376 lines
12 KiB
Plaintext
376 lines
12 KiB
Plaintext
; Listing generated by Microsoft (R) Optimizing Compiler Version 19.44.35220.0
|
|
|
|
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
|
|
__836A8CFC_stdint@h DB 01H
|
|
__714FE6D7_sn_target@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
|
|
__F562FA57_stb_image@h DB 01H
|
|
__C95DC661_khrplatform@h DB 01H
|
|
__6C577136_gl@h DB 01H
|
|
__EE638A2C_wgl@h DB 01H
|
|
__8EF9E35C_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
|
|
__F49407C4_sn_intrinsic@matkernel@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
|
|
__E15488A2_vector@c DB 01H
|
|
msvcjmc ENDS
|
|
PUBLIC f32_normalize
|
|
PUBLIC __JustMyCode_Default
|
|
EXTRN _RTC_InitBase:PROC
|
|
EXTRN _RTC_Shutdown:PROC
|
|
EXTRN __CheckForDebuggerJustMyCode:PROC
|
|
; COMDAT pdata
|
|
pdata SEGMENT
|
|
$pdata$f32_normalize DD imagerel $LN3
|
|
DD imagerel $LN3+417
|
|
DD imagerel $unwind$f32_normalize
|
|
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
|
|
$unwind$f32_normalize DD 025051801H
|
|
DD 01132318H
|
|
DD 0700c0047H
|
|
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:\platform-dev\StgSharp.Native\sn-src\vector.c
|
|
; COMDAT f32_normalize
|
|
_TEXT SEGMENT
|
|
s$ = 16
|
|
a_length_sq$ = 64
|
|
a_inv_length$ = 112
|
|
dot_product$ = 160
|
|
b_orthogonal$ = 208
|
|
b_length_sq$ = 256
|
|
b_inv_length$ = 304
|
|
source$ = 560
|
|
target$ = 568
|
|
f32_normalize PROC ; COMDAT
|
|
|
|
; 5 : {
|
|
|
|
$LN3:
|
|
00000 48 89 54 24 10 mov QWORD PTR [rsp+16], rdx
|
|
00005 48 89 4c 24 08 mov QWORD PTR [rsp+8], rcx
|
|
0000a 55 push rbp
|
|
0000b 57 push rdi
|
|
0000c 48 81 ec 38 02
|
|
00 00 sub rsp, 568 ; 00000238H
|
|
00013 48 8d 6c 24 20 lea rbp, QWORD PTR [rsp+32]
|
|
00018 48 8d 0d 00 00
|
|
00 00 lea rcx, OFFSET FLAT:__E15488A2_vector@c
|
|
0001f e8 00 00 00 00 call __CheckForDebuggerJustMyCode
|
|
00024 90 npad 1
|
|
|
|
; 6 : register __m128 s = _mm_load_ps(&source->f32_x);
|
|
|
|
00025 48 8b 85 30 02
|
|
00 00 mov rax, QWORD PTR source$[rbp]
|
|
0002c 0f 10 00 movups xmm0, XMMWORD PTR [rax]
|
|
0002f 0f 29 45 10 movaps XMMWORD PTR s$[rbp], xmm0
|
|
|
|
; 7 : register __m128 a_length_sq = _mm_mul_ps(s, s);
|
|
|
|
00033 0f 28 45 10 movaps xmm0, XMMWORD PTR s$[rbp]
|
|
00037 0f 59 45 10 mulps xmm0, XMMWORD PTR s$[rbp]
|
|
0003b 0f 29 45 40 movaps XMMWORD PTR a_length_sq$[rbp], xmm0
|
|
|
|
; 8 : a_length_sq = _mm_add_ps(a_length_sq,
|
|
|
|
0003f 0f 28 45 40 movaps xmm0, XMMWORD PTR a_length_sq$[rbp]
|
|
00043 0f c6 45 40 b1 shufps xmm0, XMMWORD PTR a_length_sq$[rbp], 177 ; 000000b1H
|
|
00048 0f 28 4d 40 movaps xmm1, XMMWORD PTR a_length_sq$[rbp]
|
|
0004c 0f 58 c8 addps xmm1, xmm0
|
|
0004f 0f 28 c1 movaps xmm0, xmm1
|
|
00052 0f 29 45 40 movaps XMMWORD PTR a_length_sq$[rbp], xmm0
|
|
|
|
; 9 : _mm_shuffle_ps(a_length_sq, a_length_sq, _MM_SHUFFLE(2, 3, 0, 1)));
|
|
; 10 : a_length_sq = _mm_add_ss(a_length_sq,
|
|
|
|
00056 0f 28 45 40 movaps xmm0, XMMWORD PTR a_length_sq$[rbp]
|
|
0005a 0f c6 45 40 0e shufps xmm0, XMMWORD PTR a_length_sq$[rbp], 14
|
|
0005f 0f 28 4d 40 movaps xmm1, XMMWORD PTR a_length_sq$[rbp]
|
|
00063 f3 0f 58 c8 addss xmm1, xmm0
|
|
00067 0f 29 4d 40 movaps XMMWORD PTR a_length_sq$[rbp], xmm1
|
|
|
|
; 11 : _mm_shuffle_ps(a_length_sq, a_length_sq, _MM_SHUFFLE(0, 0, 3, 2)));
|
|
; 12 :
|
|
; 13 : __m128 a_inv_length = _mm_rsqrt_ps(a_length_sq);
|
|
|
|
0006b 0f 52 45 40 rsqrtps xmm0, XMMWORD PTR a_length_sq$[rbp]
|
|
0006f 0f 29 45 70 movaps XMMWORD PTR a_inv_length$[rbp], xmm0
|
|
|
|
; 14 : s = _mm_mul_ps(s, a_inv_length);
|
|
|
|
00073 0f 28 45 10 movaps xmm0, XMMWORD PTR s$[rbp]
|
|
00077 0f 59 45 70 mulps xmm0, XMMWORD PTR a_inv_length$[rbp]
|
|
0007b 0f 29 45 10 movaps XMMWORD PTR s$[rbp], xmm0
|
|
|
|
; 15 :
|
|
; 16 : __m128 dot_product = _mm_mul_ps(target->f32_x[0], s);
|
|
|
|
0007f b8 10 00 00 00 mov eax, 16
|
|
00084 48 6b c0 00 imul rax, rax, 0
|
|
00088 48 8b 8d 38 02
|
|
00 00 mov rcx, QWORD PTR target$[rbp]
|
|
0008f 0f 10 04 01 movups xmm0, XMMWORD PTR [rcx+rax]
|
|
00093 0f 59 45 10 mulps xmm0, XMMWORD PTR s$[rbp]
|
|
00097 0f 29 85 a0 00
|
|
00 00 movaps XMMWORD PTR dot_product$[rbp], xmm0
|
|
|
|
; 17 : dot_product = _mm_add_ps(dot_product,
|
|
|
|
0009e 0f 28 85 a0 00
|
|
00 00 movaps xmm0, XMMWORD PTR dot_product$[rbp]
|
|
000a5 0f c6 85 a0 00
|
|
00 00 b1 shufps xmm0, XMMWORD PTR dot_product$[rbp], 177 ; 000000b1H
|
|
000ad 0f 28 8d a0 00
|
|
00 00 movaps xmm1, XMMWORD PTR dot_product$[rbp]
|
|
000b4 0f 58 c8 addps xmm1, xmm0
|
|
000b7 0f 28 c1 movaps xmm0, xmm1
|
|
000ba 0f 29 85 a0 00
|
|
00 00 movaps XMMWORD PTR dot_product$[rbp], xmm0
|
|
|
|
; 18 : _mm_shuffle_ps(dot_product, dot_product, _MM_SHUFFLE(2, 3, 0, 1)));
|
|
; 19 : dot_product = _mm_add_ss(dot_product,
|
|
|
|
000c1 0f 28 85 a0 00
|
|
00 00 movaps xmm0, XMMWORD PTR dot_product$[rbp]
|
|
000c8 0f c6 85 a0 00
|
|
00 00 0e shufps xmm0, XMMWORD PTR dot_product$[rbp], 14
|
|
000d0 0f 28 8d a0 00
|
|
00 00 movaps xmm1, XMMWORD PTR dot_product$[rbp]
|
|
000d7 f3 0f 58 c8 addss xmm1, xmm0
|
|
000db 0f 29 8d a0 00
|
|
00 00 movaps XMMWORD PTR dot_product$[rbp], xmm1
|
|
|
|
; 20 : _mm_shuffle_ps(dot_product, dot_product, _MM_SHUFFLE(0, 0, 3, 2)));
|
|
; 21 :
|
|
; 22 : __m128 b_orthogonal = _mm_sub_ps(target->f32_x[0], _mm_mul_ps(s, dot_product));
|
|
|
|
000e2 0f 28 45 10 movaps xmm0, XMMWORD PTR s$[rbp]
|
|
000e6 0f 59 85 a0 00
|
|
00 00 mulps xmm0, XMMWORD PTR dot_product$[rbp]
|
|
000ed b8 10 00 00 00 mov eax, 16
|
|
000f2 48 6b c0 00 imul rax, rax, 0
|
|
000f6 48 8b 8d 38 02
|
|
00 00 mov rcx, QWORD PTR target$[rbp]
|
|
000fd 0f 10 0c 01 movups xmm1, XMMWORD PTR [rcx+rax]
|
|
00101 0f 5c c8 subps xmm1, xmm0
|
|
00104 0f 28 c1 movaps xmm0, xmm1
|
|
00107 0f 29 85 d0 00
|
|
00 00 movaps XMMWORD PTR b_orthogonal$[rbp], xmm0
|
|
|
|
; 23 :
|
|
; 24 : __m128 b_length_sq = _mm_mul_ps(b_orthogonal, b_orthogonal);
|
|
|
|
0010e 0f 28 85 d0 00
|
|
00 00 movaps xmm0, XMMWORD PTR b_orthogonal$[rbp]
|
|
00115 0f 59 85 d0 00
|
|
00 00 mulps xmm0, XMMWORD PTR b_orthogonal$[rbp]
|
|
0011c 0f 29 85 00 01
|
|
00 00 movaps XMMWORD PTR b_length_sq$[rbp], xmm0
|
|
|
|
; 25 : b_length_sq = _mm_add_ps(b_length_sq,
|
|
|
|
00123 0f 28 85 00 01
|
|
00 00 movaps xmm0, XMMWORD PTR b_length_sq$[rbp]
|
|
0012a 0f c6 85 00 01
|
|
00 00 b1 shufps xmm0, XMMWORD PTR b_length_sq$[rbp], 177 ; 000000b1H
|
|
00132 0f 28 8d 00 01
|
|
00 00 movaps xmm1, XMMWORD PTR b_length_sq$[rbp]
|
|
00139 0f 58 c8 addps xmm1, xmm0
|
|
0013c 0f 28 c1 movaps xmm0, xmm1
|
|
0013f 0f 29 85 00 01
|
|
00 00 movaps XMMWORD PTR b_length_sq$[rbp], xmm0
|
|
|
|
; 26 : _mm_shuffle_ps(b_length_sq, b_length_sq, _MM_SHUFFLE(2, 3, 0, 1)));
|
|
; 27 : b_length_sq = _mm_add_ss(b_length_sq,
|
|
|
|
00146 0f 28 85 00 01
|
|
00 00 movaps xmm0, XMMWORD PTR b_length_sq$[rbp]
|
|
0014d 0f c6 85 00 01
|
|
00 00 0e shufps xmm0, XMMWORD PTR b_length_sq$[rbp], 14
|
|
00155 0f 28 8d 00 01
|
|
00 00 movaps xmm1, XMMWORD PTR b_length_sq$[rbp]
|
|
0015c f3 0f 58 c8 addss xmm1, xmm0
|
|
00160 0f 29 8d 00 01
|
|
00 00 movaps XMMWORD PTR b_length_sq$[rbp], xmm1
|
|
|
|
; 28 : _mm_shuffle_ps(b_length_sq, b_length_sq, _MM_SHUFFLE(0, 0, 3, 2)));
|
|
; 29 :
|
|
; 30 : __m128 b_inv_length = _mm_rsqrt_ps(b_length_sq);
|
|
|
|
00167 0f 52 85 00 01
|
|
00 00 rsqrtps xmm0, XMMWORD PTR b_length_sq$[rbp]
|
|
0016e 0f 29 85 30 01
|
|
00 00 movaps XMMWORD PTR b_inv_length$[rbp], xmm0
|
|
|
|
; 31 : target->f32_x[0] = _mm_mul_ps(b_orthogonal, b_inv_length);
|
|
|
|
00175 0f 28 85 d0 00
|
|
00 00 movaps xmm0, XMMWORD PTR b_orthogonal$[rbp]
|
|
0017c 0f 59 85 30 01
|
|
00 00 mulps xmm0, XMMWORD PTR b_inv_length$[rbp]
|
|
00183 b8 10 00 00 00 mov eax, 16
|
|
00188 48 6b c0 00 imul rax, rax, 0
|
|
0018c 48 8b 8d 38 02
|
|
00 00 mov rcx, QWORD PTR target$[rbp]
|
|
00193 0f 11 04 01 movups XMMWORD PTR [rcx+rax], xmm0
|
|
|
|
; 32 : }
|
|
|
|
00197 48 8d a5 18 02
|
|
00 00 lea rsp, QWORD PTR [rbp+536]
|
|
0019e 5f pop rdi
|
|
0019f 5d pop rbp
|
|
001a0 c3 ret 0
|
|
f32_normalize ENDP
|
|
_TEXT ENDS
|
|
END
|