daily update on 20240226 by nitload youmu
This commit is contained in:
@@ -0,0 +1,890 @@
|
||||
; Listing generated by Microsoft (R) Optimizing Compiler Version 19.39.33519.0
|
||||
|
||||
include listing.inc
|
||||
|
||||
INCLUDELIB MSVCRTD
|
||||
INCLUDELIB OLDNAMES
|
||||
|
||||
_DATA SEGMENT
|
||||
COMM infolog:QWORD:0200H
|
||||
_DATA ENDS
|
||||
msvcjmc SEGMENT
|
||||
__0ED14FEE_winpackagefamily@h DB 01H
|
||||
__F92EF742_winapifamily@h DB 01H
|
||||
__22E515EB_sdkddkver@h DB 01H
|
||||
__4B783B3C_concurrencysal@h DB 01H
|
||||
__137BB1BE_sal@h DB 01H
|
||||
__805F4069_vadefs@h DB 01H
|
||||
__1EAACA9B_vcruntime@h DB 01H
|
||||
__1227BCE2_excpt@h DB 01H
|
||||
__CC8D90C0_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
|
||||
__1B0272D9_corecrt@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
|
||||
__0504B279_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
|
||||
__FA48C214_ssgc_framework@h DB 01H
|
||||
__6C418DCA_ssgc_internal@h DB 01H
|
||||
__F8522E0F_stdint@h DB 01H
|
||||
__761E84A4_KHRplatform@h DB 01H
|
||||
__D5035E09_glad_gl@h DB 01H
|
||||
__58BCF33E_glad_wgl@h DB 01H
|
||||
__13F5961F_stddef@h DB 01H
|
||||
__8E16B543_glfw3@h DB 01H
|
||||
__5BC5AF51_mmintrin@h DB 01H
|
||||
__45DED006_corecrt_malloc@h DB 01H
|
||||
__FA77ED4C_malloc@h DB 01H
|
||||
__9E7FA6CF_xmmintrin@h DB 01H
|
||||
__FB1D2F03_emmintrin@h DB 01H
|
||||
__24BC478A_pmmintrin@h DB 01H
|
||||
__94653408_tmmintrin@h DB 01H
|
||||
__BD5E218B_smmintrin@h DB 01H
|
||||
__D83CA847_nmmintrin@h DB 01H
|
||||
__0D875209_wmmintrin@h DB 01H
|
||||
__C6131F0E_zmmintrin@h DB 01H
|
||||
__F107BDC4_immintrin@h DB 01H
|
||||
__44420885_corecrt_stdio_config@h DB 01H
|
||||
__AB601D8D_corecrt_wstdio@h DB 01H
|
||||
__C1EC35D3_stdio@h DB 01H
|
||||
__FD6DBD68_corecrt_search@h DB 01H
|
||||
__E2E33A61_corecrt_wstdlib@h DB 01H
|
||||
__46240030_limits@h DB 01H
|
||||
__6E5ACABA_stdlib@h DB 01H
|
||||
__A27F041F_stb_image@h DB 01H
|
||||
__3E20A821_StgSharpC@h DB 01H
|
||||
__C8529849_mat3@c DB 01H
|
||||
msvcjmc ENDS
|
||||
PUBLIC transpose3to2
|
||||
PUBLIC transpose3to3
|
||||
PUBLIC transpose3to4
|
||||
PUBLIC det_mat3
|
||||
PUBLIC det_mat3_internal
|
||||
PUBLIC __JustMyCode_Default
|
||||
EXTRN _RTC_CheckStackVars:PROC
|
||||
EXTRN _RTC_InitBase:PROC
|
||||
EXTRN _RTC_Shutdown:PROC
|
||||
EXTRN __CheckForDebuggerJustMyCode:PROC
|
||||
EXTRN defaultVecPtr:QWORD
|
||||
EXTRN _fltused:DWORD
|
||||
; COMDAT pdata
|
||||
pdata SEGMENT
|
||||
$pdata$transpose3to2 DD imagerel $LN3
|
||||
DD imagerel $LN3+238
|
||||
DD imagerel $unwind$transpose3to2
|
||||
pdata ENDS
|
||||
; COMDAT pdata
|
||||
pdata SEGMENT
|
||||
$pdata$transpose3to3 DD imagerel $LN3
|
||||
DD imagerel $LN3+407
|
||||
DD imagerel $unwind$transpose3to3
|
||||
pdata ENDS
|
||||
; COMDAT pdata
|
||||
pdata SEGMENT
|
||||
$pdata$transpose3to4 DD imagerel $LN3
|
||||
DD imagerel $LN3+439
|
||||
DD imagerel $unwind$transpose3to4
|
||||
pdata ENDS
|
||||
; COMDAT pdata
|
||||
pdata SEGMENT
|
||||
$pdata$det_mat3 DD imagerel $LN3
|
||||
DD imagerel $LN3+148
|
||||
DD imagerel $unwind$det_mat3
|
||||
pdata ENDS
|
||||
; COMDAT pdata
|
||||
pdata SEGMENT
|
||||
$pdata$det_mat3_internal DD imagerel $LN3
|
||||
DD imagerel $LN3+370
|
||||
DD imagerel $unwind$det_mat3_internal
|
||||
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
|
||||
$unwind$det_mat3_internal DD 035053601H
|
||||
DD 0118331dH
|
||||
DD 070110049H
|
||||
DD 05010H
|
||||
xdata ENDS
|
||||
; COMDAT CONST
|
||||
CONST SEGMENT
|
||||
det_mat3_internal$rtcName$0 DB 072H
|
||||
DB 065H
|
||||
DB 074H
|
||||
DB 00H
|
||||
ORG $+12
|
||||
det_mat3_internal$rtcVarDesc DD 0160H
|
||||
DD 010H
|
||||
DQ FLAT:det_mat3_internal$rtcName$0
|
||||
ORG $+48
|
||||
det_mat3_internal$rtcFrameData DD 01H
|
||||
DD 00H
|
||||
DQ FLAT:det_mat3_internal$rtcVarDesc
|
||||
CONST ENDS
|
||||
; COMDAT xdata
|
||||
xdata SEGMENT
|
||||
$unwind$det_mat3 DD 035051301H
|
||||
DD 010e3313H
|
||||
DD 070070031H
|
||||
DD 05006H
|
||||
xdata ENDS
|
||||
; COMDAT xdata
|
||||
xdata SEGMENT
|
||||
$unwind$transpose3to4 DD 025051801H
|
||||
DD 01132318H
|
||||
DD 0700c0035H
|
||||
DD 0500bH
|
||||
xdata ENDS
|
||||
; COMDAT xdata
|
||||
xdata SEGMENT
|
||||
$unwind$transpose3to3 DD 025051801H
|
||||
DD 01132318H
|
||||
DD 0700c0035H
|
||||
DD 0500bH
|
||||
xdata ENDS
|
||||
; COMDAT xdata
|
||||
xdata SEGMENT
|
||||
$unwind$transpose3to2 DD 025051801H
|
||||
DD 01132318H
|
||||
DD 0700c0029H
|
||||
DD 0500bH
|
||||
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:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\实验\contexttest\StgSharpC\mat3.c
|
||||
; COMDAT det_mat3_internal
|
||||
_TEXT SEGMENT
|
||||
addc1$ = 16
|
||||
addc2$ = 64
|
||||
subc1$ = 112
|
||||
subc2$ = 160
|
||||
cache$ = 208
|
||||
sub$ = 256
|
||||
ret$ = 304
|
||||
c0$ = 560
|
||||
c1$ = 568
|
||||
c2$ = 576
|
||||
det_mat3_internal PROC ; COMDAT
|
||||
|
||||
; 42 : {
|
||||
|
||||
$LN3:
|
||||
00000 4c 89 44 24 18 mov QWORD PTR [rsp+24], r8
|
||||
00005 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
|
||||
0000a 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
|
||||
0000f 55 push rbp
|
||||
00010 57 push rdi
|
||||
00011 48 81 ec 48 02
|
||||
00 00 sub rsp, 584 ; 00000248H
|
||||
00018 48 8d 6c 24 30 lea rbp, QWORD PTR [rsp+48]
|
||||
0001d 48 8d 7c 24 30 lea rdi, QWORD PTR [rsp+48]
|
||||
00022 b9 56 00 00 00 mov ecx, 86 ; 00000056H
|
||||
00027 b8 cc cc cc cc mov eax, -858993460 ; ccccccccH
|
||||
0002c f3 ab rep stosd
|
||||
0002e 48 8b 8c 24 68
|
||||
02 00 00 mov rcx, QWORD PTR [rsp+616]
|
||||
00036 48 8d 0d 00 00
|
||||
00 00 lea rcx, OFFSET FLAT:__C8529849_mat3@c
|
||||
0003d e8 00 00 00 00 call __CheckForDebuggerJustMyCode
|
||||
|
||||
; 43 :
|
||||
; 44 : //let adding row aligned
|
||||
; 45 : __m128 addc1 = _mm_shuffle_ps(c1, c1, _MM_SHUFFLE(3, 0, 2, 1));
|
||||
|
||||
00042 48 8b 85 38 02
|
||||
00 00 mov rax, QWORD PTR c1$[rbp]
|
||||
00049 48 8b 8d 38 02
|
||||
00 00 mov rcx, QWORD PTR c1$[rbp]
|
||||
00050 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
00053 0f c6 01 c9 shufps xmm0, XMMWORD PTR [rcx], 201 ; 000000c9H
|
||||
00057 0f 29 45 10 movaps XMMWORD PTR addc1$[rbp], xmm0
|
||||
|
||||
; 46 : __m128 addc2 = _mm_shuffle_ps(c2, c2, _MM_SHUFFLE(3, 1, 0, 2));
|
||||
|
||||
0005b 48 8b 85 40 02
|
||||
00 00 mov rax, QWORD PTR c2$[rbp]
|
||||
00062 48 8b 8d 40 02
|
||||
00 00 mov rcx, QWORD PTR c2$[rbp]
|
||||
00069 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
0006c 0f c6 01 d2 shufps xmm0, XMMWORD PTR [rcx], 210 ; 000000d2H
|
||||
00070 0f 29 45 40 movaps XMMWORD PTR addc2$[rbp], xmm0
|
||||
|
||||
; 47 :
|
||||
; 48 : //let sub row aligned
|
||||
; 49 : __m128 subc1 = _mm_shuffle_ps(c1, c1, _MM_SHUFFLE(3, 1, 0, 2));
|
||||
|
||||
00074 48 8b 85 38 02
|
||||
00 00 mov rax, QWORD PTR c1$[rbp]
|
||||
0007b 48 8b 8d 38 02
|
||||
00 00 mov rcx, QWORD PTR c1$[rbp]
|
||||
00082 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
00085 0f c6 01 d2 shufps xmm0, XMMWORD PTR [rcx], 210 ; 000000d2H
|
||||
00089 0f 29 45 70 movaps XMMWORD PTR subc1$[rbp], xmm0
|
||||
|
||||
; 50 : __m128 subc2 = _mm_shuffle_ps(c2, c2, _MM_SHUFFLE(3, 0, 2, 1));
|
||||
|
||||
0008d 48 8b 85 40 02
|
||||
00 00 mov rax, QWORD PTR c2$[rbp]
|
||||
00094 48 8b 8d 40 02
|
||||
00 00 mov rcx, QWORD PTR c2$[rbp]
|
||||
0009b 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
0009e 0f c6 01 c9 shufps xmm0, XMMWORD PTR [rcx], 201 ; 000000c9H
|
||||
000a2 0f 29 85 a0 00
|
||||
00 00 movaps XMMWORD PTR subc2$[rbp], xmm0
|
||||
|
||||
; 51 :
|
||||
; 52 : //multple all add row
|
||||
; 53 : __m128 cache = _mm_mul_ps(c0, addc1);
|
||||
|
||||
000a9 48 8b 85 30 02
|
||||
00 00 mov rax, QWORD PTR c0$[rbp]
|
||||
000b0 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
000b3 0f 59 45 10 mulps xmm0, XMMWORD PTR addc1$[rbp]
|
||||
000b7 0f 29 85 d0 00
|
||||
00 00 movaps XMMWORD PTR cache$[rbp], xmm0
|
||||
|
||||
; 54 : cache = _mm_mul_ps(cache, addc2);
|
||||
|
||||
000be 0f 28 85 d0 00
|
||||
00 00 movaps xmm0, XMMWORD PTR cache$[rbp]
|
||||
000c5 0f 59 45 40 mulps xmm0, XMMWORD PTR addc2$[rbp]
|
||||
000c9 0f 29 85 d0 00
|
||||
00 00 movaps XMMWORD PTR cache$[rbp], xmm0
|
||||
|
||||
; 55 :
|
||||
; 56 : //multiple all sub row
|
||||
; 57 : __m128 sub = _mm_mul_ps(c0, subc1);
|
||||
|
||||
000d0 48 8b 85 30 02
|
||||
00 00 mov rax, QWORD PTR c0$[rbp]
|
||||
000d7 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
000da 0f 59 45 70 mulps xmm0, XMMWORD PTR subc1$[rbp]
|
||||
000de 0f 29 85 00 01
|
||||
00 00 movaps XMMWORD PTR sub$[rbp], xmm0
|
||||
|
||||
; 58 : sub = _mm_mul_ps(sub, subc2);
|
||||
|
||||
000e5 0f 28 85 00 01
|
||||
00 00 movaps xmm0, XMMWORD PTR sub$[rbp]
|
||||
000ec 0f 59 85 a0 00
|
||||
00 00 mulps xmm0, XMMWORD PTR subc2$[rbp]
|
||||
000f3 0f 29 85 00 01
|
||||
00 00 movaps XMMWORD PTR sub$[rbp], xmm0
|
||||
|
||||
; 59 :
|
||||
; 60 : //main-corner sub sub-corner
|
||||
; 61 : cache = _mm_sub_ps(cache, sub);
|
||||
|
||||
000fa 0f 28 85 d0 00
|
||||
00 00 movaps xmm0, XMMWORD PTR cache$[rbp]
|
||||
00101 0f 5c 85 00 01
|
||||
00 00 subps xmm0, XMMWORD PTR sub$[rbp]
|
||||
00108 0f 29 85 d0 00
|
||||
00 00 movaps XMMWORD PTR cache$[rbp], xmm0
|
||||
|
||||
; 62 :
|
||||
; 63 : __m128 ret = _mm_hadd_ps(cache, cache); //element 1 add element 2
|
||||
|
||||
0010f 0f 28 85 d0 00
|
||||
00 00 movaps xmm0, XMMWORD PTR cache$[rbp]
|
||||
00116 f2 0f 7c 85 d0
|
||||
00 00 00 haddps xmm0, XMMWORD PTR cache$[rbp]
|
||||
0011e 0f 29 85 30 01
|
||||
00 00 movaps XMMWORD PTR ret$[rbp], xmm0
|
||||
|
||||
; 64 : ret = _mm_add_ps(ret, cache); //(e1 add e2) add element 3
|
||||
|
||||
00125 0f 28 85 30 01
|
||||
00 00 movaps xmm0, XMMWORD PTR ret$[rbp]
|
||||
0012c 0f 58 85 d0 00
|
||||
00 00 addps xmm0, XMMWORD PTR cache$[rbp]
|
||||
00133 0f 29 85 30 01
|
||||
00 00 movaps XMMWORD PTR ret$[rbp], xmm0
|
||||
|
||||
; 65 :
|
||||
; 66 :
|
||||
; 67 : return ret.m128_f32[2];
|
||||
|
||||
0013a b8 04 00 00 00 mov eax, 4
|
||||
0013f 48 6b c0 02 imul rax, rax, 2
|
||||
00143 f3 0f 10 84 05
|
||||
30 01 00 00 movss xmm0, DWORD PTR ret$[rbp+rax]
|
||||
|
||||
; 68 : }
|
||||
|
||||
0014c f3 0f 7f 44 24
|
||||
20 movdqu XMMWORD PTR [rsp+32], xmm0
|
||||
00152 48 8d 4d d0 lea rcx, QWORD PTR [rbp-48]
|
||||
00156 48 8d 15 00 00
|
||||
00 00 lea rdx, OFFSET FLAT:det_mat3_internal$rtcFrameData
|
||||
0015d e8 00 00 00 00 call _RTC_CheckStackVars
|
||||
00162 f3 0f 6f 44 24
|
||||
20 movdqu xmm0, XMMWORD PTR [rsp+32]
|
||||
00168 48 8d a5 18 02
|
||||
00 00 lea rsp, QWORD PTR [rbp+536]
|
||||
0016f 5f pop rdi
|
||||
00170 5d pop rbp
|
||||
00171 c3 ret 0
|
||||
det_mat3_internal ENDP
|
||||
_TEXT ENDS
|
||||
; Function compile flags: /Odtp /RTCsu /ZI
|
||||
; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\实验\contexttest\StgSharpC\mat3.c
|
||||
; COMDAT det_mat3
|
||||
_TEXT SEGMENT
|
||||
$T1 = 208
|
||||
$T2 = 256
|
||||
$T3 = 304
|
||||
mat$ = 368
|
||||
det_mat3 PROC ; COMDAT
|
||||
|
||||
; 71 : {
|
||||
|
||||
$LN3:
|
||||
00000 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
|
||||
00005 55 push rbp
|
||||
00006 57 push rdi
|
||||
00007 48 81 ec 88 01
|
||||
00 00 sub rsp, 392 ; 00000188H
|
||||
0000e 48 8d 6c 24 30 lea rbp, QWORD PTR [rsp+48]
|
||||
00013 48 8d 0d 00 00
|
||||
00 00 lea rcx, OFFSET FLAT:__C8529849_mat3@c
|
||||
0001a e8 00 00 00 00 call __CheckForDebuggerJustMyCode
|
||||
|
||||
; 72 : return det_mat3_internal(
|
||||
|
||||
0001f b8 10 00 00 00 mov eax, 16
|
||||
00024 48 6b c0 02 imul rax, rax, 2
|
||||
00028 48 8b 8d 70 01
|
||||
00 00 mov rcx, QWORD PTR mat$[rbp]
|
||||
0002f 0f 10 04 01 movups xmm0, XMMWORD PTR [rcx+rax]
|
||||
00033 0f 11 85 30 01
|
||||
00 00 movups XMMWORD PTR $T3[rbp], xmm0
|
||||
0003a b8 10 00 00 00 mov eax, 16
|
||||
0003f 48 6b c0 01 imul rax, rax, 1
|
||||
00043 48 8b 8d 70 01
|
||||
00 00 mov rcx, QWORD PTR mat$[rbp]
|
||||
0004a 0f 10 04 01 movups xmm0, XMMWORD PTR [rcx+rax]
|
||||
0004e 0f 11 85 00 01
|
||||
00 00 movups XMMWORD PTR $T2[rbp], xmm0
|
||||
00055 b8 10 00 00 00 mov eax, 16
|
||||
0005a 48 6b c0 00 imul rax, rax, 0
|
||||
0005e 48 8b 8d 70 01
|
||||
00 00 mov rcx, QWORD PTR mat$[rbp]
|
||||
00065 0f 10 04 01 movups xmm0, XMMWORD PTR [rcx+rax]
|
||||
00069 0f 11 85 d0 00
|
||||
00 00 movups XMMWORD PTR $T1[rbp], xmm0
|
||||
00070 4c 8d 85 30 01
|
||||
00 00 lea r8, QWORD PTR $T3[rbp]
|
||||
00077 48 8d 95 00 01
|
||||
00 00 lea rdx, QWORD PTR $T2[rbp]
|
||||
0007e 48 8d 8d d0 00
|
||||
00 00 lea rcx, QWORD PTR $T1[rbp]
|
||||
00085 e8 00 00 00 00 call det_mat3_internal
|
||||
|
||||
; 73 : mat->colum[0],
|
||||
; 74 : mat->colum[1],
|
||||
; 75 : mat->colum[2]
|
||||
; 76 : );
|
||||
; 77 : }
|
||||
|
||||
0008a 48 8d a5 58 01
|
||||
00 00 lea rsp, QWORD PTR [rbp+344]
|
||||
00091 5f pop rdi
|
||||
00092 5d pop rbp
|
||||
00093 c3 ret 0
|
||||
det_mat3 ENDP
|
||||
_TEXT ENDS
|
||||
; Function compile flags: /Odtp /RTCsu /ZI
|
||||
; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\实验\contexttest\StgSharpC\mat3.c
|
||||
; COMDAT transpose3to4
|
||||
_TEXT SEGMENT
|
||||
t0$ = 16
|
||||
t1$ = 64
|
||||
t2$ = 112
|
||||
t3$ = 160
|
||||
source$ = 416
|
||||
target$ = 424
|
||||
transpose3to4 PROC ; COMDAT
|
||||
|
||||
; 29 : {
|
||||
|
||||
$LN3:
|
||||
00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
|
||||
00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
|
||||
0000a 55 push rbp
|
||||
0000b 57 push rdi
|
||||
0000c 48 81 ec a8 01
|
||||
00 00 sub rsp, 424 ; 000001a8H
|
||||
00013 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
|
||||
00018 48 8d 0d 00 00
|
||||
00 00 lea rcx, OFFSET FLAT:__C8529849_mat3@c
|
||||
0001f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
|
||||
|
||||
; 30 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->m[0][0]), ALIGN(source->m[0][1]), _MM_SHUFFLE(1, 0, 1, 0));
|
||||
|
||||
00024 b8 0c 00 00 00 mov eax, 12
|
||||
00029 48 6b c0 00 imul rax, rax, 0
|
||||
0002d 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00034 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
00039 b9 04 00 00 00 mov ecx, 4
|
||||
0003e 48 6b c9 01 imul rcx, rcx, 1
|
||||
00042 48 03 c1 add rax, rcx
|
||||
00045 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
00048 b8 0c 00 00 00 mov eax, 12
|
||||
0004d 48 6b c0 00 imul rax, rax, 0
|
||||
00051 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00058 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
0005d b9 04 00 00 00 mov ecx, 4
|
||||
00062 48 6b c9 00 imul rcx, rcx, 0
|
||||
00066 48 03 c1 add rax, rcx
|
||||
00069 0f 10 08 movups xmm1, XMMWORD PTR [rax]
|
||||
0006c 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H
|
||||
00070 0f 28 c1 movaps xmm0, xmm1
|
||||
00073 0f 29 45 10 movaps XMMWORD PTR t0$[rbp], xmm0
|
||||
|
||||
; 31 : __m128 t1 = _mm_shuffle_ps(ALIGN(source->m[0][0]), ALIGN(source->m[0][1]), _MM_SHUFFLE(3, 2, 3, 2));
|
||||
|
||||
00077 b8 0c 00 00 00 mov eax, 12
|
||||
0007c 48 6b c0 00 imul rax, rax, 0
|
||||
00080 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00087 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
0008c b9 04 00 00 00 mov ecx, 4
|
||||
00091 48 6b c9 01 imul rcx, rcx, 1
|
||||
00095 48 03 c1 add rax, rcx
|
||||
00098 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
0009b b8 0c 00 00 00 mov eax, 12
|
||||
000a0 48 6b c0 00 imul rax, rax, 0
|
||||
000a4 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
000ab 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
000b0 b9 04 00 00 00 mov ecx, 4
|
||||
000b5 48 6b c9 00 imul rcx, rcx, 0
|
||||
000b9 48 03 c1 add rax, rcx
|
||||
000bc 0f 10 08 movups xmm1, XMMWORD PTR [rax]
|
||||
000bf 0f c6 c8 ee shufps xmm1, xmm0, 238 ; 000000eeH
|
||||
000c3 0f 28 c1 movaps xmm0, xmm1
|
||||
000c6 0f 29 45 40 movaps XMMWORD PTR t1$[rbp], xmm0
|
||||
|
||||
; 32 : __m128 t2 = _mm_shuffle_ps(ALIGN(source->m[0][2]), *defaultVecPtr, _MM_SHUFFLE(1, 0, 1, 0));
|
||||
|
||||
000ca b8 0c 00 00 00 mov eax, 12
|
||||
000cf 48 6b c0 00 imul rax, rax, 0
|
||||
000d3 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
000da 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
000df b9 04 00 00 00 mov ecx, 4
|
||||
000e4 48 6b c9 02 imul rcx, rcx, 2
|
||||
000e8 48 03 c1 add rax, rcx
|
||||
000eb 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
000ee 48 8b 05 00 00
|
||||
00 00 mov rax, QWORD PTR defaultVecPtr
|
||||
000f5 0f c6 00 44 shufps xmm0, XMMWORD PTR [rax], 68 ; 00000044H
|
||||
000f9 0f 29 45 70 movaps XMMWORD PTR t2$[rbp], xmm0
|
||||
|
||||
; 33 : __m128 t3 = _mm_shuffle_ps(ALIGN(source->m[0][2]), *defaultVecPtr, _MM_SHUFFLE(3, 2, 3, 2));
|
||||
|
||||
000fd b8 0c 00 00 00 mov eax, 12
|
||||
00102 48 6b c0 00 imul rax, rax, 0
|
||||
00106 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
0010d 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
00112 b9 04 00 00 00 mov ecx, 4
|
||||
00117 48 6b c9 02 imul rcx, rcx, 2
|
||||
0011b 48 03 c1 add rax, rcx
|
||||
0011e 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
00121 48 8b 05 00 00
|
||||
00 00 mov rax, QWORD PTR defaultVecPtr
|
||||
00128 0f c6 00 ee shufps xmm0, XMMWORD PTR [rax], 238 ; 000000eeH
|
||||
0012c 0f 29 85 a0 00
|
||||
00 00 movaps XMMWORD PTR t3$[rbp], xmm0
|
||||
|
||||
; 34 :
|
||||
; 35 : target->colum[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0));
|
||||
|
||||
00133 0f 28 45 10 movaps xmm0, XMMWORD PTR t0$[rbp]
|
||||
00137 0f c6 45 70 88 shufps xmm0, XMMWORD PTR t2$[rbp], 136 ; 00000088H
|
||||
0013c b8 10 00 00 00 mov eax, 16
|
||||
00141 48 6b c0 00 imul rax, rax, 0
|
||||
00145 48 8b 8d a8 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
0014c 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 36 : target->colum[1] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 1, 3, 1));
|
||||
|
||||
00150 0f 28 45 10 movaps xmm0, XMMWORD PTR t0$[rbp]
|
||||
00154 0f c6 45 70 dd shufps xmm0, XMMWORD PTR t2$[rbp], 221 ; 000000ddH
|
||||
00159 b8 10 00 00 00 mov eax, 16
|
||||
0015e 48 6b c0 01 imul rax, rax, 1
|
||||
00162 48 8b 8d a8 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
00169 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 37 : target->colum[2] = _mm_shuffle_ps(t1, t3, _MM_SHUFFLE(2, 0, 2, 0));
|
||||
|
||||
0016d 0f 28 45 40 movaps xmm0, XMMWORD PTR t1$[rbp]
|
||||
00171 0f c6 85 a0 00
|
||||
00 00 88 shufps xmm0, XMMWORD PTR t3$[rbp], 136 ; 00000088H
|
||||
00179 b8 10 00 00 00 mov eax, 16
|
||||
0017e 48 6b c0 02 imul rax, rax, 2
|
||||
00182 48 8b 8d a8 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
00189 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 38 : target->colum[3] = _mm_shuffle_ps(t1, t3, _MM_SHUFFLE(3, 1, 3, 1));
|
||||
|
||||
0018d 0f 28 45 40 movaps xmm0, XMMWORD PTR t1$[rbp]
|
||||
00191 0f c6 85 a0 00
|
||||
00 00 dd shufps xmm0, XMMWORD PTR t3$[rbp], 221 ; 000000ddH
|
||||
00199 b8 10 00 00 00 mov eax, 16
|
||||
0019e 48 6b c0 03 imul rax, rax, 3
|
||||
001a2 48 8b 8d a8 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
001a9 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 39 : }
|
||||
|
||||
001ad 48 8d a5 88 01
|
||||
00 00 lea rsp, QWORD PTR [rbp+392]
|
||||
001b4 5f pop rdi
|
||||
001b5 5d pop rbp
|
||||
001b6 c3 ret 0
|
||||
transpose3to4 ENDP
|
||||
_TEXT ENDS
|
||||
; Function compile flags: /Odtp /RTCsu /ZI
|
||||
; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\实验\contexttest\StgSharpC\mat3.c
|
||||
; COMDAT transpose3to3
|
||||
_TEXT SEGMENT
|
||||
t0$ = 16
|
||||
t1$ = 64
|
||||
t2$ = 112
|
||||
t3$ = 160
|
||||
source$ = 416
|
||||
target$ = 424
|
||||
transpose3to3 PROC ; COMDAT
|
||||
|
||||
; 16 : {
|
||||
|
||||
$LN3:
|
||||
00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
|
||||
00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
|
||||
0000a 55 push rbp
|
||||
0000b 57 push rdi
|
||||
0000c 48 81 ec a8 01
|
||||
00 00 sub rsp, 424 ; 000001a8H
|
||||
00013 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
|
||||
00018 48 8d 0d 00 00
|
||||
00 00 lea rcx, OFFSET FLAT:__C8529849_mat3@c
|
||||
0001f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
|
||||
|
||||
; 17 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->m[0][0]), ALIGN(source->m[0][1]), _MM_SHUFFLE(1, 0, 1, 0));
|
||||
|
||||
00024 b8 0c 00 00 00 mov eax, 12
|
||||
00029 48 6b c0 00 imul rax, rax, 0
|
||||
0002d 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00034 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
00039 b9 04 00 00 00 mov ecx, 4
|
||||
0003e 48 6b c9 01 imul rcx, rcx, 1
|
||||
00042 48 03 c1 add rax, rcx
|
||||
00045 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
00048 b8 0c 00 00 00 mov eax, 12
|
||||
0004d 48 6b c0 00 imul rax, rax, 0
|
||||
00051 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00058 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
0005d b9 04 00 00 00 mov ecx, 4
|
||||
00062 48 6b c9 00 imul rcx, rcx, 0
|
||||
00066 48 03 c1 add rax, rcx
|
||||
00069 0f 10 08 movups xmm1, XMMWORD PTR [rax]
|
||||
0006c 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H
|
||||
00070 0f 28 c1 movaps xmm0, xmm1
|
||||
00073 0f 29 45 10 movaps XMMWORD PTR t0$[rbp], xmm0
|
||||
|
||||
; 18 : __m128 t1 = _mm_shuffle_ps(ALIGN(source->m[0][0]), ALIGN(source->m[0][1]), _MM_SHUFFLE(3, 2, 3, 2));
|
||||
|
||||
00077 b8 0c 00 00 00 mov eax, 12
|
||||
0007c 48 6b c0 00 imul rax, rax, 0
|
||||
00080 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00087 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
0008c b9 04 00 00 00 mov ecx, 4
|
||||
00091 48 6b c9 01 imul rcx, rcx, 1
|
||||
00095 48 03 c1 add rax, rcx
|
||||
00098 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
0009b b8 0c 00 00 00 mov eax, 12
|
||||
000a0 48 6b c0 00 imul rax, rax, 0
|
||||
000a4 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
000ab 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
000b0 b9 04 00 00 00 mov ecx, 4
|
||||
000b5 48 6b c9 00 imul rcx, rcx, 0
|
||||
000b9 48 03 c1 add rax, rcx
|
||||
000bc 0f 10 08 movups xmm1, XMMWORD PTR [rax]
|
||||
000bf 0f c6 c8 ee shufps xmm1, xmm0, 238 ; 000000eeH
|
||||
000c3 0f 28 c1 movaps xmm0, xmm1
|
||||
000c6 0f 29 45 40 movaps XMMWORD PTR t1$[rbp], xmm0
|
||||
|
||||
; 19 : __m128 t2 = _mm_shuffle_ps(ALIGN(source->m[0][2]), *defaultVecPtr, _MM_SHUFFLE(1, 0, 1, 0));
|
||||
|
||||
000ca b8 0c 00 00 00 mov eax, 12
|
||||
000cf 48 6b c0 00 imul rax, rax, 0
|
||||
000d3 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
000da 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
000df b9 04 00 00 00 mov ecx, 4
|
||||
000e4 48 6b c9 02 imul rcx, rcx, 2
|
||||
000e8 48 03 c1 add rax, rcx
|
||||
000eb 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
000ee 48 8b 05 00 00
|
||||
00 00 mov rax, QWORD PTR defaultVecPtr
|
||||
000f5 0f c6 00 44 shufps xmm0, XMMWORD PTR [rax], 68 ; 00000044H
|
||||
000f9 0f 29 45 70 movaps XMMWORD PTR t2$[rbp], xmm0
|
||||
|
||||
; 20 : __m128 t3 = _mm_shuffle_ps(ALIGN(source->m[0][2]), *defaultVecPtr, _MM_SHUFFLE(3, 2, 3, 2));
|
||||
|
||||
000fd b8 0c 00 00 00 mov eax, 12
|
||||
00102 48 6b c0 00 imul rax, rax, 0
|
||||
00106 48 8b 8d a0 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
0010d 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
00112 b9 04 00 00 00 mov ecx, 4
|
||||
00117 48 6b c9 02 imul rcx, rcx, 2
|
||||
0011b 48 03 c1 add rax, rcx
|
||||
0011e 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
00121 48 8b 05 00 00
|
||||
00 00 mov rax, QWORD PTR defaultVecPtr
|
||||
00128 0f c6 00 ee shufps xmm0, XMMWORD PTR [rax], 238 ; 000000eeH
|
||||
0012c 0f 29 85 a0 00
|
||||
00 00 movaps XMMWORD PTR t3$[rbp], xmm0
|
||||
|
||||
; 21 :
|
||||
; 22 : target->colum[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0));
|
||||
|
||||
00133 0f 28 45 10 movaps xmm0, XMMWORD PTR t0$[rbp]
|
||||
00137 0f c6 45 70 88 shufps xmm0, XMMWORD PTR t2$[rbp], 136 ; 00000088H
|
||||
0013c b8 10 00 00 00 mov eax, 16
|
||||
00141 48 6b c0 00 imul rax, rax, 0
|
||||
00145 48 8b 8d a8 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
0014c 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 23 : target->colum[1] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 1, 3, 1));
|
||||
|
||||
00150 0f 28 45 10 movaps xmm0, XMMWORD PTR t0$[rbp]
|
||||
00154 0f c6 45 70 dd shufps xmm0, XMMWORD PTR t2$[rbp], 221 ; 000000ddH
|
||||
00159 b8 10 00 00 00 mov eax, 16
|
||||
0015e 48 6b c0 01 imul rax, rax, 1
|
||||
00162 48 8b 8d a8 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
00169 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 24 : target->colum[2] = _mm_shuffle_ps(t1, t3, _MM_SHUFFLE(2, 0, 2, 0));
|
||||
|
||||
0016d 0f 28 45 40 movaps xmm0, XMMWORD PTR t1$[rbp]
|
||||
00171 0f c6 85 a0 00
|
||||
00 00 88 shufps xmm0, XMMWORD PTR t3$[rbp], 136 ; 00000088H
|
||||
00179 b8 10 00 00 00 mov eax, 16
|
||||
0017e 48 6b c0 02 imul rax, rax, 2
|
||||
00182 48 8b 8d a8 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
00189 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 25 : }
|
||||
|
||||
0018d 48 8d a5 88 01
|
||||
00 00 lea rsp, QWORD PTR [rbp+392]
|
||||
00194 5f pop rdi
|
||||
00195 5d pop rbp
|
||||
00196 c3 ret 0
|
||||
transpose3to3 ENDP
|
||||
_TEXT ENDS
|
||||
; Function compile flags: /Odtp /RTCsu /ZI
|
||||
; File D:\NitloadIndustry\研究项目\1-Flagship\TH2020\ProjectFile\实验\contexttest\StgSharpC\mat3.c
|
||||
; COMDAT transpose3to2
|
||||
_TEXT SEGMENT
|
||||
t0$ = 16
|
||||
t2$ = 64
|
||||
source$ = 320
|
||||
target$ = 328
|
||||
transpose3to2 PROC ; COMDAT
|
||||
|
||||
; 7 : {
|
||||
|
||||
$LN3:
|
||||
00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
|
||||
00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
|
||||
0000a 55 push rbp
|
||||
0000b 57 push rdi
|
||||
0000c 48 81 ec 48 01
|
||||
00 00 sub rsp, 328 ; 00000148H
|
||||
00013 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
|
||||
00018 48 8d 0d 00 00
|
||||
00 00 lea rcx, OFFSET FLAT:__C8529849_mat3@c
|
||||
0001f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
|
||||
|
||||
; 8 : __m128 t0 = _mm_shuffle_ps(ALIGN(source->m[0][0]), ALIGN(source->m[0][1]), _MM_SHUFFLE(1, 0, 1, 0));
|
||||
|
||||
00024 b8 0c 00 00 00 mov eax, 12
|
||||
00029 48 6b c0 00 imul rax, rax, 0
|
||||
0002d 48 8b 8d 40 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00034 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
00039 b9 04 00 00 00 mov ecx, 4
|
||||
0003e 48 6b c9 01 imul rcx, rcx, 1
|
||||
00042 48 03 c1 add rax, rcx
|
||||
00045 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
00048 b8 0c 00 00 00 mov eax, 12
|
||||
0004d 48 6b c0 00 imul rax, rax, 0
|
||||
00051 48 8b 8d 40 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00058 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
0005d b9 04 00 00 00 mov ecx, 4
|
||||
00062 48 6b c9 00 imul rcx, rcx, 0
|
||||
00066 48 03 c1 add rax, rcx
|
||||
00069 0f 10 08 movups xmm1, XMMWORD PTR [rax]
|
||||
0006c 0f c6 c8 44 shufps xmm1, xmm0, 68 ; 00000044H
|
||||
00070 0f 28 c1 movaps xmm0, xmm1
|
||||
00073 0f 29 45 10 movaps XMMWORD PTR t0$[rbp], xmm0
|
||||
|
||||
; 9 : __m128 t2 = _mm_shuffle_ps(ALIGN(source->m[0][2]), *defaultVecPtr, _MM_SHUFFLE(1, 0, 1, 0));
|
||||
|
||||
00077 b8 0c 00 00 00 mov eax, 12
|
||||
0007c 48 6b c0 00 imul rax, rax, 0
|
||||
00080 48 8b 8d 40 01
|
||||
00 00 mov rcx, QWORD PTR source$[rbp]
|
||||
00087 48 8d 44 01 30 lea rax, QWORD PTR [rcx+rax+48]
|
||||
0008c b9 04 00 00 00 mov ecx, 4
|
||||
00091 48 6b c9 02 imul rcx, rcx, 2
|
||||
00095 48 03 c1 add rax, rcx
|
||||
00098 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
||||
0009b 48 8b 05 00 00
|
||||
00 00 mov rax, QWORD PTR defaultVecPtr
|
||||
000a2 0f c6 00 44 shufps xmm0, XMMWORD PTR [rax], 68 ; 00000044H
|
||||
000a6 0f 29 45 40 movaps XMMWORD PTR t2$[rbp], xmm0
|
||||
|
||||
; 10 :
|
||||
; 11 : target->colum[0] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(2, 0, 2, 0));
|
||||
|
||||
000aa 0f 28 45 10 movaps xmm0, XMMWORD PTR t0$[rbp]
|
||||
000ae 0f c6 45 40 88 shufps xmm0, XMMWORD PTR t2$[rbp], 136 ; 00000088H
|
||||
000b3 b8 10 00 00 00 mov eax, 16
|
||||
000b8 48 6b c0 00 imul rax, rax, 0
|
||||
000bc 48 8b 8d 48 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
000c3 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 12 : target->colum[1] = _mm_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 1, 3, 1));
|
||||
|
||||
000c7 0f 28 45 10 movaps xmm0, XMMWORD PTR t0$[rbp]
|
||||
000cb 0f c6 45 40 dd shufps xmm0, XMMWORD PTR t2$[rbp], 221 ; 000000ddH
|
||||
000d0 b8 10 00 00 00 mov eax, 16
|
||||
000d5 48 6b c0 01 imul rax, rax, 1
|
||||
000d9 48 8b 8d 48 01
|
||||
00 00 mov rcx, QWORD PTR target$[rbp]
|
||||
000e0 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
||||
|
||||
; 13 : }
|
||||
|
||||
000e4 48 8d a5 28 01
|
||||
00 00 lea rsp, QWORD PTR [rbp+296]
|
||||
000eb 5f pop rdi
|
||||
000ec 5d pop rbp
|
||||
000ed c3 ret 0
|
||||
transpose3to2 ENDP
|
||||
_TEXT ENDS
|
||||
END
|
||||
Reference in New Issue
Block a user