expr-codegen 0.16.6__tar.gz → 0.17.0__tar.gz

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: expr_codegen
-Version: 0.16.6
+Version: 0.17.0
 Summary: symbol expression to polars expression tool
 Author-email: wukan <wu-kan@163.com>
 License: BSD 3-Clause License
@@ -116,19 +116,19 @@ def _code_block_2():
     CPV = cs_zscore(_corr) + cs_zscore(_beta)
 
 
-code = codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by', output_file=sys.stdout)  # 打印代码
-code = codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by', output_file="output.py")  # 保存到文件
-code = codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by')  # 只执行，不保存代码
+code = codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by', output_file=sys.stdout)  # 打印代码
+code = codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by', output_file="output.py")  # 保存到文件
+code = codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by')  # 只执行，不保存代码
 
 code = StringIO()
-codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by', output_file=code)  # 保存到字符串
+codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by', output_file=code)  # 保存到字符串
 code.seek(0)
 code.read()  # 读取代码
 
 # TODO 替换成合适的数据
 df = pl.DataFrame()
-df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='partition_by').collect()  # Lazy CPU
-df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='partition_by').collect(engine="gpu")  # Lazy GPU
+df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='order_by').collect()  # Lazy CPU
+df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='order_by').collect(engine="gpu")  # Lazy GPU
 
 ```
 
@@ -217,11 +217,11 @@ X3 = (ts_returns(CLOSE, 3)).over(_ASSET_, order_by=_DATE_),
 
 第2种开头的`null`区域，是否影响结果由算子所决定，特别时是多列输入时`null`区域可能有数据
 
-1. `over_null='partition_by'`。分到两个区域
-2. `over_null='order_by'`。分到一个区域，`null`排在前面
+1. `over_null='partition_by'`。分到两个区域。多`ts`嵌套时，分区导致数据区变短，输出的有效数据变得更短
+2. `over_null='order_by'`。分到一个区域，`null`排在前面。多`ts`嵌套时，数据区长度不变，输出的有效数据符合算法
 3. `over_null=None`。不处理，直接调用，速度更快。如果确信不会中段产生`null`建议使用此参数
 
-`codegen_exec(over_null='partition_by')`为全局使用`partition_by`。但遇到`ts_count_nulls`这类`null`
+`codegen_exec(over_null='order_by')`为全局使用`order_by`。但遇到`ts_count_nulls`这类`null`
 函数就得使用`over_null=None`，所以本工具还新添了注释功能来指定单行表达式参数
 
 1. `# --over_null partition_by`。单行`over_null='partition_by'`
@@ -233,18 +233,28 @@ X3 = (ts_returns(CLOSE, 3)).over(_ASSET_, order_by=_DATE_),
 
 1. `# --over_null`传参注释只能写在单行表达式的后面，不能独立成一行，否则会被忽略
 2. `# --over_null # --over_null=order_by`多个`#`时，只取第一个有效
-3. 只对最外层`ts`函数有效。如果`ts`函数不在外层，需要人工提炼。例如：
+3. 只对`ts`函数有效。并能向上游传播。例如：
    ```python
-   X1 = cs_rank(ts_mean(CLOSE, 3)) # --over_null=order_by # 应用在cs_rank上，没有意义
-   X2 = ts_rank(ts_mean(CLOSE, 3), 5) # --over_null=order_by # 本以为应用在ts_rank(ts_mean)上，但由于出现了公共ts_mean，其实是应用在ts_rank(_x_0)上
+    X1 = ts_mean(ts_mean(CLOSE, 2), 2) # --over_null order_by
+    X2 = ts_mean(CLOSE, 2) # --over_null=order_by
+    X3 = ts_mean(CLOSE, 2) # 取`codegen_exec`参数传入的`over_null`值
+    X4 = cs_rank(CLOSE) + ts_mean(CLOSE, 2) # --over_null None
    ```
 
-   需写成
+   会自动翻译成
 
    ```python
-   _x_0 = ts_mean(CLOSE, 3)  # --over_null=order_by 
-   X1 = cs_rank(_x_0)
-   X2 = ts_rank(_x_0, 5)
+    #========================================func_0_cs__date
+    _x_1 = cs_rank(CLOSE) # --over_null=None
+    #========================================func_0_ts__asset
+    _x_0 = ts_mean(CLOSE, 2) # --over_null=order_by
+    _x_2 = ts_mean(CLOSE, 2) # --over_null=None
+    #========================================func_0_ts__asset
+    X1 = ts_mean(_x_0, 2) # --over_null=order_by
+    #========================================func_1_cl
+    X2 = _x_0 # --over_null=order_by
+    X3 = _x_0 # --over_null=order_by
+    X4 = _x_1 + _x_2 # --over_null=None
    ```
 4. 由于很容易搞错，强烈建议生成`output_file`,检查生成的代码是否正确。
 

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/README.md

@@ -66,19 +66,19 @@ def _code_block_2():
     CPV = cs_zscore(_corr) + cs_zscore(_beta)
 
 
-code = codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by', output_file=sys.stdout)  # 打印代码
-code = codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by', output_file="output.py")  # 保存到文件
-code = codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by')  # 只执行，不保存代码
+code = codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by', output_file=sys.stdout)  # 打印代码
+code = codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by', output_file="output.py")  # 保存到文件
+code = codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by')  # 只执行，不保存代码
 
 code = StringIO()
-codegen_exec(None, _code_block_1, _code_block_2, over_null='partition_by', output_file=code)  # 保存到字符串
+codegen_exec(None, _code_block_1, _code_block_2, over_null='order_by', output_file=code)  # 保存到字符串
 code.seek(0)
 code.read()  # 读取代码
 
 # TODO 替换成合适的数据
 df = pl.DataFrame()
-df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='partition_by').collect()  # Lazy CPU
-df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='partition_by').collect(engine="gpu")  # Lazy GPU
+df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='order_by').collect()  # Lazy CPU
+df = codegen_exec(df.lazy(), _code_block_1, _code_block_2, over_null='order_by').collect(engine="gpu")  # Lazy GPU
 
 ```
 
@@ -167,11 +167,11 @@ X3 = (ts_returns(CLOSE, 3)).over(_ASSET_, order_by=_DATE_),
 
 第2种开头的`null`区域，是否影响结果由算子所决定，特别时是多列输入时`null`区域可能有数据
 
-1. `over_null='partition_by'`。分到两个区域
-2. `over_null='order_by'`。分到一个区域，`null`排在前面
+1. `over_null='partition_by'`。分到两个区域。多`ts`嵌套时，分区导致数据区变短，输出的有效数据变得更短
+2. `over_null='order_by'`。分到一个区域，`null`排在前面。多`ts`嵌套时，数据区长度不变，输出的有效数据符合算法
 3. `over_null=None`。不处理，直接调用，速度更快。如果确信不会中段产生`null`建议使用此参数
 
-`codegen_exec(over_null='partition_by')`为全局使用`partition_by`。但遇到`ts_count_nulls`这类`null`
+`codegen_exec(over_null='order_by')`为全局使用`order_by`。但遇到`ts_count_nulls`这类`null`
 函数就得使用`over_null=None`，所以本工具还新添了注释功能来指定单行表达式参数
 
 1. `# --over_null partition_by`。单行`over_null='partition_by'`
@@ -183,18 +183,28 @@ X3 = (ts_returns(CLOSE, 3)).over(_ASSET_, order_by=_DATE_),
 
 1. `# --over_null`传参注释只能写在单行表达式的后面，不能独立成一行，否则会被忽略
 2. `# --over_null # --over_null=order_by`多个`#`时，只取第一个有效
-3. 只对最外层`ts`函数有效。如果`ts`函数不在外层，需要人工提炼。例如：
+3. 只对`ts`函数有效。并能向上游传播。例如：
    ```python
-   X1 = cs_rank(ts_mean(CLOSE, 3)) # --over_null=order_by # 应用在cs_rank上，没有意义
-   X2 = ts_rank(ts_mean(CLOSE, 3), 5) # --over_null=order_by # 本以为应用在ts_rank(ts_mean)上，但由于出现了公共ts_mean，其实是应用在ts_rank(_x_0)上
+    X1 = ts_mean(ts_mean(CLOSE, 2), 2) # --over_null order_by
+    X2 = ts_mean(CLOSE, 2) # --over_null=order_by
+    X3 = ts_mean(CLOSE, 2) # 取`codegen_exec`参数传入的`over_null`值
+    X4 = cs_rank(CLOSE) + ts_mean(CLOSE, 2) # --over_null None
    ```
 
-   需写成
+   会自动翻译成
 
    ```python
-   _x_0 = ts_mean(CLOSE, 3)  # --over_null=order_by 
-   X1 = cs_rank(_x_0)
-   X2 = ts_rank(_x_0, 5)
+    #========================================func_0_cs__date
+    _x_1 = cs_rank(CLOSE) # --over_null=None
+    #========================================func_0_ts__asset
+    _x_0 = ts_mean(CLOSE, 2) # --over_null=order_by
+    _x_2 = ts_mean(CLOSE, 2) # --over_null=None
+    #========================================func_0_ts__asset
+    X1 = ts_mean(_x_0, 2) # --over_null=order_by
+    #========================================func_1_cl
+    X2 = _x_0 # --over_null=order_by
+    X3 = _x_0 # --over_null=order_by
+    X4 = _x_1 + _x_2 # --over_null=None
    ```
 4. 由于很容易搞错，强烈建议生成`output_file`,检查生成的代码是否正确。
 

expr_codegen-0.17.0/expr_codegen/_version.py

@@ -0,0 +1 @@
+__version__ = "0.17.0"

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/expr_codegen/expr.py

@@ -200,7 +200,7 @@ def get_children(func, func_kwargs, expr, output_exprs, output_symbols, date, as
     expr
         表达式
     output_exprs
-        输出分割后的了表达式
+        输出分割后的子表达式
     output_symbols
         输出每个子表达式中的符号
     date
@@ -236,17 +236,21 @@ def get_children(func, func_kwargs, expr, output_exprs, output_symbols, date, as
             # 外部与内部不同，需处理
             for i, child in enumerate(children):
                 # 不在子中即表示不同
-                if curr in child:
-                    continue
+                # if curr in child:
+                #     continue
                 if expr.args[i].is_Atom:
                     # print(expr.args[i], 'is_Atom 2')
                     continue
                 output_exprs = append_node(expr.args[i], output_exprs)
         else:
-            # ts_sum(OPEN, 5)*ts_sum(RETURNS, 5) ('cl',) [{('ts', 'asset', 'date')}, {('ts', 'asset', 'date')}] 6 alpha_008
-            pass
-            # if isinstance(expr, Mul):
-            #     output_exprs = append_node(expr, output_exprs)
+            for i, child in enumerate(children):
+                if not child:  # 跳过空集合
+                    continue
+                if next(iter(child))[0] == CL:
+                    continue
+                if expr.args[i].is_Atom:
+                    continue
+                output_exprs = append_node(expr.args[i], output_exprs)
 
         # 按需返回，当前是基础算子就返回下一层信息，否则返回当前
         if curr[0] == CL:

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/expr_codegen/model.py

@@ -349,6 +349,8 @@ def skip_expr_node(G: nx.DiGraph, node, keep_nodes):
         e = G.nodes[s]['expr']
         e = e.xreplace({symbol: expr})
         G.nodes[s]['expr'] = e
+        if d['comment'] != "#":
+            G.nodes[s]['comment'] = d['comment']
 
     # 这里用了product生成多个关联边
     G.add_edges_from(product(pred, succ))
@@ -365,14 +367,11 @@ def dag_start(exprs_list, func, func_kwargs, date, asset):
     return G
 
 
-def dag_middle(G, exprs_names, skip_columns, func, func_kwargs, date, asset, skip_simplify):
+def dag_middle(G, exprs_names, skip_columns, func, func_kwargs, date, asset):
     """删除几个没有必要的节点"""
     # 以下划线开头的节点，不保留
     keep_nodes = [k for k in exprs_names if not k.startswith('_')]
 
-    if not skip_simplify:
-        # ts_rank(-RET - -RET, 20),防止替换成ts_rank(0, 20)
-        G = merge_nodes_1(G, keep_nodes, *keep_nodes)
     G = merge_nodes_2(G, keep_nodes, *keep_nodes)
 
     # 移除0出度的节点，但保留部分
@@ -405,3 +404,30 @@ def dag_end(G):
     exprs_ldl._list = exprs_ldl.values()[1:]
 
     return exprs_ldl, G
+
+
+def propagate_comment_recursive(G):
+    """按拓扑排序反序遍历，确保从终点开始"""
+    from collections import deque
+
+    # Kahn算法求拓扑排序
+    in_degree = {node: G.in_degree(node) for node in G.nodes()}
+    queue = deque([node for node in G.nodes() if in_degree[node] == 0])
+    topo_order = []
+
+    while queue:
+        node = queue.popleft()
+        topo_order.append(node)
+        for succ in G.successors(node):
+            in_degree[succ] -= 1
+            if in_degree[succ] == 0:
+                queue.append(succ)
+
+    # 反序遍历拓扑排序（从终点开始）
+    for node in reversed(topo_order):
+        comment = G.nodes[node]['comment']
+        if comment == "#":
+            continue
+        for pred in G.predecessors(node):
+            if G.nodes[pred]['comment'] == "#":
+                G.nodes[pred]['comment'] = comment

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/expr_codegen/polars/code.py

@@ -44,7 +44,7 @@ def codegen(exprs_ldl: ListDictList, exprs_src, syms_dst,
             filename,
             date='date', asset='asset',
             extra_codes: Sequence[str] = (),
-            over_null: Literal['order_by', 'partition_by', None] = 'partition_by',
+            over_null: Literal['order_by', 'partition_by', None] = 'order_by',
             ge_date_idx: int = 0,
             **kwargs):
     """基于模板的代码生成"""
@@ -83,14 +83,14 @@ def codegen(exprs_ldl: ListDictList, exprs_src, syms_dst,
                     exprs_dst.append(f"#" + '=' * 40 + func_name)
                 else:
                     va, ex, sym, comment = kv
-                    # 多个#时，只取第一个#后的参数
-                    args, argv = parser.parse_known_args(args=comment.split("#")[1].split(" "))
                     s1 = str(ex)
                     s2 = p.doprint(ex)
                     if s1 != s2:
                         # 不想等，打印注释，显示会更直观察
                         func_code.append(f"# {va} = {s1}")
                     if k[0] == TS:
+                        # 多个#时，只取第一个#后的参数
+                        args, argv = parser.parse_known_args(args=comment.split("#")[1].split(" "))
                         ts_func_name = func_name
                         # https://github.com/pola-rs/polars/issues/12925#issuecomment-2552764629
                         _sym = [f"{s}.is_not_null()" for s in set(sym)]

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/expr_codegen/rust/code.py

@@ -15,7 +15,7 @@ def get_groupby_from_tuple(tup, func_name, drop_cols):
     prefix2, *_ = tup
     if len(drop_cols) > 0:
         drop_cols = [f'"{c}".into()' for c in drop_cols]
-        drop_str = f'.drop(Selector::ByName {{ names: Arc::new([{','.join(drop_cols)}]), strict: true }})'
+        drop_str = f'.drop(Selector::ByName {{ names: Arc::new([{",".join(drop_cols)}]), strict: true }})'
     else:
         drop_str = ""
 
@@ -44,7 +44,7 @@ def codegen(exprs_ldl: ListDictList, exprs_src, syms_dst,
             filename,
             date='date', asset='asset',
             extra_codes: Sequence[str] = (),
-            over_null: Literal['order_by', 'partition_by', None] = 'partition_by',
+            over_null: Literal['order_by', 'partition_by', None] = 'order_by',
             ge_date_idx: int = 0,
             **kwargs):
     """基于模板的代码生成"""
@@ -83,14 +83,14 @@ def codegen(exprs_ldl: ListDictList, exprs_src, syms_dst,
                     exprs_dst.append(f"#" + '=' * 40 + func_name)
                 else:
                     va, ex, sym, comment = kv
-                    # 多个#时，只取第一个#后的参数
-                    args, argv = parser.parse_known_args(args=comment.split("#")[1].split(" "))
                     s1 = str(ex)
                     s2 = p.doprint(ex)
                     if s1 != s2:
                         # 不想等，打印注释，显示会更直观察
                         func_code.append(f"// {va} = {s1}")
                     if k[0] == TS:
+                        # 多个#时，只取第一个#后的参数
+                        args, argv = parser.parse_known_args(args=comment.split("#")[1].split(" "))
                         ts_func_name = func_name
                         # https://github.com/pola-rs/polars/issues/12925#issuecomment-2552764629
                         _sym = [f'col("{s}").is_not_null()' for s in set(sym)]

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/expr_codegen/sql/code.py

@@ -14,7 +14,7 @@ def codegen(exprs_ldl: ListDictList, exprs_src, syms_dst,
             filename,
             date='date', asset='asset',
             extra_codes: Sequence[str] = (),
-            over_null: Literal['order_by', 'partition_by', None] = 'partition_by',
+            over_null: Literal['order_by', 'partition_by', None] = 'order_by',
             table_name: str = 'self',
             ge_date_idx: int = 0,
             **kwargs):

{expr_codegen-0.16.6 → expr_codegen-0.17.0}/expr_codegen/tool.py

@@ -1,9 +1,10 @@
+import argparse
 import inspect
 import os
 import pathlib
 from functools import lru_cache
 from io import TextIOBase
-from typing import Sequence, Union, TypeVar, Optional, Literal, Iterable, Dict, Tuple
+from typing import Sequence, Union, TypeVar, Optional, Literal, Iterable
 
 import polars as pl
 from black import Mode, format_str
@@ -15,7 +16,7 @@ from sympy.simplify import cse_opts
 
 from expr_codegen.codes import sources_to_exprs
 from expr_codegen.expr import get_current_by_prefix, get_children, replace_exprs
-from expr_codegen.model import dag_start, dag_end, dag_middle, _RESERVED_WORD_
+from expr_codegen.model import dag_start, dag_end, dag_middle, _RESERVED_WORD_, propagate_comment_recursive, create_dag_exprs
 
 try:
     from pandas import DataFrame as _pd_DataFrame
@@ -61,6 +62,20 @@ def simplify2(expr, skip_simplify: bool):
     return expr
 
 
+def comment_adjust(expr_src, over_null):
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--over_null", type=str, nargs="?", default=over_null)
+
+    expr_dst = []
+    for k, v, c in expr_src:
+        args, argv = parser.parse_known_args(args=c.split("#")[1].split(" "))
+        c = f"# --over_null={args.over_null}"
+
+        expr_dst.append((k, v, c))
+
+    return expr_dst
+
+
 class ExprTool:
 
     def __init__(self):
@@ -135,21 +150,7 @@ class ExprTool:
         syms = [str(s) for s in syms]
         return exprs, syms
 
-    def reduce(self, repl, redu):
-        """减少中间变量数量，有利用减少内存占用"""
-
-        exprs_list = []
-
-        # cse前简化一次，cse后不再简化
-        # (~开盘涨停 & 昨收涨停) | (~收盘涨停 & 最高涨停)
-        for k, v in repl:
-            exprs_list.append((k, v, "#"))
-        for k, v, c in redu:
-            exprs_list.append((k, v, c))
-
-        return exprs_list
-
-    def cse(self, exprs, symbols_repl=None, exprs_src=None, skip_simplify=False):
+    def cse(self, exprs, symbols_repl=None, exprs_src=None, skip_simplify=False, date='date', asset='asset'):
         """多个子公式+长公式，提取公共公式
 
         Parameters
@@ -171,38 +172,26 @@ class ExprTool:
             表达式
 
         """
-        self.exprs_names = [k for k, v, c in exprs_src]
-        # 包含了注释信息
-        _exprs = [k for k, v in exprs]
-
-        # 注意：对于表达式右边相同，左边不同的情况，会当成一个处理
-        if skip_simplify:
-            repl, redu = cse(_exprs, symbols_repl, optimizations=[])
-        else:
-            repl, redu = cse(_exprs, symbols_repl, optimizations=[(cse_opts.sub_pre, cse_opts.sub_post), ])
-
-        outputs_len = len(exprs_src)
-
-        new_redu = []
-        symbols_redu = iter(exprs_src)
-        for expr in redu[-outputs_len:]:
-            # 可能部分表达式只在之前出现过，后面完全用不到如，ts_rank(ts_decay_linear(x_147, 11.4157), 6.72611)
-            variable = next(symbols_redu)
-            a = symbols(variable[0])
-            new_redu.append((a, expr, variable[2]))
-
-        self.exprs_list = self.reduce(repl, new_redu)
-
-        # with open("exprs.pickle", "wb") as file:
-        #     pickle.dump(exprs_dict, file)
-
+        self.exprs_names = [k for k, _, _ in exprs_src]
+        used = set(self.exprs_names) | {date, asset} | _RESERVED_WORD_
+        used.update(str(s) for expr, _ in exprs for s in expr.free_symbols)
+        pool = symbols_repl if symbols_repl is not None else numbered_symbols('_x_')
+        replacements = (s for s in pool if str(s) not in used)
+        optimizations = [] if skip_simplify else [(cse_opts.sub_pre, cse_opts.sub_post)]
+        self.exprs_list = []
+        for comment in dict.fromkeys(c for _, _, c in exprs_src):
+            outputs = [(k, v, c) for k, v, c in exprs_src if c == comment]
+            selected = [expr for expr, c in exprs if c == comment]
+            repl, redu = cse(selected, replacements, optimizations=optimizations)
+            self.exprs_list.extend((k, v, comment) for k, v in repl)
+            self.exprs_list.extend((symbols(k), expr, c) for (k, _, c), expr in zip(outputs, redu[-len(outputs):]))
         return self.exprs_list
 
-    def dag(self, merge: bool, skip_columns, date, asset, skip_simplify):
+    def dag(self, skip_columns, date, asset):
         """生成DAG"""
         G = dag_start(self.exprs_list, self.get_current_func, self.get_current_func_kwargs, date, asset)
-        if merge:
-            G = dag_middle(G, self.exprs_names, skip_columns, self.get_current_func, self.get_current_func_kwargs, date, asset, skip_simplify)
+        G = dag_middle(G, self.exprs_names, skip_columns, self.get_current_func, self.get_current_func_kwargs, date, asset)
+        propagate_comment_recursive(G)
         return dag_end(G)
 
     def all(self, exprs_src, style: Literal['pandas', 'polars', 'sql', 'rust'] = 'polars',
@@ -210,7 +199,7 @@ class ExprTool:
             replace: bool = True, regroup: bool = False, format: bool = True,
             date='date', asset='asset',
             extra_codes: Sequence[object] = (),
-            over_null: Literal['order_by', 'partition_by', None] = 'partition_by',
+            over_null: Literal['order_by', 'partition_by', None] = 'order_by',
             table_name: str = 'self',
             ge_date_idx: int = 0,
             skip_simplify: bool = False,
@@ -254,14 +243,19 @@ class ExprTool:
         if replace:
             exprs_src = replace_exprs(exprs_src)
 
+        source_graph = create_dag_exprs((symbols(k), v, c) for k, v, c in exprs_src)
+        propagate_comment_recursive(source_graph)
+        exprs_context = [(k, v, source_graph.nodes[k]['comment']) for k, v, _ in exprs_src]
+        exprs_context = comment_adjust(exprs_context, over_null)
+
         # 子表达式在前，原表式在最后
-        exprs_dst, syms_dst = self.merge(date, asset, exprs_src, skip_simplify=skip_simplify)
+        exprs_dst, syms_dst = self.merge(date, asset, exprs_context, skip_simplify=skip_simplify)
         syms_dst = list(set(syms_dst) - _RESERVED_WORD_)
 
         # 提取公共表达式
-        self.cse(exprs_dst, symbols_repl=numbered_symbols('_x_'), exprs_src=exprs_src, skip_simplify=skip_simplify)
+        self.cse(exprs_dst, symbols_repl=numbered_symbols('_x_'), exprs_src=exprs_context, skip_simplify=skip_simplify, date=date, asset=asset)
         # 有向无环图流转
-        exprs_ldl, G = self.dag(True, skip_columns, date, asset, skip_simplify=skip_simplify)
+        exprs_ldl, G = self.dag(skip_columns, date, asset)
 
         if regroup:
             exprs_ldl.optimize(merge=style != 'sql')
@@ -306,7 +300,7 @@ class ExprTool:
                   style: Literal['pandas', 'polars', 'sql', 'rust'] = 'polars',
                   template_file: Optional[str] = None,
                   date: str = 'date', asset: str = 'asset',
-                  over_null: Literal['order_by', 'partition_by', None] = 'partition_by',
+                  over_null: Literal['order_by', 'partition_by', None] = 'order_by',
                   table_name: str = 'self',
                   ge_date_idx: int = 0,
                   skip_simplify: bool = False,

expr_codegen-0.16.6/expr_codegen/_version.py

@@ -1 +0,0 @@
-__version__ = "0.16.6"