// RUN: tf-opt -tfl-prepare-tf=allow-bf16-and-f16-type-legalization=true %s | FileCheck %s

module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {

// CHECK-LABEL: conv_2d_bf16
func.func @conv_2d_bf16(%arg0 : tensor<256x32x32x3xbf16>, %arg1 : tensor<3x3x3x16xbf16>) -> tensor<256x8x7x16xbf16> {

// the LICENSE file.

package main

//go:noinline
func f8(x int32) bool {
	return byte(x&0xc0) == 64
}

//go:noinline
func f16(x int32) bool {
	return uint16(x&0x8040) == 64
}

func main() {
	if !f8(64) {
		panic("wanted true, got false")
	}
	if !f16(64) {
		panic("wanted true, got false")
	}

	return int64(int8(x))
}

func main() {
	if got := f32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1); got != 0xffffffff {
		println("bad f32", got)
	}
	if got := f16(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1); got != 0xffff {
		println("bad f16", got)
	}
	if got := f8(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1); got != 0xff {
		println("bad f8", got)
	}

	if t := x / y; t != q {
		fmt.Printf("%d/%d = %d, want %d\n", x, y, t, q)
		panic("divide")
	}
	if t := x % y; t != r {
		fmt.Printf("%d%%%d = %d, want %d\n", x, y, t, r)
		panic("divide")
	}
}

func f16(x, y, q, r int16) {
	if t := x / y; t != q {
		fmt.Printf("%d/%d = %d, want %d\n", x, y, t, q)
		panic("divide")
	}
	if t := x % y; t != r {
		fmt.Printf("%d%%%d = %d, want %d\n", x, y, t, r)
		panic("divide")
	}

		len(a.B.C1.D2.E2.F6) != 0 ||
		len(a.B.C1.D2.E2.F7) != 0 ||
		len(a.B.C1.D2.E2.F8) != 0 ||
		len(a.B.C1.D2.E2.F9) != 0 ||
		len(a.B.C1.D2.E2.F10) != 0 ||
		len(a.B.C1.D2.E2.F11) != 0 ||
		len(a.B.C1.D2.E2.F16) != 0 {
		panic("bad")
	}
	C.malloc(100)
}

type A struct {
	B
}

type B struct {
	C1 C
	C2 C
}

type C struct {
	D1 D
	D2 D
}

type D struct {
	E1 E
	E2 E

	var f14 float32 = 14.0
	var f15 float32 = 15.0
	var f16 float32 = 16.0
	i0 = int(f0)
	i1 = int(f1)
	i2 = int(f2)
	i3 = int(f3)
	i4 = int(f4)
	i5 = int(f5)
	i6 = int(f6)
	i7 = int(f7)
	i8 = int(f8)
	i9 = int(f9)
	i10 = int(f10)
	i11 = int(f11)
	i12 = int(f12)
	i13 = int(f13)
	i14 = int(f14)
	i15 = int(f15)
	i16 = int(f16)
	if f16 != count {
		panic("fail")
	}
	count -= 1

	FDIVD	F14, F15	// F15 = (r*c)/(2-c)
	FSUBD	F15, F5, F15	// lo-(r*c)/(2-c)
	FSUBD	F4, F15, F15	// (lo-(r*c)/(2-c))-hi
	FSUBD	F15, F1, F16	// F16 = y = 1-((lo-(r*c)/(2-c))-hi)
	// inline Ldexp(y, k), benefit:
	// 1, no parameter pass overhead.
	// 2, skip unnecessary checks for Inf/NaN/Zero
	FMOVD	F16, R0
	AND	$FracMask, R0, R2	// fraction
	LSR	$52, R0, R5	// exponent
	ADD	R1, R5		// R1 = int(k)
	CMP	$1, R5
	BGE	normal

      // Create new ConstantOp-ConvertOp-Operation sequences. At this moment,
      // old ConstantOp is guaranteed to have one F32->F16 convert op regardless
      // of its number of users.
      rewriter.setInsertionPointAfter(op);
      // create new F16 constant op in that location
      ConstantOp new_const = rewriter.create<ConstantOp>(
          op->getLoc(), new_result_type, new_value_attr);

		len(a.B.C1.D2.E2.F6) != 0 ||
		len(a.B.C1.D2.E2.F7) != 0 ||
		len(a.B.C1.D2.E2.F8) != 0 ||
		len(a.B.C1.D2.E2.F9) != 0 ||
		len(a.B.C1.D2.E2.F10) != 0 ||
		len(a.B.C1.D2.E2.F11) != 0 ||
		len(a.B.C1.D2.E2.F16) != 0 {
		panic("bad")
	}
}

type A struct {
	B
}

type B struct {
	C1 C
	C2 C
}

type C struct {
	D1 D
	D2 D
}

type D struct {
	E1 E
	E2 E
	E3 E
	E4 E

				{1, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
			},
			outputs: []outputInfo{
				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31