Lennart’s Blog

There are many ways in which programming languages may support runtime error handling. Languages like Python, C++ and Java support exceptions, but this feature has become less popular over the last few years, as it leads to problems with freeing up resources. Rust and Zig require you to handle errors at each call level and they allow you to simply return any error to the calling function with minimal additional code.

Some early BASIC interpreters would simply abort when they detected a runtime error. Turbo Pascal would do the same. It came with a mini-spreadsheet. While it did check for division by zero and avoided that error, a floating point overflow error would abort the program without any opportunity to save your data that you entered. Even GW-BASIC had an ON ERROR GOTO statement, which allowed you to keep the program running if an error occurred.

Error Detection

There are two ways a runtime error can be detected:

• A hardware trap. For example, most modern CPUs would trap on accessing memory addresses out of range and on division by zero. Some could even trap on integer overflow. On Unix systems, these hardware traps would cause your program to receive a ‘signal’. Your program would be aborted by default, but you can handle the signal by calling a handling function when it occurs.
• A check in software. Most I/O functions return a result code. For example if you try to open a non-existing file, your OS detects in software that the file does not exists. The open system call returns an error code. The C library function fopen will return a NULL pointer instead of a pointer to a valid FILE data structure. A hardware trap is never involved in this case.

Early Strategies

Many versions of BASIC had an ON ERROR GOTO statement. The error handler at that line would then try to fix the error condition and continue the program with RESUME or RESUME NEXT. The latter variant would skip the statement that caused the error, instead of retrying it. Instead of restarting, you could try to end the program more gracefully, possibly allowing the user to save unsaved data first.

C and Unix use the signal functions, that allow you to handle errors like out-of-bounds memory access and division by zero. C also has the infamous setjmp/longjmp functions. These functions implement a very crude way to handle exceptions. The function setjmp stores the contents of registers in a jmp_buf structure, including the current value of the stack pointer and the return address the setjmp function would return to. After this, setjmp returns the value 0. The function longjmp would restore the registers from the jmp_buf data structure and jump back to the point where setjmp would have returned to, but now returning a non-zero value. Every function called after setjmp could later call longjmp and return to the location of setjmp. The longjmp function could be called by a signal handler to throw the program back to a defined state, from which it could safely recover.

I/O errors in C would typically be handled by error results from function calls and each function had to propagate the error back to its caller.

There is also the possibility not to disrupt the control flow in case of an error, but to store a special result value like NaN or Infinity. This is typically used for floating point computations. The program is allowed to continue to run and at the end, some results are invalid, but others may be valid and useful.

Exceptions

Some languages implement exceptions, which act a bit like the setjmp/longjmp functions in C, but now there is a nicer language syntax around it. Plus you can have the exception handled at multiple levels. For example, if function A catches an exception and it calls function B, which also catches that exception, then B would handle the exception when it occurs. After B returns, the exception would again be handled by A.

In a simplistic implementation, when an exception occurs, the registers like the stack pointer are restored to the point where the exception was last caught. Any memory allocated at intermediate call levels would not be freed and would likely be leaked, And I’m not even mentioning other resources like open files, network connections and windows in a GUI. A bit more sophisticated implementation would visit each stack frame in turn and free all memory that was known to need freeing when that function returned. Garbage collected languages would not free the memory anyway and would leave it to the garbage collector at a later time. But even they would typically not release other resources.

Handling an exception properly, under all circumstances, is very hard. It is therefore a much less popular feature than a few decades ago. Exceptions are still great when you want to handle error conditions that are typically detected by a hardware trap, such as integer division by zero. Or for integer overflow in general.

It may be feasible to add error handling code to each I/O function call, but it would be very unwieldy to add this to every arithmetic expression that might overflow. Exceptions may still be a good solution for these.

Error Propagation

Rust and Zig use error propagation instead of exceptions. There are special result types that are logically the union of an error result and a regular result. So a function can return either an error result or a regular result. Function returns are checked at every call, but there is easy syntax to just return from the function early and just propagate the error result to the caller. Function A calls function B (that can return an error result), and function B opens a file. With some convenient syntax, function B can just return the error to A when the file open call returns an error result. In these languages, each function always returns to its caller and there is no magic stack unwinding.

Zig (and C3 and Odin) have the defer statement to specify that some statements have to be executed before returning from the function (or leaving another scope), regardless of how the function returns. These defer statements will be executed even in case of error propagation. These defer statements are typically used to free up resources that were allocated during that function.

Conclusion

Error handling is complex and no matter what strategy you use for it, error conditions are always the least tested aspects of a program. Whether it is acceptable to abort the program when a serious error occurs, depends on the situation. Embedded control systems often do not have the option to abort and they must keep the system under control under all circumstances. Your fly by wire system is never allowed to drop the plane from the sky. Of course a single program could abort in this case (and probably be restarted), as long as a fallback system is in place. Coping with error situations in cars, planes and nuclear power plants is an engineering discipline of its own.

Programs like text editors or spreadsheets, that allow a user to enter large amounts of data, must be designed to allow that data to be saved. Losing a few hours worth of spreadsheet data may not be as bad as a crashed plane, but it is certainly rude to lose that data without putting up a fight to save it.

Ignoring an error is never a good idea. Checking inputs to ensure that buffer overflows cannot happen and integers cannot overflow, is generally a good idea. It is generally better to refuse to complete a transaction because it fails range checks (even though the values are valid after all) than to complete it with a silent overflow and a bogus result.

In the early days of PC compatible computers, MS-DOS came with GW-BASIC, a BASIC interpreter with many built-in commands for graphics, event handling and file access. .We had commands like:

1000 ON KEY(1) GOSUB 100: REM call a subroutine when F1 key is pressed
1010 LINE (100,200)-(300,400): REM DRAW a line on the screen
1020 OPEN "myfile.txt" FOR INPUT AS #1 : REM speaks for itself
1030 PRINT USING "####.##";A : REM print number with specific formatting
1040 PRINT #1,"The numbers are:";A,B: REM print to a file, different parameter separators.

Each of these command had its own special syntax. Apart from ON KEY we also had ON PEN to specify an action for the light pen, a truly forgotten device, that nevertheless got its own keyword in the syntax. The parentheses around the end points in the LINE command were part of the syntax and the “-” did not mean subtraction, but it was there suggest from..to. The PRINT command had an optional USING part, but also an optional ‘#’ part to specify a file and different separators between parameters: “;” to specify that the next parameter had to be printed immediately after the previous one and “,’ to specify that it had to be printed at the next “tab stop”.

At the same time, GW-BASIC did not allow you to create new procedures, only the humble GOSUB, which is in some respects below the level of assembly language. QBASIC did add named subroutines, but calls to them did not look anything like the built-in commands.

COBOL also has a very large number of commands built into the language, each with its own syntax.

These languages represent one end of the extensibility spectrum. They came with many features included, but what you got, was all you would ever have.

Suffice it to say that no modern programming language comes with built-in commands to draw lines and circles or to bind keypress events to functions. If you need this type of functionality, you can load a library for it.

Pascal

Pascal is a nice step up from the horrors of GW-BASIC. It lets you define your own functions and procedures and even your own data types. But if you look closer to Pascal, you will find that built-in procedures can do many things that user-defined procedures can’t:

• The read and write procedures can take parameters of many different types. The same is true for some other built-in procedures, like reset, rewrite, new and dispose.
• The read and write procedures can take a variable number of parameters.
• The write procedure has special formatting syntax with colon characters to describe how a number should be printed. like write(a:12:3);

Pascal clearly has its I/O functions baked into the language, using many privileged procedures.

Small versus Big Languages

Almost any modern programming languages has its I/O functionality outside the core language. Any I/O functionality in the standard library, could also be implemented in a library you could implement in the language itself. Programs in those languages can run on embedded systems that do not have terminal or file I/O, but very different I/O functionality instead.

C is a small language, but flexible enough to write its own I/O library in C itself. C. Functions from stdio.h, such as fopen and fread are ordinary functions that you can just write in C. Even the printf function, with its variable number and types of parameters, can be written in C, though it requires some trickery.. Better yet, heap allocation functions like malloc and free can just be written in C.

If you write code in C and don’t call any functions outside of your program, the resulting program will include very few external functions when it is linked. If you are on a RISC system without integer division, you get an integer division function linked with your program and you may get memcpy or its equivalent when you assign struct variables to one another. Software floating point functions are another thing you might get if your CPU has no hardware floating point and your program uses floating point. You can use C to write operating system kernels, boot loaders, embedded firmware, real-time applications and more.

Zig, C3 and Odin are similarly small languages that don’t force you to link with an excessively sized runtime library..

On the other hand, if your language has a garbage collector, your compiled programs will depend on a large piece of external code, the garbage collator itself. It takes away the control that you need when you write an operating system kernel or a real time embedded application.

C does not have all the extension features that exist in programming languages. It has no operator overloading, no generics, no true modules and the preprocessor macros are very crude by today’s standards. A language like C3 has more extensibility features in that respect, even though this language is still not very big.

Operator overloading makes a language more complex, but at the same time, it allows you to add data types like matrices, vectors, complex numbers and arbitrary size integers later, complete with algebraic operators, so we do not have to choose between including them in the base language (like Odin does) or not having the convenience of algebraic expressions with these types at all.

Ada, C++, Java, Go, Rust and D are much larger languages than for example C. Some of these have a garbage collector, Rust achieves memory safety using extensive compiler checks. Sine if these languages support exceptions, so an errors can cause the program to unwind the stack through many call levels and then return to the level at which the exception is caught.

While perl has hash tables and array lists as built-in types, most modern languages have the extensibility features to define them in their standard libraries, without them being necessary in the core language. Even a small language can have these features at its disposal, if it has enough extensibility features to add them.

The Upper End of the Spectrum

In terms of extensibility, LISP is probably the best language. While some languages brag about object-oriented features, LISP has always had the ability to be extended in a way to add them.

LISP has macros that can rewrite programs in an arbitrary way. You can add new control structures, like more advanced loops or switch statements. Very few programming languages have that ability.

LISP is one of very few languages that has no infix expressions. That makes its parser very simple. Once parsed, the LISP program is just a bunch of nested lists, a direct representation of the parse tree.

A honourable mention goes to FORTH, Like LISP, it sacrifices infix expressions, but unlike LISP, it does not parse an expression into a nested list, instead it directly executes each token and lets the stack do all the work. While LISP uses prefix notation with parentheses to indicate nesting, FORTH uses postfix notation without parentheses. FORTH can also build new control structures. FORTH taps into some of the unique extensibility features that LISP also has, but its implementation is considerably simpler. It could work on machines with very little memory and it would execute much faster than interpreted BASIC or LISP.. I should devote s separate post to this little language.

The C preprocessor may be the worst macro system for any programming language. It is complex enough to seriously obfuscate programs with it, but it is not Turing complete, so you cannot meaningfully implement loops in it. The preprocessor supports conditional compilation using #if and #ifdef, but these conditionals cannot appear in the expansion of a macro. As the preprocessor is a separate pass from the C compiler itself. preprocessor conditionals cannot contain values that are only known to the compiler (and not to the preprocessor), such as the size of a data item, enum values or the value of an object declared ‘const’. So for instance you cannot do something like this:

#if sizeof(struct1) != sizeof(struct2)
#error "struct1 must be the same size as struct2"
#endif

Some tricks exist to do a similar check at compile time, like:

int unused_array[-(sizeof(struct1) != sizeof(struct2)];

This declares an unused array of size 0 (permitted) i n case the sizes are equal, but of size -1 (compile time error) if the sizes are not equal.

Header files are not real “module interface” files, but hey are textually included by the preprocessor. We need the “include guard” pattern to prevent them from being include more than once, like this:

#ifndef MYHEADER_H_
#define MYHEADER_H_
...

#endif

If you define a macro expression that looks like a function call, you need to put parentheses around any parameters and around the expression itself. If the macro expands to a sequence of statements, you have to wrap in in a do .. while(0) construct. This is annoying and once in a thousand times you make a mistakes and another nasty bug is born.

If macros can’t do what you want, you could use a more powerful macro processor, such as m4, or you could use a dedicated program to generate the C files for you. Think for example of the ‘yacc’ program to generate parsers. Apparently there is no standard way to convert a binary file into a C header file containing an initialised constant array with all the bytes of that file. If have written scripts to do that just too often. If these arrays are large, the memory consumption of the C compiler is usually terrible.

Generics

In C++ you have templates. These are useful for generating data structures that may contain different data types as their payload. So you can instantiate a hash table that contains integers or one that contains “struct foo” items. In fact the std::map template is generic over two types: one for the keys and one for the stored values. You can declare a map like this:

std::map<std::string, u32> mytable;

The parameters are in angle brackets and they are all handled at compile time. Templates perform some tasks that macros could do, but they are part of the compiler itself, not of some separate preprocessor. In C++, it is pretty easy to make use of pre-defined templates, but it is usually very hard to create your own templates.

Other languages, like Ada, also have generics. The Ada programming language has no macro preprocessor, but it does have powerful generics.

Other Languages

Many programming languages do not have any macro system or preprocessor at all, not even generics. Languages like Pascal, FORTRAN and BASIC fall into this category.

Python has no micros or generics either, but as it is dynamically typed, you do not need generics to implement generic container classes. You can use operators like ‘+’ and ‘<‘ in a function and they will automatically just work for any data types that support these operators, making the function a generic one.

Systems like LISP support macros that allow you to build arbitrary LISP programs under program control. This way you can for example add totally new control structures to the language.

Some programming systems have a clear distinction between “compile time” and “execution time”. When the C compiler runs, it translates the C program to assembly language, but it cannot itself run C code. The compiler can run on a system different from where the compiled code is to run. You can run a C compiler on an x86 PC and let it generate code that will run on a Raspberry Pi Pico (ARM Thumb 2 core). With interpreted languages, like Python, you can run Python code at load time. When you import a Python module, it will usually add new functions and/or classes to your program, but it could also run code immediately when it loads. It could build a complex data structure when it loads. The Python interpreter can always run Python code.

FORTH has an interesting combination of features. When it compiles new functions (colon definitions), it converts FORTH source code into threaded code, which can later be interpreted. When not compiling new functions, it just runs your FORTH code. In FORTH, you can extend the compiler, so you can define new control structures. This gives you the same flexibility as LISP macros. Extending the compiler is done by marking some FORTH functions IMMEDIATE, so they will execute, even while a new function is being compiled. By default, the compiler would just add the function’s address to the threaded code.

Rust has macros. For example println! is a macro. Using those macros is fairly easy, but defining new ones is a hard job, requiring you to learn much about compiler internals.

Assemblers have had macros for a long time. A macro could expand to a sequence of instructions and depending on the exact assembler you were using, you could give the macro parameters (for example specifying a register to be used on some of the instructions), the macro expansion could contain conditional assembly or even loops and the macro expansion could contain local labels, so the labels inside different instances of the macro would not conflict with one another.

Zig

Zig is the latest development in macro technology.,First of all, Zig is a true compiler. It converts your source code to compiled machine code, to be executed later, possibly on a completely different machine. Therefore it has clearly distinct compile time and run time phases. But even at compile time, the compiler can run Zig code in your source files. Zig replaces any dedicated macro language that programming languages may have. The build script is Zig code, generics are implemented using Zig code. This can work because data types are values that can be parameters and returned results of compile time Zig functions. Zig code can generate arbitrarily complex data tables at compile time.

One example of compile time behaviour in Zig: the format strings for formatted printing are parsed at compile time and calls to specific output functions are generated, depending on the desired format and the types of the parameters passed to that function.

In the beginning we had 16-bit integers. £327.67 was still a lot of money and the total amount in a single sale of your shop was unlikely to be more than that. Until it was and you got a negative amount on your sales slip. When the amount was represented in decimal, you still found yourself in for surprises when amounts started to exceed £999.99 or whatever you thought the maximum would be.

How computer systems handle integer overflow, depends very much on the programming language used. Integer BASIC versions usually aborted the program with an overflow error. That sucked, but at least you knew there was something wrong. Many compiled languages on the other hand, silently ignored the error and just wrapped the number around. Compiled C typically ignores the error. Most CPUs do have an overflow flag, so the machine has the possibility to check for overflow and throw an error in this case. But the designers of RISC-V, in all their wisdom, decided to leave out the overflow status flag because C was the only language worth dealing with and nobody ever checked for integer overflow. Of course you can still do it on RISC-V, but it takes multiple instructions.

In some cases, the wrap-around behaviour of integer overflow is actually desired, especially in hash functions or when emulating the behaviour of real hardware. This mostly applies to unsigned integers. The C standard specifies wrap-around behaviour for unsigned integers and (until recently) specified signed integer overflow as undefined behaviour.

Possible Handling

There are many ways to deal with integer overflow:

• Ignore it and wrap around, which is the default behaviour in C.
• Set an integer status bit, which could be checked after every addition and subtraction. Most CPUs, such as x86 and ARM have this.
• Trap the overflow in hardware. Some CPUs, like MIPS, can do this.
• Have a ‘sticky’ overflow status bit. It could be cleared before a computation and could be checked at the end of a computation. If any operation caused overflow, the bit would be set. This is what the exception flags in floating point hardware actually do.
• Use the maximum negative value 0x8000, 0x80000000 or 0x8000000000000000 as a special overflow value. If any of the inputs of an addition or subtraction have this value, or if an overflow occurs, the result is this value. It would be similar to the NaN value for floating point operations.
• Redo the operation with floating point (some versions of BASIC) or arbitrary size integers (Python).
• Throw an exception on integer overflow.
• Panic (abort the program) on integer overflow
• Saturate, i.e. let the result be either the maximum positive value (0x7ffffffff) or the maximum negative value (0x80000000). Some DSPs or SIMD instructions in some general purpose CPUs can do this.

Consequences

If you ignore integer overflow and just wrap around, integer addition retains its associative properties (a+b)+c is equal to a+(b+c). As soon as you start to saturate or trap the overflow condition, the associative property no longer holds. (MAXINT+1)-1 is not equal to MAXINT+(1-1). In the former case the result will be MAXINT-1 (in the case of saturation) or an error condition. In the latter case the result will be MAXINT. Therefore, in some cases, you are saved by not detecting overflow, but this is not something you should be relying on.

One early observation was that (A-B) < 0 is not a reliable way to check A < B. A-B could overflow and you would get the wrong result. For 16-bit signed integers, 30000 – -10000 overflows and gives a negative number, wile of course 30000 > -10000. Most CPUs have dedicated compare and conditional branch instructions that give you the correct result in all cases. The 8080 did not have an overflow flag at all and the 6502 and Z80 had an overflow flag, but no dedicated Branch-if-less-than instruction. You had to do multiple conditional branches to properly test for less-than of signed quantities. The 6809 on the other hand did have a BLT (Branch if less than) instruction, as did the 8086, 68000, ARM and anything more modern.

Computing the average of two integers can fail due to overflow, if done in a straightforward way. There are tricks to make it work under all circumstances. For unsigned integers we have the expression (a&b) + ((a^b)>>1)

Even if the computation itself does not overflow (as it is done in 32-bit on most modern machines), storing the result in a shorter variable (8 or 16 bits) could still overflow. Compilers may or may not add checks for this. For C they typically don’t,

In some cases you have to be very careful to cast the operands to a wider type before doing the operations. For example:

uint32_t a=1000000;
uint32_t b=2000000;
uint64_t c = (uint64_t)a * (uint64_t)b;

If we forget to cast, we do a 32×32 bit multiplication that overflows.

The Zig programming language is very picky about overflow checking. It even checks unsigned operations. If you do want wrap-around behaviour (for example for hash functions), you have to explicitly use operations that wrap around, like ‘+%’ and ‘-%’.

If a multiplication is part of the size computation for an allocation, we could allocate a way too small buffer if the multiplication overflows, which in turn could lead to memory corruption. It is therefore important to detect the overflow in this case. (but in the case of C, this is often not possible).

Avoiding Overflow

Integer overflow can often be avoided by choosing the data types for your computation sufficiently wide and to range-check your inputs.

For example, in a webshop (for consumers), it makes no sense to accept orders of more than 100 of the same item. If you allow the user to type any integer, even 1 billion, and do not check this is within reasonable limits, you could cause an overflow when multiplying it by the item price. If both the item price and the number of items are within a set range, you can prove that the multiplication cannot overflow.

With simple formulas, it is often feasible to assure that no overflow occurs whatever the input values, when the inputs are all within the acceptable range. If you also range-check intermediate values (like the running total of your order), you can prevent overflow from happening and possibly error out with a sensible message, such as “Orders above £10000 are not accepted”.

Dealing with Overflow

If you cannot avoid overflow (or stumble across it despite best efforts), you have to deal with it. It depends entirely on the situation what you should do. If your job is controlling an aircraft, it’s never acceptable to just abort and let the plane drop from the sky. It your job is computing this year’s tax statement, it might be good to abort, so somebody could recompile the program with larger integer types and redo the calculations. If your program contains lots of unsaved user data, you should at least try to save that data before aborting. A spreadsheet may convert the integer values to floating point values and redo the calculation that way.

Maybe the database format needs to be changed and all existing databases converted, because some fixed width fields are now too small to hold values that can occur after all. You should see these things coming in advance so you can plan the conversion ahead. For example the Unix timestamp overflow in 2038.

As with any error handling, there is never one good way that works in all cases.

Last time I talked about garbage collection, a feature that some programming languages have and other don’t. Garbage collection adds runtime overhead, but makes memory management safer. There is also a third way, used by Rust, in which the compiler makes checks at compile time and guarantees memory safety that way, but at the cost of very complex restrictions, hard to grasp by new programmers.

This time I will take about operator overloading. Some people like the feature, but others insist on it being left out of their favourite programming language because it adds unnecessary complexity to the language.

Infix operators

In the mid 1950s, FORTRAN introduced infix operators, complete with precedence rules ( in A+B*C the multiplication between B and C is performed first, before adding the result to A) and parentheses (In A*(B+C), the addition between B and C is performed first, before multiplying the result by A). This infix notation closely follows algebraic formulas used in mathematics for centuries.

Nearly all programming languages adopted infix expressions. Notable exceptions are FORTH (that uses Reverse Polish Notation, like B C + A *) or LISP (that puts the operator first and uses lots of parentheses, like (* A (+ B C))).

Infix expressions were a step up from assembly, where we used to write things like:

ADD T, B, C
MUL T, T, A

Operands in infix expressions can be variables, constants, array elements and function calls. Infix expressions are typically used on the right hand side of an assignment. An assignment statement in FORTRAN may look like:

A = A + COS(PHI) + 2*SIN(B(I))

Where B is an array (indexed by I), A and PHI are variables and COS and SIN are functions.

Operator Overloading

Languages like FORTRAN, BASIC, Pascal and C can use operators on a fixed number of data types, always including integers and real numbers. In Pascal we can use + and * operators for sets, in Basic we can use + to concatenate strings and in C we can use + to add to pointers. Nearly all languages have relational operations like < and > and also Boolean operators like AND and OR. But the types and semantics, as well as the data types that can be used, are hardcoded into the programming language. For example, it is not possible to define a COMPLEX number type and define new infix operators for them, equivalent to what the algebraic operators are for complex numbers in mathematics (some of these languages have COMPLEX data types, but these are then also hardcoded into the language).

Algol-68 was one of the first programming languages to have operator overloading. It could redefine existing operators for new types, it allowed you to add completely new infix operators and you could specify the priorities of any of these operators. This might have been too much flexibility and room for abuse.

Most large programming languages, like Python, C++, Ada and Rust do allow you to overload infix operators, but none of them allows to to add completely new infix operators or to redefine operator priorities.

Benefits

In mathematics, algebraic expressions are not just used with numbers, but for example also with vectors and matrices. Especially for those with knowledge of the problem domain, infix expressions are very readable. A single infix expression with vectors and matrices, can replace a rat’s nest of hard to read function calls. An expression with vectors and matrices is certainly more readable than a loop over the separate array values or two nested loops for matrix multiplication.

Types that are suitable for operator overloading:

• Vectors and matrices, with scalar multiplication, inner product, matrix multiplication and matrix-vector multiplication.
• New numeric types, like arbitrary size integers or complex numbers.
• Sets
• Lists and strings for concatenation (using + operator) and sometimes the * operator for replicating n times.
• Arbitrary algebraic types as mathematicians define them, as long as they can be represented in a computer and operations can be implemented.

Drawbacks

To an unsuspecting reader, an expression with infix operators may look like it’s just adding and multiplying integers (or maybe floating point numbers), while in reality it is doing arbitrarily complex operations on arbitrarily complex data structures.

Sometimes, operators are overloaded to do operations that are totally unrelated to their original meaning, for example the << operation in C++, that was originally “left shift”, but it is used to output something on an output stream.

In some languages, such as C++, you can also overload operations like assignment or array subscripting. This is all fine, as long as you do this to implement sane assignment or subscripting semantics for them, but it gets pretty ugly if you implement totally unrelated functionality for these operators.

In some languages, such as C++, you can also overload operations like assignment or array subscripting. This is all fine, as long as you do this to implement sane assignment or subscripting semantics for them, but it gets pretty ugly if you implement totally unrelated functionality for these operators.

Even in cases where it does make sense to overload, there are some drawbacks:

• It is not always clear what an overloaded operator will do. For example: is the ‘*’ operator between matrices doing matrix multiplication or component-wise multiplication instead?
• Overloaded operators and the function calls that the compiler invokes for these operations, may in many cases not be the optimal way to perform a task. Dedicated function calls for a specific task may run more efficiently.

What Some Languages Do

Three “small” languages that aim to replace C, chose different solutions:

• Zig does not have operator overloading. It has some infix operations on vectors though.
• Odin does not have operator overloading, but it has many operations on vectors, matrices, complex numbers and even quaternions. AFAIK, this is the only language that has quaternions as a built-in type.
• C3 does allow operator overloading. The documentation contains a plea to use it only in useful ways, but that might not help too much.