VPCOMPRESSB Y24, K1, Y24                           // 62027d2963c0
	VPCOMPRESSB Y21, K1, Y24                           // 62827d2963e8
	VPCOMPRESSB Y5, K1, -17(BP)(SI*8)                  // 62f27d29636cf5ef
	VPCOMPRESSB Y24, K1, -17(BP)(SI*8)                 // 62627d296344f5ef
	VPCOMPRESSB Y21, K1, -17(BP)(SI*8)                 // 62e27d29636cf5ef
	VPCOMPRESSB Y5, K1, (R15)                          // 62d27d29632f

	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

	{"$~15", "$-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)"},

	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

	VPOPCNTB Z18, K7, Z13                              // 62327d4f54ea
	VPOPCNTB Z8, K7, Z13                               // 62527d4f54e8
	VPOPCNTB 17(SP)(BP*8), K7, Z13                     // 62727d4f54acec11000000
	VPOPCNTB 17(SP)(BP*4), K7, Z13                     // 62727d4f54acac11000000
	VPOPCNTW X20, K3, X11                              // 6232fd0b54dc
	VPOPCNTW X5, K3, X11                               // 6272fd0b54dd

	VRSQRT28PS 17(SP)(BP*8), K1, Z12                   // 62727d49cca4ec11000000
	VRSQRT28PS 17(SP)(BP*4), K1, Z12                   // 62727d49cca4ac11000000
	VRSQRT28PS Z11, K1, Z22                            // 62c27d49ccf3
	VRSQRT28PS Z5, K1, Z22                             // 62e27d49ccf5
	VRSQRT28PS 17(SP)(BP*8), K1, Z22                   // 62e27d49ccb4ec11000000

	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

	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

			"#define KEYROUND(xmm, load, off, r1, r2, index) \\",
			"\tMOVBLZX	(BP)(DX*4),	R8 \\",
			"\tload((off+1), r2) \\",
			"\tMOVB	R8,		(off*4)(R12) \\",
			"\tPINSRW	$index, (BP)(R8*4), xmm",
			"#define LOAD(off, reg) \\",
			"\tMOVBLZX	(off*4)(R12),	reg \\",
			"\tADDB	reg,		DX",
			"KEYROUND(X0, LOAD, 8, AX, BX, 0)",
		),

·blockAMD64(SB), $264-32 MOVQ p_base+8(FP), SI MOVQ p_len+16(FP), DX SHRQ $0x06, DX SHLQ $0x06, DX LEAQ (SI)(DX*1), DI MOVQ DI, 256(SP) CMPQ SI, DI JEQ end MOVQ dig+0(FP), BP MOVL (BP), R8 MOVL 4(BP), R9 MOVL 8(BP), R10 MOVL 12(BP), R11 MOVL 16(BP), R12 MOVL 20(BP), R13 MOVL 24(BP), R14 MOVL 28(BP), R15 loop: MOVQ SP, BP MOVL (SI), AX BSWAPL AX MOVL AX, (BP) ADDL AX, R15 MOVL R12, AX ADDL $0x428a2f98, R15 MOVL R12, CX RORL $0x06, AX MOVL R12, DX RORL $0x0b, CX XORL CX, AX MOVL R12, CX RORL $0x19, DX ANDL R13, CX...