[Pitch] constexpr (i.e. compile-time constant expressions)

I’m not a language/library designer, just a wannabe designer; I’m hitting the limits of my expertise. So unless I lucked out, one of the main designers will have to pick up this ball and clean it up. I’ll be here if you need clarification on stuff.

The seed of this idea was coming up with a equivalent to C++’s “static_assert” after someone querying me about “constexpr” in one of my array threads. Using these expressions during a generic where clause is like using C++ static-assert during template instantiation. Using them during an #if/#elseif/#else surrounding an #error directive (explained in a separate post a couple days ago) acts like using C++ static-asserts in non-template contexts. (This is what inspired the #error proposal.)

Arbitrary Compiler Constant Expressions
Proposal: SE-NNNN <file:///Users/daryle/NNNN-filename.md>
Authors: Daryle Walker <https://github.com/CTMacUser&gt;, Author 2 <https://github.com/swiftdev&gt;
Review Manager: TBD
Status: Awaiting review
During the review process, add the following fields as needed:

Decision Notes: Rationale <https://lists.swift.org/pipermail/swift-evolution/&gt;, Additional Commentary <https://lists.swift.org/pipermail/swift-evolution/&gt;
Bugs: SR-NNNN <Issues · apple/swift-issues · GitHub, SR-MMMM <Issues · apple/swift-issues · GitHub;
Previous Revision: 1 <https://github.com/apple/swift-evolution/blob/...commit-ID.../proposals/NNNN-filename.md&gt;
Previous Proposal: SE-XXXX <file:///Users/daryle/XXXX-filename.md>
Introduction
This proposal adds compile-time constant expressions, and how to generate them, to the language

Swift-evolution thread: Discussion thread topic for that proposal <https://lists.swift.org/pipermail/swift-evolution/&gt;
Motivation
Conditional compilation blocks can shape the design of code based on criteria known at compile time. Generic where clauses do the same for limiting which generics can get instantiated. The criteria for either is extremely limited; Boolean combinations of environment state for conditional compilation blocks, and type equivalence, inheritance, or conformance for generic where clauses.

Other languages, like C++, allow values computed from specially-marked ("constexpr") objects and functions. Adding this to Swift permits user-side generated tests today, and is a first step to non-Boolean compile-time values, like value-based generic parameters or static-sized array extents.

Proposed solution
Define a new declaration attribute that can mark off data and code as being compatible with compile-time expressions (CTE). The definitions of the marked items may have restrictions to ensure realization at compile-time.
Designate parts of certain constructs that they can only take CTEs. Expressions outside of those constructs are evaluated at run-time as usual.
Specify the compile-time realization engine.
Speculate on what parts of the existing standard library should be updated to support CTEs.
(I have no concrete examples here. But I do think about if we had something like C++'s type-traits library, then we could replace direct type-testing generic where clauses with ones that check the desired traits from our type-traits equivalent structure.)

Detailed design
Computation

The language to process the Swift source file for CTEs before compilation is... the compiler itself! The Swift compiler can translate Swift as a compiled language or an interpreted language. CTEs are computed in an interpreted pass of the source code. If the translation is a Swift interpretation, then the translation goes over the CTE directly instead of a second interpreter.

There cannot be any recursion. If the CTE is within a statement or part of a generic where clause, then the CTE cannot refer either directly or indirectly to the function its modifying. If the CTE is within an initializer, then it cannot refer to the object being initialized. A fatal diagnostic shall be emitted. One shall also be emitted if CTE evaluation reaches a nonreturning function, an uncaught exception, or dereferenced nil optional or similar.

Floating-point operations of the interpreter do not have to exactly match the ones from the compiler. The interpreter will make its best interpolation to a value supported by the compiler.

The directly-referred symbols of a CTE (the ones seen at the CTE-accepting construct) can only be constant declarations and/or routines that do not have inout parameters or mutating self. Constants have to be literals, top-level objects, or type-level properties. A type-level variable computed property can be used if it's getter-only.

Complier implementations may place restrictions on the complexity, resource usage, time taken, etc., of CTE evaluation. (Put reasonable minimal limits here.)

Attribute

The define attribute shall indicate its owning declaration can be included in CTEs. The corresponding definition must be available through the source code; the attribute is ignored for items with their realization only as binary code.

Such items still have their run-time realization, as if the attribute was never assigned. The attribute can be applied to these declarations:

constant
variable
function (including operator definitions and closures)
initializer
subscript
Constants and variables must be of a CTE-compatible type. Routines must have parameters, returns, and self (if applicable) of a CTE-compatible type. Declarations within a protocol declaration cannot have the attribute, but ones defined in an extension or conforming type can. For methods in an extension for a protocol, the attribute is ignored if the conforming type's implementation methods cannot support it.

A compile-time expression can be of any type except:

One without any CTE-compatible initializers
Tuple types with any non-CTE members
Static-sized array types (if added) with a non-CTE element type
class types
struct types with instance-level stored properties of a non-CTE type
Raw-value enum types with a non-CTE implementation type
Union enum types with a non-CTE type as an associated value in at least one case
Types can have non-CTE (type-level and/or computed) properties and/or methods with non-CTE parameters/returns; they just can't be accessed/called during a CTE context.

Minimum Symbol Set Needed to Be CTE-Certified

Operators and other feasible operations associated with the default Boolean, string/character, and numeric types
The variables and methods of MemoryLayout
(Probably other stuff I haven't considered)
CTE Clients

CTE can only be used (for now) in conditional compilation blocks and generic where clauses. Since the former currently uses a custom hard-coded Boolean calculus, CTEs will be inserted with a wrapping syntax so the whole dynamic doesn't have to be converted at once.

For both conditional compilation blocks and generic where clauses, the CTE has to evaluate to a Boolean-compatible type.

Add to the "Grammar of a Conditional Compilation Block":

compilation-condition → inline ( compilation-expression )

compilation-expression → expression
Add to the "Grammar of a Generic Parameter Clause":

requirement → compilation-expression
Source compatibility
The changes in this proposal should be additive, so existing source should work as-is.

Effect on ABI stability
CTE computation happens at compile-time, so there is no direct effect on the ABI. Generic types and functions from the standard library may change their availability if they take on additional generic where clauses with CTE-based tests.

Effect on API resilience
The API would not be affected by this proposal, modulo any changes future standard library authors make to take advantage of complex tests.

Alternatives considered
The main alternative is to do nothing. That would keep conditional compilation blocks and generic where clauses as-is. Value-based generic arguments couldn't be implemented and static-sized arrays would use integer literals and/or custom syntax for their extents list.

···


Daryle Walker
Mac, Internet, and Video Game Junkie
darylew AT mac DOT com

Daryle, there have been discussions of this idea here in the past, and some
very good points have been brought up. Recently, there was also an attempt
to resurrect discussion on pure functions, which are of course a very much
related topic. Then, with the luxury of time, i was able to pull together a
good collection of links to the very extensive past conversations.
Unfortunately, I can’t do that for you today, but I’d highly recommend
taking a look at those conversations.

if you’re serious about this idea, putting in the hard but necessary work
of finding, digesting, and summarizing prior discussions for those who come
afterward is, I would argue, a key step. Admittedly, the mailing list
format is not at all easy to search, but one of the delights of the written
medium is that we’re called to engage past conversations in the present.

···

On Mon, Jun 12, 2017 at 07:20 Daryle Walker via swift-evolution < swift-evolution@swift.org> wrote:

I’m not a language/library designer, just a wannabe designer; I’m hitting
the limits of my expertise. So unless I lucked out, one of the main
designers will have to pick up this ball and clean it up. I’ll be here if
you need clarification on stuff.

The seed of this idea was coming up with a equivalent to C++’s
“static_assert” after someone querying me about “constexpr” in one of my
array threads. Using these expressions during a generic where clause is
like using C++ static-assert during template instantiation. Using them
during an if/#elseif/#else surrounding an error directive (explained in a
separate post a couple days ago) acts like using C++ static-asserts in
non-template contexts. (This is what inspired the error proposal.)

Arbitrary Compiler Constant Expressions

   - Proposal: SE-NNNN
   - Authors: Daryle Walker <https://github.com/CTMacUser&gt;, Author 2
   <https://github.com/swiftdev&gt;
   - Review Manager: TBD
   - Status: *Awaiting review*

*During the review process, add the following fields as needed:*

   - Decision Notes: Rationale
   <https://lists.swift.org/pipermail/swift-evolution/&gt;, Additional
   Commentary <https://lists.swift.org/pipermail/swift-evolution/&gt;
   - Bugs: SR-NNNN <Issues · apple/swift-issues · GitHub, SR-MMMM
   <Issues · apple/swift-issues · GitHub;
   - Previous Revision: 1
   <https://github.com/apple/swift-evolution/blob/...commit-ID.../proposals/NNNN-filename.md&gt;
   - Previous Proposal: SE-XXXX

Introduction

This proposal adds compile-time constant expressions, and how to generate
them, to the language

Swift-evolution thread: Discussion thread topic for that proposal
<https://lists.swift.org/pipermail/swift-evolution/&gt;
Motivation

Conditional compilation blocks can shape the design of code based on
criteria known at compile time. Generic where clauses do the same for
limiting which generics can get instantiated. The criteria for either is
extremely limited; Boolean combinations of environment state for
conditional compilation blocks, and type equivalence, inheritance, or
conformance for generic where clauses.

Other languages, like C++, allow values computed from specially-marked ("
constexpr") objects and functions. Adding this to Swift permits user-side
generated tests today, and is a first step to non-Boolean compile-time
values, like value-based generic parameters or static-sized array extents.
Proposed solution

   1. Define a new declaration attribute that can mark off data and code
   as being compatible with compile-time expressions (CTE). The definitions of
   the marked items may have restrictions to ensure realization at
   compile-time.
   2. Designate parts of certain constructs that they can only take CTEs.
   Expressions outside of those constructs are evaluated at run-time as usual.
   3. Specify the compile-time realization engine.
   4. Speculate on what parts of the existing standard library should be
   updated to support CTEs.

(I have no concrete examples here. But I do think about if we had
something like C++'s type-traits library, then we could replace direct
type-testing generic where clauses with ones that check the desired traits
from our type-traits equivalent structure.)
Detailed designComputation

The language to process the Swift source file for CTEs before compilation
is... the compiler itself! The Swift compiler can translate Swift as a
compiled language or an interpreted language. CTEs are computed in an
interpreted pass of the source code. If the translation is a Swift
interpretation, then the translation goes over the CTE directly instead of
a second interpreter.

There cannot be any recursion. If the CTE is within a statement or part of
a generic where clause, then the CTE cannot refer either directly or
indirectly to the function its modifying. If the CTE is within an
initializer, then it cannot refer to the object being initialized. A fatal
diagnostic shall be emitted. One shall also be emitted if CTE evaluation
reaches a nonreturning function, an uncaught exception, or dereferenced nil
optional or similar.

Floating-point operations of the interpreter do not have to exactly match
the ones from the compiler. The interpreter will make its best
interpolation to a value supported by the compiler.

The directly-referred symbols of a CTE (the ones seen at the CTE-accepting
construct) can only be constant declarations and/or routines that do not
have inout parameters or mutating self. Constants have to be literals,
top-level objects, or type-level properties. A type-level variable computed
property can be used if it's getter-only.

Complier implementations may place restrictions on the complexity,
resource usage, time taken, etc., of CTE evaluation. (Put reasonable
minimal limits here.)
Attribute

The define attribute shall indicate its owning declaration can be
included in CTEs. The corresponding definition must be available through
the source code; the attribute is ignored for items with their realization
only as binary code.

Such items still have their run-time realization, as if the attribute was
never assigned. The attribute can be applied to these declarations:

   - constant
   - variable
   - function (including operator definitions and closures)
   - initializer
   - subscript

Constants and variables must be of a CTE-compatible type. Routines must
have parameters, returns, and self (if applicable) of a CTE-compatible
type. Declarations within a protocol declaration cannot have the attribute,
but ones defined in an extension or conforming type can. For methods in an
extension for a protocol, the attribute is ignored if the conforming type's
implementation methods cannot support it.

A compile-time expression can be of any type except:

   - One without any CTE-compatible initializers
   - Tuple types with any non-CTE members
   - Static-sized array types (if added) with a non-CTE element type
   - class types
   - struct types with instance-level stored properties of a non-CTE type
   - Raw-value enum types with a non-CTE implementation type
   - Union enum types with a non-CTE type as an associated value in at
   least one case

Types can have non-CTE (type-level and/or computed) properties and/or
methods with non-CTE parameters/returns; they just can't be accessed/called
during a CTE context.
Minimum Symbol Set Needed to Be CTE-Certified

   - Operators and other feasible operations associated with the default
   Boolean, string/character, and numeric types
   - The variables and methods of MemoryLayout
   - (Probably other stuff I haven't considered)

CTE Clients

CTE can only be used (for now) in conditional compilation blocks and
generic where clauses. Since the former currently uses a custom hard-coded
Boolean calculus, CTEs will be inserted with a wrapping syntax so the whole
dynamic doesn't have to be converted at once.

For both conditional compilation blocks and generic where clauses, the CTE
has to evaluate to a Boolean-compatible type.

Add to the "Grammar of a Conditional Compilation Block":

*compilation-condition* → *inline* *(* *compilation-expression* *)*

*compilation-expression* → *expression*

Add to the "Grammar of a Generic Parameter Clause":

*requirement* → *compilation-expression*

Source compatibility

The changes in this proposal should be additive, so existing source should
work as-is.
Effect on ABI stability

CTE computation happens at compile-time, so there is no direct effect on
the ABI. Generic types and functions from the standard library may change
their availability if they take on additional generic where clauses with
CTE-based tests.
Effect on API resilience

The API would not be affected by this proposal, modulo any changes future
standard library authors make to take advantage of complex tests.
Alternatives considered

The main alternative is to do nothing. That would keep conditional
compilation blocks and generic where clauses as-is. Value-based generic
arguments couldn't be implemented and static-sized arrays would use integer
literals and/or custom syntax for their extents list.


Daryle Walker
Mac, Internet, and Video Game Junkie
darylew AT mac DOT com

_______________________________________________
swift-evolution mailing list
swift-evolution@swift.org
https://lists.swift.org/mailman/listinfo/swift-evolution