Skip to content

Latest commit

 

History

History
705 lines (570 loc) · 34.2 KB

p1083.md

File metadata and controls

705 lines (570 loc) · 34.2 KB
Raw

Arithmetic expressions

Pull request

Table of contents

Problem

Carbon needs a set of arithmetic operators in order to perform basic calculations on various kinds of built-in and user-defined numbers and number-like values: integers, real numbers, complex numbers, vectors, matrices, and so on.

Background

Symbols

Conventions for arithmetic operators have a very long tradition. The following symbols have common meaning across C, C++, Java, JavaScript, Rust, Swift, and many other languages:

  • + and - mean addition and subtraction, as in mathematics. Unary - forms a negated value -- an additive inverse. Sometimes, unary + is permitted, as a no-op.
  • * means multiplication, diverging from the use of '×', '⋅', or simply juxtaposition in normal mathematical notation. However, this symbol does have some visual similarity to '×'.
  • / means division, diverging from the use of '÷', fraction notation, or an exponent of -1 in mathematics. However, this symbol does somewhat visually resemble fractional notation that has "fallen over" to the left.
  • % means remainder, diverging from the use of the word "mod" in mathematics, and, perhaps confusingly, resembling the '÷' division operator from mathematics.

Semantics

Efficient integer types often have finite bounds on the numbers they can represent, and we expect Carbon's to be no different. This presents a problem for operations that might produce values outside those bounds: what should the behavior of a primitive arithmetic operation be if the result cannot be represented? Moreover, some operations, such as division by zero, have no mathematically-defined result.

Different languages take different approaches to this problem.

  • In C and C++, signed integer overflow and division by zero -- including floating-point division by zero -- have undefined behavior, meaning that there are no constraints on the behavior of a program that performs such a calculation.
  • In Rust, these conditions are classified as being errors, which, depending on various factors, will either panic or give a two's complement result. Rust also provides a Wrapping<T> type, where T is a signed or unsigned integer type, that provides arithmetic with guaranteed two's complement wrapping semantics.
  • In Swift, overflow triggers a runtime fatal error. Arithmetic operators can be prefixed with &, such as big_num &* other_big_num, to request two's complement wrapping behavior instead.
  • In Java, integer arithmetic has two's complement wrapping behavior. Division by zero throws an exception.
  • In JavaScript, all arithmetic is (at least notionally) performed in IEEE 754 double-precision floating-point, so all operations have defined results, although some integer operations may produce non-integer results, such as infinities or NaNs, and some produce incorrect integer results. For example, in JavaScript, 100000001 * 100000001 evaluates to 10000000200000000 not to 10000000200000001.
  • In LLVM, the result of overflow is a poison value, which results in undefined behavior only when it is observed, for example by a branch, and otherwise propagates to dependent values. This permits speculative execution of arithmetic that might overflow.

Proposal

Carbon will provide the five usual binary arithmetic operators -- +, -, *, /, and % -- with their usual semantics. A unary minus - operator will also be provided, but no unary plus + operator.

Notation Meaning
-a Negation
a + b Addition
a - b Subtraction
a * b Multiplication
a / b Division
a % b Modulo / remainder

These operators follow the same precedence rules as in C-family languages -- multiplicative operators bind tighter than additive operators, and negation binds tighter than multiplicative operators. Unlike in other C-family languages, no precedence is defined between % and other binary operators, so a + b % c * d is an error, and does not mean a + (b % c) * d as it would in C++.

Unlike in C++, lossy conversions are never performed. Instead, built-in arithmetic is not permitted unless one of the operand types can represent all the values of the other operand. The calculation is performed in that operand type; there is no implicit widening to Carbon's equivalent of int.

Unsigned integer types uN model arithmetic modulo 2N, and we strongly advise that they are not used except when those semantics are desired, such as in random number generation, cryptography, hashing, and similar domains.

For signed integer types iN, it is a programming error if overflow occurs. We guarantee that in development build modes such errors will result in a runtime trap, and that in hardened build modes the result will either be a trap or the two's complement value.

Initially, no support will be provided for wrapping signed integer arithmetic, nor for non-wrapping unsigned integer arithmetic. This can be added by a future proposal if we find there is sufficient demand for either.

Floating-point types use IEEE 754 semantics, with round-to-nearest rounding mode, no signaling NaNs, and no floating-point exceptions.

This proposal takes no position on whether the + operator is supported on strings to perform concatenation.

See the changes to the design for more details; the rest of this proposal will focus on the rationale and alternatives.

Rationale based on Carbon's goals

  • Language tools and ecosystem
    • Treating overflow as a programming error empowers static and dynamic analysis tools to distinguish arithmetic bugs from intentional wraparound.
  • Performance-critical software
    • Allowing the optimizer to assume that signed integer overflow does not occur in performance build modes allows certain important loop optimizations to fire that would otherwise be incorrect.
    • Avoiding extending into larger types when performing arithmetic on small operands aids in the ability to vectorize.
  • Software and language evolution
    • Each subexpression of an arithmetic expression is given a specific type, rather than being computed in a type of sufficient width for any possible value, in order to make it simple to factor out subexpressions.
  • Code that is easy to read, understand, and write
    • Using the same operators as other languages, with largely the same semantics as C++, should aid readability and writeability especially for those with less Carbon-specific knowledge.
    • Performing built-in arithmetic in the larger operand type and refusing cases where a lossy conversion would be performed in C++ reduces the scope for confusion and surprises.
  • Practical safety and testing mechanisms
    • Treating signed integer overflow as an error condition, and having it produce a runtime error in some build modes, will assist with certain forms of testing and with analysis of program behavior. For example, fuzz testing can be used to locate overflow bugs, with confidence that any overflow detected is unintentional.
  • Modern OS platforms, hardware architectures, and environments
    • The choice of division and modulo semantics aims to provide fast execution and small code size on modern hardware architectures.
  • Interoperability with and migration from existing C++ code
    • Using the same operator set as C++ may aid migration of C++ developers and make code using interop between Carbon and C++ easier to follow.
    • Making similar choices for the behavior of signed and unsigned types will make a correct but unidiomatic conversion of C++ into Carbon easier; however, in idiomatic Carbon, signed types are expected to be used more frequently.

Choice of operator symbols

We want Carbon to provide a close analogue of standard mathematical expression notation, with some unsurprising set of operators. While accessibility to beginners is important, the most important audience for which the operators set should be unsurprising is programmers with C++ or Rust experience. The choice of +, -, *, /, and % is ubiquitous among programming languages, and any significant change to this symbol set for the benefit of programming novices with a mathematical background would be a detriment to those with a background in other programming languages.

Therefore we choose the conventional symbol set.

Signed integer semantics

Carbon prioritizes fast and predictable performance above all other factors. Allowing an optimizer to assume that certain forms of arithmetic do not result in overflow can improve program performance and allow the generation of smaller code. Moreover, while in general we only want to provide the means for accessing the fastest possible implementation of an operation, we expect integer arithmetic to be so ubiquitous that it's important that the most efficient option be the default option.

At the same time, it's important that Carbon code can be used in safety-critical scenarios where unbounded undefined behavior on integer arithmetic is problematic, so in a hardened build mode, we provide stronger guarantees.

Alternatives considered

Use a sufficiently wide result type to avoid overflow

We could define that integer operators always produce a sufficiently wide result type such that overflow never occurs, and defer all overflow handling -- checking for out-of-bounds values, wrapping around, or a programmer assertion that overflow simply doesn't happen -- until the end of the computation or some explicit step.

For example:

fn F(a: i32, b: i32, c: i32) {
  // a * b is computed in i63,
  // ... + c is computed in i64.
  // =! is an assertion that the value is in-range.
  let d: i32 =! a * b + c;

  // Same, but =% says to wrap around.
  let e: i32 =% a * b + c;

  // If the value doesn't fit in the specified type, pattern-matching fails.
  let f: i32 = a * b + c else { return; }
}

We could put a limit on how large an intermediate type can be, and reject if it would require a larger intermediate operand than the largest we can efficiently support.

This approach seems quite promising, but close inspection finds a number of non-trivial issues.

Advantages:

  • No implicit undefined behavior or incorrect results on overflow.
  • Forces developer to think about the possibility of overflow and write down what they intend to happen.
  • If the final assignment is =% or =!, all intermediate arithmetic other than division and remainder can be done in the result type, avoiding the need for wide computations.

Disadvantages:

  • This approach is novel and it's unknown whether developers would accept its ergonomic burden.
  • Decreases the uniformity of the model. While we can generalize =! to mean "assign assuming the RHS fits into the LHS", there doesn't seem to be any good generalization of =% to other situations and types.
  • The largest type for which we can truly efficiently perform arithmetic is i64 / u64. While 128-bit arithmetic is typically available on our target architectures, use of it will often be less efficient and may increase register pressure. Calculations as simple as (a + b) / c may be rejected if the operands are already in the largest efficient type.
  • Operations such as negation and division can increase the width of the operand: -i32.Min and i32.Min / -1 don't fit in i32. The following approaches to this problem were considered:
    • Remove the Min value from iN types, so the negative and positive range are identical. However, this would severely violate programmer expectations, for example in some important bit-manipulation cases.
    • Give integer types a range of values rather than simply a bit-width. For example, we can say that negation on i32 produces a type that can represent [-231+1, 231], which still fits in 32 bits. However, this would add significant complexity to the type system, and with this approach, division would still increase the bit width: for example, a / b, where a and b are iNs, has 2N+1 distinct possible values. This is especially surprising because integer division is usually expected to make a number smaller!
  • Refactoring code becomes more challenging, as the appropriate intermediate type must be determined. Mitigating this, the type system would inform the programmer when they make a mistake.
  • The ! and % annotations become quite viral: they would be needed not only in initialization and assignment, but likely also in case, return, and other initialization contexts not using =. As an alternative, the annotation could be put on the outermost arithmetic operator, but that is likely to be syntactically awkward, especially in unparenthesized expressions such as a + b +% c.
  • It will likely become idiomatic in many communities to either always use % or always use !.
    • Always using ! means that the developers see no benefit compared to the behavior as proposed here, and nonetheless pay an ergonomic cost.
    • Always using % means that we lose any ability to distinguish between intentional wraparound and overflow, making a class of bugs that would otherwise be easily detectible be undetectable without making the program behavior correct.
  • Division by zero is still not handled.

Guarantee that the program never proceeds with an incorrect value after overflow

We could say that overflow errors always result in program termination, possibly with the permission for the compiler to give the mathematically correct result instead if it so chooses.

This would result in slower and larger code being generated in a lot of cases. Some optimizations would still be possible if the optimizer could ensure that it only increased the set of cases for which the correct result is given, but current optimizers are not well-suited to perform that task.

Given Carbon's emphasis on performance, this approach is rejected without prejudice until we have a demonstration that no important performance metric is harmed by it.

Guarantee that all integer arithmetic is two's complement

Instead of making overflow a programming error, we could define it as two's complement. This is, for example, the approach taken by Java.

The major problem with this approach is that it makes erroneous wrapping and intentional wrapping indistinguishable. This would make finding such bugs much harder for readers of the code, and all but impossible for static and dynamic analysis tools. Making overflow issues programming errors allows problems to be caught earlier and more reliably.

Treat overflow as an error but don't optimize on it

We could follow Rust and say that overflow is an error, but that we promise we'll either catch it or give two's complement semantics. This approach is currently rejected for the same reason we reject guaranteeing we catch all cases where we can't give a correct result.

Don't let Unsigned arithmetic wrap

We could treat Unsigned(N) like Integer(N), and make it an error by default if it overflows. However, this doesn't seem like a great fit for the problem domain.

There are, broadly speaking, two different classes of use cases we want to support:

  • Cases where the developer wants a number. They might have some expectation of the range of the number -- eg, non-negative, or strictly positive, or between -1 and 100 -- or they might just want a number and not have explicit bounds. We expect iN to be used for all such cases, and do not want to treat the non-negative cases as a distinct and special type.
  • Cases where the developer wants the ring ℤ/2nℤ of integers modulo 2n, for example in cryptography, hashing, or random number generation.

The second class is certainly rarer than the first, but contains many important use cases.

The typical arguments for using an unsigned type for the first class of use cases are:

  • Better expression of developer intent. It is generally preferable to make invalid states unrepresentable, and if negative numbers are invalid, then unsigned types are better suited than signed types. However, supporting this use case with the same types used to support the wrapping use cases results in a situation where one side or the other has to make compromises. The alternative would be to have three different kinds of type: signed, unsigned, and modulo. But in that setup, it's not clear that the value added by unsigned types is worthwhile. Also, it's common to want to take the difference of such unsigned quantities, and it's generally preferable for such subtractions to produce a negative result rather than a subtle bug. Moreover, while a restriction to non-negative values is common, supporting only the case of a range restriction to [0, 2N-1], but not any other range, does not do a good job of addressing the general desire to capture intent and to make invalid states unrepresentable.
  • Ability to reduce storage size. Spending a sign bit every time a number is stored, even when it's known to be non-negative is wasteful. This is an important concern, and one we should address, but it's thought to be better to address this by annotating a field with a description of how it should be packed -- such as in a bit-field -- rather than by changing the type of the data and possibly the behavior of operations on it.

Additionally, providing unsigned types for which overflow is an error introduces a much larger risk of that error state being inadvertently reached than for signed types, because calculations typically involve numbers that are close to zero, which is far from overflowing in signed types but near to overflowing in unsigned types. For example, a calculation such as a - b + c may be known to always produce a non-negative result, and the developer may be tempted to use non-negative non-wrapping types for a, b, and c, but doing so introduces the risk that the intermediate a - b calculation is negative. By contrast, overflow would only occur in a signed calculation if the numbers involved were very large.

Provide separate wrapping types

We could provide distinct types for wrapping versus non-wrapping use cases, independent of the choice of signedness. This approach is being followed by Rust.

Advantages:

  • Avoids coupling two decisions that are logically independent.
  • Allows developer intent to be expressed more explicitly, in the case where the intent is a non-negative but non-wrapping integer.

Disadvantages:

  • Twice as many kinds of integer types, likely meaning twice as many impls need to be defined for each operation, twice as many overloads in each overload set, and so on.
  • The presence of unsigned types for which overflow is an error has a greater risk of inadvertent overflow, as described in the previous section.
  • Less familiar to those coming from C++.

Do not provide an ordering or division for uN

For arithmetic purposes, we treat uN as the integers modulo 2N. There is no single canonical total order for that ring, and because multiplication by even numbers loses information by discarding the high-order bit, not all non-zero numbers have multiplicative inverses, and so division is not well-defined. We could be more mathematically pure by refusing to implement < and friends for uN, and similarly refusing to support /.

However, the uN types exist as much for pragmatic purposes as for mathematical ones. Making the choice to treat all uN values as non-negative is somewhat arbitrary, but is both unsurprising to those coming from C++ and useful for the cases where some ordering is desired.

Give unary - lower precedence

In this proposal, - a * b is interpreted as (-a) * b. We could instead interpret it as - (a * b).

Advantages:

  • Can be argued as better following mathematical convention.

Disadvantages:

  • Unfamiliar to people coming from most other programming languages, including C++.
  • If we followed our normal precedence partial ordering rule, this would mean that a * -b is invalid, because * has higher precedence than -.

Comparison to other languages:

  • C and C++ give all unary operators (including unary - and unary +) higher precedence than any binary operator.
  • Go, Rust, and Swift give unary - higher precedence than multiplication.
  • Python binds unary - more tightly than multiplication but less tightly than exponentiation, so - a * b is (- a) * b but - a ** b is - (a ** b).
  • Haskell gives unary - the same precedence as binary - and rejects a * - b.

Include a unary plus operator

C and C++ include a unary + operator. In principle, this operator permits lists of numbers to be written with an operator attached to each:

int arr[] = {
  -100,
  -50,
  +20,
  +400
};

... but in practice this use case is rare at best, and unary + in C++ is instead mainly used to coerce operands to prvalues of built-in types. For example, auto *p = +[]{ /*...*/ }; might be used to create a function pointer (auto deduction would fail without the unary + operator), and min(+Class::static_member, x) might be used to force Class::static_member to be loaded, to avoid requiring the static member to be defined.

Such usage of + is a design wart that we need not replicate. If we want a mechanism to decay an operand, we can pick a better name for it than +.

Floating-point modulo operator

It would be possible and sometimes useful to support the % operator for floating-point types.

Advantages:

  • When desired, f1 % f2 would be a more concise notation than a call to an fmod function (however it is named).

Disadvantages:

  • Uses of this operation would likely be rare and unfamiliar enough that people would assume an integer operation is being performed when they see the operator.
  • This operation is not available in hardware in many modern architectures, and providing it as a built-in operator may create a false impression of its implementation as a non-trivial library function.
  • We lack sufficient evidence of utility to propose it at this time.

Provide different division operators

We could follow Python3 and provide an / operator for integers that produces a floating-point (or perhaps rational) type. This would be mathematically clean: ignoring overflow and precision loss, all standard properties for division would be maintained. Or we could follow Haskell and refuse to provide an / operator between integers on the basis that such division is not mathematically defined for that type. Or we could provide a division operator that produces a pair of dividend and modulo, or an Optional(Int) that is absent whenever the division is inexact. In each case, functionality not available through operators could be provided with named functions instead.

All of these options are likely to be surprising to programmers coming from C++, to a level that outweighs the perceived benefit.

Use different division and modulo semantics

There are multiple somewhat-reasonable ways to define division operations for signed integers (whether we provide those operations as operators or library functions). Assuming operator notation for now, and that we define modulo as a % b == a - a / b * b, the following options all have merit:

Property Round towards zero (truncating division) Round towards negative infinity (floor division) Round based on sign of divisor[1] (Euclidean division)
(-a) / b ==
a / (-b)		 👍 Yes 👍 Yes No
(-a) / b ==
-(a / b)		 👍 Yes No 👍 Yes
a / (-b) ==
-(a / b)		 👍 Yes No No
(a + k * b) / b
== a / b + k		 No 👍 Yes 👍 Yes
Sign of a % b Same as a 👍 Same as b 👍 Never negative
x86 instruction? 👍 Yes: cqo (or similar) + idiv First option + fixup:
s * (a / (s * b)) where s is sign(a) * sign(b)[2]		 First option + fixup:
a / b - (a % b < 0)
LLVM IR + optimization support 👍 Yes No No
Use in existing languages C, C++, Rust, Swift
quotRem in Haskell		 // and % in Python
/ in Python 2 only
divMod in Haskell		 None?

The cells marked 👍 suggest generally desirable properties. For further reading, see this Microsoft Research paper.

Our options here are as follows:

  • Pick one of the two interpretations as the meaning of /, and (optionally) pick one of the above interpretations as the meaning of %); perhaps provide the others as library functions or as additional operators.
  • Do not provide any of these operators for integers and provide only named functions.

Note that we are not required to provide a % that is consistent with our chosen / operator. (We could pick truncate-towards-zero for / and truncate-towards-negative-infinity for %, for example.) There is long-established tradition here, but it's unclear to what extent practicing programmers really care about the relationship between / and %.

It is likely that most Carbon code that performs division and modulo between signed integer types does not actually care about what happens when either operand is negative. Therefore, following our goals of supporting high-performance code and current CPU architectures, we will choose to implement / and % as division with truncation towards zero. The other variants can be provided by a library function, if at all.

[1]: That is: for positive divisors, round towards negative infinity; for negative divisors, round towards positive infinity.

[2]: Here, sign(a) is 1 if a >= 0 and is -1 if a < 0.

Use different precedence groups for division and multiplication

Under this proposal, division and multiplication are in the same precedence group, and are left-associative: a * b / c * d is ((a * b) / c) / d, and not (a * b) / (c * d) or some other grouping.

It's not feasible to provide a different interpretation here due to the risk of confusing developers migrating from other languages. However, we could reject such expressions and require explicit parentheses.

While the value of accepting code such as this is relatively low, the fact that both existing programming languages and common mathematical education treat multiplication and division as the same, left-associative, precedence group means that it's unlikely to be a significant burden to expect Carbon developers to remember this precedence rule.

Use the same precedence group for modulo and multiplication

In most languages with the set of arithmetic operators discussed in this proposal, % is considered a multiplicative operator, and so a sequence of *, /, and % operators is processed left-to-right. In some sense this is reasonable: % is notionally performing a division, after all. Moreover, in a code search, I was unable to find evidence that it's common for precedence errors with % to be checked in to source control, and C++ compilers don't have warnings for mixing % with + without parentheses.

Giving % and / the same precedence also allows some kinds of code to be written symmetrically:

char two_digit_number[] = {
  '0' + m / 10,
  '0' + m % 10,
  0
};
char four_digit_number[] = {
  '0' + n / 1000,
  '0' + n / 100 % 10,
  '0' + n / 10 % 10,
  '0' + n % 10,
  0
};

With minimal parentheses, that example would be written as follows under this proposal:

var two_digit_number: Array(Char, 3) = (
  '0' + m / 10,
  '0' + (m % 10),
  0
);
var four_digit_number: Array(Char, 5) = (
  '0' + n / 1000,
  '0' + ((n / 100) % 10),
  '0' + ((n / 10) % 10),
  '0' + (n % 10),
  0
);

We could use the same precedence rule as other languages, and permit examples similar to the above to be written without any parentheses.

However, our rule for precedence is:

For every combination of operators, either it should be reasonable to expect most or all developers who regularly use Carbon to reliably remember the precedence, or there should not be a precedence rule.

It is not clear that a precedence rule that gives a defined meaning to, for example, a + b % c + d would satisfy this rule. Moreover, in mathematics, the "mod" operator generally binds very loosely: in 'n = m + 1 (mod 5)', the '(mod 5)' applies to the entire equality, certainly not to the '1'.

Therefore, we do not give modulo higher precedence than addition, and require parentheses when mixing the two. This decision should be revisited if it is a significant source of friction in practice.

Use a different spelling for modulo

We could use a different spelling for the modulo operation. For example, we could spell it as mod.

Advantages:

  • The percent sign has no relation to modulo or remainder. The only known justification for this particular choice of symbol is that it resembles the '÷' symbol; however, that symbol means division, not remainder.
  • Would free up the % symbol for other uses that may be more prevalent than modulo. However, we would need to be cautious when adding any alternative uses to avoid confusion for people who expect % to mean modulo.

Disadvantages:

  • This would be unfamiliar to developers coming from C++ and other languages with similar operator sets.
  • There is no other common established symbol for this operation. Using a keyword such as mod would break our loose convention of using keywords for non-overloadable operators and operator symbols for overloadable operators. Using a function would substantially increase the verbosity of certain kinds of code.

Future work

Provide separate wrapping operators

We could provide distinct operators with wrapping semantics for types where overflow is normally a programming error. For example, Swift provides overflow operators &+, &-, and &* for this purpose.

This proposal takes no position on whether this would be useful. However, given that in this proposal, unsigned types already have wrapping semantics, the pressure to provide operators for signed types with those semantics is somewhat reduced.

Provide separate operations to detect overflow

It would be useful to provide a way to perform an arithmetic operation if possible, and to provide a separate codepath to handle the case where the arithmetic would have overflowed. For example, we could imagine the Carbon standard library providing a facility such as:

fn AddWithOverflow[N:! BigInt](a: Integer(N), b: Integer(N)) -> Optional(Integer(N));

While this would undoubtedly be useful, this proposal provides no facility for this operation.