1 | """
|
2 | control_flow_pass.py - AST pass that builds a control flow graph.
|
3 | """
|
4 | import collections
|
5 | from typing import overload, Union, Optional, Dict
|
6 |
|
7 | import mypy
|
8 | from mypy.nodes import (Block, Expression, Statement, ExpressionStmt, StrExpr,
|
9 | CallExpr, FuncDef, IfStmt, NameExpr, MemberExpr,
|
10 | IndexExpr, TupleExpr, IntExpr)
|
11 |
|
12 | from mypy.types import CallableType, Instance, Type, UnionType, NoneTyp, TupleType
|
13 |
|
14 | from mycpp.crash import catch_errors
|
15 | from mycpp.util import join_name, split_py_name
|
16 | from mycpp.visitor import SimpleVisitor, T
|
17 | from mycpp import util
|
18 | from mycpp import pass_state
|
19 |
|
20 |
|
21 | class UnsupportedException(Exception):
|
22 | pass
|
23 |
|
24 |
|
25 | def GetObjectTypeName(t: Type) -> util.SymbolPath:
|
26 | if isinstance(t, Instance):
|
27 | return split_py_name(t.type.fullname)
|
28 |
|
29 | elif isinstance(t, UnionType):
|
30 | assert len(t.items) == 2
|
31 | if isinstance(t.items[0], NoneTyp):
|
32 | return GetObjectTypeName(t.items[1])
|
33 |
|
34 | return GetObjectTypeName(t.items[0])
|
35 |
|
36 | assert False, t
|
37 |
|
38 |
|
39 | class Build(SimpleVisitor):
|
40 |
|
41 | def __init__(self, types: Dict[Expression, Type], virtual, local_vars,
|
42 | dot_exprs):
|
43 |
|
44 | self.types = types
|
45 | self.cfgs = collections.defaultdict(pass_state.ControlFlowGraph)
|
46 | self.current_statement_id = None
|
47 | self.current_class_name = None
|
48 | self.current_func_node = None
|
49 | self.loop_stack = []
|
50 | self.virtual = virtual
|
51 | self.local_vars = local_vars
|
52 | self.dot_exprs = dot_exprs
|
53 | self.callees = {} # statement object -> SymbolPath of the callee
|
54 |
|
55 | def current_cfg(self):
|
56 | if not self.current_func_node:
|
57 | return None
|
58 |
|
59 | return self.cfgs[split_py_name(self.current_func_node.fullname)]
|
60 |
|
61 | def resolve_callee(self, o: CallExpr) -> Optional[util.SymbolPath]:
|
62 | """
|
63 | Returns the fully qualified name of the callee in the given call
|
64 | expression.
|
65 |
|
66 | Member functions are prefixed by the names of the classes that contain
|
67 | them. For example, the name of the callee in the last statement of the
|
68 | snippet below is `module.SomeObject.Foo`.
|
69 |
|
70 | x = module.SomeObject()
|
71 | x.Foo()
|
72 |
|
73 | Free-functions defined in the local module are referred to by their
|
74 | normal fully qualified names. The function `foo` in a module called
|
75 | `moduleA` would is named `moduleA.foo`. Calls to free-functions defined
|
76 | in imported modules are named the same way.
|
77 | """
|
78 |
|
79 | if isinstance(o.callee, NameExpr):
|
80 | return split_py_name(o.callee.fullname)
|
81 |
|
82 | elif isinstance(o.callee, MemberExpr):
|
83 | if isinstance(o.callee.expr, NameExpr):
|
84 | is_module = isinstance(self.dot_exprs.get(o.callee),
|
85 | pass_state.ModuleMember)
|
86 | if is_module:
|
87 | return split_py_name(
|
88 | o.callee.expr.fullname) + (o.callee.name, )
|
89 |
|
90 | elif o.callee.expr.name == 'self':
|
91 | assert self.current_class_name
|
92 | return self.current_class_name + (o.callee.name, )
|
93 |
|
94 | else:
|
95 | local_type = None
|
96 | for name, t in self.local_vars.get(self.current_func_node,
|
97 | []):
|
98 | if name == o.callee.expr.name:
|
99 | local_type = t
|
100 | break
|
101 |
|
102 | if local_type:
|
103 | if isinstance(local_type, str):
|
104 | return split_py_name(local_type) + (
|
105 | o.callee.name, )
|
106 |
|
107 | elif isinstance(local_type, Instance):
|
108 | return split_py_name(
|
109 | local_type.type.fullname) + (o.callee.name, )
|
110 |
|
111 | elif isinstance(local_type, UnionType):
|
112 | assert len(local_type.items) == 2
|
113 | return split_py_name(
|
114 | local_type.items[0].type.fullname) + (
|
115 | o.callee.expr.name, )
|
116 |
|
117 | else:
|
118 | assert not isinstance(local_type, CallableType)
|
119 | # primitive type or string. don't care.
|
120 | return None
|
121 |
|
122 | else:
|
123 | # context or exception handler. probably safe to ignore.
|
124 | return None
|
125 |
|
126 | else:
|
127 | t = self.types.get(o.callee.expr)
|
128 | if isinstance(t, Instance):
|
129 | return split_py_name(t.type.fullname) + (o.callee.name, )
|
130 |
|
131 | elif isinstance(t, UnionType):
|
132 | assert len(t.items) == 2
|
133 | return split_py_name(
|
134 | t.items[0].type.fullname) + (o.callee.name, )
|
135 |
|
136 | elif o.callee.expr and getattr(o.callee.expr, 'fullname',
|
137 | None):
|
138 | return split_py_name(
|
139 | o.callee.expr.fullname) + (o.callee.name, )
|
140 |
|
141 | else:
|
142 | # constructors of things that we don't care about.
|
143 | return None
|
144 |
|
145 | # Don't currently get here
|
146 | raise AssertionError()
|
147 |
|
148 | def get_place_or_ref_name(self, expr: Expression) -> Optional[util.SymbolPath]:
|
149 | """
|
150 | To do dataflow analysis we need to track changes to objects, which
|
151 | requires naming them. This function returns the name of the object
|
152 | referred to by the given expression. If the expression doesn't refer to
|
153 | an object or variable it returns None.
|
154 |
|
155 | Objects are named slightly differently than they appear in the source
|
156 | code.
|
157 |
|
158 | Objects referenced by local variables are referred to by the name of the
|
159 | local. For example, the name of the object in both statements below is
|
160 | `x`.
|
161 |
|
162 | x = module.SomeObject()
|
163 | x = None
|
164 |
|
165 | Member expressions are named after the parent object's type. For
|
166 | example, the names of the objects in the member assignment statements
|
167 | below are both `module.SomeObject.member_a`. This makes it possible to
|
168 | track data flow across object members without having to track individual
|
169 | heap objects, which would increase the search space for analyses and
|
170 | slow things down.
|
171 |
|
172 | x = module.SomeObject()
|
173 | y = module.SomeObject()
|
174 | x.member_a = 'foo'
|
175 | y.member_a = 'bar'
|
176 |
|
177 | Index expressions are named after their bases, for the same reasons as
|
178 | member expressions. The coarse-grained precision should lead to an
|
179 | over-approximation of where objects are in use, but should not miss any
|
180 | references. This should be fine for our purposes. In the snippet below
|
181 | the last two assignments are named `x` and `module.SomeObject.a_list`.
|
182 |
|
183 | x = [None] # list[Thing]
|
184 | y = module.SomeObject()
|
185 | x[0] = Thing()
|
186 | y.a_list[1] = Blah()
|
187 |
|
188 | Index expressions over tuples are treated differently, though. Tuples
|
189 | have a fixed size, tend to be small, and their elements have distinct
|
190 | types. So, each element can be (and probably needs to be) individually
|
191 | named. In the snippet below, the name of the RHS in the second
|
192 | assignment is `t.0`.
|
193 |
|
194 | t = (1, 2, 3, 4)
|
195 | x = t[0]
|
196 |
|
197 | The examples above all deal with assignments, but these rules apply to
|
198 | any expression that uses an object or variable.
|
199 | """
|
200 | if isinstance(expr,
|
201 | NameExpr) and expr.name not in {'True', 'False', 'None'}:
|
202 | return (expr.name, )
|
203 |
|
204 | elif isinstance(expr, MemberExpr):
|
205 | dot_expr = self.dot_exprs[expr]
|
206 | if isinstance(dot_expr, pass_state.ModuleMember):
|
207 | return dot_expr.module_path + (dot_expr.member, )
|
208 |
|
209 | elif isinstance(dot_expr, pass_state.HeapObjectMember):
|
210 | return GetObjectTypeName(
|
211 | dot_expr.object_type) + (dot_expr.member, )
|
212 |
|
213 | elif isinstance(dot_expr, pass_state.StackObjectMember):
|
214 | return GetObjectTypeName(
|
215 | dot_expr.object_type) + (dot_expr.member, )
|
216 |
|
217 | elif isinstance(expr, IndexExpr):
|
218 | if isinstance(self.types[expr.base], TupleType):
|
219 | assert isinstance(expr.index, IntExpr)
|
220 | return self.get_place_or_ref_name(expr.base) + (str(expr.index.value),)
|
221 |
|
222 | return self.get_place_or_ref_name(expr.base)
|
223 |
|
224 | return None
|
225 |
|
226 | #
|
227 | # COPIED from IRBuilder
|
228 | #
|
229 |
|
230 | @overload
|
231 | def accept(self, node: Expression) -> T:
|
232 | ...
|
233 |
|
234 | @overload
|
235 | def accept(self, node: Statement) -> None:
|
236 | ...
|
237 |
|
238 | def accept(self, node: Union[Statement, Expression]) -> Optional[T]:
|
239 | with catch_errors(self.module_path, node.line):
|
240 | if isinstance(node, Expression):
|
241 | try:
|
242 | res = node.accept(self)
|
243 | #res = self.coerce(res, self.node_type(node), node.line)
|
244 |
|
245 | # If we hit an error during compilation, we want to
|
246 | # keep trying, so we can produce more error
|
247 | # messages. Generate a temp of the right type to keep
|
248 | # from causing more downstream trouble.
|
249 | except UnsupportedException:
|
250 | res = self.alloc_temp(self.node_type(node))
|
251 | return res
|
252 | else:
|
253 | try:
|
254 | cfg = self.current_cfg()
|
255 | # Most statements have empty visitors because they don't
|
256 | # require any special logic. Create statements for them
|
257 | # here. Don't create statements from blocks to avoid
|
258 | # stuttering.
|
259 | if cfg and not isinstance(node, Block):
|
260 | self.current_statement_id = cfg.AddStatement()
|
261 |
|
262 | node.accept(self)
|
263 | except UnsupportedException:
|
264 | pass
|
265 | return None
|
266 |
|
267 | # Not in superclasses:
|
268 |
|
269 | def visit_mypy_file(self, o: 'mypy.nodes.MypyFile') -> T:
|
270 | if util.ShouldSkipPyFile(o):
|
271 | return
|
272 |
|
273 | self.module_path = o.path
|
274 |
|
275 | for node in o.defs:
|
276 | # skip module docstring
|
277 | if isinstance(node, ExpressionStmt) and isinstance(
|
278 | node.expr, StrExpr):
|
279 | continue
|
280 | self.accept(node)
|
281 |
|
282 | # Statements
|
283 |
|
284 | def visit_for_stmt(self, o: 'mypy.nodes.ForStmt') -> T:
|
285 | cfg = self.current_cfg()
|
286 | with pass_state.CfgLoopContext(
|
287 | cfg, entry=self.current_statement_id) as loop:
|
288 | self.accept(o.expr)
|
289 | self.loop_stack.append(loop)
|
290 | self.accept(o.body)
|
291 | self.loop_stack.pop()
|
292 |
|
293 | def _handle_switch(self, expr, o, cfg):
|
294 | assert len(o.body.body) == 1, o.body.body
|
295 | if_node = o.body.body[0]
|
296 | assert isinstance(if_node, IfStmt), if_node
|
297 | cases = []
|
298 | default_block = util._collect_cases(self.module_path, if_node, cases)
|
299 | with pass_state.CfgBranchContext(
|
300 | cfg, self.current_statement_id) as branch_ctx:
|
301 | for expr, body in cases:
|
302 | self.accept(expr)
|
303 | assert expr is not None, expr
|
304 | with branch_ctx.AddBranch():
|
305 | self.accept(body)
|
306 |
|
307 | if default_block:
|
308 | with branch_ctx.AddBranch():
|
309 | self.accept(default_block)
|
310 |
|
311 | def visit_with_stmt(self, o: 'mypy.nodes.WithStmt') -> T:
|
312 | cfg = self.current_cfg()
|
313 | assert len(o.expr) == 1, o.expr
|
314 | expr = o.expr[0]
|
315 | assert isinstance(expr, CallExpr), expr
|
316 | self.accept(expr)
|
317 |
|
318 | callee_name = expr.callee.name
|
319 | if callee_name == 'switch':
|
320 | self._handle_switch(expr, o, cfg)
|
321 | elif callee_name == 'str_switch':
|
322 | self._handle_switch(expr, o, cfg)
|
323 | elif callee_name == 'tagswitch':
|
324 | self._handle_switch(expr, o, cfg)
|
325 | else:
|
326 | with pass_state.CfgBlockContext(cfg, self.current_statement_id):
|
327 | self.accept(o.body)
|
328 |
|
329 | def visit_func_def(self, o: 'mypy.nodes.FuncDef') -> T:
|
330 | if o.name == '__repr__': # Don't translate
|
331 | return
|
332 |
|
333 | # For virtual methods, pretend that the method on the base class calls
|
334 | # the same method on every subclass. This way call sites using the
|
335 | # abstract base class will over-approximate the set of call paths they
|
336 | # can take when checking if they can reach MaybeCollect().
|
337 | if self.current_class_name and self.virtual.IsVirtual(
|
338 | self.current_class_name, o.name):
|
339 | key = (self.current_class_name, o.name)
|
340 | base = self.virtual.virtuals[key]
|
341 | if base:
|
342 | sub = join_name(self.current_class_name + (o.name, ),
|
343 | delim='.')
|
344 | base_key = base[0] + (base[1], )
|
345 | cfg = self.cfgs[base_key]
|
346 | cfg.AddFact(0, pass_state.FunctionCall(sub))
|
347 |
|
348 | self.current_func_node = o
|
349 | cfg = self.current_cfg()
|
350 | for arg in o.arguments:
|
351 | cfg.AddFact(0, pass_state.Definition((arg.variable.name,)))
|
352 |
|
353 | self.accept(o.body)
|
354 | self.current_func_node = None
|
355 | self.current_statement_id = None
|
356 |
|
357 | def visit_class_def(self, o: 'mypy.nodes.ClassDef') -> T:
|
358 | self.current_class_name = split_py_name(o.fullname)
|
359 | for stmt in o.defs.body:
|
360 | # Ignore things that look like docstrings
|
361 | if (isinstance(stmt, ExpressionStmt) and
|
362 | isinstance(stmt.expr, StrExpr)):
|
363 | continue
|
364 |
|
365 | if isinstance(stmt, FuncDef) and stmt.name == '__repr__':
|
366 | continue
|
367 |
|
368 | self.accept(stmt)
|
369 |
|
370 | self.current_class_name = None
|
371 |
|
372 | def visit_while_stmt(self, o: 'mypy.nodes.WhileStmt') -> T:
|
373 | cfg = self.current_cfg()
|
374 | with pass_state.CfgLoopContext(
|
375 | cfg, entry=self.current_statement_id) as loop:
|
376 | self.accept(o.expr)
|
377 | self.loop_stack.append(loop)
|
378 | self.accept(o.body)
|
379 | self.loop_stack.pop()
|
380 |
|
381 | def visit_return_stmt(self, o: 'mypy.nodes.ReturnStmt') -> T:
|
382 | cfg = self.current_cfg()
|
383 | if cfg:
|
384 | cfg.AddDeadend(self.current_statement_id)
|
385 |
|
386 | if o.expr:
|
387 | self.accept(o.expr)
|
388 |
|
389 | def visit_if_stmt(self, o: 'mypy.nodes.IfStmt') -> T:
|
390 | cfg = self.current_cfg()
|
391 |
|
392 | if util.ShouldVisitIfExpr(o):
|
393 | for expr in o.expr:
|
394 | self.accept(expr)
|
395 |
|
396 | with pass_state.CfgBranchContext(
|
397 | cfg, self.current_statement_id) as branch_ctx:
|
398 | if util.ShouldVisitIfBody(o):
|
399 | with branch_ctx.AddBranch():
|
400 | for node in o.body:
|
401 | self.accept(node)
|
402 |
|
403 | if util.ShouldVisitElseBody(o):
|
404 | with branch_ctx.AddBranch():
|
405 | self.accept(o.else_body)
|
406 |
|
407 | def visit_break_stmt(self, o: 'mypy.nodes.BreakStmt') -> T:
|
408 | if len(self.loop_stack):
|
409 | self.loop_stack[-1].AddBreak(self.current_statement_id)
|
410 |
|
411 | def visit_continue_stmt(self, o: 'mypy.nodes.ContinueStmt') -> T:
|
412 | if len(self.loop_stack):
|
413 | self.loop_stack[-1].AddContinue(self.current_statement_id)
|
414 |
|
415 | def visit_raise_stmt(self, o: 'mypy.nodes.RaiseStmt') -> T:
|
416 | cfg = self.current_cfg()
|
417 | if cfg:
|
418 | cfg.AddDeadend(self.current_statement_id)
|
419 |
|
420 | if o.expr:
|
421 | self.accept(o.expr)
|
422 |
|
423 | def visit_try_stmt(self, o: 'mypy.nodes.TryStmt') -> T:
|
424 | cfg = self.current_cfg()
|
425 | with pass_state.CfgBranchContext(cfg,
|
426 | self.current_statement_id) as try_ctx:
|
427 | with try_ctx.AddBranch() as try_block:
|
428 | self.accept(o.body)
|
429 |
|
430 | for t, v, handler in zip(o.types, o.vars, o.handlers):
|
431 | with try_ctx.AddBranch(try_block.exit):
|
432 | self.accept(handler)
|
433 |
|
434 | def visit_assignment_stmt(self, o: 'mypy.nodes.AssignmentStmt') -> T:
|
435 | cfg = self.current_cfg()
|
436 | if cfg:
|
437 | assert len(o.lvalues) == 1
|
438 | lval = o.lvalues[0]
|
439 | lval_names = []
|
440 | if isinstance(lval, TupleExpr):
|
441 | lval_names.extend(
|
442 | [self.get_place_or_ref_name(item) for item in lval.items])
|
443 |
|
444 | else:
|
445 | lval_names.append(self.get_place_or_ref_name(lval))
|
446 |
|
447 | assert lval_names, o
|
448 |
|
449 | rval_type = self.types[o.rvalue]
|
450 | rval_names = []
|
451 | if isinstance(o.rvalue, CallExpr):
|
452 | # The RHS is either an object constructor or something that
|
453 | # returns a primitive type (e.g. Tuple[int, int] or str).
|
454 | # XXX: When we add inter-procedural analysis we should treat
|
455 | # these not as definitions but as some new kind of assignment.
|
456 | rval_names = [None for _ in lval_names]
|
457 |
|
458 | elif isinstance(o.rvalue, TupleExpr) and len(lval_names) == 1:
|
459 | # We're constructing a tuple. Since tuples have have a fixed
|
460 | # (and usually small) size, we can name each of the
|
461 | # elements.
|
462 | base = lval_names[0]
|
463 | lval_names = [
|
464 | base + (str(i), ) for i in range(len(o.rvalue.items))
|
465 | ]
|
466 | rval_names = [
|
467 | self.get_place_or_ref_name(item) for item in o.rvalue.items
|
468 | ]
|
469 |
|
470 | elif isinstance(rval_type, TupleType):
|
471 | # We're unpacking a tuple. Like the tuple construction case,
|
472 | # give each element a name.
|
473 | rval_name = self.get_place_or_ref_name(o.rvalue)
|
474 | assert rval_name, o.rvalue
|
475 | rval_names = [
|
476 | rval_name + (str(i), ) for i in range(len(lval_names))
|
477 | ]
|
478 |
|
479 | else:
|
480 | rval_names = [self.get_place_or_ref_name(o.rvalue)]
|
481 |
|
482 | assert len(rval_names) == len(lval_names)
|
483 |
|
484 | for lhs, rhs in zip(lval_names, rval_names):
|
485 | assert lhs, lval
|
486 | if rhs:
|
487 | # In this case rhe RHS is another variable. Record the
|
488 | # assignment so we can keep track of aliases.
|
489 | cfg.AddFact(self.current_statement_id,
|
490 | pass_state.Assignment(lhs, rhs))
|
491 | else:
|
492 | # In this case the RHS is either some kind of literal (e.g.
|
493 | # [] or 'foo') or a call to an object constructor. Mark this
|
494 | # statement as an (re-)definition of a variable.
|
495 | cfg.AddFact(
|
496 | self.current_statement_id,
|
497 | pass_state.Definition(lhs),
|
498 | )
|
499 |
|
500 | for lval in o.lvalues:
|
501 | self.accept(lval)
|
502 |
|
503 | self.accept(o.rvalue)
|
504 |
|
505 | # Expressions
|
506 |
|
507 | def visit_call_expr(self, o: 'mypy.nodes.CallExpr') -> T:
|
508 | cfg = self.current_cfg()
|
509 | if self.current_func_node:
|
510 | full_callee = self.resolve_callee(o)
|
511 | if full_callee:
|
512 | self.callees[o] = full_callee
|
513 | cfg.AddFact(
|
514 | self.current_statement_id,
|
515 | pass_state.FunctionCall(join_name(full_callee, delim='.')))
|
516 |
|
517 | self.accept(o.callee)
|
518 | for arg in o.args:
|
519 | self.accept(arg)
|