FilesBlogsAbout
Register
File lapack_564f5cd6dd718723530ccdd421a63387e76d5fcdb30f49d1f0425db7bc512296.dll
Function:
    37b718c80
    37b718c80strevc3_:
    37b718c80push r15
    37b718c82push r14
    37b718c84push r13
    37b718c86push r12
    37b718c88push rbp
    37b718c89push rdi
    37b718c8apush rsi
    37b718c8bpush rbx
    37b718c8csub rsp, 1128
    37b718c93movups [rsp+960], x6
    37b718c9bmovups [rsp+976], x7
    37b718ca3movups [rsp+992], x8
    37b718cacmovups [rsp+1008], x9
    37b718cb5movups [rsp+1024], x10
    37b718cbemovups [rsp+1040], x11
    37b718cc7movups [rsp+1056], x12
    37b718cd0movups [rsp+1072], x13
    37b718cd9movups [rsp+1088], x14
    37b718ce2movups [rsp+1104], x15
    37b718cebmov rax, [rsp+1240]
    37b718cf3mov r15, [rsp+1296]
    37b718cfbmov r13l, [rax]
    37b718cfemov rax, [rsp+1272]
    37b718d06mov r12l, [rax]
    37b718d09mov rsi, rdx
    37b718d0cmov rbx, rcx
    37b718d0fmov r14, r9
    37b718d12mov rax, [rsp+1256]
    37b718d1amov [rsp+1216], r8
    37b718d22mov r9l, 1
    37b718d28mov r8l, 1
    37b718d2elea rdx, [rip+0xd1d56b]
    37b718d35mov eax, [rax]
    37b718d37mov [rsp+224], eax
    37b718d3ecall lsame_
    37b718d43mov r9l, 1
    37b718d49mov r8l, 1
    37b718d4fmov rcx, rbx
    37b718d52lea rdx, [rip+0xd1d548]
    37b718d59mov edi, eax
    37b718d5bcall lsame_
    37b718d60mov r9l, 1
    37b718d66mov r8l, 1
    37b718d6cmov rcx, rbx
    37b718d6for eax, edi
    37b718d71lea rdx, [rip+0xd1d52a]
    37b718d78mov ebp, eax
    37b718d7acall lsame_
    37b718d7fmov rcx, rsi
    37b718d82mov r9l, 1
    37b718d88mov r8l, 1
    37b718d8eor eax, edi
    37b718d90lea rdx, [rip+0xd1d50c]
    37b718d97mov [rsp+216], eax
    37b718d9ecall lsame_
    37b718da3mov rcx, rsi
    37b718da6mov r9l, 1
    37b718dacmov r8l, 1
    37b718db2lea rdx, [rip+0xd1d4e7]
    37b718db9mov edi, eax
    37b718dbbcall lsame_
    37b718dc0mov rcx, rsi
    37b718dc3mov r9l, 1
    37b718dc9mov r8l, 1
    37b718dcflea rdx, [rip+0xd1d4ce]
    37b718dd6mov [rsp+268], eax
    37b718dddcall lsame_
    37b718de2mov rcx, [rsp+1312]
    37b718deamov r9, rbx
    37b718dedmov r8l, 1
    37b718df3mov [rsp+264], eax
    37b718dfalea rax, [rsp+448]
    37b718e02mov [rcx], 0
    37b718e08mov rdx, rax
    37b718e0bmov ecx, 2
    37b718e10mov [rsp+40], rsi
    37b718e15mov rsi, rax
    37b718e18mov [rsp+32], 1
    37b718e21mov [rsp+312], rax
    37b718e29call _gfortran_concat_string
    37b718e2elea rax, [rip+0xd1d47b]
    37b718e35mov r9, r14
    37b718e38mov r8, rsi
    37b718e3bmov [rsp+64], 2
    37b718e44lea rdx, [rip+0xd1d45a]
    37b718e4blea rcx, [rip+0xd1d45a]
    37b718e52mov [rsp+56], 6
    37b718e5bmov [rsp+48], rax
    37b718e60mov [rsp+40], rax
    37b718e65mov [rsp+32], rax
    37b718e6acall ilaenv_
    37b718e6fmov edx, [r14]
    37b718e72mov ecx, 1
    37b718e77pxor x0, x0
    37b718e7bimul eax, edx
    37b718e7elea eax, [rdx+2*rax]
    37b718e81test eax, eax
    37b718e83cmovle eax, ecx
    37b718e86cvtsi2ss x0, eax
    37b718e8amovss [r15], x0
    37b718e8ftest ebp, ebp
    37b718e91jne 0x37b718f50
    37b718e97mov r8l, [rsp+216]
    37b718e9fmov ecx, -1
    37b718ea4mov eax, 1
    37b718ea9test r8l, r8l
    37b718eacjne 0x37b718f50
    37b718eb2
    37b718eb2loc_37b718eb2:
    37b718eb2mov rdi, [rsp+1312]
    37b718ebamov [rdi], ecx
    37b718ebc
    37b718ebcloc_37b718ebc:
    37b718ebcmov rdx, [rsp+312]
    37b718ec4mov r8l, 7
    37b718ecalea rcx, [rip+0xd1d3e3]
    37b718ed1mov [rsp+448], eax
    37b718ed8call xerbla_
    37b718eddnop
    37b718ede
    37b718edeloc_37b718ede:
    37b718edemovups x6, [rsp+960]
    37b718ee6movups x7, [rsp+976]
    37b718eeemovups x8, [rsp+992]
    37b718ef7movups x9, [rsp+1008]
    37b718f00movups x10, [rsp+1024]
    37b718f09movups x11, [rsp+1040]
    37b718f12movups x12, [rsp+1056]
    37b718f1bmovups x13, [rsp+1072]
    37b718f24movups x14, [rsp+1088]
    37b718f2dmovups x15, [rsp+1104]
    37b718f36add rsp, 1128
    37b718f3dpop rbx
    37b718f3epop rsi
    37b718f3fpop rdi
    37b718f40pop rbp
    37b718f41pop r12
    37b718f43pop r13
    37b718f45pop r14
    37b718f47pop r15
    37b718f49ret
    37b718f4a
    37b718f4a 66 0f 1f 44 00 00 f..D..
    37b718f50
    37b718f50loc_37b718f50:
    37b718f50or edi, [rsp+268]
    37b718f57jne 0x37b718f68
    37b718f59mov eax, [rsp+264]
    37b718f60test eax, eax
    37b718f62je 0x37b719018
    37b718f68
    37b718f68loc_37b718f68:
    37b718f68test edx, edx
    37b718f6ajs 0x37b71903f
    37b718f70mov rdi, [rsp+1240]
    37b718f78mov eax, 1
    37b718f7dcmovg eax, edx
    37b718f80cmp [rdi], eax
    37b718f82jl 0x37b719030
    37b718f88mov rax, [rsp+1256]
    37b718f90mov eax, [rax]
    37b718f92test eax, eax
    37b718f94jle 0x37b719057
    37b718f9acmp edx, eax
    37b718f9cjle 0x37b718fac
    37b718f9etest [rsp+216], 1
    37b718fa6jne 0x37b719075
    37b718fac
    37b718facloc_37b718fac:
    37b718facmov rax, [rsp+1272]
    37b718fb4mov eax, [rax]
    37b718fb6test eax, eax
    37b718fb8jle 0x37b719066
    37b718fbecmp edx, eax
    37b718fc0jle 0x37b718fcc
    37b718fc2test bpl, 1
    37b718fc6jne 0x37b71d825
    37b718fcc
    37b718fccloc_37b718fcc:
    37b718fccmov rax, [rsp+1304]
    37b718fd4lea r10l, [rdx+rdx]
    37b718fd8mov ecx, 1
    37b718fddmov r8l, [rax]
    37b718fe0lea eax, [r10+rdx]
    37b718fe4test eax, eax
    37b718fe6cmovle eax, ecx
    37b718fe9cmp r8l, eax
    37b718fecjge 0x37b71908d
    37b718ff2cmp r8l, -1
    37b718ff6je 0x37b71908d
    37b718ffcmov rax, [rsp+1312]
    37b719004mov [rax], -14
    37b71900amov eax, 14
    37b71900fjmp 0x37b718ebc
    37b719014
    37b719014 0f 1f 40 00 ..@.
    37b719018
    37b719018loc_37b719018:
    37b719018mov rax, [rsp+1312]
    37b719020mov [rax], -2
    37b719026mov eax, 2
    37b71902bjmp 0x37b718ebc
    37b719030
    37b719030loc_37b719030:
    37b719030mov ecx, -6
    37b719035mov eax, 6
    37b71903ajmp 0x37b718eb2
    37b71903f
    37b71903floc_37b71903f:
    37b71903fmov rax, [rsp+1312]
    37b719047mov [rax], -4
    37b71904dmov eax, 4
    37b719052jmp 0x37b718ebc
    37b719057
    37b719057loc_37b719057:
    37b719057mov ecx, -8
    37b71905cmov eax, 8
    37b719061jmp 0x37b718eb2
    37b719066
    37b719066loc_37b719066:
    37b719066mov ecx, -10
    37b71906bmov eax, 10
    37b719070jmp 0x37b718eb2
    37b719075
    37b719075loc_37b719075:
    37b719075mov rax, [rsp+1312]
    37b71907dmov [rax], -8
    37b719083mov eax, 8
    37b719088jmp 0x37b718ebc
    37b71908d
    37b71908dloc_37b71908d:
    37b71908dmovsxd rsi, r13l
    37b719090xor eax, eax
    37b719092mov r13l, [rsp+264]
    37b71909atest rsi, rsi
    37b71909dcmovns rax, rsi
    37b7190a1mov [rsp+208], rax
    37b7190a9test r13l, r13l
    37b7190acje 0x37b719136
    37b7190b2mov rdi, [rsp+1288]
    37b7190bamov [rdi], 0
    37b7190c0test edx, edx
    37b7190c2je 0x37b71d8a7
    37b7190c8lea r11, [4*rax+4]
    37b7190d0mov rax, [rsp+1232]
    37b7190d8mov ecx, 1
    37b7190ddpxor x0, x0
    37b7190e1lea r9, [rax-4]
    37b7190e5movsxd rax, edx
    37b7190e8lea rbx, [rax-1]
    37b7190ecmov rax, [rsp+1216]
    37b7190f4jmp 0x37b719117
    37b7190f6
    37b7190f6loc_37b7190f6:
    37b7190f6mov esi, [rax+4*rbx]
    37b7190f9test esi, esi
    37b7190fbje 0x37b719108
    37b7190fd
    37b7190fdloc_37b7190fd:
    37b7190fdmov rsi, [rsp+1288]
    37b719105add [rsi], 1
    37b719108
    37b719108loc_37b719108:
    37b719108add rcx, 1
    37b71910cadd r9, r11
    37b71910fcmp edx, ecx
    37b719111jl 0x37b71d816
    37b719117
    37b719117loc_37b719117:
    37b719117cmp edx, ecx
    37b719119jle 0x37b7190f6
    37b71911bucomiss x0, [r9+8]
    37b719120mov esi, [rax+4*rcx-4]
    37b719124jp 0x37b71d83d
    37b71912ajne 0x37b71d83d
    37b719130test esi, esi
    37b719132je 0x37b719108
    37b719134jmp 0x37b7190fd
    37b719136
    37b719136loc_37b719136:
    37b719136mov rax, [rsp+1288]
    37b71913emov [rax], edx
    37b719140mov eax, edx
    37b719142
    37b719142loc_37b719142:
    37b719142mov rcx, [rsp+1280]
    37b71914acmp [rcx], eax
    37b71914cjge 0x37b719160
    37b71914emov ecx, -11
    37b719153mov eax, 11
    37b719158jmp 0x37b718eb2
    37b71915d
    37b71915d 0f 1f 00 ...
    37b719160
    37b719160loc_37b719160:
    37b719160mov rax, [rsp+1312]
    37b719168mov eax, [rax]
    37b71916atest eax, eax
    37b71916cjne 0x37b71d8ae
    37b719172cmp r8l, -1
    37b719176je 0x37b718ede
    37b71917ctest edx, edx
    37b71917eje 0x37b718ede
    37b719184mov rax, [rsp+208]
    37b71918cmov ecx, [rsp+268]
    37b719193mov [rsp+236], 1
    37b71919enot rax
    37b7191a1mov [rsp+240], rax
    37b7191a9test ecx, ecx
    37b7191abje 0x37b7191bd
    37b7191admov eax, edx
    37b7191afshl eax, 4
    37b7191b2add eax, edx
    37b7191b4cmp r8l, eax
    37b7191b7jge 0x37b71cd00
    37b7191bd
    37b7191bdloc_37b7191bd:
    37b7191bdmov edx, 12
    37b7191c2lea rcx, [rip+0xd1d0f8]
    37b7191c9pxor x13, x13
    37b7191cecall slamch_
    37b7191d3mov edx, 9
    37b7191d8lea rcx, [rip+0xd1d0ee]
    37b7191dfmovaps x9, x0
    37b7191e3call slamch_
    37b7191e8mov r13l, [r14]
    37b7191ebmov [r15], 0
    37b7191f2movss [rsp+304], x0
    37b7191fbcvtsi2ss x13, r13l
    37b719200divss x13, x0
    37b719205mulss x13, x9
    37b71920amovss x9, [rip+0xd1d0d9]
    37b719213movaps x14, x9
    37b719217subss x14, x0
    37b71921cdivss x14, x13
    37b719221cmp r13l, 1
    37b719225jle 0x37b71d6fb
    37b71922bmov rax, [rsp+240]
    37b719233mov rbx, [rsp+208]
    37b71923blea r9l, [r13-2]
    37b71923flea rcx, [r15+4]
    37b719243mov rdi, [rsp+1232]
    37b71924blea r10, [r15+4*r9+8]
    37b719250lea r8, [rax+2*rbx]
    37b719254lea r11, [4*rbx+4]
    37b71925clea rax, [4*r8]
    37b719264lea r9, [rdi+4]
    37b719268lea rdx, [rdi+rax+8]
    37b71926dnop [rax]
    37b719270
    37b719270loc_37b719270:
    37b719270add rax, r9
    37b719273pxor x1, x1
    37b719277nop [rax+rax]
    37b719280
    37b719280loc_37b719280:
    37b719280movss x0, [rax]
    37b719284andps x0, [rip+0xd1d065]
    37b71928badd rax, 4
    37b71928faddss x1, x0
    37b719293cmp rax, rdx
    37b719296jne 0x37b719280
    37b719298movss [rcx], x1
    37b71929cadd rcx, 4
    37b7192a0add r8, rbx
    37b7192a3lea rdx, [rax+r11]
    37b7192a7cmp rcx, r10
    37b7192aaje 0x37b7192c0
    37b7192aclea rax, [4*r8]
    37b7192b4jmp 0x37b719270
    37b7192b6
    37b7192b6 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b7192c0
    37b7192c0loc_37b7192c0:
    37b7192c0test ebp, ebp
    37b7192c2je 0x37b719b23
    37b7192c8cmp [rsp+236], 2
    37b7192d0jg 0x37b71c9df
    37b7192d6mov rax, [rsp+1288]
    37b7192demov [rsp+392], r13l
    37b7192e6mov [rsp+388], 2
    37b7192f1mov eax, [rax]
    37b7192f3mov [rsp+232], eax
    37b7192fa
    37b7192faloc_37b7192fa:
    37b7192famovss x12, [rip+0xd1cfed]
    37b719303movsxd rsi, r12l
    37b719306xor eax, eax
    37b719308movss x10, [rip+0xd1cfef]
    37b719311test rsi, rsi
    37b719314cmovns rax, rsi
    37b719318mov [rsp+256], rax
    37b719320not rax
    37b719323mov [rsp+272], rax
    37b71932bmov rax, [rsp+208]
    37b719333neg rax
    37b719336shl rax, 2
    37b71933amov [rsp+144], rax
    37b719342mov rax, [rsp+240]
    37b71934ashl rax, 2
    37b71934emov [rsp+152], rax
    37b719356jmp 0x37b7193c0
    37b719358
    37b719358 0f 1f 84 00 00 00 00 00 ........
    37b719360
    37b719360loc_37b719360:
    37b719360mov edx, [rsp+264]
    37b719367lea edi, [r13-2]
    37b71936btest edx, edx
    37b71936dje 0x37b71b2b0
    37b719373mov rsi, [rsp+1216]
    37b71937bmovsxd rdx, edi
    37b71937emov r12l, [rsi+4*rdx]
    37b719382test r12l, r12l
    37b719385jne 0x37b71b2b0
    37b71938bmov eax, [rsp+392]
    37b719392sub eax, 1
    37b719395mov [rsp+392], eax
    37b71939ctest eax, eax
    37b71939ejle 0x37b719b23
    37b7193a4
    37b7193a4loc_37b7193a4:
    37b7193a4mov eax, [rsp+392]
    37b7193ablea r13l, [rax-1]
    37b7193afmov [rsp+392], r13l
    37b7193b7test r13l, r13l
    37b7193bajle 0x37b719b23
    37b7193c0
    37b7193c0loc_37b7193c0:
    37b7193c0lea rax, [rsp+408]
    37b7193c8mov [rsp+176], rax
    37b7193d0lea rax, [rsp+384]
    37b7193d8mov [rsp+184], rax
    37b7193e0
    37b7193e0loc_37b7193e0:
    37b7193e0cmp r13l, 1
    37b7193e4je 0x37b71b2a0
    37b7193eamovsxd r9, r13l
    37b7193edmov rdx, r9
    37b7193f0
    37b7193f0loc_37b7193f0:
    37b7193f0mov r8, [rsp+208]
    37b7193f8lea eax, [r13-1]
    37b7193fcmov rdi, [rsp+240]
    37b719404pxor x0, x0
    37b719408movsxd rcx, eax
    37b71940bimul r8, rcx
    37b71940flea rbp, [r8+rdi]
    37b719413mov rdi, [rsp+1232]
    37b71941blea r10, [rbp+r9]
    37b719420movss x1, [rdi+4*r10]
    37b719426ucomiss x1, x0
    37b719429jp 0x37b719360
    37b71942fjne 0x37b719360
    37b719435
    37b719435loc_37b719435:
    37b719435mov eax, [rsp+264]
    37b71943ctest eax, eax
    37b71943ejne 0x37b719a20
    37b719444
    37b719444loc_37b719444:
    37b719444mov r9, [rsp+208]
    37b71944cpxor x1, x1
    37b719450mov r10l, [r14]
    37b719453movaps x11, x12
    37b719457mov rax, [rsp+1232]
    37b71945fmov esi, [rsp+388]
    37b719466imul r9, rdx
    37b71946aadd r9, [rsp+240]
    37b719472add rdx, r9
    37b719475imul r10l, esi
    37b719479movss x0, [rax+4*rdx]
    37b71947emov [rsp+412], 0
    37b719489movss [rsp+416], x0
    37b719492andps x0, x12
    37b719496lea eax, [r10+r13-1]
    37b71949baddss x0, x1
    37b71949fcdqe
    37b7194a1mulss x0, [rsp+304]
    37b7194aamovss [r15+4*rax], x9
    37b7194b0maxss x0, x13
    37b7194b5movss [rsp+408], x0
    37b7194becmp r13l, 1
    37b7194c2je 0x37b719a77
    37b7194c8lea eax, [r13-2]
    37b7194cccmp eax, 2
    37b7194cfjbe 0x37b71d743
    37b7194d5lea eax, [r13-1]
    37b7194d9mov rdi, [rsp+1232]
    37b7194e1movsxd r10, r10l
    37b7194e4xor edx, edx
    37b7194e6mov ecx, eax
    37b7194e8movss x1, [rip+0xd1ce10]
    37b7194f0lea r8, [r15+4*r10]
    37b7194f4shr ecx, 2
    37b7194f7lea r11, [rdi+4*r9+4]
    37b7194fcshl rcx, 4
    37b719500shufps x1, x1, 0
    37b719504nop [rax+rax]
    37b71950anop [rax+rax]
    37b719515nop [rax+rax]
    37b719520
    37b719520loc_37b719520:
    37b719520movups x0, [r11+rdx]
    37b719525xorps x0, x1
    37b719528movups [r8+rdx], x0
    37b71952dadd rdx, 16
    37b719531cmp rcx, rdx
    37b719534jne 0x37b719520
    37b719536mov r12l, eax
    37b719539test al, 3
    37b71953bje 0x37b719574
    37b71953dmov edx, eax
    37b71953fand edx, -4
    37b719542add edx, 1
    37b719545
    37b719545loc_37b719545:
    37b719545mov rdi, [rsp+1232]
    37b71954dmovsxd rdx, edx
    37b719550lea r8, [r15+4*r10]
    37b719554lea rcx, [rdi+4*r9]
    37b719558
    37b719558loc_37b719558:
    37b719558movss x0, [rcx+4*rdx]
    37b71955dxorps x0, x10
    37b719561movss [r8+4*rdx-4], x0
    37b719568add rdx, 1
    37b71956ccmp r13l, edx
    37b71956fjg 0x37b719558
    37b719571mov r12l, eax
    37b719574
    37b719574loc_37b719574:
    37b719574lea edx, [rax-1]
    37b719577movsxd r13, r13l
    37b71957amovsxd rbx, eax
    37b71957dmov [rsp+160], esi
    37b719584movsxd rdx, edx
    37b719587mov rdi, [rsp+240]
    37b71958fimul rdx, [rsp+208]
    37b719598mov rcx, [rsp+1232]
    37b7195a0lea rax, [rdx+rdi]
    37b7195a4lea rbp, [rcx+4*rax+4]
    37b7195a9lea rax, [r15+4*r13]
    37b7195adadd r13, rdi
    37b7195b0add r13, rdx
    37b7195b3mov rsi, rax
    37b7195b6lea rdi, [rcx+4*r13-4]
    37b7195bbjmp 0x37b7197a0
    37b7195c0
    37b7195c0loc_37b7195c0:
    37b7195c0lea rdx, [rsp+420]
    37b7195c8mov rax, [rsp+184]
    37b7195d0lea r11, [rsp+404]
    37b7195d8mov [rsp+80], r14
    37b7195ddmov [rsp+128], rdx
    37b7195e5lea rdx, [rsp+432]
    37b7195edlea r8, [rip+0xd1ccb8]
    37b7195f4mov r9, [rsp+176]
    37b7195fcmov [rsp+104], rdx
    37b719601lea rdx, [rsp+416]
    37b719609mov [rsp+88], rdx
    37b71960elea edx, [rcx+rbx-1]
    37b719612mov rcx, [rsp+1240]
    37b71961amov [rsp+136], rax
    37b719622movsxd rdx, edx
    37b719625lea rax, [rip+0xd1ccac]
    37b71962cmov [rsp+112], rax
    37b719631lea rdx, [r15+4*rdx]
    37b719635lea rax, [rip+0xd1cc80]
    37b71963cmov [rsp+96], rax
    37b719641lea rax, [rip+0xd1cc98]
    37b719648mov [rsp+72], rdx
    37b71964dmov rdx, rdi
    37b719650sub rdx, [rsp+144]
    37b719658mov [rsp+120], r11
    37b71965dmov [rsp+48], rcx
    37b719662lea rcx, [rip+0xd1cc73]
    37b719669mov [rsp+40], rdx
    37b71966emov rdx, r8
    37b719671mov [rsp+64], rax
    37b719676mov [rsp+56], rax
    37b71967bmov [rsp+32], rax
    37b719680mov [rsp+168], r11
    37b719688movups [rsp+432], x15
    37b719691call slaln2_
    37b719696mov r11, [rsp+168]
    37b71969emovss x0, [rsp+420]
    37b7196a7movss x15, [rsp+432]
    37b7196b1movss x7, [rsp+436]
    37b7196bamovss x6, [rsp+440]
    37b7196c3comiss x0, x9
    37b7196c7movss x8, [rsp+444]
    37b7196d1movss x1, [rsp+404]
    37b7196dajbe 0x37b719700
    37b7196dcmovaps x2, x14
    37b7196e0movss x3, [rsi-8]
    37b7196e5divss x2, x0
    37b7196e9comiss x3, x2
    37b7196ecjbe 0x37b719700
    37b7196eedivss x1, x0
    37b7196f2divss x15, x0
    37b7196f7movss [rsp+404], x1
    37b719700
    37b719700loc_37b719700:
    37b719700ucomiss x1, x9
    37b719704jp 0x37b71bf6a
    37b71970ajne 0x37b71bf6a
    37b719710
    37b719710loc_37b719710:
    37b719710mov r8l, [rsp+160]
    37b719718imul r8l, [r14]
    37b71971clea rax, [rip+0xd1cb89]
    37b719723movaps x0, x15
    37b719727xorps x0, x10
    37b71972blea rcx, [rsp+424]
    37b719733mov r9, rax
    37b719736mov [rsp+424], r13l
    37b71973emovss [rsp+428], x0
    37b719747lea edx, [rbx+r8-1]
    37b71974cmovsxd r8, r8l
    37b71974fmovsxd rdx, edx
    37b719752lea r8, [r15+4*r8]
    37b719756movss [r15+4*rdx], x15
    37b71975clea rdx, [rsp+428]
    37b719764mov [rsp+32], r8
    37b719769mov r8, rbp
    37b71976csub r8, [rsp+144]
    37b719774mov [rsp+40], rax
    37b719779call saxpy_
    37b71977enop
    37b719780
    37b719780loc_37b719780:
    37b719780sub rbx, 1
    37b719784add rbp, [rsp+144]
    37b71978csub rsi, 4
    37b719790add rdi, [rsp+152]
    37b719798test ebx, ebx
    37b71979aje 0x37b719a70
    37b7197a0
    37b7197a0loc_37b7197a0:
    37b7197a0cmp r12l, ebx
    37b7197a3jl 0x37b719780
    37b7197a5unpcklps x6, x8
    37b7197a9unpcklps x15, x7
    37b7197adlea r13l, [rbx-1]
    37b7197b1mov edx, ebx
    37b7197b3mov ecx, [rsp+160]
    37b7197bamov [rsp+168], ebx
    37b7197c1movlhps x15, x6
    37b7197c5imul ecx, [r14]
    37b7197c9cmp ebx, 1
    37b7197ccje 0x37b71afc0
    37b7197d2pxor x0, x0
    37b7197d6ucomiss x0, [rdi]
    37b7197d9jp 0x37b7197e4
    37b7197dbmov r12l, r13l
    37b7197deje 0x37b7195c0
    37b7197e4
    37b7197e4loc_37b7197e4:
    37b7197e4lea r12l, [rdx-2]
    37b7197e8lea rdx, [rsp+420]
    37b7197f0mov [rsp+80], r14
    37b7197f5mov rax, [rsp+184]
    37b7197fdmov [rsp+128], rdx
    37b719805lea rdx, [rsp+432]
    37b71980dlea r11, [rsp+404]
    37b719815mov r9, [rsp+176]
    37b71981dmov [rsp+104], rdx
    37b719822lea rdx, [rsp+416]
    37b71982alea r8, [rip+0xd1ca7b]
    37b719831mov [rsp+88], rdx
    37b719836lea edx, [rcx+r13-1]
    37b71983bmov rcx, [rsp+1240]
    37b719843mov [rsp+136], rax
    37b71984bmovsxd rdx, edx
    37b71984elea rax, [rip+0xd1ca83]
    37b719855mov [rsp+112], rax
    37b71985alea rdx, [r15+4*rdx]
    37b71985elea rax, [rip+0xd1ca57]
    37b719865mov [rsp+96], rax
    37b71986alea rax, [rip+0xd1ca6f]
    37b719871mov [rsp+72], rdx
    37b719876lea rdx, [rdi-4]
    37b71987amov [rsp+120], r11
    37b71987fmov [rsp+48], rcx
    37b719884lea rcx, [rip+0xd1ca51]
    37b71988bmov [rsp+40], rdx
    37b719890lea rdx, [rip+0xd1ca41]
    37b719897mov [rsp+64], rax
    37b71989cmov [rsp+56], rax
    37b7198a1mov [rsp+32], rax
    37b7198a6mov [rsp+168], r11
    37b7198aemovups [rsp+432], x15
    37b7198b7call slaln2_
    37b7198bcmov r11, [rsp+168]
    37b7198c4movss x0, [rsp+420]
    37b7198cdmovss x15, [rsp+432]
    37b7198d7movss x7, [rsp+436]
    37b7198e0movss x6, [rsp+440]
    37b7198e9comiss x0, x9
    37b7198edmovss x8, [rsp+444]
    37b7198f7movss x1, [rsp+404]
    37b719900jbe 0x37b71992f
    37b719902movaps x2, x14
    37b719906movss x3, [rsi-12]
    37b71990bmaxss x3, [rsi-8]
    37b719910divss x2, x0
    37b719914comiss x3, x2
    37b719917jbe 0x37b71992f
    37b719919divss x1, x0
    37b71991ddivss x15, x0
    37b719922movss [rsp+404], x1
    37b71992bdivss x7, x0
    37b71992f
    37b71992floc_37b71992f:
    37b71992fucomiss x1, x9
    37b719933jp 0x37b71b0f0
    37b719939jne 0x37b71b0f0
    37b71993f
    37b71993floc_37b71993f:
    37b71993fmov ecx, [rsp+160]
    37b719946imul ecx, [r14]
    37b71994amovaps x0, x15
    37b71994emov r8, rbp
    37b719951lea rax, [rip+0xd1c954]
    37b719958xorps x0, x10
    37b71995cmov [rsp+424], r12l
    37b719964movss [rsp+428], x0
    37b71996dmov r9, rax
    37b719970lea edx, [rcx+r13-1]
    37b719975movsxd rdx, edx
    37b719978movss [r15+4*rdx], x15
    37b71997elea edx, [rcx+rbx-1]
    37b719982movsxd rcx, ecx
    37b719985lea rcx, [r15+4*rcx]
    37b719989movsxd rdx, edx
    37b71998cmovss [r15+4*rdx], x7
    37b719992lea rdx, [rsp+428]
    37b71999amov [rsp+40], rax
    37b71999fmov [rsp+32], rcx
    37b7199a4lea rcx, [rsp+424]
    37b7199acmov [rsp+168], rdx
    37b7199b4call saxpy_
    37b7199b9lea rax, [rip+0xd1c8ec]
    37b7199c0movaps x0, x7
    37b7199c3mov ecx, [rsp+160]
    37b7199camov [rsp+40], rax
    37b7199cfimul ecx, [r14]
    37b7199d3xorps x0, x10
    37b7199d7mov r8, rbp
    37b7199damov rdx, [rsp+168]
    37b7199e2mov r9, rax
    37b7199e5sub r8, [rsp+144]
    37b7199edmov [rsp+424], r12l
    37b7199f5movss [rsp+428], x0
    37b7199femovsxd rcx, ecx
    37b719a01lea rcx, [r15+4*rcx]
    37b719a05mov [rsp+32], rcx
    37b719a0alea rcx, [rsp+424]
    37b719a12call saxpy_
    37b719a17jmp 0x37b719780
    37b719a1c
    37b719a1c 0f 1f 40 00 ..@.
    37b719a20
    37b719a20loc_37b719a20:
    37b719a20mov rax, [rsp+1216]
    37b719a28mov eax, [rax+4*rdx-4]
    37b719a2ctest eax, eax
    37b719a2ejne 0x37b719444
    37b719a34mov eax, [rsp+392]
    37b719a3blea r13l, [rax-1]
    37b719a3fmov [rsp+392], r13l
    37b719a47test r13l, r13l
    37b719a4ajle 0x37b719b23
    37b719a50mov edx, 1
    37b719a55cmp r13l, 1
    37b719a59je 0x37b719a20
    37b719a5bmovsxd rdx, r13l
    37b719a5emov r9, rdx
    37b719a61jmp 0x37b7193f0
    37b719a66
    37b719a66 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b719a70
    37b719a70loc_37b719a70:
    37b719a70mov esi, [rsp+160]
    37b719a77
    37b719a77loc_37b719a77:
    37b719a77movsxd rax, [rsp+232]
    37b719a7fmov ebp, [rsp+268]
    37b719a86lea edi, [rax-1]
    37b719a89test ebp, ebp
    37b719a8bje 0x37b71d418
    37b719a91mov ebx, [rsp+392]
    37b719a98cmp [rsp+236], 1
    37b719aa0jne 0x37b71c081
    37b719aa6mov rax, [rsp+1264]
    37b719aaemovsxd r12, ebx
    37b719ab1imul r12, [rsp+256]
    37b719abaadd r12, [rsp+272]
    37b719ac2lea rbp, [rax+4*r12+4]
    37b719ac7cmp ebx, 1
    37b719acajg 0x37b71bf98
    37b719ad0lea r8, [rip+0xd1c7d5]
    37b719ad7mov rdx, rbp
    37b719adamov rcx, r14
    37b719addcall isamax_
    37b719ae2movaps x1, x9
    37b719ae6mov r8, rbp
    37b719ae9mov rcx, r14
    37b719aecmov rdi, [rsp+1264]
    37b719af4cdqe
    37b719af6lea rdx, [rsp+400]
    37b719afelea r9, [rip+0xd1c7a7]
    37b719b05add rax, r12
    37b719b08movss x0, [rdi+4*rax]
    37b719b0dandps x0, x12
    37b719b11divss x1, x0
    37b719b15movss [rsp+400], x1
    37b719b1ecall sscal_
    37b719b23
    37b719b23loc_37b719b23:
    37b719b23mov ebx, [rsp+216]
    37b719b2atest ebx, ebx
    37b719b2cje 0x37b718ede
    37b719b32mov [rsp+388], 1
    37b719b3dmov eax, [r14]
    37b719b40mov [rsp+392], 1
    37b719b4bmov [rsp+308], eax
    37b719b52test eax, eax
    37b719b54jle 0x37b718ede
    37b719b5amovsxd rdi, [rsp+224]
    37b719b62xor eax, eax
    37b719b64mov [rsp+176], 3
    37b719b6fmov r12l, 2
    37b719b75mov [rsp+248], -1
    37b719b81movss x10, [rip+0xd1c766]
    37b719b8amovss x11, [rip+0xd1c76d]
    37b719b93mov [rsp+216], 1
    37b719b9ftest rdi, rdi
    37b719ba2mov [rsp+328], 1
    37b719badcmovns rax, rdi
    37b719bb1mov [rsp+272], rax
    37b719bb9not rax
    37b719bbcmov [rsp+376], rax
    37b719bc4mov rax, [rsp+208]
    37b719bcclea rdi, [4*rax+4]
    37b719bd4shl rax, 2
    37b719bd8mov [rsp+168], rdi
    37b719be0mov rdi, [rsp+1232]
    37b719be8mov [rsp+184], rax
    37b719bf0lea rsi, [rdi-4]
    37b719bf4xor edi, edi
    37b719bf6mov [rsp+256], rsi
    37b719bfelea rsi, [r15+4]
    37b719c02mov [rsp+352], rsi
    37b719c0amov [rsp+192], rsi
    37b719c12jmp 0x37b719d04
    37b719c17
    37b719c17 66 0f 1f 84 00 00 00 00 00 f........
    37b719c20
    37b719c20loc_37b719c20:
    37b719c20lea eax, [r12-1]
    37b719c25movsxd rcx, [r14]
    37b719c28mov [rsp+160], eax
    37b719c2fmov rax, [rsp+216]
    37b719c37mov [rsp+224], ecx
    37b719c3emov [rsp+332], eax
    37b719c45cmp ecx, eax
    37b719c47je 0x37b71afd0
    37b719c4dmov rdi, [rsp+256]
    37b719c55mov r11l, [rsp+264]
    37b719c5dmovss x0, [rdi+8]
    37b719c62test r11l, r11l
    37b719c65je 0x37b719d40
    37b719c6bmov rdi, [rsp+1216]
    37b719c73mov r10l, [rdi+4*rax-4]
    37b719c78test r10l, r10l
    37b719c7bjne 0x37b719d40
    37b719c81pxor x1, x1
    37b719c85xor edi, edi
    37b719c87lea r13, [r12+1]
    37b719c8cmov eax, 1
    37b719c91ucomiss x0, x1
    37b719c94mov rsi, [rsp+248]
    37b719c9csetp dil
    37b719ca0cmovne edi, eax
    37b719ca3mov rax, [rsp+272]
    37b719cablea rbx, [rax+rsi]
    37b719cafnop
    37b719cb0
    37b719cb0loc_37b719cb0:
    37b719cb0mov eax, [rsp+232]
    37b719cb7mov r12, r13
    37b719cbamov rsi, [rsp+168]
    37b719cc2mov [rsp+248], rbx
    37b719ccaadd [rsp+216], 1
    37b719cd3mov [rsp+392], eax
    37b719cdalea eax, [r13-1]
    37b719cdeadd [rsp+256], rsi
    37b719ce6add [rsp+192], 4
    37b719cefadd [rsp+176], 1
    37b719cf7cmp [rsp+308], eax
    37b719cfejl 0x37b718ede
    37b719d04
    37b719d04loc_37b719d04:
    37b719d04mov [rsp+232], r12l
    37b719d0ccmp edi, 1
    37b719d0fjne 0x37b719c20
    37b719d15mov rax, [rsp+272]
    37b719d1dmov edi, -1
    37b719d22lea r13, [r12+1]
    37b719d27mov rsi, [rsp+248]
    37b719d2flea rbx, [rax+rsi]
    37b719d33jmp 0x37b719cb0
    37b719d38
    37b719d38 0f 1f 84 00 00 00 00 00 ........
    37b719d40
    37b719d40loc_37b719d40:
    37b719d40mov rax, [rsp+256]
    37b719d48pxor x1, x1
    37b719d4cmov [rsp+412], 0
    37b719d57ucomiss x0, x1
    37b719d5amovss x5, [rax+4]
    37b719d5fmovss [rsp+416], x5
    37b719d68jp 0x37b719d70
    37b719d6aje 0x37b71a711
    37b719d70
    37b719d70loc_37b719d70:
    37b719d70mov edi, [rsp+388]
    37b719d77mov esi, [rsp+160]
    37b719d7emovaps x1, x0
    37b719d81andps x5, x10
    37b719d85mov rbx, [rsp+208]
    37b719d8dandps x1, x10
    37b719d91movups [rsp+336], x10
    37b719d9amov eax, edi
    37b719d9cmovaps x12, x1
    37b719da0mov [rsp+152], edi
    37b719da7add edi, 1
    37b719daaimul eax, ecx
    37b719dadsqrtss x12, x12
    37b719db2mov [rsp+200], edi
    37b719db9lea r10l, [rsi+rax-1]
    37b719dbemov rsi, [rsp+256]
    37b719dc6lea r8l, [rcx+rax]
    37b719dcalea r9l, [r8+r12-1]
    37b719dcfmovsxd r10, r10l
    37b719dd2movss x3, [rsi+4*rbx+4]
    37b719dd8movsxd r9, r9l
    37b719ddbmovaps x2, x3
    37b719ddeandps x2, x10
    37b719de2movaps x4, x2
    37b719de5comiss x2, x1
    37b719de8sqrtss x4, x4
    37b719decmulss x12, x4
    37b719df1addss x5, x12
    37b719df6mulss x5, [rsp+304]
    37b719dffmaxss x5, x13
    37b719e04jb 0x37b71bcdd
    37b719e0amovaps x1, x12
    37b719e0emovaps x0, x9
    37b719e12divss x1, x3
    37b719e16
    37b719e16loc_37b719e16:
    37b719e16mov edi, [rsp+160]
    37b719e1dlea edx, [rax+r12-1]
    37b719e22movss [r15+4*r10], x1
    37b719e28mov esi, [rsp+176]
    37b719e2fmovsxd rdx, edx
    37b719e32movss [r15+4*r9], x0
    37b719e38mov [r15+4*rdx], 0
    37b719e40lea edx, [rdi+r8-1]
    37b719e45movss [rsp+412], x12
    37b719e4fmovsxd rdx, edx
    37b719e52mov [r15+4*rdx], 0
    37b719e5amovss [rsp+408], x5
    37b719e63cmp ecx, esi
    37b719e65jl 0x37b71ae75
    37b719e6bmov rdx, [rsp+208]
    37b719e73mov r8l, esi
    37b719e76cdqe
    37b719e78mov rdi, [rsp+216]
    37b719e80mov rbx, [rsp+240]
    37b719e88mov rsi, [rsp+1232]
    37b719e90imul rdx, r8
    37b719e94lea r11, [rax+rdi]
    37b719e98mov rax, [rsp+352]
    37b719ea0add rbx, rdi
    37b719ea3lea rax, [rax+4*r11]
    37b719ea7add rdx, rbx
    37b719eaalea rdx, [rsi+4*rdx]
    37b719eaelea esi, [rcx-2]
    37b719eb1sub esi, edi
    37b719eb3mov rdi, [rsp+184]
    37b719ebbadd rsi, r11
    37b719ebemovsxd r11, ecx
    37b719ec1lea rsi, [r15+4*rsi+8]
    37b719ec6nop [rax]
    37b719ec9nop [rax+rax]
    37b719ed4nop [rax+rax]
    37b719edfnop [rax+rax]
    37b719eeanop [rax+rax]
    37b719ef5nop [rax+rax]
    37b719f00
    37b719f00loc_37b719f00:
    37b719f00movss x0, [r15+4*r10]
    37b719f06mulss x0, [rdx]
    37b719f0axorps x0, x11
    37b719f0emovss [rax], x0
    37b719f12movss x0, [r15+4*r9]
    37b719f18mulss x0, [rdx+4]
    37b719f1dadd rdx, rdi
    37b719f20xorps x0, x11
    37b719f24movss [rax+4*r11], x0
    37b719f2aadd rax, 4
    37b719f2ecmp rsi, rax
    37b719f31jne 0x37b719f00
    37b719f33mov r9, [rsp+208]
    37b719f3bxor r13l, r13l
    37b719f3emov rbp, r14
    37b719f41mov rax, [rsp+1232]
    37b719f49mov r10, [rsp+184]
    37b719f51mov [rsp+368], r12
    37b719f59movaps x2, x14
    37b719f5dxorps x12, x11
    37b719f61movss [rsp+360], x13
    37b719f6bimul r9, r8
    37b719f6fmovaps x13, x9
    37b719f73add rbx, r9
    37b719f76neg r9
    37b719f79lea rsi, [rax+4*rbx+8]
    37b719f7elea rax, [r8+1]
    37b719f82imul r10, rax
    37b719f86lea rdi, [r10+4*r9]
    37b719f8amov r10l, [rsp+176]
    37b719f92mov [rsp+320], rdi
    37b719f9amov rdi, rsi
    37b719f9djmp 0x37b71a2a8
    37b719fa2
    37b719fa2 66 0f 1f 44 00 00 f..D..
    37b719fa8
    37b719fa8loc_37b719fa8:
    37b719fa8comiss x0, x2
    37b719fabja 0x37b71bd6e
    37b719fb1lea rcx, [rsp+428]
    37b719fb9mov [rsp+144], rcx
    37b719fc1
    37b719fc1loc_37b719fc1:
    37b719fc1mov r14l, [rsp+152]
    37b719fc9mov [rsp+288], rax
    37b719fd1mov eax, [rsp+176]
    37b719fd8mov [rsp+428], r13l
    37b719fe0imul edx, r14l
    37b719fe4lea ecx, [rbx+rdx]
    37b719fe7lea edx, [rdx+rax-1]
    37b719febmovsxd rcx, ecx
    37b719feemovsxd rdx, edx
    37b719ff1movss x1, [r15+4*rcx-4]
    37b719ff8lea rcx, [rip+0xd1c2ad]
    37b719ffflea r9, [r15+4*rdx]
    37b71a003mov rdx, rsi
    37b71a006mov [rsp+32], rcx
    37b71a00bmov r8, rcx
    37b71a00emov rcx, [rsp+144]
    37b71a016movss [rsp+280], x1
    37b71a01fcall sdot_
    37b71a024mov r8l, [rbp]
    37b71a028mov edx, r14l
    37b71a02bmovss x1, [rsp+280]
    37b71a034mov eax, [rsp+176]
    37b71a03blea r10, [rip+0xd1c26a]
    37b71a042mov [rsp+428], r13l
    37b71a04alea r14, [rsp+404]
    37b71a052imul edx, r8l
    37b71a056subss x1, x0
    37b71a05alea ecx, [rbx+rdx]
    37b71a05dadd edx, r8l
    37b71a060mov r8, r10
    37b71a063movsxd rcx, ecx
    37b71a066movss [r15+4*rcx-4], x1
    37b71a06dlea ecx, [rbx+rdx]
    37b71a070lea edx, [rdx+rax-1]
    37b71a074movsxd rcx, ecx
    37b71a077movsxd rdx, edx
    37b71a07amovss x1, [r15+4*rcx-4]
    37b71a081mov [rsp+32], r10
    37b71a086lea r9, [r15+4*rdx]
    37b71a08amov rdx, rsi
    37b71a08dmov rcx, [rsp+144]
    37b71a095movss [rsp+280], x1
    37b71a09ecall sdot_
    37b71a0a3mov r8l, [rbp]
    37b71a0a7mov edx, [rsp+200]
    37b71a0aelea r10, [rip+0xd1c223]
    37b71a0b5mov rax, [rsp+144]
    37b71a0bdmovss x1, [rsp+280]
    37b71a0c6lea r9, [rsp+408]
    37b71a0cemovss [rsp+432], x15
    37b71a0d8movss [rsp+436], x7
    37b71a0e1imul edx, r8l
    37b71a0e5movss [rsp+440], x6
    37b71a0eesubss x1, x0
    37b71a0f2movss [rsp+444], x8
    37b71a0fcmovss [rsp+428], x12
    37b71a106lea ecx, [rbx+rdx]
    37b71a109sub edx, r8l
    37b71a10cmov r8, r10
    37b71a10fmovsxd rcx, ecx
    37b71a112add edx, ebx
    37b71a114movss [r15+4*rcx-4], x1
    37b71a11blea rcx, [rsp+384]
    37b71a123movsxd rdx, edx
    37b71a126mov [rsp+136], rcx
    37b71a12elea rcx, [rsp+420]
    37b71a136lea rdx, [r15+4*rdx-4]
    37b71a13bmov [rsp+128], rcx
    37b71a143lea rcx, [rsp+432]
    37b71a14bmov [rsp+104], rcx
    37b71a150lea rcx, [rsp+416]
    37b71a158mov [rsp+88], rcx
    37b71a15dmov rcx, [rsp+1240]
    37b71a165mov [rsp+96], rax
    37b71a16alea rax, [rip+0xd1c16f]
    37b71a171mov [rsp+72], rdx
    37b71a176lea rdx, [rip+0xd1c12f]
    37b71a17dmov [rsp+64], rax
    37b71a182mov [rsp+56], rax
    37b71a187mov [rsp+48], rcx
    37b71a18clea rcx, [rip+0xd1c149]
    37b71a193mov [rsp+32], rax
    37b71a198mov [rsp+80], rbp
    37b71a19dmov [rsp+40], rdi
    37b71a1a2mov [rsp+120], r14
    37b71a1a7mov [rsp+112], r10
    37b71a1accall slaln2_
    37b71a1b1mov rax, [rsp+288]
    37b71a1b9ucomiss x9, [rsp+404]
    37b71a1c2movss x15, [rsp+432]
    37b71a1ccmovss x7, [rsp+436]
    37b71a1d5movss x6, [rsp+440]
    37b71a1demovss x8, [rsp+444]
    37b71a1e8jp 0x37b71be8c
    37b71a1eejne 0x37b71be8c
    37b71a1f4
    37b71a1f4loc_37b71a1f4:
    37b71a1f4movsxd rcx, [rbp]
    37b71a1f8mov edx, [rsp+152]
    37b71a1ffmovaps x2, x6
    37b71a202mov r10l, r12l
    37b71a205andps x2, x10
    37b71a209imul edx, ecx
    37b71a20clea r8l, [rbx+rdx]
    37b71a210add edx, ecx
    37b71a212movsxd r8, r8l
    37b71a215add edx, ebx
    37b71a217sub r8, 1
    37b71a21bmovsxd rdx, edx
    37b71a21emovss [r15+4*r8], x15
    37b71a224movss [r15+4*rdx-4], x6
    37b71a22bmovss x0, [r15+4*r8]
    37b71a231andps x0, x10
    37b71a235maxss x0, x2
    37b71a239movaps x2, x14
    37b71a23dmaxss x13, x0
    37b71a242divss x2, x13
    37b71a247nop
    37b71a249nop [rax+rax]
    37b71a254nop [rax+rax]
    37b71a25fnop [rax+rax]
    37b71a26anop [rax+rax]
    37b71a275nop [rax+rax]
    37b71a280
    37b71a280loc_37b71a280:
    37b71a280add r13, 1
    37b71a284mov r8, rax
    37b71a287add rsi, [rsp+184]
    37b71a28fadd rdi, [rsp+168]
    37b71a297cmp [rsp+224], eax
    37b71a29ejl 0x37b71ae60
    37b71a2a4add rax, 1
    37b71a2a8
    37b71a2a8loc_37b71a2a8:
    37b71a2a8cmp r8l, r10l
    37b71a2abjl 0x37b71a280
    37b71a2admov r11, [rsp+192]
    37b71a2b5mov r10l, r8l
    37b71a2b8mov ebx, r8l
    37b71a2bbmov edx, ecx
    37b71a2bdmov r9l, [rsp+152]
    37b71a2c5lea r12l, [r8+1]
    37b71a2c9movss x0, [r11+4*r13+4]
    37b71a2d0imul r9l, ecx
    37b71a2d4cmp r8l, ecx
    37b71a2d7jge 0x37b719fa8
    37b71a2ddpxor x1, x1
    37b71a2e1ucomiss x1, [rdi+4]
    37b71a2e5jp 0x37b71a2ed
    37b71a2e7je 0x37b719fa8
    37b71a2ed
    37b71a2edloc_37b71a2ed:
    37b71a2edmov r11, [rsp+192]
    37b71a2f5add r10l, 2
    37b71a2f9maxss x0, [r11+4*r13+8]
    37b71a300comiss x0, x2
    37b71a303ja 0x37b71b1d0
    37b71a309lea rcx, [rsp+428]
    37b71a311mov [rsp+144], rcx
    37b71a319
    37b71a319loc_37b71a319:
    37b71a319mov r14l, [rsp+152]
    37b71a321mov [rsp+296], rax
    37b71a329mov eax, [rsp+176]
    37b71a330mov [rsp+288], r10l
    37b71a338imul edx, r14l
    37b71a33cmov [rsp+428], r13l
    37b71a344lea ecx, [rbx+rdx]
    37b71a347lea edx, [rdx+rax-1]
    37b71a34bmovsxd rcx, ecx
    37b71a34emovsxd rdx, edx
    37b71a351movss x1, [r15+4*rcx-4]
    37b71a358lea rcx, [rip+0xd1bf4d]
    37b71a35flea r9, [r15+4*rdx]
    37b71a363mov rdx, rsi
    37b71a366mov [rsp+32], rcx
    37b71a36bmov r8, rcx
    37b71a36emov rcx, [rsp+144]
    37b71a376movss [rsp+280], x1
    37b71a37fcall sdot_
    37b71a384mov r8l, [rbp]
    37b71a388mov edx, r14l
    37b71a38bmovss x1, [rsp+280]
    37b71a394mov eax, [rsp+176]
    37b71a39blea r11, [rip+0xd1bf0a]
    37b71a3a2mov [rsp+152], r14l
    37b71a3aaimul edx, r8l
    37b71a3aesubss x1, x0
    37b71a3b2mov [rsp+428], r13l
    37b71a3balea ecx, [rbx+rdx]
    37b71a3bdadd edx, r8l
    37b71a3c0mov r8, r11
    37b71a3c3movsxd rcx, ecx
    37b71a3c6movss [r15+4*rcx-4], x1
    37b71a3cdlea ecx, [rbx+rdx]
    37b71a3d0lea edx, [rdx+rax-1]
    37b71a3d4movsxd rcx, ecx
    37b71a3d7movsxd rdx, edx
    37b71a3damovss x1, [r15+4*rcx-4]
    37b71a3e1mov [rsp+32], r11
    37b71a3e6lea r9, [r15+4*rdx]
    37b71a3eamov rdx, rsi
    37b71a3edmov rcx, [rsp+144]
    37b71a3f5movss [rsp+280], x1
    37b71a3fecall sdot_
    37b71a403mov ecx, [rbp]
    37b71a406mov r10l, [rsp+200]
    37b71a40elea r11, [rip+0xd1be97]
    37b71a415movss x1, [rsp+280]
    37b71a41emov eax, [rsp+176]
    37b71a425mov [rsp+428], r13l
    37b71a42dimul r10l, ecx
    37b71a431mov r9, [rsp+320]
    37b71a439subss x1, x0
    37b71a43dlea r14, [rsi+r9]
    37b71a441mov edx, r10l
    37b71a444lea r8l, [rbx+r10]
    37b71a448sub edx, ecx
    37b71a44amovsxd r8, r8l
    37b71a44dlea ecx, [rdx+r12]
    37b71a451movss [r15+4*r8-4], x1
    37b71a458lea edx, [rdx+rax-1]
    37b71a45cmov r8, r11
    37b71a45fmovsxd rcx, ecx
    37b71a462movsxd rdx, edx
    37b71a465movss x1, [r15+4*rcx-4]
    37b71a46cmov [rsp+32], r11
    37b71a471lea r9, [r15+4*rdx]
    37b71a475mov rdx, r14
    37b71a478mov rcx, [rsp+144]
    37b71a480movss [rsp+280], x1
    37b71a489call sdot_
    37b71a48emov r8l, [rbp]
    37b71a492mov edx, [rsp+152]
    37b71a499lea r11, [rip+0xd1be0c]
    37b71a4a0movss x1, [rsp+280]
    37b71a4a9mov eax, [rsp+176]
    37b71a4b0mov [rsp+428], r13l
    37b71a4b8imul edx, r8l
    37b71a4bcsubss x1, x0
    37b71a4c0lea ecx, [rdx+r12]
    37b71a4c4add edx, r8l
    37b71a4c7mov r8, r11
    37b71a4camovsxd rcx, ecx
    37b71a4cdmovss [r15+4*rcx-4], x1
    37b71a4d4lea ecx, [rdx+r12]
    37b71a4d8lea edx, [rdx+rax-1]
    37b71a4dcmovsxd rcx, ecx
    37b71a4dfmovsxd rdx, edx
    37b71a4e2movss x1, [r15+4*rcx-4]
    37b71a4e9mov [rsp+32], r11
    37b71a4eelea r9, [r15+4*rdx]
    37b71a4f2mov rdx, r14
    37b71a4f5mov rcx, [rsp+144]
    37b71a4fdlea r14, [rsp+404]
    37b71a505movss [rsp+280], x1
    37b71a50ecall sdot_
    37b71a513mov ecx, [rbp]
    37b71a516mov edx, [rsp+200]
    37b71a51dlea r10, [rip+0xd1bdb4]
    37b71a524mov rax, [rsp+144]
    37b71a52cmovss x1, [rsp+280]
    37b71a535lea r9, [rsp+408]
    37b71a53dmovss [rsp+432], x15
    37b71a547movss [rsp+436], x7
    37b71a550imul edx, ecx
    37b71a553movss [rsp+440], x6
    37b71a55csubss x1, x0
    37b71a560movss [rsp+444], x8
    37b71a56amovss [rsp+428], x12
    37b71a574lea r8l, [rdx+r12]
    37b71a578sub edx, ecx
    37b71a57amov rcx, [rsp+1240]
    37b71a582movsxd r8, r8l
    37b71a585add edx, ebx
    37b71a587movss [r15+4*r8-4], x1
    37b71a58emovsxd rdx, edx
    37b71a591lea r8, [rsp+384]
    37b71a599mov [rsp+136], r8
    37b71a5a1lea rdx, [r15+4*rdx-4]
    37b71a5a6lea r8, [rsp+420]
    37b71a5aemov [rsp+96], rax
    37b71a5b3lea rax, [rip+0xd1bd26]
    37b71a5bamov [rsp+128], r8
    37b71a5c2lea r8, [rsp+432]
    37b71a5camov [rsp+72], rdx
    37b71a5cfmov rdx, r10
    37b71a5d2mov [rsp+64], rax
    37b71a5d7mov [rsp+56], rax
    37b71a5dcmov [rsp+48], rcx
    37b71a5e1lea rcx, [rip+0xd1bd00]
    37b71a5e8mov [rsp+32], rax
    37b71a5edmov [rsp+80], rbp
    37b71a5f2mov [rsp+40], rdi
    37b71a5f7mov [rsp+104], r8
    37b71a5fclea r8, [rsp+416]
    37b71a604mov [rsp+88], r8
    37b71a609mov r8, r10
    37b71a60cmov [rsp+112], r10
    37b71a611mov [rsp+120], r14
    37b71a616call slaln2_
    37b71a61bmov r10l, [rsp+288]
    37b71a623ucomiss x9, [rsp+404]
    37b71a62cmovss x15, [rsp+432]
    37b71a636movss x7, [rsp+436]
    37b71a63fmovss x6, [rsp+440]
    37b71a648mov rax, [rsp+296]
    37b71a650movss x8, [rsp+444]
    37b71a65ajp 0x37b71b120
    37b71a660jne 0x37b71b120
    37b71a666
    37b71a666loc_37b71a666:
    37b71a666movaps x0, x15
    37b71a66amovaps x2, x6
    37b71a66dmovsxd rcx, [rbp]
    37b71a671mov edx, [rsp+152]
    37b71a678andps x2, x10
    37b71a67candps x0, x10
    37b71a680imul edx, ecx
    37b71a683maxss x0, x2
    37b71a687movaps x2, x8
    37b71a68bandps x2, x10
    37b71a68fmovaps x1, x0
    37b71a692movaps x0, x7
    37b71a695lea r8l, [rbx+rdx]
    37b71a699andps x0, x10
    37b71a69dmaxss x1, x13
    37b71a6a2movsxd r8, r8l
    37b71a6a5maxss x0, x2
    37b71a6a9movss [r15+4*r8-4], x15
    37b71a6b0lea r8l, [rcx+rdx]
    37b71a6b4add edx, r12l
    37b71a6b7add ebx, r8l
    37b71a6baadd r8l, r12l
    37b71a6bdmovaps x2, x14
    37b71a6c1movsxd rdx, edx
    37b71a6c4movsxd rbx, ebx
    37b71a6c7movsxd r8, r8l
    37b71a6camaxss x1, x0
    37b71a6cemovss [r15+4*rbx-4], x6
    37b71a6d5movss [r15+4*rdx-4], x7
    37b71a6dcmovss [r15+4*r8-4], x8
    37b71a6e3divss x2, x1
    37b71a6e7movaps x13, x1
    37b71a6ebjmp 0x37b71a280
    37b71a6f0
    37b71a6f0loc_37b71a6f0:
    37b71a6f0mov rax, [rsp+256]
    37b71a6f8movss x5, [rax+4]
    37b71a6fdmov [rsp+412], 0
    37b71a708movss [rsp+416], x5
    37b71a711
    37b71a711loc_37b71a711:
    37b71a711andps x5, x10
    37b71a715pxor x0, x0
    37b71a719mov eax, [rsp+388]
    37b71a720mov edi, [rsp+160]
    37b71a727addss x0, x5
    37b71a72blea r13, [r12+1]
    37b71a730mulss x0, [rsp+304]
    37b71a739movups [rsp+336], x10
    37b71a742mov [rsp+200], eax
    37b71a749imul eax, ecx
    37b71a74clea edx, [rax+rdi-1]
    37b71a750maxss x0, x13
    37b71a755movsxd rdx, edx
    37b71a758movss [r15+4*rdx], x9
    37b71a75emovss [rsp+408], x0
    37b71a767cmp ecx, r12l
    37b71a76ajl 0x37b71ad8a
    37b71a770mov rdx, [rsp+208]
    37b71a778mov rsi, [rsp+216]
    37b71a780cdqe
    37b71a782lea r9l, [rcx-1]
    37b71a786mov rdi, [rsp+240]
    37b71a78eimul rdx, r12
    37b71a792lea r8, [rax+rsi]
    37b71a796sub r9l, esi
    37b71a799lea r10, [rdi+rsi]
    37b71a79dmov rdi, [rsp+1232]
    37b71a7a5add r9, r8
    37b71a7a8lea rax, [r15+4*r8]
    37b71a7acadd rdx, r10
    37b71a7aflea rdx, [rdi+4*rdx]
    37b71a7b3mov rdi, [rsp+352]
    37b71a7bblea r8, [rdi+4*r9]
    37b71a7bfmov r9, [rsp+184]
    37b71a7c7nop [rax]
    37b71a7canop [rax+rax]
    37b71a7d5nop [rax+rax]
    37b71a7e0
    37b71a7e0loc_37b71a7e0:
    37b71a7e0movss x0, [rdx]
    37b71a7e4add rax, 4
    37b71a7e8add rdx, r9
    37b71a7ebxorps x0, x11
    37b71a7efmovss [rax-4], x0
    37b71a7f4cmp r8, rax
    37b71a7f7jne 0x37b71a7e0
    37b71a7f9mov rdx, [rsp+208]
    37b71a801lea r13, [r12+1]
    37b71a806mov rbx, r12
    37b71a809xor edi, edi
    37b71a80bmov r8, [rsp+184]
    37b71a813movaps x1, x14
    37b71a817movaps x12, x9
    37b71a81bmov rax, [rsp+1232]
    37b71a823imul rdx, r12
    37b71a827mov [rsp+360], r13
    37b71a82flea r12, [rip+0xd1baaa]
    37b71a836imul r8, r13
    37b71a83amov [rsp+224], ecx
    37b71a841movss [rsp+320], x13
    37b71a84badd r10, rdx
    37b71a84eneg rdx
    37b71a851lea rsi, [rax+4*r10+4]
    37b71a856lea rax, [r8+4*rdx]
    37b71a85amov r10l, [rsp+232]
    37b71a862mov rbp, rsi
    37b71a865mov r13, rsi
    37b71a868mov [rsp+296], rax
    37b71a870mov esi, [rsp+200]
    37b71a877jmp 0x37b71aaa5
    37b71a87c
    37b71a87c 0f 1f 40 00 ..@.
    37b71a880
    37b71a880loc_37b71a880:
    37b71a880comiss x0, x1
    37b71a883ja 0x37b71be29
    37b71a889lea rax, [rsp+428]
    37b71a891mov [rsp+152], rax
    37b71a899
    37b71a899loc_37b71a899:
    37b71a899imul edx, esi
    37b71a89cmov [rsp+428], edi
    37b71a8a3lea ecx, [rbx+rdx]
    37b71a8a6movsxd rcx, ecx
    37b71a8a9movss x13, [r15+4*rcx-4]
    37b71a8b0mov ecx, [rsp+232]
    37b71a8b7lea edx, [rdx+rcx-1]
    37b71a8bblea rcx, [rip+0xd1b9ea]
    37b71a8c2mov [rsp+32], rcx
    37b71a8c7mov r8, rcx
    37b71a8camovsxd rdx, edx
    37b71a8cdmov rcx, [rsp+152]
    37b71a8d5lea r9, [r15+4*rdx]
    37b71a8d9mov rdx, r13
    37b71a8dccall sdot_
    37b71a8e1mov edx, [r14]
    37b71a8e4movaps x1, x13
    37b71a8e8lea rcx, [rsp+384]
    37b71a8f0subss x1, x0
    37b71a8f4lea rax, [rip+0xd1b9dd]
    37b71a8fbmovss [rsp+432], x15
    37b71a905imul edx, esi
    37b71a908lea r8, [rip+0xd1b99d]
    37b71a90fmovss [rsp+436], x7
    37b71a918lea r11, [rsp+404]
    37b71a920lea r9, [rsp+408]
    37b71a928movss [rsp+440], x6
    37b71a931mov [rsp+280], r11
    37b71a939movss [rsp+444], x8
    37b71a943add edx, ebx
    37b71a945movsxd rdx, edx
    37b71a948sub rdx, 1
    37b71a94cmovss [r15+4*rdx], x1
    37b71a952lea rdx, [r15+4*rdx]
    37b71a956mov [rsp+136], rcx
    37b71a95elea rcx, [rsp+420]
    37b71a966mov [rsp+112], rax
    37b71a96blea rax, [rip+0xd1b94a]
    37b71a972mov [rsp+128], rcx
    37b71a97alea rcx, [rsp+432]
    37b71a982mov [rsp+96], rax
    37b71a987mov rax, [rsp+1240]
    37b71a98fmov [rsp+104], rcx
    37b71a994lea rcx, [rsp+416]
    37b71a99cmov [rsp+88], rcx
    37b71a9a1lea rcx, [rip+0xd1b934]
    37b71a9a8mov [rsp+72], rdx
    37b71a9admov rdx, r8
    37b71a9b0mov [rsp+80], r14
    37b71a9b5mov [rsp+64], r12
    37b71a9bamov [rsp+56], r12
    37b71a9bfmov [rsp+48], rax
    37b71a9c4mov [rsp+40], rbp
    37b71a9c9mov [rsp+32], r12
    37b71a9cemov [rsp+120], r11
    37b71a9d3call slaln2_
    37b71a9d8mov r11, [rsp+280]
    37b71a9e0ucomiss x9, [rsp+404]
    37b71a9e9movss x15, [rsp+432]
    37b71a9f3movss x7, [rsp+436]
    37b71a9fcmovss x6, [rsp+440]
    37b71aa05movss x8, [rsp+444]
    37b71aa0fjp 0x37b71bf28
    37b71aa15jne 0x37b71bf28
    37b71aa1b
    37b71aa1bloc_37b71aa1b:
    37b71aa1bmovaps x0, x15
    37b71aa1fmovaps x1, x14
    37b71aa23mov edx, [r14]
    37b71aa26mov r10l, [rsp+144]
    37b71aa2eandps x0, x10
    37b71aa32imul edx, esi
    37b71aa35maxss x12, x0
    37b71aa3adivss x1, x12
    37b71aa3fadd edx, ebx
    37b71aa41movsxd rdx, edx
    37b71aa44movss [r15+4*rdx-4], x15
    37b71aa4bnop [rax+rax]
    37b71aa54nop [rax+rax]
    37b71aa5fnop [rax+rax]
    37b71aa6anop [rax+rax]
    37b71aa75nop [rax+rax]
    37b71aa80
    37b71aa80loc_37b71aa80:
    37b71aa80add rdi, 1
    37b71aa84add rbx, 1
    37b71aa88add r13, [rsp+184]
    37b71aa90add rbp, [rsp+168]
    37b71aa98cmp [rsp+224], ebx
    37b71aa9fjl 0x37b71ad78
    37b71aaa5
    37b71aaa5loc_37b71aaa5:
    37b71aaa5cmp r10l, ebx
    37b71aaa8jg 0x37b71aa80
    37b71aaaalea eax, [rbx+1]
    37b71aaadmov edx, [r14]
    37b71aab0mov ecx, esi
    37b71aab2mov r10l, ebx
    37b71aab5mov [rsp+144], eax
    37b71aabcmov rax, [rsp+192]
    37b71aac4imul ecx, edx
    37b71aac7movss x0, [rax+4*rdi]
    37b71aacccmp edx, ebx
    37b71aacejle 0x37b71a880
    37b71aad4pxor x2, x2
    37b71aad8ucomiss x2, [rbp+4]
    37b71aadcjp 0x37b71aae4
    37b71aadeje 0x37b71a880
    37b71aae4
    37b71aae4loc_37b71aae4:
    37b71aae4mov rax, [rsp+192]
    37b71aaecadd r10l, 2
    37b71aaf0maxss x0, [rax+4*rdi+4]
    37b71aaf6comiss x0, x1
    37b71aaf9ja 0x37b71b078
    37b71aafflea rax, [rsp+428]
    37b71ab07mov [rsp+152], rax
    37b71ab0f
    37b71ab0floc_37b71ab0f:
    37b71ab0fimul edx, esi
    37b71ab12lea r11, [rip+0xd1b793]
    37b71ab19mov [rsp+288], r10l
    37b71ab21mov r8, r11
    37b71ab24mov [rsp+428], edi
    37b71ab2blea ecx, [rbx+rdx]
    37b71ab2emovsxd rcx, ecx
    37b71ab31movss x13, [r15+4*rcx-4]
    37b71ab38mov ecx, [rsp+232]
    37b71ab3fmov [rsp+32], r11
    37b71ab44lea edx, [rdx+rcx-1]
    37b71ab48mov rcx, [rsp+152]
    37b71ab50movsxd rdx, edx
    37b71ab53lea r9, [r15+4*rdx]
    37b71ab57mov rdx, r13
    37b71ab5acall sdot_
    37b71ab5fmov edx, [r14]
    37b71ab62movaps x1, x13
    37b71ab66mov r9l, [rsp+144]
    37b71ab6esubss x1, x0
    37b71ab72mov r10l, [rsp+232]
    37b71ab7alea r11, [rip+0xd1b72b]
    37b71ab81mov [rsp+428], edi
    37b71ab88imul edx, esi
    37b71ab8bmov r8, r11
    37b71ab8elea ecx, [rbx+rdx]
    37b71ab91movsxd rcx, ecx
    37b71ab94movss [r15+4*rcx-4], x1
    37b71ab9blea ecx, [rdx+r9]
    37b71ab9flea edx, [rdx+r10-1]
    37b71aba4movsxd rcx, ecx
    37b71aba7movss x13, [r15+4*rcx-4]
    37b71abaemovsxd rcx, edx
    37b71abb1mov [rsp+32], r11
    37b71abb6mov rdx, [rsp+296]
    37b71abbelea r9, [r15+4*rcx]
    37b71abc2mov rcx, [rsp+152]
    37b71abcaadd rdx, r13
    37b71abcdcall sdot_
    37b71abd2mov edx, [r14]
    37b71abd5mov r9l, [rsp+144]
    37b71abddmovaps x1, x13
    37b71abe1subss x1, x0
    37b71abe5mov rax, [rsp+1240]
    37b71abedlea r11, [rsp+404]
    37b71abf5movss [rsp+432], x15
    37b71abffimul edx, esi
    37b71ac02lea r8, [rip+0xd1b6a3]
    37b71ac09mov [rsp+280], r11
    37b71ac11movss [rsp+436], x7
    37b71ac1amovss [rsp+440], x6
    37b71ac23movss [rsp+444], x8
    37b71ac2dlea ecx, [rdx+r9]
    37b71ac31add edx, ebx
    37b71ac33lea r9, [rsp+408]
    37b71ac3bmovsxd rcx, ecx
    37b71ac3emovsxd rdx, edx
    37b71ac41movss [r15+4*rcx-4], x1
    37b71ac48lea rcx, [rsp+384]
    37b71ac50lea rdx, [r15+4*rdx-4]
    37b71ac55mov [rsp+136], rcx
    37b71ac5dlea rcx, [rsp+420]
    37b71ac65mov [rsp+128], rcx
    37b71ac6dlea rcx, [rip+0xd1b664]
    37b71ac74mov [rsp+112], rcx
    37b71ac79lea rcx, [rsp+432]
    37b71ac81mov [rsp+104], rcx
    37b71ac86lea rcx, [rip+0xd1b62f]
    37b71ac8dmov [rsp+96], rcx
    37b71ac92lea rcx, [rsp+416]
    37b71ac9amov [rsp+88], rcx
    37b71ac9flea rcx, [rip+0xd1b642]
    37b71aca6mov [rsp+72], rdx
    37b71acablea rdx, [rip+0xd1b626]
    37b71acb2mov [rsp+80], r14
    37b71acb7mov [rsp+64], r12
    37b71acbcmov [rsp+56], r12
    37b71acc1mov [rsp+48], rax
    37b71acc6mov [rsp+40], rbp
    37b71accbmov [rsp+32], r12
    37b71acd0mov [rsp+120], r11
    37b71acd5call slaln2_
    37b71acdamov r11, [rsp+280]
    37b71ace2ucomiss x9, [rsp+404]
    37b71acebmovss x15, [rsp+432]
    37b71acf5movss x7, [rsp+436]
    37b71acfemovss x6, [rsp+440]
    37b71ad07mov r10l, [rsp+288]
    37b71ad0fmovss x8, [rsp+444]
    37b71ad19jp 0x37b71b020
    37b71ad1fjne 0x37b71b020
    37b71ad25
    37b71ad25loc_37b71ad25:
    37b71ad25mov ecx, [r14]
    37b71ad28movaps x1, x7
    37b71ad2bandps x1, x10
    37b71ad2fimul ecx, esi
    37b71ad32lea edx, [rbx+rcx]
    37b71ad35add ecx, [rsp+144]
    37b71ad3cmovsxd rdx, edx
    37b71ad3fmovsxd rcx, ecx
    37b71ad42sub rdx, 1
    37b71ad46movss [r15+4*rdx], x15
    37b71ad4cmovss [r15+4*rcx-4], x7
    37b71ad53movss x0, [r15+4*rdx]
    37b71ad59andps x0, x10
    37b71ad5dmaxss x0, x1
    37b71ad61movaps x1, x14
    37b71ad65maxss x12, x0
    37b71ad6adivss x1, x12
    37b71ad6fjmp 0x37b71aa80
    37b71ad74
    37b71ad74 0f 1f 40 00 ..@.
    37b71ad78
    37b71ad78loc_37b71ad78:
    37b71ad78movss x13, [rsp+320]
    37b71ad82mov r13, [rsp+360]
    37b71ad8a
    37b71ad8aloc_37b71ad8a:
    37b71ad8amov eax, [rsp+328]
    37b71ad91mov r9l, [rsp+268]
    37b71ad99add eax, 1
    37b71ad9cmov [rsp+144], eax
    37b71ada3test r9l, r9l
    37b71ada6je 0x37b71d517
    37b71adaccmp [rsp+236], 1
    37b71adb4jne 0x37b71c367
    37b71adbamov rdi, [rsp+248]
    37b71adc2mov rsi, [rsp+1248]
    37b71adcamov rbx, [rsp+272]
    37b71add2mov eax, [r14]
    37b71add5lea rsi, [rsi+4*rdi+4]
    37b71addaadd rbx, rdi
    37b71adddcmp eax, [rsp+332]
    37b71ade4jg 0x37b71cc5b
    37b71adea
    37b71adealoc_37b71adea:
    37b71adealea r8, [rip+0xd1b4bb]
    37b71adf1mov rdx, rsi
    37b71adf4mov rcx, r14
    37b71adf7call isamax_
    37b71adfcmovaps x1, x9
    37b71ae00mov r8, rsi
    37b71ae03mov rcx, r14
    37b71ae06mov rdi, [rsp+1248]
    37b71ae0ecdqe
    37b71ae10add rax, [rsp+248]
    37b71ae18lea rbp, [rsp+400]
    37b71ae20lea r9, [rip+0xd1b485]
    37b71ae27mov rdx, rbp
    37b71ae2amovss x0, [rdi+4*rax]
    37b71ae2fandps x0, x10
    37b71ae33divss x1, x0
    37b71ae37movss [rsp+400], x1
    37b71ae40call sscal_
    37b71ae45
    37b71ae45loc_37b71ae45:
    37b71ae45mov eax, [rsp+144]
    37b71ae4cxor edi, edi
    37b71ae4emov [rsp+328], eax
    37b71ae55jmp 0x37b719cb0
    37b71ae5a
    37b71ae5a 66 0f 1f 44 00 00 f..D..
    37b71ae60
    37b71ae60loc_37b71ae60:
    37b71ae60movss x13, [rsp+360]
    37b71ae6amov r12, [rsp+368]
    37b71ae72mov r14, rbp
    37b71ae75
    37b71ae75loc_37b71ae75:
    37b71ae75mov r8l, [rsp+268]
    37b71ae7dtest r8l, r8l
    37b71ae80je 0x37b71ce46
    37b71ae86cmp [rsp+236], 1
    37b71ae8eje 0x37b71c6e6
    37b71ae94cmp [rsp+216], 1
    37b71ae9dje 0x37b71aef6
    37b71ae9fmov edx, [rsp+152]
    37b71aea6mov rsi, [rsp+216]
    37b71aeaeimul edx, ecx
    37b71aeb1lea r8l, [rsi-2]
    37b71aeb5mov rsi, [rsp+352]
    37b71aebdmovsxd rdx, edx
    37b71aec0lea rax, [r15+4*rdx]
    37b71aec4add rdx, r8
    37b71aec7lea rdx, [rsi+4*rdx]
    37b71aecbnop [rax+rax]
    37b71aed5nop [rax+rax]
    37b71aee0
    37b71aee0loc_37b71aee0:
    37b71aee0mov [rax], 0
    37b71aee6mov [rax+4*rcx], 0
    37b71aeedadd rax, 4
    37b71aef1cmp rdx, rax
    37b71aef4jne 0x37b71aee0
    37b71aef6
    37b71aef6loc_37b71aef6:
    37b71aef6movsxd rax, [rsp+152]
    37b71aefemov rsi, [rsp+312]
    37b71af06mov rdi, [rip+0xd1b403]
    37b71af0dmov [rsi+4*rax-4], rdi
    37b71af12mov eax, [rsp+200]
    37b71af19mov [rsp+388], eax
    37b71af20mov eax, [rsp+328]
    37b71af27add eax, 1
    37b71af2amov [rsp+144], eax
    37b71af31cmp [rsp+236], 1
    37b71af39jle 0x37b71c9c1
    37b71af3f
    37b71af3floc_37b71af3f:
    37b71af3fmov eax, [rsp+232]
    37b71af46mov edi, 1
    37b71af4blea r13, [r12+1]
    37b71af50mov [rsp+160], eax
    37b71af57
    37b71af57loc_37b71af57:
    37b71af57mov esi, [rsp+236]
    37b71af5emov eax, [r14]
    37b71af61lea edx, [rsi-1]
    37b71af64mov esi, [rsp+200]
    37b71af6bcmp edx, esi
    37b71af6djle 0x37b71c3e1
    37b71af73lea edx, [rsi+1]
    37b71af76cmp [rsp+160], eax
    37b71af7dje 0x37b71c3e1
    37b71af83
    37b71af83loc_37b71af83:
    37b71af83mov [rsp+388], edx
    37b71af8atest edi, edi
    37b71af8cje 0x37b71bc89
    37b71af92mov rax, [rsp+272]
    37b71af9amov rsi, [rsp+248]
    37b71afa2lea rbx, [rax+rsi]
    37b71afa6
    37b71afa6loc_37b71afa6:
    37b71afa6add [rsp+328], 2
    37b71afaemov edi, 1
    37b71afb3jmp 0x37b719cb0
    37b71afb8
    37b71afb8 0f 1f 84 00 00 00 00 00 ........
    37b71afc0
    37b71afc0loc_37b71afc0:
    37b71afc0xor r12l, r12l
    37b71afc3jmp 0x37b7195c0
    37b71afc8
    37b71afc8 0f 1f 84 00 00 00 00 00 ........
    37b71afd0
    37b71afd0loc_37b71afd0:
    37b71afd0mov edx, [rsp+264]
    37b71afd7test edx, edx
    37b71afd9je 0x37b71a6f0
    37b71afdfmov rax, [rsp+1216]
    37b71afe7mov rsi, [rsp+216]
    37b71afefmov eax, [rax+4*rsi-4]
    37b71aff3test eax, eax
    37b71aff5jne 0x37b71a6f0
    37b71affbmov rax, [rsp+272]
    37b71b003mov rsi, [rsp+248]
    37b71b00bxor edi, edi
    37b71b00dlea r13, [r12+1]
    37b71b012lea rbx, [rax+rsi]
    37b71b016jmp 0x37b719cb0
    37b71b01b
    37b71b01b 0f 1f 44 00 00 ..D..
    37b71b020
    37b71b020loc_37b71b020:
    37b71b020mov edx, [r14]
    37b71b023mov eax, [rsp+160]
    37b71b02alea r9, [rip+0xd1b27b]
    37b71b031mov [rsp+280], r10l
    37b71b039mov ecx, edx
    37b71b03bimul edx, esi
    37b71b03esub ecx, eax
    37b71b040add ecx, 1
    37b71b043lea edx, [rax+rdx-1]
    37b71b047mov [rsp+428], ecx
    37b71b04emov rcx, [rsp+152]
    37b71b056movsxd rdx, edx
    37b71b059lea r8, [r15+4*rdx]
    37b71b05dmov rdx, r11
    37b71b060call sscal_
    37b71b065mov r10l, [rsp+280]
    37b71b06djmp 0x37b71ad25
    37b71b072
    37b71b072 66 0f 1f 44 00 00 f..D..
    37b71b078
    37b71b078loc_37b71b078:
    37b71b078movaps x0, x9
    37b71b07cmov eax, [rsp+160]
    37b71b083lea r9, [rip+0xd1b222]
    37b71b08amov [rsp+280], r10l
    37b71b092divss x0, x12
    37b71b097movaps x12, x9
    37b71b09bsub edx, eax
    37b71b09dlea ecx, [rax+rcx-1]
    37b71b0a1lea rax, [rsp+428]
    37b71b0a9add edx, 1
    37b71b0acmovsxd rcx, ecx
    37b71b0afmov [rsp+152], rax
    37b71b0b7mov [rsp+428], edx
    37b71b0belea r8, [r15+4*rcx]
    37b71b0c2lea rdx, [rsp+396]
    37b71b0camov rcx, rax
    37b71b0cdmovss [rsp+396], x0
    37b71b0d6call sscal_
    37b71b0dbmov edx, [r14]
    37b71b0demov r10l, [rsp+280]
    37b71b0e6jmp 0x37b71ab0f
    37b71b0eb
    37b71b0eb 0f 1f 44 00 00 ..D..
    37b71b0f0
    37b71b0f0loc_37b71b0f0:
    37b71b0f0mov edx, [rsp+160]
    37b71b0f7imul edx, [r14]
    37b71b0fblea rcx, [rsp+392]
    37b71b103lea r9, [rip+0xd1b1a2]
    37b71b10amovsxd rdx, edx
    37b71b10dlea r8, [r15+4*rdx]
    37b71b111mov rdx, r11
    37b71b114call sscal_
    37b71b119jmp 0x37b71993f
    37b71b11e
    37b71b11e 66 90 f.
    37b71b120
    37b71b120loc_37b71b120:
    37b71b120mov edx, [rbp]
    37b71b123mov [rsp+288], rax
    37b71b12blea r9, [rip+0xd1b17a]
    37b71b132mov eax, [rsp+160]
    37b71b139mov [rsp+280], r10l
    37b71b141mov ecx, edx
    37b71b143imul edx, [rsp+152]
    37b71b14bsub ecx, eax
    37b71b14dadd ecx, 1
    37b71b150lea edx, [rax+rdx-1]
    37b71b154mov [rsp+428], ecx
    37b71b15bmov rcx, [rsp+144]
    37b71b163movsxd rdx, edx
    37b71b166lea r8, [r15+4*rdx]
    37b71b16amov rdx, r14
    37b71b16dcall sscal_
    37b71b172mov edx, [rbp]
    37b71b175mov eax, [rsp+160]
    37b71b17clea r9, [rip+0xd1b129]
    37b71b183mov r11l, edx
    37b71b186imul edx, [rsp+200]
    37b71b18esub r11l, eax
    37b71b191lea ecx, [r11+1]
    37b71b195lea edx, [rax+rdx-1]
    37b71b199mov [rsp+428], ecx
    37b71b1a0mov rcx, [rsp+144]
    37b71b1a8movsxd rdx, edx
    37b71b1ablea r8, [r15+4*rdx]
    37b71b1afmov rdx, r14
    37b71b1b2call sscal_
    37b71b1b7mov rax, [rsp+288]
    37b71b1bfmov r10l, [rsp+280]
    37b71b1c7jmp 0x37b71a666
    37b71b1cc
    37b71b1cc 0f 1f 40 00 ..@.
    37b71b1d0
    37b71b1d0loc_37b71b1d0:
    37b71b1d0movaps x0, x9
    37b71b1d4mov [rsp+288], rax
    37b71b1dcmov eax, [rsp+160]
    37b71b1e3divss x0, x13
    37b71b1e8mov [rsp+280], r10l
    37b71b1f0movaps x13, x9
    37b71b1f4sub ecx, eax
    37b71b1f6lea edx, [rcx+1]
    37b71b1f9lea rcx, [rsp+428]
    37b71b201mov [rsp+428], edx
    37b71b208lea edx, [rax+r9-1]
    37b71b20dlea r9, [rip+0xd1b098]
    37b71b214movsxd rdx, edx
    37b71b217mov [rsp+144], rcx
    37b71b21flea r8, [r15+4*rdx]
    37b71b223lea rdx, [rsp+396]
    37b71b22bmovss [rsp+396], x0
    37b71b234call sscal_
    37b71b239mov edx, [rbp]
    37b71b23cmov eax, [rsp+160]
    37b71b243lea r9, [rip+0xd1b062]
    37b71b24amov r10l, edx
    37b71b24dimul edx, [rsp+200]
    37b71b255sub r10l, eax
    37b71b258lea ecx, [r10+1]
    37b71b25clea edx, [rax+rdx-1]
    37b71b260mov [rsp+428], ecx
    37b71b267mov rcx, [rsp+144]
    37b71b26fmovsxd rdx, edx
    37b71b272lea r8, [r15+4*rdx]
    37b71b276lea rdx, [rsp+396]
    37b71b27ecall sscal_
    37b71b283mov edx, [rbp]
    37b71b286mov rax, [rsp+288]
    37b71b28emov r10l, [rsp+280]
    37b71b296jmp 0x37b71a319
    37b71b29b
    37b71b29b 0f 1f 44 00 00 ..D..
    37b71b2a0
    37b71b2a0loc_37b71b2a0:
    37b71b2a0mov edx, 1
    37b71b2a5jmp 0x37b719435
    37b71b2aa
    37b71b2aa 66 0f 1f 44 00 00 f..D..
    37b71b2b0
    37b71b2b0loc_37b71b2b0:
    37b71b2b0mov rsi, [rsp+240]
    37b71b2b8movaps x2, x1
    37b71b2bbadd r8, [rsp+208]
    37b71b2c3movss x11, [rip+0xd1b024]
    37b71b2ccandps x2, x11
    37b71b2d0lea r10, [r8+rsi]
    37b71b2d4mov rsi, [rsp+1232]
    37b71b2dcmovaps x0, x2
    37b71b2dflea rdx, [r9+r10]
    37b71b2e3sqrtss x0, x0
    37b71b2e7add rcx, r10
    37b71b2eamovss x3, [rsi+4*rdx]
    37b71b2efmov rsi, [rsp+1232]
    37b71b2f7movss [rsp+416], x3
    37b71b300movss x4, [rsi+4*rcx]
    37b71b305movsxd rcx, [r14]
    37b71b308andps x3, x11
    37b71b30cmov esi, [rsp+388]
    37b71b313movss [rsp+160], x4
    37b71b31cmovaps x5, x4
    37b71b31fandps x5, x11
    37b71b323mov r11l, esi
    37b71b326mov [rsp+168], esi
    37b71b32dsub esi, 1
    37b71b330movaps x4, x5
    37b71b333imul r11l, ecx
    37b71b337sqrtss x4, x4
    37b71b33bmov ebx, r11l
    37b71b33elea r8l, [r13+r11-1]
    37b71b343mulss x0, x4
    37b71b347sub ebx, ecx
    37b71b349movsxd r8, r8l
    37b71b34ccomiss x5, x2
    37b71b34flea edx, [rbx+rax-1]
    37b71b353movsxd rdx, edx
    37b71b356movss [rsp+412], x0
    37b71b35faddss x3, x0
    37b71b363mulss x3, [rsp+304]
    37b71b36cmaxss x3, x13
    37b71b371movss [rsp+408], x3
    37b71b37ajb 0x37b71bcc6
    37b71b380divss x0, [rsp+160]
    37b71b389movaps x1, x9
    37b71b38d
    37b71b38dloc_37b71b38d:
    37b71b38dlea r12l, [r13+rbx-1]
    37b71b392movss [r15+4*rdx], x1
    37b71b398movsxd r12, r12l
    37b71b39bmovss [r15+4*r8], x0
    37b71b3a1mov [r15+4*r12], 0
    37b71b3a9lea r12l, [r11+rax-1]
    37b71b3aemovsxd r12, r12l
    37b71b3b1mov [r15+4*r12], 0
    37b71b3b9test edi, edi
    37b71b3bbje 0x37b71bae2
    37b71b3c1mov rcx, [rsp+1232]
    37b71b3c9movsxd rbx, ebx
    37b71b3ccmovsxd r11, r11l
    37b71b3cfmov eax, eax
    37b71b3d1lea rbx, [r15+4*rbx]
    37b71b3d5lea r11, [r15+4*r11]
    37b71b3d9lea rbp, [rcx+4*rbp]
    37b71b3ddlea r10, [rcx+4*r10]
    37b71b3e1mov ecx, 1
    37b71b3e6nop [rax]
    37b71b3eanop [rax+rax]
    37b71b3f5nop [rax+rax]
    37b71b400
    37b71b400loc_37b71b400:
    37b71b400movss x0, [r15+4*rdx]
    37b71b406mulss x0, [rbp+4*rcx]
    37b71b40cxorps x0, [rip+0xd1aeed]
    37b71b413movss [rbx+4*rcx-4], x0
    37b71b419movss x0, [r15+4*r8]
    37b71b41fmulss x0, [r10+4*rcx]
    37b71b425xorps x0, [rip+0xd1aed4]
    37b71b42cmovss [r11+4*rcx-4], x0
    37b71b433add rcx, 1
    37b71b437cmp rcx, rax
    37b71b43ajne 0x37b71b400
    37b71b43clea eax, [r13-3]
    37b71b440mov rcx, [rsp+240]
    37b71b448movsxd rbx, edi
    37b71b44bmov r10, [rsp+1232]
    37b71b453cdqe
    37b71b455mov [rsp+160], esi
    37b71b45cimul rax, [rsp+208]
    37b71b465lea rdx, [rax+rcx]
    37b71b469lea rbp, [r10+4*rdx+4]
    37b71b46elea r10, [r15+4*r9]
    37b71b472add r9, rcx
    37b71b475add r9, rax
    37b71b478mov rax, [rsp+1232]
    37b71b480mov rsi, r10
    37b71b483lea rax, [rax+4*r9-8]
    37b71b488jmp 0x37b71b740
    37b71b48d
    37b71b48d 0f 1f 00 ...
    37b71b490
    37b71b490loc_37b71b490:
    37b71b490mov rcx, [rsp+184]
    37b71b498lea edx, [r12+rdx-1]
    37b71b49dmov [rsp+80], r14
    37b71b4a2lea r8, [rip+0xd1ae2f]
    37b71b4a9movsxd rdx, edx
    37b71b4acmov r9, [rsp+176]
    37b71b4b4mov [rsp+200], rax
    37b71b4bcmov [rsp+136], rcx
    37b71b4c4lea rdx, [r15+4*rdx]
    37b71b4c8lea rcx, [rsp+420]
    37b71b4d0mov [rsp+128], rcx
    37b71b4d8lea rcx, [rsp+404]
    37b71b4e0mov [rsp+192], rcx
    37b71b4e8mov [rsp+120], rcx
    37b71b4edlea rcx, [rip+0xd1ade4]
    37b71b4f4mov [rsp+112], rcx
    37b71b4f9lea rcx, [rsp+432]
    37b71b501mov [rsp+104], rcx
    37b71b506lea rcx, [rsp+412]
    37b71b50emov [rsp+72], rdx
    37b71b513mov rdx, [rsp+1240]
    37b71b51bmov [rsp+96], rcx
    37b71b520lea rcx, [rsp+416]
    37b71b528mov [rsp+88], rcx
    37b71b52dlea rcx, [rip+0xd1adac]
    37b71b534mov [rsp+48], rdx
    37b71b539mov rdx, rax
    37b71b53csub rdx, [rsp+144]
    37b71b544mov [rsp+64], rcx
    37b71b549mov [rsp+56], rcx
    37b71b54emov [rsp+40], rdx
    37b71b553lea rdx, [rip+0xd1ad52]
    37b71b55amov [rsp+32], rcx
    37b71b55flea rcx, [rip+0xd1ad76]
    37b71b566movss [rsp+432], x15
    37b71b570movss [rsp+436], x7
    37b71b579movss [rsp+440], x6
    37b71b582movss [rsp+444], x8
    37b71b58ccall slaln2_
    37b71b591mov rax, [rsp+200]
    37b71b599movss x0, [rsp+420]
    37b71b5a2movss x15, [rsp+432]
    37b71b5acmovss x7, [rsp+436]
    37b71b5b5movss x6, [rsp+440]
    37b71b5becomiss x0, x9
    37b71b5c2movss x8, [rsp+444]
    37b71b5ccmovss x1, [rsp+404]
    37b71b5d5jbe 0x37b71b5ff
    37b71b5d7movaps x2, x14
    37b71b5dbmovss x3, [rsi-12]
    37b71b5e0divss x2, x0
    37b71b5e4comiss x3, x2
    37b71b5e7jbe 0x37b71b5ff
    37b71b5e9divss x1, x0
    37b71b5eddivss x15, x0
    37b71b5f2movss [rsp+404], x1
    37b71b5fbdivss x6, x0
    37b71b5ff
    37b71b5ffloc_37b71b5ff:
    37b71b5ffucomiss x1, x9
    37b71b603jp 0x37b71bcf5
    37b71b609jne 0x37b71bcf5
    37b71b60f
    37b71b60floc_37b71b60f:
    37b71b60fmov ecx, [r14]
    37b71b612mov edx, [rsp+160]
    37b71b619movaps x0, x15
    37b71b61dmov [rsp+248], rax
    37b71b625lea rax, [rip+0xd1ac80]
    37b71b62cxorps x0, [rip+0xd1accd]
    37b71b633mov [rsp+424], r13l
    37b71b63bimul edx, ecx
    37b71b63emov r9, rax
    37b71b641movss [rsp+428], x0
    37b71b64aadd ecx, edx
    37b71b64clea r8l, [r12+rdx-1]
    37b71b651movsxd rdx, edx
    37b71b654lea ecx, [r12+rcx-1]
    37b71b659movsxd r8, r8l
    37b71b65clea rdx, [r15+4*rdx]
    37b71b660movsxd rcx, ecx
    37b71b663movss [r15+4*r8], x15
    37b71b669mov r8, rbp
    37b71b66csub r8, [rsp+144]
    37b71b674movss [r15+4*rcx], x6
    37b71b67alea rcx, [rsp+424]
    37b71b682mov [rsp+40], rax
    37b71b687mov [rsp+32], rdx
    37b71b68clea rdx, [rsp+428]
    37b71b694mov [rsp+200], r8
    37b71b69cmov [rsp+192], rcx
    37b71b6a4call saxpy_
    37b71b6a9lea rax, [rip+0xd1abfc]
    37b71b6b0movaps x0, x6
    37b71b6b3mov edx, [rsp+168]
    37b71b6bamov [rsp+40], rax
    37b71b6bfimul edx, [r14]
    37b71b6c3mov r9, rax
    37b71b6c6xorps x0, [rip+0xd1ac33]
    37b71b6cdmov r8, [rsp+200]
    37b71b6d5mov [rsp+424], r13l
    37b71b6ddmov rcx, [rsp+192]
    37b71b6e5movss [rsp+428], x0
    37b71b6eemovsxd rdx, edx
    37b71b6f1lea rdx, [r15+4*rdx]
    37b71b6f5mov [rsp+32], rdx
    37b71b6falea rdx, [rsp+428]
    37b71b702call saxpy_
    37b71b707mov rax, [rsp+248]
    37b71b70fnop [rax+rax]
    37b71b715nop [rax+rax]
    37b71b720
    37b71b720loc_37b71b720:
    37b71b720sub rbx, 1
    37b71b724add rbp, [rsp+144]
    37b71b72csub rsi, 4
    37b71b730add rax, [rsp+152]
    37b71b738test ebx, ebx
    37b71b73ajle 0x37b71bad8
    37b71b740
    37b71b740loc_37b71b740:
    37b71b740cmp edi, ebx
    37b71b742jl 0x37b71b720
    37b71b744mov edx, [rsp+160]
    37b71b74bmov ecx, ebx
    37b71b74dimul edx, [r14]
    37b71b751mov r12l, ebx
    37b71b754lea r13l, [rbx-1]
    37b71b758cmp ebx, 1
    37b71b75bje 0x37b71bc00
    37b71b761pxor x0, x0
    37b71b765ucomiss x0, [rax]
    37b71b768jp 0x37b71b773
    37b71b76amov edi, r13l
    37b71b76dje 0x37b71b490
    37b71b773
    37b71b773loc_37b71b773:
    37b71b773lea edi, [rcx-2]
    37b71b776mov rcx, [rsp+184]
    37b71b77elea edx, [rdx+r13-1]
    37b71b783mov [rsp+80], r14
    37b71b788movsxd rdx, edx
    37b71b78blea r8, [rip+0xd1ab46]
    37b71b792mov r9, [rsp+176]
    37b71b79amov [rsp+200], rax
    37b71b7a2mov [rsp+136], rcx
    37b71b7aalea rdx, [r15+4*rdx]
    37b71b7aelea rcx, [rsp+420]
    37b71b7b6mov [rsp+128], rcx
    37b71b7belea rcx, [rsp+404]
    37b71b7c6mov [rsp+192], rcx
    37b71b7cemov [rsp+120], rcx
    37b71b7d3lea rcx, [rip+0xd1aafe]
    37b71b7damov [rsp+112], rcx
    37b71b7dflea rcx, [rsp+432]
    37b71b7e7mov [rsp+104], rcx
    37b71b7eclea rcx, [rsp+412]
    37b71b7f4mov [rsp+72], rdx
    37b71b7f9mov rdx, [rsp+1240]
    37b71b801mov [rsp+96], rcx
    37b71b806lea rcx, [rsp+416]
    37b71b80emov [rsp+88], rcx
    37b71b813lea rcx, [rip+0xd1aac6]
    37b71b81amov [rsp+48], rdx
    37b71b81flea rdx, [rax-4]
    37b71b823mov [rsp+64], rcx
    37b71b828mov [rsp+56], rcx
    37b71b82dmov [rsp+40], rdx
    37b71b832mov rdx, r8
    37b71b835mov [rsp+32], rcx
    37b71b83alea rcx, [rip+0xd1aa9b]
    37b71b841movss [rsp+432], x15
    37b71b84bmovss [rsp+436], x7
    37b71b854movss [rsp+440], x6
    37b71b85dmovss [rsp+444], x8
    37b71b867call slaln2_
    37b71b86cmov rax, [rsp+200]
    37b71b874movss x0, [rsp+420]
    37b71b87dmovss x15, [rsp+432]
    37b71b887movss x7, [rsp+436]
    37b71b890movss x6, [rsp+440]
    37b71b899comiss x0, x9
    37b71b89dmovss x8, [rsp+444]
    37b71b8a7movss x1, [rsp+404]
    37b71b8b0jbe 0x37b71b8f9
    37b71b8b2movaps x2, x14
    37b71b8b6movss x3, [rsi-16]
    37b71b8bbmaxss x3, [rsi-12]
    37b71b8c0divss x2, x0
    37b71b8c4comiss x3, x2
    37b71b8c7jbe 0x37b71b8f9
    37b71b8c9movaps x2, x9
    37b71b8cddivss x2, x0
    37b71b8d1mulss x1, x2
    37b71b8d5movss [rsp+396], x2
    37b71b8demulss x15, x2
    37b71b8e3mulss x6, x2
    37b71b8e7mulss x7, x2
    37b71b8ebmulss x8, x2
    37b71b8f0movss [rsp+404], x1
    37b71b8f9
    37b71b8f9loc_37b71b8f9:
    37b71b8f9ucomiss x1, x9
    37b71b8fdjp 0x37b71bc10
    37b71b903jne 0x37b71bc10
    37b71b909
    37b71b909loc_37b71b909:
    37b71b909mov ecx, [r14]
    37b71b90cmov edx, [rsp+160]
    37b71b913movaps x0, x15
    37b71b917mov [rsp+200], rax
    37b71b91flea rax, [rip+0xd1a986]
    37b71b926xorps x0, [rip+0xd1a9d3]
    37b71b92dmov [rsp+424], edi
    37b71b934imul edx, ecx
    37b71b937mov r9, rax
    37b71b93amovss [rsp+428], x0
    37b71b943lea r8l, [rdx+r13-1]
    37b71b948add ecx, edx
    37b71b94amovsxd r8, r8l
    37b71b94dmovss [r15+4*r8], x15
    37b71b953lea r8l, [r12+rdx-1]
    37b71b958movsxd rdx, edx
    37b71b95bmovsxd r8, r8l
    37b71b95elea rdx, [r15+4*rdx]
    37b71b962movss [r15+4*r8], x7
    37b71b968lea r8l, [rcx+r13-1]
    37b71b96dlea ecx, [r12+rcx-1]
    37b71b972movsxd r8, r8l
    37b71b975movsxd rcx, ecx
    37b71b978movss [r15+4*r8], x6
    37b71b97emov r8, rbp
    37b71b981movss [r15+4*rcx], x8
    37b71b987lea rcx, [rsp+424]
    37b71b98fmov [rsp+40], rax
    37b71b994mov [rsp+32], rdx
    37b71b999lea rdx, [rsp+428]
    37b71b9a1call saxpy_
    37b71b9a6lea rax, [rip+0xd1a8ff]
    37b71b9admovaps x0, x7
    37b71b9b0mov r11, rbp
    37b71b9b3mov [rsp+40], rax
    37b71b9b8mov edx, [rsp+160]
    37b71b9bfmov r9, rax
    37b71b9c2lea rcx, [rsp+424]
    37b71b9caimul edx, [r14]
    37b71b9cesub r11, [rsp+144]
    37b71b9d6mov [rsp+424], edi
    37b71b9ddxorps x0, [rip+0xd1a91c]
    37b71b9e4mov r8, r11
    37b71b9e7mov [rsp+192], r11
    37b71b9efmovss [rsp+428], x0
    37b71b9f8movsxd rdx, edx
    37b71b9fblea rdx, [r15+4*rdx]
    37b71b9ffmov [rsp+32], rdx
    37b71ba04lea rdx, [rsp+428]
    37b71ba0ccall saxpy_
    37b71ba11lea rax, [rip+0xd1a894]
    37b71ba18movaps x0, x6
    37b71ba1bmov r8, rbp
    37b71ba1emov [rsp+40], rax
    37b71ba23mov edx, [rsp+168]
    37b71ba2amov r9, rax
    37b71ba2dlea rcx, [rsp+424]
    37b71ba35imul edx, [r14]
    37b71ba39xorps x0, [rip+0xd1a8c0]
    37b71ba40mov [rsp+424], edi
    37b71ba47movss [rsp+428], x0
    37b71ba50movsxd rdx, edx
    37b71ba53lea rdx, [r15+4*rdx]
    37b71ba57mov [rsp+32], rdx
    37b71ba5clea rdx, [rsp+428]
    37b71ba64call saxpy_
    37b71ba69lea rax, [rip+0xd1a83c]
    37b71ba70mov edx, [rsp+168]
    37b71ba77mov r8, [rsp+192]
    37b71ba7fmov [rsp+40], rax
    37b71ba84imul edx, [r14]
    37b71ba88movaps x0, x8
    37b71ba8cmov r9, rax
    37b71ba8fxorps x0, [rip+0xd1a86a]
    37b71ba96lea rcx, [rsp+424]
    37b71ba9emov [rsp+424], edi
    37b71baa5movss [rsp+428], x0
    37b71baaemovsxd rdx, edx
    37b71bab1lea rdx, [r15+4*rdx]
    37b71bab5mov [rsp+32], rdx
    37b71babalea rdx, [rsp+428]
    37b71bac2call saxpy_
    37b71bac7mov rax, [rsp+200]
    37b71bacfjmp 0x37b71b720
    37b71bad4
    37b71bad4 0f 1f 40 00 ..@.
    37b71bad8
    37b71bad8loc_37b71bad8:
    37b71bad8mov esi, [rsp+160]
    37b71badfmovsxd rcx, [r14]
    37b71bae2
    37b71bae2loc_37b71bae2:
    37b71bae2mov edi, [rsp+268]
    37b71bae9test edi, edi
    37b71baebje 0x37b71d19f
    37b71baf1mov edi, [rsp+392]
    37b71baf8lea ebx, [rdi-1]
    37b71bafbmov r13l, ebx
    37b71bafecmp [rsp+236], 1
    37b71bb06je 0x37b71c0e4
    37b71bb0ccmp ecx, edi
    37b71bb0ejle 0x37b71bb56
    37b71bb10mov edx, ecx
    37b71bb12mov r8l, ecx
    37b71bb15movsxd rax, edi
    37b71bb18imul edx, esi
    37b71bb1bsub r8l, edi
    37b71bb1esub r8l, 1
    37b71bb22movsxd rdx, edx
    37b71bb25add rdx, rax
    37b71bb28add r8, rdx
    37b71bb2blea rax, [r15+4*rdx]
    37b71bb2flea rdx, [r15+4*r8+4]
    37b71bb34nop
    37b71bb35nop [rax+rax]
    37b71bb40
    37b71bb40loc_37b71bb40:
    37b71bb40mov [rax], 0
    37b71bb46mov [rax+4*rcx], 0
    37b71bb4dadd rax, 4
    37b71bb51cmp rax, rdx
    37b71bb54jne 0x37b71bb40
    37b71bb56
    37b71bb56loc_37b71bb56:
    37b71bb56mov rdx, [rip+0xd1a7b3]
    37b71bb5dmov rdi, [rsp+312]
    37b71bb65movsxd rax, esi
    37b71bb68mov [rsp+388], esi
    37b71bb6fmov [rdi+4*rax-4], rdx
    37b71bb74cmp [rsp+236], 1
    37b71bb7cjle 0x37b71c35a
    37b71bb82mov eax, [rsp+232]
    37b71bb89lea r13l, [rax-1]
    37b71bb8d
    37b71bb8dloc_37b71bb8d:
    37b71bb8dmov ebp, [rsp+232]
    37b71bb94mov [rsp+232], r13l
    37b71bb9cmov r10l, -1
    37b71bba2
    37b71bba2loc_37b71bba2:
    37b71bba2cmp esi, 2
    37b71bba5jle 0x37b71ca0b
    37b71bbabcmp ebx, 1
    37b71bbaeje 0x37b71ca0b
    37b71bbb4sub esi, 1
    37b71bbb7mov r13l, [rsp+392]
    37b71bbbfmov [rsp+388], esi
    37b71bbc6
    37b71bbc6loc_37b71bbc6:
    37b71bbc6sub r13l, 1
    37b71bbcatest r10l, r10l
    37b71bbcdje 0x37b71bcb0
    37b71bbd3mov r12l, ebp
    37b71bbd6
    37b71bbd6loc_37b71bbd6:
    37b71bbd6mov [rsp+392], r13l
    37b71bbdetest r13l, r13l
    37b71bbe1jle 0x37b719b23
    37b71bbe7lea eax, [r12-2]
    37b71bbecmov [rsp+232], eax
    37b71bbf3jmp 0x37b7193a4
    37b71bbf8
    37b71bbf8 0f 1f 84 00 00 00 00 00 ........
    37b71bc00
    37b71bc00loc_37b71bc00:
    37b71bc00xor edi, edi
    37b71bc02jmp 0x37b71b490
    37b71bc07
    37b71bc07 66 0f 1f 84 00 00 00 00 00 f........
    37b71bc10
    37b71bc10loc_37b71bc10:
    37b71bc10mov edx, [rsp+160]
    37b71bc17imul edx, [r14]
    37b71bc1blea rcx, [rsp+392]
    37b71bc23mov [rsp+248], rax
    37b71bc2blea r9, [rip+0xd1a67a]
    37b71bc32mov [rsp+200], rcx
    37b71bc3amovsxd rdx, edx
    37b71bc3dlea r8, [r15+4*rdx]
    37b71bc41mov rdx, [rsp+192]
    37b71bc49call sscal_
    37b71bc4emov eax, [rsp+168]
    37b71bc55imul eax, [r14]
    37b71bc59lea r9, [rip+0xd1a64c]
    37b71bc60mov rcx, [rsp+200]
    37b71bc68movsxd rdx, eax
    37b71bc6blea r8, [r15+4*rdx]
    37b71bc6fmov rdx, [rsp+192]
    37b71bc77call sscal_
    37b71bc7cmov rax, [rsp+248]
    37b71bc84jmp 0x37b71b909
    37b71bc89
    37b71bc89loc_37b71bc89:
    37b71bc89mov eax, [rsp+144]
    37b71bc90mov rsi, [rsp+248]
    37b71bc98mov [rsp+328], eax
    37b71bc9fmov rax, [rsp+272]
    37b71bca7lea rbx, [rax+rsi]
    37b71bcabjmp 0x37b719cb0
    37b71bcb0
    37b71bcb0loc_37b71bcb0:
    37b71bcb0mov [rsp+392], r13l
    37b71bcb8test r13l, r13l
    37b71bcbbjg 0x37b7193e0
    37b71bcc1jmp 0x37b719b23
    37b71bcc6
    37b71bcc6loc_37b71bcc6:
    37b71bcc6divss x0, x1
    37b71bccamovaps x1, x0
    37b71bccdmovaps x0, x9
    37b71bcd1xorps x1, [rip+0xd1a628]
    37b71bcd8jmp 0x37b71b38d
    37b71bcdd
    37b71bcddloc_37b71bcdd:
    37b71bcddmovaps x1, x12
    37b71bce1divss x1, x0
    37b71bce5xorps x1, x11
    37b71bce9movaps x0, x1
    37b71bcecmovaps x1, x9
    37b71bcf0jmp 0x37b719e16
    37b71bcf5
    37b71bcf5loc_37b71bcf5:
    37b71bcf5mov edx, [rsp+160]
    37b71bcfcimul edx, [r14]
    37b71bd00lea rcx, [rsp+392]
    37b71bd08mov [rsp+248], rax
    37b71bd10lea r9, [rip+0xd1a595]
    37b71bd17mov [rsp+200], rcx
    37b71bd1fmovsxd rdx, edx
    37b71bd22lea r8, [r15+4*rdx]
    37b71bd26mov rdx, [rsp+192]
    37b71bd2ecall sscal_
    37b71bd33mov eax, [rsp+168]
    37b71bd3aimul eax, [r14]
    37b71bd3elea r9, [rip+0xd1a567]
    37b71bd45mov rcx, [rsp+200]
    37b71bd4dmovsxd rdx, eax
    37b71bd50lea r8, [r15+4*rdx]
    37b71bd54mov rdx, [rsp+192]
    37b71bd5ccall sscal_
    37b71bd61mov rax, [rsp+248]
    37b71bd69jmp 0x37b71b60f
    37b71bd6e
    37b71bd6eloc_37b71bd6e:
    37b71bd6emovaps x0, x9
    37b71bd72mov [rsp+280], rax
    37b71bd7amov eax, [rsp+160]
    37b71bd81divss x0, x13
    37b71bd86movaps x13, x9
    37b71bd8asub ecx, eax
    37b71bd8clea edx, [rcx+1]
    37b71bd8flea rcx, [rsp+428]
    37b71bd97mov [rsp+428], edx
    37b71bd9elea edx, [rax+r9-1]
    37b71bda3lea r9, [rip+0xd1a502]
    37b71bdaamovsxd rdx, edx
    37b71bdadmov [rsp+144], rcx
    37b71bdb5lea r8, [r15+4*rdx]
    37b71bdb9lea rdx, [rsp+396]
    37b71bdc1movss [rsp+396], x0
    37b71bdcacall sscal_
    37b71bdcfmov edx, [rbp]
    37b71bdd2mov eax, [rsp+160]
    37b71bdd9lea r9, [rip+0xd1a4cc]
    37b71bde0mov r11l, edx
    37b71bde3imul edx, [rsp+200]
    37b71bdebsub r11l, eax
    37b71bdeelea ecx, [r11+1]
    37b71bdf2lea edx, [rax+rdx-1]
    37b71bdf6mov [rsp+428], ecx
    37b71bdfdmov rcx, [rsp+144]
    37b71be05movsxd rdx, edx
    37b71be08lea r8, [r15+4*rdx]
    37b71be0clea rdx, [rsp+396]
    37b71be14call sscal_
    37b71be19mov edx, [rbp]
    37b71be1cmov rax, [rsp+280]
    37b71be24jmp 0x37b719fc1
    37b71be29
    37b71be29loc_37b71be29:
    37b71be29movaps x0, x9
    37b71be2dmov eax, [rsp+160]
    37b71be34lea r9, [rip+0xd1a471]
    37b71be3bdivss x0, x12
    37b71be40movaps x12, x9
    37b71be44sub edx, eax
    37b71be46lea ecx, [rax+rcx-1]
    37b71be4alea rax, [rsp+428]
    37b71be52add edx, 1
    37b71be55movsxd rcx, ecx
    37b71be58mov [rsp+152], rax
    37b71be60mov [rsp+428], edx
    37b71be67lea r8, [r15+4*rcx]
    37b71be6blea rdx, [rsp+396]
    37b71be73mov rcx, rax
    37b71be76movss [rsp+396], x0
    37b71be7fcall sscal_
    37b71be84mov edx, [r14]
    37b71be87jmp 0x37b71a899
    37b71be8c
    37b71be8cloc_37b71be8c:
    37b71be8cmov edx, [rbp]
    37b71be8fmov [rsp+280], rax
    37b71be97lea r9, [rip+0xd1a40e]
    37b71be9emov eax, [rsp+160]
    37b71bea5mov ecx, edx
    37b71bea7imul edx, [rsp+152]
    37b71beafsub ecx, eax
    37b71beb1add ecx, 1
    37b71beb4lea edx, [rax+rdx-1]
    37b71beb8mov [rsp+428], ecx
    37b71bebfmov rcx, [rsp+144]
    37b71bec7movsxd rdx, edx
    37b71becalea r8, [r15+4*rdx]
    37b71becemov rdx, r14
    37b71bed1call sscal_
    37b71bed6mov edx, [rbp]
    37b71bed9mov eax, [rsp+160]
    37b71bee0lea r9, [rip+0xd1a3c5]
    37b71bee7mov r10l, edx
    37b71beeaimul edx, [rsp+200]
    37b71bef2sub r10l, eax
    37b71bef5lea ecx, [r10+1]
    37b71bef9lea edx, [rax+rdx-1]
    37b71befdmov [rsp+428], ecx
    37b71bf04mov rcx, [rsp+144]
    37b71bf0cmovsxd rdx, edx
    37b71bf0flea r8, [r15+4*rdx]
    37b71bf13mov rdx, r14
    37b71bf16call sscal_
    37b71bf1bmov rax, [rsp+280]
    37b71bf23jmp 0x37b71a1f4
    37b71bf28
    37b71bf28loc_37b71bf28:
    37b71bf28mov edx, [r14]
    37b71bf2bmov eax, [rsp+160]
    37b71bf32lea r9, [rip+0xd1a373]
    37b71bf39mov ecx, edx
    37b71bf3bimul edx, esi
    37b71bf3esub ecx, eax
    37b71bf40add ecx, 1
    37b71bf43lea edx, [rax+rdx-1]
    37b71bf47mov [rsp+428], ecx
    37b71bf4emov rcx, [rsp+152]
    37b71bf56movsxd rdx, edx
    37b71bf59lea r8, [r15+4*rdx]
    37b71bf5dmov rdx, r11
    37b71bf60call sscal_
    37b71bf65jmp 0x37b71aa1b
    37b71bf6a
    37b71bf6aloc_37b71bf6a:
    37b71bf6amov edx, [rsp+160]
    37b71bf71imul edx, [r14]
    37b71bf75lea rcx, [rsp+392]
    37b71bf7dlea r9, [rip+0xd1a328]
    37b71bf84movsxd rdx, edx
    37b71bf87lea r8, [r15+4*rdx]
    37b71bf8bmov rdx, r11
    37b71bf8ecall sscal_
    37b71bf93jmp 0x37b719710
    37b71bf98
    37b71bf98loc_37b71bf98:
    37b71bf98imul esi, [r14]
    37b71bf9cmov [rsp+72], rbp
    37b71bfa1lea r13l, [rbx-1]
    37b71bfa5mov rdx, r14
    37b71bfa8lea rcx, [rip+0xd1a2fd]
    37b71bfaflea r9, [rip+0xd1a32a]
    37b71bfb6mov [rsp+88], 1
    37b71bfbfmov [rsp+80], rcx
    37b71bfc4lea r8, [rsp+428]
    37b71bfcclea eax, [rsi+rbx-1]
    37b71bfd0movsxd rsi, esi
    37b71bfd3mov [rsp+56], rcx
    37b71bfd8lea rcx, [rip+0xd1a305]
    37b71bfdfcdqe
    37b71bfe1mov [rsp+428], r13l
    37b71bfe9lea rax, [r15+4*rax]
    37b71bfedmov [rsp+64], rax
    37b71bff2lea rax, [r15+4*rsi]
    37b71bff6mov [rsp+48], rax
    37b71bffbmov rax, [rsp+1272]
    37b71c003mov [rsp+40], rax
    37b71c008mov rax, [rsp+1264]
    37b71c010mov [rsp+32], rax
    37b71c015call sgemv_
    37b71c01alea r8, [rip+0xd1a28b]
    37b71c021mov rdx, rbp
    37b71c024mov rcx, r14
    37b71c027call isamax_
    37b71c02cmovaps x1, x9
    37b71c030mov r8, rbp
    37b71c033mov rcx, r14
    37b71c036mov rsi, [rsp+1264]
    37b71c03ecdqe
    37b71c040lea r9, [rip+0xd1a265]
    37b71c047lea rdx, [rsp+400]
    37b71c04fadd rax, r12
    37b71c052movss x0, [rsi+4*rax]
    37b71c057andps x0, x12
    37b71c05bdivss x1, x0
    37b71c05fmovss [rsp+400], x1
    37b71c068call sscal_
    37b71c06dmov [rsp+392], r13l
    37b71c075mov [rsp+232], edi
    37b71c07cjmp 0x37b7193e0
    37b71c081
    37b71c081loc_37b71c081:
    37b71c081mov eax, [r14]
    37b71c084lea ecx, [rbx+1]
    37b71c087cmp ecx, eax
    37b71c089jg 0x37b71c0b2
    37b71c08bmov edx, eax
    37b71c08dimul eax, esi
    37b71c090movsxd rcx, ecx
    37b71c093sub edx, ebx
    37b71c095sub edx, 1
    37b71c098cdqe
    37b71c09alea r8, [4*rdx+4]
    37b71c0a2xor edx, edx
    37b71c0a4lea rax, [rax+rcx-1]
    37b71c0a9lea rcx, [r15+4*rax]
    37b71c0adcall memset
    37b71c0b2
    37b71c0b2loc_37b71c0b2:
    37b71c0b2movsxd rax, esi
    37b71c0b5mov [rsp+4*rax+444], 0
    37b71c0c0cmp [rsp+236], 1
    37b71c0c8jle 0x37b71d17e
    37b71c0ce
    37b71c0celoc_37b71c0ce:
    37b71c0cemov ebp, [rsp+232]
    37b71c0d5xor r10l, r10l
    37b71c0d8mov [rsp+232], edi
    37b71c0dfjmp 0x37b71bba2
    37b71c0e4
    37b71c0e4loc_37b71c0e4:
    37b71c0e4mov rdx, [rsp+256]
    37b71c0ecmovsxd rax, edi
    37b71c0efmov r10, [rsp+272]
    37b71c0f7imul ecx, esi
    37b71c0faimul rax, rdx
    37b71c0felea rbx, [r10+rax]
    37b71c102sub rax, rdx
    37b71c105lea edx, [rcx+r13]
    37b71c109lea r12, [r10+rax]
    37b71c10dmov rax, [rsp+1264]
    37b71c115lea rbp, [rax+4*r12+4]
    37b71c11alea rsi, [rax+4*rbx+4]
    37b71c11fcmp edi, 2
    37b71c122jle 0x37b71d671
    37b71c128sub edx, 1
    37b71c12bmovsxd rcx, ecx
    37b71c12elea eax, [rdi-2]
    37b71c131mov [rsp+72], rbp
    37b71c136movsxd rdx, edx
    37b71c139mov [rsp+428], eax
    37b71c140lea r8, [rsp+428]
    37b71c148lea r9, [rip+0xd1a191]
    37b71c14flea rdx, [r15+4*rdx]
    37b71c153mov [rsp+176], eax
    37b71c15alea rax, [rip+0xd1a14b]
    37b71c161mov [rsp+64], rdx
    37b71c166lea rdx, [r15+4*rcx]
    37b71c16alea rcx, [rip+0xd1a173]
    37b71c171mov [rsp+48], rdx
    37b71c176mov rdx, [rsp+1272]
    37b71c17emov [rsp+80], rax
    37b71c183mov [rsp+40], rdx
    37b71c188mov rdx, [rsp+1264]
    37b71c190mov [rsp+56], rax
    37b71c195mov [rsp+32], rdx
    37b71c19amov rdx, r14
    37b71c19dmov [rsp+88], 1
    37b71c1a6mov [rsp+160], r8
    37b71c1aecall sgemv_
    37b71c1b3mov edx, [rsp+168]
    37b71c1baimul edx, [r14]
    37b71c1belea rcx, [rip+0xd1a0e7]
    37b71c1c5mov eax, [rsp+176]
    37b71c1ccmov [rsp+80], rcx
    37b71c1d1lea r9, [rip+0xd1a108]
    37b71c1d8mov [rsp+56], rcx
    37b71c1ddmov r8, [rsp+160]
    37b71c1e5lea rcx, [rip+0xd1a0f8]
    37b71c1ecmov [rsp+428], eax
    37b71c1f3movsxd rax, edx
    37b71c1f6lea edx, [rax+rdi-1]
    37b71c1falea rax, [r15+4*rax]
    37b71c1femov [rsp+72], rsi
    37b71c203mov [rsp+48], rax
    37b71c208movsxd rdx, edx
    37b71c20bmov rax, [rsp+1272]
    37b71c213mov [rsp+88], 1
    37b71c21clea rdx, [r15+4*rdx]
    37b71c220mov [rsp+40], rax
    37b71c225mov rax, [rsp+1264]
    37b71c22dmov [rsp+64], rdx
    37b71c232mov rdx, r14
    37b71c235mov [rsp+32], rax
    37b71c23acall sgemv_
    37b71c23f
    37b71c23floc_37b71c23f:
    37b71c23fmov r9l, [r14]
    37b71c242test r9l, r9l
    37b71c245jle 0x37b71d7b5
    37b71c24blea eax, [r9-1]
    37b71c24fcmp eax, 2
    37b71c252jbe 0x37b71d7f1
    37b71c258mov rax, [rsp+1264]
    37b71c260mov edx, r9l
    37b71c263movss x2, [rip+0xd1a085]
    37b71c26bpxor x1, x1
    37b71c26fshr edx, 2
    37b71c272lea r8, [rax+4*r12+4]
    37b71c277lea rcx, [rax+4*rbx+4]
    37b71c27cshl rdx, 4
    37b71c280xor eax, eax
    37b71c282shufps x2, x2, 0
    37b71c286nop [rax]
    37b71c28anop [rax+rax]
    37b71c295nop [rax+rax]
    37b71c2a0
    37b71c2a0loc_37b71c2a0:
    37b71c2a0movups x0, [rcx+rax]
    37b71c2a4movups x3, [r8+rax]
    37b71c2a9add rax, 16
    37b71c2adandps x0, x2
    37b71c2b0andps x3, x2
    37b71c2b3addps x0, x3
    37b71c2b6maxps x1, x0
    37b71c2b9cmp rax, rdx
    37b71c2bcjne 0x37b71c2a0
    37b71c2bemovaps x0, x1
    37b71c2c1movhlps x0, x1
    37b71c2c4maxps x0, x1
    37b71c2c7movaps x1, x0
    37b71c2cashufps x1, x0, 85
    37b71c2cemaxps x1, x0
    37b71c2d1test r9b, 3
    37b71c2d5je 0x37b71c315
    37b71c2d7mov eax, r9l
    37b71c2daand eax, -4
    37b71c2ddadd eax, 1
    37b71c2e0
    37b71c2e0loc_37b71c2e0:
    37b71c2e0mov rdi, [rsp+1264]
    37b71c2e8cdqe
    37b71c2ealea rcx, [rdi+4*r12]
    37b71c2eelea rdx, [rdi+4*rbx]
    37b71c2f2
    37b71c2f2loc_37b71c2f2:
    37b71c2f2movss x0, [rdx+4*rax]
    37b71c2f7movss x2, [rcx+4*rax]
    37b71c2fcadd rax, 1
    37b71c300andps x0, x11
    37b71c304andps x2, x11
    37b71c308addss x0, x2
    37b71c30cmaxss x1, x0
    37b71c310cmp r9l, eax
    37b71c313jge 0x37b71c2f2
    37b71c315
    37b71c315loc_37b71c315:
    37b71c315movaps x0, x9
    37b71c319mov r8, rbp
    37b71c31cmov rcx, r14
    37b71c31fdivss x0, x1
    37b71c323lea r9, [rip+0xd19f82]
    37b71c32alea rdx, [rsp+400]
    37b71c332movss [rsp+400], x0
    37b71c33bcall sscal_
    37b71c340lea r9, [rip+0xd19f65]
    37b71c347mov r8, rsi
    37b71c34amov rcx, r14
    37b71c34dlea rdx, [rsp+400]
    37b71c355call sscal_
    37b71c35a
    37b71c35aloc_37b71c35a:
    37b71c35amov r12l, [rsp+232]
    37b71c362jmp 0x37b71bbd6
    37b71c367
    37b71c367loc_37b71c367:
    37b71c367cmp [rsp+216], 1
    37b71c370je 0x37b71c39e
    37b71c372mov edx, [rsp+200]
    37b71c379imul edx, [r14]
    37b71c37dmov rax, [rsp+216]
    37b71c385movsxd rdx, edx
    37b71c388sub eax, 1
    37b71c38blea rcx, [r15+4*rdx]
    37b71c38flea r8, [4*rax]
    37b71c397xor edx, edx
    37b71c399call memset
    37b71c39e
    37b71c39eloc_37b71c39e:
    37b71c39emovsxd rax, [rsp+200]
    37b71c3a6mov [rsp+4*rax+444], 0
    37b71c3b1cmp [rsp+236], 1
    37b71c3b9jle 0x37b71c3c8
    37b71c3bb
    37b71c3bbloc_37b71c3bb:
    37b71c3bbxor edi, edi
    37b71c3bdjmp 0x37b71af57
    37b71c3c2
    37b71c3c2 66 0f 1f 44 00 00 f..D..
    37b71c3c8
    37b71c3c8loc_37b71c3c8:
    37b71c3c8mov rax, [rsp+272]
    37b71c3d0mov rsi, [rsp+248]
    37b71c3d8lea rbx, [rax+rsi]
    37b71c3dcjmp 0x37b71ae45
    37b71c3e1
    37b71c3e1loc_37b71c3e1:
    37b71c3e1mov esi, [rsp+160]
    37b71c3e8mov ecx, [rsp+200]
    37b71c3efmov edx, eax
    37b71c3f1mov [rsp+96], r14
    37b71c3f6mov ebx, [rsp+236]
    37b71c3fdmov [rsp+72], r14
    37b71c402mov r8, r14
    37b71c405lea r10, [rsp+428]
    37b71c40dsub edx, esi
    37b71c40fsub esi, ecx
    37b71c411mov [rsp+112], 1
    37b71c41alea r9, [rsp+388]
    37b71c422add edx, ecx
    37b71c424add ebx, 1
    37b71c427mov [rsp+104], 1
    37b71c430lea rbp, [rsp+400]
    37b71c438mov [rsp+428], edx
    37b71c43fmov ecx, ebx
    37b71c441lea edx, [rsi+1]
    37b71c444lea rsi, [rip+0xd19e71]
    37b71c44bimul ecx, eax
    37b71c44eadd eax, edx
    37b71c450movsxd rdx, edx
    37b71c453mov [rsp+80], rsi
    37b71c458cdqe
    37b71c45amov rsi, [rsp+1248]
    37b71c462imul rdx, [rsp+272]
    37b71c46bmov [rsp+32], r10
    37b71c470lea rax, [r15+4*rax-4]
    37b71c475mov [rsp+152], ebx
    37b71c47cmov [rsp+64], rax
    37b71c481mov rax, [rsp+1256]
    37b71c489movsxd rcx, ecx
    37b71c48clea rcx, [r15+4*rcx]
    37b71c490mov [rsp+200], r9
    37b71c498mov [rsp+56], rax
    37b71c49dmov rax, [rsp+376]
    37b71c4a5mov [rsp+88], rcx
    37b71c4aalea rax, [rax+rdx+1]
    37b71c4aflea rdx, [rip+0xd19e2e]
    37b71c4b6lea rax, [rsi+4*rax]
    37b71c4bamov rcx, rdx
    37b71c4bdmov [rsp+48], rax
    37b71c4c2lea rax, [rip+0xd19e17]
    37b71c4c9mov [rsp+40], rax
    37b71c4cecall sgemm_
    37b71c4d3mov r12l, [rsp+388]
    37b71c4dbtest r12l, r12l
    37b71c4dejle 0x37b71c591
    37b71c4e4mov [rsp+224], edi
    37b71c4ebmov edi, 1
    37b71c4f0mov rsi, [rsp+312]
    37b71c4f8movups x12, [rsp+336]
    37b71c501jmp 0x37b71c535
    37b71c503
    37b71c503 0f 1f 44 00 00 ..D..
    37b71c508
    37b71c508loc_37b71c508:
    37b71c508cmp edx, 1
    37b71c50bje 0x37b71c615
    37b71c511cdqe
    37b71c513add ebx, 1
    37b71c516
    37b71c516loc_37b71c516:
    37b71c516lea r8, [r15+4*rax]
    37b71c51amov rdx, rbp
    37b71c51dmov rcx, r14
    37b71c520add rdi, 1
    37b71c524lea r9, [rip+0xd19d81]
    37b71c52bcall sscal_
    37b71c530cmp r12l, edi
    37b71c533jl 0x37b71c58a
    37b71c535
    37b71c535loc_37b71c535:
    37b71c535mov ecx, [r14]
    37b71c538mov edx, [rsi+4*rdi-4]
    37b71c53cmov eax, ebx
    37b71c53eimul eax, ecx
    37b71c541test edx, edx
    37b71c543jne 0x37b71c508
    37b71c545cdqe
    37b71c547lea r8, [rip+0xd19d5e]
    37b71c54emov rcx, r14
    37b71c551lea rdx, [r15+4*rax]
    37b71c555call isamax_
    37b71c55amovaps x1, x9
    37b71c55emov edx, eax
    37b71c560mov eax, [r14]
    37b71c563imul eax, ebx
    37b71c566add ebx, 1
    37b71c569add edx, eax
    37b71c56bcdqe
    37b71c56dmovsxd rdx, edx
    37b71c570movss x0, [r15+4*rdx-4]
    37b71c577andps x0, x12
    37b71c57bdivss x1, x0
    37b71c57fmovss [rsp+400], x1
    37b71c588jmp 0x37b71c516
    37b71c58a
    37b71c58aloc_37b71c58a:
    37b71c58amov edi, [rsp+224]
    37b71c591
    37b71c591loc_37b71c591:
    37b71c591mov edx, [rsp+160]
    37b71c598mov rsi, [rsp+1256]
    37b71c5a0lea rcx, [rip+0xd19d19]
    37b71c5a7mov eax, [rsp+152]
    37b71c5aeimul eax, [r14]
    37b71c5b2mov [rsp+56], 1
    37b71c5bbsub edx, r12l
    37b71c5bemov [rsp+48], rsi
    37b71c5c3mov rsi, [rsp+376]
    37b71c5cbadd edx, 1
    37b71c5cemov [rsp+32], r14
    37b71c5d3mov r8, [rsp+200]
    37b71c5dbmovsxd rdx, edx
    37b71c5decdqe
    37b71c5e0imul rdx, [rsp+272]
    37b71c5e9lea r9, [r15+4*rax]
    37b71c5edlea rdx, [rsi+rdx+1]
    37b71c5f2mov rsi, [rsp+1248]
    37b71c5falea rdx, [rsi+4*rdx]
    37b71c5femov [rsp+40], rdx
    37b71c603mov rdx, r14
    37b71c606call slacpy_
    37b71c60bmov edx, 1
    37b71c610jmp 0x37b71af83
    37b71c615
    37b71c615loc_37b71c615:
    37b71c615test ecx, ecx
    37b71c617jle 0x37b71d7be
    37b71c61dadd ebx, 1
    37b71c620lea r9l, [rcx-1]
    37b71c624mov r8l, ebx
    37b71c627imul r8l, ecx
    37b71c62bcmp r9l, 2
    37b71c62fjbe 0x37b71d7d5
    37b71c635movsxd r10, r8l
    37b71c638mov r8l, ecx
    37b71c63bmovss x2, [rip+0xd19cad]
    37b71c643cdqe
    37b71c645shr r8l, 2
    37b71c649lea r11, [r15+4*rax]
    37b71c64dlea r9, [r15+4*r10]
    37b71c651xor edx, edx
    37b71c653shl r8, 4
    37b71c657pxor x1, x1
    37b71c65bshufps x2, x2, 0
    37b71c65fnop
    37b71c660
    37b71c660loc_37b71c660:
    37b71c660movups x0, [r9+rdx]
    37b71c665movups x3, [r11+rdx]
    37b71c66aadd rdx, 16
    37b71c66eandps x0, x2
    37b71c671andps x3, x2
    37b71c674addps x0, x3
    37b71c677maxps x1, x0
    37b71c67acmp rdx, r8
    37b71c67djne 0x37b71c660
    37b71c67fmovaps x0, x1
    37b71c682mov edx, ecx
    37b71c684movhlps x0, x1
    37b71c687maxps x0, x1
    37b71c68aand edx, -4
    37b71c68dadd edx, 1
    37b71c690movaps x1, x0
    37b71c693shufps x1, x0, 85
    37b71c697maxps x1, x0
    37b71c69atest cl, 3
    37b71c69dje 0x37b71c6d0
    37b71c69f
    37b71c69floc_37b71c69f:
    37b71c69fmovsxd rdx, edx
    37b71c6a2lea r9, [r15+4*rax]
    37b71c6a6lea r8, [r15+4*r10]
    37b71c6aa
    37b71c6aaloc_37b71c6aa:
    37b71c6aamovss x0, [r8+4*rdx-4]
    37b71c6b1movss x2, [r9+4*rdx-4]
    37b71c6b8add rdx, 1
    37b71c6bcandps x0, x12
    37b71c6c0andps x2, x12
    37b71c6c4addss x0, x2
    37b71c6c8maxss x1, x0
    37b71c6cccmp ecx, edx
    37b71c6cejge 0x37b71c6aa
    37b71c6d0
    37b71c6d0loc_37b71c6d0:
    37b71c6d0movaps x0, x9
    37b71c6d4divss x0, x1
    37b71c6d8movss [rsp+400], x0
    37b71c6e1jmp 0x37b71c516
    37b71c6e6
    37b71c6e6loc_37b71c6e6:
    37b71c6e6mov rax, [rsp+248]
    37b71c6eemov r10, [rsp+272]
    37b71c6f6lea r8l, [rcx-1]
    37b71c6famov r11l, [rsp+160]
    37b71c702mov rsi, [rsp+1248]
    37b71c70alea rbp, [4*rax]
    37b71c712lea rbx, [r10+rax]
    37b71c716mov eax, [rsp+152]
    37b71c71dlea r13, [4*rbx]
    37b71c725lea rdi, [rsi+rbp+4]
    37b71c72aimul eax, ecx
    37b71c72dlea rsi, [rsi+r13+4]
    37b71c732lea edx, [r11+rax-1]
    37b71c737movsxd rdx, edx
    37b71c73alea rdx, [r15+4*rdx]
    37b71c73ecmp r8l, [rsp+332]
    37b71c746jle 0x37b71d6b6
    37b71c74csub ecx, r11l
    37b71c74fmov [rsp+64], rdx
    37b71c754mov rdx, r14
    37b71c757sub ecx, 1
    37b71c75amov [rsp+72], rdi
    37b71c75fmov [rsp+428], ecx
    37b71c766mov ecx, [rsp+176]
    37b71c76dmov [rsp+88], 1
    37b71c776mov r9, rcx
    37b71c779imul rcx, r10
    37b71c77dmov r10, [rsp+376]
    37b71c785lea eax, [r9+rax-1]
    37b71c78alea r9, [rip+0xd19b4f]
    37b71c791cdqe
    37b71c793lea rcx, [r10+rcx+1]
    37b71c798lea rax, [r15+4*rax]
    37b71c79cmov r10, [rsp+1248]
    37b71c7a4mov [rsp+48], rax
    37b71c7a9mov rax, [rsp+1256]
    37b71c7b1lea r11, [r10+4*rcx]
    37b71c7b5lea rcx, [rip+0xd19af0]
    37b71c7bcmov [rsp+80], rcx
    37b71c7c1lea r10, [rsp+428]
    37b71c7c9mov [rsp+56], rcx
    37b71c7cemov r8, r10
    37b71c7d1lea rcx, [rip+0xd19b0c]
    37b71c7d8mov [rsp+40], rax
    37b71c7ddmov [rsp+32], r11
    37b71c7e2mov [rsp+152], r11
    37b71c7eamov [rsp+144], r10
    37b71c7f2call sgemv_
    37b71c7f7mov eax, [r14]
    37b71c7famov r9l, [rsp+176]
    37b71c802lea rcx, [rip+0xd19aa3]
    37b71c809mov r11, [rsp+152]
    37b71c811mov [rsp+80], rcx
    37b71c816mov edx, eax
    37b71c818sub edx, [rsp+160]
    37b71c81fmov [rsp+56], rcx
    37b71c824lea rcx, [rip+0xd19ab9]
    37b71c82bimul eax, [rsp+200]
    37b71c833sub edx, 1
    37b71c836mov [rsp+88], 1
    37b71c83fmov [rsp+428], edx
    37b71c846mov edx, [rsp+232]
    37b71c84dmov [rsp+72], rsi
    37b71c852mov r8, [rsp+144]
    37b71c85alea edx, [rax+rdx-1]
    37b71c85elea eax, [rax+r9-1]
    37b71c863mov [rsp+32], r11
    37b71c868cdqe
    37b71c86amovsxd rdx, edx
    37b71c86dlea r9, [rip+0xd19a6c]
    37b71c874lea rax, [r15+4*rax]
    37b71c878lea rdx, [r15+4*rdx]
    37b71c87cmov [rsp+48], rax
    37b71c881mov rax, [rsp+1256]
    37b71c889mov [rsp+64], rdx
    37b71c88emov rdx, r14
    37b71c891mov [rsp+40], rax
    37b71c896call sgemv_
    37b71c89b
    37b71c89bloc_37b71c89b:
    37b71c89bmov edx, [r14]
    37b71c89etest edx, edx
    37b71c8a0jle 0x37b71d7cc
    37b71c8a6lea eax, [rdx-1]
    37b71c8a9cmp eax, 2
    37b71c8acjbe 0x37b71d7e3
    37b71c8b2mov ecx, edx
    37b71c8b4movss x1, [rip+0xd19a34]
    37b71c8bcxor eax, eax
    37b71c8bepxor x2, x2
    37b71c8c2shr ecx, 2
    37b71c8c5shl rcx, 4
    37b71c8c9shufps x1, x1, 0
    37b71c8cdnop [rax+rax]
    37b71c8d5nop [rax+rax]
    37b71c8e0
    37b71c8e0loc_37b71c8e0:
    37b71c8e0movups x0, [rsi+rax]
    37b71c8e4movups x3, [rdi+rax]
    37b71c8e8add rax, 16
    37b71c8ecandps x0, x1
    37b71c8efandps x3, x1
    37b71c8f2addps x0, x3
    37b71c8f5maxps x2, x0
    37b71c8f8cmp rax, rcx
    37b71c8fbjne 0x37b71c8e0
    37b71c8fdmovaps x0, x2
    37b71c900movhlps x0, x2
    37b71c903maxps x0, x2
    37b71c906movaps x1, x0
    37b71c909shufps x1, x0, 85
    37b71c90dmaxps x1, x0
    37b71c910test dl, 3
    37b71c913je 0x37b71c958
    37b71c915mov eax, edx
    37b71c917and eax, -4
    37b71c91aadd eax, 1
    37b71c91d
    37b71c91dloc_37b71c91d:
    37b71c91dmov r8, [rsp+1248]
    37b71c925mov rcx, [rsp+1248]
    37b71c92dcdqe
    37b71c92fadd r8, rbp
    37b71c932add rcx, r13
    37b71c935
    37b71c935loc_37b71c935:
    37b71c935movss x0, [rcx+4*rax]
    37b71c93amovss x2, [r8+4*rax]
    37b71c940add rax, 1
    37b71c944andps x0, x10
    37b71c948andps x2, x10
    37b71c94caddss x0, x2
    37b71c950maxss x1, x0
    37b71c954cmp edx, eax
    37b71c956jge 0x37b71c935
    37b71c958
    37b71c958loc_37b71c958:
    37b71c958movaps x0, x9
    37b71c95cmov r8, rdi
    37b71c95flea r13, [r12+1]
    37b71c964mov rcx, r14
    37b71c967divss x0, x1
    37b71c96blea rbp, [rsp+400]
    37b71c973lea r9, [rip+0xd19932]
    37b71c97amov rdx, rbp
    37b71c97dmovss [rsp+400], x0
    37b71c986call sscal_
    37b71c98bmov r8, rsi
    37b71c98emov rdx, rbp
    37b71c991mov rcx, r14
    37b71c994lea r9, [rip+0xd19911]
    37b71c99bcall sscal_
    37b71c9a0jmp 0x37b71afa6
    37b71c9a5
    37b71c9a5loc_37b71c9a5:
    37b71c9a5mov eax, [rsp+152]
    37b71c9acmov [rsp+200], eax
    37b71c9b3cmp [rsp+236], 1
    37b71c9bbjg 0x37b71af3f
    37b71c9c1
    37b71c9c1loc_37b71c9c1:
    37b71c9c1mov rax, [rsp+272]
    37b71c9c9mov rsi, [rsp+248]
    37b71c9d1lea r13, [r12+1]
    37b71c9d6lea rbx, [rax+rsi]
    37b71c9dajmp 0x37b71afa6
    37b71c9df
    37b71c9dfloc_37b71c9df:
    37b71c9dfmov eax, [rsp+236]
    37b71c9e6mov [rsp+392], r13l
    37b71c9eemov [rsp+388], eax
    37b71c9f5mov rax, [rsp+1288]
    37b71c9fdmov eax, [rax]
    37b71c9ffmov [rsp+232], eax
    37b71ca06jmp 0x37b7192fa
    37b71ca0b
    37b71ca0bloc_37b71ca0b:
    37b71ca0bmov edi, [rsp+236]
    37b71ca12lea rcx, [rip+0xd198a3]
    37b71ca19mov r8, r14
    37b71ca1clea r9, [rsp+424]
    37b71ca24mov [rsp+192], r10l
    37b71ca2cmov eax, edi
    37b71ca2elea r13l, [rdi+rsi]
    37b71ca32sub eax, esi
    37b71ca34mov edx, r13l
    37b71ca37add eax, 1
    37b71ca3amov [rsp+160], eax
    37b71ca41mov [rsp+424], eax
    37b71ca48lea eax, [rbx+rdi]
    37b71ca4bsub eax, esi
    37b71ca4dmov [rsp+428], eax
    37b71ca54mov eax, [r14]
    37b71ca57mov [rsp+80], rcx
    37b71ca5cmov rcx, [rsp+1272]
    37b71ca64imul edx, eax
    37b71ca67mov [rsp+96], r14
    37b71ca6cimul eax, esi
    37b71ca6fmov [rsp+56], rcx
    37b71ca74mov rcx, [rsp+1264]
    37b71ca7cmov [rsp+112], 1
    37b71ca85mov [rsp+48], rcx
    37b71ca8amovsxd rdx, edx
    37b71ca8dlea rcx, [rip+0xd1984c]
    37b71ca94cdqe
    37b71ca96mov [rsp+40], rcx
    37b71ca9blea rdx, [r15+4*rdx]
    37b71ca9flea rcx, [rsp+428]
    37b71caa7lea rax, [r15+4*rax]
    37b71caabmov [rsp+88], rdx
    37b71cab0lea rdx, [rip+0xd1982d]
    37b71cab7mov [rsp+32], rcx
    37b71cabcmov [rsp+104], 1
    37b71cac5mov [rsp+72], r14
    37b71cacamov [rsp+64], rax
    37b71cacfmov [rsp+168], rcx
    37b71cad7mov rcx, rdx
    37b71cadacall sgemm_
    37b71cadfcmp [rsp+236], esi
    37b71cae6mov r10l, [rsp+192]
    37b71caeejl 0x37b71cbb9
    37b71caf4mov [rsp+192], ebx
    37b71cafbmov rdi, [rsp+312]
    37b71cb03movsxd rbx, esi
    37b71cb06mov [rsp+200], ebp
    37b71cb0dmov esi, [rsp+236]
    37b71cb14mov [rsp+248], r10l
    37b71cb1cjmp 0x37b71cb4e
    37b71cb1e
    37b71cb1e 66 90 f.
    37b71cb20
    37b71cb20loc_37b71cb20:
    37b71cb20cmp edx, 1
    37b71cb23je 0x37b71cd63
    37b71cb29cdqe
    37b71cb2b
    37b71cb2bloc_37b71cb2b:
    37b71cb2blea r9, [rip+0xd1977a]
    37b71cb32lea r8, [r15+4*rax]
    37b71cb36mov rcx, r14
    37b71cb39add rbx, 1
    37b71cb3dlea rdx, [rsp+400]
    37b71cb45call sscal_
    37b71cb4acmp esi, ebx
    37b71cb4cjl 0x37b71cba3
    37b71cb4e
    37b71cb4eloc_37b71cb4e:
    37b71cb4emov ecx, [r14]
    37b71cb51lea ebp, [rsi+rbx]
    37b71cb54mov edx, [rdi+4*rbx-4]
    37b71cb58mov eax, ebp
    37b71cb5aimul eax, ecx
    37b71cb5dtest edx, edx
    37b71cb5fjne 0x37b71cb20
    37b71cb61cdqe
    37b71cb63lea r8, [rip+0xd19742]
    37b71cb6amov rcx, r14
    37b71cb6dlea rdx, [r15+4*rax]
    37b71cb71call isamax_
    37b71cb76movaps x1, x9
    37b71cb7amov edx, eax
    37b71cb7cmov eax, [r14]
    37b71cb7fimul eax, ebp
    37b71cb82add edx, eax
    37b71cb84cdqe
    37b71cb86movsxd rdx, edx
    37b71cb89movss x0, [r15+4*rdx-4]
    37b71cb90andps x0, x11
    37b71cb94divss x1, x0
    37b71cb98movss [rsp+400], x1
    37b71cba1jmp 0x37b71cb2b
    37b71cba3
    37b71cba3loc_37b71cba3:
    37b71cba3mov ebx, [rsp+192]
    37b71cbaamov ebp, [rsp+200]
    37b71cbb1mov r10l, [rsp+248]
    37b71cbb9
    37b71cbb9loc_37b71cbb9:
    37b71cbb9mov eax, [rsp+160]
    37b71cbc0mov rsi, [rsp+1272]
    37b71cbc8movsxd rbx, ebx
    37b71cbcblea rcx, [rip+0xd196ee]
    37b71cbd2mov r8, [rsp+168]
    37b71cbdamov [rsp+192], r10l
    37b71cbe2imul rbx, [rsp+256]
    37b71cbebmov [rsp+428], eax
    37b71cbf2mov eax, [r14]
    37b71cbf5mov [rsp+48], rsi
    37b71cbfamov rsi, [rsp+272]
    37b71cc02mov [rsp+32], r14
    37b71cc07mov [rsp+56], 1
    37b71cc10imul eax, r13l
    37b71cc14lea rdx, [rsi+rbx+1]
    37b71cc19mov rsi, [rsp+1264]
    37b71cc21lea rdx, [rsi+4*rdx]
    37b71cc25cdqe
    37b71cc27mov [rsp+40], rdx
    37b71cc2clea r9, [r15+4*rax]
    37b71cc30mov rdx, r14
    37b71cc33call slacpy_
    37b71cc38mov eax, [rsp+236]
    37b71cc3fmov r13l, [rsp+392]
    37b71cc47mov r10l, [rsp+192]
    37b71cc4fmov [rsp+388], eax
    37b71cc56jmp 0x37b71bbc6
    37b71cc5b
    37b71cc5bloc_37b71cc5b:
    37b71cc5bmov ecx, [rsp+200]
    37b71cc62mov edx, eax
    37b71cc64mov edi, [rsp+160]
    37b71cc6bmov [rsp+88], 1
    37b71cc74mov [rsp+72], rsi
    37b71cc79lea r9, [rip+0xd19660]
    37b71cc80lea r8, [rsp+428]
    37b71cc88imul ecx, eax
    37b71cc8bsub edx, edi
    37b71cc8dmov [rsp+428], edx
    37b71cc94mov eax, ecx
    37b71cc96lea rcx, [rip+0xd1960f]
    37b71cc9dlea edx, [rax+rdi-1]
    37b71cca1mov edi, [rsp+232]
    37b71cca8mov [rsp+80], rcx
    37b71ccadmovsxd rdx, edx
    37b71ccb0mov [rsp+56], rcx
    37b71ccb5lea rcx, [rip+0xd19628]
    37b71ccbclea eax, [rax+rdi-1]
    37b71ccc0lea rdx, [r15+4*rdx]
    37b71ccc4cdqe
    37b71ccc6mov [rsp+64], rdx
    37b71cccbmov rdx, r14
    37b71cccelea rax, [r15+4*rax]
    37b71ccd2mov [rsp+48], rax
    37b71ccd7mov rax, [rsp+1256]
    37b71ccdfmov [rsp+40], rax
    37b71cce4mov rax, [rsp+1248]
    37b71cceclea rax, [rax+4*rbx+4]
    37b71ccf1mov [rsp+32], rax
    37b71ccf6call sgemv_
    37b71ccfbjmp 0x37b71adea
    37b71cd00
    37b71cd00loc_37b71cd00:
    37b71cd00mov eax, r8l
    37b71cd03mov [rsp+48], r14
    37b71cd08mov r8, [rsp+312]
    37b71cd10lea rcx, [rip+0xd195a9]
    37b71cd17sub eax, edx
    37b71cd19mov [rsp+40], r15
    37b71cd1emov [rsp+56], 1
    37b71cd27cdq
    37b71cd28idiv r10l
    37b71cd2bmov edx, 128
    37b71cd30cmp eax, edx
    37b71cd32cmovle edx, eax
    37b71cd35lea eax, [rdx+rdx+1]
    37b71cd39mov [rsp+236], edx
    37b71cd40mov rdx, r14
    37b71cd43mov [rsp+448], eax
    37b71cd4alea rax, [rip+0xd1956b]
    37b71cd51mov [rsp+32], rax
    37b71cd56mov r9, rax
    37b71cd59call slaset_
    37b71cd5ejmp 0x37b7191bd
    37b71cd63
    37b71cd63loc_37b71cd63:
    37b71cd63test ecx, ecx
    37b71cd65jle 0x37b71d877
    37b71cd6blea r8l, [rsi+rbx+1]
    37b71cd70lea r9l, [rcx-1]
    37b71cd74imul r8l, ecx
    37b71cd78cmp r9l, 2
    37b71cd7cjbe 0x37b71d869
    37b71cd82movsxd r10, r8l
    37b71cd85mov r8l, ecx
    37b71cd88movss x2, [rip+0xd19560]
    37b71cd90cdqe
    37b71cd92shr r8l, 2
    37b71cd96lea r11, [r15+4*rax]
    37b71cd9alea r9, [r15+4*r10]
    37b71cd9exor edx, edx
    37b71cda0shl r8, 4
    37b71cda4pxor x1, x1
    37b71cda8shufps x2, x2, 0
    37b71cdacnop [rax+rax]
    37b71cdb5nop [rax+rax]
    37b71cdc0
    37b71cdc0loc_37b71cdc0:
    37b71cdc0movups x0, [r9+rdx]
    37b71cdc5movups x3, [r11+rdx]
    37b71cdcaadd rdx, 16
    37b71cdceandps x0, x2
    37b71cdd1andps x3, x2
    37b71cdd4addps x0, x3
    37b71cdd7maxps x1, x0
    37b71cddacmp rdx, r8
    37b71cdddjne 0x37b71cdc0
    37b71cddfmovaps x0, x1
    37b71cde2movhlps x0, x1
    37b71cde5maxps x0, x1
    37b71cde8movaps x1, x0
    37b71cdebshufps x1, x0, 85
    37b71cdefmaxps x1, x0
    37b71cdf2test cl, 3
    37b71cdf5je 0x37b71ce30
    37b71cdf7mov edx, ecx
    37b71cdf9and edx, -4
    37b71cdfcadd edx, 1
    37b71cdff
    37b71cdffloc_37b71cdff:
    37b71cdffmovsxd rdx, edx
    37b71ce02lea r9, [r15+4*rax]
    37b71ce06lea r8, [r15+4*r10]
    37b71ce0a
    37b71ce0aloc_37b71ce0a:
    37b71ce0amovss x0, [r8+4*rdx-4]
    37b71ce11movss x2, [r9+4*rdx-4]
    37b71ce18add rdx, 1
    37b71ce1candps x0, x11
    37b71ce20andps x2, x11
    37b71ce24addss x0, x2
    37b71ce28maxss x1, x0
    37b71ce2ccmp ecx, edx
    37b71ce2ejge 0x37b71ce0a
    37b71ce30
    37b71ce30loc_37b71ce30:
    37b71ce30movaps x0, x9
    37b71ce34divss x0, x1
    37b71ce38movss [rsp+400], x0
    37b71ce41jmp 0x37b71cb2b
    37b71ce46
    37b71ce46loc_37b71ce46:
    37b71ce46mov edi, [rsp+160]
    37b71ce4dmov eax, ecx
    37b71ce4fmov rbp, [rsp+272]
    37b71ce57lea r10, [rsp+428]
    37b71ce5fmov r13, [rsp+216]
    37b71ce67imul ecx, [rsp+152]
    37b71ce6fmov [rsp+280], r10
    37b71ce77sub eax, edi
    37b71ce79mov rbx, rbp
    37b71ce7cmov rdx, [rsp+1248]
    37b71ce84add eax, 1
    37b71ce87mov [rsp+428], eax
    37b71ce8emovsxd rax, [rsp+328]
    37b71ce96imul rbx, rax
    37b71ce9amov rax, [rsp+376]
    37b71cea2lea rsi, [rbx+rax]
    37b71cea6lea rax, [rsi+r13]
    37b71ceaalea r9, [rdx+4*rax]
    37b71ceaelea eax, [rdi+rcx-1]
    37b71ceb2lea rcx, [rip+0xd193f3]
    37b71ceb9cdqe
    37b71cebbmov [rsp+224], r9
    37b71cec3mov [rsp+32], rcx
    37b71cec8lea rdx, [r15+4*rax]
    37b71ceccmov r8, rcx
    37b71cecfmov rcx, r10
    37b71ced2call scopy_
    37b71ced7mov edx, [r14]
    37b71cedamov ecx, edi
    37b71cedcmov [rsp+216], r13
    37b71cee4mov r10, [rsp+280]
    37b71ceecmov r11l, edx
    37b71ceefsub r11l, edi
    37b71cef2mov edi, [rsp+328]
    37b71cef9lea eax, [r11+1]
    37b71cefdmov r11, [rsp+1248]
    37b71cf05add edi, 1
    37b71cf08mov [rsp+428], eax
    37b71cf0fmov rax, [rsp+376]
    37b71cf17mov [rsp+144], edi
    37b71cf1elea rdi, [rbx+rbp]
    37b71cf22lea rbp, [rdi+rax]
    37b71cf26lea rax, [rbp+r13]
    37b71cf2blea r13, [r11+4*rax]
    37b71cf2fmov r11l, [rsp+200]
    37b71cf37mov [rsp+200], r10
    37b71cf3fmov r9, r13
    37b71cf42imul r11l, edx
    37b71cf46lea eax, [rcx+r11-1]
    37b71cf4blea rcx, [rip+0xd1935a]
    37b71cf52mov [rsp+32], rcx
    37b71cf57cdqe
    37b71cf59mov r8, rcx
    37b71cf5cmov rcx, r10
    37b71cf5flea rdx, [r15+4*rax]
    37b71cf63call scopy_
    37b71cf68mov r9l, [r14]
    37b71cf6bcmp r9l, [rsp+332]
    37b71cf73mov r10, [rsp+200]
    37b71cf7bjl 0x37b71d860
    37b71cf81mov rcx, [rsp+216]
    37b71cf89mov edx, r9l
    37b71cf8csub edx, ecx
    37b71cf8ecmp edx, 2
    37b71cf91jbe 0x37b71d882
    37b71cf97mov rax, [rsp+376]
    37b71cf9flea r11l, [r9+1]
    37b71cfa3movss x1, [rip+0xd19345]
    37b71cfabpxor x2, x2
    37b71cfafsub r11l, ecx
    37b71cfb2add rax, rcx
    37b71cfb5shufps x1, x1, 0
    37b71cfb9lea rdx, [rbx+rax]
    37b71cfbdmov rbx, [rsp+1248]
    37b71cfc5add rax, rdi
    37b71cfc8lea r8, [rbx+4*rdx]
    37b71cfccmov edx, r11l
    37b71cfcflea rcx, [rbx+4*rax]
    37b71cfd3xor eax, eax
    37b71cfd5shr edx, 2
    37b71cfd8shl rdx, 4
    37b71cfdcnop [rax]
    37b71cfe0
    37b71cfe0loc_37b71cfe0:
    37b71cfe0movups x0, [rcx+rax]
    37b71cfe4movups x3, [r8+rax]
    37b71cfe9add rax, 16
    37b71cfedandps x0, x1
    37b71cff0andps x3, x1
    37b71cff3addps x0, x3
    37b71cff6maxps x2, x0
    37b71cff9cmp rdx, rax
    37b71cffcjne 0x37b71cfe0
    37b71cffemovaps x0, x2
    37b71d001movhlps x0, x2
    37b71d004maxps x0, x2
    37b71d007movaps x1, x0
    37b71d00ashufps x1, x0, 85
    37b71d00emaxps x1, x0
    37b71d011test r11b, 3
    37b71d015je 0x37b71d059
    37b71d017mov eax, r11l
    37b71d01aand eax, -4
    37b71d01dadd eax, [rsp+160]
    37b71d024
    37b71d024loc_37b71d024:
    37b71d024mov rdi, [rsp+1248]
    37b71d02ccdqe
    37b71d02elea rcx, [rdi+4*rsi]
    37b71d032lea rdx, [rdi+4*rbp]
    37b71d036
    37b71d036loc_37b71d036:
    37b71d036movss x0, [rdx+4*rax]
    37b71d03bmovss x2, [rcx+4*rax]
    37b71d040add rax, 1
    37b71d044andps x0, x10
    37b71d048andps x2, x10
    37b71d04caddss x0, x2
    37b71d050maxss x1, x0
    37b71d054cmp r9l, eax
    37b71d057jge 0x37b71d036
    37b71d059
    37b71d059loc_37b71d059:
    37b71d059movaps x0, x9
    37b71d05dmov eax, r9l
    37b71d060mov edi, [rsp+160]
    37b71d067mov r8, [rsp+224]
    37b71d06fdivss x0, x1
    37b71d073lea rbp, [rsp+400]
    37b71d07bmov rcx, r10
    37b71d07emov [rsp+160], r10
    37b71d086sub eax, edi
    37b71d088lea r9, [rip+0xd1921d]
    37b71d08fmov rdx, rbp
    37b71d092add eax, 1
    37b71d095mov [rsp+428], eax
    37b71d09cmovss [rsp+400], x0
    37b71d0a5call sscal_
    37b71d0aamov eax, [r14]
    37b71d0admov r8, r13
    37b71d0b0mov rdx, rbp
    37b71d0b3mov rcx, [rsp+160]
    37b71d0bblea r9, [rip+0xd191ea]
    37b71d0c2sub eax, edi
    37b71d0c4add eax, 1
    37b71d0c7mov [rsp+428], eax
    37b71d0cecall sscal_
    37b71d0d3mov rdi, [rsp+216]
    37b71d0dbcmp rdi, 1
    37b71d0dfje 0x37b71c9a5
    37b71d0e5mov rax, [rsp+1248]
    37b71d0edlea edx, [rdi-2]
    37b71d0f0mov rcx, [rsp+272]
    37b71d0f8add rdx, rsi
    37b71d0fblea rax, [rax+4*rsi+4]
    37b71d100mov rsi, [rsp+1248]
    37b71d108lea rdx, [rsi+4*rdx+8]
    37b71d10dnop [rax+rax]
    37b71d115nop [rax+rax]
    37b71d120
    37b71d120loc_37b71d120:
    37b71d120mov [rax], 0
    37b71d126mov [rax+4*rcx], 0
    37b71d12dadd rax, 4
    37b71d131cmp rdx, rax
    37b71d134jne 0x37b71d120
    37b71d136mov eax, [rsp+388]
    37b71d13dmov [rsp+200], eax
    37b71d144cmp [rsp+236], 1
    37b71d14cjg 0x37b71af3f
    37b71d152mov rax, [rsp+272]
    37b71d15amov rdi, [rsp+248]
    37b71d162lea r13, [r12+1]
    37b71d167lea rbx, [rax+rdi]
    37b71d16bjmp 0x37b71afa6
    37b71d170
    37b71d170loc_37b71d170:
    37b71d170cmp [rsp+236], 1
    37b71d178jg 0x37b71c0ce
    37b71d17e
    37b71d17eloc_37b71d17e:
    37b71d17elea r13l, [rbx-1]
    37b71d182mov [rsp+392], r13l
    37b71d18atest r13l, r13l
    37b71d18djle 0x37b719b23
    37b71d193mov [rsp+232], edi
    37b71d19ajmp 0x37b7193e0
    37b71d19f
    37b71d19floc_37b71d19f:
    37b71d19fmov eax, [rsp+232]
    37b71d1a6mov r12, [rsp+256]
    37b71d1aeimul ecx, esi
    37b71d1b1mov rdi, [rsp+1264]
    37b71d1b9lea r13l, [rax-1]
    37b71d1bdmov rbx, r12
    37b71d1c0movsxd rax, r13l
    37b71d1c3imul rbx, rax
    37b71d1c7mov rax, [rsp+272]
    37b71d1cflea rbp, [rbx+rax]
    37b71d1d3movsxd rax, ecx
    37b71d1d6lea rcx, [rip+0xd190cf]
    37b71d1ddmov [rsp+32], rcx
    37b71d1e2lea rdi, [rdi+4*rbp+4]
    37b71d1e7lea rdx, [r15+4*rax]
    37b71d1ebmov r8, rcx
    37b71d1eemov r9, rdi
    37b71d1f1lea rcx, [rsp+392]
    37b71d1f9call scopy_
    37b71d1femov rdx, [rsp+272]
    37b71d206lea rax, [rbx+r12]
    37b71d20amov [rsp+256], r12
    37b71d212lea rcx, [rsp+392]
    37b71d21alea r12, [rax+rdx]
    37b71d21emov eax, [rsp+168]
    37b71d225mov rdx, [rsp+1264]
    37b71d22dimul eax, [r14]
    37b71d231lea r9, [rdx+4*r12+4]
    37b71d236mov [rsp+160], r9
    37b71d23ecdqe
    37b71d240lea rdx, [r15+4*rax]
    37b71d244lea rax, [rip+0xd19061]
    37b71d24bmov [rsp+32], rax
    37b71d250mov r8, rax
    37b71d253call scopy_
    37b71d258mov r8l, [rsp+392]
    37b71d260test r8l, r8l
    37b71d263jle 0x37b71d80d
    37b71d269lea eax, [r8-1]
    37b71d26dcmp eax, 2
    37b71d270jbe 0x37b71d7ff
    37b71d276mov rcx, [rsp+1264]
    37b71d27emov rax, rbx
    37b71d281mov edx, r8l
    37b71d284sub rax, [rsp+256]
    37b71d28cmovss x2, [rip+0xd1905c]
    37b71d294shr edx, 2
    37b71d297pxor x1, x1
    37b71d29blea rcx, [rcx+4*rax]
    37b71d29fmov rax, [rsp+1264]
    37b71d2a7shl rdx, 4
    37b71d2abshufps x2, x2, 0
    37b71d2aflea r9, [rax+4*rbx]
    37b71d2b3xor eax, eax
    37b71d2b5nop [rax+rax]
    37b71d2c0
    37b71d2c0loc_37b71d2c0:
    37b71d2c0movups x0, [r9+rax]
    37b71d2c5movups x3, [rcx+rax]
    37b71d2c9add rax, 16
    37b71d2cdandps x0, x2
    37b71d2d0andps x3, x2
    37b71d2d3addps x0, x3
    37b71d2d6maxps x1, x0
    37b71d2d9cmp rax, rdx
    37b71d2dcjne 0x37b71d2c0
    37b71d2demovaps x0, x1
    37b71d2e1movhlps x0, x1
    37b71d2e4maxps x0, x1
    37b71d2e7movaps x1, x0
    37b71d2eashufps x1, x0, 85
    37b71d2eemaxps x1, x0
    37b71d2f1test r8b, 3
    37b71d2f5je 0x37b71d335
    37b71d2f7mov eax, r8l
    37b71d2faand eax, -4
    37b71d2fdadd eax, 1
    37b71d300
    37b71d300loc_37b71d300:
    37b71d300mov rbx, [rsp+1264]
    37b71d308cdqe
    37b71d30alea rcx, [rbx+4*rbp]
    37b71d30elea rdx, [rbx+4*r12]
    37b71d312
    37b71d312loc_37b71d312:
    37b71d312movss x0, [rdx+4*rax]
    37b71d317movss x2, [rcx+4*rax]
    37b71d31cadd rax, 1
    37b71d320andps x0, x11
    37b71d324andps x2, x11
    37b71d328addss x0, x2
    37b71d32cmaxss x1, x0
    37b71d330cmp r8l, eax
    37b71d333jge 0x37b71d312
    37b71d335
    37b71d335loc_37b71d335:
    37b71d335movaps x0, x9
    37b71d339mov r8, rdi
    37b71d33clea r9, [rip+0xd18f69]
    37b71d343divss x0, x1
    37b71d347lea rdx, [rsp+400]
    37b71d34flea rcx, [rsp+392]
    37b71d357movss [rsp+400], x0
    37b71d360call sscal_
    37b71d365mov r8, [rsp+160]
    37b71d36dlea rdx, [rsp+400]
    37b71d375lea r9, [rip+0xd18f30]
    37b71d37clea rcx, [rsp+392]
    37b71d384call sscal_
    37b71d389mov edi, [rsp+392]
    37b71d390mov edx, [r14]
    37b71d393cmp edi, edx
    37b71d395jge 0x37b71d3f6
    37b71d397sub edx, edi
    37b71d399movsxd rcx, edi
    37b71d39cmov rsi, [rsp+1264]
    37b71d3a4sub edx, 1
    37b71d3a7lea rax, [rcx+rbp]
    37b71d3abadd rdx, rcx
    37b71d3aemov rcx, [rsp+256]
    37b71d3b6lea rax, [rsi+4*rax+4]
    37b71d3bbadd rdx, rbp
    37b71d3belea rdx, [rsi+4*rdx+8]
    37b71d3c3nop [rax]
    37b71d3canop [rax+rax]
    37b71d3d5nop [rax+rax]
    37b71d3e0
    37b71d3e0loc_37b71d3e0:
    37b71d3e0mov [rax], 0
    37b71d3e6mov [rax+4*rcx], 0
    37b71d3edadd rax, 4
    37b71d3f1cmp rax, rdx
    37b71d3f4jne 0x37b71d3e0
    37b71d3f6
    37b71d3f6loc_37b71d3f6:
    37b71d3f6cmp [rsp+236], 1
    37b71d3fejle 0x37b71d754
    37b71d404mov esi, [rsp+168]
    37b71d40blea ebx, [rdi-1]
    37b71d40ejmp 0x37b71bb8d
    37b71d413
    37b71d413 0f 1f 44 00 00 ..D..
    37b71d418
    37b71d418loc_37b71d418:
    37b71d418mov rbx, [rsp+272]
    37b71d420lea rcx, [rsp+392]
    37b71d428imul rax, [rsp+256]
    37b71d431lea rbp, [rax+rbx]
    37b71d435mov rax, [rsp+1264]
    37b71d43dlea rbx, [rax+4*rbp+4]
    37b71d442mov eax, [r14]
    37b71d445mov r9, rbx
    37b71d448imul eax, esi
    37b71d44bcdqe
    37b71d44dlea rdx, [r15+4*rax]
    37b71d451lea rax, [rip+0xd18e54]
    37b71d458mov [rsp+32], rax
    37b71d45dmov r8, rax
    37b71d460call scopy_
    37b71d465mov rdx, rbx
    37b71d468lea r8, [rip+0xd18e3d]
    37b71d46flea rcx, [rsp+392]
    37b71d477call isamax_
    37b71d47cmov rcx, [rsp+1264]
    37b71d484mov r8, rbx
    37b71d487movaps x1, x9
    37b71d48bcdqe
    37b71d48dlea rdx, [rsp+400]
    37b71d495lea r9, [rip+0xd18e10]
    37b71d49cadd rax, rbp
    37b71d49fmovss x0, [rcx+4*rax]
    37b71d4a4lea rcx, [rsp+392]
    37b71d4acandps x0, x12
    37b71d4b0divss x1, x0
    37b71d4b4movss [rsp+400], x1
    37b71d4bdcall sscal_
    37b71d4c2mov ebx, [rsp+392]
    37b71d4c9mov eax, [r14]
    37b71d4cclea edx, [rbx+1]
    37b71d4cfcmp edx, eax
    37b71d4d1jg 0x37b71d170
    37b71d4d7sub eax, ebx
    37b71d4d9mov rsi, [rsp+1264]
    37b71d4e1movsxd rdx, edx
    37b71d4e4add rdx, rbp
    37b71d4e7sub eax, 1
    37b71d4ealea rcx, [rsi+4*rdx]
    37b71d4eelea r8, [4*rax+4]
    37b71d4f6xor edx, edx
    37b71d4f8call memset
    37b71d4fdcmp [rsp+236], 1
    37b71d505jle 0x37b71d75d
    37b71d50bmov esi, [rsp+388]
    37b71d512jmp 0x37b71c0ce
    37b71d517
    37b71d517loc_37b71d517:
    37b71d517mov edx, [r14]
    37b71d51amov edi, [rsp+160]
    37b71d521lea r10, [rsp+428]
    37b71d529lea rbp, [rsp+400]
    37b71d531mov rsi, [rsp+376]
    37b71d539mov r12, [rsp+216]
    37b71d541mov rcx, r10
    37b71d544mov [rsp+152], r10
    37b71d54cmov eax, edx
    37b71d54emov rbx, [rsp+1248]
    37b71d556sub eax, edi
    37b71d558add eax, 1
    37b71d55bmov [rsp+428], eax
    37b71d562movsxd rax, [rsp+328]
    37b71d56aimul rax, [rsp+272]
    37b71d573add rsi, rax
    37b71d576lea rax, [rsi+r12]
    37b71d57alea rbx, [rbx+4*rax]
    37b71d57emov eax, [rsp+200]
    37b71d585mov r9, rbx
    37b71d588imul eax, edx
    37b71d58blea eax, [rdi+rax-1]
    37b71d58fcdqe
    37b71d591lea rdx, [r15+4*rax]
    37b71d595lea rax, [rip+0xd18d10]
    37b71d59cmov [rsp+32], rax
    37b71d5a1mov r8, rax
    37b71d5a4call scopy_
    37b71d5a9mov eax, [r14]
    37b71d5acmov rcx, [rsp+152]
    37b71d5b4mov rdx, rbx
    37b71d5b7lea r8, [rip+0xd18cee]
    37b71d5besub eax, edi
    37b71d5c0add eax, 1
    37b71d5c3mov [rsp+428], eax
    37b71d5cacall isamax_
    37b71d5cfmovaps x1, x9
    37b71d5d3mov r8, rbx
    37b71d5d6mov rdx, rbp
    37b71d5d9lea eax, [rdi+rax-1]
    37b71d5ddmov rcx, [rsp+1248]
    37b71d5e5lea r9, [rip+0xd18cc0]
    37b71d5eccdqe
    37b71d5eeadd rax, rsi
    37b71d5f1movss x0, [rcx+4*rax]
    37b71d5f6mov eax, [r14]
    37b71d5f9mov rcx, [rsp+152]
    37b71d601andps x0, x10
    37b71d605sub eax, edi
    37b71d607divss x1, x0
    37b71d60badd eax, 1
    37b71d60emov [rsp+428], eax
    37b71d615movss [rsp+400], x1
    37b71d61ecall sscal_
    37b71d623cmp r12, 1
    37b71d627je 0x37b71d77e
    37b71d62dmov rdi, [rsp+1248]
    37b71d635lea eax, [r12-1]
    37b71d63axor edx, edx
    37b71d63clea r8, [4*rax]
    37b71d644lea rcx, [rdi+4*rsi+4]
    37b71d649call memset
    37b71d64ecmp [rsp+236], 1
    37b71d656jle 0x37b71d78c
    37b71d65cmov eax, [rsp+388]
    37b71d663xor edi, edi
    37b71d665mov [rsp+200], eax
    37b71d66cjmp 0x37b71af57
    37b71d671
    37b71d671loc_37b71d671:
    37b71d671movsxd rdx, edx
    37b71d674mov r8, rbp
    37b71d677lea r9, [rip+0xd18c2e]
    37b71d67emov rcx, r14
    37b71d681lea rdx, [r15+4*rdx-4]
    37b71d686call sscal_
    37b71d68bmov eax, [rsp+168]
    37b71d692imul eax, [r14]
    37b71d696mov r8, rsi
    37b71d699lea r9, [rip+0xd18c0c]
    37b71d6a0mov rcx, r14
    37b71d6a3add eax, edi
    37b71d6a5cdqe
    37b71d6a7lea rdx, [r15+4*rax-4]
    37b71d6accall sscal_
    37b71d6b1jmp 0x37b71c23f
    37b71d6b6
    37b71d6b6loc_37b71d6b6:
    37b71d6b6mov r8, rdi
    37b71d6b9mov rcx, r14
    37b71d6bclea r9, [rip+0xd18be9]
    37b71d6c3call sscal_
    37b71d6c8mov eax, [rsp+200]
    37b71d6cfimul eax, [r14]
    37b71d6d3mov r8, rsi
    37b71d6d6mov ecx, [rsp+232]
    37b71d6ddlea r9, [rip+0xd18bc8]
    37b71d6e4lea eax, [rcx+rax-1]
    37b71d6e8mov rcx, r14
    37b71d6ebcdqe
    37b71d6edlea rdx, [r15+4*rax]
    37b71d6f1call sscal_
    37b71d6f6jmp 0x37b71c89b
    37b71d6fb
    37b71d6fbloc_37b71d6fb:
    37b71d6fbtest ebp, ebp
    37b71d6fdje 0x37b719b23
    37b71d703mov edi, [rsp+236]
    37b71d70amov eax, 2
    37b71d70fmov [rsp+392], r13l
    37b71d717cmp edi, eax
    37b71d719cmovge eax, edi
    37b71d71cmov [rsp+388], eax
    37b71d723mov rax, [rsp+1288]
    37b71d72bmov eax, [rax]
    37b71d72dmov [rsp+232], eax
    37b71d734cmp r13l, 1
    37b71d738je 0x37b7192fa
    37b71d73ejmp 0x37b719b23
    37b71d743
    37b71d743loc_37b71d743:
    37b71d743mov edx, 1
    37b71d748lea eax, [r13-1]
    37b71d74cmovsxd r10, r10l
    37b71d74fjmp 0x37b719545
    37b71d754
    37b71d754loc_37b71d754:
    37b71d754lea r13l, [rdi-1]
    37b71d758jmp 0x37b71c35a
    37b71d75d
    37b71d75dloc_37b71d75d:
    37b71d75dlea r13l, [rbx-1]
    37b71d761mov [rsp+392], r13l
    37b71d769test r13l, r13l
    37b71d76cjle 0x37b719b23
    37b71d772mov [rsp+232], edi
    37b71d779jmp 0x37b7193c0
    37b71d77e
    37b71d77eloc_37b71d77e:
    37b71d77ecmp [rsp+236], 1
    37b71d786jg 0x37b71c3bb
    37b71d78c
    37b71d78cloc_37b71d78c:
    37b71d78cmov eax, [rsp+144]
    37b71d793mov rsi, [rsp+248]
    37b71d79bxor edi, edi
    37b71d79dmov [rsp+328], eax
    37b71d7a4mov rax, [rsp+272]
    37b71d7aclea rbx, [rax+rsi]
    37b71d7b0jmp 0x37b719cb0
    37b71d7b5
    37b71d7b5loc_37b71d7b5:
    37b71d7b5pxor x1, x1
    37b71d7b9jmp 0x37b71c315
    37b71d7be
    37b71d7beloc_37b71d7be:
    37b71d7bepxor x1, x1
    37b71d7c2cdqe
    37b71d7c4add ebx, 1
    37b71d7c7jmp 0x37b71c6d0
    37b71d7cc
    37b71d7ccloc_37b71d7cc:
    37b71d7ccpxor x1, x1
    37b71d7d0jmp 0x37b71c958
    37b71d7d5
    37b71d7d5loc_37b71d7d5:
    37b71d7d5pxor x1, x1
    37b71d7d9cdqe
    37b71d7dbmovsxd r10, r8l
    37b71d7dejmp 0x37b71c69f
    37b71d7e3
    37b71d7e3loc_37b71d7e3:
    37b71d7e3mov eax, 1
    37b71d7e8pxor x1, x1
    37b71d7ecjmp 0x37b71c91d
    37b71d7f1
    37b71d7f1loc_37b71d7f1:
    37b71d7f1mov eax, 1
    37b71d7f6pxor x1, x1
    37b71d7fajmp 0x37b71c2e0
    37b71d7ff
    37b71d7ffloc_37b71d7ff:
    37b71d7ffmov eax, 1
    37b71d804pxor x1, x1
    37b71d808jmp 0x37b71d300
    37b71d80d
    37b71d80dloc_37b71d80d:
    37b71d80dpxor x1, x1
    37b71d811jmp 0x37b71d335
    37b71d816
    37b71d816loc_37b71d816:
    37b71d816mov rax, [rsp+1288]
    37b71d81emov eax, [rax]
    37b71d820jmp 0x37b719142
    37b71d825
    37b71d825loc_37b71d825:
    37b71d825mov rax, [rsp+1312]
    37b71d82dmov [rax], -10
    37b71d833mov eax, 10
    37b71d838jmp 0x37b718ebc
    37b71d83d
    37b71d83dloc_37b71d83d:
    37b71d83dtest esi, esi
    37b71d83fjne 0x37b71d892
    37b71d841mov edi, [rax+4*rcx]
    37b71d844test edi, edi
    37b71d846jne 0x37b71d892
    37b71d848
    37b71d848loc_37b71d848:
    37b71d848add rcx, 1
    37b71d84ccmp edx, ecx
    37b71d84ejl 0x37b71d816
    37b71d850mov [rax+4*rcx-4], 0
    37b71d858add r9, r11
    37b71d85bjmp 0x37b719108
    37b71d860
    37b71d860loc_37b71d860:
    37b71d860pxor x1, x1
    37b71d864jmp 0x37b71d059
    37b71d869
    37b71d869loc_37b71d869:
    37b71d869pxor x1, x1
    37b71d86dcdqe
    37b71d86fmovsxd r10, r8l
    37b71d872jmp 0x37b71cdff
    37b71d877
    37b71d877loc_37b71d877:
    37b71d877pxor x1, x1
    37b71d87bcdqe
    37b71d87djmp 0x37b71ce30
    37b71d882
    37b71d882loc_37b71d882:
    37b71d882mov eax, [rsp+160]
    37b71d889pxor x1, x1
    37b71d88djmp 0x37b71d024
    37b71d892
    37b71d892loc_37b71d892:
    37b71d892mov rdi, [rsp+1288]
    37b71d89amov [rax+4*rcx-4], 1
    37b71d8a2add [rdi], 2
    37b71d8a5jmp 0x37b71d848
    37b71d8a7
    37b71d8a7loc_37b71d8a7:
    37b71d8a7xor eax, eax
    37b71d8a9jmp 0x37b719142
    37b71d8ae
    37b71d8aeloc_37b71d8ae:
    37b71d8aeneg eax
    37b71d8b0jmp 0x37b718ebc
    37b718c80