def timedelta_sign_impl(context, builder, sig, args):
    val, = args
    ret = alloc_timedelta_result(builder)
    zero = Constant.int(TIMEDELTA64, 0)
    with cgutils.ifelse(builder, builder.icmp(lc.ICMP_SGT, val, zero)
                        ) as (gt_zero, le_zero):
        with gt_zero:
            builder.store(Constant.int(TIMEDELTA64, 1), ret)
        with le_zero:
            with cgutils.ifelse(builder, builder.icmp(lc.ICMP_EQ, val, zero)
                                ) as (eq_zero, lt_zero):
                with eq_zero:
                    builder.store(Constant.int(TIMEDELTA64, 0), ret)
                with lt_zero:
                    builder.store(Constant.int(TIMEDELTA64, -1), ret)
    return builder.load(ret)

def getiter_range_generic(context, builder, iterobj, start, stop, step):
    diff = builder.sub(stop, start)
    intty = start.type
    zero = Constant.int(intty, 0)
    one = Constant.int(intty, 1)
    pos_diff = builder.icmp(lc.ICMP_SGT, diff, zero)
    pos_step = builder.icmp(lc.ICMP_SGT, step, zero)
    sign_differs = builder.xor(pos_diff, pos_step)
    zero_step = builder.icmp(lc.ICMP_EQ, step, zero)

    with cgutils.if_unlikely(builder, zero_step):
        # step shouldn't be zero
        context.return_errcode(builder, 1)

    with cgutils.ifelse(builder, sign_differs) as (then, orelse):
        with then:
            builder.store(zero, iterobj.count)

        with orelse:
            rem = builder.srem(diff, step)
            uneven = builder.icmp(lc.ICMP_SGT, rem, zero)
            newcount = builder.add(builder.sdiv(diff, step),
                                   builder.select(uneven, one, zero))
            builder.store(newcount, iterobj.count)

    return iterobj._getvalue()

    def to_native_optional(self, obj, typ):
        """
        Convert object *obj* to a native optional structure.
        """
        noneval = self.context.make_optional_none(self.builder, typ.type)
        is_not_none = self.builder.icmp(lc.ICMP_NE, obj, self.borrow_none())

        retptr = cgutils.alloca_once(self.builder, noneval.type)
        errptr = cgutils.alloca_once_value(self.builder, cgutils.false_bit)

        with cgutils.ifelse(self.builder, is_not_none) as (then, orelse):
            with then:
                native = self.to_native_value(obj, typ.type)
                just = self.context.make_optional_value(self.builder,
                                                        typ.type, native.value)
                self.builder.store(just, retptr)
                self.builder.store(native.is_error, errptr)

            with orelse:
                self.builder.store(ir.Constant(noneval.type, ir.Undefined),
                                   retptr)
                self.builder.store(noneval, retptr)

        if native.cleanup is not None:
            def cleanup():
                with cgutils.ifthen(self.builder, is_not_none):
                    native.cleanup()
        else:
            cleanup = None

        ret = self.builder.load(retptr)
        return NativeValue(ret, is_error=self.builder.load(errptr),
                           cleanup=cleanup)

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 cgutils.ifelse(builder, 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 get_next_int(context, builder, state_ptr, nbits):
    """
    Get the next integer with width *nbits*.
    """
    c32 = ir.Constant(nbits.type, 32)
    def get_shifted_int(nbits):
        shift = builder.sub(c32, nbits)
        y = get_next_int32(context, builder, state_ptr)
        return builder.lshr(y, builder.zext(shift, y.type))

    ret = cgutils.alloca_once_value(builder, ir.Constant(int64_t, 0))

    is_32b = builder.icmp_unsigned('<=', nbits, c32)
    with cgutils.ifelse(builder, is_32b) as (ifsmall, iflarge):
        with ifsmall:
            low = get_shifted_int(nbits)
            builder.store(builder.zext(low, int64_t), ret)
        with iflarge:
            # XXX This assumes nbits <= 64
            low = get_next_int32(context, builder, state_ptr)
            high = get_shifted_int(builder.sub(nbits, c32))
            total = builder.add(
                builder.zext(low, int64_t),
                builder.shl(builder.zext(high, int64_t), ir.Constant(int64_t, 32)))
            builder.store(total, ret)

    return builder.load(ret)

    def _long_from_native_int(self, ival, func_name, native_int_type,
                              signed):
        fnty = Type.function(self.pyobj, [native_int_type])
        fn = self._get_function(fnty, name=func_name)
        resptr = cgutils.alloca_once(self.builder, self.pyobj)

        if PYVERSION < (3, 0):
            # Under Python 2, we try to return a PyInt object whenever
            # the given number fits in a C long.
            pyint_fnty = Type.function(self.pyobj, [self.long])
            pyint_fn = self._get_function(pyint_fnty, name="PyInt_FromLong")
            long_max = Constant.int(native_int_type, _helperlib.long_max)
            if signed:
                long_min = Constant.int(native_int_type, _helperlib.long_min)
                use_pyint = self.builder.and_(
                    self.builder.icmp(lc.ICMP_SGE, ival, long_min),
                    self.builder.icmp(lc.ICMP_SLE, ival, long_max),
                    )
            else:
                use_pyint = self.builder.icmp(lc.ICMP_ULE, ival, long_max)

            with cgutils.ifelse(self.builder, use_pyint) as (then, otherwise):
                with then:
                    downcast_ival = self.builder.trunc(ival, self.long)
                    res = self.builder.call(pyint_fn, [downcast_ival])
                    self.builder.store(res, resptr)
                with otherwise:
                    res = self.builder.call(fn, [ival])
                    self.builder.store(res, resptr)
        else:
            fn = self._get_function(fnty, name=func_name)
            self.builder.store(self.builder.call(fn, [ival]), resptr)

        return self.builder.load(resptr)

def complex128_power_impl(context, builder, sig, args):
    [ca, cb] = args
    a = Complex128(context, builder, value=ca)
    b = Complex128(context, builder, value=cb)
    c = Complex128(context, builder)
    module = cgutils.get_module(builder)
    pa = a._getpointer()
    pb = b._getpointer()
    pc = c._getpointer()

    # Optimize for square because cpow looses a lot of precsiion
    TWO = context.get_constant(types.float64, 2)
    ZERO = context.get_constant(types.float64, 0)

    b_real_is_two = builder.fcmp(lc.FCMP_OEQ, b.real, TWO)
    b_imag_is_zero = builder.fcmp(lc.FCMP_OEQ, b.imag, ZERO)
    b_is_two = builder.and_(b_real_is_two, b_imag_is_zero)

    with cgutils.ifelse(builder, b_is_two) as (then, otherwise):
        with then:
            # Lower as multiplication
            res = complex_mul_impl(context, builder, sig, (ca, ca))
            cres = Complex128(context, builder, value=res)
            c.real = cres.real
            c.imag = cres.imag

        with otherwise:
            # Lower with call to external function
            fnty = Type.function(Type.void(), [pa.type] * 3)
            cpow = module.get_or_insert_function(fnty, name="numba.math.cpow")
            builder.call(cpow, (pa, pb, pc))

    return builder.load(pc)

def timedelta_abs_impl(context, builder, sig, args):
    val, = args
    ret = alloc_timedelta_result(builder)
    with cgutils.ifelse(builder,
                        cgutils.is_scalar_neg(builder, val)) as (then, otherwise):
        with then:
            builder.store(builder.neg(val), ret)
        with otherwise:
            builder.store(val, ret)
    return builder.load(ret)

def year_to_days(builder, year_val):
    """
    Given a year *year_val* (offset to 1970), return the number of days
    since the 1970 epoch.
    """
    # The algorithm below is copied from Numpy's get_datetimestruct_days()
    # (src/multiarray/datetime.c)
    ret = cgutils.alloca_once(builder, TIMEDELTA64)
    # First approximation
    days = scale_by_constant(builder, year_val, 365)
    # Adjust for leap years
    with cgutils.ifelse(builder, cgutils.is_neg_int(builder, year_val)) \
        as (if_neg, if_pos):
        with if_pos:
            # At or after 1970:
            # 1968 is the closest leap year before 1970.
            # Exclude the current year, so add 1.
            from_1968 = add_constant(builder, year_val, 1)
            # Add one day for each 4 years
            p_days = builder.add(days,
                                 unscale_by_constant(builder, from_1968, 4))
            # 1900 is the closest previous year divisible by 100
            from_1900 = add_constant(builder, from_1968, 68)
            # Subtract one day for each 100 years
            p_days = builder.sub(p_days,
                                 unscale_by_constant(builder, from_1900, 100))
            # 1600 is the closest previous year divisible by 400
            from_1600 = add_constant(builder, from_1900, 300)
            # Add one day for each 400 years
            p_days = builder.add(p_days,
                                 unscale_by_constant(builder, from_1600, 400))
            builder.store(p_days, ret)
        with if_neg:
            # Before 1970:
            # NOTE `year_val` is negative, and so will be `from_1972` and `from_2000`.
            # 1972 is the closest later year after 1970.
            # Include the current year, so subtract 2.
            from_1972 = add_constant(builder, year_val, -2)
            # Subtract one day for each 4 years (`from_1972` is negative)
            n_days = builder.add(days,
                                 unscale_by_constant(builder, from_1972, 4))
            # 2000 is the closest later year divisible by 100
            from_2000 = add_constant(builder, from_1972, -28)
            # Add one day for each 100 years
            n_days = builder.sub(n_days,
                                 unscale_by_constant(builder, from_2000, 100))
            # 2000 is also the closest later year divisible by 400
            # Subtract one day for each 400 years
            n_days = builder.add(n_days,
                                 unscale_by_constant(builder, from_2000, 400))
            builder.store(n_days, ret)
    return builder.load(ret)

 def store(retval):
     is_error = cgutils.is_null(builder, retval)
     with cgutils.ifelse(builder, is_error) as (if_error, if_ok):
         with if_error:
             msg = context.insert_const_string(pyapi.module,
                                               "object mode ufunc")
             msgobj = pyapi.string_from_string(msg)
             pyapi.err_write_unraisable(msgobj)
             pyapi.decref(msgobj)
         with if_ok:
             # Unbox
             retval = pyapi.to_native_value(retval, signature.return_type)
             # Store
             out.store_direct(retval, builder.load(store_offset))

 def impl(context, builder, sig, args):
     [va, vb] = args
     [ta, tb] = sig.args
     ret = alloc_boolean_result(builder)
     with cgutils.ifelse(builder, are_not_nat(builder, [va, vb])) as (then, otherwise):
         with then:
             norm_a, norm_b = normalize_timedeltas(context, builder, va, vb, ta, tb)
             builder.store(builder.icmp(ll_op, norm_a, norm_b), ret)
         with otherwise:
             # No scaling when comparing NaT with something else
             # (i.e. NaT is <= everything else, since it's the smallest
             #  int64 value)
             builder.store(builder.icmp(ll_op, va, vb), ret)
     return builder.load(ret)

        def from_range_state(cls, context, builder, state):
            """
            Create a RangeIter initialized from the given RangeState *state*.
            """
            self = cls(context, builder)
            start = state.start
            stop = state.stop
            step = state.step

            startptr = cgutils.alloca_once(builder, start.type)
            builder.store(start, startptr)

            countptr = cgutils.alloca_once(builder, start.type)

            self.iter = startptr
            self.stop = stop
            self.step = step
            self.count = countptr

            diff = builder.sub(stop, start)
            zero = context.get_constant(int_type, 0)
            one = context.get_constant(int_type, 1)
            pos_diff = builder.icmp(lc.ICMP_SGT, diff, zero)
            pos_step = builder.icmp(lc.ICMP_SGT, step, zero)
            sign_differs = builder.xor(pos_diff, pos_step)
            zero_step = builder.icmp(lc.ICMP_EQ, step, zero)

            with cgutils.if_unlikely(builder, zero_step):
                # step shouldn't be zero
                context.call_conv.return_user_exc(builder, ValueError,
                                                  ("range() arg 3 must not be zero",))

            with cgutils.ifelse(builder, sign_differs) as (then, orelse):
                with then:
                    builder.store(zero, self.count)

                with orelse:
                    rem = builder.srem(diff, step)
                    rem = builder.select(pos_diff, rem, builder.neg(rem))
                    uneven = builder.icmp(lc.ICMP_SGT, rem, zero)
                    newcount = builder.add(builder.sdiv(diff, step),
                                           builder.select(uneven, one, zero))
                    builder.store(newcount, self.count)

            return self

def reduce_datetime_for_unit(builder, dt_val, src_unit, dest_unit):
    dest_unit_code = npdatetime.DATETIME_UNITS[dest_unit]
    src_unit_code = npdatetime.DATETIME_UNITS[src_unit]
    if dest_unit_code < 2 or src_unit_code >= 2:
        return dt_val, src_unit
    # Need to compute the day ordinal for *dt_val*
    if src_unit_code == 0:
        # Years to days
        year_val = dt_val
        days_val = year_to_days(builder, year_val)

    else:
        # Months to days
        leap_array = cgutils.global_constant(builder, "leap_year_months_acc",
                                             leap_year_months_acc)
        normal_array = cgutils.global_constant(builder, "normal_year_months_acc",
                                               normal_year_months_acc)

        days = cgutils.alloca_once(builder, TIMEDELTA64)

        # First compute year number and month number
        year, month = cgutils.divmod_by_constant(builder, dt_val, 12)

        # Then deduce the number of days
        with cgutils.ifelse(builder,
                            is_leap_year(builder, year)) as (then, otherwise):
            with then:
                addend = builder.load(cgutils.gep(builder, leap_array,
                                                  0, month))
                builder.store(addend, days)
            with otherwise:
                addend = builder.load(cgutils.gep(builder, normal_array,
                                                  0, month))
                builder.store(addend, days)

        days_val = year_to_days(builder, year)
        days_val = builder.add(days_val, builder.load(days))

    if dest_unit_code == 2:
        # Need to scale back to weeks
        weeks, _ = cgutils.divmod_by_constant(builder, days_val, 7)
        return weeks, 'W'
    else:
        return days_val, 'D'

    def lower_expr(self, expr):
        if expr.op == 'binop':
            return self.lower_binop(expr, inplace=False)
        elif expr.op == 'inplace_binop':
            return self.lower_binop(expr, inplace=True)
        elif expr.op == 'unary':
            value = self.loadvar(expr.value.name)
            if expr.fn == '-':
                res = self.pyapi.number_negative(value)
            elif expr.fn == '+':
                res = self.pyapi.number_positive(value)
            elif expr.fn == 'not':
                res = self.pyapi.object_not(value)
                self.check_int_status(res)

                longval = self.builder.zext(res, self.pyapi.long)
                res = self.pyapi.bool_from_long(longval)
            elif expr.fn == '~':
                res = self.pyapi.number_invert(value)
            else:
                raise NotImplementedError(expr)
            self.check_error(res)
            return res
        elif expr.op == 'call':
            argvals = [self.loadvar(a.name) for a in expr.args]
            fn = self.loadvar(expr.func.name)
            if not expr.kws:
                # No keyword
                ret = self.pyapi.call_function_objargs(fn, argvals)
            else:
                # Have Keywords
                keyvalues = [(k, self.loadvar(v.name)) for k, v in expr.kws]
                args = self.pyapi.tuple_pack(argvals)
                kws = self.pyapi.dict_pack(keyvalues)
                ret = self.pyapi.call(fn, args, kws)
                self.decref(kws)
                self.decref(args)
            self.check_error(ret)
            return ret
        elif expr.op == 'getattr':
            obj = self.loadvar(expr.value.name)
            res = self.pyapi.object_getattr(obj, self._freeze_string(expr.attr))
            self.check_error(res)
            return res
        elif expr.op == 'build_tuple':
            items = [self.loadvar(it.name) for it in expr.items]
            res = self.pyapi.tuple_pack(items)
            self.check_error(res)
            return res
        elif expr.op == 'build_list':
            items = [self.loadvar(it.name) for it in expr.items]
            res = self.pyapi.list_pack(items)
            self.check_error(res)
            return res
        elif expr.op == 'build_map':
            res = self.pyapi.dict_new(expr.size)
            self.check_error(res)
            return res
        elif expr.op == 'build_set':
            items = [self.loadvar(it.name) for it in expr.items]
            res = self.pyapi.set_new()
            self.check_error(res)
            for it in items:
                ok = self.pyapi.set_add(res, it)
                self.check_int_status(ok)
            return res
        elif expr.op == 'getiter':
            obj = self.loadvar(expr.value.name)
            res = self.pyapi.object_getiter(obj)
            self.check_error(res)
            return res
        elif expr.op == 'iternext':
            iterobj = self.loadvar(expr.value.name)
            item = self.pyapi.iter_next(iterobj)
            is_valid = cgutils.is_not_null(self.builder, item)
            pair = self.pyapi.tuple_new(2)
            with cgutils.ifelse(self.builder, is_valid) as (then, otherwise):
                with then:
                    self.pyapi.tuple_setitem(pair, 0, item)
                with otherwise:
                    self.check_occurred()
                    # Make the tuple valid by inserting None as dummy
                    # iteration "result" (it will be ignored).
                    self.pyapi.tuple_setitem(pair, 0, self.pyapi.make_none())
            self.pyapi.tuple_setitem(pair, 1, self.pyapi.bool_from_bool(is_valid))
            return pair
        elif expr.op == 'pair_first':
            pair = self.loadvar(expr.value.name)
            first = self.pyapi.tuple_getitem(pair, 0)
            self.incref(first)
            return first
        elif expr.op == 'pair_second':
            pair = self.loadvar(expr.value.name)
            second = self.pyapi.tuple_getitem(pair, 1)
            self.incref(second)
            return second
        elif expr.op == 'exhaust_iter':
            iterobj = self.loadvar(expr.value.name)
            tup = self.pyapi.sequence_tuple(iterobj)
            self.check_error(tup)
#.........这里部分代码省略.........

def build_ufunc_wrapper(library, context, func, signature, objmode, env):
    """
    Wrap the scalar function with a loop that iterates over the arguments
    """
    byte_t = Type.int(8)
    byte_ptr_t = Type.pointer(byte_t)
    byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
    intp_t = context.get_value_type(types.intp)
    intp_ptr_t = Type.pointer(intp_t)

    fnty = Type.function(Type.void(), [byte_ptr_ptr_t, intp_ptr_t,
                                       intp_ptr_t, byte_ptr_t])

    wrapper_module = library.create_ir_module('')
    if objmode:
        func_type = context.call_conv.get_function_type(
            types.pyobject, [types.pyobject] * len(signature.args))
    else:
        func_type = context.call_conv.get_function_type(
            signature.return_type, signature.args)
    oldfunc = func
    func = wrapper_module.add_function(func_type,
                                       name=func.name)
    func.attributes.add("alwaysinline")

    wrapper = wrapper_module.add_function(fnty, "__ufunc__." + func.name)
    arg_args, arg_dims, arg_steps, arg_data = wrapper.args
    arg_args.name = "args"
    arg_dims.name = "dims"
    arg_steps.name = "steps"
    arg_data.name = "data"

    builder = Builder.new(wrapper.append_basic_block("entry"))

    loopcount = builder.load(arg_dims, name="loopcount")

    # Prepare inputs
    arrays = []
    for i, typ in enumerate(signature.args):
        arrays.append(UArrayArg(context, builder, arg_args, arg_steps, i, typ))

    # Prepare output
    out = UArrayArg(context, builder, arg_args, arg_steps, len(arrays),
                    signature.return_type)

    # Setup indices
    offsets = []
    zero = context.get_constant(types.intp, 0)
    for _ in arrays:
        p = cgutils.alloca_once(builder, intp_t)
        offsets.append(p)
        builder.store(zero, p)

    store_offset = cgutils.alloca_once(builder, intp_t)
    builder.store(zero, store_offset)

    unit_strided = cgutils.true_bit
    for ary in arrays:
        unit_strided = builder.and_(unit_strided, ary.is_unit_strided)

    if objmode:
        # General loop
        pyapi = context.get_python_api(builder)
        gil = pyapi.gil_ensure()
        with cgutils.for_range(builder, loopcount, intp=intp_t):
            slowloop = build_obj_loop_body(context, func, builder,
                                           arrays, out, offsets,
                                           store_offset, signature,
                                           pyapi, env)
        pyapi.gil_release(gil)
        builder.ret_void()

    else:

        with cgutils.ifelse(builder, unit_strided) as (is_unit_strided,
                                                       is_strided):

            with is_unit_strided:
                with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
                    fastloop = build_fast_loop_body(context, func, builder,
                                                    arrays, out, offsets,
                                                    store_offset, signature,
                                                    ind)
                builder.ret_void()

            with is_strided:
                # General loop
                with cgutils.for_range(builder, loopcount, intp=intp_t):
                    slowloop = build_slow_loop_body(context, func, builder,
                                                    arrays, out, offsets,
                                                    store_offset, signature)

                builder.ret_void()

        builder.ret_void()
    del builder


    # Run optimizer
    library.add_ir_module(wrapper_module)
#.........这里部分代码省略.........

#.........这里部分代码省略.........
                        # If the shape of this dimension is 1, then leave the
                        # index at 0 so that this dimension is broadcasted over
                        # the corresponding input and output dimensions.
                        cond = builder.icmp(ICMP_UGT, inp_shape[i], ONE)
                        with cgutils.ifthen(builder, cond):
                            builder.store(indices[out_ndim-inp_ndim+i], inp_indices[i])
                        if i + 1 == inp_ndim:
                            builder.branch(bb_end_inc_index)
                        else:
                            builder.branch(bb_inc_inp_index[i+1])

                builder.position_at_end(bb_end_inc_index)

            if not scalar_inp1:
                build_increment_blocks(inp1_indices, inp1_shape, inp1_ndim, '1')
            if not scalar_inp2:
                build_increment_blocks(inp2_indices, inp2_shape, inp2_ndim, '2')

            if scalar_inp1:
                x = inp1
            else:
                inds = [builder.load(index) for index in inp1_indices]
                px = cgutils.get_item_pointer2(builder,
                                               data=inp1_data,
                                               shape=inp1_shape,
                                               strides=inp1_strides,
                                               layout=inp1_layout,
                                               inds=inds)
                x = builder.load(px)

            if scalar_inp2:
                y = inp2
            else:
                inds = [builder.load(index) for index in inp2_indices]
                py = cgutils.get_item_pointer2(builder,
                                               data=inp2_data,
                                               shape=inp2_shape,
                                               strides=inp2_strides,
                                               layout=inp2_layout,
                                               inds=inds)
                y = builder.load(py)

            po = cgutils.get_item_pointer2(builder,
                                           data=out_data,
                                           shape=out_shape,
                                           strides=out_strides,
                                           layout=out_layout,
                                           inds=indices)

            if divbyzero:
                # Handle division
                iszero = cgutils.is_scalar_zero(builder, y)
                with cgutils.ifelse(builder, iszero, expect=False) as (then,
                                                                       orelse):
                    with then:
                        # Divide by zero
                        if (scalar_tyinp1 in types.real_domain or
                                scalar_tyinp2 in types.real_domain) or \
                                not numpy_support.int_divbyzero_returns_zero:
                            # If y is float and is 0 also, return Nan; else
                            # return Inf
                            outltype = context.get_data_type(result_type)
                            shouldretnan = cgutils.is_scalar_zero(builder, x)
                            nan = Constant.real(outltype, float("nan"))
                            inf = Constant.real(outltype, float("inf"))
                            tempres = builder.select(shouldretnan, nan, inf)
                            res = context.cast(builder, tempres, result_type,
                                               tyout.dtype)
                        elif tyout.dtype in types.signed_domain and \
                                not numpy_support.int_divbyzero_returns_zero:
                            res = Constant.int(context.get_data_type(tyout.dtype),
                                               0x1 << (y.type.width-1))
                        else:
                            res = Constant.null(context.get_data_type(tyout.dtype))

                        assert res.type == po.type.pointee, \
                                        (str(res.type), str(po.type.pointee))
                        builder.store(res, po)
                    with orelse:
                        # Normal
                        d_x = context.cast(builder, x, scalar_tyinp1, promote_type)
                        d_y = context.cast(builder, y, scalar_tyinp2, promote_type)
                        tempres = fnwork(builder, [d_x, d_y])
                        res = context.cast(builder, tempres, result_type, tyout.dtype)

                        assert res.type == po.type.pointee, (res.type,
                                                             po.type.pointee)
                        builder.store(res, po)
            else:
                # Handle non-division operations
                d_x = context.cast(builder, x, scalar_tyinp1, promote_type)
                d_y = context.cast(builder, y, scalar_tyinp2, promote_type)
                tempres = fnwork(builder, [d_x, d_y])
                res = context.cast(builder, tempres, result_type, tyout.dtype)

                assert res.type == po.type.pointee, (res.type,
                                                     po.type.pointee)
                builder.store(res, po)

        return out

def build_ufunc_wrapper(context, func, signature):
    """
    Wrap the scalar function with a loop that iterates over the arguments
    """
    module = func.module

    byte_t = Type.int(8)
    byte_ptr_t = Type.pointer(byte_t)
    byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
    intp_t = context.get_value_type(types.intp)
    intp_ptr_t = Type.pointer(intp_t)

    fnty = Type.function(Type.void(), [byte_ptr_ptr_t, intp_ptr_t,
                                       intp_ptr_t, byte_ptr_t])

    wrapper = module.add_function(fnty, "__ufunc__." + func.name)
    arg_args, arg_dims, arg_steps, arg_data = wrapper.args
    arg_args.name = "args"
    arg_dims.name = "dims"
    arg_steps.name = "steps"
    arg_data.name = "data"

    builder = Builder.new(wrapper.append_basic_block("entry"))

    loopcount = builder.load(arg_dims, name="loopcount")

    actual_args = context.get_arguments(func)

    # Prepare inputs
    arrays = []
    for i, typ in enumerate(signature.args):
        arrays.append(UArrayArg(context, builder, arg_args, arg_steps, i,
                                context.get_argument_type(typ)))

    # Prepare output
    valty = context.get_data_type(signature.return_type)
    out = UArrayArg(context, builder, arg_args, arg_steps, len(actual_args),
                    valty)

    # Setup indices
    offsets = []
    zero = context.get_constant(types.intp, 0)
    for _ in arrays:
        p = cgutils.alloca_once(builder, intp_t)
        offsets.append(p)
        builder.store(zero, p)

    store_offset = cgutils.alloca_once(builder, intp_t)
    builder.store(zero, store_offset)

    unit_strided = cgutils.true_bit
    for ary in arrays:
        unit_strided = builder.and_(unit_strided, ary.is_unit_strided)

    with cgutils.ifelse(builder, unit_strided) as (is_unit_strided,
                                                   is_strided):

        with is_unit_strided:
            with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
                fastloop = build_fast_loop_body(context, func, builder,
                                                arrays, out, offsets,
                                                store_offset, signature, ind)
            builder.ret_void()

        with is_strided:
            # General loop
            with cgutils.for_range(builder, loopcount, intp=intp_t):
                slowloop = build_slow_loop_body(context, func, builder,
                                                arrays, out, offsets,
                                                store_offset, signature)

            builder.ret_void()

    builder.ret_void()
    del builder

    # Set core function to internal so that it is not generated
    func.linkage = LINKAGE_INTERNAL
    # Force inline of code function
    inline_function(slowloop)
    inline_function(fastloop)
    # Run optimizer
    context.optimize(module)

    if config.DUMP_OPTIMIZED:
        print(module)

    return wrapper

    def impl(context, builder, sig, args):
        [tyinp1, tyinp2, tyout] = sig.args
        [inp1, inp2, out] = args
        if scalar_inputs:
            ndim = 1
        else:
            ndim = tyinp1.ndim

        # Temporary hack for __ftol2 llvm bug. Don't allow storing
        # float results in uint64 array on windows.
        if scalar_inputs and tyinp1 in types.real_domain and \
                tyout.dtype is types.uint64 and \
                sys.platform.startswith('win32'):
            raise TypeError('Cannot store result in uint64 array')
        if not scalar_inputs and tyinp1.dtype in types.real_domain and \
                tyout.dtype is types.uint64 and \
                sys.platform.startswith('win32'):
            raise TypeError('Cannot store result in uint64 array')

        if not scalar_inputs:
            i1ary = context.make_array(tyinp1)(context, builder, inp1)
            i2ary = context.make_array(tyinp2)(context, builder, inp2)
        oary = context.make_array(tyout)(context, builder, out)

        if asfloat and not divbyzero:
            sig = typing.signature(types.float64, types.float64, types.float64)
        else:
            if scalar_inputs:
                sig = typing.signature(tyout.dtype, tyinp1, tyinp2)
            else:
                sig = typing.signature(tyout.dtype, tyinp1.dtype, tyinp2.dtype)

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

        # TODO handle differing shape by mimicking broadcasting
        loopshape = cgutils.unpack_tuple(builder, oary.shape, ndim)

        if scalar_inputs:
            xyo_shape = [cgutils.unpack_tuple(builder, ary.shape, ndim)
                         for ary in (oary,)]
            xyo_strides = [cgutils.unpack_tuple(builder, ary.strides, ndim)
                           for ary in (oary,)]
            xyo_data = [ary.data for ary in (oary,)]
            xyo_layout = [ty.layout for ty in (tyout,)]
        else:
            xyo_shape = [cgutils.unpack_tuple(builder, ary.shape, ndim)
                         for ary in (i1ary, i2ary, oary)]
            xyo_strides = [cgutils.unpack_tuple(builder, ary.strides, ndim)
                           for ary in (i1ary, i2ary, oary)]
            xyo_data = [ary.data for ary in (i1ary, i2ary, oary)]
            xyo_layout = [ty.layout for ty in (tyinp1, tyinp2, tyout)]

        with cgutils.loop_nest(builder, loopshape, intp=intpty) as indices:
            if scalar_inputs:
                [po] = [cgutils.get_item_pointer2(builder,
                                               data=data, shape=shape,
                                               strides=strides,
                                               layout=layout,
                                               inds=indices)
                                for data, shape, strides, layout
                                in zip(xyo_data, xyo_shape, xyo_strides,
                                       xyo_layout)]
            else:
                [px, py, po] = [cgutils.get_item_pointer2(builder,
                                                          data=data, shape=shape,
                                                          strides=strides,
                                                          layout=layout,
                                                          inds=indices)
                                for data, shape, strides, layout
                                in zip(xyo_data, xyo_shape, xyo_strides,
                                       xyo_layout)]

            if scalar_inputs:
                x = inp1
                y = inp2
            else:
                x = builder.load(px)
                y = builder.load(py)
            if divbyzero:
                # Handle division
                iszero = cgutils.is_scalar_zero(builder, y)
                with cgutils.ifelse(builder, iszero, expect=False) as (then,
                                                                       orelse):
                    with then:
                        # Divide by zero
                        if ((scalar_inputs and tyinp2 in types.real_domain) or
                                (not scalar_inputs and
                                    tyinp2.dtype in types.real_domain) or
                                not numpy_support.int_divbyzero_returns_zero):
                            # If y is float and is 0 also, return Nan; else
                            # return Inf
                            outltype = context.get_data_type(tyout.dtype)
                            shouldretnan = cgutils.is_scalar_zero(builder, x)
                            nan = Constant.real(outltype, float("nan"))
                            inf = Constant.real(outltype, float("inf"))
                            res = builder.select(shouldretnan, nan, inf)
                        elif (scalar_inputs and tyout in types.signed_domain and
                                not numpy_support.int_divbyzero_returns_zero):
                            res = Constant.int(context.get_data_type(tyout),
#.........这里部分代码省略.........