-
-
Notifications
You must be signed in to change notification settings - Fork 48
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
Introduce new numerics #200
base: main
Are you sure you want to change the base?
Changes from all commits
ae3e458
bd6e0a4
fabfb84
2add507
d264e5a
ecb4606
0e08bdf
5e90235
b84afa1
103fd97
c8df400
3d71de0
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,230 @@ | ||
- Feature Name: Introduce new numerics -- `Rational`, `BigInt`, `BigFloat`, `Complex` | ||
- Start Date: 2022-02-28 | ||
- RFC PR: | ||
- Pony Issue: | ||
|
||
# Summary | ||
|
||
This RFC proposes the introduction of new numeric types; in particular the addition of a type representing a fractional number (`Rational`), arbitrary precision integer (`BigInt`), arbitrary precision float (`BigFloat`), and complex number (`Complex`). | ||
|
||
# Motivation | ||
|
||
The primary motivation for adding these types to the stdlib is to have a single canonical implementation of them which allow interoperability of numeric types across the Pony ecosystem. | ||
|
||
# Detailed design | ||
|
||
I propose we add the aforementioned numeric types into `builtin` so they exist alongside the other standard numeric types. These introduced numeric types **must** existing within the current numeric type hierarchy by being compliant with existing numeric traits. | ||
|
||
## Numeric Hierarchy | ||
|
||
Current the Pony numeric type hierarchy is as follows: | ||
|
||
```mermaid | ||
classDiagram | ||
class Any | ||
<<interface>> Any | ||
class Real | ||
<<trait>> Real | ||
class FloatingPoint | ||
<<trait>> FloatingPoint | ||
class Integer | ||
<<trait>> Integer | ||
class SignedInteger | ||
<<trait>> SignedInteger | ||
class UnsignedInteger | ||
<<trait>> UnsignedInteger | ||
|
||
Any <-- Real | ||
Real <-- FloatingPoint | ||
Real <-- Integer | ||
Integer <-- SignedInteger | ||
Integer <-- UnsignedInteger | ||
|
||
FloatingPoint <-- F32 | ||
FloatingPoint <-- F64 | ||
|
||
UnsignedInteger <-- U8 | ||
UnsignedInteger <-- U16 | ||
UnsignedInteger <-- U32 | ||
UnsignedInteger <-- U64 | ||
UnsignedInteger <-- U128 | ||
UnsignedInteger <-- USize | ||
UnsignedInteger <-- ULong | ||
|
||
SignedInteger <-- I8 | ||
SignedInteger <-- I16 | ||
SignedInteger <-- I32 | ||
SignedInteger <-- I64 | ||
SignedInteger <-- I128 | ||
SignedInteger <-- ISize | ||
SignedInteger <-- ILong | ||
``` | ||
|
||
This RFC introduces four more numeric types: `Rational`, `BigInt`, `BigFloat`, and `Complex`. These fit into the numeric type hierarchy in the following manner: | ||
|
||
```mermaid | ||
classDiagram | ||
class Any | ||
<<interface>> Any | ||
class Real | ||
<<trait>> Real | ||
class FloatingPoint | ||
<<trait>> FloatingPoint | ||
class Integer | ||
<<trait>> Integer | ||
class SignedInteger | ||
<<trait>> SignedInteger | ||
class UnsignedInteger | ||
<<trait>> UnsignedInteger | ||
|
||
Any <-- Real | ||
Any <-- Complex~Real~ | ||
Real <-- FloatingPoint | ||
Real <-- Integer | ||
Real <-- Rational~Integer~ | ||
Integer <-- SignedInteger | ||
Integer <-- UnsignedInteger | ||
Integer <-- BigInt | ||
|
||
FloatingPoint <-- F32 | ||
FloatingPoint <-- F64 | ||
FloatingPoint <-- BigFloat | ||
|
||
UnsignedInteger <-- U8 | ||
UnsignedInteger <-- U16 | ||
UnsignedInteger <-- U32 | ||
UnsignedInteger <-- U64 | ||
UnsignedInteger <-- U128 | ||
UnsignedInteger <-- USize | ||
UnsignedInteger <-- ULong | ||
|
||
SignedInteger <-- I8 | ||
SignedInteger <-- I16 | ||
SignedInteger <-- I32 | ||
SignedInteger <-- I64 | ||
SignedInteger <-- I128 | ||
SignedInteger <-- ISize | ||
SignedInteger <-- ILong | ||
``` | ||
|
||
## Methods of Concern | ||
|
||
```pony | ||
trait val Real[A: Real[A] val] is | ||
(Stringable & _ArithmeticConvertible & Comparable[A]) | ||
... | ||
new val min_value() | ||
new val max_value() | ||
... | ||
``` | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I understand that this has already been present, but given that this is re-design I'm not clear on the benefit of these constructors, but there's some obvious costs. Is there a use case for max_value() / min_value() that makes sense for all numbers? Given that this doesn't work for BigInt/BigFloat/similar types I'd be hesitant to place it at the top of the hierarchy. Can we instead make a Bounded interface? This will only break code which relies on these methods being present in any Real[A] There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I don't understand what you are suggesting. What's a "bounded interface"? There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I think he means a new interface whose name is There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. @jasoncarr0 is that what you meant? re: There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I meant a literal Pony interface, that is: There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Actually there is also an argument for removing Ordered, but only because this trait is really the lowest possible. Mostly that the implementations for things like matrices would be sketch but making them numerics makes sense and is powerful. |
||
|
||
`Real` will exist above `Rational`, `BigInt`, and `BigFloat` and as such would require defining the above methods for these types. `Rational` can be defined as minimum and maximum of the numerator, however `BigInt` and `BigFloat` by definition have arbitrary precision making defining a minimum and maximum difficult at the least -- if we define them as the minimum and maximum of a machine-sized int and float, or define them as -Inf and Inf -- or impossible at the worst -- if we define them by their possible limits which are arbitrary. | ||
|
||
```pony | ||
trait val Integer[A: Integer[A] val] is Real[A] | ||
... | ||
fun op_and(y: A): A => this and y | ||
fun op_or(y: A): A => this or y | ||
fun op_xor(y: A): A => this xor y | ||
fun op_not(): A => not this | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This can't be implemented by a BigInt, so it's odd that a BigInt can't actually be an Integer There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Sorry, I don't understand your comment. Can you try explaining in a different way? There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. You can't do a bitwise "not" operation on a You can't because the you'd expect that any bits "more significant" than the most significant bit of the value would be set to But a There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I still don't understand @jemc.
Can you explain how what you said is related to Jason's comment? I don't understand why not being able to be implemented BigInt makes it odd that BigInt can't actually be an Integer. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I'll copy Jason's words and insert parentheticals with my own commentary: This (the So it's odd that a There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Thanks @jemc. I understand now. |
||
|
||
fun bit_reverse(): A | ||
""" | ||
Reverse the order of the bits within the integer. | ||
For example, 0b11101101 (237) would return 0b10110111 (183). | ||
""" | ||
|
||
fun bswap(): A | ||
``` | ||
|
||
`Integer` will exist above `BigInt` and as such would require defining the above methods -- however `BigInt` will be defined via other numerics so these methods could be applied recursively. | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I don't understand what this means. What are "the above methods"? Those listed on the Integer trait? " |
||
|
||
```pony | ||
trait val FloatingPoint[A: FloatingPoint[A] val] is Real[A] | ||
new val min_normalised() | ||
new val epsilon() | ||
fun tag radix(): U8 | ||
fun tag precision2(): U8 | ||
fun tag precision10(): U8 | ||
fun tag min_exp2(): I16 | ||
fun tag min_exp10(): I16 | ||
fun tag max_exp2(): I16 | ||
fun tag max_exp10(): I16 | ||
... | ||
fun abs(): A | ||
fun ceil(): A | ||
fun floor(): A | ||
fun round(): A | ||
fun trunc(): A | ||
|
||
fun finite(): Bool | ||
fun infinite(): Bool | ||
fun nan(): Bool | ||
|
||
fun ldexp(x: A, exponent: I32): A | ||
fun frexp(): (A, U32) | ||
fun log(): A | ||
fun log2(): A | ||
fun log10(): A | ||
fun logb(): A | ||
|
||
fun pow(y: A): A | ||
fun powi(y: I32): A | ||
|
||
fun sqrt(): A | ||
|
||
fun sqrt_unsafe(): A | ||
""" | ||
Unsafe operation. | ||
If this is negative, the result is undefined. | ||
""" | ||
|
||
fun cbrt(): A | ||
fun exp(): A | ||
fun exp2(): A | ||
|
||
fun cos(): A | ||
fun sin(): A | ||
fun tan(): A | ||
|
||
fun cosh(): A | ||
fun sinh(): A | ||
fun tanh(): A | ||
|
||
fun acos(): A | ||
fun asin(): A | ||
fun atan(): A | ||
fun atan2(y: A): A | ||
|
||
fun acosh(): A | ||
fun asinh(): A | ||
fun atanh(): A | ||
``` | ||
|
||
`FloatingPoint` will exist above `BigFloat` and as such would require defining the above methods -- many of which are ill-defined under arbitrary precision, or are functions using C-FFI and/or LLVM intrinsics. | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I'm confused. Why would we do it this way if it has these problems? This seems like an argument against doing it this way. I feel like there's some meaning here that I am missing. |
||
|
||
# How We Teach This | ||
|
||
Adding ample documentation to these new numerics should suffice to teach Pony users how to leverage these types in their programs. I do not think any additions to the Pony Tutorial are needed, however if additions are desired than [Arithmetic](https://tutorial.ponylang.io/expressions/arithmetic.html) may be the most sensible location. | ||
|
||
# How We Test This | ||
|
||
I recommend use of `pony_check` to test all reversible operations pairs (`x+y-y == x`, `x*y/y == x`, etc), precision persistence (`Rational[U8](where numerator=x, denominator=y) * y == x`), and overflow/underflow protection (`Rational[U8](255, 1) + 1 => error`). | ||
|
||
Testing these numerics should not affect any other parts of Pony and as such standard CI should suffice. | ||
|
||
# Drawbacks | ||
|
||
+ Additional maintenance cost | ||
+ May break existing code if methods must be removed from existing numeric traits to match the suggested hierarchy placements | ||
|
||
# Alternatives | ||
|
||
Alternatively, we can introduce these types in `math` as opposed to `builtin` and/or only introduce some of the proposed new numeric types. | ||
|
||
# Unresolved questions | ||
|
||
+ Should these types be introduced in `builtin` or in `math`? | ||
+ Does `Rational` make sense as the "fractional type" or would we prefer `Fractional` to avoid confusion? | ||
+ Do we want to also include a `Decimal` type? | ||
+ How should we handle the stated "Methods of Concern"? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
For me, the preference would be to keep
builtin
as minimal as possible, so I'd prefer to put these new types in a new package (or series of packages), unless there is a strong motivation for it to go inbuiltin
.Pretty much all of the existing public types in
builtin
have hard requirements forcing them to be there, with a reason like one of the following:None
, string literals, numeric literals, etc)Array
,String
, etc)compile_intrinsic
for one or more function definitionsEnv
, which is passed to theMain
actor on entryI don't think these new types have any hard requirements forcing them to go in
builtin
, so I believe they shouldn't be there.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We can discuss this further on Sync and I will update the RFC according to our conversation. I have no particular need for these to be added to
builtin
so no objection to an agreed upon other location such as a newly creatednumerics
or adding these tomath
(as was the RFC state prior to my latest changes).