[Dy2Stat]BugFix StaticAanlysis with gast.Subscript

[Aurelius84]

PR types

Bug fixes

PR changes

Others

Describe

1. PR内容

BugFix StaticAanlysis with gast.Subscript

2. 复现样例

在有关for-loop的语法转换中，依赖的变量如果是从上游tensor slice出来的，则在StaticAnalysis阶段无法被解析为Tensor类型。原因是因为StaticAnalysis中缺少对gast.Subscript语法的处理。此PR添加此类处理：如果node.value是个Variable，则slice出来的类型也是Variable，否则返回为statement类型

def test_list_append_in_for_subscript(x):
    x = fluid.dygraph.to_variable(x)
    iter_num = paddle.shape(x)[0]   # iter_num should be consider as Variable
    a = []
    for i in range(iter_num):
        x = x + 1
        a.append(x)
    out = paddle.concat(a)
    return out[0]

Before this PR, a = [] will not be converted because iter_num is not considered as Variable. The transformed codes is:

def test_list_append_in_for_subscript(x):
    x = paddle.assign(x)
    iter_num = paddle.shape(x)[0]
    a = []
    i = 0

    def for_loop_condition_0(i, iter_num, x):
        return i < iter_num

    def for_loop_body_0(i, iter_num, x):
        x = x + 1
        paddle.jit.dy2static.convert_call(a.append)(x)
        i += 1
        return i, iter_num, x
    [i, iter_num, x] = paddle.jit.dy2static.convert_while_loop(
        for_loop_condition_0, for_loop_body_0, [i, iter_num, x])
    out = paddle.concat(a)
    return out[0]

After this PR， The transformed codes is as expected:

def test_list_append_in_for_subscript(x):
    x = paddle.assign(x)
    iter_num = paddle.shape(x)[0]
    a = paddle.tensor.create_array(dtype='float32')
    i = 0

    def for_loop_condition_0(iter_num, a, i, x):
        return i < iter_num

    def for_loop_body_0(iter_num, a, i, x):
        x = x + 1
        paddle.tensor.array_write(x=x, i=paddle.tensor.array_length(a), array=a
            )
        i += 1
        return iter_num, a, i, x
    [iter_num, a, i, x] = paddle.jit.dy2static.convert_while_loop(
        for_loop_condition_0, for_loop_body_0, [iter_num, a, i, x])
    out = paddle.concat(a)
    return out[0]

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhhsplendid]

LGTM