Yes. For example in Rust it took a long time for floating point operations to be available at compile time and even now only a subset is. The reason is that a lot of energy and thought went into the issue of producing identical output (and what identical precisely means ) even when compilation is on a different processor than where the target runs.
As far as I know this is not a concern for Zig comptime.
Zig Comptime uses softfloat so it has architecture independent determinism.
And the zig core team is absolutely concerned with bitwise determinism in the compiled artifacts, iirc this is why they rejected the sloppy bun PR to the compiler.
Determinism and host/target agreement are two different properties.
The Zig core team is apparently not concerned enough about the second point to forbid transcendentals at comptime and this is something that'd be hard to take back, because I'd break existing code.
Generally speaking, as a user of a language, stuff should just work while providing identical results to the target arch.
Which means that when cross compiling from x86 to ARM, if lets say, transcendentals provide different results then always the target archs implementations should be used, even if they have to be emulated.
If however, hypothetically, different x86 CPUs produce different results for transcendentals, and we can't control where the user will run our program, then imo the correct solution for the language is to provide a set of knobs for the user to communicate whether they favor accuracy or speed in this scenario. - but we can say there's no 'correct' decision in this case, only tradeoffs.
Forbidding transcendentals is not a correct decision as it adds a ton of compiler complexity (you have to track which functions use them transitively), and baffling UX - the user finds that comptime doesn't work because the function he uses might use transcendentals somewhere down the chain.
there has been a ton of work on making comptime a pure and deterministic execution environment. I do not expect that work to stop. I don't know where you are getting the idea that it is not an important design consideration for the language.
to expand, if two different host platforms cross compiling to the same target platform have different results, I am almost certain that would be considered a compiler bug.
if you're pointing out that a runtime operation and a compile time operation might not agree, I'd be more interested in understanding when that would ever have any meaningful impact on anything. given the compilation is supposed to be deterministic, the difference can easily be addressed by comptime branching on target architecture in the rare case that it matters for your program.
Determinism and host/target agreement are two different properties. I meant the second one.
It's something Rust guarantees (without me having to take care of it e.g. by manually branching) and Zig does not.
sure, it sounds like there is a difference here. I'm not aware of any stance by the Zig core team on the subject.
I am genuinely interested in a place that this matters for a program, or any practical consequence this has for an end user of the language.
no idea why your response was flagged originally.
Floating point hardware on different CPU architectures don’t produce bitwise identical results, unfortunately. That’s the big issue.
As far as I’m aware, they do produce semantically identical results, but something like the specific bit pattern of a NaN value can theoretically vary, and people might do fun things like encoding extra information in those bits.
About NaNs, you are right, because the standard does not mandate the use of a specific NaN for each kind of invalid operation.
Nonetheless, if you use a comparison function for which all NaNs are equivalent, different CPU architectures that are standard-compliant must produce bitwise identical results for the same sequence of operations.
Differences appear mainly when the compilers generate different operations or in a different order. Moreover, transcendental functions are computed using various approximations by the standard library, so if you use different libraries on different computers, you will get different results.
However, these are differences caused by software, not by hardware, and they happen even on the same computer when you use different versions of a compiler or of a standard library. Therefore such differences can be eliminated, if desired.
rsqrt differs between intel and amd, out-of-bounds conversions differ between x64 and arm (saturating vs sentinel), denormal handling also varies iirc (not 100% sure about that), x64 has FTZ/DAZ separately, ARM has a combined FZ flag and this list is probably not exhaustive...
The practical value is that moving a computation from runtime to compile time can then be purely an optimization, rather than potentially changing its semantics. Think generated lookup tables or numeric constants. Ideally f(x) means the same thing whether evaluated by the compiler or by the generated program.
> Ideally f(x) means the same thing whether evaluated by the compiler or by the generated program.
Absolutely, that’s why I wanted to know what types of things aren’t covered yet, and that’s also why I don’t share your certainty that they won’t be resolved later.
You'd have to read the discussions the Rust community has about this to get a picture. It's complicated.
From what I understand, there is hope for sqrt but there is no consensus on transcendentals.
I think there's some misconceptions floating around here regarding both Rust and Zig's const-evaluation philosophies.
Rust is concerned about memory-safety, yes, but the only strict requirement for memory-safety when it comes to const-evaluation is as follows: "The only guarantee the type system needs is that evaluating `some_crate::SOME_CONST` will produce consistent results if evaluation is repeated in different compilation units" ( https://rust-lang.github.io/rfcs/3514-float-semantics.html ).
Beyond that, from a philosophical standpoint, Rust takes great pains to ensure that const functions produce identical results regardless of whether or not those functions are called at compile-time or at runtime. Rust has adopted this stance because it wants to reserve the right to opportunistically evaluate const-capable functions at compile time, as a performance optimization, even if the user has not explicitly asked for it (for that matter, Rust also does its best to const-evaluate non-const functions when it can). Because of this, Rust's assumption is that users would be annoyed if their program's visible behavior depends on whether or not the optimizer has exercised its discretion to evaluate a specific function at compile-time.
However, this is only a guideline, not a strict guarantee. There is one exception to the above rule: "when a floating-point operation produces a NaN result, the resulting NaN bit pattern is some deterministic function of the operation’s inputs that satisfies the constraints placed on run-time floating point semantics. However, the exact function is not specified, and it is allowed to change across targets and Rust versions, and even with compiler flags. In particular, there is no guarantee that the choice made in const evaluation is consistent with the choice made at runtime."
In other words, calling the `.to_bits()` function on a floating-point value that happens to be NaN is allowed to produce a different result at runtime than it does at compile-time (note that all compile-time evaluations are guaranteed to always produce the same result for a given toolchain version for a given target, as required above).
This exception is made because otherwise otherwise it would be basically impossible to support floating-point math at all, thanks to the way various platforms have implemented their floating-point functions in practice.
In contrast, Zig doesn't have such a philosophical compunction against a function's result being determined by whether or not it's being evaluated at compile-time, as shown by the existence of the `@inComptime` builtin. But Zig does still broadly attempt to make comptime deterministic, including going so far as to forbid I/O, though I don't see where any specific guarantees are documented in the Zig reference.
In C++ as well, constexpr math introduced in C++23 has similar concerns regarding floating point accuracy.
What's your source for this? Comptime Zig code can do pointer casts and whatnot, all emulated as if run on the target bitness/endianness/etc. And any operations that are undefined on the target platform result in a compile error. I've never had an issue cross-compiling.
Endianness and pointer casts are the mechanical part and I would expect Zig to emulate them correctly.
Other parts, like floats are harder. This is where a difference shows. Rust is like: "Sorry, since we cannot uphold our guarantees, no transcendentals for you at comptime ", whereas Zig is chill about that and let you have your transcendentals even if results may differ between comptime and runtime. Different mindsets.
The guarantees could be provided, if there would be a way to ensure that the compiler uses the same standard math library that is used by the executable program that is created.
This could be done, for instance, if the standard math library would be dynamically linked into the compiler, so the same library would be available for the compiled program.
If you disregard performance, yes. Unfortunately forcing softfloat everywhere would result in abysmal performance, so this is not an option.
Just to note, Rust has had compile-time floats for a while now, but yes.
Yes, but no transcendentals on floats. You can add them and the like but you cannot calculate the sine.
Now I see this response, I responded to the other comment.