mirror of
https://github.com/go-gitea/gitea
synced 2024-11-15 06:34:25 +00:00
268 lines
55 KiB
ArmAsm
268 lines
55 KiB
ArmAsm
|
//+build !noasm,!appengine
|
||
|
|
|
||
|
|
TEXT ·sha256X16Avx512(SB), 7, $0
|
||
|
|
MOVQ digests+0(FP), DI
|
||
|
|
MOVQ scratch+8(FP), R12
|
||
|
|
MOVQ mask_len+32(FP), SI
|
||
|
|
MOVQ mask_base+24(FP), R13
|
||
|
|
MOVQ (R13), R14
|
||
|
|
LONG $0x92fbc1c4; BYTE $0xce
|
||
|
|
LEAQ inputs+48(FP), AX
|
||
|
|
QUAD $0xf162076f487ef162; QUAD $0x7ef162014f6f487e; QUAD $0x487ef16202576f48; QUAD $0x6f487ef162035f6f; QUAD $0x6f6f487ef1620467; QUAD $0x06776f487ef16205; LONG $0x487ef162; WORD $0x7f6f; BYTE $0x07
|
||
|
|
MOVQ table+16(FP), DX
|
||
|
|
WORD $0x3148; BYTE $0xc9
|
||
|
|
TESTQ $(1<<0), R14
|
||
|
|
JE skipInput0
|
||
|
|
MOVQ 0*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x0410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput0:
|
||
|
|
TESTQ $(1<<1), R14
|
||
|
|
JE skipInput1
|
||
|
|
MOVQ 1*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x0c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput1:
|
||
|
|
TESTQ $(1<<2), R14
|
||
|
|
JE skipInput2
|
||
|
|
MOVQ 2*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x1410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput2:
|
||
|
|
TESTQ $(1<<3), R14
|
||
|
|
JE skipInput3
|
||
|
|
MOVQ 3*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x1c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput3:
|
||
|
|
TESTQ $(1<<4), R14
|
||
|
|
JE skipInput4
|
||
|
|
MOVQ 4*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x2410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput4:
|
||
|
|
TESTQ $(1<<5), R14
|
||
|
|
JE skipInput5
|
||
|
|
MOVQ 5*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x2c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput5:
|
||
|
|
TESTQ $(1<<6), R14
|
||
|
|
JE skipInput6
|
||
|
|
MOVQ 6*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x3410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput6:
|
||
|
|
TESTQ $(1<<7), R14
|
||
|
|
JE skipInput7
|
||
|
|
MOVQ 7*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x3c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput7:
|
||
|
|
TESTQ $(1<<8), R14
|
||
|
|
JE skipInput8
|
||
|
|
MOVQ 8*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x0410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput8:
|
||
|
|
TESTQ $(1<<9), R14
|
||
|
|
JE skipInput9
|
||
|
|
MOVQ 9*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x0c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput9:
|
||
|
|
TESTQ $(1<<10), R14
|
||
|
|
JE skipInput10
|
||
|
|
MOVQ 10*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x1410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput10:
|
||
|
|
TESTQ $(1<<11), R14
|
||
|
|
JE skipInput11
|
||
|
|
MOVQ 11*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x1c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput11:
|
||
|
|
TESTQ $(1<<12), R14
|
||
|
|
JE skipInput12
|
||
|
|
MOVQ 12*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x2410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput12:
|
||
|
|
TESTQ $(1<<13), R14
|
||
|
|
JE skipInput13
|
||
|
|
MOVQ 13*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x2c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput13:
|
||
|
|
TESTQ $(1<<14), R14
|
||
|
|
JE skipInput14
|
||
|
|
MOVQ 14*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x3410; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput14:
|
||
|
|
TESTQ $(1<<15), R14
|
||
|
|
JE skipInput15
|
||
|
|
MOVQ 15*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x3c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipInput15:
|
||
|
|
lloop:
|
||
|
|
LEAQ PSHUFFLE_BYTE_FLIP_MASK<>(SB), DX
|
||
|
|
LONG $0x487e7162; WORD $0x1a6f
|
||
|
|
MOVQ table+16(FP), DX
|
||
|
|
QUAD $0xd162226f487e7162; QUAD $0x7ed16224047f487e; QUAD $0x7ed16201244c7f48; QUAD $0x7ed1620224547f48; QUAD $0x7ed16203245c7f48; QUAD $0x7ed1620424647f48; QUAD $0x7ed16205246c7f48; QUAD $0x7ed1620624747f48; QUAD $0xc1834807247c7f48; QUAD $0x44c9c6407c316240; QUAD $0x62eec1c6407ca162; QUAD $0xa16244d3c6406c31; QUAD $0x34c162eed3c6406c; QUAD $0x407ca162dddac648; QUAD $0xc6407ca16288cac6; QUAD $0xcac648345162ddc2; QUAD $0x44d5c6405ca16288; QUAD $0x62eee5c6405ca162; QUAD $0xa16244d7c6404c31; QUAD $0x6cc162eef7c6404c; QUAD $0x405ca162ddfac640; QUAD $0xc6405ca16288eec6; QUAD $0xd2c6406cc162dde6; QUAD $0x44f1c6403c816288; QUAD $0x62eec1c6403c0162; QUAD $0x016244d3c6402c11; QUAD $0x4c4162eed3c6402c; QUAD $0x403c0162dddac640; QUAD $0xc6403c016288cac6; QUAD $0xf2c6404cc162ddc2; QUAD $0x44d5c6401c016288; QUAD $0x62eee5c6401c0162; QUAD $0x016244d7c6400c11; QUAD $0x2c4162eef7c6400c; QUAD $0x401c0162ddfac640; QUAD $0xc6401c016288eec6; QUAD $0xd2c6402c4162dde6; BYTE $0x88
|
||
|
|
LEAQ PSHUFFLE_TRANSPOSE16_MASK1<>(SB), BX
|
||
|
|
LEAQ PSHUFFLE_TRANSPOSE16_MASK2<>(SB), R8
|
||
|
|
QUAD $0x2262336f487e6162; QUAD $0x487e5162f27648b5; QUAD $0xd27648b53262106f; QUAD $0xa262136f487ee162; QUAD $0x487e5162d77640e5; QUAD $0xcf7640e53262086f; QUAD $0xa2621b6f487ee162; QUAD $0x487ec162dd7640f5; QUAD $0xfd7640f5a262386f; QUAD $0xa2620b6f487ee162; QUAD $0x487ec162cc7640fd; QUAD $0xec7640fda262286f; QUAD $0x8262036f487ee162; QUAD $0x487ec162c27640cd; QUAD $0xe27640cd8262206f; QUAD $0x8262336f487ee162; QUAD $0x487e4162f77640a5; QUAD $0xd77640a50262106f; QUAD $0x02621b6f487e6162; QUAD $0x487e4162dd7640b5; QUAD $0xfd7640b50262386f; QUAD $0x02620b6f487e6162; QUAD $0x487e4162cc7640bd; QUAD $0xec7640bd0262286f; QUAD $0x62eec023408d2362; QUAD $0x236244c023408da3; QUAD $0xada362eee42348ad; QUAD $0x40c5036244e42348; QUAD $0x2340c51362eef723; QUAD $0xfd2340d5036244d7; QUAD $0x44fd2340d58362ee; QUAD $0x62eeea2348b50362; QUAD $0x036244ea2348b583; QUAD $0xe51362eed32340e5; QUAD $0x40f5036244cb2340; QUAD $0x2340f58362eed923; QUAD $0xce2340ed236244d9; QUAD $0x44ce2340eda362ee; QUAD $0xc162d16f487ec162; QUAD $0x407dc262f26f487e; QUAD $0xcb004075c262c300; QUAD $0xc262d300406dc262; QUAD $0x405dc262db004065; QUAD $0xeb004055c262e300; QUAD $0xc262f300404dc262; QUAD $0x403d4262fb004045; QUAD $0xcb0040354262c300; QUAD $0x4262d300402d4262; QUAD $0x401d4262db004025; QUAD $0xeb0040154262e300; QUAD $0x4262f300400d4262; QUAD $0x48455162fb004005; QUAD $0xcc6f487e7162c4fe; QUAD $0x6206c472482df162; QUAD $0xf1620bc4724825f1; QUAD $0x55736219c472481d; QUAD $0x483d3162cace2548; QUAD $0xd42548255362c0fe; QUAD $0x62c1fe483d516296; QUAD $0x65d162c2fe483d51; QUAD $0x724845f162d8fe48; QUAD $0xc0724825f16202c0; QUAD $0x16c072481df1620d; QUAD $0x7362c86f487e7162; QUAD $0x25d362e8ca254875; QUAD $0x4845d16296fc2548; QUAD $0xf8fe4845d162f9fe; QUAD $0x6201626f487e7162; QUAD $0x916211c672481591; QUAD $0x05916213c672480d; QUAD $0x480d53620ad67248; QUAD $0xfe407dc16296ef25; QUAD $0x62c1fe407d8162c5; QUAD $0xb16207c1724815b1; QUAD $0x05b16212c172480d; QUAD $0x480d536203d17248; QUAD $0xfe407dc16296ef25; QUAD $0x62c4fe484d5162c5; QUAD $0x2df162cb6f487e71; QUAD $0x4825f16206c37248; QUAD $0x72481df1620bc372; QUAD $0xcd25485d736219c3; QUAD $0x62c1fe483d3162ca; QUAD $0x516296d425482553; QUAD $0x483d5162c1fe483d; QUAD $0xd0fe486dd162c2fe; QUAD $0x6202c772484df162; QUAD $0xf1620dc7724825f1; QUAD $0x7e716216c772481d; QUAD $0x25487d7362cf6f48; QUAD $0xf4254825d362e8c9; QUAD $0x62f1fe484dd16296; QUAD $0x7e7162f0fe484dd1; QUAD $0x4815916202626f48; QUAD $0x72480d916211c772; QUAD $0xd7724805916213c7; QUAD $0x96ef25480d53620a; QUAD $0x8162cdfe4075c162; QUAD $0x4815b162cafe4075; QUAD $0x72480db16207c272; QUAD $0xd2724805b16212c2; QUAD $0x96ef25480d536203; QUAD $0x5162cdfe4075c162; QUAD $0x487e7162c4fe4855; QUAD $0xc272482df162ca6f; QUAD $0x0bc2724825f16206; QUAD $0x6219c272481df162; QUAD $0x3162cacc25486573; QUAD $0x48255362c2fe483d; QUAD $0xfe483d516296d425; QUAD $0x62c2fe483d5162c1; QUAD $0x55f162c8fe4875d1; QUAD $0x4825f16202c67248; QUAD $0x72481df1620dc672; QUAD $0xce6f487e716216c6; QUAD $0x62e8c82548457362; QUAD $0xd16296ec254825d3; QUAD $0x4855d162e9fe4855; QUAD $0x626f487e7162e8fe; QUAD $0x11c0724815b16203; QUAD $0x6213c072480db162; QUAD $0x53620ad0724805b1; QUAD $0x6dc16296ef25480d; QUAD $0xfe406d8162d5fe40; QUAD $0x07c3724815b162d3; QUAD $0x6212c372480db162; QUAD $0x536203d3724805b1; QUAD $0x6dc16296ef25480d; QUAD $0xfe485d5162d5fe40; QUAD $0x62c96f487e7162c4; QUAD $0xf16206c172482df1; QUAD $0x1df1620bc1724825; QUAD $0x486d736219c17248; QUAD $0xfe483d3162cacb25; QUAD $0x96d42548255362c3; QUAD $0x5162c1fe483d5162; QUAD $0x487dd162c2fe483d; QUAD $0xc572485df162c0fe; QUAD $0x0dc5724825f16202; QUAD $0x6216c572481df162; QUAD $0x4d7362cd6f487e71; QUAD $0x4825d362e8cf2548; QUAD $0xfe485dd16296e425; QUAD $0x62e0fe485dd162e1; QUAD $0xb16204626f487e71; QUAD $0x0db16211c1724815; QUAD $0x4805b16213c17248; QUAD $0x25480d53620ad172; QUAD $0xddfe4065c16296ef; QUAD $0xb162dcfe40658162; QUAD $0x0db16207c4724815; QUAD $0x4805b16212c47248; QUAD $0x25480d536203d472; QUAD $0xddfe4065c16296ef; QUAD $0x7162c4fe48655162; QUAD $0x482df162c86f487e; QUAD $0x724825f16206c0
|
||
|
|
JE lastLoop
|
||
|
|
ADDQ $8, R13
|
||
|
|
MOVQ (R13), R14
|
||
|
|
QUAD $0x7162c4fe48455162; QUAD $0x482df162cc6f487e; QUAD $0x724825f16206c472; QUAD $0xc472481df1620bc4; QUAD $0xcace254855736219; QUAD $0x5362c0fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62d8fe4865d162c2; QUAD $0xf16202c0724845f1; QUAD $0x1df1620dc0724825; QUAD $0x487e716216c07248; QUAD $0xca2548757362c86f; QUAD $0x96fc254825d362e8; QUAD $0xd162f9fe4845d162; QUAD $0x487e7162f8fe4845; WORD $0x626f; BYTE $0x31
|
||
|
|
TESTQ $(1<<0), R14
|
||
|
|
JE skipNext0
|
||
|
|
MOVQ 0*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x0410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext0:
|
||
|
|
QUAD $0x7162c4fe484d5162; QUAD $0x482df162cb6f487e; QUAD $0x724825f16206c372; QUAD $0xc372481df1620bc3; QUAD $0xcacd25485d736219; QUAD $0x5362c1fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62d0fe486dd162c2; QUAD $0xf16202c772484df1; QUAD $0x1df1620dc7724825; QUAD $0x487e716216c77248; QUAD $0xc925487d7362cf6f; QUAD $0x96f4254825d362e8; QUAD $0xd162f1fe484dd162; QUAD $0x487e7162f0fe484d; WORD $0x626f; BYTE $0x32
|
||
|
|
TESTQ $(1<<1), R14
|
||
|
|
JE skipNext1
|
||
|
|
MOVQ 1*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x0c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext1:
|
||
|
|
QUAD $0x7162c4fe48555162; QUAD $0x482df162ca6f487e; QUAD $0x724825f16206c272; QUAD $0xc272481df1620bc2; QUAD $0xcacc254865736219; QUAD $0x5362c2fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62c8fe4875d162c2; QUAD $0xf16202c6724855f1; QUAD $0x1df1620dc6724825; QUAD $0x487e716216c67248; QUAD $0xc82548457362ce6f; QUAD $0x96ec254825d362e8; QUAD $0xd162e9fe4855d162; QUAD $0x487e7162e8fe4855; WORD $0x626f; BYTE $0x33
|
||
|
|
TESTQ $(1<<2), R14
|
||
|
|
JE skipNext2
|
||
|
|
MOVQ 2*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x1410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext2:
|
||
|
|
QUAD $0x7162c4fe485d5162; QUAD $0x482df162c96f487e; QUAD $0x724825f16206c172; QUAD $0xc172481df1620bc1; QUAD $0xcacb25486d736219; QUAD $0x5362c3fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62c0fe487dd162c2; QUAD $0xf16202c572485df1; QUAD $0x1df1620dc5724825; QUAD $0x487e716216c57248; QUAD $0xcf25484d7362cd6f; QUAD $0x96e4254825d362e8; QUAD $0xd162e1fe485dd162; QUAD $0x487e7162e0fe485d; WORD $0x626f; BYTE $0x34
|
||
|
|
TESTQ $(1<<3), R14
|
||
|
|
JE skipNext3
|
||
|
|
MOVQ 3*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x1c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext3:
|
||
|
|
QUAD $0x7162c4fe48655162; QUAD $0x482df162c86f487e; QUAD $0x724825f16206c072; QUAD $0xc072481df1620bc0; QUAD $0xcaca254875736219; QUAD $0x5362c4fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62f8fe4845d162c2; QUAD $0xf16202c4724865f1; QUAD $0x1df1620dc4724825; QUAD $0x487e716216c47248; QUAD $0xce2548557362cc6f; QUAD $0x96dc254825d362e8; QUAD $0xd162d9fe4865d162; QUAD $0x487e7162d8fe4865; WORD $0x626f; BYTE $0x35
|
||
|
|
TESTQ $(1<<4), R14
|
||
|
|
JE skipNext4
|
||
|
|
MOVQ 4*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x2410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext4:
|
||
|
|
QUAD $0x7162c4fe486d5162; QUAD $0x482df162cf6f487e; QUAD $0x724825f16206c772; QUAD $0xc772481df1620bc7; QUAD $0xcac925487d736219; QUAD $0x5362c5fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62f0fe484dd162c2; QUAD $0xf16202c372486df1; QUAD $0x1df1620dc3724825; QUAD $0x487e716216c37248; QUAD $0xcd25485d7362cb6f; QUAD $0x96d4254825d362e8; QUAD $0xd162d1fe486dd162; QUAD $0x487e7162d0fe486d; WORD $0x626f; BYTE $0x36
|
||
|
|
TESTQ $(1<<5), R14
|
||
|
|
JE skipNext5
|
||
|
|
MOVQ 5*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x2c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext5:
|
||
|
|
QUAD $0x7162c4fe48755162; QUAD $0x482df162ce6f487e; QUAD $0x724825f16206c672; QUAD $0xc672481df1620bc6; QUAD $0xcac8254845736219; QUAD $0x5362c6fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62e8fe4855d162c2; QUAD $0xf16202c2724875f1; QUAD $0x1df1620dc2724825; QUAD $0x487e716216c27248; QUAD $0xcc2548657362ca6f; QUAD $0x96cc254825d362e8; QUAD $0xd162c9fe4875d162; QUAD $0x487e7162c8fe4875; WORD $0x626f; BYTE $0x37
|
||
|
|
TESTQ $(1<<6), R14
|
||
|
|
JE skipNext6
|
||
|
|
MOVQ 6*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x3410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext6:
|
||
|
|
QUAD $0x7162c4fe487d5162; QUAD $0x482df162cd6f487e; QUAD $0x724825f16206c572; QUAD $0xc572481df1620bc5; QUAD $0xcacf25484d736219; QUAD $0x5362c7fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62e0fe485dd162c2; QUAD $0xf16202c172487df1; QUAD $0x1df1620dc1724825; QUAD $0x487e716216c17248; QUAD $0xcb25486d7362c96f; QUAD $0x96c4254825d362e8; QUAD $0xd162c1fe487dd162; QUAD $0x487e7162c0fe487d; WORD $0x626f; BYTE $0x38
|
||
|
|
TESTQ $(1<<7), R14
|
||
|
|
JE skipNext7
|
||
|
|
MOVQ 7*24(AX), R9
|
||
|
|
LONG $0x487cc162; WORD $0x3c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext7:
|
||
|
|
QUAD $0x7162c4fe48455162; QUAD $0x482df162cc6f487e; QUAD $0x724825f16206c472; QUAD $0xc472481df1620bc4; QUAD $0xcace254855736219; QUAD $0x5362c0fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62d8fe4865d162c2; QUAD $0xf16202c0724845f1; QUAD $0x1df1620dc0724825; QUAD $0x487e716216c07248; QUAD $0xca2548757362c86f; QUAD $0x96fc254825d362e8; QUAD $0xd162f9fe4845d162; QUAD $0x487e7162f8fe4845; WORD $0x626f; BYTE $0x39
|
||
|
|
TESTQ $(1<<8), R14
|
||
|
|
JE skipNext8
|
||
|
|
MOVQ 8*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x0410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext8:
|
||
|
|
QUAD $0x7162c4fe484d5162; QUAD $0x482df162cb6f487e; QUAD $0x724825f16206c372; QUAD $0xc372481df1620bc3; QUAD $0xcacd25485d736219; QUAD $0x5362c1fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62d0fe486dd162c2; QUAD $0xf16202c772484df1; QUAD $0x1df1620dc7724825; QUAD $0x487e716216c77248; QUAD $0xc925487d7362cf6f; QUAD $0x96f4254825d362e8; QUAD $0xd162f1fe484dd162; QUAD $0x487e7162f0fe484d; WORD $0x626f; BYTE $0x3a
|
||
|
|
TESTQ $(1<<9), R14
|
||
|
|
JE skipNext9
|
||
|
|
MOVQ 9*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x0c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext9:
|
||
|
|
QUAD $0x7162c4fe48555162; QUAD $0x482df162ca6f487e; QUAD $0x724825f16206c272; QUAD $0xc272481df1620bc2; QUAD $0xcacc254865736219; QUAD $0x5362c2fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62c8fe4875d162c2; QUAD $0xf16202c6724855f1; QUAD $0x1df1620dc6724825; QUAD $0x487e716216c67248; QUAD $0xc82548457362ce6f; QUAD $0x96ec254825d362e8; QUAD $0xd162e9fe4855d162; QUAD $0x487e7162e8fe4855; WORD $0x626f; BYTE $0x3b
|
||
|
|
TESTQ $(1<<10), R14
|
||
|
|
JE skipNext10
|
||
|
|
MOVQ 10*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x1410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext10:
|
||
|
|
QUAD $0x7162c4fe485d5162; QUAD $0x482df162c96f487e; QUAD $0x724825f16206c172; QUAD $0xc172481df1620bc1; QUAD $0xcacb25486d736219; QUAD $0x5362c3fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62c0fe487dd162c2; QUAD $0xf16202c572485df1; QUAD $0x1df1620dc5724825; QUAD $0x487e716216c57248; QUAD $0xcf25484d7362cd6f; QUAD $0x96e4254825d362e8; QUAD $0xd162e1fe485dd162; QUAD $0x487e7162e0fe485d; WORD $0x626f; BYTE $0x3c
|
||
|
|
TESTQ $(1<<11), R14
|
||
|
|
JE skipNext11
|
||
|
|
MOVQ 11*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x1c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext11:
|
||
|
|
QUAD $0x7162c4fe48655162; QUAD $0x482df162c86f487e; QUAD $0x724825f16206c072; QUAD $0xc072481df1620bc0; QUAD $0xcaca254875736219; QUAD $0x5362c4fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62f8fe4845d162c2; QUAD $0xf16202c4724865f1; QUAD $0x1df1620dc4724825; QUAD $0x487e716216c47248; QUAD $0xce2548557362cc6f; QUAD $0x96dc254825d362e8; QUAD $0xd162d9fe4865d162; QUAD $0x487e7162d8fe4865; WORD $0x626f; BYTE $0x3d
|
||
|
|
TESTQ $(1<<12), R14
|
||
|
|
JE skipNext12
|
||
|
|
MOVQ 12*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x2410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext12:
|
||
|
|
QUAD $0x7162c4fe486d5162; QUAD $0x482df162cf6f487e; QUAD $0x724825f16206c772; QUAD $0xc772481df1620bc7; QUAD $0xcac925487d736219; QUAD $0x5362c5fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62f0fe484dd162c2; QUAD $0xf16202c372486df1; QUAD $0x1df1620dc3724825; QUAD $0x487e716216c37248; QUAD $0xcd25485d7362cb6f; QUAD $0x96d4254825d362e8; QUAD $0xd162d1fe486dd162; QUAD $0x487e7162d0fe486d; WORD $0x626f; BYTE $0x3e
|
||
|
|
TESTQ $(1<<13), R14
|
||
|
|
JE skipNext13
|
||
|
|
MOVQ 13*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x2c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext13:
|
||
|
|
QUAD $0x7162c4fe48755162; QUAD $0x482df162ce6f487e; QUAD $0x724825f16206c672; QUAD $0xc672481df1620bc6; QUAD $0xcac8254845736219; QUAD $0x5362c6fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62e8fe4855d162c2; QUAD $0xf16202c2724875f1; QUAD $0x1df1620dc2724825; QUAD $0x487e716216c27248; QUAD $0xcc2548657362ca6f; QUAD $0x96cc254825d362e8; QUAD $0xd162c9fe4875d162; QUAD $0x487e7162c8fe4875; WORD $0x626f; BYTE $0x3f
|
||
|
|
TESTQ $(1<<14), R14
|
||
|
|
JE skipNext14
|
||
|
|
MOVQ 14*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x3410; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext14:
|
||
|
|
QUAD $0x7162c4fe487d5162; QUAD $0x482df162cd6f487e; QUAD $0x724825f16206c572; QUAD $0xc572481df1620bc5; QUAD $0xcacf25484d736219; QUAD $0x5362c7fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62e0fe485dd162c2; QUAD $0xf16202c172487df1; QUAD $0x1df1620dc1724825; QUAD $0x487e716216c17248; QUAD $0xcb25486d7362c96f; QUAD $0x96c4254825d362e8; QUAD $0xd162c1fe487dd162; QUAD $0x487e7162c0fe487d; WORD $0x626f; BYTE $0x40
|
||
|
|
TESTQ $(1<<15), R14
|
||
|
|
JE skipNext15
|
||
|
|
MOVQ 15*24(AX), R9
|
||
|
|
LONG $0x487c4162; WORD $0x3c10; BYTE $0x09
|
||
|
|
|
||
|
|
skipNext15:
|
||
|
|
QUAD $0xd162d86f487e7162; QUAD $0x7dd16224046f487e; QUAD $0x6f487e7162c3fe49; QUAD $0x244c6f487ed162d9; QUAD $0x62cbfe4975d16201; QUAD $0x7ed162da6f487e71; QUAD $0x6dd1620224546f48; QUAD $0x6f487e7162d3fe49; QUAD $0x245c6f487ed162db; QUAD $0x62dbfe4965d16203; QUAD $0x7ed162dc6f487e71; QUAD $0x5dd1620424646f48; QUAD $0x6f487e7162e3fe49; QUAD $0x246c6f487ed162dd; QUAD $0x62ebfe4955d16205; QUAD $0x7ed162de6f487e71; QUAD $0x4dd1620624746f48; QUAD $0x6f487e7162f3fe49; QUAD $0x247c6f487ed162df; QUAD $0xc4fbfe4945d16207; LONG $0xce92fbc1
|
||
|
|
JMP lloop
|
||
|
|
|
||
|
|
lastLoop:
|
||
|
|
QUAD $0x7162c4fe48455162; QUAD $0x482df162cc6f487e; QUAD $0x724825f16206c472; QUAD $0xc472481df1620bc4; QUAD $0xcace254855736219; QUAD $0x5362c0fe483d3162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62d8fe4865d162c2; QUAD $0xf16202c0724845f1; QUAD $0x1df1620dc0724825; QUAD $0x487e716216c07248; QUAD $0xca2548757362c86f; QUAD $0x96fc254825d362e8; QUAD $0xd162f9fe4845d162; QUAD $0x487e7162f8fe4845; QUAD $0xfe484d516231626f; QUAD $0x62cb6f487e7162c4; QUAD $0xf16206c372482df1; QUAD $0x1df1620bc3724825; QUAD $0x485d736219c37248; QUAD $0xfe483d3162cacd25; QUAD $0x96d42548255362c1; QUAD $0x5162c1fe483d5162; QUAD $0x486dd162c2fe483d; QUAD $0xc772484df162d0fe; QUAD $0x0dc7724825f16202; QUAD $0x6216c772481df162; QUAD $0x7d7362cf6f487e71; QUAD $0x4825d362e8c92548; QUAD $0xfe484dd16296f425; QUAD $0x62f0fe484dd162f1; QUAD $0x516232626f487e71; QUAD $0x487e7162c4fe4855; QUAD $0xc272482df162ca6f; QUAD $0x0bc2724825f16206; QUAD $0x6219c272481df162; QUAD $0x3162cacc25486573; QUAD $0x48255362c2fe483d; QUAD $0xfe483d516296d425; QUAD $0x62c2fe483d5162c1; QUAD $0x55f162c8fe4875d1; QUAD $0x4825f16202c67248; QUAD $0x72481df1620dc672; QUAD $0xce6f487e716216c6; QUAD $0x62e8c82548457362; QUAD $0xd16296ec254825d3; QUAD $0x4855d162e9fe4855; QUAD $0x626f487e7162e8fe; QUAD $0x62c4fe485d516233; QUAD $0x2df162c96f487e71; QUAD $0x4825f16206c17248; QUAD $0x72481df1620bc172; QUAD $0xcb25486d736219c1; QUAD $0x62c3fe483d3162ca; QUAD $0x516296d425482553; QUAD $0x483d5162c1fe483d; QUAD $0xc0fe487dd162c2fe; QUAD $0x6202c572485df162; QUAD $0xf1620dc5724825f1; QUAD $0x7e716216c572481d; QUAD $0x25484d7362cd6f48; QUAD $0xe4254825d362e8cf; QUAD $0x62e1fe485dd16296; QUAD $0x7e7162e0fe485dd1; QUAD $0x4865516234626f48; QUAD $0xc86f487e7162c4fe; QUAD $0x6206c072482df162; QUAD $0xf1620bc0724825f1; QUAD $0x75736219c072481d; QUAD $0x483d3162caca2548; QUAD $0xd42548255362c4fe; QUAD $0x62c1fe483d516296; QUAD $0x45d162c2fe483d51; QUAD $0x724865f162f8fe48; QUAD $0xc4724825f16202c4; QUAD $0x16c472481df1620d; QUAD $0x7362cc6f487e7162; QUAD $0x25d362e8ce254855; QUAD $0x4865d16296dc2548; QUAD $0xd8fe4865d162d9fe; QUAD $0x6235626f487e7162; QUAD $0x7e7162c4fe486d51; QUAD $0x72482df162cf6f48; QUAD $0xc7724825f16206c7; QUAD $0x19c772481df1620b; QUAD $0x62cac925487d7362; QUAD $0x255362c5fe483d31; QUAD $0x483d516296d42548; QUAD $0xc2fe483d5162c1fe; QUAD $0xf162f0fe484dd162; QUAD $0x25f16202c372486d; QUAD $0x481df1620dc37248; QUAD $0x6f487e716216c372; QUAD $0xe8cd25485d7362cb; QUAD $0x6296d4254825d362; QUAD $0x6dd162d1fe486dd1; QUAD $0x6f487e7162d0fe48; QUAD $0xc4fe487551623662; QUAD $0xf162ce6f487e7162; QUAD $0x25f16206c672482d; QUAD $0x481df1620bc67248; QUAD $0x254845736219c672; QUAD $0xc6fe483d3162cac8; QUAD $0x6296d42548255362; QUAD $0x3d5162c1fe483d51; QUAD $0xfe4855d162c2fe48; QUAD $0x02c2724875f162e8; QUAD $0x620dc2724825f162; QUAD $0x716216c272481df1; QUAD $0x48657362ca6f487e; QUAD $0x254825d362e8cc25; QUAD $0xc9fe4875d16296cc; QUAD $0x7162c8fe4875d162; QUAD $0x7d516237626f487e; QUAD $0x6f487e7162c4fe48; QUAD $0x06c572482df162cd; QUAD $0x620bc5724825f162; QUAD $0x736219c572481df1; QUAD $0x3d3162cacf25484d; QUAD $0x2548255362c7fe48; QUAD $0xc1fe483d516296d4; QUAD $0xd162c2fe483d5162; QUAD $0x487df162e0fe485d; QUAD $0x724825f16202c172; QUAD $0xc172481df1620dc1; QUAD $0x62c96f487e716216; QUAD $0xd362e8cb25486d73; QUAD $0x7dd16296c4254825; QUAD $0xfe487dd162c1fe48; QUAD $0x38626f487e7162c0; QUAD $0x7162c4fe48455162; QUAD $0x482df162cc6f487e; QUAD $0x724825f16206c472; QUAD $0xc472481df1620bc4; QUAD $0xcace254855736219; QUAD $0x5362c0fe483d1162; QUAD $0x3d516296d4254825; QUAD $0xfe483d5162c1fe48; QUAD $0x62d8fe4865d162c2; QUAD $0xf16202c0724845f1; QUAD $0x1df1620dc0724825; QUAD $0x487e716216c07248; QUAD $0xca2548757362c86f; QUAD $0x96fc254825d362e8; QUAD $0xd162f9fe4845d162; QUAD $0x487e7162f8fe4845; QUAD $0xfe484d516239626f; QUAD $0x62cb6f487e7162c4; QUAD $0xf16206c372482df1; QUAD $0x1df1620bc3724825; QUAD $0x485d736219c37248; QUAD $0xfe483d1162cacd25; QUAD $0x96d42548255362c1; QUAD $0x5162c1fe483d5162; QUAD $0x486dd162c2fe483d; QUAD $0xc772484df162d0fe; QUAD
|
||
|
|
VZEROUPPER
|
||
|
|
RET
|
||
|
|
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x000(SB)/8, $0x0405060700010203
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x008(SB)/8, $0x0c0d0e0f08090a0b
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x010(SB)/8, $0x0405060700010203
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x018(SB)/8, $0x0c0d0e0f08090a0b
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x020(SB)/8, $0x0405060700010203
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x028(SB)/8, $0x0c0d0e0f08090a0b
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x030(SB)/8, $0x0405060700010203
|
||
|
|
DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x038(SB)/8, $0x0c0d0e0f08090a0b
|
||
|
|
GLOBL PSHUFFLE_BYTE_FLIP_MASK<>(SB), 8, $64
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x000(SB)/8, $0x0000000000000000
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x008(SB)/8, $0x0000000000000001
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x010(SB)/8, $0x0000000000000008
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x018(SB)/8, $0x0000000000000009
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x020(SB)/8, $0x0000000000000004
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x028(SB)/8, $0x0000000000000005
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x030(SB)/8, $0x000000000000000C
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK1<>+0x038(SB)/8, $0x000000000000000D
|
||
|
|
GLOBL PSHUFFLE_TRANSPOSE16_MASK1<>(SB), 8, $64
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x000(SB)/8, $0x0000000000000002
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x008(SB)/8, $0x0000000000000003
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x010(SB)/8, $0x000000000000000A
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x018(SB)/8, $0x000000000000000B
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x020(SB)/8, $0x0000000000000006
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x028(SB)/8, $0x0000000000000007
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x030(SB)/8, $0x000000000000000E
|
||
|
|
DATA PSHUFFLE_TRANSPOSE16_MASK2<>+0x038(SB)/8, $0x000000000000000F
|
||
|
|
GLOBL PSHUFFLE_TRANSPOSE16_MASK2<>(SB), 8, $64
|