> and divide by zero is explicitly named as undefined behavior in 5.6.4
I thought, how can this possibly be true if IEEE-754 specifically defines 1.0/-0.0 == -inf.
So I looked it up, and the actual rule seems to be
> If during the evaluation of an expression, the result is not mathematically defined or not in the range of representable values for its type, the behavior is undefined.
So divide-by-zero is only undefined behaviour if the result is not representable. But it is representable in IEEE-754 (adherence to which is implementation-defined), so it should be fine.
That said, I still don't understand how it's not a constant expression. Perhaps it's because in IEEE-754 it can sometimes cause a divide-by-zero exception, depending on flag settings? Because to check something is negative zero it's enough to do something like
So divide-by-zero is only undefined behaviour if the result is not representable. But it is representable in IEEE-754 (adherence to which is implementation-defined), so it should be fine.
That's not how I read the line of the standard you quoted. ISTM it's saying:
(Result not mathematically defined) ∨ (Result not in the range of representable values) ⇒ Behavior is undefined
In this case, 1.0/-0.0 is mathematically undefined as C++ specifies such things (I believe division-by-zero is called out specifically as such), and it doesn't help that IEEE-754 can represent it anyway. It's still a useful representation in the language because you can reach it through a series of calculations that are all included in the range of defined behavior.
I'm also skeptical that
x == 0 && 1/x == -inf
will work reliably for just that reason: I believe a compiler would be perfectly within its rights to unconditionally treat this expression as false, since x == 0 would imply UB.
Not at all an expert here, so would enjoy having a standards lawyer set me straight.
Perhaps you're right. I just find IEEE-754's approach clear, and C++'s rule confusing.
I would say that 1.0/-0.0 is mathematically defined in the pure sense: IEEE-754 defines a specific model of real numbers closed under such operations [-inf,inf]\union{NaN}.
The thing is that that test should definitely work under IEEE-754, so I would be very surprised if C++ somehow forbids you from getting the IEEE-standard behaviour. For reference, here's a proof that it's correct:
$ stack repl
λ> import Data.SBV
λ> prove $ do x <- sFloat "x"; return $ fpIsNegativeZero x <=> (x .== 0 &&& 1/x .== -infinity)
Q.E.D.
Undefined behavior means anything can happen, and "anything" includes "following IEEE-754 rules." Any given C++ implementation is free to implement this case to follow IEEE-754 and declare that this behavior is supported.
I see, that makes sense given how the C++ standard omits specifying floating-point arithmetic. But isn't there still a problem, given that in the article clang gave the error message
error: '(1.0e+0f / -0.0f)' is not a constant expression
and the proposed explanation is that "undefined behavior is unusable in a constant expression"? So it seems clang is not following IEEE-754 rules.
As I read 5.20.2.5 in N4296, undefined behaviour may not occur in constant expressions, so clang is not free to follow IEEE-754 for the division-by-zero, even though it can do that in regular code, because then it's UB. And anyway, why isn't it UB to invoke UB in a constant expression?
I think you're right, the compiler can follow IEEE-754 rules in general, but as written it couldn't do so for constant expressions.
I find that rule to be really weird in general. Does that mean that, for example, if you try to use "1 << 32" in a constant expression, then whether it's actually a constant expression will depend on the target architecture? I wonder what the rationale is there.
I find that rule to be really weird in general. Does that mean that, for example, if you try to use "1 << 32" in a constant expression, then whether it's actually a constant expression will depend on the target architecture? I wonder what the rationale is there.
Yes, that's right. The rationale is that it's not particularly different from the way constant expressions in C have always worked: The compiler knows full well whether
static const int x = 99999999999;
or
static const int x = 1 << 32;
represents a valid assignment within the range of int on the target platform, and responds accordingly. constexpr just generalizes that behavior to more of the expression machinery.
Right, but those examples still compile, they just invoke UB. Normally you want to avoid that, but sometimes you actually do want to, and the fact that the result is UB gives the compiler the freedom to support additional semantics.
For something closer to the original topic, consider:
static const double x = 1.0/0.0;
That's UB, but the compiler is free to declare that it follows IEEE-754 and fully support this. The equivalent with constexpr doesn't allow that same freedom, and the compiler must consider 1.0/0.0 to not be a constexpr.
Ah, I see what you mean. A few years of warnings-as-errors mode made me forget that the most bizarre constant expressions (not constexprs) will compile even if their results or what you do with them are complete nonsense.
I'd argue that the constexpr behavior is an improvement, in that it enables errors for large swaths of obviously buggy expressions without breaking backward compatibility. If the committee wants to remove float-division-by-0 from the set of UB expressions, they should do that independently of the rules for constexprs/constant expressions (as some Googling gives me the impression the C standard did with Annex F).
Generally, I think compiler vendors should not be distinguishing themselves by "reliably" implementing some "defined" behavior for what the standard declares to be UB[1]. At that point you're effectively implementing a proprietary language extension. "Implementation-defined" parts of the standard—and tooling/libraries, of course—are where vendors should be differentiating.
[1] I had a very traumatic experience with HP-UX and its configurable null-pointer-dereference semantics that may be biasing me here.
> I would say that 1.0/-0.0 is mathematically defined in the pure sense
`1.0 / -0.0` isn't mathematically defined though, it is certainly undefined. IEEE-754 takes that and performs "limit of `1/x` as `x` approaches zero from the left".
> IEEE-754 defines a specific model of real numbers closed under such operations [-inf,inf]\union{NaN}.
No, IEEE-754 defines a way to handle the exception by convention.
With clang 3.7.0 this results in "constexpr function never produces a constant expression", since the C-style pointer cast is essentially a reinterpret_cast (C++14 Standard §5.4p4.4), which is not a "core constant expression" (C++14 Standard §5.20p2.13) and thus must not appear in a constexpr function: C++14 Standard §7.1.5p5:
... if no argument values exist such that an invocation of the function or constructor could be an evaluated subexpression of a core constant expression, the program is ill-formed; no diagnostic required.
clang seems to go the extra mile and produces a diagnostic.
Incidentally the pointer cast would violate C++'s aliasing rules and thus invoke undefined behavior.
Idiomatic C++ is not well defined. There are so many possible ways to write C++ code that you should be skeptical if somebody promotes one version is the true one.
I would expect one can export a constexpr function from a library (that seems necessary to me if one wants to write (most of) the C++ standard library in C++).
If so, a workaround would be to hack together a library that exports a constexpr function that converts floats to their bit representation (that might require assembly or binary editing of a library containing a function that does that by casting) and calling it from the C++ constexpr. Such a workaround would support negative zero and NaNs.
Love this kind of hacking at the edges of the standard.
Which of the attempted schemes are legal (implementation-defined) C++ when used at runtime? All of them? I'm pretty sure the memcpy() approach has to be kosher, at least.
EDIT:
Don't have a copy of the standard handy, but cppreference[1] suggests using reinterpret_cast<> is UB.
If signbit() were compatible with constexpr, it would work for negative zero. Similarly, frexp() could determine the exponent, though the interface itself is a problem.
I'm glad you're aware of Carmack's association with the fast inverse square root (hopefully you also know that it doesn't originate from him) but the joke here isn't the denigration of the code itself.
The joke is that the original need for a fast inv sqrt is moot at compile-time.
> and divide by zero is explicitly named as undefined behavior in 5.6.4
I thought, how can this possibly be true if IEEE-754 specifically defines 1.0/-0.0 == -inf.
So I looked it up, and the actual rule seems to be
> If during the evaluation of an expression, the result is not mathematically defined or not in the range of representable values for its type, the behavior is undefined.
So divide-by-zero is only undefined behaviour if the result is not representable. But it is representable in IEEE-754 (adherence to which is implementation-defined), so it should be fine.
That said, I still don't understand how it's not a constant expression. Perhaps it's because in IEEE-754 it can sometimes cause a divide-by-zero exception, depending on flag settings? Because to check something is negative zero it's enough to do something like