ACP: add `scalbn` and `ilogb` float methods

[folkertdev]

Proposal

Problem statement

The implementation of complex division needs the scalbn and ilogb functions. We'll likely add these as intrinsics, and I wonder whether we want them as methods too.

Motivating examples or use cases

Complex division pre-processes the arguments like so:

https://github.com/rust-lang/compiler-builtins/blob/082387e5bd9fab846b307686ea7a9a23a74d9f94/compiler-builtins/src/float/complex/div.rs#L14-L20

let max = fmax(c.abs(), d.abs());
let mut ilogbw = 0;
if max.is_finite() && max != F::ZERO {
    ilogbw = ilogb(max);
    c = scalbn(c, -ilogbw);
    d = scalbn(d, -ilogbw);
}

The goal is to make overflow/underflow less likely to occur in subsequent operations.

Solution sketch

impl f* {
    /// Multiplies a floating-point value by 2 raised to power `exp`.
    /// A `scalbn` is always exact, and much cheaper than `self * 2.0powi(n)`.
    const fn scalbn(self, n: i32) { /* ... */ }

    /// Extracts the value of the unbiased exponent from the floating-point 
    /// argument num, and returns it as a signed integer value.
    /// 
    /// This is a cheap way of determining the (order of) magnitude of a float.
    const fn ilogb(self) -> i32 { /* ... */ }
}

For code size reasons we can usually use libcalls for the implementation. compiler-builtins already has these implementations, and Miri/const eval (via rustc_apfloat) already has support.

Alternatives

We can keep these functions just as intrinsics and not expose them.

Apparently it's been suggested before to make the << and >> operators on floats call scalbn. On integers shifts correspond to multiplication and division by powers of 2. I'm not a fan of using infix operators for non-obvious operations but it's an idea.

The names are not the best but they are standard.

Having frexp and ldexp might be useful too.

Links and related work

Tangentially related

• IRLO: Add conversions of floating point to / from exponent+mantissa

What happens now?

This issue contains an API change proposal (or ACP) and is part of the libs-api team feature lifecycle. Once this issue is filed, the libs-api team will review open proposals as capability becomes available. Current response times do not have a clear estimate, but may be up to several months.

Possible responses

The libs team may respond in various different ways. First, the team will consider the problem (this doesn't require any concrete solution or alternatives to have been proposed):

• We think this problem seems worth solving, and the standard library might be the right place to solve it.
• We think that this probably doesn't belong in the standard library.

Second, if there's a concrete solution:

• We think this specific solution looks roughly right, approved, you or someone else should implement this. (Further review will still happen on the subsequent implementation PR.)
• We're not sure this is the right solution, and the alternatives or other materials don't give us enough information to be sure about that. Here are some questions we have that aren't answered, or rough ideas about alternatives we'd want to see discussed.

[clarfonthey]

Honestly, I think that we can safely say that if C has a dedicated mathematical operation that LLVM has a dedicated intrinsic for that is not redundant with existing functionality, we probably should add it as a dedicated method. There are definitely qualifiers to that, like how floating-point methods often are going to be std-only due to dependencies on libm, but I think having the method somewhere is a good rule of thumb to have.

I second this.

[folkertdev]

that LLVM has a dedicated intrinsic for

To clarify, LLVM doesn't have an intrinsic for this operation. It defines it in compiler-rt and we have a version in our rust compiler-builtins.

But otherwise I fully agree.

[pitaj]

Can we give these friendlier, more meaningful names? With doc aliases for the C names?

[tgross35]

I'm in favor of having these operations but think we can improve on the C operations:

1. ilogb returns implementation-defined FP_ILOGB0 and FP_ILOGBNAN on 0 and NaN (our libm makes these i32::MIN). Could we use Option instead?
2. We should consider using <</>> for scalbn, since the intuition is similar as for integers. I believe @joshtriplett proposed this before.
3. As @pitaj mentioned, the C names can be a bit confusing. Perhaps ilog2 for ilogb and, if we keep a function rather than using operators, something like scale, scale_exponent, or pow2?

Having frexp and ldexp might be useful too.

Note that ldexp is the same as scalbn when using binary floating point (as opposed to base-16 or whatever)

[clarfonthey]

that LLVM has a dedicated intrinsic for

To clarify, LLVM doesn't have an intrinsic for this operation. It defines it in compiler-rt and we have a version in our rust compiler-builtins.

But otherwise I fully agree.

Ah, these both felt like low-enough-level operations that I had just assumed. Regardless, they are simple mathematical operations that feel like useful primitives to implement in the standard library.

[folkertdev]

Notably ldexp does have an LLVM intrinsic https://llvm.org/docs/LangRef.html#llvm-ldexp-intrinsic, as does frexp https://llvm.org/docs/LangRef.html#llvm-frexp-intrinsic.

[tgross35]

Was curious why, looks like there's a special lowering on AMDGPU

[clarfonthey]

Note that for example rust-lang/rust#161081 uses min/max intrinsics not because they necessarily lower to separate machine code than their naïve implementations, but because they are easier to reason about for optimisation purposes. So, I would sometimes expect float ops like these to potentially have intrinsics for that reason too, although I guess they're just not used nearly as often for that to matter.

[kennytm]

3. Perhaps ilog2 for ilogb

One slight problem is that ilogb's behavior for negative number is different from their integer counterpart. (-9_i32).ilog2() panics, but ilogb(-9.0) in C returns 3.

(Also i32::ilog2() returns an unsigned u32 but f64::ilog2() have to return a signed i32.)

So technically it would be ilog2_abs, but this name doesn't spark joy.

We could align the behavior with i32::ilog2() and also panic or return None with (-9.0).ilog2(), and requires (-9.0).abs().ilog2() for the equivalent for ilogb, letting the optimizer do their magic.

Additionally, by C's definition ilogb(INFINITY) == INT_MAX. It is mathematically wrong, but it's practically correct since 2.0^i32::MAX does round to f64::INFINITY. Though I'd argue it's also true that 2.0^i32::MIN rounds to 0.0 too, so perhaps (0.0).ilog2() == i32::MIN would be better than panicking or returning None.

[tgross35]

I'm going to nominate this: I think we're in favor of having the API but before implementation work begins, it would be good to have answers about the semantics and functions vs. shl/shr.

@rustbot label +I-libs-nominated