1 Namespace Definition
2 ====================
3
4 Module attributes are defined either at the module level or by global
5 statements.
6
7 Class attributes are defined only within class statements.
8
9 Instance attributes are defined only by assignments to attributes of self
10 within __init__ methods.
11
12 Potential Restrictions
13 ----------------------
14
15 Names of classes and functions could be restricted to only refer to those
16 objects within the same namespace. If redefinition were to occur, or if
17 multiple possibilities were present, these restrictions could be moderated as
18 follows:
19
20 * Classes assigned to the same name could provide the union of their
21 attributes. This would, however, cause a potential collision of attribute
22 definitions such as methods.
23
24 * Functions, if they share compatible signatures, could share parameter list
25 definitions.
26
27 Data Structures
28 ===============
29
30 The fundamental "value type" is a pair of references: one pointing to the
31 referenced object represented by the interchangeable value; one referring to
32 the context of the referenced object, typically the object through which the
33 referenced object was acquired as an attribute.A
34
35 Value Layout
36 ------------
37
38 0 1
39 object context
40 reference reference
41
42 Acquiring Values
43 ----------------
44
45 Values are acquired through name lookups and attribute access, yielding
46 the appropriate object reference together with a context reference as
47 indicated in the following table:
48
49 Type of Access Context Notes
50 -------------- ------- -----
51
52 Local name Preserved Functions provide no context
53
54 Global name Preserved Modules provide no context
55
56 Class-originating Accessor Methods acquire the context of their
57 attribute -or- accessor if an instance...
58 Preserved or retain the original context if the
59 accessor is a class
60
61 Instance-originating Preserved Methods retain their original context
62 attribute
63
64 There may be some scope for simplifying the above, to the detriment of Python
65 compatibility, since the unbound vs. bound methods situation can be confusing.
66
67 Objects
68 -------
69
70 Since classes, functions and instances are all "objects", each must support
71 certain features and operations in the same way.
72
73 The __class__ Attribute
74 -----------------------
75
76 All objects support the __class__ attribute:
77
78 Class: refers to the type class (type.__class__ also refers to the type class)
79 Function: refers to the function class
80 Instance: refers to the class instantiated to make the object
81
82 Invocation
83 ----------
84
85 The following actions need to be supported:
86
87 Class: create instance, call __init__ with instance, return object
88 Function: call function body, return result
89 Instance: call __call__ method, return result
90
91 Structure Layout
92 ----------------
93
94 A suitable structure layout might be something like this:
95
96 Identifier Address Type Object ...
97
98 0 1 2 3 4
99 classcode invocation __class__ attribute ...
100 reference reference reference
101
102 Here, the classcode refers to the attribute lookup table for the object. Since
103 classes and instances share the same classcode, they might resemble the
104 following:
105
106 Class C:
107
108 0 1 2 3 4
109 code for C __new__ class type attribute ...
110 reference reference reference
111
112 Instance of C:
113
114 0 1 2 3 4
115 code for C C.__call__ class C attribute ...
116 reference reference reference
117 (if exists)
118
119 The __new__ reference would lead to code consisting of the following
120 instructions:
121
122 create instance for C
123 call C.__init__(instance, ...)
124 return instance
125
126 If C has a __call__ attribute, the invocation "slot" of C instances would
127 refer to the same thing as C.__call__.
128
129 For functions, the same general layout applies:
130
131 Function f:
132
133 0 1 2 3 4
134 code for code class attribute ...
135 function reference function reference
136 reference
137
138 Here, the code reference would lead to code for the function. Note that the
139 function locals are completely distinct from this structure and are not
140 comparable to attributes.
141
142 For modules, there is no meaningful invocation reference:
143
144 Module m:
145
146 0 1 2 3 4
147 code for m (unused) module type attribute ...
148 reference (global)
149 reference
150
151 Both classes and modules have code in their definitions, but this would be
152 generated in order and not referenced externally.
153
154 Invocation Operation
155 --------------------
156
157 Consequently, regardless of the object an invocation is always done as
158 follows:
159
160 get invocation reference (at object+1)
161 jump to reference
162
163 Additional preparation is necessary before the above code: positional
164 arguments must be saved to the parameter stack, and keyword arguments must be
165 resolved and saved to the appropriate position in the parameter stack.
166
167 Attribute Operations
168 --------------------
169
170 Attribute access needs to go through the attribute lookup table. Some
171 optimisations are possible and are described in the appropriate section.
172
173 One important aspect of attribute access is the appropriate setting of the
174 context in the acquired attribute value. From the table describing the
175 acquisition of values, it is clear that the principal exception is that where
176 a class-originating attribute is accessed on an instance. Consequently, the
177 following algorithm could be employed once an attribute has been located:
178
179 1. If the attribute's context is a special value, indicating that it should
180 be replaced upon instance access, then proceed to the next step;
181 otherwise, acquire both the context and the object as they are.
182
183 2. If the accessor is an instance, use that as the value's context, acquiring
184 only the object from the attribute.
185
186 Where accesses can be determined ahead of time (as discussed in the
187 optimisations section), the above algorithm may not necessarily be employed in
188 the generated code for some accesses.
189
190 Instruction Evaluation Model
191 ============================
192
193 Programs use a value stack where evaluated instructions may save their
194 results. A value stack pointer indicates the top of this stack. In addition, a
195 separate stack is used to record the invocation frames. All stack pointers
196 refer to the next address to be used by the stack, not the address of the
197 uppermost element.
198
199 Frame Stack Value Stack
200 ----------- ----------- Address of Callable
201 -------------------
202 previous ...
203 current ------> callable -----> identifier
204 arg1 reference to code
205 arg2
206 arg3
207 local4
208 local5
209 ...
210
211 Loading local names is a matter of performing frame-relative accesses to the
212 value stack.
213
214 Invocations and Argument Evaluation
215 -----------------------------------
216
217 When preparing for an invocation, the caller first sets the invocation frame
218 pointer. Then, positional arguments are added to the stack such that the first
219 argument positions are filled. A number of stack locations for the remaining
220 arguments specified in the program are then reserved. The names of keyword
221 arguments are used (in the form of table index values) to consult the
222 parameter table and to find the location in which such arguments are to be
223 stored.
224
225 fn(a, b, d=1, e=2, c=3) -> fn(a, b, c, d, e)
226
227 Value Stack
228 -----------
229
230 ... ... ... ...
231 fn fn fn fn
232 a a a a
233 b b b b
234 ___ ___ ___ --> 3
235 ___ --> 1 1 | 1
236 ___ | ___ --> 2 | 2
237 1 ----------- 2 ----------- 3 -----------
238
239 Conceptually, the frame can be considered as a collection of attributes, as
240 seen in other kinds of structures:
241
242 Frame for invocation of fn:
243
244 0 1 2 3 4 5
245 code a b c d e
246 reference
247
248 However, where arguments are specified positionally, such "attributes" are not
249 set using a comparable approach to that employed with other structures.
250 Keyword arguments are set using an attribute-like mechanism, though, where the
251 position of each argument discovered using the parameter table.
252
253 Method invocations incorporate an implicit first argument which is obtained
254 from the context of the method:
255
256 method(a, b, d=1, e=2, c=3) -> method(self, a, b, c, d, e)
257
258 Value Stack
259 -----------
260
261 ...
262 method
263 context of method
264 a
265 b
266 3
267 1
268 2
269
270 Although it could be possible to permit any object to be provided as the first
271 argument, in order to optimise instance attribute access in methods, we should
272 seek to restrict the object type.
273
274 Verifying Supplied Arguments
275 ----------------------------
276
277 In order to ensure a correct invocation, it is also necessary to check the
278 number of supplied arguments. If the target of the invocation is known at
279 compile-time, no additional instructions need to be emitted; otherwise, the
280 generated code must test for the following situations:
281
282 1. That the number of supplied arguments is equal to the number of expected
283 parameters.
284
285 2. That no keyword argument overwrites an existing positional parameter.
286
287 Tuples, Frames and Allocation
288 -----------------------------
289
290 Using the approach where arguments are treated like attributes in some kind of
291 structure, we could choose to allocate frames in places other than a stack.
292 This would produce something somewhat similar to a plain tuple object.
293
294 Optimisations
295 =============
296
297 Some optimisations around constant objects might be possible; these depend on
298 the following:
299
300 * Reliable tracking of assignments: where assignment operations occur, the
301 target of the assignment should be determined if any hope of optimisation
302 is to be maintained. Where no guarantees can be made about the target of
303 an assignment, no assignment-related information should be written to
304 potential targets.
305
306 * Objects acting as "containers" of attributes must be regarded as "safe":
307 where assignments are recorded as occurring on an attribute, it must be
308 guaranteed that no other unforeseen ways exist to assign to such
309 attributes.
310
311 The discussion below presents certain rules which must be imposed to uphold
312 the above requirements.
313
314 Safe Containers
315 ---------------
316
317 Where attributes of modules, classes and instances are only set once and are
318 effectively constant, it should be possible to circumvent the attribute lookup
319 mechanism and to directly reference the attribute value. This technique may
320 only be considered applicable for the following "container" objects, subject
321 to the noted constraints:
322
323 1. For modules, "safety" is enforced by ensuring that assignments to module
324 attributes are only permitted within the module itself either at the
325 top-level or via names declared as globals. Thus, the following would not
326 be permitted:
327
328 another_module.this_module.attr = value
329
330 In the above, this_module is a reference to the current module.
331
332 2. For classes, "safety" is enforced by ensuring that assignments to class
333 attributes are only permitted within the class definition, outside
334 methods. This would mean that classes would be "sealed" at definition time
335 (like functions).
336
337 Unlike the property of function locals that they may only sensibly be accessed
338 within the function in which they reside, these cases demand additional
339 controls or assumptions on or about access to the stored data. Meanwhile, it
340 would be difficult to detect eligible attributes on arbitrary instances due to
341 the need for some kind of type inference or abstract execution.
342
343 Constant Attributes
344 -------------------
345
346 When accessed via "safe containers", as described above, any attribute with
347 only one recorded assignment on it can be considered a constant attribute and
348 this eligible for optimisation, the consequence of which would be the
349 replacement of a LoadAttrIndex instruction (which needs to look up an
350 attribute using the run-time details of the "container" and the compile-time
351 details of the attribute) with a LoadAttr instruction.
352
353 However, some restrictions exist on assignment operations which may be
354 regarded to cause only one assignment in the lifetime of a program:
355
356 1. For module attributes, only assignments at the top-level outside loop
357 statements can be reliably assumed to cause only a single assignment.
358
359 2. For class attributes, only assignments at the top-level within class
360 definitions and outside loop statements can be reliably assumed to cause
361 only a single assignment.
362
363 All assignments satisfying the "safe container" requirements, but not the
364 requirements immediately above, should each be recorded as causing at least
365 one assignment.
366
367 Additional Controls
368 -------------------
369
370 For the above cases for "container" objects, the following controls would need
371 to apply:
372
373 1. Modules would need to be immutable after initialisation. However, during
374 initialisation, there remains a possibility of another module attempting
375 to access the original module. For example, if ppp/__init__.py contained
376 the following...
377
378 x = 1
379 import ppp.qqq
380 print x
381
382 ...and if ppp/qqq.py contained the following...
383
384 import ppp
385 ppp.x = 2
386
387 ...then the value 2 would be printed. Since modules are objects which are
388 registered globally in a program, it would be possible to set attributes
389 in the above way.
390
391 2. Classes would need to be immutable after initialisation. However, since
392 classes are objects, any reference to a class after initialisation could
393 be used to set attributes on the class.
394
395 Solutions:
396
397 1. Insist on global scope for module attribute assignments.
398
399 2. Insist on local scope within classes.
400
401 Both of the above measures need to be enforced at run-time, since an arbitrary
402 attribute assignment could be attempted on any kind of object, yet to uphold
403 the properties of "safe containers", attempts to change attributes of such
404 objects should be denied. Since foreseen attribute assignment operations have
405 certain properties detectable at compile-time, it could be appropriate to
406 generate special instructions (or modified instructions) during the
407 initialisation of modules and classes for such foreseen assignments, whilst
408 employing normal attribute assignment operations in all other cases. Indeed,
409 the StoreAttr instruction, which is used to set attributes in "safe
410 containers" would be used exclusively for this purpose; the StoreAttrIndex
411 instruction would be used exclusively for all other attribute assignments.
412
413 To ensure the "sealing" of modules and classes, entries in the attribute
414 lookup table would encode whether a class or module is being accessed, so
415 that the StoreAttrIndex instruction could reject such accesses.
416
417 Constant Attribute Values
418 -------------------------
419
420 Where an attribute value is itself regarded as constant, is a "safe container"
421 and is used in an operation accessing its own attributes, the value can be
422 directly inspected for optimisations or employed in the generated code. For
423 the attribute values themselves, only objects of a constant nature may be
424 considered suitable for this particular optimisation:
425
426 * Classes
427 * Modules
428 * Instances defined as constant literals
429
430 This is because arbitrary objects (such as most instances) have no
431 well-defined form before run-time and cannot be investigated further at
432 compile-time or have a representation inserted into the generated code.
433
434 Class Attributes and Access via Instances
435 -----------------------------------------
436
437 Unlike module attributes, class attributes can be accessed in a number of
438 different ways:
439
440 * Using the class itself:
441
442 C.x = 123
443 cls = C; cls.x = 234
444
445 * Using a subclass of the class (for reading attributes):
446
447 class D(C):
448 pass
449 D.x # setting D.x would populate D, not C
450
451 * Using instances of the class or a subclass of the class (for reading
452 attributes):
453
454 c = C()
455 c.x # setting c.x would populate c, not C
456
457 Since assignments are only achieved using direct references to the class, and
458 since class attributes should be defined only within the class initialisation
459 process, the properties of class attributes should be consistent with those
460 desired.
461
462 Method Access via Instances
463 ---------------------------
464
465 It is desirable to optimise method access, even though most method calls are
466 likely to occur via instances. It is possible, given the properties of methods
467 as class attributes to investigate the kind of instance that the self
468 parameter/local refers to within each method: it should be an instance either
469 of the class in which the method is defined or a compatible class, although
470 situations exist where this might not be the case:
471
472 * Explicit invocation of a method:
473
474 d = D() # D is not related to C
475 C.f(d) # calling f(self) in C
476
477 If blatant usage of incompatible instances were somehow disallowed, it would
478 still be necessary to investigate the properties of an instance's class and
479 its relationship with other classes. Consider the following example:
480
481 class A:
482 def f(self): ...
483
484 class B:
485 def f(self): ...
486 def g(self):
487 self.f()
488
489 class C(A, B):
490 pass
491
492 Here, instances of B passed into the method B.g could be assumed to provide
493 access to B.f when self.f is resolved at compile-time. However, instances of C
494 passed into B.g would instead provide access to A.f when self.f is resolved at
495 compile-time (since the method resolution order is C, A, B instead of just B).
496
497 One solution might be to specialise methods for each instance type, but this
498 could be costly. Another less ambitious solution might only involve the
499 optimisation of such internal method calls if an unambiguous target can be
500 resolved.
501
502 Optimising Function Invocations
503 -------------------------------
504
505 Where an attribute value is itself regarded as constant and is a function,
506 knowledge about the parameters of the function can be employed to optimise the
507 preparation of the invocation frame.