Hi all, I have submitted PEP 667 as an alternative to PEP 558. https://www.python.org/dev/peps/pep-0667/ Nick and I have agreed to disagree on the way to fix locals() and f_locals. We are both in agreement that it needs fixing. In summary, PEP 667 has roughly the same surface behavior as PEP 558 but is simpler and more consistent internally, at the expense of some minor C API backwards incompatibility issues. PEP 558 also has backwards incompatibility issues, but claims to be more compatible at the C API level. Cheers, Mark.
Hopefully anyone is still reading python-dev. I'm going to try to summarize the differences between the two proposals, even though Mark already did so in his PEP. But I'd like to start by calling out the key point of contention. Everything here is about locals() and f_locals in *function scope*. (I use f_locals to refer to the f_locals field of frame objects as seen from Python code.) And in particular, it is about what I'll call "extra variables": the current CPython feature that you can add *new* variables to f_locals that don't exist in the frame, for example: def foo(): x = 1 locals()["y"] = 2 # or sys._getframe()["y"] = 2 My first reaction was to propose to drop this feature, but I realize it's kind of important for debuggers to be able to execute arbitrary code in function code -- assignments to locals should affect the frame, but it should also be possible to create new variables (e.g. temporaries). So I agree we should keep this. Terminology-wise, I will refer to variables that are allocated in the frame (like "x" above, and including nonlocals/cells) as "proper" variables. Both PEPs give up when it comes to locals(), declaring it to return a snapshot in this case. This is mostly to ensure better backwards compatibility, since existing code calling locals() may well assume it's a dict. Both PEPs make f_locals some kind of proxy that gives a direct read-write view on the variables in the frame (including cells used for nonlocal references), but they differ in the precise semantics. So apparently the key difference of opinion between Mark and Nick is about f_locals, and what to do with extras. In Nick's proposal when you reference f.f_locals twice in a row (for the same frame object f), you get the same proxy object, whereas in Mark's proposal you get a different object each time, but it doesn't matter, because the proxy has no state other than a reference to the frame. In Mark's proposal, if you assign a value to an extra variable, it gets stored in a hidden dict field on the frame, and when you read the proxy, the contents of that hidden dict field gets included. This hidden dict lazily created on the first store to an extra variable. (Mark shows pseudo-code to clarify this; the hidden dict is stored as _extra_locals on the frame.) In Nick's proposal, there's a cache on the frame that stores both the extras and the proper variables. This cache can get out of sync with the contents of the proper variables when some bytecode is executed (for performance reasons we don't want the bytecode to keep the cache up to date on every store), so there's an operation to sync the frame cache (sync_frame_cache(), it's not defined in which namespace this exists -- is it a builtin or in sys?). Frankly the description in Nick's PEP is hard to follow -- I am not 100% sure what is meant by "the dynamic snapshot", and it's not quite clear whether proper variables are copied into the cache (and if so, why). There are also differences in the proposed C API changes, but the differences there are solvable once we choose the semantics for f_locals. Personally, I find Mark's proposed semantics for f_locals simpler -- there's no cache, only storage for extras, so there's nothing that can get out of sync. I would even consider making locals() return the same proxy -- this is simpler and more consistent with module and class scopes, but it is less backwards compatible, and locals() is used orders of magnitude more than f_locals. (Also, we'd have to modify exec() and eval() to allow using a non-dict as globals, which would require some deep changes in the interpreter.) --Guido PS. In Mark's PEP, there's a pseudo-code version of locals() that can give a different result in class scope than the current CPython implementation: Using __prepare__, a metaclass can provide a namespace to execute the class body that's not a dict (subclass) instance. The current CPython behavior and AFAICT Nick's PEP return that namespace from locals(), but Mark's pseudo-code would return a snapshot copy. I think it's better to stick to the current semantics (and I suspect Mark overlooked this edge case). On Fri, Aug 20, 2021 at 8:23 AM Mark Shannon <mark@hotpy.org> wrote:
Hi all,
I have submitted PEP 667 as an alternative to PEP 558. https://www.python.org/dev/peps/pep-0667/
Nick and I have agreed to disagree on the way to fix locals() and f_locals. We are both in agreement that it needs fixing.
In summary, PEP 667 has roughly the same surface behavior as PEP 558 but is simpler and more consistent internally, at the expense of some minor C API backwards incompatibility issues.
PEP 558 also has backwards incompatibility issues, but claims to be more compatible at the C API level.
Cheers, Mark. _______________________________________________ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-leave@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/4RH5YCXI... Code of Conduct: http://python.org/psf/codeofconduct/
-- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
On Sun, 22 Aug 2021, 10:47 am Guido van Rossum, <guido@python.org> wrote:
Hopefully anyone is still reading python-dev.
I'm going to try to summarize the differences between the two proposals, even though Mark already did so in his PEP. But I'd like to start by calling out the key point of contention.
Everything here is about locals() and f_locals in *function scope*. (I use f_locals to refer to the f_locals field of frame objects as seen from Python code.) And in particular, it is about what I'll call "extra variables": the current CPython feature that you can add *new* variables to f_locals that don't exist in the frame, for example:
def foo(): x = 1 locals()["y"] = 2 # or sys._getframe()["y"] = 2
My first reaction was to propose to drop this feature, but I realize it's kind of important for debuggers to be able to execute arbitrary code in function code -- assignments to locals should affect the frame, but it should also be possible to create new variables (e.g. temporaries). So I agree we should keep this.
I actually tried taking this feature out in one of the PEP 558 drafts, but actually doing so breaks the pdb test suite.
So apparently the key difference of opinion between Mark and Nick is about f_locals, and what to do with extras. In Nick's proposal when you reference f.f_locals twice in a row (for the same frame object f), you get the same proxy object, whereas in Mark's proposal you get a different object each time, but it doesn't matter, because the proxy has no state other than a reference to the frame.
If PEP 558 is still giving that impression, I need to fix the wording - the proxy objects are ephemeral in both PEPs (the 558 text is slightly behind the implementation on that point, as the fast refs mapping is now stored on the frame object, so it only needs to be built once) In Mark's proposal, if you assign a value to an extra variable, it gets
stored in a hidden dict field on the frame, and when you read the proxy, the contents of that hidden dict field gets included. This hidden dict lazily created on the first store to an extra variable. (Mark shows pseudo-code to clarify this; the hidden dict is stored as _extra_locals on the frame.)
PEP 558 works essentially the same way, the difference is that it uses the existing locals dict storage rather than adding new storage just for optimised frames. In Nick's proposal, there's a cache on the frame that stores both the
extras and the proper variables. This cache can get out of sync with the contents of the proper variables when some bytecode is executed (for performance reasons we don't want the bytecode to keep the cache up to date on every store), so there's an operation to sync the frame cache (sync_frame_cache(), it's not defined in which namespace this exists -- is it a builtin or in sys?).
It's an extra method on the proxy objects. You only need it if you keep an old proxy object around - if you always retrieve a new proxy object after executing Python code, that proxy will refresh the cache when it needs to.
Frankly the description in Nick's PEP is hard to follow -- I am not 100% sure what is meant by "the dynamic snapshot", and it's not quite clear whether proper variables are copied into the cache (and if so, why).
Aye, Mark was a bit quicker with his PEP than I anticipated, so I've incorporated the implementation improvements arising from his last round of comments, but the PEP text hasn't been updated yet. Personally, I find Mark's proposed semantics for f_locals simpler --
there's no cache, only storage for extras, so there's nothing that can get out of sync.
The wording in PEP 667 undersells the cost of that simplification: "Code that uses PyEval_GetLocals() will continue to operate safely, but will need to be changed to use PyEval_Locals() to restore functionality." Code that uses PyEval_GetLocals() will NOT continue to operate safely under PEP 667: all such code will raise an exception at runtime, and need to be rewritten to use a new API with different refcounting semantics. That's essentially all code that accesses the frame locals from C, since we don't offer supported APIs for that other than PyEval_GetLocals() (directly accessing the f_locals field on the frame object is only "supported" in a very loose sense of the word, although PEP 558 mostly keeps that working, too) This means the real key difference between the two PEPs is that Mark is proposing a gratuitous compatibility break for PyEval_GetLocals() that also means that the algorithmic complexity characteristics of the proxy implementation will be completely off from those of a regular dict (e.g. len(proxy) will be O(n) in the number of variables defined on the frame rather than being O(1) after the proxy's initial cache update the way it is in PEP 558) If Mark's claim that PyEval_GetLocals() could not be fixed was true then I would be more sympathetic to his proposal, but I know it isn't true, because it still works fine in the PEP 558 implementation (it even immediately sees changes made via proxies, and proxies see changes to extra variables). The only truly unfixable public API is PyFrame_LocalsToFast(). On the code complexity front, while the cache management in PEP 558 does incur a bit of extra complexity, it also offers a lot of code simplification as many mutable mapping API operations can be delegated to the cache instead of needing to be implemented directly against the fast locals array (e.g. the keys(), values() and items() views all interact with the cache rather than the underlying frame storage, so the implementation doesn't need proxy-specific types for those). For O(n) operations, the cache is refreshed every time, while for less than O(n) operations, the cache is refreshed if it is the first time that particular proxy instance has needed it. While API clients *can* delve into the details of exactly when and how the cache gets refreshed, they can also adopt the simple principle of "if in doubt, request a new locals reference" and let the interpreter worry about the details. Cheers, Nick.
On Sat, Aug 21, 2021 at 8:52 PM Nick Coghlan <ncoghlan@gmail.com> wrote:
On Sun, 22 Aug 2021, 10:47 am Guido van Rossum, <guido@python.org> wrote:
Everything here is about locals() and f_locals in *function scope*. (I use f_locals to refer to the f_locals field of frame objects as seen from Python code.) And in particular, it is about what I'll call "extra variables": the current CPython feature that you can add *new* variables to f_locals that don't exist in the frame, for example:
def foo(): x = 1 locals()["y"] = 2 # or sys._getframe()["y"] = 2
My first reaction was to propose to drop this feature, but I realize it's kind of important for debuggers to be able to execute arbitrary code in function code -- assignments to locals should affect the frame, but it should also be possible to create new variables (e.g. temporaries). So I agree we should keep this.
I actually tried taking this feature out in one of the PEP 558 drafts, but actually doing so breaks the pdb test suite.
I wonder if we should reconsider this, given that so much of the complexity of the competing PEPs is due to this issue of "extra" variables. We can fix pdb, and other debuggers probably will be happy to make some changes (debuggers often are closely tied to the implementation anyway -- e.g. PyDev currently seems broken with 3.11). One way to fix it would be to have the debugger use a mapping implementation that acts as a proxy for f_locals, but stores extra variables in its own storage. Maybe this just moves the problem, but I feel support for extra variables will always be a bit of a wart, and many uses of f_locals don't need them. Another thing I feel we should at least have a good second look at is the "locals() returns a snapshot" behavior. This is the same in both PEPs but it is inconsistent with module and class scopes, as well as different from 3.10. I wonder if we're valuing "does it return the same type" too much over "does it exhibit the same high-level (conceptual) behavior" here. Yes, a snapshot is a dict, just like what you get from locals() in class and module scopes. But no, that dict is not an alias for the actual contents of the scope. If we made locals() return the same proxy that f_locals gives, it's no longer a dict, but it has the same *conceptual* behavior (maybe: "meaning") as for the other types of scopes.
So apparently the key difference of opinion between Mark and Nick is about
f_locals, and what to do with extras. In Nick's proposal when you reference f.f_locals twice in a row (for the same frame object f), you get the same proxy object, whereas in Mark's proposal you get a different object each time, but it doesn't matter, because the proxy has no state other than a reference to the frame.
If PEP 558 is still giving that impression, I need to fix the wording - the proxy objects are ephemeral in both PEPs (the 558 text is slightly behind the implementation on that point, as the fast refs mapping is now stored on the frame object, so it only needs to be built once)
Ah, I didn't actually find a clear indication one way or another. If your proposal *also* makes f.f_locals return a new object on each use, the difference between the two proposals really is entirely in the C API, as you bring up below.
In Mark's proposal, if you assign a value to an extra variable, it gets
stored in a hidden dict field on the frame, and when you read the proxy, the contents of that hidden dict field gets included. This hidden dict lazily created on the first store to an extra variable. (Mark shows pseudo-code to clarify this; the hidden dict is stored as _extra_locals on the frame.)
PEP 558 works essentially the same way, the difference is that it uses the existing locals dict storage rather than adding new storage just for optimised frames.
Oh, you're right. So then this doesn't really matter -- if there's no other use for the C-level field f_locals in a function scope, then we might as well use that to store the extras (assuming its NULL-ness isn't used as a flag for some other purpose). To be clear, I *think* that for a function scope where the f_locals property has never been used and locals() has never been called, the C-level f_locals field is NULL.
In Nick's proposal, there's a cache on the frame that stores both the
extras and the proper variables. This cache can get out of sync with the contents of the proper variables when some bytecode is executed (for performance reasons we don't want the bytecode to keep the cache up to date on every store), so there's an operation to sync the frame cache (sync_frame_cache(), it's not defined in which namespace this exists -- is it a builtin or in sys?).
It's an extra method on the proxy objects. You only need it if you keep an old proxy object around - if you always retrieve a new proxy object after executing Python code, that proxy will refresh the cache when it needs to.
Ah, okay. We're getting closer to the heart of the matter -- the need for this to exist feels like a pretty serious wart in your proposal.
Frankly the description in Nick's PEP is hard to follow -- I am not 100%
sure what is meant by "the dynamic snapshot", and it's not quite clear whether proper variables are copied into the cache (and if so, why).
Aye, Mark was a bit quicker with his PEP than I anticipated, so I've incorporated the implementation improvements arising from his last round of comments, but the PEP text hasn't been updated yet.
No problem, I think it is slowly getting clearer to me. (FWIW I really liked the pseudo code Mark gave for the implementation, maybe you could add something similar to your PEP). Personally, I find Mark's proposed semantics for f_locals simpler --
there's no cache, only storage for extras, so there's nothing that can get out of sync.
The wording in PEP 667 undersells the cost of that simplification:
"Code that uses PyEval_GetLocals() will continue to operate safely, but will need to be changed to use PyEval_Locals() to restore functionality."
Code that uses PyEval_GetLocals() will NOT continue to operate safely under PEP 667: all such code will raise an exception at runtime, and need to be rewritten to use a new API with different refcounting semantics.
Yeah, I did a double take too when I read what Mark wrote. He uses "safe" in a very technical sense, meaning that you get a Python exception but not a crash, and no leak or writing freed memory. And it's true, any caller to PyEval_GetLocals() should check for errors, and there are several error conditions that may occur. (The docs are incomplete, they say "returns NULL if no frame is executing" but they fail to mention that it sets an exception in that case and in other cases.) But PyEval_Locals() is in the Stable ABI (though I have no idea why), so we have essentially two options: keep it working, or make it return an error. We can't delete it. And it returns a borrowed reference, which makes it problematic to let it return a "f_locals proxy object" since those proxies are not cached on the frame. So I think Mark's solution is viable, even though his description is understated.
That's essentially all code that accesses the frame locals from C, since we don't offer supported APIs for that other than PyEval_GetLocals() (directly accessing the f_locals field on the frame object is only "supported" in a very loose sense of the word, although PEP 558 mostly keeps that working, too)
This means the real key difference between the two PEPs is that Mark is proposing a gratuitous compatibility break for PyEval_GetLocals() that also means that the algorithmic complexity characteristics of the proxy implementation will be completely off from those of a regular dict (e.g. len(proxy) will be O(n) in the number of variables defined on the frame rather than being O(1) after the proxy's initial cache update the way it is in PEP 558)
Maybe I care about the change in behavior to always return an error. But I definitely don't care about the complexity characteristics of either PyEval_GetLocals() or the f_locals proxy.
If Mark's claim that PyEval_GetLocals() could not be fixed was true then I would be more sympathetic to his proposal, but I know it isn't true, because it still works fine in the PEP 558 implementation (it even immediately sees changes made via proxies, and proxies see changes to extra variables). The only truly unfixable public API is PyFrame_LocalsToFast().
You "fixed" it at a hefty price -- leaving open the possibility of having a stale f_locals proxy object and the need to call sync_frame_cache() if it could be stale. (IIRC you have some language about "if a fresh proxy is always used you will never see out of sync data" but clearly the possibility exists to end up with a stale proxy.)
On the code complexity front, while the cache management in PEP 558 does incur a bit of extra complexity, it also offers a lot of code simplification as many mutable mapping API operations can be delegated to the cache instead of needing to be implemented directly against the fast locals array (e.g. the keys(), values() and items() views all interact with the cache rather than the underlying frame storage, so the implementation doesn't need proxy-specific types for those). For O(n) operations, the cache is refreshed every time, while for less than O(n) operations, the cache is refreshed if it is the first time that particular proxy instance has needed it.
I don't particularly care about code complexity of operations on the proxy (even if the proxy was also used for locals() I wouldn't). The locals of a function will never enter O(...) territory, people just don't write (or even generate!) code with thousands of locals. I also don't particularly care about the cost of writing the implementation, it's really not that hard, and it won't be that much code. I do care about clear semantics that are easy to explain, and Mark's version is extremely clear.
While API clients *can* delve into the details of exactly when and how the cache gets refreshed, they can also adopt the simple principle of "if in doubt, request a new locals reference" and let the interpreter worry about the details.
I think it's not that easy. I'd much rather be able to pass a mapping to some other piece of code (which doesn't need to know that it's a frame-locals proxy, only that it may change over time) than having to pass a frame with the instructions "if you need something from the frame-locals, use frm.f_locals["varname"]. But... I also care about backwards compatibility, and I have a crazy idea for making PyEval_GetLocals() work in a useful manner without compromising the behavior of the f_locals proxy: - Let's start your idea of using the C-level f_locals field to store the "extra" variables. - The Python-level f_locals proxy looks in the actual frame "fast" locals and cells, and uses the C-level f_locals field only for extras - However, PyEval_GetLocals() doesn't return the proxy. - What PyEval_GetLocals() does: it calls PyEval_FastToLocals(), which makes a pass over the frame locals and adds them to the C-level f_locals field; then it returns that field. - So the borrowed reference is owned by the frame, which is the same as currently. - The proxy only uses the f_locals field for extra variables. If a variable is deleted in the frame but exists in the f_locals field, the proxy reports it as deleted. (This requires some care but can be done, since we have the mapping from proper variable names to frame locals or cells.) - We'll still deprecate PyEval_GetLocals() and PyEval_FastToLocals(), but unless the user turns the deprecation warning into an error, they will work for another few releases. Eventually we'll make PyEval_GetLocals() always return an error (similar to Mark's proposal), since it's in the stable ABI. - For PyEval_LocalsToFast() I don't care too much whether we keep it (per Mark's proposal) or make it return an error (per yours). PS. The mapping from varname to position should be on the code object, not on the frame. This is how Mark does it (though his implementation would need to be extended to take cells into account). -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
On 23. 08. 21 5:07, Guido van Rossum wrote:
On Sat, Aug 21, 2021 at 8:52 PM Nick Coghlan <ncoghlan@gmail.com <mailto:ncoghlan@gmail.com>> wrote: [...] Code that uses PyEval_GetLocals() will NOT continue to operate safely under PEP 667: all such code will raise an exception at runtime, and need to be rewritten to use a new API with different refcounting semantics.
Yeah, I did a double take too when I read what Mark wrote. He uses "safe" in a very technical sense, meaning that you get a Python exception but not a crash, and no leak or writing freed memory. And it's true, any caller to PyEval_GetLocals() should check for errors, and there are several error conditions that may occur. (The docs are incomplete, they say "returns NULL if no frame is executing" but they fail to mention that it sets an exception in that case and in other cases.)
But PyEval_Locals() is in the Stable ABI (though I have no idea why),
This was a case of "now is better than never" – a line had to be drawn somewhere, and having a clear line is better than spending years to get the ideal line. For these PEPs, I think the discussion should stick to the desired semantics first; backwards compatibility for the Stable ABI can be bolted on to whatever solution comes up. "Regular" backwards compatibility is another matter – IMO it's important to keep things like debuggers working as much as possible.
we have essentially two options: keep it working, or make it return an error. We can't delete it. And it returns a borrowed reference, which makes it problematic to let it return a "f_locals proxy object" since those proxies are not cached on the frame.
From PEP 652 "Maintaining the Stable ABI":
Future Python versions may deprecate some members of the Stable ABI. Deprecated members will still work, but may suffer from issues like reduced performance or, in the most extreme cases, memory/resource leaks.
There are many things that can be done: - I believe we can add an extra pointer on frame objects, lazily populated, just for backward compatibility. - The bad old API can introduce a reference cycle. - We can incref the "borrowed" reference and introduce a leak. - The bad old API can start always raising an exception. (Last on the list, since if you can't fix the source and recompile an extension, there's no workaround.) In all cases, extension authors can fix things by moving to the new API.
On Mon, 23 Aug 2021 at 13:07, Guido van Rossum <guido@python.org> wrote:
But... I also care about backwards compatibility, and I have a crazy idea for making PyEval_GetLocals() work in a useful manner without compromising the behavior of the f_locals proxy:
- Let's start your idea of using the C-level f_locals field to store the "extra" variables. - The Python-level f_locals proxy looks in the actual frame "fast" locals and cells, and uses the C-level f_locals field only for extras - However, PyEval_GetLocals() doesn't return the proxy. - What PyEval_GetLocals() does: it calls PyEval_FastToLocals(), which makes a pass over the frame locals and adds them to the C-level f_locals field; then it returns that field. - So the borrowed reference is owned by the frame, which is the same as currently. - The proxy only uses the f_locals field for extra variables. If a variable is deleted in the frame but exists in the f_locals field, the proxy reports it as deleted. (This requires some care but can be done, since we have the mapping from proper variable names to frame locals or cells.) - We'll still deprecate PyEval_GetLocals() and PyEval_FastToLocals(), but unless the user turns the deprecation warning into an error, they will work for another few releases. Eventually we'll make PyEval_GetLocals() always return an error (similar to Mark's proposal), since it's in the stable ABI. - For PyEval_LocalsToFast() I don't care too much whether we keep it (per Mark's proposal) or make it return an error (per yours).
I don't think any of this is crazy, as it's how the PEP 558 reference implementation already works for individual keys (aside from only having a documented deprecation of PyEval_GetLocals(), not a programmatic one) You don't need to eliminate the cache (and hence break compatibility with PyEval_GetLocals()) to ensure it can never get out of sync - you just have to resync it every time you use it, rather than allowing the frame API consumer to make that call. The reason I don't like the proxy semantics proposed in PEP 667 is because it either makes the common case (analysing a frame from a tracing function while no application code is running) slower in order to avoid having to worry about cache consistency when trying to analyse a running frame, or else we have to write and maintain a whole lot more fast locals proxy specific code to implement the 5 different dict iteration APIs.
PS. The mapping from varname to position should be on the code object, not on the frame. This is how Mark does it (though his implementation would need to be extended to take cells into account).
It's taking cells into account that forces the lookup mapping to be on the frame: different executions of the same code object may reference different cell objects. Cheers, Nick. -- Nick Coghlan | ncoghlan@gmail.com | Brisbane, Australia
On Thu, 26 Aug 2021 at 18:29, Nick Coghlan <ncoghlan@gmail.com> wrote: [Guido wrote]
PS. The mapping from varname to position should be on the code object, not on the frame. This is how Mark does it (though his implementation would need to be extended to take cells into account).
It's taking cells into account that forces the lookup mapping to be on the frame: different executions of the same code object may reference different cell objects.
Technically, if you design the fast refs mapping so that even cell references have to do an indirection through the fast locals array, then you can put the name-to-fast-locals-offset mapping on the code object. Mine doesn't do that though, it references the cells directly instead. I don't think it makes much difference in practice though, and I don't like the idea of storing a lazily initialised cache on nominally immutable code objects. Cheers, Nick. -- Nick Coghlan | ncoghlan@gmail.com | Brisbane, Australia
On Thu, Aug 26, 2021 at 2:31 AM Nick Coghlan <ncoghlan@gmail.com> wrote:
On Thu, 26 Aug 2021 at 18:29, Nick Coghlan <ncoghlan@gmail.com> wrote: [Guido wrote]
PS. The mapping from varname to position should be on the code object, not on the frame. This is how Mark does it (though his implementation would need to be extended to take cells into account).
It's taking cells into account that forces the lookup mapping to be on the frame: different executions of the same code object may reference different cell objects.
Technically, if you design the fast refs mapping so that even cell references have to do an indirection through the fast locals array, then you can put the name-to-fast-locals-offset mapping on the code object. Mine doesn't do that though, it references the cells directly instead.
I don't think it makes much difference in practice though, and I don't like the idea of storing a lazily initialised cache on nominally immutable code objects.
Only the part visible to the user is immutable. The specializing interpreter stores all sorts of internal mutable data on there -- and at least since 3.9 we've had the inline cache stored there. So as long as this mapping is invisible to the user, please let's put it on the code object -- we have a lot more frame objects than code objects, so saving a pointer on the frame object and adding it to the code object is advantageous. The cost of the extra indirection is irrelevant, this is always going to be a slow interface meant for occasional use in a debugger. -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
On Fri, 27 Aug 2021, 2:33 am Guido van Rossum, <guido@python.org> wrote:
On Thu, Aug 26, 2021 at 2:31 AM Nick Coghlan <ncoghlan@gmail.com> wrote:
On Thu, 26 Aug 2021 at 18:29, Nick Coghlan <ncoghlan@gmail.com> wrote: [Guido wrote]
PS. The mapping from varname to position should be on the code object, not on the frame. This is how Mark does it (though his implementation would need to be extended to take cells into account).
It's taking cells into account that forces the lookup mapping to be on the frame: different executions of the same code object may reference different cell objects.
Technically, if you design the fast refs mapping so that even cell references have to do an indirection through the fast locals array, then you can put the name-to-fast-locals-offset mapping on the code object. Mine doesn't do that though, it references the cells directly instead.
I don't think it makes much difference in practice though, and I don't like the idea of storing a lazily initialised cache on nominally immutable code objects.
Only the part visible to the user is immutable. The specializing interpreter stores all sorts of internal mutable data on there -- and at least since 3.9 we've had the inline cache stored there. So as long as this mapping is invisible to the user, please let's put it on the code object -- we have a lot more frame objects than code objects, so saving a pointer on the frame object and adding it to the code object is advantageous. The cost of the extra indirection is irrelevant, this is always going to be a slow interface meant for occasional use in a debugger.
OK, that makes sense - I'll add a todo note to the implementation PR. For the main topic of PEP 667 (reducing the cache consistency question to purely a matter of PyEval_GetLocals() compatibility), I think I can see a way to make the extra code complexity of the 5 new custom accessory types (iterator, reversed iterator, keys set, values multi set, items set) worthwhile to my mind: write the latter 3 in terms of the first two and the generic mapping API, and expose them for use in other custom mapping implementations (either directly in the types module, or as optional C accelerators for the collections module). (I wouldn't make exposing them part of the PEP, I'd just aim to write them so only the forward and reverse iterators were specific to proxy objects) With that done, popitem() is trivial to rewrite to depend on the iterator code instead of the cache. That would leave len() and value comparison as the only proxy operations that don't adhere to the expected algorithmic complexity of mapping objects, and writing up the comparison with PEP 667 finally convinced me that those are quirks that API users could manage more easily than 558's current lazy caching semantics. I'd still keep the full value cache under the hood, though, as I don't see enough benefit in getting rid of it when the cost is an unnecessary API compatibility break. However, the PEP 667 write up and your own points *have* persuaded me that the extra C code needed to offer assuredly consistent views through all the proxy mapping methods is worthwhile. That will then leave the proposed C APIs as the only differences between the PEPs - the Python level behaviour will be aligned with Mark's proposal. Cheers, Nick. P.S. I don't want to rely on Python code anywhere in the fast locals proxy implementation, as that could cause weird interactions if a trace hook is enabled during the initial start up of the interpreter and tries to trace the proxy implementation code.
-- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
On Thu, Aug 26, 2021 at 5:29 PM Nick Coghlan <ncoghlan@gmail.com> wrote:
[snip] For the main topic of PEP 667 (reducing the cache consistency question to purely a matter of PyEval_GetLocals() compatibility), I think I can see a way to make the extra code complexity of the 5 new custom accessory types (iterator, reversed iterator, keys set, values multi set, items set) worthwhile to my mind: write the latter 3 in terms of the first two and the generic mapping API, and expose them for use in other custom mapping implementations (either directly in the types module, or as optional C accelerators for the collections module). (I wouldn't make exposing them part of the PEP, I'd just aim to write them so only the forward and reverse iterators were specific to proxy objects)
Yes, that makes sense. I wouldn't sweat it though.
With that done, popitem() is trivial to rewrite to depend on the iterator code instead of the cache.
That would leave len() and value comparison as the only proxy operations that don't adhere to the expected algorithmic complexity of mapping objects, and writing up the comparison with PEP 667 finally convinced me that those are quirks that API users could manage more easily than 558's current lazy caching semantics.
That's what we've been saying. :-) I'd still keep the full value cache under the hood, though, as I don't see
enough benefit in getting rid of it when the cost is an unnecessary API compatibility break.
That's for PyEval_GetLocals(), right? Mark updated PEP 667 to also keep a reference to whatever it returned on the frame. So then I think the two PEPs have converged.
However, the PEP 667 write up and your own points *have* persuaded me that the extra C code needed to offer assuredly consistent views through all the proxy mapping methods is worthwhile.
Right. That will then leave the proposed C APIs as the only differences between
the PEPs - the Python level behaviour will be aligned with Mark's proposal.
Yeah, I presume that the *new* APIs proposed are just a matter of a little bikeshedding, nothing major.
Cheers, Nick.
Thanks for your flexibility!
P.S. I don't want to rely on Python code anywhere in the fast locals proxy implementation, as that could cause weird interactions if a trace hook is enabled during the initial start up of the interpreter and tries to trace the proxy implementation code.
Of course. But it would still be interesting to have pseudo-code in your PEP showing the semantics you intend to implement -- that way we can compare 558 and 667 more easily. -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
On Thu, Aug 26, 2021 at 1:29 AM Nick Coghlan <ncoghlan@gmail.com> wrote:
On Mon, 23 Aug 2021 at 13:07, Guido van Rossum <guido@python.org> wrote:
But... I also care about backwards compatibility, and I have a crazy idea for making PyEval_GetLocals() work in a useful manner without compromising the behavior of the f_locals proxy:
[snip]
I don't think any of this is crazy, as it's how the PEP 558 reference implementation already works for individual keys
I'm not sure I understand what it means to be like my idea "for individual keys". But I'm not sure I care. (aside from only
having a documented deprecation of PyEval_GetLocals(), not a programmatic one)
Yeah, that's a detail. The deprecation could start out silent.
You don't need to eliminate the cache (and hence break compatibility with PyEval_GetLocals()) to ensure it can never get out of sync - you just have to resync it every time you use it, rather than allowing the frame API consumer to make that call.
I don't think eliminating the cache part is breaking compatibility.
The reason I don't like the proxy semantics proposed in PEP 667 is because it either makes the common case (analysing a frame from a tracing function while no application code is running) slower in order to avoid having to worry about cache consistency when trying to analyse a running frame, or else we have to write and maintain a whole lot more fast locals proxy specific code to implement the 5 different dict iteration APIs.
I just think it's weird to write a long PEP to clean up the semantics and then replace it with such a complicated solution.
PS. The mapping from varname to position should be on the code object, not on the frame. This is how Mark does it (though his implementation would need to be extended to take cells into account).
It's taking cells into account that forces the lookup mapping to be on the frame: different executions of the same code object may reference different cell objects.
But the frame stores the cells at known positions (after all, the index is embedded in the bytecode). YOu just need to interpret the mapping a bit differently. -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
Hi Nick, On 22/08/2021 4:51 am, Nick Coghlan wrote:
On Sun, 22 Aug 2021, 10:47 am Guido van Rossum, <guido@python.org <mailto:guido@python.org>> wrote:
Hopefully anyone is still reading python-dev.
I'm going to try to summarize the differences between the two proposals, even though Mark already did so in his PEP. But I'd like to start by calling out the key point of contention.
Everything here is about locals() and f_locals in *function scope*. (I use f_locals to refer to the f_locals field of frame objects as seen from Python code.) And in particular, it is about what I'll call "extra variables": the current CPython feature that you can add *new* variables to f_locals that don't exist in the frame, for example:
def foo(): x = 1 locals()["y"] = 2 # or sys._getframe()["y"] = 2
My first reaction was to propose to drop this feature, but I realize it's kind of important for debuggers to be able to execute arbitrary code in function code -- assignments to locals should affect the frame, but it should also be possible to create new variables (e.g. temporaries). So I agree we should keep this.
I actually tried taking this feature out in one of the PEP 558 drafts, but actually doing so breaks the pdb test suite.
So apparently the key difference of opinion between Mark and Nick is about f_locals, and what to do with extras. In Nick's proposal when you reference f.f_locals twice in a row (for the same frame object f), you get the same proxy object, whereas in Mark's proposal you get a different object each time, but it doesn't matter, because the proxy has no state other than a reference to the frame.
If PEP 558 is still giving that impression, I need to fix the wording - the proxy objects are ephemeral in both PEPs (the 558 text is slightly behind the implementation on that point, as the fast refs mapping is now stored on the frame object, so it only needs to be built once)
In Mark's proposal, if you assign a value to an extra variable, it gets stored in a hidden dict field on the frame, and when you read the proxy, the contents of that hidden dict field gets included. This hidden dict lazily created on the first store to an extra variable. (Mark shows pseudo-code to clarify this; the hidden dict is stored as _extra_locals on the frame.)
PEP 558 works essentially the same way, the difference is that it uses the existing locals dict storage rather than adding new storage just for optimised frames.
In Nick's proposal, there's a cache on the frame that stores both the extras and the proper variables. This cache can get out of sync with the contents of the proper variables when some bytecode is executed (for performance reasons we don't want the bytecode to keep the cache up to date on every store), so there's an operation to sync the frame cache (sync_frame_cache(), it's not defined in which namespace this exists -- is it a builtin or in sys?).
It's an extra method on the proxy objects. You only need it if you keep an old proxy object around - if you always retrieve a new proxy object after executing Python code, that proxy will refresh the cache when it needs to.
Frankly the description in Nick's PEP is hard to follow -- I am not 100% sure what is meant by "the dynamic snapshot", and it's not quite clear whether proper variables are copied into the cache (and if so, why).
Aye, Mark was a bit quicker with his PEP than I anticipated, so I've incorporated the implementation improvements arising from his last round of comments, but the PEP text hasn't been updated yet.
Personally, I find Mark's proposed semantics for f_locals simpler -- there's no cache, only storage for extras, so there's nothing that can get out of sync.
The wording in PEP 667 undersells the cost of that simplification:
"Code that uses PyEval_GetLocals() will continue to operate safely, but will need to be changed to use PyEval_Locals() to restore functionality."
Code that uses PyEval_GetLocals() will NOT continue to operate safely under PEP 667: all such code will raise an exception at runtime, and need to be rewritten to use a new API with different refcounting semantics. That's essentially all code that accesses the frame locals from C, since we don't offer supported APIs for that other than PyEval_GetLocals() (directly accessing the f_locals field on the frame object is only "supported" in a very loose sense of the word, although PEP 558 mostly keeps that working, too)
This means the real key difference between the two PEPs is that Mark is proposing a gratuitous compatibility break for PyEval_GetLocals() that also means that the algorithmic complexity characteristics of the proxy implementation will be completely off from those of a regular dict (e.g. len(proxy) will be O(n) in the number of variables defined on the frame rather than being O(1) after the proxy's initial cache update the way it is in PEP 558)
If Mark's claim that PyEval_GetLocals() could not be fixed was true then I would be more sympathetic to his proposal, but I know it isn't true, because it still works fine in the PEP 558 implementation (it even immediately sees changes made via proxies, and proxies see changes to extra variables). The only truly unfixable public API is PyFrame_LocalsToFast().
You are making claims that seem inconsistent with each other. Namely, you are claiming that: 1. That the result of locals() is ephemeral. 2. That PyEval_GetLocals() returns a borrowed reference. This seems impossible, as you can't return a borrowed reference to an emphemeral object. That's just a pointer to freed memory. Do `locals()` and `PyEval_GetLocals()` behave differently? Is the result of `PyEval_GetLocals()` cached, but `locals()` not? If that were the case, then it is a bit confusing, but could work. Would PyEval_GetLocals() be defined as something like this? (add _locals_cache attribute to the frame which is initialized to NULL). def PyEval_GetLocals(): frame._locals_cache attribute = locals() return borrow(frame._locals_cache attribute) None of this is clear (at least not to me) from PEP 558. Cheers, Mark.
On the code complexity front, while the cache management in PEP 558 does incur a bit of extra complexity, it also offers a lot of code simplification as many mutable mapping API operations can be delegated to the cache instead of needing to be implemented directly against the fast locals array (e.g. the keys(), values() and items() views all interact with the cache rather than the underlying frame storage, so the implementation doesn't need proxy-specific types for those). For O(n) operations, the cache is refreshed every time, while for less than O(n) operations, the cache is refreshed if it is the first time that particular proxy instance has needed it.
While API clients *can* delve into the details of exactly when and how the cache gets refreshed, they can also adopt the simple principle of "if in doubt, request a new locals reference" and let the interpreter worry about the details.
Cheers, Nick.
On Mon, Aug 23, 2021 at 4:38 AM Mark Shannon <mark@hotpy.org> wrote:
Hi Nick,
On 22/08/2021 4:51 am, Nick Coghlan wrote:
If Mark's claim that PyEval_GetLocals() could not be fixed was true then I would be more sympathetic to his proposal, but I know it isn't true, because it still works fine in the PEP 558 implementation (it even immediately sees changes made via proxies, and proxies see changes to extra variables). The only truly unfixable public API is PyFrame_LocalsToFast().
You are making claims that seem inconsistent with each other. Namely, you are claiming that:
1. That the result of locals() is ephemeral. 2. That PyEval_GetLocals() returns a borrowed reference.
This seems impossible, as you can't return a borrowed reference to an emphemeral object. That's just a pointer to freed memory.
Do `locals()` and `PyEval_GetLocals()` behave differently?
That is my understanding, yes. in PEP 558 locals() returns a snapshot dict, the Python-level f_locals property returns a fresh proxy that has no state except a pointer to the frame, and PyEval_GetLocals() returns a borrowed reference to the dict that's stored on the frame's C-level f_locals attribute. (In my "crazy" proposal all that is the same.)
Is the result of `PyEval_GetLocals()` cached, but `locals()` not?
I wouldn't call it a cache -- deleting it would affect the semantics, not just the performance. But yes, it returns a reference to an object that is owned by the frame, just as it does in 3.10 and before.
If that were the case, then it is a bit confusing, but could work.
Yes, see my "crazy" proposal.
Would PyEval_GetLocals() be defined as something like this?
(add _locals_cache attribute to the frame which is initialized to NULL).
def PyEval_GetLocals(): frame._locals_cache attribute = locals() return borrow(frame._locals_cache attribute)
Nah, the dict returned by PyEval_GetLocals() is stored in the frame's C-level f_locals attribute, which is consulted by the Python-level f_locals proxy -- primarily to store "extra" variables, but IIUC in Nick's latest version it is also still used to cache by that proxy. Nick's locals() just returns dict(sys._getframe().f_locals).
None of this is clear (at least not to me) from PEP 558.
One problem with PEP 558 is that it's got too many words, and it's lacking a section that crisply describes the semantics of the proposed implementation. I've suggested to Nick that he add a section with pseudo-code for the implementation, like you did in yours. (PS, did you read my PS about what locals() should do in class scope when __prepare__ returns a non-dict?) -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
Hi Guido, On 23/08/2021 3:53 pm, Guido van Rossum wrote:
On Mon, Aug 23, 2021 at 4:38 AM Mark Shannon <mark@hotpy.org <mailto:mark@hotpy.org>> wrote:
Hi Nick,
On 22/08/2021 4:51 am, Nick Coghlan wrote:
> If Mark's claim that PyEval_GetLocals() could not be fixed was true then > I would be more sympathetic to his proposal, but I know it isn't true, > because it still works fine in the PEP 558 implementation (it even > immediately sees changes made via proxies, and proxies see changes to > extra variables). The only truly unfixable public API is > PyFrame_LocalsToFast().
You are making claims that seem inconsistent with each other. Namely, you are claiming that:
1. That the result of locals() is ephemeral. 2. That PyEval_GetLocals() returns a borrowed reference.
This seems impossible, as you can't return a borrowed reference to an emphemeral object. That's just a pointer to freed memory.
Do `locals()` and `PyEval_GetLocals()` behave differently?
That is my understanding, yes. in PEP 558 locals() returns a snapshot dict, the Python-level f_locals property returns a fresh proxy that has no state except a pointer to the frame, and PyEval_GetLocals() returns a borrowed reference to the dict that's stored on the frame's C-level f_locals attribute
Can we avoid describing the C structs in any of these PEPs? It confuses readers having Python attributes and "C-level attributes" (C struct fields?). It also restricts the implementation unnecessarily. (E.g. the PyFrameObject doesn't have a `f_locals` field in 3.11: https://github.com/python/cpython/blob/main/Include/cpython/frameobject.h#L7)
(In my "crazy" proposal all that is the same.)
Is the result of `PyEval_GetLocals()` cached, but `locals()` not?
I wouldn't call it a cache -- deleting it would affect the semantics, not just the performance. But yes, it returns a reference to an object that is owned by the frame, just as it does in 3.10 and before.
If that were the case, then it is a bit confusing, but could work.
Yes, see my "crazy" proposal.
Would PyEval_GetLocals() be defined as something like this?
(add _locals_cache attribute to the frame which is initialized to NULL).
def PyEval_GetLocals(): frame._locals_cache attribute = locals() return borrow(frame._locals_cache attribute)
Nah, the dict returned by PyEval_GetLocals() is stored in the frame's C-level f_locals attribute, which is consulted by the Python-level f_locals proxy -- primarily to store "extra" variables, but IIUC in Nick's latest version it is also still used to cache by that proxy. Nick's locals() just returns dict(sys._getframe().f_locals).
The "extra" variables must be distinct from the result of locals() as that includes both extras and "proper" variables. If we want to cache the locals(), it needs to be distinct from the extra variables. A debugger setting extra variables in a function that that is also accessed by a C call to PyEval_GetLocals() is going to be incredibly rare. Let's not worry about efficiency here.
None of this is clear (at least not to me) from PEP 558.
One problem with PEP 558 is that it's got too many words, and it's lacking a section that crisply describes the semantics of the proposed implementation. I've suggested to Nick that he add a section with pseudo-code for the implementation, like you did in yours.
(PS, did you read my PS about what locals() should do in class scope when __prepare__ returns a non-dict?)
Yes, but no harm in a reminder :) I'll update my PEP to fix the semantics of locals(). Cheers, Mark.
On Mon, Aug 23, 2021 at 8:46 AM Mark Shannon <mark@hotpy.org> wrote:
Hi Guido,
On 23/08/2021 3:53 pm, Guido van Rossum wrote:
On Mon, Aug 23, 2021 at 4:38 AM Mark Shannon <mark@hotpy.org <mailto:mark@hotpy.org>> wrote:
Hi Nick,
On 22/08/2021 4:51 am, Nick Coghlan wrote:
> If Mark's claim that PyEval_GetLocals() could not be fixed was true then > I would be more sympathetic to his proposal, but I know it isn't true, > because it still works fine in the PEP 558 implementation (it even > immediately sees changes made via proxies, and proxies see changes to > extra variables). The only truly unfixable public API is > PyFrame_LocalsToFast().
You are making claims that seem inconsistent with each other. Namely, you are claiming that:
1. That the result of locals() is ephemeral. 2. That PyEval_GetLocals() returns a borrowed reference.
This seems impossible, as you can't return a borrowed reference to an emphemeral object. That's just a pointer to freed memory.
Do `locals()` and `PyEval_GetLocals()` behave differently?
That is my understanding, yes. in PEP 558 locals() returns a snapshot dict, the Python-level f_locals property returns a fresh proxy that has no state except a pointer to the frame, and PyEval_GetLocals() returns a borrowed reference to the dict that's stored on the frame's C-level f_locals attribute
Can we avoid describing the C structs in any of these PEPs?
It confuses readers having Python attributes and "C-level attributes" (C struct fields?). It also restricts the implementation unnecessarily.
(E.g. the PyFrameObject doesn't have a `f_locals` field in 3.11:
https://github.com/python/cpython/blob/main/Include/cpython/frameobject.h#L7 )
I'd be happy to. Nick's PEP still references it (and indeed it is very confusing) and I took it from him. And honestly it would be nice to have a specific short name for it, rather than circumscribing it with "an internal dynamic snapshot stored on the frame object " :-)
(In my "crazy" proposal all that is the same.)
Is the result of `PyEval_GetLocals()` cached, but `locals()` not?
I wouldn't call it a cache -- deleting it would affect the semantics, not just the performance. But yes, it returns a reference to an object that is owned by the frame, just as it does in 3.10 and before.
If that were the case, then it is a bit confusing, but could work.
Yes, see my "crazy" proposal.
Would PyEval_GetLocals() be defined as something like this?
(add _locals_cache attribute to the frame which is initialized to
NULL).
def PyEval_GetLocals(): frame._locals_cache attribute = locals() return borrow(frame._locals_cache attribute)
Nah, the dict returned by PyEval_GetLocals() is stored in the frame's C-level f_locals attribute, which is consulted by the Python-level f_locals proxy -- primarily to store "extra" variables, but IIUC in Nick's latest version it is also still used to cache by that proxy. Nick's locals() just returns dict(sys._getframe().f_locals).
The "extra" variables must be distinct from the result of locals() as that includes both extras and "proper" variables. If we want to cache the locals(), it needs to be distinct from the extra variables.
I don't care that much about caching locals(), but it seems we're bound to cache any non-NULL result from PyEval_GetLocals(), since it returns a borrowed reference. So they may be different things, with different semantics, if we don't cache locals().
A debugger setting extra variables in a function that that is also accessed by a C call to PyEval_GetLocals() is going to be incredibly rare. Let's not worry about efficiency here.
Agreed.
None of this is clear (at least not to me) from PEP 558.
One problem with PEP 558 is that it's got too many words, and it's lacking a section that crisply describes the semantics of the proposed implementation. I've suggested to Nick that he add a section with pseudo-code for the implementation, like you did in yours.
(PS, did you read my PS about what locals() should do in class scope when __prepare__ returns a non-dict?)
Yes, but no harm in a reminder :) I'll update my PEP to fix the semantics of locals().
I'm in suspense as to what semantics you chose. :-) PS. Another point where you seem to have missed some detail is in the mapping from (proper) variable names to "frame locals" -- you use co_varnames, but that (currently, at least) doesn't include cells. -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
Just adding a datapoint, searching our internal codebase at work, I only found two things that I'd consider to be uses of PyEval_GetLocals() outside of CPython itself: https://github.com/earwig/mwparserfromhell/blob/develop/src/mwparserfromhell... https://github.com/numba/numba/blob/master/numba/_dispatcher.cpp#L664 The bulk of uses are naturally with CPython's own ceval.c and similar. -gps On Mon, Aug 23, 2021 at 9:02 AM Guido van Rossum <guido@python.org> wrote:
On Mon, Aug 23, 2021 at 8:46 AM Mark Shannon <mark@hotpy.org> wrote:
Hi Guido,
On 23/08/2021 3:53 pm, Guido van Rossum wrote:
On Mon, Aug 23, 2021 at 4:38 AM Mark Shannon <mark@hotpy.org <mailto:mark@hotpy.org>> wrote:
Hi Nick,
On 22/08/2021 4:51 am, Nick Coghlan wrote:
> If Mark's claim that PyEval_GetLocals() could not be fixed was true then > I would be more sympathetic to his proposal, but I know it isn't true, > because it still works fine in the PEP 558 implementation (it even > immediately sees changes made via proxies, and proxies see changes to > extra variables). The only truly unfixable public API is > PyFrame_LocalsToFast().
You are making claims that seem inconsistent with each other. Namely, you are claiming that:
1. That the result of locals() is ephemeral. 2. That PyEval_GetLocals() returns a borrowed reference.
This seems impossible, as you can't return a borrowed reference to an emphemeral object. That's just a pointer to freed memory.
Do `locals()` and `PyEval_GetLocals()` behave differently?
That is my understanding, yes. in PEP 558 locals() returns a snapshot dict, the Python-level f_locals property returns a fresh proxy that has no state except a pointer to the frame, and PyEval_GetLocals() returns a borrowed reference to the dict that's stored on the frame's C-level f_locals attribute
Can we avoid describing the C structs in any of these PEPs?
It confuses readers having Python attributes and "C-level attributes" (C struct fields?). It also restricts the implementation unnecessarily.
(E.g. the PyFrameObject doesn't have a `f_locals` field in 3.11:
https://github.com/python/cpython/blob/main/Include/cpython/frameobject.h#L7 )
I'd be happy to. Nick's PEP still references it (and indeed it is very confusing) and I took it from him. And honestly it would be nice to have a specific short name for it, rather than circumscribing it with "an internal dynamic snapshot stored on the frame object " :-)
(In my "crazy" proposal all that is the same.)
Is the result of `PyEval_GetLocals()` cached, but `locals()` not?
I wouldn't call it a cache -- deleting it would affect the semantics, not just the performance. But yes, it returns a reference to an object that is owned by the frame, just as it does in 3.10 and before.
If that were the case, then it is a bit confusing, but could work.
Yes, see my "crazy" proposal.
Would PyEval_GetLocals() be defined as something like this?
(add _locals_cache attribute to the frame which is initialized to
NULL).
def PyEval_GetLocals(): frame._locals_cache attribute = locals() return borrow(frame._locals_cache attribute)
Nah, the dict returned by PyEval_GetLocals() is stored in the frame's C-level f_locals attribute, which is consulted by the Python-level f_locals proxy -- primarily to store "extra" variables, but IIUC in Nick's latest version it is also still used to cache by that proxy. Nick's locals() just returns dict(sys._getframe().f_locals).
The "extra" variables must be distinct from the result of locals() as that includes both extras and "proper" variables. If we want to cache the locals(), it needs to be distinct from the extra variables.
I don't care that much about caching locals(), but it seems we're bound to cache any non-NULL result from PyEval_GetLocals(), since it returns a borrowed reference. So they may be different things, with different semantics, if we don't cache locals().
A debugger setting extra variables in a function that that is also accessed by a C call to PyEval_GetLocals() is going to be incredibly rare. Let's not worry about efficiency here.
Agreed.
None of this is clear (at least not to me) from PEP 558.
One problem with PEP 558 is that it's got too many words, and it's lacking a section that crisply describes the semantics of the proposed implementation. I've suggested to Nick that he add a section with pseudo-code for the implementation, like you did in yours.
(PS, did you read my PS about what locals() should do in class scope when __prepare__ returns a non-dict?)
Yes, but no harm in a reminder :) I'll update my PEP to fix the semantics of locals().
I'm in suspense as to what semantics you chose. :-)
PS. Another point where you seem to have missed some detail is in the mapping from (proper) variable names to "frame locals" -- you use co_varnames, but that (currently, at least) doesn't include cells.
-- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...> _______________________________________________ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-leave@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/AK23IEZK... Code of Conduct: http://python.org/psf/codeofconduct/
On Tue, 24 Aug 2021 at 01:57, Guido van Rossum <guido@python.org> wrote:
On Mon, Aug 23, 2021 at 8:46 AM Mark Shannon <mark@hotpy.org> wrote:
Can we avoid describing the C structs in any of these PEPs?
It confuses readers having Python attributes and "C-level attributes" (C struct fields?). It also restricts the implementation unnecessarily.
(E.g. the PyFrameObject doesn't have a `f_locals` field in 3.11: https://github.com/python/cpython/blob/main/Include/cpython/frameobject.h#L7)
I'd be happy to. Nick's PEP still references it (and indeed it is very confusing) and I took it from him. And honestly it would be nice to have a specific short name for it, rather than circumscribing it with "an internal dynamic snapshot stored on the frame object " :-)
https://github.com/python/peps/pull/2060/files switches over to calling it the "frame value cache" in PEP 558, since that's essentially what it is on optimised frames (even today). An unrelated issue that came up while working on that update is something that affects both PEPs: calling "proxy.clear()" is *super weird* if we make it work the same way as PyEval_LocalsToFast() works today. Specifically, it can reach out and clear cells in outer frames, including the __class__ cell used by zero-arg super(). I can't see anyone being super upset if we decide to change that and say that proxy.clear() leaves free variables alone, and only clears local variables and cells owned by that particular frame. Cheers, Nick. -- Nick Coghlan | ncoghlan@gmail.com | Brisbane, Australia
On Thu, Aug 26, 2021 at 2:40 AM Nick Coghlan <ncoghlan@gmail.com> wrote:
[snip] An unrelated issue that came up while working on that update is something that affects both PEPs: calling "proxy.clear()" is *super weird* if we make it work the same way as PyEval_LocalsToFast() works today. Specifically, it can reach out and clear cells in outer frames, including the __class__ cell used by zero-arg super(). I can't see anyone being super upset if we decide to change that and say that proxy.clear() leaves free variables alone, and only clears local variables and cells owned by that particular frame.
That would be another weird corner case. The only exception I would make would be for __class__, which is only a cell for implementation convenience. But why would anyone write proxy.clear()? That would be like deleting all local variables -- what would be the use case for that? I guess to start over with a computation. But there are better ways to do that. So let's not worry too much about preventing the user from shooting themselves in the foot -- surely at the global level, "globals().clear()" will do weird shit too. :-) -- --Guido van Rossum (python.org/~guido) *Pronouns: he/him **(why is my pronoun here?)* <http://feministing.com/2015/02/03/how-using-they-as-a-singular-pronoun-can-c...>
On Sat, Aug 21, 2021 at 05:46:52PM -0700, Guido van Rossum wrote:
Hopefully anyone is still reading python-dev.
I am :-) [...]
Everything here is about locals() and f_locals in *function scope*. (I use f_locals to refer to the f_locals field of frame objects as seen from Python code.) And in particular, it is about what I'll call "extra variables": the current CPython feature that you can add *new* variables to f_locals that don't exist in the frame, for example:
def foo(): x = 1 locals()["y"] = 2 # or sys._getframe()["y"] = 2
I'm confused. I don't think it currently works, at least not in the sense that I understand "works" to mean. Sure, you can add a new key to the dict, but that doesn't add a new local variable: >>> def spam(): ... if False: y = 0 # Fool the compiler into treating y as a local. ... x = 0 ... locals()['y'] = 1 ... print(sys._getframe().f_locals) ... print(y) ... >>> spam() {'x': 0} Traceback (most recent call last): File "<stdin>", line 1, in <module> File "<stdin>", line 6, in spam UnboundLocalError: local variable 'y' referenced before assignment Am I missing something? The above is in 3.9, has something changed in 3.10 or am I just misunderstanding what you mean by "works"? Using f_locals instead of locals() doesn't change the UnboundLocalError in my testing. If you are not referring to new local variables, then I don't understand what these "extra variables" are or where they live in the current implementation. A recent thread on Discuss is maybe relevant: https://discuss.python.org/t/how-can-i-use-exec-in-functions-in-python3-6-si... Would either of these two proposals re-enable exec to work inside functions as it used to? I think Nick's PEP 558 does not, it wants to make it explicit that changes to locals will not be reflected in the local variables. Although Nick does refer to a "write back" strategy that apparently works *now*. That suggests: - there currently is a trick to writing new local variables in the function namespace; - and Nick's PEP will break it. Am I correct? Mark's PEP 667 says "However f.f_locals == f.f_locals will be True, and all changes to the underlying variables, by any means, will be always be visible" so I think that means it will allow exec to work as in Python2.x, at least if you pass sys._getframe().f_locals to exec instead of locals(). Perhaps we need an informational PEP to explain how local variables inside functions work now :-) -- Steve
I apologize. This does indeed not work. What does work, however, is setting a variable from a tracing function. There is magic in the tracing machinery that writes back the variables to the frame when the tracing function returns. This causes the bug referenced as [1] in both PEPs, and both propose to fix the bug by not writing things back that way, instead writing back whenever a key in the proxy is set. The discussion is about subtler differences between the proposals. —Guido On Mon, Aug 23, 2021 at 22:19 Steven D'Aprano <steve@pearwood.info> wrote:
On Sat, Aug 21, 2021 at 05:46:52PM -0700, Guido van Rossum wrote:
Hopefully anyone is still reading python-dev.
I am :-)
[...]
Everything here is about locals() and f_locals in *function scope*. (I use f_locals to refer to the f_locals field of frame objects as seen from Python code.) And in particular, it is about what I'll call "extra variables": the current CPython feature that you can add *new* variables to f_locals that don't exist in the frame, for example:
def foo(): x = 1 locals()["y"] = 2 # or sys._getframe()["y"] = 2
I'm confused. I don't think it currently works, at least not in the sense that I understand "works" to mean. Sure, you can add a new key to the dict, but that doesn't add a new local variable:
>>> def spam(): ... if False: y = 0 # Fool the compiler into treating y as a local. ... x = 0 ... locals()['y'] = 1 ... print(sys._getframe().f_locals) ... print(y) ... >>> spam() {'x': 0} Traceback (most recent call last): File "<stdin>", line 1, in <module> File "<stdin>", line 6, in spam UnboundLocalError: local variable 'y' referenced before assignment
Am I missing something? The above is in 3.9, has something changed in 3.10 or am I just misunderstanding what you mean by "works"?
Using f_locals instead of locals() doesn't change the UnboundLocalError in my testing.
If you are not referring to new local variables, then I don't understand what these "extra variables" are or where they live in the current implementation.
A recent thread on Discuss is maybe relevant:
https://discuss.python.org/t/how-can-i-use-exec-in-functions-in-python3-6-si...
Would either of these two proposals re-enable exec to work inside functions as it used to?
I think Nick's PEP 558 does not, it wants to make it explicit that changes to locals will not be reflected in the local variables. Although Nick does refer to a "write back" strategy that apparently works *now*. That suggests:
- there currently is a trick to writing new local variables in the function namespace;
- and Nick's PEP will break it.
Am I correct?
Mark's PEP 667 says "However f.f_locals == f.f_locals will be True, and all changes to the underlying variables, by any means, will be always be visible" so I think that means it will allow exec to work as in Python2.x, at least if you pass sys._getframe().f_locals to exec instead of locals().
Perhaps we need an informational PEP to explain how local variables inside functions work now :-)
-- Steve _______________________________________________ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-leave@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/CREZXKZW... Code of Conduct: http://python.org/psf/codeofconduct/
-- --Guido (mobile)
On Fri, Aug 20, 2021 at 04:22:56PM +0100, Mark Shannon wrote:
Hi all,
I have submitted PEP 667 as an alternative to PEP 558. https://www.python.org/dev/peps/pep-0667
Specification has a code snippet: def test(): x = 1 l()['x'] = 2 l()['y'] = 4 l()['z'] = 5 y print(locals(), x) https://www.python.org/dev/peps/pep-0667/#id10 Wouldn't that attempt to resolve global y, rather than local y? Unless there is a change to the current behaviour of the compiler, I think you need to fool the compiler: if False: y = 0 # anywhere inside the function is okay Open Issues says: "there would be backwards compatibility issues when locals is assigned to a local variable or when passed to eval." https://www.python.org/dev/peps/pep-0667/#id24 Is that eval meant to be exec? Or both eval and exec? -- Steve
On 24/08/2021 06:27, Steven D'Aprano wrote:
Wouldn't that attempt to resolve global y, rather than local y? Unless there is a change to the current behaviour of the compiler, I think you need to fool the compiler:
if False: y = 0 # anywhere inside the function is okay Time to add a `nonnonlocal` statement ;)
On 2021-08-24 17:55, Patrick Reader wrote:
On 24/08/2021 06:27, Steven D'Aprano wrote:
Wouldn't that attempt to resolve global y, rather than local y? Unless there is a change to the current behaviour of the compiler, I think you need to fool the compiler:
if False: y = 0 # anywhere inside the function is okay Time to add a `nonnonlocal` statement ;)
Or perhaps use "not nonlocal"? :-)
participants (8)
-
Gregory P. Smith -
Guido van Rossum -
Mark Shannon -
MRAB -
Nick Coghlan -
Patrick Reader -
Petr Viktorin -
Steven D'Aprano