//
// The paper gives O((M+N)*log(M)) as the number of assignments assuming a
// rotation algorithm which uses O(M+N+gcd(M+N)) assignments. The argumentation
// in the paper carries through for Swap operations, especially as the block
// swapping rotate uses only O(M+N) Swaps.
//
// symMerge assumes non-degenerate arguments: a < m && m < b.

//
// The paper gives O((M+N)*log(M)) as the number of assignments assuming a
// rotation algorithm which uses O(M+N+gcd(M+N)) assignments. The argumentation
// in the paper carries through for Swap operations, especially as the block
// swapping rotate uses only O(M+N) Swaps.
//
// symMerge assumes non-degenerate arguments: a < m && m < b.

		if meta, ok := metadata.FromIncomingContext(ac.GrpcContext); ok {
			return meta.Get(header)
		}
	} else if ac.Request != nil {
		return ac.Request.Header.Values(header)
	}
	return nil
}

// Caller carries the identity and authentication source of a caller.
type Caller struct {
	AuthSource AuthSource
	Identities []string

	KubernetesInfo KubernetesInfo
}

func RegisterDefaultKVS(kvsMap map[string]KVS) {
	for subSys, kvs := range kvsMap {
		DefaultKVS[subSys] = kvs
	}
}

// HelpSubSysMap - help for all individual KVS for each sub-systems
// also carries a special empty sub-system which dumps
// help for each sub-system key.
var HelpSubSysMap = map[string]HelpKVS{}

// RegisterHelpSubSys - this function saves
// input help KVS for each sub-system globally,

		return nil
	})
	return <-c
}

// A reporter reports output, warnings, and errors.
type reporter struct {
	prev  <-chan *reporterState
	state *reporterState
}

// reporterState carries the state of a reporter instance.
//
// Only one reporter at a time may have access to a reporterState.
type reporterState struct {
	out, err io.Writer
	exitCode int
}

	v := c.field()
	v.Set(reflect.Append(v, reflect.Zero(v.Type().Elem())))
	l := v.Len()
	reflect.Copy(v.Slice(i+1, l), v.Slice(i, l))
	v.Index(i).Set(reflect.ValueOf(n))
	c.iter.index++
}

// application carries all the shared data so we can pass it around cheaply.
type application struct {
	pre, post ApplyFunc
	cursor    Cursor
	iter      iterator
}

And we add the corresponding native jar file so that each runtime variant now carries two files: the main jar and the native jar.

In the build script, we can now request a specific variant and Gradle will fail with a selection error if more information is needed to make a decision.

//
// The paper gives O((M+N)*log(M)) as the number of assignments assuming a
// rotation algorithm which uses O(M+N+gcd(M+N)) assignments. The argumentation
// in the paper carries through for Swap operations, especially as the block
// swapping rotate uses only O(M+N) Swaps.
//
// symMerge assumes non-degenerate arguments: a < m && m < b.

def IsLegalNumpyRankedBroadcast :
    Constraint<CPred<"hlo::isLegalNumpyRankedBroadcast($0, $1, $2 ? $2.asArrayRef() : ArrayRef<int64_t>{})">,
    "broadcasting should be compatible with TF ops">;

// Return a constant op that carries the shape of the given value.
def ShapeToConst : NativeCodeCall<"ShapeToConst($_builder, $0)">;

// Check if broadcast dimensions match Tensorflow convention.

====
Choosing the right attributes to set is the hardest thing in this process, because they carry the semantics of the variant.
Therefore, before adding _new attributes_, you should always ask yourself if there isn't an attribute which carries the semantics you need.
If there isn't, then you may add a new attribute.
When adding new attributes, you must also be careful because it's possible that it creates ambiguity during selection.