DATA masks<>+0x78(SB)/8, $0x0000000000000000
DATA masks<>+0x80(SB)/8, $0xffffffffffffffff
DATA masks<>+0x88(SB)/8, $0x0000000000000000
DATA masks<>+0x90(SB)/8, $0xffffffffffffffff
DATA masks<>+0x98(SB)/8, $0x00000000000000ff
DATA masks<>+0xa0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xa8(SB)/8, $0x000000000000ffff
DATA masks<>+0xb0(SB)/8, $0xffffffffffffffff
DATA masks<>+0xb8(SB)/8, $0x0000000000ffffff

DATA shuff_00BA<>+0x08(SB)/8, $0xFFFFFFFFFFFFFFFF
DATA shuff_00BA<>+0x10(SB)/8, $0x0b0a090803020100
DATA shuff_00BA<>+0x18(SB)/8, $0xFFFFFFFFFFFFFFFF
GLOBL shuff_00BA<>(SB), 8, $32

// shuffle xDxC -> DC00
DATA shuff_DC00<>+0x00(SB)/8, $0xFFFFFFFFFFFFFFFF
DATA shuff_DC00<>+0x08(SB)/8, $0x0b0a090803020100
DATA shuff_DC00<>+0x10(SB)/8, $0xFFFFFFFFFFFFFFFF
DATA shuff_DC00<>+0x18(SB)/8, $0x0b0a090803020100

			// break the parser. Verify that integers in the opposite bases still
			// parse correctly.
			in: `go test fuzz v1
int(0x0)
int32(0x41)
int64(0xfffffffff)
uint32(0xcafef00d)
uint64(0xffffffffffffffff)
uint8(0b0000000)
byte(0x0)
byte('\000')
byte('\u0000')
byte('\'')
math.Float64frombits(9221120237041090562)
math.Float32frombits(2143289345)`,
			want: `go test fuzz v1
int(0)
rune('A')

	tests := []struct {
		// inputs
		src  RotateParams
		mask uint64

		// results
		in  *RotateParams
		out *RotateParams
	}{
		{
			src:  RotateParams{Start: 48, End: 15, Amount: 16},
			mask: 0xffffffffffffffff,
			in:   &RotateParams{Start: 48, End: 15, Amount: 16},
			out:  &RotateParams{Start: 48, End: 15, Amount: 16},
		},
		{
			src:  RotateParams{Start: 16, End: 47, Amount: 0},
			mask: 0x00000000ffffffff,

		}
	}
}

func alias_ssa(ui64 *uint64, ui32 *uint32) uint32 {
	*ui32 = 0xffffffff
	*ui64 = 0                  // store
	ret := *ui32               // load from same address, should be zero
	*ui64 = 0xffffffffffffffff // store
	return ret
}
func testdse(t *testing.T) {
	x := int64(-1)
	// construct two pointers that alias one another
	ui64 := (*uint64)(unsafe.Pointer(&x))
	ui32 := (*uint32)(unsafe.Pointer(&x))

	// Exit AVX mode.
	VZEROUPPER
	SALQ $32, CX
	ORQ CX, DX

	// Create mask to ignore overlap between previous 64 byte block
	// and the next.
	ANDQ $63, BX
	MOVQ $64, CX
	SUBQ BX, CX
	MOVQ $0xFFFFFFFFFFFFFFFF, R10
	SALQ CL, R10
	// Apply mask
	ANDQ R10, DX
	POPCNTQ DX, DX
	ADDQ DX, R12
	MOVQ R12, (R8)
	RET
endavx:
	// Exit AVX mode.
	VZEROUPPER
	MOVQ R12, (R8)

// p256One is one in the Montgomery domain.
var p256One = p256Element{0x0000000000000001, 0xffffffff00000000,
	0xffffffffffffffff, 0x00000000fffffffe}

var p256Zero = p256Element{}

// p256P is 2²⁵⁶ - 2²²⁴ + 2¹⁹² + 2⁹⁶ - 1 in the Montgomery domain.
var p256P = p256Element{0xffffffffffffffff, 0x00000000ffffffff,
	0x0000000000000000, 0xffffffff00000001}

// with any other values that might cause issues. we generate n^2 tests for each size to
// cover all cases.
var szs = []sizedTestData{
	sizedTestData{name: "uint64", sn: "64", u: []uint64{0, 1, 4294967296, 0xffffFFFFffffFFFF}},
	sizedTestData{name: "int64", sn: "64", i: []int64{-0x8000000000000000, -0x7FFFFFFFFFFFFFFF,
		-4294967296, -1, 0, 1, 4294967296, 0x7FFFFFFFFFFFFFFE, 0x7FFFFFFFFFFFFFFF}},

		expect64(t, "a", a, 127)
		expect64(t, "b", b, 255)
		expect64(t, "c", c, 32767)
		expect64(t, "d", d, 65535)
		expect64(t, "e", e, float64(int32(0x7fffffff)))
		expect64(t, "f", f, float64(uint32(0xffffffff)))
		expect64(t, "g", g, float64(int64(0x7fffffffffffffff)))
		expect64(t, "h", h, float64(uint64(0xffffffffffffffff)))

}
type szD struct {
	name   string
	sn     string
	u      []uint64
	i      []int64
	oponly string
}

var szs = []szD{
	{name: "uint64", sn: "64", u: []uint64{0, 1, 4294967296, 0x8000000000000000, 0xffffFFFFffffFFFF}},
	{name: "uint64", sn: "64", u: []uint64{3, 5, 7, 9, 10, 11, 13, 19, 21, 25, 27, 37, 41, 45, 73, 81}, oponly: "mul"},

	{name: "int64", sn: "64", i: []int64{-0x8000000000000000, -0x7FFFFFFFFFFFFFFF,