module attributes {tf.versions = {producer = 179 : i32}} {
  func.func @main(%arg0: tensor<f32>, %arg1: tensor<f32>) -> tensor<f32> {
    %0 = "tf.AddV2"(%arg0, %arg1) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    func.return %0 : tensor<f32>
  }
}

// CHECK-LABEL: HloModule main
// CHECK:       ENTRY %main.{{[0-9]+}} ([[ARG0:.*]]: f32[], [[ARG1:.*]]: f32[]) -> (f32[]) {

	RET
#ifndef GOARM64_LSE
load_store_loop:
	LDAXRW	(R0), R2
	ORR	R1, R2, R3
	STLXRW	R3, (R0), R4
	CBNZ	R4, load_store_loop
	MOVD R2, ret+16(FP)
	RET
#endif

// func And32(addr *uint32, v uint32) old uint32
TEXT ·And32(SB), NOSPLIT, $0-20
	MOVD	ptr+0(FP), R0
	MOVW	val+8(FP), R1
#ifndef GOARM64_LSE
	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
	CBZ 	R4, load_store_loop
#endif

	MOVD	ptr+0(FP), R3
	MOVW	val+8(FP), R4
	LWSYNC
again:
	LWAR	(R3), R6
	OR	R4, R6, R7
	STWCCC	R7, (R3)
	BNE	again
	MOVW	R6, ret+16(FP)
	RET

// func And32(addr *uint32, v uint32) old uint32
TEXT ·And32(SB), NOSPLIT, $0-20
	MOVD	ptr+0(FP), R3
	MOVW	val+8(FP), R4
	LWSYNC
again:
	LWAR	(R3),R6
	AND	R4, R6, R7
	STWCCC	R7, (R3)
	BNE	again
	MOVW	R6, ret+16(FP)
	RET

// func Or32(ptr *uint32, val uint32) uint32
TEXT ·Or32(SB), NOSPLIT, $0-20
	MOV	ptr+0(FP), A0
	MOVW	val+8(FP), A1
	AMOORW	A1, (A0), A2
	MOVW	A2, ret+16(FP)
	RET

// func And32(ptr *uint32, val uint32) uint32
TEXT ·And32(SB), NOSPLIT, $0-20
	MOV	ptr+0(FP), A0
	MOVW	val+8(FP), A1
	AMOANDW	A1, (A0), A2
	MOVW	A2, ret+16(FP)
	RET

// func Or64(ptr *uint64, val uint64) uint64

// Test, using compiler diagnostic flags, that inlining is working.
// Compiles but does not run.

package foo

import (
	"errors"
	"runtime"
	"unsafe"
)

func add2(p *byte, n uintptr) *byte { // ERROR "can inline add2" "leaking param: p to result"
	return (*byte)(add1(unsafe.Pointer(p), n)) // ERROR "inlining call to add1"
}

	MOVV	ptr+0(FP), R1
	MOVW	val+8(FP), R2

	SYNC
	LL	(R1), R3
	OR	R2, R3, R4
	SC	R4, (R1)
	BEQ	R4, -3(PC)
	SYNC
	MOVW	R3, ret+16(FP)
	RET

// func And32(addr *uint32, v uint32) old uint32
TEXT ·And32(SB), NOSPLIT, $0-20
	MOVV	ptr+0(FP), R1
	MOVW	val+8(FP), R2

	SYNC
	LL	(R1), R3
	AND	R2, R3, R4
	SC	R4, (R1)
	BEQ	R4, -3(PC)
	SYNC
	MOVW	R3, ret+16(FP)
	RET

    After this pass, the computation would become:

    ```mlir
    %resource_handle = "tf.VarHandleOp"()
    %init_value = "tf.ReadVariableOp"(%resource_handle)
    %1:2 = "tf_device.cluster"() ( {
      %new_value = "tf.AddV2"(%init_value, %init_value)
      tf_device.return %new_value, %new_value
    })
    "tf.AssignVariableOp"(%resource_handle, %1#1)
    ```

    You can see that there are a few main changes applied:

    %float_zp = "tf.Cast"(%zp) {Truncate = false} : (tensor<*xi32>) -> tensor<*xf32>
    %zp_plus_round_cst = "tf.AddV2"(%float_zp, %round_cst) : (tensor<*xf32>, tensor<f32>) -> tensor<*xf32>

    %div = "tf.Div"(%input, %scale) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    %add = "tf.AddV2"(%div, %zp_plus_round_cst) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    %round = "tf.Floor"(%add) : (tensor<*xf32>) -> tensor<*xf32>

			continue
		}
		if err == nil {
			addr2, err := ResolveUDPAddr(addr.Network(), addr.String())
			if !reflect.DeepEqual(addr2, tt.addr) || err != tt.err {
				t.Errorf("(%q, %q): ResolveUDPAddr(%q, %q) = %#v, %v, want %#v, %v", tt.network, tt.litAddrOrName, addr.Network(), addr.String(), addr2, err, tt.addr, tt.err)
			}
		}
	}
}

func TestWriteToUDP(t *testing.T) {

			continue
		}
		if err == nil {
			addr2, err := ResolveTCPAddr(addr.Network(), addr.String())
			if !reflect.DeepEqual(addr2, tt.addr) || err != tt.err {
				t.Errorf("(%q, %q): ResolveTCPAddr(%q, %q) = %#v, %v, want %#v, %v", tt.network, tt.litAddrOrName, addr.Network(), addr.String(), addr2, err, tt.addr, tt.err)
			}
		}
	}
}

var tcpListenerNameTests = []struct {