LogicalResult MatchReduceToArgMinMaxType2(mhlo::ReduceOp reduce_op,
                                          bool is_argmax);

// Base class for converting mhlo::ReduceOp to TF/TFL ArgMax/ArgMin ops.
template <typename Reduce, typename ArgReduce, typename BooleanReduce,
          bool is_argmax>
class ConvertReduceOpToArgMinMax : public OpConversionPattern<mhlo::ReduceOp> {
 public:
  using OpConversionPattern::OpConversionPattern;

        "tf-reduce-identity.mlir": ["no_windows"],
    },
    test_file_exts = ["mlir"],
)

# Bundle together all of the test utilities that are used by tests.
filegroup(
    name = "test_utilities",
    testonly = True,
    data = [
        ":reducer_tester",
        "//tensorflow/compiler/mlir:tf-opt",
        "//tensorflow/compiler/mlir:tf-reduce",
        "@llvm-project//llvm:FileCheck",

`Make[T]` function, which produces a reference to a canonical copy of
the value in the form of a `Handle[T]`.
Two `Handle[T]` are equal if and only if the values used to produce the
handles are equal, allowing programs to deduplicate values and reduce
their memory footprint.
Comparing two `Handle[T]` values is efficient, reducing down to a simple

	// workbufAlloc is the number of bytes to allocate at a time
	// for new workbufs. This must be a multiple of pageSize and
	// should be a multiple of _WorkbufSize.
	//
	// Larger values reduce workbuf allocation overhead. Smaller
	// values reduce heap fragmentation.
	workbufAlloc = 32 << 10
)

func init() {
	if workbufAlloc%pageSize != 0 || workbufAlloc%_WorkbufSize != 0 {
		throw("bad workbufAlloc")
	}
}

# Reduce Xmx after https://github.com/gradle/gradle-private/issues/4168 is resolved
org.gradle.jvmargs=-Xmx2700m -XX:MaxMetaspaceSize=768m -XX:+HeapDumpOnOutOfMemoryError -Dfile.encoding=UTF-8
org.gradle.parallel=true
org.gradle.caching=true
org.gradle.configuration-cache=true
systemProp.gradle.publish.skip.namespace.check=true
# Kotlin DSL settings
org.gradle.kotlin.dsl.allWarningsAsErrors=true
# Kotlin settings
kotlin.incremental.useClasspathSnapshot=true

	n := uint(len(rr.limbs))
	mLen := uint(m.BitLen())
	logR := _W * n

	// We start by computing R = 2^(_W * n) mod m. We can get pretty close, to
	// 2^⌊log₂m⌋, by setting the highest bit we can without having to reduce.
	rr.limbs[n-1] = 1 << ((mLen - 1) % _W)
	// Then we double until we reach 2^(_W * n).
	for i := mLen - 1; i < logR; i++ {
		rr.Add(rr, m)
	}

==== Compile-time dependencies

Maven automatically exposes dependencies using its implicit `compile` scope to the consumers of that project.
This behavior is undesirable, and Gradle takes steps to help library authors reduce their API footprint using the `api` and `implementation` configurations of the `java-library` plugin.

## Compiler {#compiler}

The build time overhead to building with [Profile Guided Optimization](/doc/pgo) has been reduced significantly.
Previously, large builds could see 100%+ build time increase from enabling PGO.
In Go 1.23, overhead should be in the single digit percentages.

<!-- https://go.dev/issue/62737 , https://golang.org/cl/576681,  https://golang.org/cl/577615 -->
The compiler in Go 1.23 can now overlap the stack frame slots of local variables

                                        out_dims);
  }

  Value reduced = builder.create<TF::EinsumOp>(
      loc, RankedTensorType::get(output_shape, builder.getIntegerType(32)),
      input_arguments, builder.getStringAttr(einsum_equation));

  reduced.getDefiningOp()->setAttr(
      kTfQuantCreatedEinsum,
      BoolAttr::get(reduced.getDefiningOp()->getContext(), true));

namespace mlir {
namespace TFL {
namespace {
#define GEN_PASS_DEF_OPTIMIZEOPORDERPASS
#include "tensorflow/compiler/mlir/lite/transforms/passes.h.inc"

// Dequantize ops will produce 3x larger tensors, so we want to move it after
// some passthrough ops to reduce the memory consumption.
struct PushDownDequantize : public OpRewritePattern<DequantizeOp> {
  explicit PushDownDequantize(MLIRContext* context)