uint32_t num_replicas = replicate.getN();
  llvm::SmallVector<int32_t, 4> group_assignment_val;
  for (int i = 0; i < num_replicas; ++i) group_assignment_val.push_back(i);
  Value group_assignment = builder.create<ConstOp>(
      block_arg.getLoc(),
      DenseIntElementsAttr::get(
          RankedTensorType::get({1, num_replicas}, builder.getIntegerType(32)),
          group_assignment_val));

  // then update the OperandSegmentSize attribute.
  if (send_op) {
    int32_t operand_sizes[] = {static_cast<int32_t>(send_op.getN()),
                               static_cast<int32_t>(send_op.getNN()), 1};
    auto operand_size_attr = builder.getDenseI32ArrayAttr(operand_sizes);
    NamedAttrList attrs(send_op->getAttrs());

        op->getLoc(),
        RankedTensorType::get(permuted_shape, input_type.getElementType()),
        input, permutation_tensor_op.getResult());
  };

  if (!can_convert_to_fully_connected(op)) {
    return false;
  }

  Value input_lhs = get_real_input_value(op.getX());
  Value input_rhs = get_real_input_value(op.getY());

  Value legalized_lhs =

		//	38	eattr[4]
		//	42	off[4]
		//	46	name[namelen]
		//	46+namelen+xlen+fclen - next header
		//
		if get4(buf) != zcheader {
			break
		}
		meth := get2(buf[10:])
		size := get4(buf[24:])
		namelen := get2(buf[28:])
		xlen := get2(buf[30:])
		fclen := get2(buf[32:])
		off := get4(buf[42:])
		zname := buf[46 : 46+namelen]
		buf = buf[46+namelen+xlen+fclen:]
		if string(zname) != name {
			continue

	CmdGet.Run = runGet // break init loop
	CmdGet.Flag.Var(&getD, "d", "")
	CmdGet.Flag.Var(&getU, "u", "")
}

func runGet(ctx context.Context, cmd *base.Command, args []string) {
	switch getU.version {
	case "", "upgrade", "patch":
		// ok
	default:
		base.Fatalf("go: unknown upgrade flag -u=%s", getU.rawVersion)
	}
	if getD.set {
		if !getD.value {
			base.Fatalf("go: -d flag may not be set to false")
		}

    auto l = mlir::dyn_cast<TypedValue<RankedTensorType>>(op.getX());
    auto r = mlir::dyn_cast<TypedValue<RankedTensorType>>(op.getY());
    if (!l || !r) return failure();

    auto element_type = getElementTypeOrSelf(l.getType());
    if (!element_type.isBF16()) return failure();

    auto out_type = op.getZ().getType();

    l = rewriter.create<ConvertOp>(op.getLoc(), l, rewriter.getF32Type());

//
//go:linkname runtime_setProfLabel runtime/pprof.runtime_setProfLabel
func runtime_setProfLabel(labels unsafe.Pointer) {
	// Introduce race edge for read-back via profile.
	// This would more properly use &getg().labels as the sync address,
	// but we do the read in a signal handler and can't call the race runtime then.
	//
	// This uses racereleasemerge rather than just racerelease so

Carlos Amedee <******@****.***> 1715265901 -0400

  compressDashboard "pilot-dashboard.gen.json"
  compressDashboard "istio-performance-dashboard.json"
  compressDashboard "istio-workload-dashboard.json"
  compressDashboard "istio-service-dashboard.json"
  compressDashboard "istio-mesh-dashboard.json"
  compressDashboard "istio-extension-dashboard.json"
  compressDashboard "ztunnel-dashboard.gen.json"
  echo -e "\n---\n"

}

func notetsleep(n *note, ns int64) bool {
	gp := getg()
	if gp != gp.m.g0 {
		throw("notetsleep not on g0")
	}
	semacreate(gp.m)
	return notetsleep_internal(n, ns, nil, 0)
}

// same as runtime·notetsleep, but called on user g (not g0)
// calls only nosplit functions between entersyscallblock/exitsyscall.
func notetsleepg(n *note, ns int64) bool {
	gp := getg()
	if gp == gp.m.g0 {
		throw("notetsleepg on g0")