Principle: information accumulation #875
Conversation
|
Thinking at a high level about whether there are tradeoffs not mentioned here... I think there are more disadvantages than you cover for disallowing separate declarations and definitions:
It may be worth noting that these don't seem to have much to do with top-down processing... |
| In order to support this and still permit cyclic references between entities, we | ||
| would need to permit separate declaration and definition. | ||
|
|
||
| _Comprehensibility:_ This rule is simple to explain, and has no special cases. |
There was a problem hiding this comment.
this rule is only simple to explain if declaration vs definition is simple to explain, which as noted above for C++ is not true
There was a problem hiding this comment.
I think this is really a problem with separate declaration and definition rather than with the top-down rule. I've tried to separate the two out and added a discussion of this there.
at our weekly meeting.
github-actions
bot
added
the
proposal rfc
|
@chandlerc I know this isn't your preferred direction; how opposed are you to this approach? In our most recent weekly meeting, we had a 5:2 preference for this direction over a more top-down model. |
proposals/p0875.md
Outdated
| impl fn F() {} | ||
| ``` | ||
|
|
||
| _Comprehensiblity:_ In general, determining whether two declarations declare the |
There was a problem hiding this comment.
I think this would be greatly simplified if we require overloads to be defined together with a dedicated syntax.
There was a problem hiding this comment.
Yep. And this is mentioned now below in the simple compilation area.
There was a problem hiding this comment.
Mentioning that option below makes it all the more confusing that it's not mentioned here.
Also, given that these sub-alternatives are relevant to multiple goals (at least "comprehensibility" and "compilation", but arguably some of the others as well), they should probably be introduced prior to the goal subsections.
There was a problem hiding this comment.
I think the text below and this suggestion, while superficially similar, are actually not the same thing. The text below is talking about a possible implementation strategy, particularly for name mangling, where functions are identified by ordinal position / declaration order in the API file, without restricting where in that API file they might be declared. This suggestion, as I understand it, is about matching declaration to definition by requiring that all declarations be provided together in a contiguous source utterance, and using a similar contiguous utterance to define the overload set in the same order. These options are compatible with each other, but are independent choices.
I've added a writeup of the dedicated syntax option here.
proposals/p0875.md
Outdated
| impl fn F() {} | ||
| ``` | ||
|
|
||
| _Comprehensiblity:_ In general, determining whether two declarations declare the |
There was a problem hiding this comment.
Yep. And this is mentioned now below in the simple compilation area.
proposals/p0875.md
Outdated
| impl fn F() {} | ||
| ``` | ||
|
|
||
| _Comprehensiblity:_ In general, determining whether two declarations declare the |
There was a problem hiding this comment.
Mentioning that option below makes it all the more confusing that it's not mentioned here.
Also, given that these sub-alternatives are relevant to multiple goals (at least "comprehensibility" and "compilation", but arguably some of the others as well), they should probably be introduced prior to the goal subsections.
|
@chandlerc In discussion about this, you brought up an example of Swift code where a lot of functions referenced a class which was only defined at the bottom of the file. Trying to understand those functions required reading the class definition, which was out of a logical reading order. With the accumulation rule as given, the class would more likely be defined at the top of the file, and read first. I would note though, I think the accumulation rule can also get in the way of BLUF-style reading, forcing implementation details to come first and the crux of the implementation to come last. For example, consider how we've implemented name lookup in Carbon: To summarize layout here:
From a BLUF perspective, I'd like to put the public function first: it's the high-level summary of what's going on. Then, dig into the function it calls, with the small helpers at the end. From my perspective, C++'s name lookup rule exerts a strong pressure to put small helper functions at the top of the file, and the key implementation details at the bottom -- I view this as undesirable because I think people would prefer to see the main implementation, then walk through its calls. If the leaning is that classes should have public APIs first and private APIs last, I think that reflects similar BLUF preferences: the public APIs are the main implementation, the private APIs are the helpers. I think the difference is that C++ has trained us to accept a different layout between class and free functions. This pattern continues into header files, for example when it's necessary to provide implementation-detail classes above public classes just to ensure name lookup is satisfied. Sometimes forward declarations can be used to partially address this, but the forward declarations would still be first and would still be less important to public API users; also, either language limitations or a developer's preference for minimizing forward declarations can prevent their use. Allowing more arbitrary layouts will certainly allow style issues as in the Swift example, where developers may provide wholly illogical layouts. I think flexibility in name lookup can also allow better API layouts, where developers aren't forced to put private implementation details above public APIs just because name lookup requires it. |
proposals/p0875.md
Outdated
|
|
||
| As in C++, member functions of nested classes would be deferred until the | ||
| outermost class is complete. Unlike in C++, this deferral would apply only to | ||
| the bodies of member functions, and not to default arguments or other contexts |
There was a problem hiding this comment.
I don't know what the word "contexts" here means.
There was a problem hiding this comment.
Expanded to a complete description of the difference from C++.
There was a problem hiding this comment.
Haven't finished going through all the alternatives text yet, but at least covered the main principle and some of the proposal. I'd like to see if there is any way to compress or be more brief on the alternatives without losing information. But if not, it does seem better to capture the information, it may just be worth making sure it summarizes reasonably well.
There was a problem hiding this comment.
(Just wanted to note that other than a minor wording tweak an the re-ordering of some of the alternatives, I think I'm basically happy here. But I'd like to give folks with comment threads a chance to look at the updated version and see if at least their comments are addressed -- I understand that the direction is going to be universally like, but I want to make sure the write up doesn't have actively confusing points.)
Co-authored-by: Geoff Romer <[email protected]>
There was a problem hiding this comment.
I think this looks good. I think it aligns well with the leads' decision. I know we're not going to hit full consensus here, but I'm not seeing a lot of critical comment threads left. I think this LG to land as-is. The proposal itself I think tries to capture the borderline nature of a bunch of these tradeoffs, and the principle is I think simple and to the point and well aligned with the leads decision.
Ship it!
There was a problem hiding this comment.
I think this looks good. I think it aligns well with the leads' decision. I know we're not going to hit full consensus here, but I'm not seeing a lot of critical comment threads left. I think this LG to land as-is. The proposal itself I think tries to capture the borderline nature of a bunch of these tradeoffs, and the principle is I think simple and to the point and well aligned with the leads decision.
Ship it!
github-actions
bot
added
proposal accepted
This proposal addresses the question of whether we should perform a fully top-down compilation (like in C), a mostly top-down compilation (like in C++), or whether we should allow information from later in the same source file to be used in earlier program constructs (like in Rust, Swift, Java, C#, Haskell, and so on).
The proposed direction is:
- Entities declared later in the same source file cannot be used earlier; top-down semantics apply everywhere.
- As an exception, class member function bodies are parsed as if they appeared after the class.
- Forward declarations can be used to separate interface from implementation and to allow entities to be used before they are defined.
- The behavior of the program is nonetheless required to be the same as if we had a globally-consistent rule: it's always a hard error to depend on any information that is not known or that is provided later.
…type is known. (#1352) Following #875, diagnose any use of a name prior to the point where it is introduced and its type is known. This works by performing name resolution on a top-level class, interface, or impl twice: the first pass performs name-resolution for everything other than nested function bodies, and the second pass performs name resolution on function bodies. At the moment, the second pass does a superset of the work done by the first pass, and as a consequence, some identifier expressions now have their target set twice to the same thing.
This proposal addresses the question of whether we should perform a fully top-down compilation (like in C), a mostly top-down compilation (like in C++), or whether we should allow information from later in the same source file to be used in earlier program constructs (like in Rust, Swift, Java, C#, Haskell, and so on).
The proposed direction is:
Fixes #472.