youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
11b630ffa272126ee707ee395e42fbd15d285614
stgsharp/StgSharp.Native/x64/Debug/mat2.cod
T

841 lines
25 KiB
Plaintext
Raw Blame History

; Listing generated by Microsoft (R) Optimizing Compiler Version 19.44.35215.0
include listing.inc
INCLUDELIB MSVCRTD
INCLUDELIB OLDNAMES
_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
__735960E1_corecrt@h DB 01H
__211DB995_corecrt_malloc@h DB 01H
__7CD62D9E_malloc@h DB 01H
__9DCE0DC8_xmmintrin@h DB 01H
__4C2AD0C7_winpackagefamily@h DB 01H
__9DED9ED1_winapifamily@h DB 01H
__12FEDF4E_sdkddkver@h DB 01H
__9BAEE128_excpt@h DB 01H
__B7B53233_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
__8C6AE390_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
__639A050C_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
__3D1CA2C3_limits@h DB 01H
__E8FB0A68_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
__9FB25DBE_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
__B3FFA7AA_mat2@c DB 01H
msvcjmc ENDS
PUBLIC addmatrix2_sse
PUBLIC addmatrix2_avx
PUBLIC submatrix2_sse
PUBLIC transpose23
PUBLIC transpose24
PUBLIC submatrix2_avx
PUBLIC __JustMyCode_Default
EXTRN _RTC_InitBase:PROC
EXTRN _RTC_Shutdown:PROC
EXTRN __CheckForDebuggerJustMyCode:PROC
; COMDAT pdata
pdata SEGMENT
$pdata$addmatrix2_sse DD imagerel $LN3
DD imagerel $LN3+254
DD imagerel $unwind$addmatrix2_sse
pdata ENDS
; COMDAT pdata
pdata SEGMENT
$pdata$addmatrix2_avx DD imagerel $LN3
DD imagerel $LN3+169
DD imagerel $unwind$addmatrix2_avx
pdata ENDS
; COMDAT pdata
pdata SEGMENT
$pdata$submatrix2_sse DD imagerel $LN3
DD imagerel $LN3+254
DD imagerel $unwind$submatrix2_sse
pdata ENDS
; COMDAT pdata
pdata SEGMENT
$pdata$transpose23 DD imagerel $LN3
DD imagerel $LN3+271
DD imagerel $unwind$transpose23
pdata ENDS
; COMDAT pdata
pdata SEGMENT
$pdata$transpose24 DD imagerel $LN3
DD imagerel $LN3+305
DD imagerel $unwind$transpose24
pdata ENDS
; COMDAT pdata
pdata SEGMENT
$pdata$submatrix2_avx DD imagerel $LN3
DD imagerel $LN3+169
DD imagerel $unwind$submatrix2_avx
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
$unwind$submatrix2_avx DD 02d062801H
DD 011a231fH
DD 07013002aH
DD 0d0115012H
xdata ENDS
; COMDAT xdata
xdata SEGMENT
$unwind$transpose24 DD 025051801H
DD 01132318H
DD 0700c0029H
DD 0500bH
xdata ENDS
; COMDAT xdata
xdata SEGMENT
$unwind$transpose23 DD 025051801H
DD 01132318H
DD 0700c0029H
DD 0500bH
xdata ENDS
; COMDAT xdata
xdata SEGMENT
$unwind$submatrix2_sse DD 025051d01H
DD 0118231dH
DD 070110029H
DD 05010H
xdata ENDS
; COMDAT xdata
xdata SEGMENT
$unwind$addmatrix2_avx DD 02d062801H
DD 011a231fH
DD 07013002aH
DD 0d0115012H
xdata ENDS
; COMDAT xdata
xdata SEGMENT
$unwind$addmatrix2_sse DD 025051d01H
DD 0118231dH
DD 070110029H
DD 05010H
xdata 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\mat2.c
; COMDAT submatrix2_avx
_TEXT SEGMENT
temp$ = 32
left$ = 336
right$ = 344
ans$ = 352
submatrix2_avx PROC ; COMDAT
; 54 : {
$LN3:
00000 4c 89 44 24 18 mov QWORD PTR [rsp+24], r8
00005 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
0000a 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
0000f 41 55 push r13
00011 55 push rbp
00012 57 push rdi
00013 48 81 ec 50 01
00 00 sub rsp, 336 ; 00000150H
0001a 4c 8d 6c 24 20 lea r13, QWORD PTR [rsp+32]
0001f 48 8d 6c 24 40 lea rbp, QWORD PTR [rsp+64]
00024 48 83 e5 e0 and rbp, -32 ; ffffffffffffffe0H
00028 48 8d 0d 00 00
00 00 lea rcx, OFFSET FLAT:__B3FFA7AA_mat2@c
0002f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
00034 90 npad 1
; 55 : __m256 temp = _mm256_sub_ps(ALIGN256(left->column[0]), ALIGN256(right->column[0]));
00035 b8 10 00 00 00 mov eax, 16
0003a 48 6b c0 00 imul rax, rax, 0
0003e 49 8b 8d 58 01
00 00 mov rcx, QWORD PTR right$[r13]
00045 48 03 c8 add rcx, rax
00048 48 8b c1 mov rax, rcx
0004b c5 fc 10 00 vmovups ymm0, YMMWORD PTR [rax]
0004f b8 10 00 00 00 mov eax, 16
00054 48 6b c0 00 imul rax, rax, 0
00058 49 8b 8d 50 01
00 00 mov rcx, QWORD PTR left$[r13]
0005f 48 03 c8 add rcx, rax
00062 48 8b c1 mov rax, rcx
00065 c5 fc 10 08 vmovups ymm1, YMMWORD PTR [rax]
00069 c5 f4 5c c0 vsubps ymm0, ymm1, ymm0
0006d c5 fc 11 45 20 vmovups YMMWORD PTR temp$[rbp], ymm0
; 56 : _mm256_storeu_ps(&ans->m[0][0], temp);
00072 b8 08 00 00 00 mov eax, 8
00077 48 6b c0 00 imul rax, rax, 0
0007b 49 8b 8d 60 01
00 00 mov rcx, QWORD PTR ans$[r13]
00082 48 03 c8 add rcx, rax
00085 48 8b c1 mov rax, rcx
00088 b9 04 00 00 00 mov ecx, 4
0008d 48 6b c9 00 imul rcx, rcx, 0
00091 48 03 c1 add rax, rcx
00094 c5 fc 10 45 20 vmovups ymm0, YMMWORD PTR temp$[rbp]
00099 c5 fc 11 00 vmovups YMMWORD PTR [rax], ymm0
; 57 : }
0009d 49 8d a5 30 01
00 00 lea rsp, QWORD PTR [r13+304]
000a4 5f pop rdi
000a5 5d pop rbp
000a6 41 5d pop r13
000a8 c3 ret 0
submatrix2_avx ENDP
_TEXT ENDS
; Function compile flags: /Odtp /RTCsu /ZI
; File D:\platform-dev\StgSharp.Native\sn-src\mat2.c
; COMDAT transpose24
_TEXT SEGMENT
t0$ = 16
t1$ = 64
source$ = 320
target$ = 328
transpose24 PROC ; COMDAT
; 17 : {
$LN3:
00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
0000a 55 push rbp
0000b 57 push rdi
0000c 48 81 ec 48 01
00 00 sub rsp, 328 ; 00000148H
00013 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
00018 48 8d 0d 00 00
00 00 lea rcx, OFFSET FLAT:__B3FFA7AA_mat2@c
0001f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
00024 90 npad 1
; 18 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
00025 b8 10 00 00 00 mov eax, 16
0002a 48 6b c0 01 imul rax, rax, 1
0002e 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
00035 48 03 c8 add rcx, rax
00038 48 8b c1 mov rax, rcx
0003b 0f 10 00 movups xmm0, XMMWORD PTR [rax]
0003e b8 10 00 00 00 mov eax, 16
00043 48 6b c0 00 imul rax, rax, 0
00047 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
0004e 48 03 c8 add rcx, rax
00051 48 8b c1 mov rax, rcx
00054 0f 10 08 movups xmm1, XMMWORD PTR [rax]
00057 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H
0005b 0f 28 c1 movaps xmm0, xmm1
0005e 0f 29 45 10 movaps XMMWORD PTR t0$[rbp], xmm0
; 19 : _MM_SHUFFLE(1, 0, 1, 0));
; 20 : __m128 t1 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
00062 b8 10 00 00 00 mov eax, 16
00067 48 6b c0 01 imul rax, rax, 1
0006b 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
00072 48 03 c8 add rcx, rax
00075 48 8b c1 mov rax, rcx
00078 0f 10 00 movups xmm0, XMMWORD PTR [rax]
0007b b8 10 00 00 00 mov eax, 16
00080 48 6b c0 00 imul rax, rax, 0
00084 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
0008b 48 03 c8 add rcx, rax
0008e 48 8b c1 mov rax, rcx
00091 0f 10 08 movups xmm1, XMMWORD PTR [rax]
00094 0f c6 c8 ee shufps xmm1, xmm0, 238 ; 000000eeH
00098 0f 28 c1 movaps xmm0, xmm1
0009b 0f 29 45 40 movaps XMMWORD PTR t1$[rbp], xmm0
; 21 : _MM_SHUFFLE(3, 2, 3, 2));
; 22 :
; 23 : target->column[0] = _mm_shuffle_ps(t0, zero_vec, _MM_SHUFFLE(2, 0, 2, 0));
0009f 0f 57 c0 xorps xmm0, xmm0
000a2 0f 28 4d 10 movaps xmm1, XMMWORD PTR t0$[rbp]
000a6 0f c6 c8 88 shufps xmm1, xmm0, 136 ; 00000088H
000aa 0f 28 c1 movaps xmm0, xmm1
000ad b8 10 00 00 00 mov eax, 16
000b2 48 6b c0 00 imul rax, rax, 0
000b6 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR target$[rbp]
000bd 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
; 24 : target->column[1] = _mm_shuffle_ps(t0, zero_vec, _MM_SHUFFLE(3, 1, 3, 1));
000c1 0f 57 c0 xorps xmm0, xmm0
000c4 0f 28 4d 10 movaps xmm1, XMMWORD PTR t0$[rbp]
000c8 0f c6 c8 dd shufps xmm1, xmm0, 221 ; 000000ddH
000cc 0f 28 c1 movaps xmm0, xmm1
000cf b8 10 00 00 00 mov eax, 16
000d4 48 6b c0 01 imul rax, rax, 1
000d8 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR target$[rbp]
000df 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
; 25 : target->column[2] = _mm_shuffle_ps(t1, zero_vec, _MM_SHUFFLE(2, 0, 2, 0));
000e3 0f 57 c0 xorps xmm0, xmm0
000e6 0f 28 4d 40 movaps xmm1, XMMWORD PTR t1$[rbp]
000ea 0f c6 c8 88 shufps xmm1, xmm0, 136 ; 00000088H
000ee 0f 28 c1 movaps xmm0, xmm1
000f1 b8 10 00 00 00 mov eax, 16
000f6 48 6b c0 02 imul rax, rax, 2
000fa 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR target$[rbp]
00101 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
; 26 : target->column[3] = _mm_shuffle_ps(t1, zero_vec, _MM_SHUFFLE(3, 1, 3, 1));
00105 0f 57 c0 xorps xmm0, xmm0
00108 0f 28 4d 40 movaps xmm1, XMMWORD PTR t1$[rbp]
0010c 0f c6 c8 dd shufps xmm1, xmm0, 221 ; 000000ddH
00110 0f 28 c1 movaps xmm0, xmm1
00113 b8 10 00 00 00 mov eax, 16
00118 48 6b c0 03 imul rax, rax, 3
0011c 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR target$[rbp]
00123 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
; 27 : }
00127 48 8d a5 28 01
00 00 lea rsp, QWORD PTR [rbp+296]
0012e 5f pop rdi
0012f 5d pop rbp
00130 c3 ret 0
transpose24 ENDP
_TEXT ENDS
; Function compile flags: /Odtp /RTCsu /ZI
; File D:\platform-dev\StgSharp.Native\sn-src\mat2.c
; COMDAT transpose23
_TEXT SEGMENT
t0$ = 16
t1$ = 64
source$ = 320
target$ = 328
transpose23 PROC ; COMDAT
; 5 : {
$LN3:
00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
0000a 55 push rbp
0000b 57 push rdi
0000c 48 81 ec 48 01
00 00 sub rsp, 328 ; 00000148H
00013 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
00018 48 8d 0d 00 00
00 00 lea rcx, OFFSET FLAT:__B3FFA7AA_mat2@c
0001f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
00024 90 npad 1
; 6 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
00025 b8 10 00 00 00 mov eax, 16
0002a 48 6b c0 01 imul rax, rax, 1
0002e 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
00035 48 03 c8 add rcx, rax
00038 48 8b c1 mov rax, rcx
0003b 0f 10 00 movups xmm0, XMMWORD PTR [rax]
0003e b8 10 00 00 00 mov eax, 16
00043 48 6b c0 00 imul rax, rax, 0
00047 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
0004e 48 03 c8 add rcx, rax
00051 48 8b c1 mov rax, rcx
00054 0f 10 08 movups xmm1, XMMWORD PTR [rax]
00057 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H
0005b 0f 28 c1 movaps xmm0, xmm1
0005e 0f 29 45 10 movaps XMMWORD PTR t0$[rbp], xmm0
; 7 : _MM_SHUFFLE(1, 0, 1, 0));
; 8 : __m128 t1 = _mm_shuffle_ps(ALIGN(source->column[0]), ALIGN(source->column[1]),
00062 b8 10 00 00 00 mov eax, 16
00067 48 6b c0 01 imul rax, rax, 1
0006b 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
00072 48 03 c8 add rcx, rax
00075 48 8b c1 mov rax, rcx
00078 0f 10 00 movups xmm0, XMMWORD PTR [rax]
0007b b8 10 00 00 00 mov eax, 16
00080 48 6b c0 00 imul rax, rax, 0
00084 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR source$[rbp]
0008b 48 03 c8 add rcx, rax
0008e 48 8b c1 mov rax, rcx
00091 0f 10 08 movups xmm1, XMMWORD PTR [rax]
00094 0f c6 c8 ee shufps xmm1, xmm0, 238 ; 000000eeH
00098 0f 28 c1 movaps xmm0, xmm1
0009b 0f 29 45 40 movaps XMMWORD PTR t1$[rbp], xmm0
; 9 : _MM_SHUFFLE(3, 2, 3, 2));
; 10 :
; 11 : target->column[0] = _mm_shuffle_ps(t0, zero_vec, _MM_SHUFFLE(2, 0, 2, 0));
0009f 0f 57 c0 xorps xmm0, xmm0
000a2 0f 28 4d 10 movaps xmm1, XMMWORD PTR t0$[rbp]
000a6 0f c6 c8 88 shufps xmm1, xmm0, 136 ; 00000088H
000aa 0f 28 c1 movaps xmm0, xmm1
000ad b8 10 00 00 00 mov eax, 16
000b2 48 6b c0 00 imul rax, rax, 0
000b6 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR target$[rbp]
000bd 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
; 12 : target->column[1] = _mm_shuffle_ps(t0, zero_vec, _MM_SHUFFLE(3, 1, 3, 1));
000c1 0f 57 c0 xorps xmm0, xmm0
000c4 0f 28 4d 10 movaps xmm1, XMMWORD PTR t0$[rbp]
000c8 0f c6 c8 dd shufps xmm1, xmm0, 221 ; 000000ddH
000cc 0f 28 c1 movaps xmm0, xmm1
000cf b8 10 00 00 00 mov eax, 16
000d4 48 6b c0 01 imul rax, rax, 1
000d8 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR target$[rbp]
000df 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
; 13 : target->column[2] = _mm_shuffle_ps(t1, zero_vec, _MM_SHUFFLE(2, 0, 2, 0));
000e3 0f 57 c0 xorps xmm0, xmm0
000e6 0f 28 4d 40 movaps xmm1, XMMWORD PTR t1$[rbp]
000ea 0f c6 c8 88 shufps xmm1, xmm0, 136 ; 00000088H
000ee 0f 28 c1 movaps xmm0, xmm1
000f1 b8 10 00 00 00 mov eax, 16
000f6 48 6b c0 02 imul rax, rax, 2
000fa 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR target$[rbp]
00101 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
; 14 : }
00105 48 8d a5 28 01
00 00 lea rsp, QWORD PTR [rbp+296]
0010c 5f pop rdi
0010d 5d pop rbp
0010e c3 ret 0
transpose23 ENDP
_TEXT ENDS
; Function compile flags: /Odtp /RTCsu /ZI
; File D:\platform-dev\StgSharp.Native\sn-src\mat2.c
; COMDAT submatrix2_sse
_TEXT SEGMENT
temp0$ = 16
temp1$ = 64
left$ = 320
right$ = 328
ans$ = 336
submatrix2_sse PROC ; COMDAT
; 45 : {
$LN3:
00000 4c 89 44 24 18 mov QWORD PTR [rsp+24], r8
00005 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
0000a 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
0000f 55 push rbp
00010 57 push rdi
00011 48 81 ec 48 01
00 00 sub rsp, 328 ; 00000148H
00018 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
0001d 48 8d 0d 00 00
00 00 lea rcx, OFFSET FLAT:__B3FFA7AA_mat2@c
00024 e8 00 00 00 00 call __CheckForDebuggerJustMyCode
00029 90 npad 1
; 46 : __m128 temp0 = _mm_sub_ps(ALIGN(left->column[0]), ALIGN(right->column[0]));
0002a b8 10 00 00 00 mov eax, 16
0002f 48 6b c0 00 imul rax, rax, 0
00033 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR right$[rbp]
0003a 48 03 c8 add rcx, rax
0003d 48 8b c1 mov rax, rcx
00040 0f 10 00 movups xmm0, XMMWORD PTR [rax]
00043 b8 10 00 00 00 mov eax, 16
00048 48 6b c0 00 imul rax, rax, 0
0004c 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR left$[rbp]
00053 48 03 c8 add rcx, rax
00056 48 8b c1 mov rax, rcx
00059 0f 10 08 movups xmm1, XMMWORD PTR [rax]
0005c 0f 5c c8 subps xmm1, xmm0
0005f 0f 28 c1 movaps xmm0, xmm1
00062 0f 29 45 10 movaps XMMWORD PTR temp0$[rbp], xmm0
; 47 : __m128 temp1 = _mm_sub_ps(ALIGN(left->column[1]), ALIGN(right->column[1]));
00066 b8 10 00 00 00 mov eax, 16
0006b 48 6b c0 01 imul rax, rax, 1
0006f 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR right$[rbp]
00076 48 03 c8 add rcx, rax
00079 48 8b c1 mov rax, rcx
0007c 0f 10 00 movups xmm0, XMMWORD PTR [rax]
0007f b8 10 00 00 00 mov eax, 16
00084 48 6b c0 01 imul rax, rax, 1
00088 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR left$[rbp]
0008f 48 03 c8 add rcx, rax
00092 48 8b c1 mov rax, rcx
00095 0f 10 08 movups xmm1, XMMWORD PTR [rax]
00098 0f 5c c8 subps xmm1, xmm0
0009b 0f 28 c1 movaps xmm0, xmm1
0009e 0f 29 45 40 movaps XMMWORD PTR temp1$[rbp], xmm0
; 48 :
; 49 : _mm_storeu_ps(&ans->m[0][0], temp0);
000a2 b8 08 00 00 00 mov eax, 8
000a7 48 6b c0 00 imul rax, rax, 0
000ab 48 8b 8d 50 01
00 00 mov rcx, QWORD PTR ans$[rbp]
000b2 48 03 c8 add rcx, rax
000b5 48 8b c1 mov rax, rcx
000b8 b9 04 00 00 00 mov ecx, 4
000bd 48 6b c9 00 imul rcx, rcx, 0
000c1 48 03 c1 add rax, rcx
000c4 0f 28 45 10 movaps xmm0, XMMWORD PTR temp0$[rbp]
000c8 0f 11 00 movups XMMWORD PTR [rax], xmm0
; 50 : _mm_storeu_ps(&ans->m[0][1], temp1);
000cb b8 08 00 00 00 mov eax, 8
000d0 48 6b c0 00 imul rax, rax, 0
000d4 48 8b 8d 50 01
00 00 mov rcx, QWORD PTR ans$[rbp]
000db 48 03 c8 add rcx, rax
000de 48 8b c1 mov rax, rcx
000e1 b9 04 00 00 00 mov ecx, 4
000e6 48 6b c9 01 imul rcx, rcx, 1
000ea 48 03 c1 add rax, rcx
000ed 0f 28 45 40 movaps xmm0, XMMWORD PTR temp1$[rbp]
000f1 0f 11 00 movups XMMWORD PTR [rax], xmm0
; 51 : }
000f4 48 8d a5 28 01
00 00 lea rsp, QWORD PTR [rbp+296]
000fb 5f pop rdi
000fc 5d pop rbp
000fd c3 ret 0
submatrix2_sse ENDP
_TEXT ENDS
; Function compile flags: /Odtp /RTCsu /ZI
; File D:\platform-dev\StgSharp.Native\sn-src\mat2.c
; COMDAT addmatrix2_avx
_TEXT SEGMENT
temp$ = 32
left$ = 336
right$ = 344
ans$ = 352
addmatrix2_avx PROC ; COMDAT
; 39 : {
$LN3:
00000 4c 89 44 24 18 mov QWORD PTR [rsp+24], r8
00005 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
0000a 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
0000f 41 55 push r13
00011 55 push rbp
00012 57 push rdi
00013 48 81 ec 50 01
00 00 sub rsp, 336 ; 00000150H
0001a 4c 8d 6c 24 20 lea r13, QWORD PTR [rsp+32]
0001f 48 8d 6c 24 40 lea rbp, QWORD PTR [rsp+64]
00024 48 83 e5 e0 and rbp, -32 ; ffffffffffffffe0H
00028 48 8d 0d 00 00
00 00 lea rcx, OFFSET FLAT:__B3FFA7AA_mat2@c
0002f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
00034 90 npad 1
; 40 : __m256 temp = _mm256_add_ps(ALIGN256(left->column[0]), ALIGN256(right->column[0]));
00035 b8 10 00 00 00 mov eax, 16
0003a 48 6b c0 00 imul rax, rax, 0
0003e 49 8b 8d 58 01
00 00 mov rcx, QWORD PTR right$[r13]
00045 48 03 c8 add rcx, rax
00048 48 8b c1 mov rax, rcx
0004b c5 fc 10 00 vmovups ymm0, YMMWORD PTR [rax]
0004f b8 10 00 00 00 mov eax, 16
00054 48 6b c0 00 imul rax, rax, 0
00058 49 8b 8d 50 01
00 00 mov rcx, QWORD PTR left$[r13]
0005f 48 03 c8 add rcx, rax
00062 48 8b c1 mov rax, rcx
00065 c5 fc 10 08 vmovups ymm1, YMMWORD PTR [rax]
00069 c5 f4 58 c0 vaddps ymm0, ymm1, ymm0
0006d c5 fc 11 45 20 vmovups YMMWORD PTR temp$[rbp], ymm0
; 41 : _mm256_storeu_ps(&ans->m[0][0], temp);
00072 b8 08 00 00 00 mov eax, 8
00077 48 6b c0 00 imul rax, rax, 0
0007b 49 8b 8d 60 01
00 00 mov rcx, QWORD PTR ans$[r13]
00082 48 03 c8 add rcx, rax
00085 48 8b c1 mov rax, rcx
00088 b9 04 00 00 00 mov ecx, 4
0008d 48 6b c9 00 imul rcx, rcx, 0
00091 48 03 c1 add rax, rcx
00094 c5 fc 10 45 20 vmovups ymm0, YMMWORD PTR temp$[rbp]
00099 c5 fc 11 00 vmovups YMMWORD PTR [rax], ymm0
; 42 : }
0009d 49 8d a5 30 01
00 00 lea rsp, QWORD PTR [r13+304]
000a4 5f pop rdi
000a5 5d pop rbp
000a6 41 5d pop r13
000a8 c3 ret 0
addmatrix2_avx ENDP
_TEXT ENDS
; Function compile flags: /Odtp /RTCsu /ZI
; File D:\platform-dev\StgSharp.Native\sn-src\mat2.c
; COMDAT addmatrix2_sse
_TEXT SEGMENT
temp0$ = 16
temp1$ = 64
left$ = 320
right$ = 328
ans$ = 336
addmatrix2_sse PROC ; COMDAT
; 30 : {
$LN3:
00000 4c 89 44 24 18 mov QWORD PTR [rsp+24], r8
00005 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
0000a 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
0000f 55 push rbp
00010 57 push rdi
00011 48 81 ec 48 01
00 00 sub rsp, 328 ; 00000148H
00018 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
0001d 48 8d 0d 00 00
00 00 lea rcx, OFFSET FLAT:__B3FFA7AA_mat2@c
00024 e8 00 00 00 00 call __CheckForDebuggerJustMyCode
00029 90 npad 1
; 31 : __m128 temp0 = _mm_add_ps(ALIGN(left->column[0]), ALIGN(right->column[0]));
0002a b8 10 00 00 00 mov eax, 16
0002f 48 6b c0 00 imul rax, rax, 0
00033 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR right$[rbp]
0003a 48 03 c8 add rcx, rax
0003d 48 8b c1 mov rax, rcx
00040 0f 10 00 movups xmm0, XMMWORD PTR [rax]
00043 b8 10 00 00 00 mov eax, 16
00048 48 6b c0 00 imul rax, rax, 0
0004c 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR left$[rbp]
00053 48 03 c8 add rcx, rax
00056 48 8b c1 mov rax, rcx
00059 0f 10 08 movups xmm1, XMMWORD PTR [rax]
0005c 0f 58 c8 addps xmm1, xmm0
0005f 0f 28 c1 movaps xmm0, xmm1
00062 0f 29 45 10 movaps XMMWORD PTR temp0$[rbp], xmm0
; 32 : __m128 temp1 = _mm_add_ps(ALIGN(left->column[1]), ALIGN(right->column[1]));
00066 b8 10 00 00 00 mov eax, 16
0006b 48 6b c0 01 imul rax, rax, 1
0006f 48 8b 8d 48 01
00 00 mov rcx, QWORD PTR right$[rbp]
00076 48 03 c8 add rcx, rax
00079 48 8b c1 mov rax, rcx
0007c 0f 10 00 movups xmm0, XMMWORD PTR [rax]
0007f b8 10 00 00 00 mov eax, 16
00084 48 6b c0 01 imul rax, rax, 1
00088 48 8b 8d 40 01
00 00 mov rcx, QWORD PTR left$[rbp]
0008f 48 03 c8 add rcx, rax
00092 48 8b c1 mov rax, rcx
00095 0f 10 08 movups xmm1, XMMWORD PTR [rax]
00098 0f 58 c8 addps xmm1, xmm0
0009b 0f 28 c1 movaps xmm0, xmm1
0009e 0f 29 45 40 movaps XMMWORD PTR temp1$[rbp], xmm0
; 33 :
; 34 : _mm_storeu_ps(&ans->m[0][0], temp0);
000a2 b8 08 00 00 00 mov eax, 8
000a7 48 6b c0 00 imul rax, rax, 0
000ab 48 8b 8d 50 01
00 00 mov rcx, QWORD PTR ans$[rbp]
000b2 48 03 c8 add rcx, rax
000b5 48 8b c1 mov rax, rcx
000b8 b9 04 00 00 00 mov ecx, 4
000bd 48 6b c9 00 imul rcx, rcx, 0
000c1 48 03 c1 add rax, rcx
000c4 0f 28 45 10 movaps xmm0, XMMWORD PTR temp0$[rbp]
000c8 0f 11 00 movups XMMWORD PTR [rax], xmm0
; 35 : _mm_storeu_ps(&ans->m[0][1], temp1);
000cb b8 08 00 00 00 mov eax, 8
000d0 48 6b c0 00 imul rax, rax, 0
000d4 48 8b 8d 50 01
00 00 mov rcx, QWORD PTR ans$[rbp]
000db 48 03 c8 add rcx, rax
000de 48 8b c1 mov rax, rcx
000e1 b9 04 00 00 00 mov ecx, 4
000e6 48 6b c9 01 imul rcx, rcx, 1
000ea 48 03 c1 add rax, rcx
000ed 0f 28 45 40 movaps xmm0, XMMWORD PTR temp1$[rbp]
000f1 0f 11 00 movups XMMWORD PTR [rax], xmm0
; 36 : }
000f4 48 8d a5 28 01
00 00 lea rsp, QWORD PTR [rbp+296]
000fb 5f pop rdi
000fc 5d pop rbp
000fd c3 ret 0
addmatrix2_sse ENDP
_TEXT ENDS
END
Reference in New Issue View Git Blame Copy Permalink