Trip report: Summer 2021 ISO C++ standards meeting (virtual)

On Monday, the ISO C++ committee held its third full-committee (plenary) meeting of the pandemic and adopted a few more features and improvements for draft C++23.

We had representatives from 17 voting nations at this meeting: Austria, Bulgaria, Canada, Czech Republic, Finland, France, Germany, Israel, Italy, Netherlands, Poland, Russia, Slovakia, Spain, Switzerland, United Kingdom, and United States. Slovakia is our newest national body to officially join international C++ work. Welcome!

We continue to have the same priorities and the same schedule we originally adopted for C++23, but online via Zoom during the pandemic.

This week: A few more C++23 features adopted

This week we formally adopted a third round of small features for C++23, as well as a number of bug fixes. Below, I’ll list some of the more user-noticeable changes and credit all those paper authors, but note that this is far from an exhaustive list of important contributors… even for these papers, nothing gets done without help from a lot of people and unsung heroes, so thank you first to all of the people not named here who helped the authors move their proposals forward! And thank you to everyone who worked on the adopted issue resolutions and smaller papers I didn’t include in this list.

P1938  by Barry Revzin, Richard Smith, Andrew Sutton, and Daveed Vandevoorde adds the if consteval feature to C++23. If you know about C++17 if constexpr and C++20 std::is_constant_evaluated, then you might think we already have this feature under the spelling if constexpr (std::is_constant_evaluated())… and that’s one of the reasons to add this feature, because that code actually doesn’t do what one might think. See the paper for details, and why we really want if consteval in the language.

P1401 by Andrzej Krzemieński enables testing integers as booleans in static_cast and if constexpr without having to cast the result to bool first (or test against zero). This is a small-but-nice example of removing redundant ceremony to help make C++ code that much cleaner and more readable.

P1132 by Jean-Heyd Meneide, Todor Buyukliev, and Isabella Muerte add out_ptr and inout_ptr abstractions to help with potential pointer ownership transfer when passing a smart pointer to a function that is declared with a T** “out” parameter. In a nutshell, if you’ve ever wanted to call a C API by writing something like some_c_function( &my_unique_ptr ); then these types will likely help you. The idea is that a call site can use one of these types to wrap a smart pointer argument, and then when the helper type is destroyed it automatically updates the pointer it wraps (using a reset call or semantically equivalent behavior).

P1659 by Christopher DiBella generalizes the C++20 starts_with and ends_with on string and string_view by adding the general forms ranges::starts_with and ranges::ends_with to C++23. These can work on arbitrary ranges, and also answer questions such as “are the starting elements of r1 less than the elements of r2?” and “are the final elements of r1 greater than the elements of r2?”.

P2166 by Yuriy Chernyshov helps reduce a commonly-taught pitfall with std::string. You know how since forever (C++98) you can construct a string from a string literal, like std::string("xyzzy")? But that you’d better watch out (and you’d better not cry or pout) not to pass a null pointer, like std::string(nullptr), because that’s undefined behavior where implementations aren’t required to check the pointer for null and can do just whatever they liked, including crash? That’s still the case if you pass a pointer variable whose value is null (sorry!), but with this paper, as of C++23 at least now we have overloads that reject attempts to construct or assign a std::string from nullptr specifically, as a compile-time “d’oh! don’t do that.”

We also adopted a number of other issue resolutions and small papers that made additional improvements, including a number that will be backported retroactively to C++20. Quite a few were of the “oh, you didn’t know that rare case didn’t work? now it does” variety.

Other progress

We also approved work on a second Concurrency TS. Recall that a “TS” or “Technical Specification” is like doing work in a feature branch, which can later be merged into the C++ standards (trunk).

Two related pieces of work were approved to go into the Concurrency TS: P1121 and P1122 by Paul McKenney, Maged M. Michael, Michael Wong, Geoffrey Romer, Andrew Hunter, Arthur O’Dwyer, Daisy Hollman, JF Bastien, Hans Boehm, David Goldblatt, Frank Birbacher, Erik Rigtorp, Tomasz Kamiński, and Jens Maurer add support for hazard pointers and read-copy-update (RCU) which are useful in highly concurrent applications.

What’s next

We’re going to keep meeting virtually in subgroups, and then have at least one more virtual plenary session to adopt features into the C++23 working draft in October.

The next tentatively planned ISO C++ face-to-face meeting is February 2022 in Portland, OR, USA. (Per our C++23 schedule, this is the “feature freeze” deadline for design-approving new features targeting the C++23 standard, whether the meeting is physical or virtual.) Meeting in person next February continues to look promising – barring unexpected surprises, it’s possible that by that time most ISO C++ participating nations will have been able to resume local sports/theatre/concert events with normal audiences, and removed travel restrictions among each other, so that people from most nations will be able to participate at an in-person meeting. But we still have to wait and see… we likely won’t know for sure until well into the autumn, and so we’re still calling this one “tentative” for now. You can find a list of our meeting plans on the Upcoming Meetings page.

Thank you again to the hundreds of people who are working tirelessly on C++, even in our current altered world. Your flexibility and willingness to adjust are much appreciated by all of us in the committee and by all the C++ communities! Thank you, and see you on Zoom.

GotW #102 Solution: Assertions and “UB” (Difficulty: 7/10)

This special Guru of the Week series focuses on contracts. Now that we have considered assertions, postconditions, and preconditions in GotWs #97-101, let’s pause and reflect: To what extent does a failed contract imply “UB”… either the Hidden Dragon of Undefined Behavior, or the Crouching Tiger of Unspecified Behavior?

1. Briefly, what is the difference among:

(a) undefined behavior

Undefined behavior is what happens when your program tries to do something whose meaning is not defined at all in the C++ standard language or library (illegal code and/or data). A compiler is allowed to generate an executable that does anything at all, from data corruption (objects not meeting the requirements of their types) to injecting new code to reformat your hard drive if the program is run on a Tuesday, even if there’s nothing in your source code that could possibly reformat anything. Note that undefined behavior is a global property — it always applies not only to the undefined operation, but to the whole program. [1]

(b) unspecified behavior

Unspecified behavior is what happens when your program does something for which the C++ standard doesn’t document the results. You’ll get some valid result, but you won’t know what the result is until your code looks at it. A compiler is not allowed to give you a corrupted object or to inject new code to reformat your hard drive, not even on Tuesdays.

(c) implementation-defined behavior

Implementation-defined behavior is like unspecified behavior, where the implementation additionally is required to document what the actual result will be on this particular implementation. You can’t rely on a particular answer in portable code because another implementation could choose to do something different, but you can rely on what it will be on this compiler and platform.

2. For each of the following, write a short function … where if the assertion is not checked and is false then the effect:

(a) is always undefined behavior

Easy peasy! Let’s dereference a null pointer:

// Example 2(a): If assert is violated, always undefined behavior

void deref_and_set( int* p ) {
    assert( p );
    *p = 42;

The function asserts that p is not null, and then on the next line unconditionally dereferences p and scribbles over the location it points to. If p is null and the assertion checking is off so that we can get to the next line, the compiler is allowed to make running the whole program format our hard drive.

(b) possibly results in undefined behavior

A general way to describe this class of program is that the call site has two bugs: first, it violates a precondition (so the callee’s results are always at least unspecified), and then it additionally then uses the unspecified result without checking it and/or in a dangerous way.

To make up an example, let’s bisect a numeric range:

// Example 2(b): If assert is violated, might lead to undefined behavior

int midpoint( int low, int high ) {
    assert( low <= high );
    return low + (high-low)/2;
        // less overflow-prone than “(low+high)/2”
        // more accurate than “low/2 + high/2”

The author of midpoint could have made the function more robust to take the values in either order, and thus eliminated the assertion, but assume they had a reason not to, as alluded to in the comments.

Violating the assertion does not result in undefined behavior directly. The function just doesn’t specify (ahem!) its results if call sites call it in a way that violates the precondition the assertion is testing. If the precondition is violated, then the function can add a negative number to low. But just calculating and returning some other int is not (yet) undefined behavior.

For many call sites, a bad call to midpoint won’t lead to later undefined behavior.

However, it’s possible that some call site might go on to use the unspecified result in a way that does end up being real undefined behavior, such as using it as an array index that performs an out-of-bounds access:

auto m = midpoint( low_index(arr1), high_index(arr2) );   // unspecified
   // here we expect m >= low_index(arr1) ...
stats[m-low_index(arr1)]++;                 // --> potentially undefined

This call site code has a typo, and accidentally mixes the low and high indexes of unrelated containers, which can violate the precondition and result in an index that is less than the “low” value. Then in the next line it tries to use it as an offset index into an instrumentation statistics array, which is undefined behavior for a negative number.

GUIDELINE: Remember that an unspecified result is not in itself undefined behavior, but a call site can run with it and end up with real undefined behavior later. This happen particularly when the calculated value is a pointer, or an integer used as an array index (which, remember, is basically the same thing; a pointer value is just an index into all available memory viewed as an array). If a program relies on unspecified behavior to avoid performing undefined behavior, then it has a path to undefined behavior, and so unspecified behavior is a Crouching Tiger, if you will… still dangerous, and can be turned into to the full dragon.

GUIDELINE: Don’t specify your function’s behavior (output postconditions) for invalid inputs (precondition violations), except for defense in depth (see Example 2(c)). By definition, if a function’s preconditions are violated, then the results are not specified. If you specify the outputs for precondition violations, then (a) callers will depend on the outputs, and (b) those “preconditions” aren’t really preconditions at all.

While we’re at it, here’s a second example: Let’s compare pointers in a way the C++ standard says is unspecified. This program attempts to use pointer comparisons to see whether a pointer points into the contiguous data stored in a vector, but this technique doesn’t work because today’s C++ standard only specifies the results of raw pointer comparison when the pointers point at (into, or one-past-the-end of) the same allocation, and so when ptr is not pointing into v’s buffer it’s unspecified whether either pointer comparison in this test evaluates to false:

// Example 2(b)(ii): If assert is violated, might lead to undefined behavior

// std::vector<int> v = ...;
assert(&v[0] <= ptr && ptr < (&v[0])+v.size());           // unspecified
*ptr = 42;                                  // --> potentially undefined

(c) is never undefined or unspecified behavior

An assertion violation is never undefined behavior if the function specifies what happens in every case even when the assertion is violated. Here’s an example mentioned in my paper P2064, distilled from real-world code:

// Example 2(c): If assert is violated, never undefined behavior
//               (function documents its result when x!=0)

some_result_value DoSomething( int x ) {
    assert( x != 0 );
    if    ( x == 0 ) { return error_value; }
    return sensible_result(x);

The function asserts that the parameter is not zero, to express that the call site shouldn’t do that, in a way the call site can check and test… but then it also immediately turns around and checks for the errant value and takes a well-defined fallback path anyway even if it does happen. Why? This is an example of “defense in depth,” and can be a useful technique for writing robust software. This means that even though the assertion may be violated, we are always still in a well-defined state and so this violation does not lead to undefined behavior.

GUIDELINE: Remember that violating an assertion does not necessarily lead to undefined behavior.

GUIDELINE: Function authors, always document your function’s requirements on inputs (preconditions). The caller needs to know what inputs are and aren’t valid. The requirements that are reasonably checkable should be written as code so that the caller can perform the checks when testing their code.

GUIDELINE: Always satisfy the requirements of a function you call. Otherwise, you are feeding “garbage in,” and the best you can hope for is “garbage out.” Make sure your code’s tests includes verifying all the reasonably checkable preconditions of functions that it calls.

Writing the above pattern has two problems: First, it repeats the condition, which invites copy/paste errors. Second, it makes life harder for static analysis tools, which often trust assertions to be true in order to reduce false positive results, but then will think the fallback path is unreachable and so won’t properly analyze that path. So it’s better to use a helper to express the “either assert this or check it and do a fallback operation” in one shot, which always avoids repeating the condition, and could in principle help static analysis tools that are aware of this macro (yes, it would be nicer to do it without resorting to a macro, but it’s annoyingly difficult to write the early return without a macro, because a return statement inside a lambda doesn’t mean the same thing):

// Using a helper that asserts the condition or performs the fallback

    bool b = B;                         \
    assert(b);                          \
    if(!b) ACTION;                      \

some_result_value DoSomething( int x ) {
    ASSERT_OR_FALLBACK( x != 0, return error_value; );
    return sensible_result(x);

3. Explain how your answers to Questions 1 and 2 do, or do not, correspond with each other.

In Example 2(a), violating the assertion leads to undefined behavior, 1(a).

In Example 2(b), violating the assertion leads to unspecified behavior, 1(b). At buggy call sites, this could subsequently lead to undefined behavior.

In Example 2(c), violating the assertion leads to implementation-defined behavior, 1(c), which never in itself leads to  undefined behavior.

4. BONUS: Describe a valuable service that a tool could perform for assertions that satisfy the requirement in 2(a), that is not possible for other assertions.

There are many. Here is just one example, that happens to be nice because it is perfectly accurate.

Let’s say we have all the code examples in question 2, written using C assert today (or even with those assertions missing!), and then at some future time we get a version of standard C++ that can express them as preconditions. Then only in Example 2(a), where we can see that the function body (and possibly transitively its further callees with the help of inlining) exercises undefined behavior, a tool can infer the precondition annotation and add it mechanically, and get the benefit of diagnosing existing bugs at call sites:

// What a precondition-aware tool could generate for Example 2(a)

auto f( int* p ) 
    [[pre( p )]]  // can add this automatically: because a violation
                  // leads to undefined behavior, this precondition
                  // is guaranteed to never cause a false positive
    assert( p );
    *p = 42;

For example, after some future C++2x ships with contracts, a vendor could write an automated tool that goes through every open source C++ project on GitHub and mechanically generates a pull request to insert preconditions for functions like Example 2(a) – but not (b) or (c) – whether or not the assertion already exists, just by noticing the undefined behavior. And it can inject those contract preconditions with complete confidence that none of them will ever cause a false positive, that they will purely expose existing bugs at call sites when that call site is built with contract checking enabled. I would expect such tool to identify a good number of (at least latent if not actual) bugs, and be a boon for C++ users, and it’s possible only for functions in the category of 2(a).

“Automated adoption” of at least part of a new C++ feature, combined with “automatically identifies existing bugs” in today’s code, is a pretty good value proposition.


Thank you to the following for their comments on this material: Joshua Berne, Gabriel Dos Reis, Gábor Horváth, Andrzej Krzemieński, Ville Voutilainen.


[1] In the standard, there are two flavors of undefined behavior. The basic “undefined behavior” is allowed to enter your program only once you actually try to execute the undefined part. But some code is so extremely ill-formed (with magical names like “IF-NDR”) that its very existence in the program makes the entire program invalid, whether you try to execute it or not.