LDP	(R1), (R9, R10)
	ADDS	R4, R9
	MUL	R6, R3, R14
	ADCS	R14, R10
	MUL	R7, R3, R15
	LDP	16(R1), (R11, R12)
	ADCS	R15, R11
	MUL	R8, R3, R16
	ADCS	R16, R12
	UMULH	R8, R3, R20
	ADC	$0, R20

	MUL	R5, R3, R13
	ADDS	R13, R9
	UMULH	R5, R3, R17
	ADCS	R17, R10
	UMULH	R6, R3, R21
	STP.P	(R9, R10), 16(R1)
	ADCS	R21, R11
	UMULH	R7, R3, R19
	ADCS	R19, R12
	STP.P	(R11, R12), 16(R1)

This directory contains documentation that describes Gradle's architecture and how the various pieces fit together and work.

## Architecture decision records (ADRs)

The Gradle team uses ADRs to record architectural decisions that the team has made.

See [Architecture decisions records](standards) for the list of ADRs.

#include <functional>
#include <optional>

#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/Hashing.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringMap.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"

**Experimental!**

We'd like to capture our architectural decisions about the build tool as [Architectural Decision Records (ADRs)](https://adr.github.io/).
For now we just have this global repository of ADRs.
If we see fit, we can break these out to per-platform ones, or keep a hybrid approach to having global and platform-specific ADSs.

Our aim is to keep the process lightweight and approachable.

class RuntimeDevices {
  using DeviceNameUtils = ::tensorflow::DeviceNameUtils;
  using ParsedName = ::tensorflow::DeviceNameUtils::ParsedName;

 public:
  // Adds a device with and empty metadata. Device can be of any type.
  void AddDevice(const ParsedName& device);

  // Adds a GPU device with GPU specific metadata.
  void AddGpuDevice(const ParsedName& device,
                    const GpuDeviceMetadata& metadata);

#define TENSORFLOW_COMPILER_MLIR_QUANTIZATION_TENSORFLOW_PASSES_MANIPULATE_MODEL_ATTR_H_

#include "llvm/ADT/StringRef.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"  // from @llvm-project

namespace mlir {
namespace quant {

// Adds a new input name to the `inputs` field of the `tf.entry_function`
// attribute if the attribute exist in the given function. Otherwise, no
// attribute is modified.

limitations under the License.
==============================================================================*/
#include <memory>
#include <string>

#include "absl/strings/str_cat.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/ADT/StringSet.h"
#include "llvm/Support/Debug.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"  // from @llvm-project
#include "mlir/IR/Builders.h"  // from @llvm-project

// ops.
std::unique_ptr<mlir::OperationPass<mlir::ModuleOp>> CreateTPURewritePass(
    llvm::StringRef module_name = llvm::StringRef());

// Creates a pass that adds ops which perform formatting on variables at
// run-time according to compilation result.
std::unique_ptr<OperationPass<ModuleOp>>
CreateTPUVariableRuntimeReformattingPass();

// Creates a pass that merges device variable reads/updates into the surrounded

class Pass;

namespace mhlo {

/// Converter to be used along with the fallback Tf2Xla patterns below.
class Tf2XlaTypeConverter : public TypeConverter {
 public:
  Tf2XlaTypeConverter();
};

/// Adds the TF to XLA via TF2XLA rewrite patterns to the pattern list.
/// `prefer_tf2xla` means an op will be included iff it is not in
/// `MlirLegalizedUnderPreferTf2XlaSet`. `!prefer_tf2xla` mean an op will be

# job chain to GCS and PyPI.
source "${BASH_SOURCE%/*}/utilities/setup.sh"

# Update the version numbers for Nightly only, then fetch the version numbers
# for choosing the final directory name. This adds "-devYYYYMMDD" to the end of
# the version string (.devYYYYMMDD for Python; see pypi.org/project/tf-nightly)
if [[ "$TFCI_NIGHTLY_UPDATE_VERSION_ENABLE" == 1 ]]; then