What If C++ Abandoned Backward Compatibility?
robert.ocallahan.org
robert.ocallahan.org
The C++ standards committee can standardize a backwards-incompatible "new" C++ for all I care as long as they don't call it C++.
Less implicit means safer but also more verbose and more cumbersome to write.
Rust priorities safety over everything else, also over productivity too in some case.
Absence of function overloading and automatic type conversion is a good example of that in Rust.
Is it the right choice ? Only time and multi-million line codebase will tell us.
Rust does not provide the facility to strategically obfuscate code at scale like C++-style function overloading does.
// C++
auto f() -> int32_t {
return 1;
}
// Rust
fn f() -> i32 {
return 1;
} fn i = 0;Yes, but no.
When people design a language which they know is going to be incompatible, they tend to go all the way with it. If we're breaking it anyway then might as well do a better standard library interface than the stuff designed in the 90s and all that sort of thing, right? And that's good to have for new projects.
But there is a benefit in having something which makes the smallest possible compatibility-breaking change to fix the defect in the original. So you have a "new language" with a borrow checker, but the standard library functions all have the same names and the same purpose and time complexity as the C and C++ ones, which would allow people to run the existing code through a transpiler which for 80% of the lines can be translated directly to working code in the new language, and produces diagnostic errors for the other 20% that explain what needs to be changed.
Which removes 80% of the work from transitioning an existing codebase to the new language. And then more people do that.
This x2. It's only C++ if you can optionally refactor random lines of code in a project written in compliance with C++98 using new features and still get a valid program.
Once you mess with the language in a way that your C++98 code is either not supported or requires rewrites or redesigns to comply with the new standard then quite obviously it is not C++ anymore.
Semver matters.
Java is in a similar situation, which is probably why they market the minor version.
The minor version for Java has stopped being used by Java 9.
Our shops won't cut over to something beyond Java 11 until Java 17 (LTS version) hits in a few years. Just happy that we were able to make the jump from 8 to 11.
Poor Android Java developers are now faced with being stuck with a pseudo Android Java subset of Java 8, or migrate to Kotlin.
While any Java library author that wants to share their work with the Android community is faced with the dilemma of keeping two parallel versions of their library, forget about Android, or also convert themselves to Kotlin as well, even if that isn't their thing.
You mean so much of C++'s success is due to backward compatibility with C.
Edit: Formatting.
So there is a path, but it has to take an upgrade-path into account for big legacy projects. In practice, moving to a new compiler can be a big project for companies with larger code bases. Sensible deprecation of features could (and in practice) is part of that.
Even though the standard is quite careful, in practice big C++ projects sometimes rely on non-conforming behavior of the specific compiler version they use. An example is the MSVC template support which allowed constructs that the standard didn't.
Your example is a really poor one. The volatile keyword was only used to provide hints to the compiler on whether or not it could apply some optimizations. Omiting a volatile keyword is perfectly backward compatible, as is compiling code without support for volstile. Thus the only practical consequence of adding/removing random volatile keywords from your source code was how your build could (and not should) be optimized.
That's a far cry from, say, remove malloc().
On MSVC, volatile has traditionally meant something akin to atomic. It still will, regardless of what the Standard says and other compilers do.
"Can't" is a big word when applied to the ISO committee. They would want to consider the magnitude of consequences. Since all implementations would provide a "make like before" switch, consequences would be limited.
No, at best you saw code that expects the compiler to not apply specific optimizations that C++ doesn't suggest or enforce. It's a compiler issue, thus if anything that code needs to be compiled accordingly.
https://www.cs.utah.edu/~regehr/papers/emsoft08-preprint.pdf
class foo {
void i_am_deprecated() volatile;
void i_am_also_deprecated(volatile int);
void i_am_not(volatile char*);
};
The committee does a lot of work to keep the standard compatible with c++ as it is used and this proposal seems to be a good example of them trimming out dead and bloated parts, without negatively impacting existing code.
The original proposal for that deprecation points out a lot of issues. Basically it boils down to the fact that many operations were badly specified, outright misleading, predate the c++ memory model or were artifacts of maintaining the parallel to const that volatile had in C at a large cost and without people actually using it, one example they could track down even noted how it did not behave the same way as primitives would.
Well, just FYI the "new" C++ is not the "old" C++ anymore. Languages "evolve". Try to compile an old program on the new compiler. The same is valid for other languages (C, fortran, perl).
> The C++ standards committee can standardize a backwards-incompatible "new" C++ for all I care as long as they don't call it C++.
They already do this. For C for example you have C89, C93, etc.
Same for C - that old library in C89? Still usable in modern projects.
I am sure there are broken C compilers out there which don’t implement entire language. Heck, I use one sometimes. But I don’t think this proves your point at all.
You need better examples.
Like always on the C and C++ community, "works on my compiler" and "what the standard states about it" is not the same.
This is a real power of C and C++ -- each file gets its own mode. That K&R library you made in ancient time? C++98 classes? All still works and links to the modern stuff. You can be sure that whatever code you wrote today would still be usable, possibly with minor modifications, 20 years later. Sure, the compiler names and build systems will change, but the code would work.
Honestly I would've expected an error with -std=c17 -Wall -pedantic from GCC and Clang since gets() was removed in the C11/C17 standards.
Nowadays I'm mildly annoyed if a project is on one of those "ancient" 3 but ≤3.6 Python versions and I can't use the nice features in 3.7+...
No, it is a mess and it was a failure.
It tooks more than 10 years to happen, and even so...Many tools will die with python2 and never be migrated.
If your codebase is not usable anymore because your language evolved. That's a failure.
I can still compile the C's from the 80's today if I want to. Same for Fortran, Same with C++.
Unicode change between 2 and 3 cause behaviour change on your input/output and any serialisation.
Even if your code could have been transpiled properly, the behaviour of your code would have still change, and still introducing bugs
> I can still compile the C's from the 80's today if I want to. Same for Fortran, Same with C++.
I couldn't disagree more. It's a cute trick but not useful that code from 80's is still compilable with a compiler from 40 years later.
There's so many better ways to handle legacy like that, like just freezing the toolchain. Which you've almost certainly long since done if you're actually working with code that old, as the platform it runs on is also frozen.
You can't actually run code from 80's out of the box today (16-bit legacy support has long since stopped being a thing), so why does it matter if it still compiles? And, critically, why does it matter if it compiles with the latest toolchain?
There needs to be a supported sliding window of a generously long time, but it doesn't need to be ~infinite support. That's entirely unreasonable & unnecessary. We don't expect that of any OS, library, etc... so why should we expect it from a compiler & language?
Still a good number of packages you are using right now on your Linux distribution have a core more than 15-20 years old and still they compile only with minor modifications with the last GCC.
Compatibility does matter. It should however not prevent evolution. If handle properly, nothing block us to get a proper versioning and evolution path on the toolchain we use.
And on many aspect C++ up to now has been very successful to do that.
> You can't actually run code from 80's out of the box today
Good C is close to immortal. Doom is from 93, and still you can compile it and run it with the last compiler in 2020 with minor modifications.
So now we're just debating the size of ongoing breakages instead of just their existence.
And no good C is not close to immortal. Good luck compiling K&R C with modern compilers, for example.
If you wrote your 1980s C code:
1. Using the proper C standard. 2. Using the standard library (and perhaps even common Unix libraries). 3. Were careful not to make assumptions about sizes (which, by the language standard, you shouldn't make anyway).
The your code will compile and work just fine. If you used Makefile's to build it, which you very well may have, and you were using a standard (= Unix) system, there's not a bad chance your build system might work too. :-)
You mean the standard that didn't exist until 1989?
> Using the standard library (and perhaps even common Unix libraries)
That's also very late 80s (realistically early 90s) stuff.
Which was still yes a very long time ago that C made all these breaking changes, but it was also a good ~20 years after it's release as well.
Fair point. But K&R C with System V or BSD libraries will probably-kinda-sorta work.
Only if it claimed to be eternally BW compatible. Some languages never made that claim.
To me a language is a dependency like any API, just with a much bigger surface touching my code. I'd say it's the hardest API to replace in any code. "BW compatibility" is just as much a feature as "breaking BW compatibility", I just like languages (or any API really) to be upfront about it's policy toward BW compatibility.
And I've observed a lot of langs change their attitude towards it over time. Especially considering pre-1.0 times, I find a lot of langs happily break BW compatibility in their early days.
In short time, is bad. The more time pass, is good.
The cost of STUPID C/C++/JS behavior is measured in the billons.
Is not the makers of this langs aware? Yes. They know how could be fixed? Sure. Then WHY is not fixed?
Because not considering time in the calculation (and how many MILLONS OF PEOPLE are affect). But today, are DECADES behind of 100% certainty of the cost of this mistakes, and proved beyond doubts that the langs that fixed them ARE better.
In the face of reality and facts, why resist so much?
---
The thing is HOW get out of this mess?
The big irony is that the JS world show how (but not commit with the proper force): Build transpilers. Make "Better C++" that transpile to "BAD C++". Make clear that everyone must move forward, but provide this as partial step.
Make auto converters. Fix damm stupidity like dangling IFs, (seriously some stuff is not brainer). Have a clear vision in how move forward.
And drop the ego. C/c++ not need to turn into Rust, but why believe it could not truly get better?
Now, in the case of C/C++ exist a big trouble in the case of being the "de facto" ABIs for the rest of the world. Apple have the same issue with swift -> ob-c with the case of nullability (which apis never return null even if in theory could?).
Them do annotations to the APIs. This could allow to mechanize the transformations and provide ABI translations that could be injected in the compiler.
For example:
ABI STEP 1 (annotate only):
fn evil():NullableString //Imply NullableString = Option(Str)
[safe(NullTerminatedString == Str)]
fn not_evil(): NullableString
ABI STEP 2 (rewrite):
fn ex_evil():Option<Str>
fn not_evil():Str
Eventually, when the calls are converted this step is erased and the runtime penalty removed, then all become good.
Nevertheless, in Ethernet's case, while it changed backwards incompatibly at the beginning, since then, consumer Ethernet has been backwards compatible all the way from 1990's 10BASE-T (10Mbit) to the still-rare 10GBASE-T (10Gbit). In data centers you have faster variants that require different cables, for routers... but even there I believe the connections to individual servers typically use the old twisted pair.
Maybe that's what Python 3 should have done: don't remove stuff, but just make it discouraged and let a python3only" flag treat them as errors when ready.
What we got was kind of a mess. Because python2 is very often valid python3... But not always.
1. "New C++" would not just be a subset of the syntax of old C++.
2. ABI issues.
Also:
> I think it also makes sense for C++ developers whose projects are relatively short-lived, e.g. some games.
There is a difference between games and engines - the games might be short lived, but most C++ engines are very long lived, going back decades (e.g. see how the recent Half-Life Alyx can trace its codebase back to Quake 1 or pretty much anything based on the Unreal Engine going back to the mid-90s - there are screenshots of UE1 running on what appears to be either Windows 3.1 or NT 3.51). Even when "engines" are being made "from scratch" this is often a marketing gimmick and the "from scratch" was just renaming the engine because it got a new rendering model and people can't tell the difference. EVEN when this isn't case, often new engines are based on existing code and/or libraries.
The to abandon backwards compatibility in C++ is to make a mockery even of the name of the language (see the "C" in there). If they want to create a new language they should call it something different. Willfully abandoning backwards compatibility and keeping the name is an abuse of one of the great brands in software.
I consider this more a case of the ossification of search.
Neither search engines nor specific sites like Stack Overflow deal with the fact that information can switch from right to wrong with age.
If the python 3 interpreter could still run python 2 code. If you could mix and match python 2 and 3 code on a per-module, per-file or even per file basis, then the transition could have been so much smoother.
- the languages were not that different, so no need to learn the new version
- python core devs gave 10 years to make the transition. Then extended it.
- it was possible to write code running on python 2 and 3
- python is very expressive, hence the code base have way less numbers of lines than in C++
- python cared only about one implementation, Cpython. The rest of the world needed to follow. Some didn't, like Jython and stackless, and the community didn't blink.
Despite all that, the transition was very painful.
This doesn't match reality. In reality, many Python projects have way more lines of code than equivalent C++ projects. Probably because of the 'expressiveness' you cite; you can't really showcase your love of coding and job security though artificial complexity without 'expressive' bells and whistles. (This is the idea that lead to languages like Go, I'm pretty sure.)
That said, C++ is plenty 'expressive' itself.
That is mathematically impossible. Even if the syntaxes were exactly the same (which they are not, python syntax is on average shorter), the low level nature of C++ requires your code to do operations that Python does need to do, such as memory management.
It's like stating the sky is red.
I'd also argue that operator overloading etc lets C++ be just as expressive as Python, the libraries just need to be designed with that in mind.
You mean in the same way C++ dev write tests to prove that all their code has no memory error which you get in Python for free ?
Except:
- tests are way shorter to write in python than in C++
- C++ devs often write zero tests for their code, just like python devs
- python duck typing + REPL means compiler checks are rarely necessary
- if you need to be type checks, you use type hints in python, which even then is still less verbose than c++
> I'd also argue that operator overloading etc lets C++ be just as expressive as Python, the libraries just need to be designed with that in mind.
Ok, let's say you have this json:
[{
"name": "Kévin",
"age": 23,
"hired": "2005-06-03 02:12:33",
"emails": ["kevin@foo.com", "kevin@bar.com"]
}, {
}, {
"name": "Joël",
"age": 32,
"hired": "2003-01-02 12:32:11",
"emails": ["joel@foo.com", "joel@bar.com"]
},
... other entries
]
You want to print people details in alphabetical order this way:
Joël (32) - 02/01/03:
- joel@foo.com
- joel@bar.com
Kévin (23) - 03/06/05:
- kevin@foo.com
- kevin@bar.com
... other entries
In Python 3.7, which is already 2 years old, the code would be:
import json
import datetime as dt
with open("agenda.json") as fd:
agenda = sorted(json.load(fd), key=lambda people: people["name"])
for people in agenda:
hired = dt.datetime.fromisoformat(people["hired"])
print(f'{people["name"]} ({people["age"]}) - {hired:%d/%m/%y}:')
for email in people["emails"]:
print(f" - {email}")
It's not specific to Python, you would get this expressiveness with Ruby or Perl.
I don't see in which world you would get that in regular, honest to god, day to day, portable C++.
You have to declare types, many includes, you'll have headers and a main function. You have the memory and references to manage.
It doesn't make C++ a bad language.
It doesn't make python a better language.
The C++ version will take way less RAM than the Python version for example.
It's just the nature of those languages implies that.
use chrono::NaiveDateTime;
use serde::*;
use serde_json;
#[derive(Deserialize)]
struct Person {
name: String,
age: u32,
hired: String,
emails: Vec<String>,
}
fn main() {
let data = std::fs::read_to_string("agenda.json").unwrap();
let mut people: Vec<Person> = serde_json::from_str(data).unwrap();
people.sort_by(|a, b| b.name.cmp(&a.name));
for person in &people {
let datetime = NaiveDateTime::parse_from_str(&person.hired, "%Y-%m-%d %H:%M:%S").unwrap();
println!(
"{} ({}) - {}",
person.name,
person.age,
datetime.format("%d/%m/%y")
);
for email in &person.emails {
println!(" - {}", email);
}
}
} #include <iostream>
#include <sstream>
#include <fstream>
#include <iomanip>
#include <nlohmann/json.hpp>
#include <range/v3/action/sort.hpp>
int main()
{
using namespace nlohmann;
using namespace ranges;
const json parsed = json::parse(std::ifstream("/tmp/json/test.json"));
std::vector agenda(parsed.begin(), parsed.end());
sort(agenda, {}, [] (const auto& j) { return j["name"]; });
for(const auto& people : agenda) try {
std::tm t{};
std::istringstream(people["hired"].get<std::string>()) >> std::get_time(&t, "%Y-%m-%d %H:%M:%S");
std::cout << people["name"] << " (" << people["age"] << ") - " << std::put_time(&t, "%d/%m/%y") << ": \n";
for(const auto& email : people["emails"])
std::cout << " - " << email << "\n";
} catch (...) { }
}Now I would buy your argument if C++ were as conservative as C, but it clearly isn’t. It keeps trying to push itself as a modern general purpose language, while ignoring what it actually takes to be useful as one.
We went from:
> many Python projects have way more lines of code than equivalent C++ projects
To:
> It's a bit longer in C++ but frankly not by that much
With the proof being literally __twice__ more characters, with one line being more than a 100 characters long.
I understand that C++ doesn't have a native JSON lib, and sorting is a bit out there, but giving 3rd party lib access to Python feels like cheating:
import pandas as pd
agenda = pd.read_json("agenda.json", convert_dates=["hired"])
for _, (name, age, hired, emails) in agenda.sort_values("name").iterrows():
print(f"{name} ({age}) - {hired:%d/%m/%y}:")
for email in emails:
print(f" - {email}")
Now again, this is not so say "Python is better than C++". That's not the point.
If you want to write a game engine, C++ makes sense, verbosity doesn't matter, and you don't want the Python GC to kick in. If you want to iterate on your Saas product API quickly, it's probably not the best choice.
Why pretend otherwise? What the point of stating the sky is red?
Also, the difference in tone towards jcelerier compared to steveklabnik seems not warranted here. Both IMHO contributed an educational example. Your response here equally does (or does not) apply to the Rust version, doesn't it?
And modern C++ is not 15 years old C++. You have auto, lambda, etc. I just think it's silly to try to pretend to be more expressive than scripting languages.
It's like Schwarzenegger trying to pretend it's as flexible as gymnast. What for? The guy is amazing at what he does.
So far we have a Python, a Rust and a C++ program in this thread which output the same thing given the same input, which, by the definition of expressive power of programming languages means that they are exactly as expressive as each other (given this information only and notwithstanding small differences such as error handling in this case).
Also you mentioned in your original thread :
> Even if the syntaxes were exactly the same (which they are not, python syntax is on average shorter)
indeed
> the low level nature of C++ requires your code to do operations that Python does need to do, such as memory management.
Notice that there is nothing which looks like memory management in my code. There are entire classes of programs that can be written without thinking about when to allocate or free your memory at all - just declare your variables on the stack and with std types, and that is handled for you (and that's not a modern C++ thing ! it's been like that since the very beginning). Opting in to memory management is an active, conscious effort in these cases (or a residue of people who only were taught java / c# at school and sprinkle their code with `new`).
Neither will C++, as far as I know, or do you mean that the compiler will validate the formatting of the string literal (python has linters that will validate this too!).
Hell, even with statically verifiable types, the python is shorter.
With iostreams there are no format strings so there is nothing to validate. Libraries such as https://github.com/fmtlib/fmt are allow to do compile-time parsing of the string though.
If you've done any serious (large, long, multi-team) projects in a dynamically-typed language, you know how quickly they turn into a big ball of mud where you're spending more time refactoring your refactorings than writing useful code.
Utterly pointless and reputationally ruinous. I don't do serious work in Python any more
There is a fabulously unending depth to explore in the chasm lying between these opposing world views. It would be more than possible to write a book on the topic and fail to cover it all, however here are some of the most important aspects, from my perspective at least:
* given a perfectly functional tool relied on heavily by its user to perform their job, and given that tool suddenly decides to change shape such that it no longer fits the user's hand without retraining, nor fits with the remainder of the user's toolset, including custom tools the user has invested in producing, the continued utility of the no-longer-functioning tool is called into question, along with a deserved reappraisal of the tool's applicability in the context of the user's original intended problem domain.
* when the reason for its reshaping is to solve what are highly important problems from the perspective of the tool, but much less so from the perspective of the average user, and that user's application of the tool to real world problems, it can no longer be said that the tool is simply an implement that may be called and relied upon at any point in future -- the tool develops a chaotic life and importance all of its own, and may choose to reshape once again at any future moment (and indeed in this case it has). It is no longer a tool, but some sentient entity demanding unpredictable ongoing costs and attention paid all of its own.
* given a tool that promises to cease functioning 'correctly' at any future moment based on its own whim, preferences, industry fashions and styles, in an ecosystem where many similar such tools exist that explicitly promise not to cease functioning over the same time period, it is a fool's errand to pick the tool that promises to externalize additional costs on the user when alternatives exist that avoid any such cost.
* given tool designers who externalize almost frivolously minor technical costs on to every user, where each 'minor' change is amplified perhaps 10000 times over and directly translates into expensive engineering time, the question is easily raised whether the philosophy of the tool is appropriate for its advertised utility, and whether continued reliance on the tool makes business sense. In economic terms, what was the cost to productivity of the retraining and re-tooling of users compared to any alleged future productivity improvement?
* had I written these scripts in bash, C# or C++, they would not have broken even remotely to the same degree. Of course these are not some completely unevolving entities either, however all take the promise of forwards compatibility deadly seriously, and it is more than possible to find 10-20 year old programs written in C++ or bash that continue functioning to the present day. From my perspective, they are therefore excellent and highly dependable tools.
> print vs print() is what I think the parent comment was referring to.
as another person mentioned is likely what you meant, but it was easy to take your comment as meaning that you had mismatched parens somewhere, which would imply broken python2, as well as 3.
> * given a tool that promises to cease functioning 'correctly' at any future moment based on its own whim, preferences, industry fashions and styles, in an ecosystem where many similar such tools exist that explicitly promise not to cease functioning over the same time period, it is a fool's errand to pick the tool that promises to externalize additional costs on the user when alternatives exist that avoid any such cost.
This is clearly a falsehood, if the tool provides additional value. To bring things back to the topic at hand, if C++ were allowed to break ABI compatibility in very specific ways, it could be faster. stl container types are slower than many third party ones (absl, for example).
Which is to say that if you want "the best" that C++ has to offer, you have to be willing to have your libraries make backwards incompatible changes.
To jump back to python,
> given tool designers who externalize almost frivolously minor technical costs on to every user, where each 'minor' change is amplified perhaps 10000 times over and directly translates into expensive engineering time
I disagree that this happened. The examples you give are trivially fixable with 2to3. There are harder problems that 2to3 doesn't solve, but it sounds like you don't have any, so the frivolously minor technical costs are frivolously minor, and translates into running `2to3 my_code/` one time, to fix all the missing parens and most of the moved functinos.
> bash that continue functioning to the present day
I have yet to encounter a 20 year old program written in bash that functions to this day. It might be syntactically valid, but it won't do what it intended to do.
You're welcome to absorb all the externalized costs your heart desires, but in future please consider reviewing HN's rules before bandying attacks like "smug" and "superior".
Bluntly, if I could break the ABI to get a 10% speed boost across the board, is that not worth it?
And I'm well aware of the site guidelines. I certainly don't think asking someone to tone down the holier-than-thou in their comments is a violation of them. Just the opposite, it's encouraged. So I'll continue to ask that you do so if you choose to respond.
It's just that aghast people are the most vocal. You don't hear the 90% of people that are happy. They don't take the time to speak up. But the unhappy complain all the time.
There are programs still using COBOL.
That's life in IT.
Such as? What actively developed tools or ecosystems rely or integrate with python2 exclusively?
It would be unwise to design an API that promotes a niche need like dealing with a legacy file system with corrupted file names. This is what Python 3 fixed, making the easy things easy, and the complicated things possible, not the other way around.
But Python 3 is absolutely up to the task of handling legacy file systems with random encoding mixed in, you just need to tell it explicitly you are doing so.
Let's create a file with a completely garbage name, made of random bytes, which is allowed on Unix:
>>> import os, sys
>>> sys.version_info
sys.version_info(major=3, minor=7, micro=5, releaselevel='final', serial=0)
>>> with open(os.urandom(32), 'wb') as f:
... f.write(os.urandom(200))
>>> filename = os.listdir(b'.')[0]
>>> type(filename)
<class 'bytes'>
>>> filename[:10]
b'\xf8-U\xa5\x1dq\xad?\xbf\xa2'
>>> data = open(filename, 'rb').read()
>>> data[:10]
b'E\x05\xce*M \xf5\xfeK\x18'
>>> type(data)
<class 'bytes'>
Some API in Python require text. If you want to pass the file names to those API, you can use surrogate escape, which let you convert back and forth between arbitrary bytes and utf8 text, without loosing information:
>>> as_text = filename.decode('utf8', errors='surrogateescape')
>>> as_text[:10]
'\udcf8-U\udca5\x1dq\udcad?\udcbf\udca2'
>>> type(as_text)
<class 'str'>
>>> as_text.encode('utf8', errors='surrogateescape') == filename
True
If you have to do a robust version of this with Python 2, you will have to do that anyway: at some point mixed encoding will bite you if you don't have a neutral representation for them. Python 2 gave you the illusion of robustness, because it said "yes" to most operations.
I remember quite well that a lot of Python 2 programs didn't work in Europe because your user directory would contain your name, which could be non ascii. Python 2 programs are opt in to deal with it. It's the opposite philosophy, and caused so many crashes.
If all you need is to work with arbitrary mixed bags of file names, you can just pass "str", and Python will use automatically and transparently surrogateescape everywhere.
It's not a surprise, as most softwares need iterations to get get good. Python 2.7 has not started as the amazing tool it is now, and as I started my career with 2.4, you remember some funny stuff.
This is why Python 2.7 was kept around for 13 years after Python 3 first came out.
Now, 3.6 came out in 2016. It solves many issues 2.7 had, and add tons of goodies. It's very ergonomic, can be installed easily. It's a great software.
It's 2020, let's enjoy the goodness of Python 3.
And now it has been better for the last 5 years, and does solve the problems it intended to solve.
The only thing this means is that the botched upgrade did not end up killing Python; it says nothing about whether it was done badly or not.
(I have nothing against python, I just believe it is important to understand why and how what happened happened to avoid similar errors in the future)
We are at Python 3.8.
Python 3 has been around for 13 years, during which Python 2.7 was still supported.
Except slowly people will be writing parts of it in Rust, or they would be using libraries written in Rust (because availability), and then after a while people get annoyed of using more than one language and the whole project will be rewritten in Rust. Of course, at that point, a new language will be invented and Rust code will be looked at as rusty :)
Wishful thinking.
https://kennykerr.ca/2019/11/05/rust/
https://twitter.com/benwilliamson/status/1240113606374686721
And realistically I'm not entirely sure how you can add C++ interop to Rust without also ruining Rust. It's a hard problem to say the least.
Source?
John Carmack said in 2004 that "there are still bits of early Quake code in Half-Life 2" - https://github.com/ESWAT/john-carmack-plan-archive/blob/mast...
In all seriousness, Half-Life Alyx is based on Source 2 which itself is based on Source 1 which is based on GoldSrc which is based on Quake 1.
While true the engines are also staffed and can absolutely keep up with changes. They regularly are tasked with new toolchains, platforms, OSes, and APIs to leverage already. This would just be more of the same, and if you look at the goals of the paper: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p213...
That lines up pretty damn well to what game engines want as well.
> One of the main reasons to stick with C++ (and C) is that the investment you put in the code you're writing right now, no matter how fugly it is, will be there and keep working in the future.
I think you should read the actual paper instead. http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p213...
The things they talk about actually abandoning are things like byte sizes other than 8-bits or source code in file systems that don’t support file extensions or nested directories.
Chances are any existing C++ codebase is already not compatible with such systems that they want to more formally deprecate/abandon in the first place.
> Our experience is that providing broad ABI-level stability for high-level constructs is a significant and permanent burden on their design. It becomes an impediment to evolution, which is one of our stated goals.
This does not name it, but doesn't this refers to all those discussions about changing internal representation of std::string, std::vector and so on?
This does sound pretty scary. Once you have separate classes for "C++23 std::string" and "older std::string", you can no longer link together files built with different versions.. and this means you lose a lot past compatibility.
True but this is already something that isn't really "supported" broadly. Some standard libraries work hard to keep that working, but others tend not to, and nobody knows if it's supposed to work reliably or not. And support for this binary compatibility also varies across platforms, like I don't think Microsoft bothers since you're supposed to always bundle a specific version of the runtime with your binary.
Single-version prebuilt libraries become more of a problem, but are those that widespread vs. source-based libraries?
But I think the broader point here that the paper is really pushing is for the committee to actually make a statement on things like this instead of their historically vague "it sort of works but we're not saying it will or indeed saying anything at all"
https://youtu.be/tISy7EJQPzI?t=2209
I remember watching another talk where they propose something similar for the language itself, but I can't find it at the moment.
Maybe you could do like Rust with their "Epoch" system that lets you interoperate code using various standards in order to migrate progressively without breaking backward compatibility. I suspect that it would be a lot harder to make it work for C++ however, mainly due to its extreme reliance on #includes (especially for anything using templates) and more common use of macros.
I'm not saying it's impossible but I suspect that it would fragment the ecosystem quite a bit. Removing "old features" tends to have massive side effects in a language like C++ with metaprograming, overloading, multiple inheritance, unlimited macro usage and complex symbol resolution rules.
So I think "why not just use Rust/D/Zig/Nim/Crystal" is warranted feedback for these proposals (and you could probably add Go, C#, Java and a few others).
However, there's an active (and in my understanding, decently received) proposal based on it for C++ that is called "epochs".
Such a solution is only possible if driven by the standard, otherwise it will never be fully available, just like the static analysis tooling varies by vendor.
Per Hyrum's law, someone will rely on any given behavior. So, if you choose backwards compatibility, it becomes harder for the committee to evolve the language over time.
The path of breaking changes can be made less painful, and does not, necessarily, invalidate all the code that is already written.
With that being said, the problem right now is not the decision of keeping backwards compatibility or not, but the fact that the standard should be explicit about it, so people know what to expect.
Much of this proposal would create insurmountable engineering problems for everything except the Google monorepo, or similar infrastructure at $1T companies building monoliths.
Historically, most innovation has happened outside of those environments.
Sabotaging the toolchains of smaller shops is not going to pan out well for the industry at large.
C++ 11, 14, 17, 20, and 23 are all binary compatible (modulo some possible oversights). There are many quality-of-life improvements that got voted out because then a binary compiled in C++11 wouldn't be compatible with one built in C++23 anymore.
As a simple example ([1] has many more): the only reason we have both std::scoped_lock and std::lock_guard is because changing the definition of the latter would have required re-compilation of existing code, so a new class had to be introduced.
ABI stability even overrides performance in committee decisions (see the std::regex example, but I also heard it affects the footprint of unique_ptr - edit: found it [2]).
tl;dr breaking source code compatibility seems way off if implementors don't even dare to break binary compatibility.
[1] http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p202...
I'm currently working on a C++ project that has been taken out of cold storage and when the old rewrite question pops up the tech stacks considered for the job do not feature C++ at all. The main reason is that nowadays there are better tools and frameworks to implement non-performance critical parts of an application such as the GUI, which incidental takes over more than half the total lines of code, and the rest can be implemented easily with small support services developed in whichever tech stack you choose. Thus C++'s "jack of all trades" trait is no longer a major factor in the decision. Meanwhile the lack of a stable ABI starts to feature prominently as a reason not to adopt C++.
Introducing a new ABI in competition with the current ones (e.g. ia32 and x64) would be less destructive than officially abandoning those currently in use.
Seems like a terrible thing to break this late
And an older link:
https://www.auburnsounds.com/blog/2016-11-10_Running-D-witho...
Like networking support, serialization for ubiquitous formats like JSON, etc.
These are some of the ones off the top of my head, there are more. So the fix in these cases is another name, and that has the problem of making the language needlessly more complicated and harder to learn.
Sure as time passed, the non backward compatible C++ would add features not currently in scope for C++, but the radical redesign of the fundamental language is not what this proposal suggests.
Those two languages were created without consideration of backwards compatibility with C/C++ source code and were designed for the same job. It stands to reason that given no constraints of compatibility, you'd get something similar to those languages.
Now obviously, those languages do not have all the features of C++ nor the same design principles. But as soon as you have all the features of C++ with the same principles, you have something so close to C++, that it doesn't make sense to break backwards compatibility.
But: if it's a completely new language (no backward compatibility to C++) - why do we need this? Can't we just use one of those new languages that are already available (Rust, Zig, Odin, Nim, just to name a few).
There are probably also strong non-rational components, e.g. C++ people would like to keep working on C++, leveraging their existing skills and status, rather than jump to another community.
Would backwards compatibility to "modern C++" make more sense than no backwards compatibility?
what if you actually do like templates and class-based OOP ?
C++ compiles to native code via LLVM bitcode.
(btw., since debate initiated, C++ standard attempts on garbage collection too - as well as on many powerful concepts incorporated into C# from here and there - and IL is compiled into machine code before execution, naturally, also making C# really had an intention of making a better but similar C/C++ without its crippling compatibility barriers towards improvements. C++ is more universal still - but considerably less than C - and it is all right, each have advantage here and there. You usually do not work on all use cases anyway! ;) No need to abandon backward compatibility with C++, abandoning will become a new language, we have quite a few already, it is done already!)