// [1]
  auto reduced = Cast(cpu_scope, reduction_indices_pos, DataType::DT_INT32);

  // [0, 1, 2]
  auto idx = Range(cpu_scope, zero, rank, one);

  // [0, 2]
  auto other = SetDiff1D(cpu_scope, idx, reduced).out;

  // [1, 0, 2]
  auto perm =
      Concat(cpu_scope, std::initializer_list<Input>{reduced, other}, 0);

  // 3 => [3]

// tf.data ops
//===----------------------------------------------------------------------===//

def TF_ReduceDatasetOp : TF_Op<"ReduceDataset", [SameVariadicOperandSize]> {
  let summary = [{
    Reduces the input dataset to a singleton using a reduce function.
  }];

  let arguments = (ins
    TF_VariantTensor:$input_dataset,
    Variadic<TF_Tensor>:$initial_state,
    Variadic<TF_Tensor>:$other_arguments,

    SymbolRefAttr:$f,

	// optionally, gcController.bgScanCredit. Lower values give a more
	// accurate assist ratio and make it more likely that assists will
	// successfully steal background credit. Higher values reduce memory
	// contention.
	gcCreditSlack = 2000

	// gcAssistTimeSlack is the nanoseconds of mutator assist time that
	// can accumulate on a P before updating gcController.assistTime.
	gcAssistTimeSlack = 5000

compile them as appropriate for the source language, and link the result into a binary file. For an executable component, Gradle can produce executable binary files. For a library component, Gradle can produce both static and shared library binary files. For example, when you define a library called `helloworld` and build on Linux, Gradle will, by default, produce `libhelloworld.so` and `libhelloworld.a` binaries.

In many cases, more than one binary can be produced for a component. These...

                                                    Builder *builder) {
  tensorflow::TensorFormat format;
  if (!FormatFromString(data_format.getValue().str(), &format)) return {};

  // Reduce along all dimensions except the feature dimension.
  int64_t feature_dim = GetTensorFeatureDimIndex(rank, format);
  llvm::SmallVector<int64_t, 4> dims_to_reduce(rank - 1);

	println()
}

// addb returns the byte pointer p+n.
//
//go:nowritebarrier
//go:nosplit
func addb(p *byte, n uintptr) *byte {
	// Note: wrote out full expression instead of calling add(p, n)
	// to reduce the number of temporaries generated by the
	// compiler for this trivial expression during inlining.
	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + n))
}

// subtractb returns the byte pointer p-n.
//

// Fold reshapes re-inserting reduced dimensions into the results of a reduction
// with `keep_dims=false` by changing it to one using `keep_dims=true`.
foreach ReduceOp = [TFL_MeanOp, TFL_ReduceMaxOp, TFL_ReduceMinOp,
                    TFL_ReduceProdOp, TFL_SumOp] in {
  def FoldReshapeTo#ReduceOp : Pat<
    (TFL_ReshapeOp
      (ReduceOp:$reduce $input, (Arith_ConstantOp I32ElementsAttr: $axes),

				queued:    time.Now(),
				setIndex:  er.setIndex,
				poolIndex: er.poolIndex,
			})
		}

		return
	}()

	validResp := 0
	totalResp := 0

	// minDisks value is only to reduce the number of calls
	// to the disks; this value is not accurate because we do
	// not know the storage class of the object yet
	minDisks := 0
	if p := globalStorageClass.GetParityForSC(""); p > -1 {

					triggerValue, triggerSupported = value, true
					break
				}
			}
		}
	}

	// It boils down to a tradeoff between:
	// - having it as small as possible to reduce memory usage
	// - having it large enough to ensure that watchers that need to process
	//   a bunch of changes have enough buffer to avoid from blocking other
	//   watchers on our watcher having a processing hiccup

	},
	ErrSlowDownRead: {
		Code:           "SlowDownRead",
		Description:    "Resource requested is unreadable, please reduce your request rate",
		HTTPStatusCode: http.StatusServiceUnavailable,
	},
	ErrSlowDownWrite: {
		Code:           "SlowDownWrite",
		Description:    "Resource requested is unwritable, please reduce your request rate",
		HTTPStatusCode: http.StatusServiceUnavailable,
	},
	ErrMaxVersionsExceeded: {