def init_specific(self, context, builder, arrty, arr):
                zero = context.get_constant(types.intp, 0)
                data = arr.data
                ndim = arrty.ndim
                shapes = cgutils.unpack_tuple(builder, arr.shape, ndim)

                indices = cgutils.alloca_once(builder, zero.type,
                                              size=context.get_constant(types.intp,
                                                                        arrty.ndim))
                pointers = cgutils.alloca_once(builder, data.type,
                                               size=context.get_constant(types.intp,
                                                                         arrty.ndim))
                strides = cgutils.unpack_tuple(builder, arr.strides, ndim)
                exhausted = cgutils.alloca_once_value(builder, cgutils.false_byte)

                # Initialize indices and pointers with their start values.
                for dim in range(ndim):
                    idxptr = cgutils.gep(builder, indices, dim)
                    ptrptr = cgutils.gep(builder, pointers, dim)
                    builder.store(data, ptrptr)
                    builder.store(zero, idxptr)
                    # 0-sized dimensions really indicate an empty array,
                    # but we have to catch that condition early to avoid
                    # a bug inside the iteration logic (see issue #846).
                    dim_size = shapes[dim]
                    dim_is_empty = builder.icmp(lc.ICMP_EQ, dim_size, zero)
                    with cgutils.if_unlikely(builder, dim_is_empty):
                        builder.store(cgutils.true_byte, exhausted)

                self.indices = indices
                self.pointers = pointers
                self.exhausted = exhausted

def getitem_arraynd_intp(context, builder, sig, args):
    aryty, idxty = sig.args
    ary, idx = args
    arystty = make_array(aryty)
    adapted_ary = arystty(context, builder, ary)
    ndim = aryty.ndim
    if ndim == 1:
        result = _getitem_array1d(context, builder, aryty, adapted_ary, idx,
                                  wraparound=idxty.signed)
    elif ndim > 1:
        out_ary_ty = make_array(aryty.copy(ndim = ndim - 1))
        out_ary = out_ary_ty(context, builder)
        in_shapes = cgutils.unpack_tuple(builder, adapted_ary.shape, count=ndim)
        in_strides = cgutils.unpack_tuple(builder, adapted_ary.strides,
                                          count=ndim)
        data_p = cgutils.get_item_pointer2(builder, adapted_ary.data, in_shapes,
                                           in_strides, aryty.layout, [idx],
                                           wraparound=idxty.signed)
        populate_array(out_ary,
                       data=data_p,
                       shape=cgutils.pack_array(builder, in_shapes[1:]),
                       strides=cgutils.pack_array(builder, in_strides[1:]),
                       itemsize=adapted_ary.itemsize,
                       parent=adapted_ary.parent,)
        result = out_ary._getvalue()
    else:
        raise NotImplementedError("1D indexing into %dD array" % aryty.ndim)
    return result

def dot_3_vm(context, builder, sig, args):
    """
    np.dot(vector, matrix, out)
    np.dot(matrix, vector, out)
    """
    xty, yty, outty = sig.args
    assert outty == sig.return_type
    dtype = xty.dtype

    x = make_array(xty)(context, builder, args[0])
    y = make_array(yty)(context, builder, args[1])
    out = make_array(outty)(context, builder, args[2])
    x_shapes = cgutils.unpack_tuple(builder, x.shape)
    y_shapes = cgutils.unpack_tuple(builder, y.shape)
    out_shapes = cgutils.unpack_tuple(builder, out.shape)
    if xty.ndim < yty.ndim:
        # Vector * matrix
        # Asked for x * y, we will compute y.T * x
        mty = yty
        m_shapes = y_shapes
        do_trans = yty.layout == 'F'
        m_data, v_data = y.data, x.data

        def check_args(a, b, out):
            m, = a.shape
            _m, n = b.shape
            if m != _m:
                raise ValueError("incompatible array sizes for "
                                 "np.dot(a, b) (vector * matrix)")
            if out.shape != (n,):
                raise ValueError("incompatible output array size for "
                                 "np.dot(a, b, out) (vector * matrix)")
    else:
        # Matrix * vector
        # We will compute x * y
        mty = xty
        m_shapes = x_shapes
        do_trans = xty.layout == 'C'
        m_data, v_data = x.data, y.data

        def check_args(a, b, out):
            m, _n = a.shape
            n, = b.shape
            if n != _n:
                raise ValueError("incompatible array sizes for np.dot(a, b) "
                                 "(matrix * vector)")
            if out.shape != (m,):
                raise ValueError("incompatible output array size for "
                                 "np.dot(a, b, out) (matrix * vector)")

    context.compile_internal(builder, check_args,
                             signature(types.none, *sig.args), args)
    for val in m_shapes:
        check_c_int(context, builder, val)

    call_xxgemv(context, builder, do_trans, mty, m_shapes, m_data,
                v_data, out.data)

    return impl_ret_borrowed(context, builder, sig.return_type,
                             out._getvalue())

    def impl(context, builder, sig, args):
        [tyinp, tyout] = sig.args
        [inp, out] = args
        if scalar_input:
            ndim = 1
        else:
            ndim = tyinp.ndim

        if not scalar_input:
            iary = context.make_array(tyinp)(context, builder, inp)
        oary = context.make_array(tyout)(context, builder, out)

        if asfloat:
            sig = typing.signature(types.float64, types.float64)
        else:
            if scalar_input:
                sig = typing.signature(tyout.dtype, tyinp)
            else:
                sig = typing.signature(tyout.dtype, tyinp.dtype)

        fnwork = context.get_function(funckey, sig)
        intpty = context.get_value_type(types.intp)

        # TODO handle differing shape by mimicking broadcasting
        if scalar_input:
            shape = cgutils.unpack_tuple(builder, oary.shape, ndim)
        else:
            shape = cgutils.unpack_tuple(builder, iary.shape, ndim)
        with cgutils.loop_nest(builder, shape, intp=intpty) as indices:
            if not scalar_input:
                pi = cgutils.get_item_pointer(builder, tyinp, iary, indices)
            po = cgutils.get_item_pointer(builder, tyout, oary, indices)

            if scalar_input:
                ival = inp
            else:
                ival = builder.load(pi)
            if asfloat:
                if scalar_input:
                    dval = context.cast(builder, ival, tyinp, types.float64)
                else:
                    dval = context.cast(builder, ival, tyinp.dtype, types.float64)
                dres = fnwork(builder, [dval])
                res = context.cast(builder, dres, types.float64, tyout.dtype)
            elif scalar_input and tyinp != tyout:
                tempres = fnwork(builder, [ival])
                res = context.cast(builder, tempres, tyinp, tyout.dtype)
            elif tyinp.dtype != tyout.dtype:
                tempres = fnwork(builder, [ival])
                res = context.cast(builder, tempres, tyinp.dtype, tyout.dtype)
            else:
                res = fnwork(builder, [ival])
            builder.store(res, po)

        return out

def array_nonzero(context, builder, sig, args):
    aryty = sig.args[0]
    # Return type is a N-tuple of 1D C-contiguous arrays
    retty = sig.return_type
    outaryty = retty.dtype
    ndim = aryty.ndim
    nouts = retty.count

    ary = make_array(aryty)(context, builder, args[0])
    shape = cgutils.unpack_tuple(builder, ary.shape)
    strides = cgutils.unpack_tuple(builder, ary.strides)
    data = ary.data
    layout = aryty.layout

    # First count the number of non-zero elements
    zero = context.get_constant(types.intp, 0)
    one = context.get_constant(types.intp, 1)
    count = cgutils.alloca_once_value(builder, zero)
    with cgutils.loop_nest(builder, shape, zero.type) as indices:
        ptr = cgutils.get_item_pointer2(builder, data, shape, strides,
                                        layout, indices)
        val = load_item(context, builder, aryty, ptr)
        nz = context.is_true(builder, aryty.dtype, val)
        with builder.if_then(nz):
            builder.store(builder.add(builder.load(count), one), count)

    # Then allocate output arrays of the right size
    out_shape = (builder.load(count),)
    outs = [_empty_nd_impl(context, builder, outaryty, out_shape)._getvalue()
            for i in range(nouts)]
    outarys = [make_array(outaryty)(context, builder, out) for out in outs]
    out_datas = [out.data for out in outarys]

    # And fill them up
    index = cgutils.alloca_once_value(builder, zero)
    with cgutils.loop_nest(builder, shape, zero.type) as indices:
        ptr = cgutils.get_item_pointer2(builder, data, shape, strides,
                                        layout, indices)
        val = load_item(context, builder, aryty, ptr)
        nz = context.is_true(builder, aryty.dtype, val)
        with builder.if_then(nz):
            # Store element indices in output arrays
            if not indices:
                # For a 0-d array, store 0 in the unique output array
                indices = (zero,)
            cur = builder.load(index)
            for i in range(nouts):
                ptr = cgutils.get_item_pointer2(builder, out_datas[i],
                                                out_shape, (),
                                                'C', [cur])
                store_item(context, builder, outaryty, indices[i], ptr)
            builder.store(builder.add(cur, one), index)

    tup = context.make_tuple(builder, sig.return_type, outs)
    return impl_ret_new_ref(context, builder, sig.return_type, tup)

def getitem_array_tuple(context, builder, sig, args):
    aryty, idxty = sig.args
    ary, idx = args

    arystty = make_array(aryty)
    ary = arystty(context, builder, ary)

    ndim = aryty.ndim
    if isinstance(sig.return_type, types.Array):
        # Slicing
        raw_indices = cgutils.unpack_tuple(builder, idx, aryty.ndim)
        start = []
        shapes = []
        strides = []

        oshapes = cgutils.unpack_tuple(builder, ary.shape, ndim)
        for ax, (indexval, idxty) in enumerate(zip(raw_indices, idxty)):
            if idxty == types.slice3_type:
                slice = Slice(context, builder, value=indexval)
                cgutils.normalize_slice(builder, slice, oshapes[ax])
                start.append(slice.start)
                shapes.append(cgutils.get_range_from_slice(builder, slice))
                strides.append(cgutils.get_strides_from_slice(builder, ndim,
                                                              ary.strides,
                                                              slice, ax))
            else:
                ind = context.cast(builder, indexval, idxty, types.intp)
                start.append(ind)

        dataptr = cgutils.get_item_pointer(builder, aryty, ary, start,
                                           wraparound=True)
        # Build array
        retstty = make_array(sig.return_type)
        retary = retstty(context, builder)
        populate_array(retary,
                       data=dataptr,
                       shape=cgutils.pack_array(builder, shapes),
                       strides=cgutils.pack_array(builder, strides),
                       itemsize=ary.itemsize,
                       meminfo=ary.meminfo,
                       parent=ary.parent)
        return retary._getvalue()
    else:
        # Indexing
        indices = cgutils.unpack_tuple(builder, idx, count=len(idxty))
        indices = [context.cast(builder, i, t, types.intp)
                   for t, i in zip(idxty, indices)]
        ptr = cgutils.get_item_pointer(builder, aryty, ary, indices,
                                       wraparound=True)

        return context.unpack_value(builder, aryty.dtype, ptr)

def _define_atomic_cas(module, ordering):
    """Define a llvm function for atomic compare-and-swap.
    The generated function is a direct wrapper of the LLVM cmpxchg with the
    difference that the a int indicate success (1) or failure (0) is returned
    and the last argument is a output pointer for storing the old value.

    Note
    ----
    On failure, the generated function behaves like an atomic load.  The loaded
    value is stored to the last argument.
    """
    ftype = ir.FunctionType(ir.IntType(32), [_word_type.as_pointer(),
                                             _word_type, _word_type,
                                             _word_type.as_pointer()])
    fn_cas = ir.Function(module, ftype, name="nrt_atomic_cas")

    [ptr, cmp, repl, oldptr] = fn_cas.args
    bb = fn_cas.append_basic_block()
    builder = ir.IRBuilder(bb)
    outtup = builder.cmpxchg(ptr, cmp, repl, ordering=ordering)
    old, ok = cgutils.unpack_tuple(builder, outtup, 2)
    builder.store(old, oldptr)
    builder.ret(builder.zext(ok, ftype.return_type))

    return fn_cas

def _box_class_instance(typ, val, c):
    meminfo, dataptr = cgutils.unpack_tuple(c.builder, val)

    # Create Box instance
    box_subclassed = _specialize_box(typ)
    # Note: the ``box_subclassed`` is kept alive by the cache
    int_addr_boxcls = c.context.get_constant(types.uintp, id(box_subclassed))

    box_cls = c.builder.inttoptr(int_addr_boxcls, c.pyapi.pyobj)
    box = c.pyapi.call_function_objargs(box_cls, ())

    # Initialize Box instance
    llvoidptr = ir.IntType(8).as_pointer()
    addr_meminfo = c.builder.bitcast(meminfo, llvoidptr)
    addr_data = c.builder.bitcast(dataptr, llvoidptr)

    def set_member(member_offset, value):
        # Access member by byte offset
        offset = c.context.get_constant(types.uintp, member_offset)
        ptr = cgutils.pointer_add(c.builder, box, offset)
        casted = c.builder.bitcast(ptr, llvoidptr.as_pointer())
        c.builder.store(value, casted)

    set_member(_box.box_meminfoptr_offset, addr_meminfo)
    set_member(_box.box_dataptr_offset, addr_data)
    return box

def ptx_atomic_add_tuple(context, builder, sig, args):
    aryty, indty, valty = sig.args
    ary, inds, val = args
    dtype = aryty.dtype

    indices = cgutils.unpack_tuple(builder, inds, count=len(indty))
    indices = [context.cast(builder, i, t, types.intp) for t, i in zip(indty, indices)]

    if dtype != valty:
        raise TypeError("expect %s but got %s" % (dtype, valty))

    if aryty.ndim != len(indty):
        raise TypeError("indexing %d-D array with %d-D index" % (aryty.ndim, len(indty)))

    lary = context.make_array(aryty)(context, builder, ary)
    ptr = cgutils.get_item_pointer(builder, aryty, lary, indices)

    if aryty.dtype == types.float32:
        lmod = builder.module
        return builder.call(nvvmutils.declare_atomic_add_float32(lmod), (ptr, val))
    elif aryty.dtype == types.float64:
        lmod = builder.module
        return builder.call(nvvmutils.declare_atomic_add_float64(lmod), (ptr, val))
    else:
        return builder.atomic_rmw("add", ptr, val, "monotonic")

def _gauss_impl(context, builder, sig, args, state):
    # The type for all computations (either float or double)
    ty = sig.return_type
    llty = context.get_data_type(ty)

    state_ptr = get_state_ptr(context, builder, state)
    _random = {"py": random.random,
               "np": np.random.random}[state]

    ret = cgutils.alloca_once(builder, llty, name="result")

    gauss_ptr = get_gauss_ptr(builder, state_ptr)
    has_gauss_ptr = get_has_gauss_ptr(builder, state_ptr)
    has_gauss = cgutils.is_true(builder, builder.load(has_gauss_ptr))
    with builder.if_else(has_gauss) as (then, otherwise):
        with then:
            # if has_gauss: return it
            builder.store(builder.load(gauss_ptr), ret)
            builder.store(const_int(0), has_gauss_ptr)
        with otherwise:
            # if not has_gauss: compute a pair of numbers using the Box-Muller
            # transform; keep one and return the other
            pair = context.compile_internal(builder,
                                            _gauss_pair_impl(_random),
                                            signature(types.UniTuple(ty, 2)),
                                            ())

            first, second = cgutils.unpack_tuple(builder, pair, 2)
            builder.store(first, gauss_ptr)
            builder.store(second, ret)
            builder.store(const_int(1), has_gauss_ptr)

    mu, sigma = args
    return builder.fadd(mu,
                        builder.fmul(sigma, builder.load(ret)))

def dot_2_vv(context, builder, sig, args, conjugate=False):
    """
    np.dot(vector, vector)
    np.vdot(vector, vector)
    """
    aty, bty = sig.args
    dtype = sig.return_type
    a = make_array(aty)(context, builder, args[0])
    b = make_array(bty)(context, builder, args[1])
    n, = cgutils.unpack_tuple(builder, a.shape)

    def check_args(a, b):
        m, = a.shape
        n, = b.shape
        if m != n:
            raise ValueError("incompatible array sizes for np.dot(a, b) "
                             "(vector * vector)")

    context.compile_internal(builder, check_args,
                             signature(types.none, *sig.args), args)
    check_c_int(context, builder, n)

    out = cgutils.alloca_once(builder, context.get_value_type(dtype))
    call_xxdot(context, builder, conjugate, dtype, n, a.data, b.data, out)
    return builder.load(out)

def getitem_array1d(context, builder, sig, args):
    aryty, _ = sig.args
    if aryty.ndim != 1:
        # TODO
        raise NotImplementedError("1D indexing into %dD array" % aryty.ndim)

    ary, idx = args

    arystty = make_array(aryty)
    ary = arystty(context, builder, ary)
    dataptr = ary.data

    if True or WARPAROUND:  # TODO target flag
        ZERO = context.get_constant(types.intp, 0)
        negative = builder.icmp(lc.ICMP_SLT, idx, ZERO)
        bbnormal = builder.basic_block
        with cgutils.if_unlikely(builder, negative):
            # Index is negative, wraparound
            [nelem] = cgutils.unpack_tuple(builder, ary.shape, 1)
            wrapped = builder.add(nelem, idx)
            bbwrapped = builder.basic_block

        where = builder.phi(idx.type)
        where.add_incoming(idx, bbnormal)
        where.add_incoming(wrapped, bbwrapped)

        ptr = builder.gep(dataptr, [where])
    else:
        # No wraparound
        ptr = builder.gep(dataptr, [idx])

    if context.is_struct_type(aryty.dtype):
        return ptr
    else:
        return builder.load(ptr)

def _box_class_instance(typ, val, c):
    meminfo, dataptr = cgutils.unpack_tuple(c.builder, val)

    lluintp = c.context.get_data_type(types.uintp)

    addr_meminfo = c.pyapi.from_native_value(types.uintp,
                                             c.builder.ptrtoint(meminfo,
                                                                lluintp))
    addr_dataptr = c.pyapi.from_native_value(types.uintp,
                                             c.builder.ptrtoint(dataptr,
                                                                lluintp))

    box_subclassed = _specialize_box(typ)
    # Note: the ``box_subclassed`` is kept alive by the cache
    int_addr_boxcls = c.context.get_constant(types.uintp, id(box_subclassed))

    box_cls = c.builder.inttoptr(int_addr_boxcls, c.pyapi.pyobj)

    args = [addr_meminfo, addr_dataptr]
    res = c.pyapi.call_function_objargs(box_cls, args)

    # Clean up
    c.pyapi.decref(addr_meminfo)
    c.pyapi.decref(addr_dataptr)

    return res

def array_nbytes(context, builder, typ, value):
    """
    nbytes = size * itemsize
    """
    arrayty = make_array(typ)
    array = arrayty(context, builder, value)
    dims = cgutils.unpack_tuple(builder, array.shape, typ.ndim)
    return builder.mul(array.nitems, array.itemsize)

 def codegen(context, builder, signature, args):
     # check that the return type is now defined
     arrty = signature.return_type
     assert arrty.is_precise()
     shapes = unpack_tuple(builder, args[0])
     # redirect implementation to np.empty
     res = _empty_nd_impl(context, builder, arrty, shapes)
     return impl_ret_new_ref(context, builder, arrty, res._getvalue())

def getitem_array_unituple(context, builder, sig, args):
    aryty, idxty = sig.args
    ary, idx = args

    ndim = aryty.ndim
    arystty = make_array(aryty)
    ary = arystty(context, builder, ary)

    if idxty.dtype == types.slice3_type:
        # Slicing
        raw_slices = cgutils.unpack_tuple(builder, idx, aryty.ndim)
        slices = [Slice(context, builder, value=sl) for sl in raw_slices]
        for sl, sh in zip(slices,
                          cgutils.unpack_tuple(builder, ary.shape, ndim)):
            cgutils.normalize_slice(builder, sl, sh)
        indices = [sl.start for sl in slices]
        dataptr = cgutils.get_item_pointer(builder, aryty, ary, indices,
                                           wraparound=True)
        # Build array
        retstty = make_array(sig.return_type)
        retary = retstty(context, builder)
        shapes = [cgutils.get_range_from_slice(builder, sl)
                  for sl in slices]
        strides = [cgutils.get_strides_from_slice(builder, ndim, ary.strides,
                                                  sl, i)
                   for i, sl in enumerate(slices)]
        populate_array(retary,
                       data=dataptr,
                       shape=cgutils.pack_array(builder, shapes),
                       strides=cgutils.pack_array(builder, strides),
                       itemsize=ary.itemsize,
                       meminfo=ary.meminfo,
                       parent=ary.parent)
        return retary._getvalue()
    else:
        # Indexing
        assert isinstance(idxty.dtype, types.Integer)
        indices = cgutils.unpack_tuple(builder, idx, count=len(idxty))
        indices = [context.cast(builder, i, t, types.intp)
                   for t, i in zip(idxty, indices)]
        ptr = cgutils.get_item_pointer(builder, aryty, ary, indices,
                                       wraparound=idxty.dtype.signed)

        return context.unpack_value(builder, aryty.dtype, ptr)

def iternext_numpy_flatiter(context, builder, sig, args, result):
    [flatiterty] = sig.args
    [flatiter] = args

    flatitercls = make_array_flat_cls(flatiterty)
    flatiter = flatitercls(context, builder, value=flatiter)

    arrty = flatiterty.array_type
    arrcls = context.make_array(arrty)
    arr = arrcls(context, builder, value=builder.load(flatiter.array))

    ndim = arrty.ndim
    shapes = cgutils.unpack_tuple(builder, arr.shape, ndim)
    indptr = flatiter.iters

    # Load indices and check if they are valid
    indices = []
    is_valid = cgutils.true_bit
    zero = context.get_constant(types.intp, 0)
    one = context.get_constant(types.intp, 1)
    for ax in range(ndim):
        axsize = shapes[ax]
        idxptr = builder.gep(indptr, [context.get_constant(types.intp, ax)])
        idx = builder.load(idxptr)
        ax_valid = builder.icmp(lc.ICMP_SLT, idx, axsize)

        indices.append(idx)
        is_valid = builder.and_(is_valid, ax_valid)

    result.set_valid(is_valid)

    with cgutils.if_likely(builder, is_valid):
        # Get yielded value
        valptr = cgutils.get_item_pointer(builder, arrty, arr, indices)
        yield_value = builder.load(valptr)
        result.yield_(yield_value)

        # Increment iterator indices
        carry_flags = [cgutils.true_bit]
        for ax, (idx, axsize) in reversed(list(enumerate(zip(indices,
                                                             shapes)))):
            idxptr = builder.gep(indptr, [context.get_constant(types.intp, ax)])
            lastcarry = carry_flags[-1]
            idxp1 = builder.add(idx, one)
            carry = builder.icmp(lc.ICMP_SGE, idxp1, axsize)
            idxfinal = builder.select(lastcarry,
                                      builder.select(carry, zero, idxp1),
                                      idx)
            builder.store(idxfinal, idxptr)
            carry_flags.append(builder.and_(carry, lastcarry))

        with cgutils.if_unlikely(builder, carry_flags[-1]):
            # If we have iterated all elements,
            # Set first index to out-of-bound
            idxptr = builder.gep(indptr, [context.get_constant(types.intp, 0)])
            builder.store(shapes[0], idxptr)

 def  codegen(context, builder, sig, args):
     (iterty,) = sig.args
     (value,) = args
     intp_t = context.get_value_type(types.intp)
     iterobj = context.make_helper(builder, iterty, value=value)
     arrayty = iterty.array_type
     ary = make_array(arrayty)(context, builder, value=iterobj.array)
     shape = cgutils.unpack_tuple(builder, ary.shape)
     # array iterates along the outer dimension
     return impl_ret_untracked(context, builder, intp_t, shape[0])

def _normalize_indices(context, builder, indty, inds):
    """
    Convert integer indices into tuple of intp
    """
    if indty in types.integer_domain:
        indty = types.UniTuple(dtype=indty, count=1)
        indices = [inds]
    else:
        indices = cgutils.unpack_tuple(builder, inds, count=len(indty))
    indices = [context.cast(builder, i, t, types.intp)
               for t, i in zip(indty, indices)]
    return indty, indices

def numpy_empty_nd(context, builder, sig, args):
    arrshapetype = sig.args[0]
    arrshape = args[0]
    arrtype = sig.return_type

    if isinstance(arrshapetype, types.Integer):
        shapes = [context.cast(builder, arrshape, arrshapetype, types.intp)]
    else:
        arrshape = context.cast(builder, arrshape, arrshapetype,
                                types.UniTuple(types.intp, len(arrshapetype)))
        shapes = cgutils.unpack_tuple(builder, arrshape,
                                      count=len(arrshapetype))

    return _empty_nd_impl(context, builder, arrtype, shapes)