if consteval
Document #: | P1938R1 |
Date: | 2020-02-17 |
Project: | Programming Language C++ |
Audience: |
EWG |
Reply-to: |
Barry Revzin <barry.revzin@gmail.com> Richard Smith <richard@metafoo.co.uk> Andrew Sutton <asutton@lock3software.com> Daveed Vandevoorde <daveed@edg.com> |
R0 [P1938R0] of this paper initially contained only a positive form: if consteval
. This paper additionally adds a negated form, if not consteval
.
Despite this paper missing both our respective NB comment deadlines and the mailing deadline, we still believe it provides a significant enough improvement to the status quo that it should be considered for C++20.
C++20 will have several new features to aid programmers in writing code during constant evaluation. Two of these are std::is_constant_evaluated()
[P0595R2] and consteval
[P1073R3], both adopted in San Diego 2018. consteval
is for functions that can only be invoked during constant evaluation. is_constant_evaluated()
is a magic library function to check if the current evaluation is constant evaluation to provide, for instance, a valid implementation of an algorithm for constant evaluation time and a better implementation for runtime.
However, despite being adopted at the same meeting, these features interact poorly with each other and have other issues that make them ripe for confusion.
There are two problems this paper wishes to address.
constexpr
and consteval
The first problem is the interplay between this magic library function and the new consteval
. Consider the example:
consteval int f(int i) { return i; }
constexpr int g(int i) {
if (std::is_constant_evaluated()) {
return f(i) + 1; // <==
} else {
return 42;
}
}
consteval int h(int i) {
return f(i) + 1;
}
The function h
here is basically a lifted, constant-evaluation-only version of the function g
. At constant evaluation time, they do the same thing, except that during runtime, you cannot call h
, and g
has this extra path. Maybe this code started with just h
and someone decided a runtime version would also be useful and turned it into g
.
Unfortunately, h
is well-formed while g
is ill-formed. You cannot make that call to f
(that is ominously marked with an arrow) in that location. Even though that call will only happen during constant evaluation, that’s still not enough.
With specific terms, the call to f()
inside of g()
is an immediate invocation and needs to be a constant expression and it is not. Whereas the call to f()
inside of h()
is not considered an immediate invocation because it is in an immediate function context (i.e. it’s invoked from another immediate function), so it has a weaker set of restrictions that it needs to follow.
In other words, this kind of construction of conditionally invoking a consteval
function from a constexpr
function just Does Not Work (modulo the really trivial cases - one could call f(42)
for instance, just never f(i)
).
We find this lack of composability of features to be problematic and think it can be improved.
if constexpr (std::is_constant_evaluated())
problemThe second problem is specific to is_constant_evaluated
. Once you learn what this magic function is for, the obvious usage of it is:
constexpr size_t strlen(char const* s) {
if constexpr (std::is_constant_evaluated()) {
for (const char *p = s; ; ++p) {
if (*p == '\0') {
return static_cast<std::size_t>(p - s);
}
}
} else {
__asm__("SSE 4.2 insanity");
}
}
This example, inspired by [P1045R0], has a bug: it uses if constexpr
to check the conditional is_constant_evaluated()
rather than a simple if
. You have to really deeply understand a lot about how constant evaluation works in C++ to understand that this is in fact not only not “obviously correct” but is in fact “obviously incorrect,” for some definition of obvious. This is such a likely source of error that Barry submitted bugs to both gcc and clang to encourage the compilers to warn on such improper usage. gcc 10.1 will provide a warning for the simple case:
<source>: In function 'constexpr int f(int)':
<source>:4:45: warning: 'std::is_constant_evaluated' always evaluates to true in 'if constexpr' [-Wtautological-compare]
4 | if constexpr (std::is_constant_evaluated()) {
| ~~~~~~~~~~~~~~~~~~~~~~~~~~^~
But then people have to understand why this is a warning, and what this even means. Nevertheless, a compiler warning is substantially better than silently wrong code, but it is problematic to have an API in which many users are drawn to a usage that is tautologically incorrect.
When R0 of this paper was presented in Belfast, the implementers assured that all the compilers would properly warn on all tautological uses of std::is_constant_evaluated()
- both in the always-true
and always-false
cases.
As of this writing, for instance, EDG warns on all of the following:
constexpr int f1() {
if constexpr (!std::is_constant_evaluated() && sizeof(int) == 4) { // warning: always true
return 0;
}
if (std::is_constant_evaluated()) {
return 42;
} else {
if constexpr (std::is_constant_evaluated()) { // warning: always true
return 0;
}
}
return 7;
}
consteval int f2() {
if (std::is_constant_evaluated() && f1()) { // warning: always true
return 42;
}
return 7;
}
int f3() {
if (std::is_constant_evaluated() && f1()) { // warning: always false
return 42;
}
return 7;
}
We expect the other compilers to follow suit.
We propose a new form of if
statement which is spelled:
The braces (in both the if
and the optional else
) are mandatory and there is no condition. If evaluation of this statement occurs during constant evaluation, the first substatement is executed. Otherwise, the second substatement (if there is one) is executed.
This behaves exactly as today’s:
except with three differences:
if consteval
to allow invoking immediate functions.To explain the last point a bit more, the current language rules allow you to invoke a consteval
function from inside of another consteval
function ([expr.const]/12) - we can do this by construction:
An expression or conversion is in an immediate function context if it is potentially evaluated and its innermost non-block scope is a function parameter scope of an immediate function. An expression or conversion is an immediate invocation if it is an explicit or implicit invocation of an immediate function and is not in an immediate function context. An immediate invocation shall be a constant expression.
By extending the term immediate function context to also include an if consteval
block, we can allow the second example to work:
consteval int f(int i) { return i; }
constexpr int g(int i) {
if consteval {
return f(i) + 1; // ok: immediate function context
} else {
return 42;
}
}
consteval int h(int i) {
return f(i) + 1; // ok: immediate function context
}
Additionally, such a feature would allow for an easy implementation of the original std::is_constant_evaluated()
:
Although this paper does not suggest removing the library function.
Many people have expressed the view that a negated form is also useful. That form is also proposed here, spelled:
or
With the semantics that the first substatement is executed if the context is not manifestly constant evaluated, otherwise the second substatement (if any) is executed.
As proposed, this new form of if
does not have a condition - unlike the other two we already have. While there are certainly cases where an added condition would be useful, this paper is deliberately not including such a thing. The vast majority of uses are expected to be just of the if consteval
or if not consteval
form and we do not want to clutter future design space in this area.
There are currently two uses in libstdc++ that are of the form if (is_constant_evaluated() && cond)
. One example:
if (std::is_constant_evaluated() && __n < 0)
throw "attempt to decrement a non-bidirectional iterator";
This usage is perfectly fine and doesn’t necessary need special support from this proposal. Or it could also be written as:
Or factored into a function like:
Either way, the condition form doesn’t feel strongly motivated except for consistency with if
and if constexpr
.
std::is_constant_evaluated()
One of the questions that comes up regularly in discussing this paper is: if we had if consteval
, we do we even need std::is_constant_evaluated()
, and can we just deprecate it?
This paper proposes no such deprecation. The reason is that this function is actually still occasionally useful (as in the previous section). If the standard library does not provide it, users will write their own. We’re not concerned about the implementation difficulty of it - the users that need this will definitely be able to write it correctly - but we are concerned with a proliferation of exactly this function. The advantage of having the one std::is_constant_evaluated()
is both that it becomes actually teachable and also that it becomes warnable: the warnings discussed can happen only because we know what this name means. Maybe it’s still possible to warn on if constexpr (your::is_constant_evaluated())
but that’s a much harder problem.
And note that libstdc++ already has some uses that do require the function form.
Here are a few examples from libstdc++. Today, they’re implemented uses a builtin function, and how they would look with if consteval
. It’s not a big difference, just spelling.
From libstdc++
|
Proposed
|
---|---|
As of this writing, libstdc++ has 23 uses that could be replaced by if consteval
, 2 that could be replaced by if not consteval
, and 2 that require an extra condition on the comparison.
The initial revision of the std::is_constant_evaluated()
proposal [P0595R0] was actually targeted as a language feature rather than a library feature. The original spelling was if (constexpr())
. The paper was presented in Kona 2017 and was received very favorably in the form it was presented (17-4). The poll to consider a magic library alternative was only marginally more preferred (17-3). We believe that in the two years since these polls were taken, having a dedicated language feature with an impossible-to-misuse API, that can coexist with the rest of the constant ecosystem, is the right direction.
Extend the definition of immediate function context in 7.7 [expr.const] (and use bullet points):
An expression or conversion is in an immediate function context if it is potentially evaluated and either:
- (12.1) its innermost non-block scope is a function parameter scope of an immediate function
., or- (12.2) it appears in the first compound-statement of a consteval if statement ([stmt.if]) of the form
if consteval
or the second compound-statement (if any) of a consteval if statement of the formif ! consteval
.
Change 8.5 [stmt.select] to add the new grammar:
selection-statement: if constexpropt ( init-statementopt condition ) statement if constexpropt ( init-statementopt condition ) statement else statement + if !opt consteval compound-statement + if !opt consteval compound-statement else compound-statement switch ( init-statementopt condition ) statement
Add a new clause to 8.5.1 [stmt.if]
a An
if
statement is of the formif consteval
orif ! consteval
is called a consteval if statement.b If the
if
statement is of the formif consteval
and evaluation occurs in a context that is manifestly constant-evaluated ([expr.const]), the first substatement is executed and is an immediate function context ([expr.const]). Otherwise, if theelse
part of the selection statement is present, then the second substatement is executed. Acase
ordefault
label appearing within such anif
statement shall be associated with aswitch
statement within the sameif
statement. A label declared in a substatement of an consteval if statement shall only be referred to by a statement in the same substatement.c A consteval if statement of the form
if ! consteval compound-statement
is equivalent toA consteval if statement of the form
if ! consteval compound-statement1 else compound-statement2
is equivalant to
Change 20.15.10 [meta.const.eval] to use this new functionality:
1 Returns:
true
if and only if evaluation of the call occurs within the evaluation of an expression or conversion that is manifestly constant-evaluated ([expr.const]).1 Effects: Equivalent to:
Add the macro __cpp_if_consteval
.
Thank you to David Stone and Tim Song for working through these examples.
[P0595R0] Daveed Vandevoorde. 2017. The “constexpr” Operator.
https://wg21.link/p0595r0
[P0595R2] Richard Smith, Andrew Sutton, Daveed Vandevoorde. 2018. std::is_constant_evaluated
.
https://wg21.link/p0595r2
[P1045R0] David Stone. 2018. constexpr Function Parameters.
https://wg21.link/p1045r0
[P1073R3] Richard Smith, Andrew Sutton, Daveed Vandevoorde. 2018. Immediate functions.
https://wg21.link/p1073r3
[P1938R0] Barry Revzin, Daveed Vandevoorde, Richard Smith. 2019. if consteval.
https://wg21.link/p1938r0