if optional:
          return (None, None)
        else:
          raise KeyError('OpDef does not exist: ' + f_name)
      elif len(compose_dec) > 1:
        raise KeyError('More than one TF ops decomposes for.')
      else:
        op_name = compose_dec[0].args[0].value

    op_def = op_def_registry.get(op_name)
    if not op_def:
      raise ValueError('Not a registered op: ' + op_name)

          concat_dim);
      result = concat_op.getResult();
    }

    rewriter.replaceOp(op, result);
    return success();
  }
};

// Decomposes Softmax and LogSoftmax to primitive TF ops, using the following
// formulas:
//
//     softmax = div(exp(logits), sum(exp(logits)))
//     log_softmax = sub(logits, log(sum(exp(logits))))
//

// non-generational and non-compacting. Allocation is done using size segregated per P allocation
// areas to minimize fragmentation while eliminating locks in the common case.
//
// The algorithm decomposes into several steps.
// This is a high level description of the algorithm being used. For an overview of GC a good
// place to start is Richard Jones' gchandbook.org.
//

// RUN: tf-opt %s -tf-device-decompose-resource-ops | FileCheck -check-prefixes=CHECK,ALWAYS-DECOMPOSE %s
// RUN: tf-opt %s -tf-device-decompose-resource-ops-in-cluster | FileCheck -check-prefixes=CHECK,CLUSTER-ONLY %s

// Tests different behavior of two passes if the op is outside cluster
// tf-device-decompose-resource-ops would always decompose the resource op.
// tf-device-decompose-resource-ops-in-cluster would not decompose ops outside

```
### `-tf-data-optimization`

_Performs tf.data optimizations_

### `-tf-decompose-reduce-dataset`

_Decomposes ReduceDataset op into dataset operations._

  Decomposes ReduceDataset op into a while loop that iterates the dataset and calls
  into the reduction function.  This decomposition is only done if the

	}

	if state.varParts == nil {
		state.varParts = make(map[*ir.Name][]SlotID)
	} else {
		for n := range state.varParts {
			delete(state.varParts, n)
		}
	}

	// Recompose any decomposed variables, and establish the canonical
	// IDs for each var and slot by filling out state.vars and state.slots.

	state.slots = state.slots[:0]
	state.vars = state.vars[:0]
	for i, slot := range f.Names {

 *
 * The path identifies a specific resource on the host. Paths have a hierarchical structure like
 * "/square/okhttp/issues/1486" and decompose into a list of segments like `["square", "okhttp",
 * "issues", "1486"]`.
 *
 * This class offers methods to compose and decompose paths by segment. It composes each path
 * from a list of segments by alternating between "/" and the encoded segment. For example the

	// cond: is64BitInt(v.Type) && !config.BigEndian && v.Type.IsSigned() && !(b.Func.pass.name == "decompose builtin")
	// result: (Int64Make (Arg <typ.Int32> {n} [off+4]) (Arg <typ.UInt32> {n} [off]))
	for {
		off := auxIntToInt32(v.AuxInt)
		n := auxToSym(v.Aux)
		if !(is64BitInt(v.Type) && !config.BigEndian && v.Type.IsSigned() && !(b.Func.pass.name == "decompose builtin")) {
			break
		}
		v.reset(OpInt64Make)

    for (auto v : return_op.getOperands()) {
      if (absl::c_count(return_value_from_host, v) == 0) {
        return_value_from_device.push_back(v);
      }
    }
  }
}

// Decompose control flow in `device_cluster` into device computation and host
// (outside compiled) computation into two separate control flow ops with
// communication between the device/host for data dependencies.  Both device and

				if c := q.Cmp(got); c != 0 {
					t.Errorf("%v: round trip from decimal back to quantity is not comparable: %d: %#v vs %#v", item.input, c, got, q)
				}
			}

			// verify that we can decompose the input and get the same result by building up from the base.
			positive, _, num, denom, suffix, err := parseQuantityString(item.input)
			if err != nil {
				t.Errorf("%v: unexpected error: %v", item.input, err)