Should I use memcmp or chained equal-to operations when both give the same result?

问题

Precondition: Consider such a class or struct T, that for two objects a and b of type T

memcmp(&a, &b, sizeof(T)) == 0

yields the same result as

a.member1 == b.member1 && a.member2 == b.member2 && ...

(memberN is a non-static member variable of T).

Question: When should memcmp be used to compare a and b for equality, and when should the chained ==s be used?

Here's a simple example:

struct vector
{
    int x, y;
};

To overload operator == for vector, there are two possibilities (if they're guaranteed to give the same result):

bool operator==(vector lhs, vector rhs)
{ return lhs.x == rhs.x && lhs.y == rhs.y; }

bool operator==(vector lhs, vector rhs)
{ return memcmp(&lhs, &rhs, sizeof(vector)) == 0; }

Now if a new member were to be added to vector, for example a z component:

If ==s were used to implement operator==, it would have to be modified.
If memcmp was used instead, operator== wouldn't have to be modified at all.

But I think using chained ==s conveys a clearer meaning. Although for a large T with many members memcmp is more tempting. Additionally, is there a performance improvement from using memcmp over ==s? Anything else to consider?

回答1:

Regarding the precondition of memcmp yielding the same result as member-wise comparisons with ==, while this precondition is often fulfilled in practice, it's somewhat brittle.

Changing compilers or compiler options can in theory break that precondition. Of more concern, code maintenance (and 80% of all programming work is maintenance, IIRC) can break it by adding or removing members, making the class polymorphic, adding custom == overloads, etc. And as mentioned in one of the comments, the precondition can hold for static variables while it doesn't hold for automatic variables, and then maintenance work that creates non-static objects can do Bad Things™.

And regarding the question of whether to use memcmp or member-wise == to implement an == operator for the class, first, this is a false dichotomy, for those are not the only options.

For example, it can be less work and more maintainable to use automatic generation of relational operator overloads, in terms of a compare function. The std::string::compare function is an example of such a function.

Secondly, the answer to what implementation to choose depends strongly on what you consider important, e.g.:

should one seek to maximize runtime efficiency, or
should one seek to create clearest code, or
should one seek the most terse, fastest to write code, or
should one seek to make the class most safe to use, or
something else, perhaps?

Generating relational operators.

You may have heard of CRTP, the Curiously Recurring Template Pattern. As I recall it was invented to deal with the requirement of generating relational operator overloads. I may possibly be conflating that with something else, though, but anyway:

template< class Derived >
struct Relops_from_compare
{
    friend
    auto operator!=( const Derived& a, const Derived& b )
        -> bool
    { return compare( a, b ) != 0; }

    friend
    auto operator<( const Derived& a, const Derived& b )
        -> bool
    { return compare( a, b ) < 0; }

    friend
    auto operator<=( const Derived& a, const Derived& b )
        -> bool
    { return compare( a, b ) <= 0; }

    friend
    auto operator==( const Derived& a, const Derived& b )
        -> bool
    { return compare( a, b ) == 0; }

    friend
    auto operator>=( const Derived& a, const Derived& b )
        -> bool
    { return compare( a, b ) >= 0; }

    friend
    auto operator>( const Derived& a, const Derived& b )
        -> bool
    { return compare( a, b ) > 0; }
};

Given the above support, we can investigate the options available for your question.

Implementation A: comparison by subtraction.

This is a class providing a full set of relational operators without using either memcmp or ==:

struct Vector
    : Relops_from_compare< Vector >
{
    int x, y, z;

    // This implementation assumes no overflow occurs.
    friend
    auto compare( const Vector& a, const Vector& b )
        -> int
    {
        if( const auto r = a.x - b.x ) { return r; }
        if( const auto r = a.y - b.y ) { return r; }
        return a.z - b.z;
    }

    Vector( const int _x, const int _y, const int _z )
        : x( _x ), y( _y ), z( _z )
    {}
};

Implementation B: comparison via `memcmp`.

This is the same class implemented using memcmp; I think you'll agree that this code scales better and is simpler:

struct Vector
    : Relops_from_compare< Vector >
{
    int x, y, z;

    // This implementation requires that there is no padding.
    // Also, it doesn't deal with negative numbers for < or >.
    friend
    auto compare( const Vector& a, const Vector& b )
        -> int
    {
        static_assert( sizeof( Vector ) == 3*sizeof( x ), "!" );
        return memcmp( &a, &b, sizeof( Vector ) );
    }

    Vector( const int _x, const int _y, const int _z )
        : x( _x ), y( _y ), z( _z )
    {}
};

Implementation C: comparison member by member.

This is an implementation using member-wise comparisons. It doesn't impose any special requirements or assumptions. But it's more source code.

struct Vector
    : Relops_from_compare< Vector >
{
    int x, y, z;

    friend
    auto compare( const Vector& a, const Vector& b )
        -> int
    {
        if( a.x < b.x ) { return -1; }
        if( a.x > b.x ) { return +1; }
        if( a.y < b.y ) { return -1; }
        if( a.y > b.y ) { return +1; }
        if( a.z < b.z ) { return -1; }
        if( a.z > b.z ) { return +1; }
        return 0;
    }

    Vector( const int _x, const int _y, const int _z )
        : x( _x ), y( _y ), z( _z )
    {}
};

Implementation D: `compare` in terms of relational operators.

This is an implementation sort of reversing the natural order of things, by implementing compare in terms of < and ==, which are provided directly and implemented in terms of std::tuple comparisons (using std::tie).

struct Vector
{
    int x, y, z;

    friend
    auto operator<( const Vector& a, const Vector& b )
        -> bool
    {
        using std::tie;
        return tie( a.x, a.y, a.z ) < tie( b.x, b.y, b.z );
    }

    friend
    auto operator==( const Vector& a, const Vector& b )
        -> bool
    {
        using std::tie;
        return tie( a.x, a.y, a.z ) == tie( b.x, b.y, b.z );
    }

    friend
    auto compare( const Vector& a, const Vector& b )
        -> int
    {
        return (a < b? -1 : a == b? 0 : +1);
    }

    Vector( const int _x, const int _y, const int _z )
        : x( _x ), y( _y ), z( _z )
    {}
};

As given, client code using e.g. > needs a using namespace std::rel_ops;.

Alternatives include adding all other operators to the above (much more code), or using a CRTP operator generation scheme that implements the other operators in terms of < and = (possibly inefficiently).

Implementation E: comparision by manual use of `<` and `==`.

This implementation is the result not applying any abstraction, just banging away at the keyboard and writing directly what the machine should do:

struct Vector
{
    int x, y, z;

    friend
    auto operator<( const Vector& a, const Vector& b )
        -> bool
    {
        return (
            a.x < b.x ||
            a.x == b.x && (
                a.y < b.y ||
                a.y == b.y && (
                    a.z < b.z
                    )
                )
            );
    }

    friend
    auto operator==( const Vector& a, const Vector& b )
        -> bool
    {
        return
            a.x == b.x &&
            a.y == b.y &&
            a.z == b.z;
    }

    friend
    auto compare( const Vector& a, const Vector& b )
        -> int
    {
        return (a < b? -1 : a == b? 0 : +1);
    }

    Vector( const int _x, const int _y, const int _z )
        : x( _x ), y( _y ), z( _z )
    {}
};

What to choose.

Considering the list of possible aspects to value most, like safety, clarity, efficiency, shortness, evaluate each approach above.

Then choose the one that to you is clearly best, or one of the approaches that seem about equally best.

Guidance: For safety you would not want to choose approach A, subtraction, since it relies on an assumption about the values. Note that also option B, memcmp, is unsafe as an implementation for the general case, but can do well for just == and !=. For efficiency you should better MEASURE, with relevant compiler options and environment, and remember Donald Knuth's adage: “premature optimization is the root of all evil” (i.e. spending time on that may be counter-productive).

回答2:

If, as you say, you've chosen types such that the two solutions yield the same results (presumably, then, you have no indirect data and the alignment/padding is all the same), then clearly you can use whichever solution you like.

Things to consider:

Performance: I doubt you'll see much if any difference, but measure it to be sure, if you care;
Safety: Well you say the two solutions are the same for your T, but are they? Are they really? On all systems? Is your memcmp approach portable? Probably not;
Clarity: If your preconditions ever change and you did not adequately comment-describe your memcmp usage, then your program is liable to break — you've therefore made it fragile;
Consistency: Presumably you use == elsewhere; certainly you'll have to do it for every T that doesn't meet your preconditions; unless this is a deliberate optimising specialisation for T, you may consider sticking to a single approach throughout your program;
Ease of use: Of course, it's pretty easy to miss out a member from chained ==, especially if your list of members ever grows.

回答3:

If two solutions are both correct, prefer the more readable one. I'd say that for a C++ programmer, == is more readable than memcmp. I would go so far as to use std::tie instead of chaining:

bool operator==(const vector &lhs, const vector &rhs)
{ return std::tie(lhs.x, lhs.y) == std::tie(rhs.x, rhs.y); }

回答4:

If any only if the structure is POD and if it is safely memcmp comparable (not even all numeric types are...) the result is the same and the question is about readability and performance.

Readability? This is a rather opinion based question I think but I'd prefer operator==.

Performance? operator== is a short-circuit operator. You have more control over your program here because you can reorder the comparison sequence.

Although a == b && c == d and c == d && a == b are equivalent in terms of algorithmic logic (result is the same) they aren't equivalent in terms of produced assembly, "background logic" and possibly performance.

You can influence your program if you can forsee some points.

In example:

If both statements are roughly equally likely to yield false, you'll want to have the cheaper statement first to skip the more complex comparison if possible.
If both statements are roughly equally complex and you know in advance that c == d is more likely to be false than a == b, you should compare c and d first.

It is possible to adjust the comparison sequence in a problem-dependant fashion using operator== whereas memcmp does not give you this kind of freedom.

PS: You would want to measure it but for a small struct with 3 members, MS VS 2013 produces slightly more complex assembly for the memcmp case. I'd expect the operator== solution to have a higher performace (if the impact would be measurable) in this case.

-/edith-

Note: Even POD struct members can have overloaded operator==.

Consider:

#include <iostream>
#include <iomanip>

struct A { int * p; };

bool operator== (A const &a, A const &b) { return *(a.p) == *(b.p); }

struct B { A m; };

bool operator== (B const &a, B const &b) { return a.m == b.m; }

int main()
{
  int a(1), b(1);
  B x, y;
  x.m.p = &a;
  y.m.p = &b;
  std::cout << std::boolalpha;
  std::cout << (memcmp(&x, &y, sizeof(B)) == 0) << "\n";
  std::cout << (x == y) << "\n";
  return 0;
}

Prints

false
true

Even if -in turn- all the members are fundamental types I would prefer operator== and leave it to the compiler to consider optimizing the comparison into whatever assembly it considers to be preferable.

回答5:

== is better, because memcmp compares pure memory data(comparing that way can be wrong in many situations, such as std::string, array-imitiating classes or types that can be equal even if they aren't perfectly identical). Since inside your classes there may be such types, you should always use their own operators instead of comparing raw memory data.

== is also better because it's more readable than some weird-looking function.

回答6:

You imposed a very strong condition that there is no padding (I assume neither between the members of the class, nor inside these members). I presume that you also intended to exclude any "hidden" housekeeping data from the class. In addition the question itself implies that we always compare objects of exactly the same type. Under such strong conditions there's probably no way to come up with a counterexample that would make the memcmp-based comparison to differ from == comparison.

Whether it makes it worth it to use memcmp for performance reasons... well, if you really have a good reason to aggressively optimize some critical piece of code and profiling shows that there's improvement after switching from == to memcmp, then definitely go ahead. But I wouldn't use it a a routine technique for writing comparison operators, even if your class satisfies the requirements.

来源：https://stackoverflow.com/questions/28858359/should-i-use-memcmp-or-chained-equal-to-operations-when-both-give-the-same-resul

标签

c++

memcmp