File lapack_564f5cd6dd718723530ccdd421a63387e76d5fcdb30f49d1f0425db7bc512296.dll
Function:
    37b7acc80
    37b7acc80slasd1_:
    37b7acc80push r15
    37b7acc82push r14
    37b7acc84push r13
    37b7acc86push r12
    37b7acc88push rbp
    37b7acc89push rdi
    37b7acc8apush rsi
    37b7acc8bpush rbx
    37b7acc8csub rsp, 296
    37b7acc93mov rax, [rsp+472]
    37b7acc9bmov rdi, [rsp+400]
    37b7acca3mov rsi, [rsp+408]
    37b7accabmov rbp, [rsp+464]
    37b7accb3mov [rax], 0
    37b7accb9mov [rsp+376], rdx
    37b7accc1movsxd rdx, [rcx]
    37b7accc4mov rbx, r9
    37b7accc7mov [rsp+368], rcx
    37b7acccfmov [rsp+384], r8
    37b7accd7test edx, edx
    37b7accd9jle 0x37b7ad1d0
    37b7accdfmov rax, [rsp+376]
    37b7acce7mov eax, [rax]
    37b7acce9test eax, eax
    37b7accebjle 0x37b7acd48
    37b7accedmov r12l, [r8]
    37b7accf0mov ecx, 3
    37b7accf5mov r8l, -3
    37b7accfbcmp r12l, 1
    37b7accffjbe 0x37b7acd58
    37b7acd01
    37b7acd01loc_37b7acd01:
    37b7acd01mov rax, [rsp+472]
    37b7acd09mov [rsp+284], ecx
    37b7acd10lea rdx, [rsp+284]
    37b7acd18lea rcx, [rip+0xc8bcc5]
    37b7acd1fmov [rax], r8l
    37b7acd22mov r8l, 6
    37b7acd28call xerbla_
    37b7acd2dnop
    37b7acd2e
    37b7acd2eloc_37b7acd2e:
    37b7acd2eadd rsp, 296
    37b7acd35pop rbx
    37b7acd36pop rsi
    37b7acd37pop rdi
    37b7acd38pop rbp
    37b7acd39pop r12
    37b7acd3bpop r13
    37b7acd3dpop r14
    37b7acd3fpop r15
    37b7acd41ret
    37b7acd42
    37b7acd42 66 0f 1f 44 00 00 f..D..
    37b7acd48
    37b7acd48loc_37b7acd48:
    37b7acd48mov r8l, -2
    37b7acd4emov ecx, 2
    37b7acd53jmp 0x37b7acd01
    37b7acd55
    37b7acd55 0f 1f 00 ...
    37b7acd58
    37b7acd58loc_37b7acd58:
    37b7acd58lea ecx, [rdx+rax]
    37b7acd5bmovss x1, [rdi]
    37b7acd5fmovss x0, [rsi]
    37b7acd63mov [r9+4*rdx], 0
    37b7acd6blea r15l, [rcx+1]
    37b7acd6fandps x1, [rip+0xc8bc7a]
    37b7acd76mov rdx, r9
    37b7acd79mov r8l, 1
    37b7acd7fadd r12l, r15l
    37b7acd82mov r14l, r15l
    37b7acd85andps x0, [rip+0xc8bc64]
    37b7acd8cmov [rsp+272], r15l
    37b7acd94imul r14l, r15l
    37b7acd98mov eax, r12l
    37b7acd9blea r13l, [r15+r12+1]
    37b7acda0mov [rsp+264], r15l
    37b7acda8imul eax, r12l
    37b7acdacmaxss x1, x0
    37b7acdb0mov [rsp+268], r12l
    37b7acdb8add r14l, r13l
    37b7acdbbmovss [rsp+284], x1
    37b7acdc4add eax, r14l
    37b7acdc7mov [rsp+220], eax
    37b7acdcelea eax, [rcx+r15+2]
    37b7acdd3movsxd rcx, ecx
    37b7acdd6lea r11l, [r15+rax]
    37b7acddalea r9, [r9+4*rcx+4]
    37b7acddfxor ecx, ecx
    37b7acde1mov [rsp+224], r11l
    37b7acde9nop
    37b7acdeanop [rax+rax]
    37b7acdf5nop [rax+rax]
    37b7ace00
    37b7ace00loc_37b7ace00:
    37b7ace00movss x0, [rdx]
    37b7ace04andps x0, [rip+0xc8bbe5]
    37b7ace0bcomiss x0, x1
    37b7ace0emaxss x0, x1
    37b7ace12cmova ecx, r8l
    37b7ace16add rdx, 4
    37b7ace1amovaps x1, x0
    37b7ace1dcmp r9, rdx
    37b7ace20jne 0x37b7ace00
    37b7ace22test cl, cl
    37b7ace24je 0x37b7ace2f
    37b7ace26movss [rsp+284], x0
    37b7ace2f
    37b7ace2floc_37b7ace2f:
    37b7ace2fmov rcx, [rsp+472]
    37b7ace37lea rdx, [rip+0xc8bb9e]
    37b7ace3elea r10, [rsp+272]
    37b7ace46mov [rsp+80], 1
    37b7ace4fmov [rsp+64], r10
    37b7ace54movsxd r15, r15l
    37b7ace57sub r14l, 1
    37b7ace5bsub r13l, 1
    37b7ace5fmov [rsp+72], rcx
    37b7ace64movsxd r14, r14l
    37b7ace67movsxd r13, r13l
    37b7ace6amovsxd r12, r12l
    37b7ace6dmov [rsp+40], r10
    37b7ace72lea r13, [rbp+4*r13]
    37b7ace77lea r12, [rbp+4*r12]
    37b7ace7clea rcx, [rip+0xc8bb55]
    37b7ace83mov [rsp+32], rdx
    37b7ace88lea r8, [rip+0xc8bb41]
    37b7ace8fmov [rsp+48], rcx
    37b7ace94mov rdx, r8
    37b7ace97lea rcx, [rip+0xc8bb36]
    37b7ace9emov [rsp+56], rbx
    37b7acea3mov [rsp+200], eax
    37b7aceaalea rax, [rsp+284]
    37b7aceb2mov r9, rax
    37b7aceb5mov [rsp+248], r10
    37b7acebdmov [rsp+208], rax
    37b7acec5call slascl_
    37b7acecamovss x0, [rdi]
    37b7acecemovss x1, [rsp+284]
    37b7aced7lea r9, [rsp+256]
    37b7acedfmov eax, [rsp+200]
    37b7acee6mov rdx, [rsp+456]
    37b7aceeedivss x0, x1
    37b7acef2mov ecx, [rsp+224]
    37b7acef9sub eax, 1
    37b7acefccdqe
    37b7acefelea r11, [rdx+4*rax]
    37b7acf02lea rax, [rdx+4*r15]
    37b7acf06mov rdx, [rsp+472]
    37b7acf0emov [rsp+200], rax
    37b7acf16lea r15, [rbp+4*r14]
    37b7acf1blea r14, [rsp+264]
    37b7acf23mov [rsp+240], r11
    37b7acf2bmovss [rdi], x0
    37b7acf2fmovss x0, [rsi]
    37b7acf33divss x0, x1
    37b7acf37movss [rsi], x0
    37b7acf3bmov [rsp+176], rdx
    37b7acf43mov rdx, [rsp+448]
    37b7acf4bmov [rsp+152], rax
    37b7acf53lea eax, [rcx-1]
    37b7acf56mov [rsp+160], rdx
    37b7acf5emov rdx, [rsp+456]
    37b7acf66cdqe
    37b7acf68mov [rsp+168], r11
    37b7acf70lea rax, [rdx+4*rax]
    37b7acf74mov [rsp+144], rdx
    37b7acf7cmov rdx, [rsp+440]
    37b7acf84mov [rsp+136], rax
    37b7acf8clea rax, [rsp+268]
    37b7acf94mov [rsp+88], rdx
    37b7acf99mov rdx, [rsp+432]
    37b7acfa1mov [rsp+128], rax
    37b7acfa9mov [rsp+80], rdx
    37b7acfaemov rdx, [rsp+424]
    37b7acfb6mov [rsp+232], rax
    37b7acfbemov [rsp+72], rdx
    37b7acfc3mov rdx, [rsp+416]
    37b7acfcbmov [rsp+120], r15
    37b7acfd0mov [rsp+112], r14
    37b7acfd5mov [rsp+104], r13
    37b7acfdamov [rsp+96], r12
    37b7acfdfmov [rsp+64], rdx
    37b7acfe4mov r8, [rsp+384]
    37b7acfecmov [rsp+56], rsi
    37b7acff1mov rdx, [rsp+376]
    37b7acff9mov [rsp+48], rdi
    37b7acffemov rcx, [rsp+368]
    37b7ad006mov [rsp+40], rbp
    37b7ad00bmov [rsp+32], rbx
    37b7ad010mov [rsp+224], r9
    37b7ad018call slasd2_
    37b7ad01dmov rax, [rsp+200]
    37b7ad025mov rsi, [rsp+472]
    37b7ad02dmov [rsp+32], rbx
    37b7ad032mov edx, [rsp+256]
    37b7ad039mov rdi, [rsp+440]
    37b7ad041mov [rsp+144], rbp
    37b7ad049mov [rsp+128], rax
    37b7ad051mov rax, [rsp+232]
    37b7ad059mov [rsp+152], rsi
    37b7ad061mov rsi, [rsp+432]
    37b7ad069mov [rsp+120], rax
    37b7ad06elea rax, [rsp+260]
    37b7ad076mov r11, [rsp+240]
    37b7ad07emov [rsp+48], rax
    37b7ad083mov eax, [rsp+220]
    37b7ad08amov [rsp+260], edx
    37b7ad091mov r9, [rsp+224]
    37b7ad099sub eax, 1
    37b7ad09cmov [rsp+104], rdi
    37b7ad0a1mov r8, [rsp+384]
    37b7ad0a9mov rdi, [rsp+424]
    37b7ad0b1mov [rsp+96], rsi
    37b7ad0b6cdqe
    37b7ad0b8mov rsi, [rsp+416]
    37b7ad0c0lea rax, [rbp+4*rax]
    37b7ad0c5mov rdx, [rsp+376]
    37b7ad0cdmov [rsp+136], r11
    37b7ad0d5mov rcx, [rsp+368]
    37b7ad0ddmov [rsp+72], rdi
    37b7ad0e2mov [rsp+40], rax
    37b7ad0e7mov [rsp+112], r15
    37b7ad0ecmov [rsp+88], r14
    37b7ad0f1mov [rsp+80], r13
    37b7ad0f6mov [rsp+64], rsi
    37b7ad0fbmov [rsp+56], r12
    37b7ad100call slasd3_
    37b7ad105mov rdi, [rsp+472]
    37b7ad10dmov eax, [rdi]
    37b7ad10ftest eax, eax
    37b7ad111jne 0x37b7acd2e
    37b7ad117mov r10, [rsp+248]
    37b7ad11flea rcx, [rip+0xc8b8b2]
    37b7ad126mov rax, [rsp+208]
    37b7ad12emov [rsp+80], 1
    37b7ad137mov [rsp+48], rcx
    37b7ad13clea r8, [rip+0xc8b88d]
    37b7ad143lea r9, [rip+0xc8b892]
    37b7ad14amov [rsp+32], rax
    37b7ad14fmov rdx, r8
    37b7ad152lea rcx, [rip+0xc8b87b]
    37b7ad159mov [rsp+72], rdi
    37b7ad15emov [rsp+64], r10
    37b7ad163mov [rsp+56], rbx
    37b7ad168mov [rsp+40], r10
    37b7ad16dcall slascl_
    37b7ad172mov edx, [rsp+256]
    37b7ad179mov r8, rbx
    37b7ad17cmov eax, [rsp+272]
    37b7ad183lea rcx, [rsp+276]
    37b7ad18blea r9, [rip+0xc8b846]
    37b7ad192sub eax, edx
    37b7ad194mov [rsp+276], edx
    37b7ad19blea rdx, [rsp+280]
    37b7ad1a3mov [rsp+280], eax
    37b7ad1aamov rax, [rsp+448]
    37b7ad1b2mov [rsp+40], rax
    37b7ad1b7lea rax, [rip+0xc8b822]
    37b7ad1bemov [rsp+32], rax
    37b7ad1c3call slamrg_
    37b7ad1c8jmp 0x37b7acd2e
    37b7ad1cd
    37b7ad1cd 0f 1f 00 ...
    37b7ad1d0
    37b7ad1d0loc_37b7ad1d0:
    37b7ad1d0mov r8l, -1
    37b7ad1d6mov ecx, 1
    37b7ad1dbjmp 0x37b7acd01
    37b7acc80
    37b7ad1e0
    37b7ad1e0 // align 0x40
    37b7ad200
    37b7ad200slasd2_:
    37b7ad200push r15
    37b7ad202push r14
    37b7ad204push r13
    37b7ad206push r12
    37b7ad208push rbp
    37b7ad209push rdi
    37b7ad20apush rsi
    37b7ad20bpush rbx
    37b7ad20csub rsp, 328
    37b7ad213movups [rsp+272], x6
    37b7ad21bmovups [rsp+288], x7
    37b7ad223movups [rsp+304], x8
    37b7ad22cmov rsi, [rsp+488]
    37b7ad234mov rax, [rsp+576]
    37b7ad23cmov r15, [rsp+432]
    37b7ad244mov rbx, [rsp+440]
    37b7ad24cmov [rax], 0
    37b7ad252mov rbp, [rsp+552]
    37b7ad25amov r11, [rsp+568]
    37b7ad262mov r8l, [r8]
    37b7ad265mov [rsp+400], rcx
    37b7ad26dmov ecx, [rsi]
    37b7ad26fmov rsi, [rsp+512]
    37b7ad277mov [rsp+424], r9
    37b7ad27fmov [rsp+408], rdx
    37b7ad287mov rdi, [rsp+408]
    37b7ad28fmov r9l, [rsi]
    37b7ad292mov rsi, [rsp+472]
    37b7ad29amov r14l, [rdi]
    37b7ad29dmov edx, [rsi]
    37b7ad29fmov rsi, [rsp+528]
    37b7ad2a7mov esi, [rsi]
    37b7ad2a9mov [rsp+188], esi
    37b7ad2b0mov rsi, [rsp+400]
    37b7ad2b8mov esi, [rsi]
    37b7ad2batest esi, esi
    37b7ad2bcjle 0x37b7ad3a8
    37b7ad2c2test r14l, r14l
    37b7ad2c5jle 0x37b7ad388
    37b7ad2cbcmp r8l, 1
    37b7ad2cfjbe 0x37b7ad2d7
    37b7ad2d1mov [rax], -3
    37b7ad2d7
    37b7ad2d7loc_37b7ad2d7:
    37b7ad2d7lea r13l, [rsi+r14]
    37b7ad2dblea edi, [r13+1]
    37b7ad2dflea r12l, [rdi+r8]
    37b7ad2e3mov [rsp+72], edi
    37b7ad2e7mov [rsp+224], edi
    37b7ad2eemov [rsp+220], r12l
    37b7ad2f6cmp edx, edi
    37b7ad2f8jl 0x37b7ad398
    37b7ad2fecmp ecx, r12l
    37b7ad301jl 0x37b7ad3b8
    37b7ad307cmp r9l, edi
    37b7ad30ajl 0x37b7ad3d0
    37b7ad310cmp [rsp+188], r12l
    37b7ad318jge 0x37b7ad3df
    37b7ad31emov ecx, -17
    37b7ad323mov edx, 17
    37b7ad328nop [rax+rax]
    37b7ad330
    37b7ad330loc_37b7ad330:
    37b7ad330mov [rax], ecx
    37b7ad332
    37b7ad332loc_37b7ad332:
    37b7ad332mov [rsp+256], edx
    37b7ad339mov r8l, 6
    37b7ad33flea rdx, [rsp+256]
    37b7ad347lea rcx, [rip+0xc8b6b2]
    37b7ad34ecall xerbla_
    37b7ad353nop
    37b7ad354
    37b7ad354loc_37b7ad354:
    37b7ad354movups x6, [rsp+272]
    37b7ad35cmovups x7, [rsp+288]
    37b7ad364movups x8, [rsp+304]
    37b7ad36dadd rsp, 328
    37b7ad374pop rbx
    37b7ad375pop rsi
    37b7ad376pop rdi
    37b7ad377pop rbp
    37b7ad378pop r12
    37b7ad37apop r13
    37b7ad37cpop r14
    37b7ad37epop r15
    37b7ad380ret
    37b7ad381
    37b7ad381 0f 1f 80 00 00 00 00 .......
    37b7ad388
    37b7ad388loc_37b7ad388:
    37b7ad388mov [rax], -2
    37b7ad38ejmp 0x37b7ad2d7
    37b7ad393
    37b7ad393 0f 1f 44 00 00 ..D..
    37b7ad398
    37b7ad398loc_37b7ad398:
    37b7ad398mov ecx, -10
    37b7ad39dmov edx, 10
    37b7ad3a2jmp 0x37b7ad330
    37b7ad3a4
    37b7ad3a4 0f 1f 40 00 ..@.
    37b7ad3a8
    37b7ad3a8loc_37b7ad3a8:
    37b7ad3a8mov [rax], -1
    37b7ad3aejmp 0x37b7ad2d7
    37b7ad3b3
    37b7ad3b3 0f 1f 44 00 00 ..D..
    37b7ad3b8
    37b7ad3b8loc_37b7ad3b8:
    37b7ad3b8mov ecx, -12
    37b7ad3bdmov edx, 12
    37b7ad3c2jmp 0x37b7ad330
    37b7ad3c7
    37b7ad3c7 66 0f 1f 84 00 00 00 00 00 f........
    37b7ad3d0
    37b7ad3d0loc_37b7ad3d0:
    37b7ad3d0mov ecx, -15
    37b7ad3d5mov edx, 15
    37b7ad3dajmp 0x37b7ad330
    37b7ad3df
    37b7ad3dfloc_37b7ad3df:
    37b7ad3dfmovsxd rdi, ecx
    37b7ad3e2xor ecx, ecx
    37b7ad3e4mov eax, [rax]
    37b7ad3e6test rdi, rdi
    37b7ad3e9cmovns rcx, rdi
    37b7ad3edmovsxd rdi, r9l
    37b7ad3f0mov [rsp+152], rcx
    37b7ad3f8not rcx
    37b7ad3fbmov [rsp+144], rcx
    37b7ad403xor ecx, ecx
    37b7ad405test rdi, rdi
    37b7ad408cmovns rcx, rdi
    37b7ad40cmovsxd rdi, edx
    37b7ad40fxor edx, edx
    37b7ad411test rdi, rdi
    37b7ad414cmovns rdx, rdi
    37b7ad418mov [rsp+168], rcx
    37b7ad420not rcx
    37b7ad423mov [rsp+176], rcx
    37b7ad42bmov [rsp+112], rdx
    37b7ad430not rdx
    37b7ad433mov [rsp+120], rdx
    37b7ad438test eax, eax
    37b7ad43ajne 0x37b7ae723
    37b7ad440lea edi, [rsi+2]
    37b7ad443lea eax, [rsi+1]
    37b7ad446mov [rsp+184], edi
    37b7ad44dmov rdi, [rsp+448]
    37b7ad455mov [rsp+84], eax
    37b7ad459cdqe
    37b7ad45bmovss x3, [rdi]
    37b7ad45fmov rdi, [rsp+152]
    37b7ad467imul rdi, rax
    37b7ad46bmov [rsp+88], rdi
    37b7ad470add rdi, [rsp+144]
    37b7ad478add rax, rdi
    37b7ad47bmov [rsp+96], rdi
    37b7ad480mov rdi, [rsp+480]
    37b7ad488movss x0, [rdi+4*rax]
    37b7ad48dmulss x0, x3
    37b7ad491movss [rsp+232], x0
    37b7ad49amovss [rbx], x0
    37b7ad49etest esi, esi
    37b7ad4a0jle 0x37b7ae6b0
    37b7ad4a6lea edi, [rsi-1]
    37b7ad4a9cmp edi, 2
    37b7ad4acjle 0x37b7ae6ce
    37b7ad4b2movsxd rax, esi
    37b7ad4b5mov rcx, [rsp+480]
    37b7ad4bdpcmpeqd x1, x1
    37b7ad4c1movaps x2, x3
    37b7ad4c4lea rdx, [rax-3]
    37b7ad4c8mov [rsp+136], rax
    37b7ad4d0mov rax, [rsp+88]
    37b7ad4d5shufps x2, x2, 0
    37b7ad4d9psrld x1, 31
    37b7ad4deadd rax, rdx
    37b7ad4e1add rax, [rsp+144]
    37b7ad4e9shl rdx, 2
    37b7ad4edlea r10, [rcx+4*rax]
    37b7ad4f1mov ecx, esi
    37b7ad4f3mov rax, [rsp+560]
    37b7ad4fblea r9, [rbx+rdx]
    37b7ad4ffshr ecx, 2
    37b7ad502neg rcx
    37b7ad505lea r8, [rax+rdx-4]
    37b7ad50aadd rdx, rax
    37b7ad50dxor eax, eax
    37b7ad50fshl rcx, 4
    37b7ad513nop
    37b7ad514nop [rax+rax]
    37b7ad51fnop [rax+rax]
    37b7ad52anop [rax+rax]
    37b7ad535nop [rax+rax]
    37b7ad540
    37b7ad540loc_37b7ad540:
    37b7ad540movups x0, [r10+rax]
    37b7ad545movdqu x4, [r8+rax]
    37b7ad54bshufps x0, x0, 27
    37b7ad54fmulps x0, x2
    37b7ad552shufps x0, x0, 27
    37b7ad556movups [r9+rax], x0
    37b7ad55bpshufd x0, x4, 27
    37b7ad560paddd x0, x1
    37b7ad564pshufd x0, x0, 27
    37b7ad569movups [rdx+rax], x0
    37b7ad56dsub rax, 16
    37b7ad571cmp rax, rcx
    37b7ad574jne 0x37b7ad540
    37b7ad576test sil, 3
    37b7ad57aje 0x37b7ad5bc
    37b7ad57cmov eax, esi
    37b7ad57eand eax, 3
    37b7ad581
    37b7ad581loc_37b7ad581:
    37b7ad581mov rcx, [rsp+480]
    37b7ad589mov rdx, [rsp+96]
    37b7ad58emov r8, [rsp+560]
    37b7ad596lea rcx, [rcx+4*rdx]
    37b7ad59a
    37b7ad59aloc_37b7ad59a:
    37b7ad59amovss x0, [rcx+4*rax]
    37b7ad59fmov edx, [r8+4*rax-4]
    37b7ad5a4mulss x0, x3
    37b7ad5a8add edx, 1
    37b7ad5abmov [r8+4*rax], edx
    37b7ad5afmovss [rbx+4*rax], x0
    37b7ad5b4sub rax, 1
    37b7ad5b8test eax, eax
    37b7ad5bajg 0x37b7ad59a
    37b7ad5bc
    37b7ad5bcloc_37b7ad5bc:
    37b7ad5bcmov rax, [rsp+136]
    37b7ad5c4movsxd rcx, edi
    37b7ad5c7mov [rsp+568], r11
    37b7ad5cfneg rcx
    37b7ad5d2lea r8, [4*rax]
    37b7ad5dashl rcx, 2
    37b7ad5delea rdx, [r8+rcx-4]
    37b7ad5e3add rcx, r8
    37b7ad5e6add rdx, r15
    37b7ad5e9add rcx, r15
    37b7ad5eccall memmove
    37b7ad5f1cmp r12l, [rsp+184]
    37b7ad5f9mov r11, [rsp+568]
    37b7ad601jl 0x37b7ad6de
    37b7ad607
    37b7ad607loc_37b7ad607:
    37b7ad607mov rax, [rsp+456]
    37b7ad60fmov r10, [rsp+88]
    37b7ad614add r10, [rsp+152]
    37b7ad61cmovss x2, [rax]
    37b7ad620mov eax, r12l
    37b7ad623sub eax, esi
    37b7ad625sub eax, 2
    37b7ad628cmp eax, 2
    37b7ad62bjbe 0x37b7ae6ed
    37b7ad631mov rdi, [rsp+136]
    37b7ad639mov edx, r12l
    37b7ad63cmov rcx, [rsp+480]
    37b7ad644movaps x1, x2
    37b7ad647sub edx, esi
    37b7ad649shufps x1, x1, 0
    37b7ad64dsub edx, 1
    37b7ad650lea rax, [r10+rdi+2]
    37b7ad655add rax, [rsp+144]
    37b7ad65dlea r9, [rcx+4*rax]
    37b7ad661mov ecx, edx
    37b7ad663lea r8, [rbx+4*rdi+4]
    37b7ad668xor eax, eax
    37b7ad66ashr ecx, 2
    37b7ad66dshl rcx, 4
    37b7ad671nop [rax]
    37b7ad675nop [rax+rax]
    37b7ad680
    37b7ad680loc_37b7ad680:
    37b7ad680movups x0, [r9+rax]
    37b7ad685mulps x0, x1
    37b7ad688movups [r8+rax], x0
    37b7ad68dadd rax, 16
    37b7ad691cmp rax, rcx
    37b7ad694jne 0x37b7ad680
    37b7ad696test dl, 3
    37b7ad699je 0x37b7ad6d4
    37b7ad69band edx, -4
    37b7ad69eadd edx, [rsp+184]
    37b7ad6a5
    37b7ad6a5loc_37b7ad6a5:
    37b7ad6a5mov rax, [rsp+480]
    37b7ad6adadd r10, [rsp+144]
    37b7ad6b5movsxd rdx, edx
    37b7ad6b8lea rax, [rax+4*r10]
    37b7ad6bc
    37b7ad6bcloc_37b7ad6bc:
    37b7ad6bcmovss x0, [rax+4*rdx]
    37b7ad6c1mulss x0, x2
    37b7ad6c5movss [rbx+4*rdx-4], x0
    37b7ad6cbadd rdx, 1
    37b7ad6cfcmp r12l, edx
    37b7ad6d2jge 0x37b7ad6bc
    37b7ad6d4
    37b7ad6d4loc_37b7ad6d4:
    37b7ad6d4cmp [rsp+84], 1
    37b7ad6d9jle 0x37b7ad736
    37b7ad6dblea edi, [rsi-1]
    37b7ad6de
    37b7ad6deloc_37b7ad6de:
    37b7ad6demov edi, edi
    37b7ad6e0lea rax, [r11+4]
    37b7ad6e4lea rdx, [r11+4*rdi+8]
    37b7ad6e9mov rcx, rdx
    37b7ad6ecsub rcx, rax
    37b7ad6efand ecx, 4
    37b7ad6f2je 0x37b7ad720
    37b7ad6f4mov [rax], 1
    37b7ad6falea rax, [r11+8]
    37b7ad6fecmp rax, rdx
    37b7ad701je 0x37b7ad736
    37b7ad703nop [rax]
    37b7ad70anop [rax+rax]
    37b7ad715nop [rax+rax]
    37b7ad720
    37b7ad720loc_37b7ad720:
    37b7ad720mov [rax], 1
    37b7ad726add rax, 8
    37b7ad72amov [rax-4], 1
    37b7ad731cmp rax, rdx
    37b7ad734jne 0x37b7ad720
    37b7ad736
    37b7ad736loc_37b7ad736:
    37b7ad736mov esi, [rsp+184]
    37b7ad73dcmp [rsp+72], esi
    37b7ad741jl 0x37b7ad848
    37b7ad747mov rsi, [rsp+136]
    37b7ad74flea edx, [r14-1]
    37b7ad753mov rcx, rdx
    37b7ad756add rdx, rsi
    37b7ad759lea r8, [4*rsi]
    37b7ad761lea rdx, [r11+4*rdx+4]
    37b7ad766lea rax, [r8+r11]
    37b7ad76amov r9, rdx
    37b7ad76dsub r9, rax
    37b7ad770and r9l, 4
    37b7ad774je 0x37b7ad7a0
    37b7ad776mov [rax+4], 2
    37b7ad77dadd rax, 4
    37b7ad781cmp rax, rdx
    37b7ad784je 0x37b7ad7b6
    37b7ad786nop [rax]
    37b7ad78anop [rax+rax]
    37b7ad795nop [rax+rax]
    37b7ad7a0
    37b7ad7a0loc_37b7ad7a0:
    37b7ad7a0mov [rax+4], 2
    37b7ad7a7add rax, 8
    37b7ad7abmov [rax], 2
    37b7ad7b1cmp rax, rdx
    37b7ad7b4jne 0x37b7ad7a0
    37b7ad7b6
    37b7ad7b6loc_37b7ad7b6:
    37b7ad7b6cmp ecx, 2
    37b7ad7b9jbe 0x37b7ae6e1
    37b7ad7bfmov rax, [rsp+560]
    37b7ad7c7mov edx, r14l
    37b7ad7camovd x5, [rsp+84]
    37b7ad7d0shr edx, 2
    37b7ad7d3lea rax, [rax+r8+4]
    37b7ad7d8shl rdx, 4
    37b7ad7dcpshufd x1, x5, 0
    37b7ad7e1add rdx, rax
    37b7ad7e4nop [rax+rax]
    37b7ad7eanop [rax+rax]
    37b7ad7f5nop [rax+rax]
    37b7ad800
    37b7ad800loc_37b7ad800:
    37b7ad800movdqu x0, [rax]
    37b7ad804add rax, 16
    37b7ad808paddd x0, x1
    37b7ad80cmovups [rax-16], x0
    37b7ad810cmp rax, rdx
    37b7ad813jne 0x37b7ad800
    37b7ad815test r14b, 3
    37b7ad819je 0x37b7ad848
    37b7ad81bmov eax, r14l
    37b7ad81eand eax, -4
    37b7ad821add eax, [rsp+184]
    37b7ad828
    37b7ad828loc_37b7ad828:
    37b7ad828mov r8l, [rsp+84]
    37b7ad82dmov rdx, [rsp+560]
    37b7ad835cdqe
    37b7ad837
    37b7ad837loc_37b7ad837:
    37b7ad837add [rdx+4*rax-4], r8l
    37b7ad83cadd rax, 1
    37b7ad840lea ecx, [rax-1]
    37b7ad843cmp r13l, ecx
    37b7ad846jge 0x37b7ad837
    37b7ad848
    37b7ad848loc_37b7ad848:
    37b7ad848movsxd rdi, r13l
    37b7ad84blea esi, [r13+2]
    37b7ad84fcmp [rsp+72], 1
    37b7ad854jle 0x37b7ae5d0
    37b7ad85amov rcx, [rsp+496]
    37b7ad862mov r8, [rsp+504]
    37b7ad86amov r12l, r13l
    37b7ad86dxor eax, eax
    37b7ad86fmov r9, [rsp+560]
    37b7ad877nop [rax+rax]
    37b7ad880
    37b7ad880loc_37b7ad880:
    37b7ad880movsxd rdx, [r9+4*rax+4]
    37b7ad885sub rdx, 1
    37b7ad889movss x0, [r15+4*rdx]
    37b7ad88fmovss [rcx+4*rax+4], x0
    37b7ad895movss x0, [rbx+4*rdx]
    37b7ad89amov edx, [r11+4*rdx]
    37b7ad89emovss [r8+4*rax+4], x0
    37b7ad8a5mov [rbp+4*rax+4], edx
    37b7ad8a9add rax, 1
    37b7ad8adcmp rax, r12
    37b7ad8b0jne 0x37b7ad880
    37b7ad8b2mov rax, [rsp+544]
    37b7ad8balea rcx, [rip+0xc8b147]
    37b7ad8c1mov rdx, [rsp+408]
    37b7ad8c9mov [rsp+568], r11
    37b7ad8d1mov [rsp+32], rcx
    37b7ad8d6mov r9, rcx
    37b7ad8d9add rax, 4
    37b7ad8ddmov [rsp+96], rcx
    37b7ad8e2mov rcx, [rsp+400]
    37b7ad8eamov [rsp+40], rax
    37b7ad8efmov rax, [rsp+496]
    37b7ad8f7lea r8, [rax+4]
    37b7ad8fbcall slamrg_
    37b7ad900mov rcx, [rsp+496]
    37b7ad908xor eax, eax
    37b7ad90amov r8, [rsp+504]
    37b7ad912mov r9, [rsp+544]
    37b7ad91amov r11, [rsp+568]
    37b7ad922nop [rax+rax]
    37b7ad92anop [rax+rax]
    37b7ad935nop [rax+rax]
    37b7ad940
    37b7ad940loc_37b7ad940:
    37b7ad940movsxd rdx, [r9+4*rax+4]
    37b7ad945movss x0, [rcx+4*rdx]
    37b7ad94amovss [r15+4*rax+4], x0
    37b7ad951movss x0, [r8+4*rdx]
    37b7ad957mov edx, [rbp+4*rdx]
    37b7ad95bmovss [rbx+4*rax+4], x0
    37b7ad961mov [r11+4*rax+4], edx
    37b7ad966add rax, 1
    37b7ad96acmp rax, r12
    37b7ad96djne 0x37b7ad940
    37b7ad96fmov edx, 7
    37b7ad974lea rcx, [rip+0xc8b091]
    37b7ad97bmov [rsp+568], r11
    37b7ad983call slamch_
    37b7ad988mov rax, [rsp+448]
    37b7ad990movss x7, [rip+0xc8b088]
    37b7ad998mulss x0, [rip+0xc8b090]
    37b7ad9a0mov r11, [rsp+568]
    37b7ad9a8movss x6, [rax]
    37b7ad9acmov rax, [rsp+456]
    37b7ad9b4movss x1, [rax]
    37b7ad9b8andps x6, x7
    37b7ad9bbmov rax, [rsp+424]
    37b7ad9c3andps x1, x7
    37b7ad9c6mov [rax], 1
    37b7ad9ccmov rax, [rsp+536]
    37b7ad9d4maxss x6, x1
    37b7ad9d8movss x1, [r15+4*rdi]
    37b7ad9delea rdx, [rax+4*rdi]
    37b7ad9e2mov eax, 2
    37b7ad9e7andps x1, x7
    37b7ad9eamaxss x6, x1
    37b7ad9eemulss x6, x0
    37b7ad9f2jmp 0x37b7ada1f
    37b7ad9f4
    37b7ad9f4 90 66 66 2e 0f 1f 84 00 00 00 00 00 .ff.........
    37b7ada00
    37b7ada00loc_37b7ada00:
    37b7ada00sub esi, 1
    37b7ada03mov [rdx], eax
    37b7ada05mov [r11+4*rax-4], 4
    37b7ada0ecmp esi, 2
    37b7ada11je 0x37b7ae550
    37b7ada17add rax, 1
    37b7ada1bsub rdx, 4
    37b7ada1f
    37b7ada1floc_37b7ada1f:
    37b7ada1fmovss x1, [rbx+4*rax-4]
    37b7ada25andps x1, x7
    37b7ada28comiss x6, x1
    37b7ada2bjae 0x37b7ada00
    37b7ada2dmov [rsp+128], eax
    37b7ada34
    37b7ada34loc_37b7ada34:
    37b7ada34mov eax, [rsp+128]
    37b7ada3bcmp r13l, eax
    37b7ada3ejl 0x37b7ae6f9
    37b7ada44movsxd rdx, [rsp+128]
    37b7ada4clea rcx, [rsp+224]
    37b7ada54mov rax, [rsp+544]
    37b7ada5cmov [rsp+432], r15
    37b7ada64mov [rsp+88], rcx
    37b7ada69lea rcx, [rsp+220]
    37b7ada71lea r13, [4*rdx]
    37b7ada79mov [rsp+104], rcx
    37b7ada7elea edi, [rdx+1]
    37b7ada81add rax, r13
    37b7ada84lea r14, [r11+r13]
    37b7ada88lea r12, [rbx+r13]
    37b7ada8cmov [rsp+568], r11
    37b7ada94lea rcx, [rsp+228]
    37b7ada9cadd r13, r15
    37b7ada9fmov [rsp+552], rbp
    37b7adaa7mov r11, [rsp+536]
    37b7adaafmov [rsp+160], rcx
    37b7adab7mov r15, rax
    37b7adabajmp 0x37b7adaf4
    37b7adabc
    37b7adabc 0f 1f 40 00 ..@.
    37b7adac0
    37b7adac0loc_37b7adac0:
    37b7adac0sub esi, 1
    37b7adac3mov [r14], 4
    37b7adacamovsxd rax, esi
    37b7adacdmov [r11+4*rax-4], edi
    37b7adad2
    37b7adad2loc_37b7adad2:
    37b7adad2add edi, 1
    37b7adad5add r12, 4
    37b7adad9add r14, 4
    37b7adaddadd r15, 4
    37b7adae1mov eax, [rsp+224]
    37b7adae8add r13, 4
    37b7adaeccmp eax, edi
    37b7adaeejl 0x37b7add50
    37b7adaf4
    37b7adaf4loc_37b7adaf4:
    37b7adaf4movss x1, [r12]
    37b7adafamovaps x0, x1
    37b7adafdandps x0, x7
    37b7adb00comiss x6, x0
    37b7adb03jae 0x37b7adac0
    37b7adb05mov rax, [rsp+432]
    37b7adb0dlea rbp, [rdx-1]
    37b7adb11movss x0, [r13]
    37b7adb17movss x3, [rbx+4*rbp]
    37b7adb1cmovss x2, [rax+4*rbp]
    37b7adb21subss x0, x2
    37b7adb25andps x0, x7
    37b7adb28comiss x6, x0
    37b7adb2bjae 0x37b7adb80
    37b7adb2dmov rax, [rsp+424]
    37b7adb35mov rcx, [rsp+424]
    37b7adb3dmovsxd rax, [rax]
    37b7adb40lea edx, [rax+1]
    37b7adb43mov [rcx], edx
    37b7adb45mov rcx, [rsp+504]
    37b7adb4dmovsxd rdx, edi
    37b7adb50movss [rcx+4*rax], x3
    37b7adb55mov rcx, [rsp+496]
    37b7adb5dmovss [rcx+4*rax], x2
    37b7adb62mov ecx, [rsp+128]
    37b7adb69mov [rsp+128], edi
    37b7adb70mov [r11+4*rax], ecx
    37b7adb74jmp 0x37b7adad2
    37b7adb79
    37b7adb79 0f 1f 80 00 00 00 00 .......
    37b7adb80
    37b7adb80loc_37b7adb80:
    37b7adb80lea rax, [rsp+216]
    37b7adb88mov rdx, [rsp+160]
    37b7adb90movss [rsp+228], x3
    37b7adb99mov rcx, rax
    37b7adb9cmov [rsp+72], rax
    37b7adba1mov [rsp+536], r11
    37b7adba9movss [rsp+216], x1
    37b7adbb2call slapy2_
    37b7adbb7mov r9, [rsp+544]
    37b7adbbfxor edx, edx
    37b7adbc1mov r11l, [rsp+84]
    37b7adbc6movaps x1, x0
    37b7adbc9lea r8, [rip+0xc8ae38]
    37b7adbd0movss x0, [rsp+216]
    37b7adbd9movsxd rax, [r9+4*rbp]
    37b7adbddmov r9, [rsp+560]
    37b7adbe5movss [r12], x1
    37b7adbebdivss x0, x1
    37b7adbefmov [rbx+4*rbp], 0
    37b7adbf6mov r10l, [r9+4*rax]
    37b7adbfamovsxd rax, [r15]
    37b7adbfdcmp r11l, r10l
    37b7adc00mov eax, [r9+4*rax]
    37b7adc04mov r9, [rsp+112]
    37b7adc09setge dl
    37b7adc0csub r10l, edx
    37b7adc0fxor edx, edx
    37b7adc11cmp r11l, eax
    37b7adc14mov r11, [rsp+120]
    37b7adc19setge dl
    37b7adc1cmovsxd r10, r10l
    37b7adc1fsub eax, edx
    37b7adc21mov rdx, r9
    37b7adc24mov [rsp+192], r10
    37b7adc2ccdqe
    37b7adc2eimul r9, r10
    37b7adc32mov r10, [rsp+160]
    37b7adc3aimul rdx, rax
    37b7adc3emov [rsp+200], rax
    37b7adc46mov rax, [rsp+72]
    37b7adc4bmov [rsp+48], r10
    37b7adc50mov [rsp+40], rax
    37b7adc55lea rcx, [r11+rdx+1]
    37b7adc5alea rdx, [r9+r11+1]
    37b7adc5fmov r11, [rsp+96]
    37b7adc64mov r9, [rsp+464]
    37b7adc6cmov [rsp+32], r11
    37b7adc71lea rdx, [r9+4*rdx]
    37b7adc75lea r9, [r9+4*rcx]
    37b7adc79mov rcx, [rsp+88]
    37b7adc7emovss [rsp+216], x0
    37b7adc87movss x0, [rsp+228]
    37b7adc90divss x0, x1
    37b7adc94xorps x0, [rip+0xc8ada5]
    37b7adc9bmovss [rsp+228], x0
    37b7adca4call srot_
    37b7adca9mov rax, [rsp+72]
    37b7adcaemov rcx, [rsp+480]
    37b7adcb6mov r10, [rsp+192]
    37b7adcbemov r8, [rsp+488]
    37b7adcc6mov [rsp+40], rax
    37b7adccbmov rax, [rsp+488]
    37b7adcd3lea rdx, [rcx+4*r10-4]
    37b7adcd8mov r10, [rsp+160]
    37b7adce0mov [rsp+32], rax
    37b7adce5mov rax, [rsp+200]
    37b7adcedmov [rsp+48], r10
    37b7adcf2lea r9, [rcx+4*rax-4]
    37b7adcf7mov rcx, [rsp+104]
    37b7adcfccall srot_
    37b7add01mov rax, [rsp+568]
    37b7add09mov r11, [rsp+536]
    37b7add11mov eax, [rax+4*rbp]
    37b7add14cmp [r14], eax
    37b7add17je 0x37b7add20
    37b7add19mov [r14], 3
    37b7add20
    37b7add20loc_37b7add20:
    37b7add20mov rax, [rsp+568]
    37b7add28mov ecx, [rsp+128]
    37b7add2fsub esi, 1
    37b7add32movsxd rdx, edi
    37b7add35mov [rsp+128], edi
    37b7add3cmov [rax+4*rbp], 4
    37b7add43movsxd rax, esi
    37b7add46mov [r11+4*rax-4], ecx
    37b7add4bjmp 0x37b7adad2
    37b7add50
    37b7add50loc_37b7add50:
    37b7add50mov [rsp+72], eax
    37b7add54mov r15, [rsp+432]
    37b7add5cmov rbp, [rsp+552]
    37b7add64mov r11, [rsp+568]
    37b7add6c
    37b7add6cloc_37b7add6c:
    37b7add6cmov rax, [rsp+424]
    37b7add74mov rsi, [rsp+424]
    37b7add7csub rdx, 1
    37b7add80mov [rsp+240], 0
    37b7add8cmovss x0, [rbx+4*rdx]
    37b7add91mov edi, [rsp+128]
    37b7add98mov [rsp+248], 0
    37b7adda4movsxd rax, [rax]
    37b7adda7lea ecx, [rax+1]
    37b7addaamov [rsi], ecx
    37b7addacmov rsi, [rsp+504]
    37b7addb4movss [rsi+4*rax], x0
    37b7addb9mov rsi, [rsp+496]
    37b7addc1movss x0, [r15+4*rdx]
    37b7addc7movss [rsi+4*rax], x0
    37b7addccmov rsi, [rsp+536]
    37b7addd4mov [rsi+4*rax], edi
    37b7addd7cmp [rsp+72], 1
    37b7adddcjle 0x37b7ae690
    37b7adde2
    37b7adde2loc_37b7adde2:
    37b7adde2mov esi, [rsp+72]
    37b7adde6lea rax, [r11+4]
    37b7addealea r8l, [rsi-2]
    37b7addeelea rcx, [r11+4*r8+8]
    37b7addf3nop
    37b7addf5nop [rax+rax]
    37b7ade00
    37b7ade00loc_37b7ade00:
    37b7ade00movsxd rdx, [rax]
    37b7ade03add rax, 4
    37b7ade07add [rsp+4*rdx+236], 1
    37b7ade0fcmp rax, rcx
    37b7ade12jne 0x37b7ade00
    37b7ade14mov eax, [rsp+240]
    37b7ade1bmov r10, [rsp+536]
    37b7ade23add r8, 3
    37b7ade27mov [rsp+256], 2
    37b7ade32movdqu x8, [rsp+240]
    37b7ade3cadd eax, 2
    37b7ade3fmov [rsp+260], eax
    37b7ade46add eax, [rsp+244]
    37b7ade4dmov [rsp+264], eax
    37b7ade54add eax, [rsp+248]
    37b7ade5bmov [rsp+268], eax
    37b7ade62mov eax, 2
    37b7ade67nop [rax]
    37b7ade6anop [rax+rax]
    37b7ade75nop [rax+rax]
    37b7ade80
    37b7ade80loc_37b7ade80:
    37b7ade80movsxd rdx, [r10+4*rax-4]
    37b7ade85movsxd rdx, [r11+4*rdx-4]
    37b7ade8asub rdx, 1
    37b7ade8emovsxd r9, [rsp+4*rdx+256]
    37b7ade96mov rcx, r9
    37b7ade99mov [rbp+4*r9-4], eax
    37b7ade9eadd rax, 1
    37b7adea2add ecx, 1
    37b7adea5mov [rsp+4*rdx+256], ecx
    37b7adeaccmp r8, rax
    37b7adeafjne 0x37b7ade80
    37b7adeb1mov rax, [rsp+168]
    37b7adeb9mov rsi, [rsp+176]
    37b7adec1mov [rsp+568], r11
    37b7adec9mov [rsp+552], rbp
    37b7aded1mov r13, [rsp+552]
    37b7aded9mov ebp, 2
    37b7adedelea rdi, [4*rax]
    37b7adee6lea rax, [rsi+2*rax]
    37b7adeeamov rsi, [rsp+504]
    37b7adef2mov [rsp+440], rbx
    37b7adefamov rbx, [rsp+536]
    37b7adf02lea r12, [rsi+4]
    37b7adf06lea r14, [r12+4*rax]
    37b7adf0amov rax, [rsp+520]
    37b7adf12mov [rsp+128], r12
    37b7adf1amov r12, r14
    37b7adf1dlea rsi, [rax+4]
    37b7adf21nop [rax]
    37b7adf25nop [rax+rax]
    37b7adf30
    37b7adf30loc_37b7adf30:
    37b7adf30movsxd rax, [rbx+4*rbp-4]
    37b7adf35mov rcx, [rsp+544]
    37b7adf3dxor edx, edx
    37b7adf3fmov r9, r12
    37b7adf42lea r8, [rip+0xc8aabf]
    37b7adf49movss x0, [r15+4*rax-4]
    37b7adf50mov rax, [rsp+496]
    37b7adf58movss [rax+4*rbp-4], x0
    37b7adf5emovsxd rax, [r13+4*rbp-4]
    37b7adf63movsxd rax, [rbx+4*rax-4]
    37b7adf68movsxd rax, [rcx+4*rax-4]
    37b7adf6dmov rcx, [rsp+560]
    37b7adf75mov eax, [rcx+4*rax]
    37b7adf78cmp [rsp+84], eax
    37b7adf7csetge dl
    37b7adf7fmov rcx, [rsp+88]
    37b7adf84add rbp, 1
    37b7adf88add r12, rdi
    37b7adf8bsub eax, edx
    37b7adf8dmov rdx, [rsp+112]
    37b7adf92movsxd r14, eax
    37b7adf95mov rax, [rsp+120]
    37b7adf9aimul rdx, r14
    37b7adf9elea rdx, [rax+rdx+1]
    37b7adfa3mov rax, [rsp+464]
    37b7adfablea rdx, [rax+4*rdx]
    37b7adfafmov rax, [rsp+96]
    37b7adfb4mov [rsp+32], rax
    37b7adfb9call scopy_
    37b7adfbemov rax, [rsp+480]
    37b7adfc6mov rcx, [rsp+104]
    37b7adfcbmov r9, rsi
    37b7adfcemov r8, [rsp+488]
    37b7adfd6add rsi, 4
    37b7adfdalea rdx, [rax+4*r14-4]
    37b7adfdfmov rax, [rsp+528]
    37b7adfe7mov [rsp+32], rax
    37b7adfeccall scopy_
    37b7adff1cmp [rsp+72], ebp
    37b7adff5jge 0x37b7adf30
    37b7adffbmov eax, [rsp+224]
    37b7ae002mov rbx, [rsp+440]
    37b7ae00amov r11, [rsp+568]
    37b7ae012mov r12, [rsp+128]
    37b7ae01amov [rsp+72], eax
    37b7ae01e
    37b7ae01eloc_37b7ae01e:
    37b7ae01emovss x1, [rip+0xc8aa2a]
    37b7ae026mov rax, [rsp+496]
    37b7ae02emulss x1, x6
    37b7ae032movss x0, [rax+4]
    37b7ae037mov [rax], 0
    37b7ae03dandps x0, x7
    37b7ae040comiss x1, x0
    37b7ae043jb 0x37b7ae04a
    37b7ae045movss [rax+4], x1
    37b7ae04a
    37b7ae04aloc_37b7ae04a:
    37b7ae04amov esi, [rsp+220]
    37b7ae051cmp esi, [rsp+72]
    37b7ae055jg 0x37b7ae398
    37b7ae05bmovss x0, [rsp+232]
    37b7ae064andps x7, x0
    37b7ae067comiss x6, x7
    37b7ae06ajb 0x37b7ae380
    37b7ae070movss x7, [rip+0xc8a9dc]
    37b7ae078movss [rbx], x6
    37b7ae07c
    37b7ae07cloc_37b7ae07c:
    37b7ae07cmov rax, [rsp+424]
    37b7ae084lea r9, [rbx+4]
    37b7ae088mov rdx, r12
    37b7ae08blea rbp, [rsp+236]
    37b7ae093lea r8, [rip+0xc8a96e]
    37b7ae09amov rcx, rbp
    37b7ae09dmov [rsp+568], r11
    37b7ae0a5mov eax, [rax]
    37b7ae0a7mov [rsp+72], eax
    37b7ae0absub eax, 1
    37b7ae0aemov [rsp+236], eax
    37b7ae0b5mov rax, [rsp+96]
    37b7ae0bamov [rsp+32], rax
    37b7ae0bfcall scopy_
    37b7ae0c4mov rax, [rsp+512]
    37b7ae0ccmov [rsp+56], 1
    37b7ae0d5lea r8, [rip+0xc8a92c]
    37b7ae0dcmov rdx, [rsp+88]
    37b7ae0e1lea rcx, [rip+0xc8a930]
    37b7ae0e8mov [rsp+48], rax
    37b7ae0edmov rax, [rsp+504]
    37b7ae0f5mov [rsp+40], rax
    37b7ae0falea rax, [rip+0xc8a913]
    37b7ae101mov [rsp+32], rax
    37b7ae106mov r9, rax
    37b7ae109call slaset_
    37b7ae10emov rbx, [rsp+136]
    37b7ae116mov rax, [rsp+504]
    37b7ae11emov r11, [rsp+568]
    37b7ae126movss [rax+4*rbx], x7
    37b7ae12bmov ebx, [rsp+224]
    37b7ae132cmp esi, ebx
    37b7ae134jle 0x37b7ae318
    37b7ae13amovsxd r8, [rsp+188]
    37b7ae142xor eax, eax
    37b7ae144test r8, r8
    37b7ae147cmovs r8, rax
    37b7ae14bmov eax, [rsp+84]
    37b7ae14ftest eax, eax
    37b7ae151jle 0x37b7ae71b
    37b7ae157mov rax, [rsp+152]
    37b7ae15fmov r14, [rsp+136]
    37b7ae167movsxd r13, esi
    37b7ae16alea r10, [4*r8]
    37b7ae172mov rcx, [rsp+520]
    37b7ae17amov r9, r13
    37b7ae17dmovss x2, [rsp+228]
    37b7ae186mov edx, 1
    37b7ae18blea rdi, [4*rax]
    37b7ae193mov rax, [rsp+480]
    37b7ae19bmov r12l, [rsp+184]
    37b7ae1a3sub r9, r14
    37b7ae1a6movss x1, [rsp+216]
    37b7ae1afsub rcx, 4
    37b7ae1b3lea rax, [rax+4*r14]
    37b7ae1b7nop [rax+rax]
    37b7ae1c0
    37b7ae1c0loc_37b7ae1c0:
    37b7ae1c0movss x0, [rax]
    37b7ae1c4add edx, 1
    37b7ae1c7mulss x0, x2
    37b7ae1cbxorps x0, [rip+0xc8a86e]
    37b7ae1d2movss [rax+4*r9-4], x0
    37b7ae1d9movss x0, [rax]
    37b7ae1ddadd rax, rdi
    37b7ae1e0mulss x0, x1
    37b7ae1e4movss [rcx+4], x0
    37b7ae1e9add rcx, r10
    37b7ae1eccmp r12l, edx
    37b7ae1efjne 0x37b7ae1c0
    37b7ae1f1
    37b7ae1f1loc_37b7ae1f1:
    37b7ae1f1cmp esi, [rsp+184]
    37b7ae1f8jl 0x37b7ae2a8
    37b7ae1femovsxd r9, [rsp+184]
    37b7ae206mov rax, [rsp+152]
    37b7ae20eadd esi, 1
    37b7ae211mov rdx, [rsp+480]
    37b7ae219mov r10, [rsp+520]
    37b7ae221mov rdi, r9
    37b7ae224lea rcx, [4*rax]
    37b7ae22cmovss x3, [rsp+228]
    37b7ae235movss x2, [rsp+216]
    37b7ae23eimul rax, r9
    37b7ae242add rax, [rsp+144]
    37b7ae24aimul r9, r8
    37b7ae24eadd rax, r13
    37b7ae251lea rax, [rdx+4*rax]
    37b7ae255lea rdx, [4*r8]
    37b7ae25dnot r8
    37b7ae260add r9, r8
    37b7ae263lea r8, [r10+4*r9]
    37b7ae267mov r9l, edi
    37b7ae26anop [rax+rax]
    37b7ae275nop [rax+rax]
    37b7ae280
    37b7ae280loc_37b7ae280:
    37b7ae280movss x0, [rax]
    37b7ae284add r9l, 1
    37b7ae288movaps x1, x0
    37b7ae28bmulss x1, x3
    37b7ae28fmulss x0, x2
    37b7ae293movss [r8+4], x1
    37b7ae299add r8, rdx
    37b7ae29cmovss [rax], x0
    37b7ae2a0add rax, rcx
    37b7ae2a3cmp r9l, esi
    37b7ae2a6jne 0x37b7ae280
    37b7ae2a8
    37b7ae2a8loc_37b7ae2a8:
    37b7ae2a8mov rsi, [rsp+528]
    37b7ae2b0mov rdx, [rsp+480]
    37b7ae2b8lea rax, [4*r13-4]
    37b7ae2c0mov [rsp+568], r11
    37b7ae2c8mov r8, [rsp+488]
    37b7ae2d0mov rcx, [rsp+104]
    37b7ae2d5mov [rsp+32], rsi
    37b7ae2damov rsi, [rsp+520]
    37b7ae2e2add rdx, rax
    37b7ae2e5lea r9, [rsi+rax]
    37b7ae2e9call scopy_
    37b7ae2eemov r11, [rsp+568]
    37b7ae2f6
    37b7ae2f6loc_37b7ae2f6:
    37b7ae2f6mov rax, [rsp+424]
    37b7ae2femovsxd rax, [rax]
    37b7ae301cmp ebx, eax
    37b7ae303jg 0x37b7ae3f8
    37b7ae309
    37b7ae309loc_37b7ae309:
    37b7ae309movups [r11], x8
    37b7ae30djmp 0x37b7ad354
    37b7ae312
    37b7ae312 66 0f 1f 44 00 00 f..D..
    37b7ae318
    37b7ae318loc_37b7ae318:
    37b7ae318mov rax, [rsp+480]
    37b7ae320mov rsi, [rsp+136]
    37b7ae328mov [rsp+568], r11
    37b7ae330mov r9, [rsp+520]
    37b7ae338mov r8, [rsp+488]
    37b7ae340lea rdx, [rax+4*rsi]
    37b7ae344mov rax, [rsp+528]
    37b7ae34cmov rcx, [rsp+104]
    37b7ae351mov [rsp+32], rax
    37b7ae356call scopy_
    37b7ae35bmov esi, [rsp+220]
    37b7ae362mov r11, [rsp+568]
    37b7ae36acmp ebx, esi
    37b7ae36cjge 0x37b7ae2f6
    37b7ae36emovsxd r13, esi
    37b7ae371jmp 0x37b7ae2a8
    37b7ae376
    37b7ae376 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b7ae380
    37b7ae380loc_37b7ae380:
    37b7ae380movss x7, [rip+0xc8a6cc]
    37b7ae388movss [rbx], x0
    37b7ae38cjmp 0x37b7ae07c
    37b7ae391
    37b7ae391 0f 1f 80 00 00 00 00 .......
    37b7ae398
    37b7ae398loc_37b7ae398:
    37b7ae398movsxd r13, esi
    37b7ae39blea rcx, [rsp+232]
    37b7ae3a3mov [rsp+568], r11
    37b7ae3absub r13, 1
    37b7ae3aflea rdx, [rbx+4*r13]
    37b7ae3b3call slapy2_
    37b7ae3b8mov r11, [rsp+568]
    37b7ae3c0comiss x6, x0
    37b7ae3c3movss [rbx], x0
    37b7ae3c7jb 0x37b7ae590
    37b7ae3cdmovss x7, [rip+0xc8a67f]
    37b7ae3d5movss [rbx], x6
    37b7ae3d9mov [rsp+228], 0
    37b7ae3e4movss [rsp+216], x7
    37b7ae3edjmp 0x37b7ae07c
    37b7ae3f2
    37b7ae3f2 66 0f 1f 44 00 00 f..D..
    37b7ae3f8
    37b7ae3f8loc_37b7ae3f8:
    37b7ae3f8mov rsi, [rsp+96]
    37b7ae3fdmov edx, ebx
    37b7ae3fflea r9, [r15+4*rax]
    37b7ae403mov rcx, rbp
    37b7ae406sub edx, eax
    37b7ae408lea r8, [rip+0xc8a5f9]
    37b7ae40fmov [rsp+568], r11
    37b7ae417mov [rsp+32], rsi
    37b7ae41cmov [rsp+236], edx
    37b7ae423lea rdx, [4*rax]
    37b7ae42badd rdx, [rsp+496]
    37b7ae433call scopy_
    37b7ae438mov rax, [rsp+424]
    37b7ae440mov r8, rbp
    37b7ae443mov rdx, [rsp+168]
    37b7ae44blea rcx, [rip+0xc8a5c6]
    37b7ae452mov eax, [rax]
    37b7ae454mov [rsp+56], 1
    37b7ae45dsub ebx, eax
    37b7ae45fadd eax, 1
    37b7ae462cdqe
    37b7ae464mov [rsp+236], ebx
    37b7ae46bmov rbx, [rsp+176]
    37b7ae473imul rdx, rax
    37b7ae477imul rax, [rsp+112]
    37b7ae47dlea rdx, [rbx+rdx+1]
    37b7ae482mov rbx, [rsp+472]
    37b7ae48amov [rsp+48], rbx
    37b7ae48fmov rbx, [rsp+120]
    37b7ae494lea rax, [rbx+rax+1]
    37b7ae499mov rbx, [rsp+464]
    37b7ae4a1lea rax, [rbx+4*rax]
    37b7ae4a5mov [rsp+40], rax
    37b7ae4aamov rax, [rsp+512]
    37b7ae4b2mov [rsp+32], rax
    37b7ae4b7mov rax, [rsp+504]
    37b7ae4bflea r9, [rax+4*rdx]
    37b7ae4c3mov rdx, [rsp+88]
    37b7ae4c8call slacpy_
    37b7ae4cdmov rax, [rsp+424]
    37b7ae4d5mov r8, [rsp+104]
    37b7ae4dalea rcx, [rip+0xc8a537]
    37b7ae4e1mov rbx, [rsp+488]
    37b7ae4e9mov edx, [rsp+224]
    37b7ae4f0movsxd rax, [rax]
    37b7ae4f3mov [rsp+56], 1
    37b7ae4fcmov [rsp+48], rbx
    37b7ae501mov rbx, [rsp+528]
    37b7ae509sub edx, eax
    37b7ae50bmov [rsp+236], edx
    37b7ae512lea rdx, [4*rax]
    37b7ae51aadd rdx, [rsp+480]
    37b7ae522mov [rsp+40], rdx
    37b7ae527mov rdx, rbp
    37b7ae52amov [rsp+32], rbx
    37b7ae52fmov rbx, [rsp+520]
    37b7ae537lea r9, [rbx+4*rax]
    37b7ae53bcall slacpy_
    37b7ae540mov r11, [rsp+568]
    37b7ae548jmp 0x37b7ae309
    37b7ae54d
    37b7ae54d 0f 1f 00 ...
    37b7ae550
    37b7ae550loc_37b7ae550:
    37b7ae550lea rax, [rsp+224]
    37b7ae558mov [rsp+240], 0
    37b7ae564mov [rsp+88], rax
    37b7ae569lea rax, [rsp+220]
    37b7ae571mov [rsp+248], 0
    37b7ae57dmov [rsp+104], rax
    37b7ae582jmp 0x37b7adde2
    37b7ae587
    37b7ae587 66 0f 1f 84 00 00 00 00 00 f........
    37b7ae590
    37b7ae590loc_37b7ae590:
    37b7ae590movss x1, [rsp+232]
    37b7ae599movss x7, [rip+0xc8a4b3]
    37b7ae5a1divss x1, x0
    37b7ae5a5movss [rsp+216], x1
    37b7ae5aemovss x1, [rbx+4*r13]
    37b7ae5b4divss x1, x0
    37b7ae5b8movss [rsp+228], x1
    37b7ae5c1jmp 0x37b7ae07c
    37b7ae5c6
    37b7ae5c6 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b7ae5d0
    37b7ae5d0loc_37b7ae5d0:
    37b7ae5d0mov rax, [rsp+544]
    37b7ae5d8lea rcx, [rip+0xc8a429]
    37b7ae5dfmov rdx, [rsp+408]
    37b7ae5e7mov [rsp+568], r11
    37b7ae5efmov [rsp+32], rcx
    37b7ae5f4mov r9, rcx
    37b7ae5f7add rax, 4
    37b7ae5fbmov [rsp+96], rcx
    37b7ae600mov rcx, [rsp+400]
    37b7ae608mov [rsp+40], rax
    37b7ae60dmov rax, [rsp+496]
    37b7ae615lea r8, [rax+4]
    37b7ae619call slamrg_
    37b7ae61emov edx, 7
    37b7ae623lea rcx, [rip+0xc8a3e2]
    37b7ae62acall slamch_
    37b7ae62fmov rax, [rsp+448]
    37b7ae637movss x7, [rip+0xc8a3e1]
    37b7ae63fmulss x0, [rip+0xc8a3e9]
    37b7ae647mov r11, [rsp+568]
    37b7ae64fmovss x6, [rax]
    37b7ae653mov rax, [rsp+456]
    37b7ae65bmovss x1, [rax]
    37b7ae65fandps x6, x7
    37b7ae662mov rax, [rsp+424]
    37b7ae66aandps x1, x7
    37b7ae66dmov [rax], 1
    37b7ae673maxss x6, x1
    37b7ae677movss x1, [r15+4*rdi]
    37b7ae67dandps x1, x7
    37b7ae680maxss x6, x1
    37b7ae684mulss x6, x0
    37b7ae688jmp 0x37b7ada34
    37b7ae68d
    37b7ae68d 0f 1f 00 ...
    37b7ae690
    37b7ae690loc_37b7ae690:
    37b7ae690mov rax, [rsp+504]
    37b7ae698pxor x8, x8
    37b7ae69dlea r12, [rax+4]
    37b7ae6a1jmp 0x37b7ae01e
    37b7ae6a6
    37b7ae6a6 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b7ae6b0
    37b7ae6b0loc_37b7ae6b0:
    37b7ae6b0movsxd rax, esi
    37b7ae6b3cmp r12l, [rsp+184]
    37b7ae6bbmov [rsp+136], rax
    37b7ae6c3jge 0x37b7ad607
    37b7ae6c9jmp 0x37b7ad736
    37b7ae6ce
    37b7ae6celoc_37b7ae6ce:
    37b7ae6cemovsxd rcx, esi
    37b7ae6d1movsxd rax, esi
    37b7ae6d4mov [rsp+136], rcx
    37b7ae6dcjmp 0x37b7ad581
    37b7ae6e1
    37b7ae6e1loc_37b7ae6e1:
    37b7ae6e1mov eax, [rsp+184]
    37b7ae6e8jmp 0x37b7ad828
    37b7ae6ed
    37b7ae6edloc_37b7ae6ed:
    37b7ae6edmov edx, [rsp+184]
    37b7ae6f4jmp 0x37b7ad6a5
    37b7ae6f9
    37b7ae6f9loc_37b7ae6f9:
    37b7ae6f9movsxd rdx, eax
    37b7ae6fclea rax, [rsp+224]
    37b7ae704mov [rsp+88], rax
    37b7ae709lea rax, [rsp+220]
    37b7ae711mov [rsp+104], rax
    37b7ae716jmp 0x37b7add6c
    37b7ae71b
    37b7ae71bloc_37b7ae71b:
    37b7ae71bmovsxd r13, esi
    37b7ae71ejmp 0x37b7ae1f1
    37b7ae723
    37b7ae723loc_37b7ae723:
    37b7ae723neg eax
    37b7ae725mov edx, eax
    37b7ae727jmp 0x37b7ad332
    37b7ad200
    37b7ae72c
    37b7ae72c // align 0x20
    37b7ae740
    37b7ae740slasd3_:
    37b7ae740push r15
    37b7ae742push r14
    37b7ae744push r13
    37b7ae746push r12
    37b7ae748push rbp
    37b7ae749push rdi
    37b7ae74apush rsi
    37b7ae74bpush rbx
    37b7ae74csub rsp, 296
    37b7ae753movups [rsp+272], x6
    37b7ae75bmov rax, [rsp+488]
    37b7ae763mov rsi, [rsp+416]
    37b7ae76bmov rbx, [rsp+520]
    37b7ae773mov r15, [rsp+400]
    37b7ae77bmov eax, [rax]
    37b7ae77dmov esi, [rsi]
    37b7ae77fmov [rbx], 0
    37b7ae785mov r14, [rsp+424]
    37b7ae78dmov [rsp+232], eax
    37b7ae794mov rax, [rsp+472]
    37b7ae79cmov [rsp+368], rcx
    37b7ae7a4mov r13, [rsp+464]
    37b7ae7acmov ecx, [rax]
    37b7ae7aemov rax, [rsp+456]
    37b7ae7b6mov [rsp+200], esi
    37b7ae7bdmov rsi, [rsp+368]
    37b7ae7c5mov [rsp+376], rdx
    37b7ae7cdmov eax, [rax]
    37b7ae7cfmov [rsp+384], r8
    37b7ae7d7mov edi, [rsi]
    37b7ae7d9mov [rsp+392], r9
    37b7ae7e1mov [rsp+224], eax
    37b7ae7e8mov rax, [rsp+440]
    37b7ae7f0mov esi, [rdx]
    37b7ae7f2mov [rsp+212], edi
    37b7ae7f9mov rbp, [rsp+512]
    37b7ae801mov eax, [rax]
    37b7ae803mov edx, [r8]
    37b7ae806test edi, edi
    37b7ae808jle 0x37b7ae8d0
    37b7ae80etest esi, esi
    37b7ae810jle 0x37b7ae8c0
    37b7ae816cmp edx, 1
    37b7ae819jbe 0x37b7ae821
    37b7ae81bmov [rbx], -3
    37b7ae821
    37b7ae821loc_37b7ae821:
    37b7ae821mov edi, [rsp+212]
    37b7ae828add esi, edi
    37b7ae82alea r12l, [rsi+1]
    37b7ae82eadd edx, r12l
    37b7ae831mov [rsp+256], r12l
    37b7ae839mov [rsp+228], edx
    37b7ae840mov [rsp+252], edx
    37b7ae847lea edx, [rdi+1]
    37b7ae84amov rdi, [rsp+392]
    37b7ae852mov [rsp+260], edx
    37b7ae859mov edi, [rdi]
    37b7ae85btest edi, edi
    37b7ae85djle 0x37b7ae8e0
    37b7ae863cmp r12l, edi
    37b7ae866jl 0x37b7ae930
    37b7ae86ccmp [rsp+200], edi
    37b7ae873jl 0x37b7ae93e
    37b7ae879cmp eax, r12l
    37b7ae87cjl 0x37b7ae94c
    37b7ae882cmp [rsp+224], r12l
    37b7ae88ajl 0x37b7ae95a
    37b7ae890mov edx, [rsp+228]
    37b7ae897cmp ecx, edx
    37b7ae899jl 0x37b7afc4c
    37b7ae89fcmp [rsp+232], edx
    37b7ae8a6jge 0x37b7ae970
    37b7ae8acmov edx, -16
    37b7ae8b1mov eax, 16
    37b7ae8b6mov [rbx], edx
    37b7ae8b8jmp 0x37b7ae8eb
    37b7ae8ba
    37b7ae8ba 66 0f 1f 44 00 00 f..D..
    37b7ae8c0
    37b7ae8c0loc_37b7ae8c0:
    37b7ae8c0mov [rbx], -2
    37b7ae8c6jmp 0x37b7ae821
    37b7ae8cb
    37b7ae8cb 0f 1f 44 00 00 ..D..
    37b7ae8d0
    37b7ae8d0loc_37b7ae8d0:
    37b7ae8d0mov [rbx], -1
    37b7ae8d6jmp 0x37b7ae821
    37b7ae8db
    37b7ae8db 0f 1f 44 00 00 ..D..
    37b7ae8e0
    37b7ae8e0loc_37b7ae8e0:
    37b7ae8e0mov [rbx], -4
    37b7ae8e6mov eax, 4
    37b7ae8eb
    37b7ae8ebloc_37b7ae8eb:
    37b7ae8eblea rdx, [rsp+268]
    37b7ae8f3mov r8l, 6
    37b7ae8f9lea rcx, [rip+0xc8a160]
    37b7ae900mov [rsp+268], eax
    37b7ae907call xerbla_
    37b7ae90cnop
    37b7ae90d
    37b7ae90dloc_37b7ae90d:
    37b7ae90dmovups x6, [rsp+272]
    37b7ae915add rsp, 296
    37b7ae91cpop rbx
    37b7ae91dpop rsi
    37b7ae91epop rdi
    37b7ae91fpop rbp
    37b7ae920pop r12
    37b7ae922pop r13
    37b7ae924pop r14
    37b7ae926pop r15
    37b7ae928ret
    37b7ae929
    37b7ae929 0f 1f 80 00 00 00 00 .......
    37b7ae930
    37b7ae930loc_37b7ae930:
    37b7ae930mov edx, -4
    37b7ae935mov eax, 4
    37b7ae93amov [rbx], edx
    37b7ae93cjmp 0x37b7ae8eb
    37b7ae93e
    37b7ae93eloc_37b7ae93e:
    37b7ae93emov edx, -7
    37b7ae943mov eax, 7
    37b7ae948mov [rbx], edx
    37b7ae94ajmp 0x37b7ae8eb
    37b7ae94c
    37b7ae94cloc_37b7ae94c:
    37b7ae94cmov edx, -10
    37b7ae951mov eax, 10
    37b7ae956mov [rbx], edx
    37b7ae958jmp 0x37b7ae8eb
    37b7ae95a
    37b7ae95aloc_37b7ae95a:
    37b7ae95amov edx, -12
    37b7ae95fmov eax, 12
    37b7ae964mov [rbx], edx
    37b7ae966jmp 0x37b7ae8eb
    37b7ae968
    37b7ae968 0f 1f 84 00 00 00 00 00 ........
    37b7ae970
    37b7ae970loc_37b7ae970:
    37b7ae970mov edx, [rbx]
    37b7ae972test edx, edx
    37b7ae974jne 0x37b7afc5d
    37b7ae97acmp edi, 1
    37b7ae97dje 0x37b7af720
    37b7ae983mov [rsp+128], eax
    37b7ae98amovsxd rax, ecx
    37b7ae98dxor edx, edx
    37b7ae98flea rdi, [rip+0xc8a0d2]
    37b7ae996test rax, rax
    37b7ae999mov [rsp+32], rdi
    37b7ae99emov r8, rdi
    37b7ae9a1mov r9, [rsp+408]
    37b7ae9a9cmovns rdx, rax
    37b7ae9admov eax, [rsp+212]
    37b7ae9b4mov rcx, [rsp+392]
    37b7ae9bclea rsi, [rsp+268]
    37b7ae9c4mov [rsp+168], rdx
    37b7ae9ccnot rdx
    37b7ae9cfadd eax, 2
    37b7ae9d2mov [rsp+136], rdx
    37b7ae9damov rdx, rbp
    37b7ae9ddmov [rsp+236], eax
    37b7ae9e4call scopy_
    37b7ae9e9mov rcx, [rsp+392]
    37b7ae9f1mov r8, rdi
    37b7ae9f4mov rdx, rbp
    37b7ae9f7call snrm2_
    37b7ae9fcmov rax, [rsp+392]
    37b7aea04mov [rsp+72], rbx
    37b7aea09mov r9, rsi
    37b7aea0cmov [rsp+80], 1
    37b7aea15lea r8, [rip+0xc8a050]
    37b7aea1clea rcx, [rip+0xc8a04d]
    37b7aea23mov [rsp+64], rax
    37b7aea28mov rdx, r8
    37b7aea2bmov [rsp+40], rax
    37b7aea30lea rax, [rip+0xc8a03d]
    37b7aea37mov [rsp+32], rax
    37b7aea3cmov [rsp+56], rbp
    37b7aea41mov [rsp+48], rdi
    37b7aea46movss [rsp+268], x0
    37b7aea4fcall slascl_
    37b7aea54mov rax, [rsp+392]
    37b7aea5cmovss x0, [rsp+268]
    37b7aea65mov [rsp+244], 1
    37b7aea70mulss x0, x0
    37b7aea74mov edx, [rax]
    37b7aea76movss [rsp+268], x0
    37b7aea7ftest edx, edx
    37b7aea81jle 0x37b7afc37
    37b7aea87movsxd rax, [rsp+128]
    37b7aea8flea r12, [rsp+244]
    37b7aea97mov [rsp+424], r14
    37b7aea9fmov [rsp+144], r12
    37b7aeaa7mov r14, [rsp+432]
    37b7aeaafmov rdi, rax
    37b7aeab2mov r12, [rsp+168]
    37b7aeabamov [rsp+128], rax
    37b7aeac2mov eax, 1
    37b7aeac7mov [rsp+512], rbp
    37b7aeacfnot rdi
    37b7aead2mov rbp, rbx
    37b7aead5mov ebx, edx
    37b7aead7nop [rax+rax]
    37b7aeae0
    37b7aeae0loc_37b7aeae0:
    37b7aeae0cdqe
    37b7aeae2mov rcx, [rsp+136]
    37b7aeaeamov [rsp+64], rbp
    37b7aeaefmov rdx, rax
    37b7aeaf2mov [rsp+40], rsi
    37b7aeaf7mov r8, [rsp+424]
    37b7aeaffimul rdx, r12
    37b7aeb03mov r9, [rsp+512]
    37b7aeb0blea rdx, [rcx+rdx+1]
    37b7aeb10mov rcx, [rsp+392]
    37b7aeb18lea rdx, [r13+4*rdx]
    37b7aeb1dmov [rsp+56], rdx
    37b7aeb22lea rdx, [r15+4*rax-4]
    37b7aeb27imul rax, [rsp+128]
    37b7aeb30mov [rsp+48], rdx
    37b7aeb35mov rdx, [rsp+144]
    37b7aeb3dlea rax, [rdi+rax+1]
    37b7aeb42lea rax, [r14+4*rax]
    37b7aeb46mov [rsp+32], rax
    37b7aeb4bcall slasd4_
    37b7aeb50mov r8l, [rbp]
    37b7aeb54test r8l, r8l
    37b7aeb57jne 0x37b7ae90d
    37b7aeb5dmov eax, [rsp+244]
    37b7aeb64add eax, 1
    37b7aeb67mov [rsp+244], eax
    37b7aeb6ecmp eax, ebx
    37b7aeb70jle 0x37b7aeae0
    37b7aeb76mov rax, [rsp+392]
    37b7aeb7emov r14, [rsp+424]
    37b7aeb86mov rbp, [rsp+512]
    37b7aeb8emov eax, [rax]
    37b7aeb90mov [rsp+144], eax
    37b7aeb97test eax, eax
    37b7aeb99jle 0x37b7afc37
    37b7aeb9fmov rsi, [rsp+128]
    37b7aeba7mov r12, r13
    37b7aebaamov r11, r13
    37b7aebadmovsxd rax, [rsp+144]
    37b7aebb5mov [rsp+464], r13
    37b7aebbdmov r8l, 1
    37b7aebc3movss x3, [rip+0xc89ed5]
    37b7aebcblea rbx, [4*rsi+4]
    37b7aebd3mov rcx, rax
    37b7aebd6mov [rsp+512], rbp
    37b7aebdemov [rsp+152], rbx
    37b7aebe6mov rbx, [rsp+168]
    37b7aebeeadd ecx, 1
    37b7aebf1mov [rsp+184], rcx
    37b7aebf9lea rdx, [4*rbx+4]
    37b7aec01mov [rsp+160], rdx
    37b7aec09mov rdx, rsi
    37b7aec0cshl rsi, 2
    37b7aec10imul rdx, rax
    37b7aec14imul rax, rbx
    37b7aec18add rax, [rsp+136]
    37b7aec20lea rax, [r13+4*rax]
    37b7aec25add rdx, rdi
    37b7aec28mov rdi, [rsp+432]
    37b7aec30mov [rsp+176], rax
    37b7aec38lea rax, [r14+4]
    37b7aec3cmov [rsp+192], rax
    37b7aec44lea r15, [rdi+4*rdx]
    37b7aec48lea rdi, [4*rbx]
    37b7aec50mov rbx, [rsp+432]
    37b7aec58mov r10, rbx
    37b7aec5bnop [rax+rax]
    37b7aec60
    37b7aec60loc_37b7aec60:
    37b7aec60mov rax, [rsp+176]
    37b7aec68movss x2, [r15+4*r8]
    37b7aec6emov r13l, r8l
    37b7aec71mov ebp, r8l
    37b7aec74mulss x2, [rax+4*r8]
    37b7aec7acmp r8, 1
    37b7aec7eje 0x37b7aecf2
    37b7aec80mov rcx, [rsp+192]
    37b7aec88lea edx, [r8-2]
    37b7aec8cmovss x4, [r14+4*r8-4]
    37b7aec93mov rax, r14
    37b7aec96lea r9, [rcx+4*rdx]
    37b7aec9amov rcx, r11
    37b7aec9dmov rdx, r10
    37b7aeca0nop [rax+rax]
    37b7aecaanop [rax+rax]
    37b7aecb5nop [rax+rax]
    37b7aecc0
    37b7aecc0loc_37b7aecc0:
    37b7aecc0movss x1, [rax]
    37b7aecc4movss x0, [rdx]
    37b7aecc8movaps x5, x4
    37b7aeccbadd rax, 4
    37b7aeccfmulss x0, [rcx]
    37b7aecd3add rdx, rsi
    37b7aecd6add rcx, rdi
    37b7aecd9subss x5, x1
    37b7aecddaddss x1, x4
    37b7aece1divss x0, x5
    37b7aece5divss x0, x1
    37b7aece9mulss x2, x0
    37b7aecedcmp rax, r9
    37b7aecf0jne 0x37b7aecc0
    37b7aecf2
    37b7aecf2loc_37b7aecf2:
    37b7aecf2mov ecx, [rsp+144]
    37b7aecf9cmp ecx, ebp
    37b7aecfbjle 0x37b7aed72
    37b7aecfdmov edx, ecx
    37b7aecffmovss x4, [r14+4*r8-4]
    37b7aed06lea rax, [r14+4*r8]
    37b7aed0amov rcx, r12
    37b7aed0dsub edx, r13l
    37b7aed10add rdx, r8
    37b7aed13lea r9, [r14+4*rdx]
    37b7aed17mov rdx, rbx
    37b7aed1anop [rax+rax]
    37b7aed1fnop [rax+rax]
    37b7aed2anop [rax+rax]
    37b7aed35nop [rax+rax]
    37b7aed40
    37b7aed40loc_37b7aed40:
    37b7aed40movss x1, [rax]
    37b7aed44movss x0, [rdx]
    37b7aed48movaps x5, x4
    37b7aed4badd rax, 4
    37b7aed4fmulss x0, [rcx]
    37b7aed53add rdx, rsi
    37b7aed56add rcx, rdi
    37b7aed59subss x5, x1
    37b7aed5daddss x1, x4
    37b7aed61divss x0, x5
    37b7aed65divss x0, x1
    37b7aed69mulss x2, x0
    37b7aed6dcmp rax, r9
    37b7aed70jne 0x37b7aed40
    37b7aed72
    37b7aed72loc_37b7aed72:
    37b7aed72andps x2, [rip+0xc89d17]
    37b7aed79movaps x0, x3
    37b7aed7cadd r10, 4
    37b7aed80add r11, 4
    37b7aed84mov rax, [rsp+408]
    37b7aed8cadd rbx, [rsp+152]
    37b7aed94sqrtss x2, x2
    37b7aed98add r12, [rsp+160]
    37b7aeda0movss x1, [rax+4*r8-4]
    37b7aeda7mov rax, [rsp+512]
    37b7aedafandnps x0, x2
    37b7aedb2andps x1, x3
    37b7aedb5orps x0, x1
    37b7aedb8movss [rax+4*r8-4], x0
    37b7aedbfadd r8, 1
    37b7aedc3cmp r8, [rsp+184]
    37b7aedcbjne 0x37b7aec60
    37b7aedd1mov rax, [rsp+432]
    37b7aedd9mov r13, [rsp+464]
    37b7aede1mov [rsp+192], rsi
    37b7aede9mov r15, -1
    37b7aedf0mov rbp, [rsp+512]
    37b7aedf8mov edx, [rsp+144]
    37b7aedffmov r10, -1
    37b7aee06mov r11l, 1
    37b7aee0clea rbx, [rax-4]
    37b7aee10movsxd rax, [rsp+200]
    37b7aee18mov [rsp+200], rdi
    37b7aee20lea r12, [r13-4]
    37b7aee24movss x6, [rip+0xc89c84]
    37b7aee2cmov rsi, [rsp+496]
    37b7aee34mov [rsp+184], rax
    37b7aee3cmov rax, [rsp+408]
    37b7aee44add rax, 8
    37b7aee48mov [rsp+216], rax
    37b7aee50
    37b7aee50loc_37b7aee50:
    37b7aee50movss x0, [rbp]
    37b7aee55divss x0, [rbx+4]
    37b7aee5amovss [rbx+4], x6
    37b7aee5fdivss x0, [r12+4]
    37b7aee66movss [r12+4], x0
    37b7aee6dcmp edx, 1
    37b7aee70jle 0x37b7aef2a
    37b7aee76lea eax, [rdx-2]
    37b7aee79cmp eax, 2
    37b7aee7cjbe 0x37b7af716
    37b7aee82lea ecx, [rdx-1]
    37b7aee85lea r9, [rbx+8]
    37b7aee89xor eax, eax
    37b7aee8bmov edi, ecx
    37b7aee8dlea r8, [r12+8]
    37b7aee92shr edi, 2
    37b7aee95shl rdi, 4
    37b7aee99nop [rax+rax]
    37b7aee9fnop [rax+rax]
    37b7aeeaanop [rax+rax]
    37b7aeeb5nop [rax+rax]
    37b7aeec0
    37b7aeec0loc_37b7aeec0:
    37b7aeec0movups x3, [r9+rax]
    37b7aeec5movups x0, [rbp+rax+4]
    37b7aeecadivps x0, x3
    37b7aeecdmovups x3, [r8+rax]
    37b7aeed2divps x0, x3
    37b7aeed5movups x3, [r14+rax+4]
    37b7aeedbmovups [r8+rax], x0
    37b7aeee0mulps x0, x3
    37b7aeee3movups [r9+rax], x0
    37b7aeee8add rax, 16
    37b7aeeeccmp rdi, rax
    37b7aeeefjne 0x37b7aeec0
    37b7aeef1test cl, 3
    37b7aeef4je 0x37b7aef2a
    37b7aeef6and ecx, -4
    37b7aeef9add ecx, 2
    37b7aeefc
    37b7aeefcloc_37b7aeefc:
    37b7aeefcmovsxd rcx, ecx
    37b7aeeff
    37b7aeeffloc_37b7aeeff:
    37b7aeeffmovss x0, [rbp+4*rcx-4]
    37b7aef05divss x0, [rbx+4*rcx]
    37b7aef0adivss x0, [r12+4*rcx]
    37b7aef10movss [r12+4*rcx], x0
    37b7aef16mulss x0, [r14+4*rcx-4]
    37b7aef1dmovss [rbx+4*rcx], x0
    37b7aef22add rcx, 1
    37b7aef26cmp edx, ecx
    37b7aef28jge 0x37b7aeeff
    37b7aef2a
    37b7aef2aloc_37b7aef2a:
    37b7aef2amov rcx, [rsp+392]
    37b7aef32lea rdx, [rbx+4]
    37b7aef36lea r8, [rip+0xc89b2b]
    37b7aef3dmov [rsp+160], r10
    37b7aef45mov [rsp+152], r11l
    37b7aef4dcall snrm2_
    37b7aef52mov rax, [rsp+408]
    37b7aef5amov r10, [rsp+160]
    37b7aef62movaps x1, x0
    37b7aef65movss x0, [rbx+4]
    37b7aef6amov r11l, [rsp+152]
    37b7aef72divss x0, x1
    37b7aef76movss [rax+4*r10+4], x0
    37b7aef7dmov rax, [rsp+392]
    37b7aef85mov edx, [rax]
    37b7aef87cmp edx, 1
    37b7aef8ajle 0x37b7af07f
    37b7aef90lea eax, [rdx-2]
    37b7aef93cmp eax, 2
    37b7aef96jbe 0x37b7af703
    37b7aef9clea r8l, [rdx-1]
    37b7aefa0mov rax, [rsp+216]
    37b7aefa8mov [rsp+152], edx
    37b7aefaflea rcx, [4*r10]
    37b7aefb7mov edi, r8l
    37b7aefbamov [rsp+160], r11l
    37b7aefc2movaps x3, x1
    37b7aefc5mov [rsp+176], r12
    37b7aefcdshr edi, 2
    37b7aefd0lea r13, [rax+rcx]
    37b7aefd4xor eax, eax
    37b7aefd6shl rdi, 4
    37b7aefdashufps x3, x3, 0
    37b7aefdenop
    37b7aefe0
    37b7aefe0loc_37b7aefe0:
    37b7aefe0lea rdx, [rsi+rax]
    37b7aefe4movsxd r9, [rdx+4]
    37b7aefe8movsxd r11, [rdx+8]
    37b7aefecmovsxd r12, [rdx+12]
    37b7aeff0movsxd rdx, [rdx+16]
    37b7aeff4movss x4, [rbx+4*r11]
    37b7aeffamovss x0, [rbx+4*rdx]
    37b7aefffmovss x2, [rbx+4*r12]
    37b7af005unpcklps x2, x0
    37b7af008movss x0, [rbx+4*r9]
    37b7af00eunpcklps x0, x4
    37b7af011movlhps x0, x2
    37b7af014divps x0, x3
    37b7af017movups [r13+rax], x0
    37b7af01dadd rax, 16
    37b7af021cmp rdi, rax
    37b7af024jne 0x37b7aefe0
    37b7af026mov edx, [rsp+152]
    37b7af02dmov r11l, [rsp+160]
    37b7af035mov r12, [rsp+176]
    37b7af03dtest r8b, 3
    37b7af041je 0x37b7af07f
    37b7af043and r8l, -4
    37b7af047add r8l, 2
    37b7af04b
    37b7af04bloc_37b7af04b:
    37b7af04bmov r9, [rsp+432]
    37b7af053movsxd r8, r8l
    37b7af056add rcx, [rsp+408]
    37b7af05e
    37b7af05eloc_37b7af05e:
    37b7af05emovsxd rax, [rsi+4*r8-4]
    37b7af063add rax, r15
    37b7af066movss x0, [r9+4*rax]
    37b7af06cdivss x0, x1
    37b7af070movss [rcx+4*r8], x0
    37b7af076add r8, 1
    37b7af07acmp edx, r8l
    37b7af07djge 0x37b7af05e
    37b7af07f
    37b7af07floc_37b7af07f:
    37b7af07fadd r11l, 1
    37b7af083add rbx, [rsp+192]
    37b7af08badd r12, [rsp+200]
    37b7af093add r15, [rsp+128]
    37b7af09badd r10, [rsp+184]
    37b7af0a3cmp [rsp+144], r11l
    37b7af0abjge 0x37b7aee50
    37b7af0b1mov r13, [rsp+464]
    37b7af0b9cmp edx, 2
    37b7af0bcje 0x37b7afb0c
    37b7af0c2
    37b7af0c2loc_37b7af0c2:
    37b7af0c2mov rax, [rsp+504]
    37b7af0camovsxd rbx, [rsp+224]
    37b7af0d2mov eax, [rax]
    37b7af0d4mov rsi, rbx
    37b7af0d7not rsi
    37b7af0datest eax, eax
    37b7af0dcjg 0x37b7af7fe
    37b7af0e2mov rdi, [rsp+504]
    37b7af0eamov edx, [rdi+8]
    37b7af0edtest edx, edx
    37b7af0efjle 0x37b7afbc0
    37b7af0f5mov rcx, [rsp+440]
    37b7af0fdmov edx, [rdi+4]
    37b7af100lea rdi, [rip+0xc89971]
    37b7af107mov [rsp+112], 1
    37b7af110mov [rsp+144], rdi
    37b7af118mov [rsp+96], rcx
    37b7af11dlea eax, [rax+rdx+2]
    37b7af121mov rcx, [rsp+432]
    37b7af129mov [rsp+104], 1
    37b7af132cdqe
    37b7af134mov [rsp+88], rcx
    37b7af139lea rcx, [rip+0xc8993c]
    37b7af140mov [rsp+152], rcx
    37b7af148mov [rsp+80], rcx
    37b7af14dmov rcx, [rsp+416]
    37b7af155mov [rsp+72], rcx
    37b7af15amov rcx, [rsp+408]
    37b7af162lea rdx, [rcx+4*rax-4]
    37b7af167mov rcx, [rsp+456]
    37b7af16fimul rax, rbx
    37b7af173mov [rsp+64], rdx
    37b7af178mov [rsp+56], rcx
    37b7af17dmov rcx, [rsp+448]
    37b7af185lea rax, [rsi+rax+1]
    37b7af18alea rax, [rcx+4*rax]
    37b7af18emov [rsp+48], rax
    37b7af193lea rax, [rip+0xc898da]
    37b7af19amov [rsp+40], rax
    37b7af19f
    37b7af19floc_37b7af19f:
    37b7af19fmov rax, [rsp+504]
    37b7af1a7mov r9, [rsp+392]
    37b7af1afmov rdx, rdi
    37b7af1b2mov rcx, rdi
    37b7af1b5mov r8, [rsp+368]
    37b7af1bdadd rax, 8
    37b7af1c1mov [rsp+32], rax
    37b7af1c6call sgemm_
    37b7af1cb
    37b7af1cbloc_37b7af1cb:
    37b7af1cbmov rax, [rsp+440]
    37b7af1d3movsxd rdi, [rsp+212]
    37b7af1dbmov r8, [rsp+416]
    37b7af1e3mov rdx, [rsp+408]
    37b7af1ebmov [rsp+32], rax
    37b7af1f0mov rax, [rsp+432]
    37b7af1f8shl rdi, 2
    37b7af1fcmov rcx, [rsp+392]
    37b7af204lea r9, [rax+rdi]
    37b7af208call scopy_
    37b7af20dmov rcx, [rsp+504]
    37b7af215mov rax, [rsp+504]
    37b7af21dmov r9, [rsp+392]
    37b7af225mov r8, [rsp+376]
    37b7af22dmov edx, [rcx+8]
    37b7af230add edx, [rcx+4]
    37b7af233mov rcx, [rsp+440]
    37b7af23bmov eax, [rax]
    37b7af23dmov [rsp+240], edx
    37b7af244mov [rsp+112], 1
    37b7af24dmov [rsp+96], rcx
    37b7af252mov rcx, [rsp+432]
    37b7af25amov [rsp+128], eax
    37b7af261add eax, 2
    37b7af264mov [rsp+104], 1
    37b7af26dlea rdx, [rcx+rdi+4]
    37b7af272cdqe
    37b7af274mov rdi, [rsp+152]
    37b7af27cmov [rsp+88], rdx
    37b7af281movsxd rcx, [rsp+236]
    37b7af289mov [rsp+80], rdi
    37b7af28emov rdi, [rsp+416]
    37b7af296mov [rsp+72], rdi
    37b7af29bmov rdi, [rsp+408]
    37b7af2a3lea rdx, [rdi+4*rax-4]
    37b7af2a8mov rdi, [rsp+456]
    37b7af2b0imul rax, rbx
    37b7af2b4mov [rsp+64], rdx
    37b7af2b9mov [rsp+56], rdi
    37b7af2beadd rax, rsi
    37b7af2c1mov rsi, [rsp+448]
    37b7af2c9add rax, rcx
    37b7af2ccmov rcx, [rsp+144]
    37b7af2d4lea rax, [rsi+4*rax]
    37b7af2d8mov [rsp+48], rax
    37b7af2ddlea rax, [rip+0xc89790]
    37b7af2e4mov rdx, rcx
    37b7af2e7mov [rsp+40], rax
    37b7af2eclea rax, [rsp+240]
    37b7af2f4mov [rsp+32], rax
    37b7af2f9call sgemm_
    37b7af2fe
    37b7af2feloc_37b7af2fe:
    37b7af2femov rax, [rsp+184]
    37b7af306mov rsi, [rsp+392]
    37b7af30emov rbx, rax
    37b7af311mov r15l, [rsi]
    37b7af314not rbx
    37b7af317mov [rsp+160], rbx
    37b7af31ftest r15l, r15l
    37b7af322jle 0x37b7af449
    37b7af328mov rsi, rax
    37b7af32blea rax, [rbx+2*rax]
    37b7af32fmov rbx, [rsp+408]
    37b7af337mov [rsp+128], r15l
    37b7af33flea rbp, [4*rsi]
    37b7af347mov r15l, 1
    37b7af34dlea rsi, [r13+4]
    37b7af351mov rdi, -1
    37b7af358lea r14, [rbx+4*rax+4]
    37b7af35dmov rax, [rsp+496]
    37b7af365mov rbx, [rsp+392]
    37b7af36dlea r12, [rax+8]
    37b7af371mov rax, rbp
    37b7af374mov rbp, r14
    37b7af377mov r14, rax
    37b7af37anop [rax+rax]
    37b7af380
    37b7af380loc_37b7af380:
    37b7af380lea rdx, [rsi+4*rdi]
    37b7af384lea r8, [rip+0xc896dd]
    37b7af38bmov rcx, rbx
    37b7af38ecall snrm2_
    37b7af393mov rax, [rsp+408]
    37b7af39bmov r9l, [rbx]
    37b7af39emovaps x1, x0
    37b7af3a1movss x0, [r13+4*rdi+4]
    37b7af3a8divss x0, x1
    37b7af3acmovss [rax+4*r15-4], x0
    37b7af3b3cmp r9l, 1
    37b7af3b7jle 0x37b7af421
    37b7af3b9mov rax, [rsp+496]
    37b7af3c1lea edx, [r9-2]
    37b7af3c5mov rcx, rbp
    37b7af3c8lea r8, [r12+4*rdx]
    37b7af3ccadd rax, 4
    37b7af3d0nop [rax]
    37b7af3d4nop [rax+rax]
    37b7af3dfnop [rax+rax]
    37b7af3eanop [rax+rax]
    37b7af3f5nop [rax+rax]
    37b7af400
    37b7af400loc_37b7af400:
    37b7af400movsxd rdx, [rax]
    37b7af403add rax, 4
    37b7af407add rdx, rdi
    37b7af40amovss x0, [r13+4*rdx]
    37b7af411divss x0, x1
    37b7af415movss [rcx], x0
    37b7af419add rcx, r14
    37b7af41ccmp r8, rax
    37b7af41fjne 0x37b7af400
    37b7af421
    37b7af421loc_37b7af421:
    37b7af421add r15, 1
    37b7af425add rbp, 4
    37b7af429add rdi, [rsp+168]
    37b7af431cmp [rsp+128], r15l
    37b7af439jge 0x37b7af380
    37b7af43fcmp r9l, 2
    37b7af443je 0x37b7afa3d
    37b7af449
    37b7af449loc_37b7af449:
    37b7af449movsxd rbx, [rsp+232]
    37b7af451xor eax, eax
    37b7af453mov rcx, [rsp+144]
    37b7af45blea r9, [rsp+260]
    37b7af463mov r8, [rsp+392]
    37b7af46btest rbx, rbx
    37b7af46emov rdx, rcx
    37b7af471cmovs rbx, rax
    37b7af475mov rax, [rsp+504]
    37b7af47dmov eax, [rax]
    37b7af47fmov [rsp+88], r13
    37b7af484mov rbp, rbx
    37b7af487mov [rsp+112], 1
    37b7af490not rbp
    37b7af493mov [rsp+128], eax
    37b7af49aadd eax, 1
    37b7af49dmov [rsp+248], eax
    37b7af4a4mov rax, [rsp+472]
    37b7af4acmov [rsp+104], 1
    37b7af4b5mov [rsp+96], rax
    37b7af4bamov rax, [rsp+152]
    37b7af4c2mov [rsp+80], rax
    37b7af4c7mov rax, [rsp+488]
    37b7af4cfmov [rsp+72], rax
    37b7af4d4mov rax, [rsp+480]
    37b7af4dcmov [rsp+64], rax
    37b7af4e1mov rax, [rsp+416]
    37b7af4e9mov [rsp+56], rax
    37b7af4eemov rax, [rsp+408]
    37b7af4f6mov [rsp+48], rax
    37b7af4fblea rax, [rip+0xc89572]
    37b7af502mov [rsp+40], rax
    37b7af507lea rax, [rsp+248]
    37b7af50fmov [rsp+32], rax
    37b7af514call sgemm_
    37b7af519mov rax, [rsp+504]
    37b7af521mov rdi, [rsp+488]
    37b7af529movsxd r8, [rax]
    37b7af52cmov r11l, [rax+4]
    37b7af530lea eax, [r8+r11+2]
    37b7af535cmp eax, [rdi]
    37b7af537jle 0x37b7af96e
    37b7af53d
    37b7af53dloc_37b7af53d:
    37b7af53dmov rsi, [rsp+384]
    37b7af545mov r9, [rsp+184]
    37b7af54dlea eax, [r8+1]
    37b7af551mov edx, [rsi]
    37b7af553mov rsi, [rsp+376]
    37b7af55badd edx, [rsi]
    37b7af55dmovsxd rsi, eax
    37b7af560imul r9, rsi
    37b7af564mov [rsp+264], edx
    37b7af56badd r9, [rsp+160]
    37b7af573cmp eax, 1
    37b7af576jle 0x37b7af961
    37b7af57cmov rax, [rsp+392]
    37b7af584mov edx, [rax]
    37b7af586test edx, edx
    37b7af588jle 0x37b7af5b4
    37b7af58amov rax, [rsp+408]
    37b7af592sub edx, 1
    37b7af595lea rdx, [rax+4*rdx+4]
    37b7af59anop [rax+rax]
    37b7af5a0
    37b7af5a0loc_37b7af5a0:
    37b7af5a0movss x0, [rax]
    37b7af5a4movss [rax+4*r9+4], x0
    37b7af5abadd rax, 4
    37b7af5afcmp rax, rdx
    37b7af5b2jne 0x37b7af5a0
    37b7af5b4
    37b7af5b4loc_37b7af5b4:
    37b7af5b4mov edi, [rsp+236]
    37b7af5bbmovsxd rcx, edi
    37b7af5becmp [rsp+228], edi
    37b7af5c5jl 0x37b7af617
    37b7af5c7movsxd rcx, [rsp+236]
    37b7af5cfmov r10, [rsp+480]
    37b7af5d7lea rdi, [4*rbx]
    37b7af5dfmov rax, rcx
    37b7af5e2mov rdx, rcx
    37b7af5e5imul rax, rbx
    37b7af5e9add rax, rbp
    37b7af5eclea rax, [r10+4*rax]
    37b7af5f0mov r10l, [rsp+228]
    37b7af5f8add r10l, 1
    37b7af5fcnop [rax]
    37b7af600
    37b7af600loc_37b7af600:
    37b7af600movss x0, [rax+4]
    37b7af605add edx, 1
    37b7af608movss [rax+4*r8+4], x0
    37b7af60fadd rax, rdi
    37b7af612cmp r10l, edx
    37b7af615jne 0x37b7af600
    37b7af617
    37b7af617loc_37b7af617:
    37b7af617mov rax, [rsp+504]
    37b7af61fmov rdi, [rsp+136]
    37b7af627mov r8, [rsp+392]
    37b7af62fmov eax, [rax+8]
    37b7af632mov [rsp+112], 1
    37b7af63bmov [rsp+104], 1
    37b7af644lea eax, [rax+r11+1]
    37b7af649mov [rsp+240], eax
    37b7af650mov rax, [rsp+472]
    37b7af658mov [rsp+96], rax
    37b7af65dmov rax, [rsp+168]
    37b7af665imul rax, rcx
    37b7af669imul rcx, rbx
    37b7af66dlea rax, [rdi+rax+1]
    37b7af672lea rax, [r13+4*rax]
    37b7af677mov [rsp+88], rax
    37b7af67cmov rax, [rsp+152]
    37b7af684mov [rsp+80], rax
    37b7af689mov rax, [rsp+488]
    37b7af691mov [rsp+72], rax
    37b7af696lea rax, [rcx+rbp]
    37b7af69amov rcx, [rsp+144]
    37b7af6a2add rax, rsi
    37b7af6a5mov rsi, [rsp+480]
    37b7af6admov rdx, rcx
    37b7af6b0lea rax, [rsi+4*rax]
    37b7af6b4mov [rsp+64], rax
    37b7af6b9mov rax, [rsp+416]
    37b7af6c1mov [rsp+56], rax
    37b7af6c6mov rax, [rsp+408]
    37b7af6celea rax, [rax+4*r9+4]
    37b7af6d3lea r9, [rsp+264]
    37b7af6dbmov [rsp+48], rax
    37b7af6e0lea rax, [rip+0xc8938d]
    37b7af6e7mov [rsp+40], rax
    37b7af6eclea rax, [rsp+240]
    37b7af6f4mov [rsp+32], rax
    37b7af6f9call sgemm_
    37b7af6fejmp 0x37b7ae90d
    37b7af703
    37b7af703loc_37b7af703:
    37b7af703mov r8l, 2
    37b7af709lea rcx, [4*r10]
    37b7af711jmp 0x37b7af04b
    37b7af716
    37b7af716loc_37b7af716:
    37b7af716mov ecx, 2
    37b7af71bjmp 0x37b7aeefc
    37b7af720
    37b7af720loc_37b7af720:
    37b7af720movss x0, [rbp]
    37b7af725mov rax, [rsp+472]
    37b7af72dlea rcx, [rsp+252]
    37b7af735mov r9, r13
    37b7af738andps x0, [rip+0xc89351]
    37b7af73fmov r8, [rsp+488]
    37b7af747mov rdx, [rsp+480]
    37b7af74fmovss [r15], x0
    37b7af754mov [rsp+32], rax
    37b7af759call scopy_
    37b7af75emovss x1, [rbp]
    37b7af763pxor x0, x0
    37b7af767comiss x1, x0
    37b7af76aja 0x37b7afadb
    37b7af770cmp esi, 2
    37b7af773jbe 0x37b7af7c8
    37b7af775mov edx, r12l
    37b7af778movss x1, [rip+0xc89320]
    37b7af780xor eax, eax
    37b7af782mov rcx, [rsp+432]
    37b7af78ashr edx, 2
    37b7af78dmov r8, [rsp+448]
    37b7af795shl rdx, 4
    37b7af799shufps x1, x1, 0
    37b7af79dnop [rax]
    37b7af7a0
    37b7af7a0loc_37b7af7a0:
    37b7af7a0movups x0, [r8+rax]
    37b7af7a5xorps x0, x1
    37b7af7a8movups [rcx+rax], x0
    37b7af7acadd rax, 16
    37b7af7b0cmp rax, rdx
    37b7af7b3jne 0x37b7af7a0
    37b7af7b5test r12b, 3
    37b7af7b9je 0x37b7ae90d
    37b7af7bfand r12l, -4
    37b7af7c3lea edi, [r12+1]
    37b7af7c8
    37b7af7c8loc_37b7af7c8:
    37b7af7c8mov rdx, [rsp+432]
    37b7af7d0mov rcx, [rsp+448]
    37b7af7d8movsxd rdi, edi
    37b7af7db
    37b7af7dbloc_37b7af7db:
    37b7af7dbmovss x0, [rcx+4*rdi-4]
    37b7af7e1xorps x0, [rip+0xc892b8]
    37b7af7e8movss [rdx+4*rdi-4], x0
    37b7af7eeadd rdi, 1
    37b7af7f2lea eax, [rdi-1]
    37b7af7f5cmp esi, eax
    37b7af7f7jge 0x37b7af7db
    37b7af7f9jmp 0x37b7ae90d
    37b7af7fe
    37b7af7feloc_37b7af7fe:
    37b7af7femov rax, [rsp+440]
    37b7af806lea r15, [rip+0xc89267]
    37b7af80dmov rcx, [rsp+448]
    37b7af815mov [rsp+104], 1
    37b7af81emov [rsp+40], r15
    37b7af823lea rdi, [rip+0xc8924e]
    37b7af82amov r9, [rsp+392]
    37b7af832mov [rsp+96], rax
    37b7af837mov rax, [rsp+432]
    37b7af83fmov rdx, rdi
    37b7af842mov r8, [rsp+368]
    37b7af84amov [rsp+144], rdi
    37b7af852mov [rsp+88], rax
    37b7af857lea rax, [rip+0xc8921e]
    37b7af85emov [rsp+152], rax
    37b7af866mov [rsp+80], rax
    37b7af86bmov rax, [rsp+416]
    37b7af873mov [rsp+112], 1
    37b7af87cmov [rsp+72], rax
    37b7af881mov rax, [rsp+408]
    37b7af889add rax, 4
    37b7af88dmov [rsp+64], rax
    37b7af892mov rax, [rsp+456]
    37b7af89amov [rsp+56], rax
    37b7af89flea rax, [rsi+2*rbx+1]
    37b7af8a4lea rax, [rcx+4*rax]
    37b7af8a8mov rcx, rdi
    37b7af8abmov [rsp+48], rax
    37b7af8b0mov rax, [rsp+504]
    37b7af8b8mov [rsp+32], rax
    37b7af8bdcall sgemm_
    37b7af8c2mov rax, [rsp+504]
    37b7af8camov ecx, [rax+8]
    37b7af8cdtest ecx, ecx
    37b7af8cfjle 0x37b7af1cb
    37b7af8d5mov rcx, [rsp+440]
    37b7af8ddmov edx, [rax]
    37b7af8dfmov eax, [rax+4]
    37b7af8e2mov [rsp+80], r15
    37b7af8e7mov [rsp+96], rcx
    37b7af8ecmov rcx, [rsp+432]
    37b7af8f4lea eax, [rdx+rax+2]
    37b7af8f8mov [rsp+40], r15
    37b7af8fdmov [rsp+88], rcx
    37b7af902cdqe
    37b7af904mov rcx, [rsp+416]
    37b7af90cmov [rsp+112], 1
    37b7af915mov [rsp+72], rcx
    37b7af91amov rcx, [rsp+408]
    37b7af922mov [rsp+104], 1
    37b7af92blea rdx, [rcx+4*rax-4]
    37b7af930mov rcx, [rsp+456]
    37b7af938imul rax, rbx
    37b7af93cmov [rsp+64], rdx
    37b7af941mov [rsp+56], rcx
    37b7af946mov rcx, [rsp+448]
    37b7af94elea rax, [rsi+rax+1]
    37b7af953lea rax, [rcx+4*rax]
    37b7af957mov [rsp+48], rax
    37b7af95cjmp 0x37b7af19f
    37b7af961
    37b7af961loc_37b7af961:
    37b7af961movsxd rcx, [rsp+236]
    37b7af969jmp 0x37b7af617
    37b7af96e
    37b7af96eloc_37b7af96e:
    37b7af96emov rcx, [rsp+488]
    37b7af976cdqe
    37b7af978mov [rsp+88], r13
    37b7af97dlea r9, [rsp+260]
    37b7af985mov rdi, [rsp+472]
    37b7af98dmov r8, [rsp+392]
    37b7af995mov [rsp+112], 1
    37b7af99emov [rsp+72], rcx
    37b7af9a3mov rcx, [rsp+480]
    37b7af9abmov [rsp+96], rdi
    37b7af9b0lea rdi, [rip+0xc890bd]
    37b7af9b7lea rdx, [rcx+4*rax-4]
    37b7af9bcmov rcx, [rsp+416]
    37b7af9c4mov [rsp+80], rdi
    37b7af9c9imul rax, [rsp+184]
    37b7af9d2mov [rsp+64], rdx
    37b7af9d7mov [rsp+56], rcx
    37b7af9dcmov rcx, [rsp+160]
    37b7af9e4mov [rsp+104], 1
    37b7af9edlea rax, [rcx+rax+1]
    37b7af9f2mov rcx, [rsp+408]
    37b7af9famov [rsp+40], rdi
    37b7af9fflea rax, [rcx+4*rax]
    37b7afa03mov rcx, [rsp+144]
    37b7afa0bmov [rsp+48], rax
    37b7afa10mov rax, [rsp+504]
    37b7afa18mov rdx, rcx
    37b7afa1badd rax, 8
    37b7afa1fmov [rsp+32], rax
    37b7afa24call sgemm_
    37b7afa29mov rax, [rsp+504]
    37b7afa31movsxd r8, [rax]
    37b7afa34mov r11l, [rax+4]
    37b7afa38jmp 0x37b7af53d
    37b7afa3d
    37b7afa3dloc_37b7afa3d:
    37b7afa3dmov rax, [rsp+472]
    37b7afa45mov [rsp+88], r13
    37b7afa4alea r9, [rsp+252]
    37b7afa52mov rcx, [rsp+144]
    37b7afa5amov [rsp+112], 1
    37b7afa63mov [rsp+96], rax
    37b7afa68mov rax, [rsp+152]
    37b7afa70mov [rsp+104], 1
    37b7afa79mov rdx, rcx
    37b7afa7cmov [rsp+80], rax
    37b7afa81mov rax, [rsp+488]
    37b7afa89mov [rsp+72], rax
    37b7afa8emov rax, [rsp+480]
    37b7afa96mov [rsp+64], rax
    37b7afa9bmov rax, [rsp+416]
    37b7afaa3mov [rsp+56], rax
    37b7afaa8mov rax, [rsp+408]
    37b7afab0mov [rsp+48], rax
    37b7afab5lea rax, [rip+0xc88fb8]
    37b7afabcmov [rsp+40], rax
    37b7afac1mov rax, [rsp+392]
    37b7afac9mov [rsp+32], rax
    37b7afacemov r8, rax
    37b7afad1call sgemm_
    37b7afad6jmp 0x37b7ae90d
    37b7afadb
    37b7afadbloc_37b7afadb:
    37b7afadblea rax, [rip+0xc88f86]
    37b7afae2mov r9, [rsp+432]
    37b7afaeamov rdx, [rsp+448]
    37b7afaf2lea rcx, [rsp+256]
    37b7afafamov [rsp+32], rax
    37b7afaffmov r8, rax
    37b7afb02call scopy_
    37b7afb07jmp 0x37b7ae90d
    37b7afb0c
    37b7afb0cloc_37b7afb0c:
    37b7afb0cmov rsi, [rsp+440]
    37b7afb14lea rcx, [rip+0xc88f5d]
    37b7afb1bmov [rsp+112], 1
    37b7afb24lea r8, [rsp+256]
    37b7afb2cmov [rsp+104], 1
    37b7afb35mov rdx, rcx
    37b7afb38mov [rsp+96], rsi
    37b7afb3dmov rsi, [rsp+432]
    37b7afb45mov [rsp+144], rcx
    37b7afb4dmov [rsp+88], rsi
    37b7afb52lea rsi, [rip+0xc88f23]
    37b7afb59mov [rsp+152], rsi
    37b7afb61mov [rsp+80], rsi
    37b7afb66mov rsi, [rsp+416]
    37b7afb6emov [rsp+72], rsi
    37b7afb73mov rsi, [rsp+408]
    37b7afb7bmov [rsp+64], rsi
    37b7afb80mov rsi, [rsp+456]
    37b7afb88mov [rsp+56], rsi
    37b7afb8dmov rsi, [rsp+448]
    37b7afb95mov [rsp+48], rsi
    37b7afb9alea rsi, [rip+0xc88ed3]
    37b7afba1mov [rsp+40], rsi
    37b7afba6mov rsi, [rsp+392]
    37b7afbaemov [rsp+32], rsi
    37b7afbb3mov r9, rsi
    37b7afbb6call sgemm_
    37b7afbbbjmp 0x37b7af2fe
    37b7afbc0
    37b7afbc0loc_37b7afbc0:
    37b7afbc0mov rax, [rsp+440]
    37b7afbc8mov r9, [rsp+448]
    37b7afbd0mov [rsp+56], 1
    37b7afbd9lea rcx, [rip+0xc88ea0]
    37b7afbe0mov r8, [rsp+392]
    37b7afbe8mov rdx, [rsp+368]
    37b7afbf0mov [rsp+48], rax
    37b7afbf5mov rax, [rsp+432]
    37b7afbfdmov [rsp+40], rax
    37b7afc02mov rax, [rsp+456]
    37b7afc0amov [rsp+32], rax
    37b7afc0fcall slacpy_
    37b7afc14lea rax, [rip+0xc88e5d]
    37b7afc1bmov [rsp+144], rax
    37b7afc23lea rax, [rip+0xc88e52]
    37b7afc2amov [rsp+152], rax
    37b7afc32jmp 0x37b7af1cb
    37b7afc37
    37b7afc37loc_37b7afc37:
    37b7afc37movsxd rax, [rsp+200]
    37b7afc3fmov [rsp+184], rax
    37b7afc47jmp 0x37b7af0c2
    37b7afc4c
    37b7afc4cloc_37b7afc4c:
    37b7afc4cmov edx, -14
    37b7afc51mov eax, 14
    37b7afc56mov [rbx], edx
    37b7afc58jmp 0x37b7ae8eb
    37b7afc5d
    37b7afc5dloc_37b7afc5d:
    37b7afc5dmov eax, edx
    37b7afc5fneg eax
    37b7afc61jmp 0x37b7ae8eb
    37b7ae740