Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add signed(T), unsigned(T). #30445

Closed
wants to merge 12 commits into from
Closed

Conversation

NHDaly
Copy link
Member

@NHDaly NHDaly commented Dec 19, 2018

Converts Types into their corresponding signed/unsigned equivalents, maintaining storage size, but changing signed-ness.

Fixes #30444.

@ghost ghost mentioned this pull request Dec 19, 2018
Converts Types into their corresponding signed/unsigned equivalents,
maintaining storage size, but changing signed-ness.
@NHDaly NHDaly force-pushed the signed_T__unsigned_T branch from 5e7915e to 1a5a012 Compare December 19, 2018 00:07
@NHDaly
Copy link
Member Author

NHDaly commented Dec 19, 2018

(Still needs tests, etc, but just wanted to see what y'all think first! 😄)

Copy link
Contributor

@JeffreySarnoff JeffreySarnoff left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I use these extensions often. For the builtin types, I favor

unsigned(::Type{Int64}) = UInt64
unsigned(::Type{Float32}) = UInt32  # important for low level floating point work
unsigned(x::T) where {T<:Union{Int8..Int128}} = reinterpret(unsigned(T), x)
# etc and cover signed() and e.g. floating(::Type{UInt32}) = Float32 ..

Copy link
Member

@timholy timholy left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I approve of adding this, once it has tests.

@JeffreySarnoff, I don't doubt being able to derive the "plain bits" type corresponding to a floating-point type is useful, but I don't think that operation can be called unsigned. Maybe call that planbitstype or something.

base/int.jl Outdated Show resolved Hide resolved
base/int.jl Outdated Show resolved Hide resolved
@NHDaly
Copy link
Member Author

NHDaly commented Dec 19, 2018

I approve of adding this, once it has tests.

:) Thanks Tim! I've added Tests. This will probably also need a News change before it could be merged.

@JeffreySarnoff, I don't doubt being able to derive the "plain bits" type corresponding to a floating-point type is useful, but I don't think that operation can be called unsigned. Maybe call that plainbitstype or something.

I have to agree with Tim about that. Converting a float to an int seems beyond the scope of unsigned/signed. Also, I think you could already achieve that with something like Int(3.0) / UInt(3.0)?

Also, since I restored the original definitions, I also reordered the
methods back to their original places.
Added unit tests, and also added the missing implementations for calling
`signed` on already signed types and same for `unsigned`.
@NHDaly NHDaly force-pushed the signed_T__unsigned_T branch from 2c6b04f to 6a34de1 Compare December 19, 2018 18:24
@test_throws InexactError signed(UInt(-3))
@test signed(UInt(0) - 1) === -1
@test_throws InexactError signed(UInt(-3)) # Note that UInt() throws, not signed().
end
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I changed all these tests to be === from ==, because they would be equal even if signed() did nothing. Lemme know if you don't like that! :)

Also, I was a bit confused by the @test_throws here. I clarified its behavior with this # Note, but maybe it's better to fix/clarify the test itself? I'm not reallly sure what it's testing though, unfortunately.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't understand that test either. Does git blame shed any light?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks like the test was added here: #11962, during v0.4.0.

Maybe the behavior was different back then, or maybe it was a mistake? (It was David's first PR ever after all! 😋) I'm happy to either remove the whole line or just leave the comment; whatever you prefer!

base/int.jl Outdated Show resolved Hide resolved
@test_throws InexactError signed(UInt(-3))
@test signed(UInt(0) - 1) === -1
@test_throws InexactError signed(UInt(-3)) # Note that UInt() throws, not signed().
end
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't understand that test either. Does git blame shed any light?

@NHDaly
Copy link
Member Author

NHDaly commented Dec 20, 2018

Also: is this news-worthy? I'm ambivalent on that, but here's a draft for what i can put there:

* There are new method overloads `unsigned(T::Type)` and `signed(T::Type)` that accept a `Type`,
  and return its corresponding signed/unsigned equivalent type ([#30445]).

base/int.jl Outdated Show resolved Hide resolved
base/int.jl Outdated Show resolved Hide resolved
base/int.jl Show resolved Hide resolved
base/int.jl Outdated Show resolved Hide resolved
base/int.jl Outdated Show resolved Hide resolved
base/int.jl Show resolved Hide resolved
fredrikekre and others added 2 commits December 20, 2018 17:49
Fixes docstrings for `unsigned(T::Type)` and `signed(T::Type)`.

Co-Authored-By: NHDaly <[email protected]>
@NHDaly
Copy link
Member Author

NHDaly commented Dec 20, 2018

(K, i added a NEWS entry based on your thumbsup)

NHDaly added 2 commits January 2, 2019 10:25
Line 202 had trailing whitespace in the docstring which broke the tests. Fixed.
Now that I understand we're not supposed to generate these every time, i've removed the generated links to prevent conflicts.
@NHDaly NHDaly closed this Jan 24, 2019
@NHDaly NHDaly reopened this Jan 24, 2019
@NHDaly
Copy link
Member Author

NHDaly commented Jan 27, 2019

I reran the tests, and they're passing now. :)

Please take another look when you get a chance! :)

@NHDaly
Copy link
Member Author

NHDaly commented Feb 28, 2019

Friendly ping, this should be good to merge. Please take another look?

@JeffreySarnoff
Copy link
Contributor

I am wondering what benefit there is to avoiding symmetry in these two functional definitions.
To my eye, this is more natural and tends better toward future-proof:

unsigned(::Type{T}) where {T<:Unsigned} = T
unsigned(::Type{T}) where {T<:BitSigned} = typeof(unsigned(zero(T)))

signed(::Type{T}) where {T<:Signed} = T
signed(::Type{T}) where {T<:BitUnsigned} = typeof(signed(zero(T)))

(I don't have v1.2 available to run these .. if the foregoing does not work, there are bigger issues)

@NHDaly
Copy link
Member Author

NHDaly commented Mar 1, 2019

I am wondering what benefit there is to avoiding symmetry in these two functional definitions.

@JeffreySarnoff I completely agree. We discussed it above, here: #30445 (comment) -- the reason is to mirror the existing asymmetry in the definitions of unsigned and signed. As you pointed out there, I guess they're like that due to a difference regarding BigInt...

I opened #30473 to discuss this further. I think we should probably make signed and unsigned symmetric in the way you suggest. Either make them both take Signed and Unsigned, or both take BitSigned and BitUnsigned. I'd actually lean towards the first option i think. But yes, i agree, and i think we should change that for the existing signed/unsigned as well.

But that also seems orthogonal to this PR, which is why i separated it into #30473.

@JeffreySarnoff
Copy link
Contributor

I want to see this resolved as much as you do (modulo keypress as unit of effort). There is nothing in my experience that suggests spackle over a conceptual inelegance is a winning approach -- there are situations wherein it is the only approach available, fortunately this is not so for us.

Rather than orthogonal concerns there is a dependance relationship. Get the underpinning correct and this PR is implemented with clarity, coherence and aplomb. Leaving the underpinning for resolution after the band-aid is applied ... well, who enjoys ripping off a bandage when the alternative is no bandage?

I don't mind that there are split into two PR. It is my considered view that we need to come to an agreement on the other first. I am willing to help do that.

@JeffreySarnoff
Copy link
Contributor

I have a feeling that's not all. @timholy the first version of SaferIntegers was a conversion of your approach writing the pkg for Ints that round themselves -- and I put a premium on minimizing changes. I remember spending days on why you had to use discrepant approaches to the sogned and the unsigned types and wondering how to bring both under a shared umbrella Integer type that would support the expressiveness we like.

@NHDaly
Copy link
Member Author

NHDaly commented Dec 31, 2019

Now that #33459 was merged, I've (finally) updated this PR to the simple version suggested by @JeffreySarnoff above.
This is now fully symmetric, and simple. Thanks for the patience! :)

@JeffBezanson on the review of #33459, you noted that you're "really surprised we still have signed and unsigned. It seems like those should be x % Signed instead."
Is it okay that we move forward then with extending these methods in this way? I suppose instead of overloading like this, we could simply rename these functions to something else, like signedtype and unsignedtype (or signed_type and unsigned_type)?

@NHDaly
Copy link
Member Author

NHDaly commented Dec 31, 2019

Is it okay that we move forward then with extending these methods in this way? I suppose instead of overloading like this, we could simply rename these functions to something else, like signedtype and unsignedtype (or signed_type and unsigned_type)?

Though note that, the whole motivation for this PR was the observation that Base already includes an (undocumented) method for unsigned(::Type), just not one for signed(::Type):
#30444 (comment)

(Which, actually, this makes me wonder why i didn't see a method redefinition warning.. Ah, never mind yeah there is one. Lemme fix that now. EDIT: Fixed.)

So I guess if we rename these, either we'd want to keep that weird undocumented asymmetry, or remove the undocumented method (which is "breaking")?

@JeffreySarnoff
Copy link
Contributor

fwiw I am happier with Unsigned(T) and Signed(T) than unsigned(T) and signed(T) for this role.

@NHDaly
Copy link
Member Author

NHDaly commented Jan 10, 2020

fwiw I am happier with Unsigned(T) and Signed(T) than unsigned(T) and signed(T) for this role.

I can see the appeal of that, but as a counter-argument, Signed and Unsigned read like constructors for types. It seems unexpected to me for the Unsigned constructor to return a value that is not <: Unsigned.

Perhaps the best name for these would simply be the more explicit:

  • unsignedtype(T) and signedtype(T)?

@NHDaly
Copy link
Member Author

NHDaly commented Jan 10, 2020

I think that would be my vote now: unsignedtype(T) and signedtype(T). It doesn't address the question of why we even have those unsigned(x) and signed(x) functions laying around, but oh well such is life. No reason to pile on to that decision.

unsignedtype(T) and signedtype(T) is clear, explicit, and discoverable, i think.

@Keno
Copy link
Member

Keno commented Jun 3, 2020

This needs a rebase, but looks otherwise reasonable to me.

@NHDaly
Copy link
Member Author

NHDaly commented Jun 4, 2020

@Keno do you have a preference regarding renaming these to unsignedtype(T) and signedtype(T)? If i was to rewrite this PR, i would advocate for those names (except that there's already the [undocumented] unsigned(T), described in #30444)

@Keno
Copy link
Member

Keno commented Jun 4, 2020

If it's just about @JeffBezanson's comment regarding the value of version these functions, then I don't think it's a problem, because the type method doesn't quite have the same semantics. It's not really a value conversion, it's a type property query. So I'm fine with the PR as is. Did I miss another reason to use a different generic function?

@musm
Copy link
Contributor

musm commented Dec 12, 2020

unsignedtype(T) and signedtype(T) is clear, explicit, and discoverable, i think.

I also agree that these names would be more descriptive. I'm not sure on the policy on renaming exported function in 1.x releases, but someone should chime in before we merge this.

@kimikage
Copy link
Contributor

BTW, PR #34864 was merged. So, at least #30444 could be closed.

The discussion in this PR may be still worth continuing, but it seems to be a good idea to change the subject of this PR.

cc: @NHDaly, @JeffreySarnoff

@JeffreySarnoff
Copy link
Contributor

what do you suggest?

@kimikage
Copy link
Contributor

To be honest, I am satisfied with the current signed(T)/unsigned(T), so I don't feel the need to discuss new function names or interfaces.

However, I think that adding functions such as unsignedtype will affect PR #36526.

@JeffreySarnoff
Copy link
Contributor

That makes sense. Certainly we should move the direction of these two so we are following one more consistent approach.

@StefanKarpinski StefanKarpinski added the triage This should be discussed on a triage call label Apr 7, 2021
@StefanKarpinski
Copy link
Member

Marked for triage to decide whether to do this or not.

@musm
Copy link
Contributor

musm commented Apr 7, 2021

Sounds good, triage should also considered this PR and the methods names in light of #36526 , as brought up by @kimikage and @JeffreySarnoff

@JeffBezanson JeffBezanson removed the triage This should be discussed on a triage call label Apr 8, 2021
@JeffBezanson
Copy link
Member

I believe we added this already.

@NHDaly NHDaly deleted the signed_T__unsigned_T branch August 16, 2024 02:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

unsigned(::Type) and signed(::Type)
9 participants