Regression Tests (Python 3.9) #85
Description
I've been running some regression tests against the online version at pylingual.io, and have so far collected the following test cases which fail using the published source code on the dev branch:
# test.py
def if_else(A):
"""
3 0 LOAD_FAST 0 (A)
2 POP_JUMP_IF_FALSE 6
4 4 JUMP_FORWARD 4 (to 10)
6 >> 6 LOAD_CONST 1 (False)
8 RETURN_VALUE
7 >> 10 LOAD_CONST 2 (True)
12 RETURN_VALUE
"""
# expect pass
if A:
pass
else:
return False
return True
def pop_block_return_value_in_if_else(A):
"""
11 0 SETUP_FINALLY 16 (to 18)
12 2 LOAD_FAST 0 (A)
4 POP_JUMP_IF_FALSE 8
13 6 JUMP_FORWARD 6 (to 14)
15 >> 8 POP_BLOCK
10 LOAD_CONST 1 (False)
12 RETURN_VALUE
>> 14 POP_BLOCK
16 JUMP_FORWARD 12 (to 30)
16 >> 18 POP_TOP
20 POP_TOP
22 POP_TOP
17 24 POP_EXCEPT
26 JUMP_FORWARD 2 (to 30)
28 RERAISE
18 >> 30 LOAD_CONST 2 (True)
32 RETURN_VALUE
"""
# expect fail
try:
if A:
pass
else:
# <8> POP_BLOCK
# BlockTemplate[
# <10> LOAD_CONST False
# <12> RETURN_VALUE
# ]
#
# Since the BlockTemplate node does not have the same
# out[EdgeCategory.Exception] as POP_BLOCK, it is not merged.
# The tail node referenced by if_body and else_body diverges:
# - if_body.tail: 14 POP_BLOCK
# - else_body.tail: BlockTemplate[
# <10> LOAD_CONST False
# <12> RETURN_VALUE
# ]
#
# Attempted solution: merge POP_BLOCK into BlockTemplate if it
# ends with RETURN_VALUE
return False
except:
pass
return True
def pop_block_return_global_in_if_else(A):
"""
23 0 SETUP_FINALLY 16 (to 18)
24 2 LOAD_FAST 0 (A)
4 POP_JUMP_IF_FALSE 8
25 6 JUMP_FORWARD 6 (to 14)
27 >> 8 LOAD_GLOBAL 0 (result)
10 POP_BLOCK
12 RETURN_VALUE
>> 14 POP_BLOCK
16 JUMP_FORWARD 12 (to 30)
28 >> 18 POP_TOP
20 POP_TOP
22 POP_TOP
29 24 POP_EXCEPT
26 JUMP_FORWARD 2 (to 30)
28 RERAISE
30 >> 30 LOAD_CONST 1 (True)
32 RETURN_VALUE
"""
# expect fail
global result
try:
if A:
pass
else:
# BlockTemplate [
# <8> LOAD_GLOBAL
# <10> POP_BLOCK
# ]
# <12> RETURN_VALUE
#
# Similar to pop_block_return_value_in_if_else, but RETURN_VALUE is
# an InstTemplate.
#
# Attempted solution: merge 2 consecutive {Inst,Block}Templates
# if the first one ends with POP_BLOCK and the second one with
# RETURN_VALUE
return result
except:
pass
return True
def pop_block_return_value_in_for_loop(A):
"""
34 0 LOAD_FAST 0 (A)
2 GET_ITER
>> 4 FOR_ITER 30 (to 36)
6 STORE_FAST 1 (x)
35 8 SETUP_FINALLY 8 (to 18)
36 10 POP_BLOCK
12 POP_TOP
14 LOAD_CONST 0 (None)
16 RETURN_VALUE
37 >> 18 POP_TOP
20 POP_TOP
22 POP_TOP
38 24 POP_EXCEPT
26 JUMP_ABSOLUTE 4
28 POP_EXCEPT
30 JUMP_ABSOLUTE 4
32 RERAISE
34 JUMP_ABSOLUTE 4
>> 36 LOAD_CONST 0 (None)
38 RETURN_VALUE
"""
# expect pass
for x in A:
try:
return
except:
continue
def pop_block_jump_in_for_loop(A):
"""
42 0 LOAD_FAST 0 (A)
2 GET_ITER
>> 4 FOR_ITER 32 (to 38)
6 STORE_FAST 1 (x)
43 8 SETUP_FINALLY 10 (to 20)
44 10 POP_BLOCK
12 POP_TOP
14 JUMP_ABSOLUTE 38
16 POP_BLOCK
18 JUMP_ABSOLUTE 4
45 >> 20 POP_TOP
22 POP_TOP
24 POP_TOP
46 26 POP_EXCEPT
28 JUMP_ABSOLUTE 4
30 POP_EXCEPT
32 JUMP_ABSOLUTE 4
34 RERAISE
36 JUMP_ABSOLUTE 4
>> 38 LOAD_CONST 0 (None)
40 RETURN_VALUE
"""
# expect fail
for x in A:
try:
break
except:
continueCompiled to test.pyc with:
PYENV_VERSION=3.9 pyenv exec python -c 'import py_compile;py_compile.compile("test.py",cfile="test.pyc")'The online version passes all the tests: https://pylingual.io/view_chimera?identifier=bea4f3d02f7d5e19b525b83b633607e6a28e22598612eb0303089be1f6abbb2b
The open-source version fails for pop_block_return_value_in_if_else, pop_block_return_global_in_if_else, and pop_block_jump_in_for_loop.
I have attempted a naïve fix, based on my very limited understanding of the control flow reconstructor, by registering the following template in Block.py:
@register_template(2, 21)
class ExcBlockTemplate(ControlFlowTemplate):
@override
@classmethod
def try_match(cls, cfg, node) -> ControlFlowTemplate | None:
out = out_edge_dict(cfg, node)
if not ending_instructions("POP_BLOCK")(cfg, node):
return
if not (next_node := out[EdgeCategory.Natural]) or out[EdgeCategory.Conditional]:
return
if not with_top_level_instructions("RETURN_VALUE", "RETURN_CONST", "JUMP_ABSOLUTE")(cfg, next_node):
return
match node, next_node:
case InstTemplate(), InstTemplate():
members = [node, next_node]
new_node = BlockTemplate(members)
in_edges = [(src, new_node, prop) for src, _, prop in cfg.in_edges(node, data=True) if src not in members]
out_edges = [(new_node, dst, prop) for _, dst, prop in cfg.out_edges(next_node, data=True)]
out_edges += [(new_node, dst, prop) for _, dst, prop in cfg.out_edges(node, data=True) if prop['kind'] == EdgeKind.Exception]
cfg.remove_nodes_from(members)
cfg.add_node(new_node)
cfg.add_edges_from(chain(in_edges, out_edges))
return new_node
case BlockTemplate(), InstTemplate():
out_edges = [(node, dst, prop) for _, dst, prop in cfg.out_edges(next_node, data=True)]
cfg.remove_node(next_node)
cfg.add_edges_from(out_edges)
node.members.append(next_node)
return node
case BlockTemplate(), BlockTemplate():
in_edges = [(src, next_node, prop) for src, _, prop in cfg.in_edges(node, data=True) if src not in next_node.members]
out_edges = [(next_node, dst, prop) for _, dst, prop in cfg.out_edges(node, data=True) if prop['kind'] == EdgeKind.Exception]
cfg.remove_node(node)
cfg.add_edges_from(chain(in_edges, out_edges))
next_node.members[:] = [*node.members, *next_node.members]
return next_node
case InstTemplate(), BlockTemplate():
in_edges = [(src, next_node, prop) for src, _, prop in cfg.in_edges(node, data=True) if src not in next_node.members]
out_edges = [(next_node, dst, prop) for _, dst, prop in cfg.out_edges(node, data=True) if prop['kind'] == EdgeKind.Exception]
cfg.remove_node(node)
cfg.add_edges_from(chain(in_edges, out_edges))
next_node.members.insert(0, node)
return next_nodeEssentially, I violate the assumption made by BlockTemplate.try_match, that out[EdgeCategory.Exception] == exc for all members in the block, but only in the case where POP_BLOCK is followed by a RETURN_* or a JUMP_ABSOLUTE.
This results in passing the tests pop_block_return_value_in_if_else and pop_block_return_global_in_if_else, and it yields a different but incorrect output for pop_block_jump_in_for_loop.
The fix is far from correct or well-founded, but I hope it helps identify the root cause for these regressions.
Metadata
Metadata
Assignees
Labels
No labels