VPGATHERDQ Y0, 664(X14*8), Y6           // c4a2fd9034f598020000
	VPGATHERQQ X2, (BP)(X7*2), X1           // c4e2e9914c7d00
	VPGATHERQQ Y2, (BP)(Y7*2), Y1           // c4e2ed914c7d00
	VPGATHERQQ X12, (R13)(X14*2), X11       // c40299915c7500
	VPGATHERQQ Y12, (R13)(Y14*2), Y11       // c4029d915c7500
	VPGATHERQQ X2, (BP)(X7*2), X1           // c4e2e9914c7d00
	VPGATHERQQ Y2, (BP)(Y7*2), Y1           // c4e2ed914c7d00

	VAESDECLAST -17(BP)(SI*8), Y11, Y24                // 62622528df84f5efffffff or 6262a528df84f5efffffff
	VAESDECLAST (R15), Y11, Y24                        // 62422528df07 or 6242a528df07
	VAESDECLAST Z8, Z23, Z23                           // 62c24540dff8 or 62c2c540dff8
	VAESDECLAST Z28, Z23, Z23                          // 62824540dffc or 6282c540dffc

	{"$~15", "$-16"},
	{"(((8)&0xf)*4)(SP)", "32(SP)"},
	{"(((8-14)&0xf)*4)(SP)", "40(SP)"},
	{"(6+8)(AX)", "14(AX)"},
	{"(8*4)(BP)", "32(BP)"},
	{"(AX)", "(AX)"},
	{"(AX)(CX*8)", "(AX)(CX*8)"},
	{"(BP)(CX*4)", "(BP)(CX*4)"},
	{"(BP)(DX*4)", "(BP)(DX*4)"},
	{"(BP)(R8*4)", "(BP)(R8*4)"},
	{"(BX)", "(BX)"},
	{"(DI)", "(DI)"},
	{"(DI)(BX*1)", "(DI)(BX*1)"},
	{"(DX)", "(DX)"},
	{"(R9)", "(R9)"},

	KMOVQ K2, -17(BP)(SI*4)                            // c4e1f89154b5ef
	KMOVQ K7, -17(BP)(SI*4)                            // c4e1f8917cb5ef
	KMOVQ K0, DX                                       // c4e1fb93d0
	KMOVQ K5, DX                                       // c4e1fb93d5
	KMOVQ K0, BP                                       // c4e1fb93e8
	KMOVQ K5, BP                                       // c4e1fb93ed

	VRANGEPD $11, 17(SP)(BP*4), X31, K2, X12           // 6273850250a4ac110000000b
	VRANGEPD $11, X24, X23, K2, X16                    // 6283c50250c00b
	VRANGEPD $11, X14, X23, K2, X16                    // 62c3c50250c60b
	VRANGEPD $11, X0, X23, K2, X16                     // 62e3c50250c00b
	VRANGEPD $11, 17(SP)(BP*8), X23, K2, X16           // 62e3c5025084ec110000000b

	VPMADD52HUQ X0, X11, K1, X18                       // 62e2a509b5d0
	VPMADD52HUQ 17(SP)(BP*2), X11, K1, X18             // 62e2a509b5946c11000000
	VPMADD52HUQ -7(DI)(R8*4), X11, K1, X18             // 62a2a509b59487f9ffffff
	VPMADD52HUQ X7, X31, K1, X18                       // 62e28501b5d7
	VPMADD52HUQ X0, X31, K1, X18                       // 62e28501b5d0
	VPMADD52HUQ 17(SP)(BP*2), X31, K1, X18             // 62e28501b5946c11000000

	VPERMB 17(SP)(BP*1), Y31, K7, Y22                  // 62e205278db42c11000000
	VPERMB -7(CX)(DX*8), Y31, K7, Y22                  // 62e205278db4d1f9ffffff
	VPERMB Y5, Y5, K7, Y22                             // 62e2552f8df5
	VPERMB Y19, Y5, K7, Y22                            // 62a2552f8df3
	VPERMB Y31, Y5, K7, Y22                            // 6282552f8df7

	VGATHERPF0QPS K2, (BP)(Z10*2)                      // 62b27d4ac74c5500
	VGATHERPF0QPS K2, (R10)(Z29*8)                     // 62927d42c70cea
	VGATHERPF0QPS K2, (R14)(Z29*8)                     // 62927d42c70cee
	VGATHERPF1DPD K2, (R14)(Y29*8)                     // 6292fd42c614ee
	VGATHERPF1DPD K2, (AX)(Y4*1)                       // 62f2fd4ac61420
	VGATHERPF1DPD K2, (BP)(Y10*2)                      // 62b2fd4ac6545500

	VPCLMULQDQ $127, X7, X21, X15                      // 6273550044ff7f or 6273d50044ff7f
	VPCLMULQDQ $127, X19, X21, X15                     // 6233550044fb7f or 6233d50044fb7f
	VPCLMULQDQ $127, -17(BP)(SI*8), X21, X15           // 6273550044bcf5efffffff7f or 6273d50044bcf5efffffff7f
	VPCLMULQDQ $127, (R15), X21, X15                   // 62535500443f7f or 6253d500443f7f

	VCVTPH2PS 17(SP)(BP*8), K1, Z1                     // 62f27d49138cec11000000
	VCVTPH2PS 17(SP)(BP*4), K1, Z1                     // 62f27d49138cac11000000
	VCVTPH2PS Y14, K1, Z16                             // 62c27d4913c6
	VCVTPH2PS 17(SP)(BP*8), K1, Z16                    // 62e27d491384ec11000000
	VCVTPH2PS 17(SP)(BP*4), K1, Z16                    // 62e27d491384ac11000000