, qword ptr [rsi+rax] addpd xmm3, xmm12 ; IMUL_R r6, r7 imul r14, r15 ; FSUB_R f0, a0 subpd xmm0, xmm8 ; FNEG_R f2 xorps xmm2, xmm15 ; ISUB_R r6, r4 sub r14, r12 ; FADD_R f1, a1 addpd xmm1, xmm9 ; IXOR_R r0, r5 xor r8, r13 ; FADD_R f2, a1 addpd xmm2, xmm9 ; ISWAP_R r7, r5 xchg r15, r13 ; FMUL_R e3, a2 mulpd xmm7, xmm10 ; IADD_RC r3, r6, 2977336568 lea r11, [r11+r14-1317630728] ; IMUL_R r2, r3 imul r10, r11 ; IADD_RC r1, r4, 894105694 lea r9, [r9+r12+894105694] ; IMUL_9C r7, 504293473 lea r15, [r15+r15*8+504293473] ; FSUB_R f1, a0 subpd xmm1, xmm8 ; IMUL_R r7, r1 imul r15, r9 ; IXOR_R r2, r4 xor r10, r12 ; IADD_RC r0, r1, 392362094 lea r8, [r8+r9+392362094] ; IDIV_C r4, 1645771433 mov rax, 376097195048767223 mul r12 shr rdx, 25 add r12, rdx ; ISUB_R r4, r3 sub r12, r11 ; ISUB_M r7, L1[r4] mov eax, r12d and eax, 16376 sub r15, qword ptr [rsi+rax] ; IMUL_M r5, L1[r7] mov eax, r15d and eax, 16376 imul r13, qword ptr [rsi+rax] ; IROR_R r1, r7 mov ecx, r15d ror r9, cl ; INEG_R r4 neg r12 ; IMUL_R r3, 1863959234 imul r11, 1863959234 ; IROR_R r4, 59 ror r12, 59 ; IMUL_M r1, L3[363256] imul r9, qword ptr [rsi+363256] ; ISTORE L2[r6], r7 mov eax, r14d and eax, 262136 mov qword ptr [rsi+rax], r15 ; ISTORE L1[r1], r5 mov eax, r9d and eax, 16376 mov qword ptr [rsi+rax], r13 ; FNEG_R f0 xorps xmm0, xmm15 ; FSQRT_R e2 sqrtpd xmm6, xmm6 ; FMUL_R e0, a3 mulpd xmm4, xmm11 ; FMUL_R e3, a2 mulpd xmm7, xmm10 ; IROR_R r5, r2 mov ecx, r10d ror r13, cl ; IADD_R r0, r4 add r8, r12