Files
stgsharp/StgSharp.Native/x64/Debug/FactorialSIMD.cod
T
2025-08-25 10:11:20 +08:00

390 lines
11 KiB
Plaintext

; Listing generated by Microsoft (R) Optimizing Compiler Version 19.44.35215.0
include listing.inc
INCLUDELIB MSVCRTD
INCLUDELIB OLDNAMES
PUBLIC facIncrement
PUBLIC facBase
_DATA SEGMENT
COMM infolog:BYTE:0200H
_DATA ENDS
msvcjmc SEGMENT
__5C8CA4D3_mmintrin@h DB 01H
__8C83BECC_concurrencysal@h DB 01H
__608E3340_sal@h DB 01H
__FB67E29A_vadefs@h DB 01H
__1D1B619C_vcruntime@h DB 01H
__1B0272D9_corecrt@h DB 01H
__45DED006_corecrt_malloc@h DB 01H
__FA77ED4C_malloc@h DB 01H
__9DCE0DC8_xmmintrin@h DB 01H
__0ED14FEE_winpackagefamily@h DB 01H
__F92EF742_winapifamily@h DB 01H
__22E515EB_sdkddkver@h DB 01H
__9BAEE128_excpt@h DB 01H
__B7B53233_stdarg@h DB 01H
__29F49F40_specstrings_undef@h DB 01H
__D0F42932_specstrings_strict@h DB 01H
__F00405FD_sdv_driverspecs@h DB 01H
__173E09CA_driverspecs@h DB 01H
__9A452361_specstrings@h DB 01H
__74484FB3_corecrt_wctype@h DB 01H
__1EC467ED_ctype@h DB 01H
__98517906_kernelspecs@h DB 01H
__D3497022_basetsd@h DB 01H
__61CC25A1_errno@h DB 01H
__8C6AE390_vcruntime_string@h DB 01H
__E4C945B4_corecrt_memcpy_s@h DB 01H
__8EAFF484_corecrt_memory@h DB 01H
__6490AEC5_corecrt_wstring@h DB 01H
__E8295E1E_string@h DB 01H
__7FCB5136_guiddef@h DB 01H
__FF8A9E11_pshpack4@h DB 01H
__D5438902_poppack@h DB 01H
__FB07E2A3_pshpack2@h DB 01H
__F6906775_pshpack8@h DB 01H
__F9415CFA_pshpack1@h DB 01H
__E58C6F1C_apiset@h DB 01H
__A5C52262_ktmtypes@h DB 01H
__FBBA8F9D_winnt@h DB 01H
__A32D3D2C_minwindef@h DB 01H
__7D798299_windef@h DB 01H
__9AE108E9_apisetcconv@h DB 01H
__9CE53D2C_minwinbase@h DB 01H
__5D0AA210_apiquery2@h DB 01H
__4DD4E726_processenv@h DB 01H
__3575CC96_fileapi@h DB 01H
__8EA220E6_fileapifromapp@h DB 01H
__BF9B31FE_debugapi@h DB 01H
__C3AFE134_utilapiset@h DB 01H
__76F69E51_handleapi@h DB 01H
__7E63D784_errhandlingapi@h DB 01H
__FFA7003C_fibersapi@h DB 01H
__8971CECF_namedpipeapi@h DB 01H
__46C9D8A6_profileapi@h DB 01H
__98D4F0A3_heapapi@h DB 01H
__50EF2A88_ioapiset@h DB 01H
__92EA1464_synchapi@h DB 01H
__A53D8AED_interlockedapi@h DB 01H
__B506CAAF_processthreadsapi@h DB 01H
__FF735694_sysinfoapi@h DB 01H
__29589789_memoryapi@h DB 01H
__E9C7A663_enclaveapi@h DB 01H
__D035F9F7_threadpoollegacyapiset@h DB 01H
__C899130C_threadpoolapiset@h DB 01H
__E0441EE7_jobapi@h DB 01H
__CEF86838_jobapi2@h DB 01H
__35AC6B9A_wow64apiset@h DB 01H
__C8AE41A5_libloaderapi@h DB 01H
__CC9E232D_securitybaseapi@h DB 01H
__C9D0C9B8_namespaceapi@h DB 01H
__30FF9CAD_systemtopologyapi@h DB 01H
__4D358795_processtopologyapi@h DB 01H
__40446D10_securityappcontainer@h DB 01H
__3A952EDF_realtimeapiset@h DB 01H
__F4865F4B_winerror@h DB 01H
__53C1BD93_timezoneapi@h DB 01H
__1FFBEDBF_winbase@h DB 01H
__43D3FA34_wingdi@h DB 01H
__D24A7391_tvout@h DB 01H
__7713A6C8_winuser@h DB 01H
__0B345F71_datetimeapi@h DB 01H
__9AC6960C_winnls@h DB 01H
__973A40A8_stringapiset@h DB 01H
__CC46FEE2_wincontypes@h DB 01H
__97D70BE1_consoleapi@h DB 01H
__27EC5E18_consoleapi2@h DB 01H
__262E342F_consoleapi3@h DB 01H
__641DCD70_wincon@h DB 01H
__5A96C7A7_verrsrc@h DB 01H
__B69C6FF2_winver@h DB 01H
__7AEB8C37_reason@h DB 01H
__59F1A8A9_winreg@h DB 01H
__EE0E2B7F_wnnc@h DB 01H
__DCDB7054_winnetwk@h DB 01H
__81DDEDE3_stralign@h DB 01H
__2134D1B4_winsvc@h DB 01H
__11C07376_mcx@h DB 01H
__577CE3E7_ime_cmodes@h DB 01H
__1E72C31E_imm@h DB 01H
__E15BF856_windows@h DB 01H
__639A050C_sn_internal@h DB 01H
__44420885_corecrt_stdio_config@h DB 01H
__AB601D8D_corecrt_wstdio@h DB 01H
__C1EC35D3_stdio@h DB 01H
__13F5961F_stddef@h DB 01H
__FD6DBD68_corecrt_search@h DB 01H
__E2E33A61_corecrt_wstdlib@h DB 01H
__3D1CA2C3_limits@h DB 01H
__6E5ACABA_stdlib@h DB 01H
__C363A436_stb_image@h DB 01H
__836A8CFC_stdint@h DB 01H
__C95DC661_khrplatform@h DB 01H
__6C577136_gl@h DB 01H
__EE638A2C_wgl@h DB 01H
__E76F0EB7_glfw3@h DB 01H
__F8AC8404_emmintrin@h DB 01H
__270DEC8D_pmmintrin@h DB 01H
__97D49F0F_tmmintrin@h DB 01H
__BEEF8A8C_smmintrin@h DB 01H
__DB8D0340_nmmintrin@h DB 01H
__0E36F90E_wmmintrin@h DB 01H
__C5A2B409_zmmintrin@h DB 01H
__F2B616C3_immintrin@h DB 01H
__19139D6C_assert@h DB 01H
__4E664295_glfw_function@h DB 01H
__64C0EF11_StgSharpNative@h DB 01H
__47E843D2_intrin0@inl@h DB 01H
__130964FF_setjmp@h DB 01H
__4875F786_ammintrin@h DB 01H
__353FD888_intrin@h DB 01H
__5FE1E9E3_isa_availability@h DB 01H
__AA840D82_sn_intrinsic@h DB 01H
__FEE4A61F_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$rtcVarDesc
CONST ENDS
; Function compile flags: /Odt
; COMDAT __JustMyCode_Default
_TEXT SEGMENT
__JustMyCode_Default PROC ; COMDAT
00000 c2 00 00 ret 0
__JustMyCode_Default ENDP
_TEXT ENDS
; Function compile flags: /Odtp /RTCsu /ZI
; File D:\platform-dev\StgSharp.Native\sn-src\FactorialSIMD.c
; COMDAT factorial_simd_sse
_TEXT SEGMENT
baseResult$ = 16
facBaseLocal$ = 64
cycle$ = 100
rest$ = 132
i$4 = 168
result$ = 200
i$5 = 232
n$ = 480
factorial_simd_sse PROC ; COMDAT
; 8 : {
$LN9:
00000 89 4c 24 08 mov DWORD PTR [rsp+8], ecx
00004 55 push rbp
00005 57 push rdi
00006 48 81 ec e8 01
00 00 sub rsp, 488 ; 000001e8H
0000d 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
00012 48 8d 7c 24 20 lea rdi, QWORD PTR [rsp+32]
00017 b9 42 00 00 00 mov ecx, 66 ; 00000042H
0001c b8 cc cc cc cc mov eax, -858993460 ; ccccccccH
00021 f3 ab rep stosd
00023 8b 8c 24 08 02
00 00 mov ecx, DWORD PTR [rsp+520]
0002a 48 8d 0d 00 00
00 00 lea rcx, OFFSET FLAT:__FEE4A61F_FactorialSIMD@c
00031 e8 00 00 00 00 call __CheckForDebuggerJustMyCode
00036 90 npad 1
; 9 : __m128i baseResult = facBase;
00037 66 0f 6f 05 00
00 00 00 movdqa xmm0, XMMWORD PTR facBase
0003f 66 0f 7f 45 10 movdqa XMMWORD PTR baseResult$[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