VPCONFLICTQ 99(R15)(R15*8), K4, Y24                // 6202fd2cc484ff63000000
	VPCONFLICTQ 7(AX)(CX*8), K4, Y24                   // 6262fd2cc484c807000000
	VPCONFLICTQ Y5, K4, Y1                             // 62f2fd2cc4cd
	VPCONFLICTQ Y18, K4, Y1                            // 62b2fd2cc4ca
	VPCONFLICTQ Y20, K4, Y1                            // 62b2fd2cc4cc
	VPCONFLICTQ 99(R15)(R15*8), K4, Y1                 // 6292fd2cc48cff63000000

	V4FNMADDPS 99(R15)(R15*1), [Z1-Z4], K3, Z15        // 6212774baabc3f63000000
	V4FNMADDPS (DX), [Z1-Z4], K3, Z15                  // 6272774baa3a
	V4FNMADDPS 99(R15)(R15*1), [Z11-Z14], K3, Z15      // 6212274baabc3f63000000
	V4FNMADDPS (DX), [Z11-Z14], K3, Z15                // 6272274baa3a
	V4FNMADDPS 99(R15)(R15*1), [Z21-Z24], K3, Z15      // 62125743aabc3f63000000

	VPCOMPRESSW Z3, K7, -15(R14)(R15*2)                // 6292fd4f639c7ef1ffffff
	VPCOMPRESSW Z0, K7, -15(R14)(R15*2)                // 6292fd4f63847ef1ffffff
	VPEXPANDB X16, K1, X6                              // 62b27d0962f0
	VPEXPANDB X28, K1, X6                              // 62927d0962f4
	VPEXPANDB X8, K1, X6                               // 62d27d0962f0
	VPEXPANDB 99(R15)(R15*4), K1, X6                   // 62927d096274bf63

	VGATHERDPD 360(R15)(Y30*2), K6, Z20 // 6282fd469264772d
	VGATHERDPD 640(R15)(Y20*2), K6, Z10 // 6252fd4692546750
	VGATHERDPD 960(R15)(Y10*2), K6, Z20 // 6282fd4e92645778
	VGATHERDPD 1280(R15)(Y0*2), K6, Z10 // 6252fd4e92944700050000
	VGATHERDPS 360(R15)(X30*2), K6, X20 // 62827d069264775a
	VGATHERDPS 640(R15)(X20*2), K6, X10 // 62527d0692946780020000
	VGATHERDPS 960(R15)(X10*2), K6, X20 // 62827d0e92a457c0030000

	VGF2P8AFFINEINVQB $47, 99(R15)(R15*1), Z14, K2, Z28 // 62038d4acfa43f630000002f
	VGF2P8AFFINEINVQB $47, (DX), Z14, K2, Z28          // 62638d4acf222f
	VGF2P8AFFINEINVQB $47, Z3, Z28, K2, Z28            // 62639d42cfe32f
	VGF2P8AFFINEINVQB $47, Z12, Z28, K2, Z28           // 62439d42cfe42f
	VGF2P8AFFINEINVQB $47, 99(R15)(R15*1), Z28, K2, Z28 // 62039d42cfa43f630000002f

	VEXP2PS Z9, K1, Z1                                 // 62d27d49c8c9
	VEXP2PS 99(R15)(R15*4), K1, Z1                     // 62927d49c88cbf63000000
	VEXP2PS 15(DX), K1, Z1                             // 62f27d49c88a0f000000
	VEXP2PS Z20, K1, Z9                                // 62327d49c8cc
	VEXP2PS Z9, K1, Z9                                 // 62527d49c8c9
	VEXP2PS 99(R15)(R15*4), K1, Z9                     // 62127d49c88cbf63000000

	{"R11", "R11"},
	{"R12", "R12"},
	{"R13", "R13"},
	{"R14", "R14"},
	{"R15", "R15"},
	{"R1<<2(R3)", "R1<<2(R3)"},
	{"R(1)<<2(R(3))", "R1<<2(R3)"},
	{"R2", "R2"},
	{"R3", "R3"},
	{"R4", "R4"},
	{"R(4)", "R4"},
	{"R5", "R5"},
	{"R6", "R6"},
	{"R7", "R7"},
	{"R8", "R8"},
	{"[R0,R1,g,R15]", "[R0,R1,g,R15]"},
	{"[R0-R7]", "[R0,R1,R2,R3,R4,R5,R6,R7]"},
	{"[R(0)-R(7)]", "[R0,R1,R2,R3,R4,R5,R6,R7]"},

	CLREX $0                                   // 5f3003d5
	CLSW R15, R6                               // e615c05a
	CLS R15, ZR                                // ff15c0da
	CLZW R1, R14                               // 2e10c05a
	CLZ R21, R9                                // a912c0da
	CMNW R21.UXTB<<4, R15                      // ff11352b
	CMN R0.UXTW<<4, R16                        // 1f5220ab

	VPDPBUSD -15(R14)(R15*1), X16, K2, X6              // 62927d0250b43ef1ffffff
	VPDPBUSD -15(BX), X16, K2, X6                      // 62f27d0250b3f1ffffff
	VPDPBUSD X15, X28, K2, X6                          // 62d21d0250f7
	VPDPBUSD X11, X28, K2, X6                          // 62d21d0250f3
	VPDPBUSD X1, X28, K2, X6                           // 62f21d0250f1
	VPDPBUSD -15(R14)(R15*1), X28, K2, X6              // 62921d0250b43ef1ffffff

ADDQ R8, R9 ADCQ R15, R10 ADCQ AX, R11 ADCQ DX, R12 XORQ R13, R13 // Second stage MOVQ R9, AX MOVQ R9, R15 SHLQ $0x20, R9 MULQ p256const1<>+0(SB) SHRQ $0x20, R15 ADDQ R9, R10 ADCQ R15, R11 ADCQ AX, R12 ADCQ DX, R13 XORQ R8, R8 // Third stage MOVQ R10, AX MOVQ R10, R15 SHLQ $0x20, R10 MULQ p256const1<>+0(SB) SHRQ $0x20, R15 ADDQ R10, R11 ADCQ R15, R12 ADCQ AX, R13 ADCQ DX, R8 XORQ R9, R9 // Last stage MOVQ R11, AX MOVQ R11, R15 SHLQ $0x20, R11 MULQ p256const1<>+0(SB) SHRQ $0x20, R15 ADDQ R11, R12 ADCQ...