390 lines
11 KiB
Plaintext
390 lines
11 KiB
Plaintext
; Listing generated by Microsoft (R) Optimizing Compiler Version 19.44.35213.0
|
|
|
|
include listing.inc
|
|
|
|
INCLUDELIB MSVCRTD
|
|
INCLUDELIB OLDNAMES
|
|
|
|
PUBLIC facIncrement
|
|
PUBLIC facBase
|
|
_DATA SEGMENT
|
|
COMM infolog:BYTE:0200H
|
|
_DATA ENDS
|
|
msvcjmc SEGMENT
|
|
__1F8170E1_mmintrin@h DB 01H
|
|
__39F22AB7_concurrencysal@h DB 01H
|
|
__B3864887_sal@h DB 01H
|
|
__AD974A83_vadefs@h DB 01H
|
|
__D58F3DC8_vcruntime@h DB 01H
|
|
__735960E1_corecrt@h DB 01H
|
|
__211DB995_corecrt_malloc@h DB 01H
|
|
__7CD62D9E_malloc@h DB 01H
|
|
__555A519C_xmmintrin@h DB 01H
|
|
__4C2AD0C7_winpackagefamily@h DB 01H
|
|
__9DED9ED1_winapifamily@h DB 01H
|
|
__12FEDF4E_sdkddkver@h DB 01H
|
|
__D8C2DB1D_excpt@h DB 01H
|
|
__E1459A2A_stdarg@h DB 01H
|
|
__6B04FCB3_specstrings_undef@h DB 01H
|
|
__F4027AF7_specstrings_strict@h DB 01H
|
|
__B90340D4_sdv_driverspecs@h DB 01H
|
|
__BFFF5AD3_driverspecs@h DB 01H
|
|
__32847078_specstrings@h DB 01H
|
|
__108B2620_corecrt_wctype@h DB 01H
|
|
__6CD6613D_ctype@h DB 01H
|
|
__30902A1F_kernelspecs@h DB 01H
|
|
__376FD5CA_basetsd@h DB 01H
|
|
__13DE2371_errno@h DB 01H
|
|
__B7C3A181_vcruntime_string@h DB 01H
|
|
__8ACB2152_corecrt_memcpy_s@h DB 01H
|
|
__EA6C9D17_corecrt_memory@h DB 01H
|
|
__0DF2AF52_corecrt_wstring@h DB 01H
|
|
__6E889ECC_string@h DB 01H
|
|
__9BEDF4DE_guiddef@h DB 01H
|
|
__51BFD2FE_pshpack4@h DB 01H
|
|
__31652CEA_poppack@h DB 01H
|
|
__5532AE4C_pshpack2@h DB 01H
|
|
__58A52B9A_pshpack8@h DB 01H
|
|
__57741015_pshpack1@h DB 01H
|
|
__CDE68C90_apiset@h DB 01H
|
|
__0BF06E8D_ktmtypes@h DB 01H
|
|
__40C200C0_winnt@h DB 01H
|
|
__9336F789_minwindef@h DB 01H
|
|
__55136115_windef@h DB 01H
|
|
__32205BF0_apisetcconv@h DB 01H
|
|
__B48FDEA0_minwinbase@h DB 01H
|
|
__3551B028_apiquery2@h DB 01H
|
|
__65BE04AA_processenv@h DB 01H
|
|
__4767CA46_fileapi@h DB 01H
|
|
__282E6C4B_fileapifromapp@h DB 01H
|
|
__393AF12C_debugapi@h DB 01H
|
|
__EBC502B8_utilapiset@h DB 01H
|
|
__1EAD8C69_handleapi@h DB 01H
|
|
__D8EF9B29_errhandlingapi@h DB 01H
|
|
__97FC1204_fibersapi@h DB 01H
|
|
__27448220_namedpipeapi@h DB 01H
|
|
__6EA33B2A_profileapi@h DB 01H
|
|
__EAC6F673_heapapi@h DB 01H
|
|
__D64EEA5A_ioapiset@h DB 01H
|
|
__144BD4B6_synchapi@h DB 01H
|
|
__03B1C640_interlockedapi@h DB 01H
|
|
__DC64CB38_processthreadsapi@h DB 01H
|
|
__D719B518_sysinfoapi@h DB 01H
|
|
__410385B1_memoryapi@h DB 01H
|
|
__C1AD45EF_enclaveapi@h DB 01H
|
|
__F4C3AA32_threadpoollegacyapiset@h DB 01H
|
|
__AC5A7A9F_threadpoolapiset@h DB 01H
|
|
__F5254B21_jobapi@h DB 01H
|
|
__BCEA6EE8_jobapi2@h DB 01H
|
|
__D18ACE72_wow64apiset@h DB 01H
|
|
__669B0D4A_libloaderapi@h DB 01H
|
|
__645F7034_securitybaseapi@h DB 01H
|
|
__67E58557_namespaceapi@h DB 01H
|
|
__599D9D3A_systemtopologyapi@h DB 01H
|
|
__2337E373_processtopologyapi@h DB 01H
|
|
__02BFF239_securityappcontainer@h DB 01H
|
|
__9C196272_realtimeapiset@h DB 01H
|
|
__5AB313A4_winerror@h DB 01H
|
|
__B7E7187B_timezoneapi@h DB 01H
|
|
__6DE9EB6F_winbase@h DB 01H
|
|
__56B2AFF2_wingdi@h DB 01H
|
|
__BA1161A9_tvout@h DB 01H
|
|
__0501A018_winuser@h DB 01H
|
|
__EF12FA99_datetimeapi@h DB 01H
|
|
__8FA7C3CA_winnls@h DB 01H
|
|
__390F0C47_stringapiset@h DB 01H
|
|
__28605B0A_wincontypes@h DB 01H
|
|
__BFBDE86D_consoleapi@h DB 01H
|
|
__C3CAFBF0_consoleapi2@h DB 01H
|
|
__C20891C7_consoleapi3@h DB 01H
|
|
__717C98B6_wincon@h DB 01H
|
|
__2884C177_verrsrc@h DB 01H
|
|
__A3FD3A34_winver@h DB 01H
|
|
__6F8AD9F1_reason@h DB 01H
|
|
__4C90FD6F_winreg@h DB 01H
|
|
__68AFEBAD_wnnc@h DB 01H
|
|
__5A7AB086_winnetwk@h DB 01H
|
|
__2FE8A10C_stralign@h DB 01H
|
|
__34558472_winsvc@h DB 01H
|
|
__058845F1_mcx@h DB 01H
|
|
__7F16006B_ime_cmodes@h DB 01H
|
|
__0A3AF599_imm@h DB 01H
|
|
__9349FE86_windows@h DB 01H
|
|
__AD5C15D8_sn_internal@h DB 01H
|
|
__60B45B40_corecrt_stdio_config@h DB 01H
|
|
__CFA3741E_corecrt_wstdio@h DB 01H
|
|
__B3FE3303_stdio@h DB 01H
|
|
__955456CD_stddef@h DB 01H
|
|
__99AED4FB_corecrt_search@h DB 01H
|
|
__8B813BF6_corecrt_wstdlib@h DB 01H
|
|
__6BEC0ADA_limits@h DB 01H
|
|
__E8FB0A68_stdlib@h DB 01H
|
|
__500EB2D2_stb_image@h DB 01H
|
|
__D59A24E5_stdint@h DB 01H
|
|
__079BD6B5_khrplatform@h DB 01H
|
|
__67B0FD84_gl@h DB 01H
|
|
__CB07BF00_wgl@h DB 01H
|
|
__EC888205_glfw3@h DB 01H
|
|
__3038D850_emmintrin@h DB 01H
|
|
__EF99B0D9_pmmintrin@h DB 01H
|
|
__5F40C35B_tmmintrin@h DB 01H
|
|
__767BD6D8_smmintrin@h DB 01H
|
|
__13195F14_nmmintrin@h DB 01H
|
|
__C6A2A55A_wmmintrin@h DB 01H
|
|
__0D36E85D_zmmintrin@h DB 01H
|
|
__3A224A97_immintrin@h DB 01H
|
|
__9FB25DBE_assert@h DB 01H
|
|
__B6859C2C_glfw_function@h DB 01H
|
|
__D68507E7_StgSharpNative@h DB 01H
|
|
__FBE20EC9_intrin0@inl@h DB 01H
|
|
__45F9CCE6_setjmp@h DB 01H
|
|
__80E1ABD2_ammintrin@h DB 01H
|
|
__63CF7091_intrin@h DB 01H
|
|
__6448ABF2_isa_availability@h DB 01H
|
|
__2BF4DD5F_sn_intrinsic@h DB 01H
|
|
__7F9476C2_FactorialSIMD@c DB 01H
|
|
msvcjmc ENDS
|
|
_DATA SEGMENT
|
|
facIncrement DB 04H
|
|
DB 04H
|
|
DB 04H
|
|
DB 04H
|
|
ORG $+12
|
|
facBase DB 01H
|
|
DB 02H
|
|
DB 03H
|
|
DB 04H
|
|
ORG $+12
|
|
_DATA ENDS
|
|
PUBLIC factorial_simd_sse
|
|
PUBLIC __JustMyCode_Default
|
|
EXTRN _RTC_CheckStackVars:PROC
|
|
EXTRN _RTC_InitBase:PROC
|
|
EXTRN _RTC_Shutdown:PROC
|
|
EXTRN __CheckForDebuggerJustMyCode:PROC
|
|
; COMDAT pdata
|
|
pdata SEGMENT
|
|
$pdata$factorial_simd_sse DD imagerel $LN9
|
|
DD imagerel $LN9+377
|
|
DD imagerel $unwind$factorial_simd_sse
|
|
pdata ENDS
|
|
; COMDAT rtc$TMZ
|
|
rtc$TMZ SEGMENT
|
|
_RTC_Shutdown.rtc$TMZ DQ FLAT:_RTC_Shutdown
|
|
rtc$TMZ ENDS
|
|
; COMDAT rtc$IMZ
|
|
rtc$IMZ SEGMENT
|
|
_RTC_InitBase.rtc$IMZ DQ FLAT:_RTC_InitBase
|
|
rtc$IMZ ENDS
|
|
; COMDAT xdata
|
|
xdata SEGMENT
|
|
$unwind$factorial_simd_sse DD 025052a01H
|
|
DD 010d2312H
|
|
DD 07006003dH
|
|
DD 05005H
|
|
xdata ENDS
|
|
; COMDAT CONST
|
|
CONST SEGMENT
|
|
factorial_simd_sse$rtcName$0 DB 062H
|
|
DB 061H
|
|
DB 073H
|
|
DB 065H
|
|
DB 052H
|
|
DB 065H
|
|
DB 073H
|
|
DB 075H
|
|
DB 06cH
|
|
DB 074H
|
|
DB 00H
|
|
ORG $+5
|
|
factorial_simd_sse$rtcVarDesc DD 030H
|
|
DD 010H
|
|
DQ FLAT:factorial_simd_sse$rtcName$0
|
|
ORG $+48
|
|
factorial_simd_sse$rtcFrameData DD 01H
|
|
DD 00H
|
|
DQ FLAT:factorial_simd_sse$rtcVarDesc
|
|
CONST ENDS
|
|
; Function compile flags: /Odt
|
|
; COMDAT __JustMyCode_Default
|
|
_TEXT SEGMENT
|
|
__JustMyCode_Default PROC ; COMDAT
|
|
00000 c2 00 00 ret 0
|
|
__JustMyCode_Default ENDP
|
|
_TEXT ENDS
|
|
; Function compile flags: /Odtp /RTCsu /ZI
|
|
; File D:\NitloadIndustry\Ñо¿ÏîÄ¿\1-Flagship\TH2020\ProjectFile\ÄÚºËÔ´´úÂë\StgSharp.Native\sn-src\FactorialSIMD.c
|
|
; COMDAT factorial_simd_sse
|
|
_TEXT SEGMENT
|
|
baseResult$ = 16
|
|
facBaseLocal$ = 64
|
|
cycle$ = 100
|
|
rest$ = 132
|
|
i$4 = 168
|
|
result$ = 200
|
|
i$5 = 232
|
|
n$ = 480
|
|
factorial_simd_sse PROC ; COMDAT
|
|
|
|
; 8 : {
|
|
|
|
$LN9:
|
|
00000 89 4c 24 08 mov DWORD PTR [rsp+8], ecx
|
|
00004 55 push rbp
|
|
00005 57 push rdi
|
|
00006 48 81 ec e8 01
|
|
00 00 sub rsp, 488 ; 000001e8H
|
|
0000d 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
|
|
00012 48 8d 7c 24 20 lea rdi, QWORD PTR [rsp+32]
|
|
00017 b9 42 00 00 00 mov ecx, 66 ; 00000042H
|
|
0001c b8 cc cc cc cc mov eax, -858993460 ; ccccccccH
|
|
00021 f3 ab rep stosd
|
|
00023 8b 8c 24 08 02
|
|
00 00 mov ecx, DWORD PTR [rsp+520]
|
|
0002a 48 8d 0d 00 00
|
|
00 00 lea rcx, OFFSET FLAT:__7F9476C2_FactorialSIMD@c
|
|
00031 e8 00 00 00 00 call __CheckForDebuggerJustMyCode
|
|
00036 90 npad 1
|
|
|
|
; 9 : __m128i baseResult = facBase;
|
|
|
|
00037 66 0f 6f 05 00
|
|
00 00 00 movdqa xmm0, XMMWORD PTR facBase
|
|
0003f 66 0f 7f 45 10 movdqa XMMWORD PTR baseResult$[rbp], xmm0
|
|
|
|
; 10 : __m128i facBaseLocal = facBase;
|
|
|
|
00044 66 0f 6f 05 00
|
|
00 00 00 movdqa xmm0, XMMWORD PTR facBase
|
|
0004c 66 0f 7f 45 40 movdqa XMMWORD PTR facBaseLocal$[rbp], xmm0
|
|
|
|
; 11 : const int cycle = (n - 3) >> 2;
|
|
|
|
00051 8b 85 e0 01 00
|
|
00 mov eax, DWORD PTR n$[rbp]
|
|
00057 83 e8 03 sub eax, 3
|
|
0005a c1 f8 02 sar eax, 2
|
|
0005d 89 45 64 mov DWORD PTR cycle$[rbp], eax
|
|
|
|
; 12 : const uint32_t rest = (cycle << 2) + 1;
|
|
|
|
00060 8b 45 64 mov eax, DWORD PTR cycle$[rbp]
|
|
00063 8d 04 85 01 00
|
|
00 00 lea eax, DWORD PTR [rax*4+1]
|
|
0006a 89 85 84 00 00
|
|
00 mov DWORD PTR rest$[rbp], eax
|
|
|
|
; 13 : for (size_t i = 0; i < cycle; i++) {
|
|
|
|
00070 48 c7 85 a8 00
|
|
00 00 00 00 00
|
|
00 mov QWORD PTR i$4[rbp], 0
|
|
0007b eb 11 jmp SHORT $LN4@factorial_
|
|
$LN2@factorial_:
|
|
0007d 48 8b 85 a8 00
|
|
00 00 mov rax, QWORD PTR i$4[rbp]
|
|
00084 48 ff c0 inc rax
|
|
00087 48 89 85 a8 00
|
|
00 00 mov QWORD PTR i$4[rbp], rax
|
|
$LN4@factorial_:
|
|
0008e 48 63 45 64 movsxd rax, DWORD PTR cycle$[rbp]
|
|
00092 48 39 85 a8 00
|
|
00 00 cmp QWORD PTR i$4[rbp], rax
|
|
00099 73 24 jae SHORT $LN3@factorial_
|
|
|
|
; 14 : facBaseLocal = _mm_add_epi32(facBaseLocal, facIncrement);
|
|
|
|
0009b 66 0f 6f 45 40 movdqa xmm0, XMMWORD PTR facBaseLocal$[rbp]
|
|
000a0 66 0f fe 05 00
|
|
00 00 00 paddd xmm0, XMMWORD PTR facIncrement
|
|
000a8 66 0f 7f 45 40 movdqa XMMWORD PTR facBaseLocal$[rbp], xmm0
|
|
|
|
; 15 : baseResult = _mm_mullo_epi32(baseResult, facBaseLocal);
|
|
|
|
000ad 66 0f 6f 45 10 movdqa xmm0, XMMWORD PTR baseResult$[rbp]
|
|
000b2 66 0f 38 40 45
|
|
40 pmulld xmm0, XMMWORD PTR facBaseLocal$[rbp]
|
|
000b8 66 0f 7f 45 10 movdqa XMMWORD PTR baseResult$[rbp], xmm0
|
|
|
|
; 16 : }
|
|
|
|
000bd eb be jmp SHORT $LN2@factorial_
|
|
$LN3@factorial_:
|
|
|
|
; 17 : uint64_t result = (uint64_t)baseResult.m128i_i32[0] * (uint64_t)baseResult.m128i_i32[1] *
|
|
|
|
000bf b8 04 00 00 00 mov eax, 4
|
|
000c4 48 6b c0 00 imul rax, rax, 0
|
|
000c8 48 63 44 05 10 movsxd rax, DWORD PTR baseResult$[rbp+rax]
|
|
000cd b9 04 00 00 00 mov ecx, 4
|
|
000d2 48 6b c9 01 imul rcx, rcx, 1
|
|
000d6 48 63 4c 0d 10 movsxd rcx, DWORD PTR baseResult$[rbp+rcx]
|
|
000db 48 0f af c1 imul rax, rcx
|
|
000df b9 04 00 00 00 mov ecx, 4
|
|
000e4 48 6b c9 02 imul rcx, rcx, 2
|
|
000e8 48 63 4c 0d 10 movsxd rcx, DWORD PTR baseResult$[rbp+rcx]
|
|
000ed 48 0f af c1 imul rax, rcx
|
|
000f1 b9 04 00 00 00 mov ecx, 4
|
|
000f6 48 6b c9 03 imul rcx, rcx, 3
|
|
000fa 48 63 4c 0d 10 movsxd rcx, DWORD PTR baseResult$[rbp+rcx]
|
|
000ff 48 0f af c1 imul rax, rcx
|
|
00103 48 89 85 c8 00
|
|
00 00 mov QWORD PTR result$[rbp], rax
|
|
|
|
; 18 : (uint64_t)baseResult.m128i_i32[2] * (uint64_t)baseResult.m128i_i32[3];
|
|
; 19 : for (uint64_t i = rest; i < n; i++) {
|
|
|
|
0010a 8b 85 84 00 00
|
|
00 mov eax, DWORD PTR rest$[rbp]
|
|
00110 48 89 85 e8 00
|
|
00 00 mov QWORD PTR i$5[rbp], rax
|
|
00117 eb 11 jmp SHORT $LN7@factorial_
|
|
$LN5@factorial_:
|
|
00119 48 8b 85 e8 00
|
|
00 00 mov rax, QWORD PTR i$5[rbp]
|
|
00120 48 ff c0 inc rax
|
|
00123 48 89 85 e8 00
|
|
00 00 mov QWORD PTR i$5[rbp], rax
|
|
$LN7@factorial_:
|
|
0012a 48 63 85 e0 01
|
|
00 00 movsxd rax, DWORD PTR n$[rbp]
|
|
00131 48 39 85 e8 00
|
|
00 00 cmp QWORD PTR i$5[rbp], rax
|
|
00138 73 18 jae SHORT $LN6@factorial_
|
|
|
|
; 20 : result *= i ;
|
|
|
|
0013a 48 8b 85 c8 00
|
|
00 00 mov rax, QWORD PTR result$[rbp]
|
|
00141 48 0f af 85 e8
|
|
00 00 00 imul rax, QWORD PTR i$5[rbp]
|
|
00149 48 89 85 c8 00
|
|
00 00 mov QWORD PTR result$[rbp], rax
|
|
|
|
; 21 : }
|
|
|
|
00150 eb c7 jmp SHORT $LN5@factorial_
|
|
$LN6@factorial_:
|
|
|
|
; 22 : return result;
|
|
|
|
00152 48 8b 85 c8 00
|
|
00 00 mov rax, QWORD PTR result$[rbp]
|
|
|
|
; 23 : }
|
|
|
|
00159 48 8b f8 mov rdi, rax
|
|
0015c 48 8d 4d e0 lea rcx, QWORD PTR [rbp-32]
|
|
00160 48 8d 15 00 00
|
|
00 00 lea rdx, OFFSET FLAT:factorial_simd_sse$rtcFrameData
|
|
00167 e8 00 00 00 00 call _RTC_CheckStackVars
|
|
0016c 48 8b c7 mov rax, rdi
|
|
0016f 48 8d a5 c8 01
|
|
00 00 lea rsp, QWORD PTR [rbp+456]
|
|
00176 5f pop rdi
|
|
00177 5d pop rbp
|
|
00178 c3 ret 0
|
|
factorial_simd_sse ENDP
|
|
_TEXT ENDS
|
|
END
|