ab.Put1(byte(op))
				ab.asmand(ctxt, cursym, p, &p.From, &p.To)

			case Z_m_r:
				ab.Put1(byte(op))
				ab.asmand(ctxt, cursym, p, p.GetFrom3(), &p.To)

			case Zm2_r:
				ab.Put2(byte(op), o.op[z+1])
				ab.asmand(ctxt, cursym, p, &p.From, &p.To)

			case Zm_r_xm:
				ab.mediaop(ctxt, o, op, int(yt.zoffset), z)
				ab.asmand(ctxt, cursym, p, &p.From, &p.To)

			case Zm_r_xm_nr:

						falign = 32
					}
				}
			}

			pc += int64(m)
		}

		c.cursym.Size = pc
	}

	c.cursym.Size = pc
	c.cursym.Func().Align = falign
	c.cursym.Grow(c.cursym.Size)

	// lay out the code, emitting code and data relocations.

	bp := c.cursym.P
	var i int32
	for p := c.cursym.Func().Text.Link; p != nil; p = p.Link {
		c.pc = p.Pc
		o = c.oplook(p)

				}
			}
		}
		loop++
	}

	c.cursym.Size = int64(len(buffer))
	if c.cursym.Size%funcAlign != 0 {
		c.cursym.Size += funcAlign - (c.cursym.Size % funcAlign)
	}
	c.cursym.Grow(c.cursym.Size)
	copy(c.cursym.P, buffer)

	// Mark nonpreemptible instruction sequences.
	// We use REGTMP as a scratch register during call injection,
	// so instruction sequences that use REGTMP are unsafe to

				}
			}

			pc += int64(m)
		}
	}

	pc += -pc & (funcAlign - 1)
	c.cursym.Size = pc

	/*
	 * lay out the code, emitting code and data relocations.
	 */
	c.cursym.Grow(c.cursym.Size)
	bp := c.cursym.P
	psz := int32(0)
	var i int
	var out [6]uint32
	for p := c.cursym.Func().Text.Link; p != nil; p = p.Link {
		c.pc = p.Pc
		o = c.oplook(p)
		sz := o.size(c.ctxt, p)

  // CHECK-LABEL: tranpose_arg64
  // CHECK: "tfl.transpose"
}

func.func @cumsum(%arg0: tensor<3x3xf32>, %arg1: tensor<i32>) -> tensor<3x3xf32> {
  %0 = "tf.Cumsum"(%arg0, %arg1) {exclusive = false, reverse = false} : (tensor<3x3xf32>, tensor<i32>) -> tensor<3x3xf32>
  func.return %0 : tensor<3x3xf32>
  // CHECK-LABEL: cumsum

				vaTmp := va
				for j := i; j < len(ctxt.Textp); j++ {
					curSym := ctxt.Textp[j]
					if symPkg := ldr.SymPkg(curSym); symPkg == "" || curPkg != symPkg {
						break
					}
					// We do not pass big to assignAddress here, as this
					// can result in side effects such as section splitting.
					sect, n, vaTmp = assignAddress(ctxt, sect, n, curSym, vaTmp, false, false)

      return TF::ArraysAreCastCompatible(l, r);
    }

  }];
}

def TFL_CumsumOp: TFL_Op<"cumsum", [
    Pure,
    PredOpTrait<"input and output must have same element type",
      TFL_TCresVTEtIsSameAsOp<0, 0>>,
    TFL_OperandHasRank<1, 0>]> {
  let summary = "Cumsum operator";

  let description = [{
    Compute the cumulative sum of the tensor x along axis.
  }];