VGF2P8AFFINEQB $82, X19, X0, K1, X16               // 62a3fd09cec352
	VGF2P8AFFINEQB $82, 17(SP)(BP*2), X0, K1, X16      // 62e3fd09ce846c1100000052
	VGF2P8AFFINEQB $82, -7(DI)(R8*4), X0, K1, X16      // 62a3fd09ce8487f9ffffff52
	VGF2P8AFFINEQB $82, X22, X28, K1, X16              // 62a39d01cec652
	VGF2P8AFFINEQB $82, X7, X28, K1, X16               // 62e39d01cec752

	VRSQRT28SS X28, X22, K7, X16                       // 62824d07cdc4
	VRSQRT28SS X8, X22, K7, X16                        // 62c24d07cdc0
	VRSQRT28SS X16, X12, K7, X16                       // 62a21d0fcdc0
	VRSQRT28SS X28, X12, K7, X16                       // 62821d0fcdc4
	VRSQRT28SS X8, X12, K7, X16                        // 62c21d0fcdc0
	VRSQRT28SS X16, X6, K7, X6                         // 62b24d0fcdf0

	VAESENC X14, X16, X13                              // 62527d00dcee or 6252fd00dcee
	VAESENC X19, X16, X13                              // 62327d00dceb or 6232fd00dceb
	VAESENC X8, X16, X13                               // 62527d00dce8 or 6252fd00dce8
	VAESENC 99(R15)(R15*8), X16, X13                   // 62127d00dcacff63000000 or 6212fd00dcacff63000000

	VPERMI2B 17(SP), X8, K7, X16                       // 62e23d0f75842411000000
	VPERMI2B -17(BP)(SI*4), X8, K7, X16                // 62e23d0f7584b5efffffff
	VPERMI2B X15, X1, K7, X16                          // 62c2750f75c7
	VPERMI2B X0, X1, K7, X16                           // 62e2750f75c0
	VPERMI2B X16, X1, K7, X16                          // 62a2750f75c0
	VPERMI2B 17(SP), X1, K7, X16                       // 62e2750f75842411000000

	x12, x11 = bits.Mul64(x4, arg2[0])
	var x13 uint64
	var x14 uint64
	x13, x14 = bits.Add64(x12, x9, uint64(0x0))
	var x15 uint64
	var x16 uint64
	x15, x16 = bits.Add64(x10, x7, uint64(p224Uint1(x14)))
	var x17 uint64
	var x18 uint64
	x17, x18 = bits.Add64(x8, x5, uint64(p224Uint1(x16)))
	x19 := (uint64(p224Uint1(x18)) + x6)
	var x20 uint64
	_, x20 = bits.Mul64(x11, 0xffffffffffffffff)
	var x22 uint64
	var x23 uint64

	x12, x11 = bits.Mul64(x4, arg2[0])
	var x13 uint64
	var x14 uint64
	x13, x14 = bits.Add64(x12, x9, uint64(0x0))
	var x15 uint64
	var x16 uint64
	x15, x16 = bits.Add64(x10, x7, uint64(fiatScalarUint1(x14)))
	var x17 uint64
	var x18 uint64
	x17, x18 = bits.Add64(x8, x5, uint64(fiatScalarUint1(x16)))
	x19 := (uint64(fiatScalarUint1(x18)) + x6)
	var x20 uint64
	_, x20 = bits.Mul64(x11, 0xd2b51da312547e1b)
	var x22 uint64

TEXT asmtest_avx512_vnni(SB), NOSPLIT, $0
	VPDPBUSD X15, X16, K2, X6                          // 62d27d0250f7
	VPDPBUSD X11, X16, K2, X6                          // 62d27d0250f3
	VPDPBUSD X1, X16, K2, X6                           // 62f27d0250f1
	VPDPBUSD -15(R14)(R15*1), X16, K2, X6              // 62927d0250b43ef1ffffff
	VPDPBUSD -15(BX), X16, K2, X6                      // 62f27d0250b3f1ffffff

	if x8[i+1] < x8[0] {
		return 0
	}
	// amd64: `MOVWLZX\t2\([A-Z]+[0-9]*\)\([A-Z]+[0-9]*\*2\), [A-Z]+[0-9]*`
	if x16[i+1] < x16[0] {
		return 0
	}
	// amd64: `MOVWLZX\t2\([A-Z]+[0-9]*\)\([A-Z]+[0-9]*\*[12]\), [A-Z]+[0-9]*`
	if x16[16*i+1] < x16[0] {
		return 0
	}
	// amd64: `MOVL\t4\([A-Z]+[0-9]*\)\([A-Z]+[0-9]*\*4\), [A-Z]+[0-9]*`
	if x32[i+1] < x32[0] {
		return 0
	}

	x12, x11 = bits.Mul64(x4, arg2[0])
	var x13 uint64
	var x14 uint64
	x13, x14 = bits.Add64(x12, x9, uint64(0x0))
	var x15 uint64
	var x16 uint64
	x15, x16 = bits.Add64(x10, x7, uint64(p256Uint1(x14)))
	var x17 uint64
	var x18 uint64
	x17, x18 = bits.Add64(x8, x5, uint64(p256Uint1(x16)))
	x19 := (uint64(p256Uint1(x18)) + x6)
	var x20 uint64
	var x21 uint64
	x21, x20 = bits.Mul64(x11, 0xffffffff00000001)
	var x22 uint64

TEXT asmtest_avx512_bitalg(SB), NOSPLIT, $0
	VPOPCNTB X14, K4, X16                              // 62c27d0c54c6
	VPOPCNTB X19, K4, X16                              // 62a27d0c54c3
	VPOPCNTB X8, K4, X16                               // 62c27d0c54c0
	VPOPCNTB 15(R8)(R14*1), K4, X16                    // 62827d0c5484300f000000
	VPOPCNTB 15(R8)(R14*2), K4, X16                    // 62827d0c5484700f000000