and scales with the number of available CPU cores.
Decompression speed is typically at least 1GB/s.

This means that in cases where raw IO is below these numbers
compression will not only reduce disk usage but also help increase system throughput.
Typically, enabling compression on spinning disk systems
will increase speed when the content can be compressed.

## Get Started

### 1. Prerequisites

  // CHECK: %[[INDEX:.*]] = "mhlo.dynamic_iota"(%[[SHAPE]]) <{iota_dimension = 1 : i64}> : (tensor<2xindex>) -> tensor<3x7xi64>
  // CHECK: %[[REDUCE:.*]]:2 = mhlo.reduce(%arg0 init: %[[INIT]]), (%[[INDEX]] init: %[[INDEX_INIT]])
  // CHECK: return %[[REDUCE]]#1 : tensor<3xi64>
  %axis = "tf.Const"() { value = dense<1> : tensor<i32> } : () -> tensor<i32>

     */
    static String createMavenVersionString(Properties buildProperties) {
        String timestamp = reduce(buildProperties.getProperty("timestamp"));
        String version = reduce(buildProperties.getProperty(BUILD_VERSION_PROPERTY));
        String rev = reduce(buildProperties.getProperty("buildNumber"));
        String distributionName = reduce(buildProperties.getProperty("distributionName"));

        String msg = distributionName + " ";

    auto output_type = mlir::cast<ShapedType>(op.getOutput().getType());

    if (!input_type.hasRank() || !output_type.hasRank()) return failure();

    // The pattern attempts to reduce the rank of the input to BroadcastTo.
    // Thus, we fail to match if the consuming reshape rank is larger.
    ArrayRef<int64_t> input_shape = input_type.getShape();

		// executing this request.
		//
		// To mitigate the impact of it, we're capping the maximum number of
		// seats that can be assigned to a given request. Thanks to it:
		// 1) we reduce the amount of seat-seconds that are "wasted" during
		//    dispatching and executing initial phase of the request
		// 2) we are not changing the finalWork estimate - just potentially

package main

func Map[F, T any](s []F, f func(F) T) []T { return nil }

func Reduce[Elem1, Elem2 any](s []Elem1, initializer Elem2, f func(Elem2, Elem1) Elem2) Elem2 { var x Elem2; return x }

func main() {
	var s []int
	var f1 func(int) float64
	var f2 func(float64, int) float64
	_ = Map[int](s, f1)
	_ = Map(s, f1)
	_ = Reduce[int](s, 0, f2)
	_ = Reduce(s, 0, f2)

See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

// This is a pass to reduce operands without changing the outcome.

#include <vector>

#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/STLExtras.h"

        "@llvm-project//mlir:Support",
        "@llvm-project//mlir:TransformUtils",
        "@local_xla//xla/mlir_hlo",
    ],
)

cc_library(
    name = "reduce",
    srcs = [
        "reduce.cc",
    ],
    hdrs = [
        "reduce.h",
    ],
    deps = [
        "//tensorflow/compiler/mlir/lite/stablehlo:hlo_matchers",
        "@llvm-project//llvm:Support",
        "@llvm-project//mlir:ArithDialect",

	// compiled Go functions.
	//
	// Requires wrappers (to do ABI translation), and reflect (so
	// reflection calls use registers).
	RegabiArgs bool

	// HeapMinimum512KiB reduces the minimum heap size to 512 KiB.
	//
	// This was originally reduced as part of PacerRedesign, but
	// has been broken out to its own experiment that is disabled
	// by default.
	HeapMinimum512KiB bool

	*v = *feZero
	return v
}

var feOne = &Element{1, 0, 0, 0, 0}

// One sets v = 1, and returns v.
func (v *Element) One() *Element {
	*v = *feOne
	return v
}

// reduce reduces v modulo 2^255 - 19 and returns it.
func (v *Element) reduce() *Element {
	v.carryPropagate()

	// After the light reduction we now have a field element representation
	// v < 2^255 + 2^13 * 19, but need v < 2^255 - 19.