How can I get a warning when comparing unsigned integers of different size in C and C++?

There does not appear to be a warning option built in to gcc or
clang that does what is requested. However, we can use
clang-query
instead.

Below is a clang-query command that will report comparison of
32-bit and 64-bit integers, on the assumption that int is 32 bits and
long is 64 bits. (More about that below.)

#!/bin/sh

PATH=$HOME/opt/clang+llvm-14.0.0-x86_64-linux-gnu-ubuntu-18.04/bin:$PATH

# In this query, the comments are ignored because clang-query (not the
# shell) recognizes and discards them.
query='m
  binaryOperator(                            # Find a binary operator expression
    anyOf(                                   #  such that any of:
      hasOperatorName("<"),                  #   is operator <, or
      hasOperatorName("<="),                 #   is operator <=, or
      hasOperatorName(">"),                  #   is operator >, or
      hasOperatorName(">="),                 #   is operator >=, or
      hasOperatorName("=="),                 #   is operator ==, or
      hasOperatorName("!=")                  #   is operator !=;
    ),

    hasEitherOperand(                        #  and where either operand
      implicitCastExpr(                      #   is an implicit cast
        has(                                 #    from
          expr(                              #     an expression
            hasType(                         #      whose type
              hasCanonicalType(              #       after resolving typedefs
                anyOf(                       #        is either
                  asString("int"),           #         int or
                  asString("unsigned int")   #         unsigned int,
                )
              )
            )
          )
        ),
        hasImplicitDestinationType(          #    and to a type
          hasCanonicalType(                  #     that after typedefs
            anyOf(                           #      is either
              asString("long"),              #       long or
              asString("unsigned long")      #       unsigned long.
            )
          )
        )
      ).bind("operand")
    )
  )
'

# Run the query on test.c.
clang-query 
  -c="set bind-root false" 
  -c="$query" 
  test.c -- -w

# EOF

When run on the following test.c it reports all of the indicated cases:

// test.c
// Demonstrate reporting comparisons of different-size operands.

#include <stddef.h>          // size_t
#include <stdint.h>          // int32_t, etc.

void test(int32_t i32, int64_t i64, uint32_t u32, uint64_t u64)
{
  i32 < i32;                 // Not reported: same sizes.
  i32 < i64;                 // reported
  i64 < i64;

  u32 < u32;
  u32 < u64;                 // reported
  u64 < u64;

  i32 < u64;                 // reported
  u32 < i64;                 // reported

  i32 <= i64;                // reported

  i64 > i32;                 // reported
  i64 >= i32;                // reported

  i32 == i64;                // reported
  u64 != u32;                // reported

  i32 + i64;                 // Not reported: not a comparison operator.

  ((int64_t)i32) < i64;      // Not reported: explicit cast.

  // Example #1 in question.
  size_t n = 0;
  for (unsigned int i = 0; i < n; i++) {}        // reported
}

// Example #2 in question.
void f(uint64_t n)
{
  for (uint32_t i = 0; i < n; ++i) {             // reported
  }
}

// EOF

Some details about the clang-query command:

• The command passes -w to clang-tidy to suppress other warnings.
  That’s just because I wrote the test in a way that provokes warnings
  about unused values, and is not necessary with normal code.

• It passes set bind-root false so the only reported site is the
  operand of interest rather than also reporting the entire expression.

The unsatisfying aspect of the query is it explicitly lists the source
and destination types. Unfortunately, clang-query does not have a
matcher to, say, report any 32-bit type, so they have to be listed
individually. You might want to add [unsigned] long long on the
destination side. You might also need to remove [unsigned] long if running this
code with compiler options that target an IL32 platform like Windows.

Relatedly, note that clang-query accepts compiler options after
the --, or alternatively in a
compile_commands.json
file.

Finally, I’ll note that I haven’t done any "tuning" of this query on
real code. It is likely to produce a lot of noise.

This does not directly answer the question (provide a warning), but would you consider an alternative which avoids the problem entirely?

    size_t n = // ...

    for (typeof(n) i = 0; i < n; i++) // ...

It now doesn’t matter what type n is, since i will always be the same type as n, you should never have trouble with infinite loops resulting from i being a smaller type or having a smaller range than n.

PVS Studio can issue such warning (and many more), here is almost identical example from their docs:

https://pvs-studio.com/en/docs/warnings/v104/

It is a paid tool, but they give free license to Open Source projects.

I did not find such a warning in Clang-tidy, a free linter tool from LLVM project, but it would be very simple to add a check for comparison of integers of different sizes (a later reply by Scott McPeak with excellent clang-query did most of the work – the remaining part is just plugging this query to clang-tidy). It would be very noisy check though. One can restrict the noise by limiting the check to conditions of loops, that can be done with Clang-tidy too, but a bit more work with AST matchers.

Recent versions of gcc seem to support -Warith-conversion for this purpose:

-Warith-conversion

Do warn about implicit conversions from arithmetic operations even when conversion of the operands to the same type cannot change their values. This affects warnings from -Wconversion, -Wfloat-conversion, and -Wsign-conversion.

void f (char c, int i)
{
    c = c + i; // warns with -Wconversion
    c = c + 1; // only warns with -Warith-conversion
}

Yet it does not work for your example, probably because i < n is not an arithmetic expression. There does not seem to be a variant of this warning for generic binary expressions.

For C++ you may be able to do even better than a compiler warning, assuming n is a compile time constant. This also works for non-gcc compilers. This logic is not available for C code though.

The idea is basically encoding the value information in the variable type instead of the variable value.

template<std::integral T, auto N>
constexpr bool operator<(T value, std::integral_constant<decltype(N), N>)
{
    static_assert(std::is_signed_v<T> == std::is_signed_v<decltype(N)>, "the types have different signs");
    static_assert((std::numeric_limits<T>::max)() >= N, "the maximum of type T is smaller than N");
    return value < N;
}

// todo: overload with swapped operator parameter types

int main()
{
    constexpr std::integral_constant<size_t, 500'000'000> n; // go with 5'000'000'000, and you'll get a compiler error 

    for (unsigned int i = 0; i < n; i++)
    {

    }
}

If the value is not a compile time constant, you could still create a wrapper template type for the integer and overload the < operator for comparisons with integral values, adding the static_asserts into the body of this operator.

template<std::integral T>
class IntWrapper
{
    T m_value;
public:
    constexpr IntWrapper(T value)
        : m_value(value)
    {}

    template<std::integral U>
    friend constexpr bool operator<(U o1, IntWrapper o2)
    {
        static_assert(std::is_signed_v<U> == std::is_signed_v<T>, "types have different signedness");
        static_assert((std::numeric_limits<U>::max)() >= (std::numeric_limits<T>::max)(),
            "the comparison may never yield false because of the maxima of the types involved");

        return o1 < o2.m_value;
    }
};

void f(IntWrapper<uint64_t> n)
{
    for (uint32_t i = 0; i < n; ++i) {
    }
}

Note that the necessity of changing the type for one of the operands of the comparison operator can be both a benefit and a drawback: it requires you to modify the code, but it also allows you to apply check on a per-variable basis…

Follow a coding standard which stops it at source

Most coding standards for embedded software prohibit the use of "int", for exactly the reasons you say. The C standard requires the existence of equivalent data types of fixed lengths (8, 16 and 32 bits have been mandatory for years; 64-bit is newer but still supported basically everywhere). It is good practise to use them everywhere, but in safety-related software it is normally mandatory. Many tools exist for popular coding standards such as MISRA-C which will catch these issues for you.

Your example apparently seems obscure to many commenters above, because it needs 2^32 iterations to overflow. What many modern coders forget is that int can also be 16-bit, and this made overflows much easier in the past. The Ariane-5 disaster was caused by a 64-bit value overflowing when truncated to 16-bit. The Therac-25 disaster was also partly caused by an integer overflow which cleared the error status.

Examples do exist in 32-bit too, though. Windows 95 and 98 famously crashed after 49.7 days caused by overflow of a 32-bit millisecond timer. And in 15 years time we have the Y2038 problem to look forward to. Most modern systems are Y2038-ready, but it’s likely we’ll get there and have some surprises!

All of these form part of the institutional history of software engineering as an engineering discipline, in the same way as the Tay Bridge and Tacoma Narrows Bridge form part of the institutional history of civil engineering. I’m somewhat shocked to see someone above saying they’d never heard of this in 40 years of C coding. It is certainly possible to be just a coder and be unaware of this, just as it is possible to be just a builder and be unaware of civil engineering principles. An engineer must be aware of the deeper principles behind their designs though. This issue, as trivial as it may seem, is one thing which clearly differentiates expertise levels between software engineers and mere coders. Given Ariane-5 and Therac-25, I make no apology for being judgemental about this.