%2 = stablehlo.divide %arg0, %0 : tensor<2xf32>
  return %2 : tensor<2xf32>
}

// CHECK-LABEL: divideToMulReciprocalSplat
// CHECK: stablehlo.constant dense<5.000000e-01> : tensor<2xf32>
// CHECK: stablehlo.multiply

// -----

func.func @divideToMulReciprocal(%arg0: tensor<2xf32>) -> tensor<2xf32> {
  %0 = stablehlo.constant dense<[2.0, 3.0]> : tensor<2xf32>
  %2 = stablehlo.divide %arg0, %0 : tensor<2xf32>

def TFL_ArithmeticCount : OpInterface<"TflArithmeticCountOpInterface"> {
  let description = [{
    Interface for TFLite ops to calculate arithmetic count (Multiply-Add Count).
  }];

  let methods = [
    StaticInterfaceMethod<
      [{Returns an integer representing the op's arithmetic count (Multiply-Add
      Count), return -1 if the arithmetic count cannot be determined.}],
       "int64_t", "GetArithmeticCount", (ins "Operation*":$op)
    >,

	// This does not seem worthwhile, at least for Go: not having any high
	// bits in the multiplier reduces the effect of low bits on the highest bits,
	// and it only saves 1 multiply out of 3.
	// (On 32-bit systems, it saves 1 out of 6, since Mul64 is doing 4.)
	const (
		mulHi = 2549297995355413924
		mulLo = 4865540595714422341
		incHi = 6364136223846793005
		incLo = 1442695040888963407
	)

// It also runs ssa/check and gccheck to be sure that those are checked at least a
// little in each run.bash.  It does not check or run the generated code.
// The test file is however a useful example of fused-vs-cascaded multiply-add.
func TestFmaHash(t *testing.T) {
	switch runtime.GOOS {
	case "linux", "darwin":
	default:
		t.Skipf("Slow test, usually avoid it, os=%s not linux or darwin", runtime.GOOS)
	}
	switch runtime.GOARCH {

	err := syscall.Sysinfo(&info)
	if err != nil {
		errorList = append(errorList, errors.Wrapf(err, "failed to get system info"))
	}

	// Totalram holds the total usable memory. Unit holds the size of a memory unit in bytes. Multiply them and convert to MB
	actual := uint64(info.Totalram) * uint64(info.Unit) / 1024 / 1024
	if actual < mc.Mem {
		errorList = append(errorList, errors.Errorf("the system RAM (%d MB) is less than the minimum %d MB", actual, mc.Mem))

// pages into objects of the given size wastes at most 12.5% (1.125x)
// of the memory. It is not necessary that the cutoff here be
// the same as above.
//
// The two sources of waste multiply, so the worst possible case
// for the above constraints would be that allocations of some
// size might have a 26.6% (1.266x) overhead.
// In practice, only one of the wastes comes into play for a

		runtime.HandleError(fmt.Errorf("feature %q is not implemented in DefaultFeatureSupportChecker", feature))
	}
}

func (f *defaultFeatureSupportChecker) checkClient(ctx context.Context, c client) {
	// start with 10 ms, multiply by 2 each step, until 15 s and stays on 15 seconds.
	delayFunc := wait.Backoff{
		Duration: 10 * time.Millisecond,
		Cap:      15 * time.Second,
		Factor:   2.0,
		Steps:    11}.DelayFunc()
	f.lock.Lock()

            matmul_op.getAdjY() ? shape_y.back() : *(shape_y.rbegin() + 1);
        const int y_col =
            !matmul_op.getAdjY() ? shape_y.back() : *(shape_y.rbegin() + 1);

        // Checks that matrix multiply can perform a valid contraction.
        if (x_col != y_row) {
          result_shape.clear();
          return false;
        }

        result_shape.push_back(x_row);
        result_shape.push_back(y_col);

  // will overcount the divisors.
  //    Following from (LEMMA 1) the only windows which contain overcounted
  // divisors are the ones on the outside right and bottom edge. We can iterate
  // over these windows and multiply the corresponding out element by
  // `kernel_size / X` where `X` is the number of elements in the padded input
  // tensor not in the newly padded zone. This corrects the overcounting of