Is * calling *= (or *= calling *) slower than writing seperate functions (for math library)?

Question

I have some vector classes where the arithmetic functions looks like this:

template<typename T, typename U>
auto operator*(const Vector3<T>& lhs, const Vector3<U>& rhs)
{
    return Vector3<decltype(lhs.x*rhs.x)>(
        lhs.x + rhs.x,
        lhs.y + rhs.y,
        lhs.z + rhs.z
        );
}

template<typename T, typename U>
Vector3<T>& operator*=(Vector3<T>& lhs, const Vector3<U>& rhs)
{
    lhs.x *= rhs.x;
    lhs.y *= rhs.y;
    lhs.z *= rhs.z;

    return lhs;
}

I want to do a bit of cleanup to remove duplicated code. Basically, I want to convert all the operator* functions to call operator*= functions like this:

template<typename T, typename U>
auto operator*(const Vector3<T>& lhs, const Vector3<U>& rhs)
{
    Vector3<decltype(lhs.x*rhs.x)> result = lhs;
    result *= rhs;
    return result;
}

But I'm concerned as to whether it will incur any additional overhead from the extra function call.

Is it a good idea? Bad idea?

Answer 1

In practice, no additional overhead will be incurred. In C++, small functions are usually inlined by the compiler as an optimization, so the resulting assembly will have all the operations at the callsite— the functions won't call each other, since the functions won't exist in the final code, only the mathematical operations.

Depending on the compiler, you may see one of these functions calling the other with no or low optimization (as with debug builds). At higher optimizations level though (release builds), they'll be optimized down to just the math.

If you'd still like to be pedantic about it (say you're creating a library), adding the inline keyword to operator*() (and similar wrapper functions) may hint your compiler to perform the inline, or using compiler-specific flags/syntax like: -finline-small-functions, -finline-functions, -findirect-inlining, __attribute__((always_inline)) (credit to @Stephane Hockenhull's helpful info in the comments).  Personally, I tend to follow what the framework/libs I'm using do— if I'm using GLKit's math library, I'll just use the GLK_INLINE macro that it provides too.

---

Double-checking using Clang (Xcode 7.2's Apple LLVM version 7.0.2 / clang-700.1.81), the following main() function (in combination with your functions and a naive Vector3<T> implementation):

int main(int argc, const char * argv[])
{
    Vector3<int> a = { 1, 2, 3 };
    Vector3<int> b;
    scanf("%d", &b.x);
    scanf("%d", &b.y);
    scanf("%d", &b.z);

    Vector3<int> c = a * b;

    printf("%d, %d, %d\n", c.x, c.y, c.z);

    return 0;
}

compiles to this assembly using optimization flag -O0:

    .section    __TEXT,__text,regular,pure_instructions
    .globl  _main
    .align  4, 0x90
_main:                                  ## @main
Lfunc_begin0:
    .loc    6 30 0                  ## main.cpp:30:0
    .cfi_startproc
## BB#0:
    pushq   %rbp
Ltmp0:
    .cfi_def_cfa_offset 16
Ltmp1:
    .cfi_offset %rbp, -16
    movq    %rsp, %rbp
Ltmp2:
    .cfi_def_cfa_register %rbp
    subq    $128, %rsp
    leaq    L_.str1(%rip), %rax
    ##DEBUG_VALUE: main:argc <- undef
    ##DEBUG_VALUE: main:argv <- undef
    movl    $0, -4(%rbp)
    movl    %edi, -8(%rbp)
    movq    %rsi, -16(%rbp)
    .loc    6 31 15 prologue_end    ## main.cpp:31:15
Ltmp3:
    movl    l__ZZ4mainE1a+8(%rip), %edi
    movl    %edi, -24(%rbp)
    movq    l__ZZ4mainE1a(%rip), %rsi
    movq    %rsi, -32(%rbp)
    .loc    6 33 2                  ## main.cpp:33:2
    leaq    L_.str(%rip), %rsi
    xorl    %edi, %edi
    movb    %dil, %cl
    leaq    -48(%rbp), %rdx
    movq    %rsi, %rdi
    movq    %rsi, -88(%rbp)         ## 8-byte Spill
    movq    %rdx, %rsi
    movq    %rax, -96(%rbp)         ## 8-byte Spill
    movb    %cl, %al
    movb    %cl, -97(%rbp)          ## 1-byte Spill
    movq    %rdx, -112(%rbp)        ## 8-byte Spill
    callq   _scanf
    .loc    6 34 17                 ## main.cpp:34:17
    leaq    -44(%rbp), %rsi
    .loc    6 34 2 is_stmt 0        ## main.cpp:34:2
    movq    -88(%rbp), %rdi         ## 8-byte Reload
    movb    -97(%rbp), %cl          ## 1-byte Reload
    movl    %eax, -116(%rbp)        ## 4-byte Spill
    movb    %cl, %al
    callq   _scanf
    .loc    6 35 17 is_stmt 1       ## main.cpp:35:17
    leaq    -40(%rbp), %rsi
    .loc    6 35 2 is_stmt 0        ## main.cpp:35:2
    movq    -88(%rbp), %rdi         ## 8-byte Reload
    movb    -97(%rbp), %cl          ## 1-byte Reload
    movl    %eax, -120(%rbp)        ## 4-byte Spill
    movb    %cl, %al
    callq   _scanf
    leaq    -32(%rbp), %rdi
    .loc    6 37 21 is_stmt 1       ## main.cpp:37:21
    movq    -112(%rbp), %rsi        ## 8-byte Reload
    movl    %eax, -124(%rbp)        ## 4-byte Spill
    callq   __ZmlIiiE7Vector3IDTmldtfp_1xdtfp0_1xEERKS0_IT_ERKS0_IT0_E
    movl    %edx, -72(%rbp)
    movq    %rax, -80(%rbp)
    movq    -80(%rbp), %rax
    movq    %rax, -64(%rbp)
    movl    -72(%rbp), %edx
    movl    %edx, -56(%rbp)
    .loc    6 39 27                 ## main.cpp:39:27
    movl    -64(%rbp), %esi
    .loc    6 39 32 is_stmt 0       ## main.cpp:39:32
    movl    -60(%rbp), %edx
    .loc    6 39 37                 ## main.cpp:39:37
    movl    -56(%rbp), %ecx
    .loc    6 39 2                  ## main.cpp:39:2
    movq    -96(%rbp), %rdi         ## 8-byte Reload
    movb    $0, %al
    callq   _printf
    xorl    %ecx, %ecx
    .loc    6 41 5 is_stmt 1        ## main.cpp:41:5
    movl    %eax, -128(%rbp)        ## 4-byte Spill
    movl    %ecx, %eax
    addq    $128, %rsp
    popq    %rbp
    retq
Ltmp4:
Lfunc_end0:
    .cfi_endproc

In the above, __ZmlIiiE7Vector3IDTmldtfp_1xdtfp0_1xEERKS0_IT_ERKS0_IT0_E is your operator*() function and ends up callqing another __…Vector3… function.  It amounts to quite a lot of assembly.  Compiling with -O1 is almost the same, still calling out to __…Vector3… functions.

However, when we bump it up to -O2, the callqs to __…Vector3… disappear, replaced with a imull instruction (the * a.z ≈ * 3), an addl instruction (the * a.y ≈ * 2), and just using the b.x value straight-up (because * a.x ≈ * 1).

    .section    __TEXT,__text,regular,pure_instructions
    .globl  _main
    .align  4, 0x90
_main:                                  ## @main
Lfunc_begin0:
    .loc    6 30 0                  ## main.cpp:30:0
    .cfi_startproc
## BB#0:
    pushq   %rbp
Ltmp0:
    .cfi_def_cfa_offset 16
Ltmp1:
    .cfi_offset %rbp, -16
    movq    %rsp, %rbp
Ltmp2:
    .cfi_def_cfa_register %rbp
    .loc    6 33 2 prologue_end     ## main.cpp:33:2
Ltmp3:
    pushq   %rbx
    subq    $24, %rsp
Ltmp4:
    .cfi_offset %rbx, -24
    ##DEBUG_VALUE: main:argc <- EDI
    ##DEBUG_VALUE: main:argv <- RSI
    leaq    L_.str(%rip), %rbx
    leaq    -24(%rbp), %rsi
Ltmp5:
    ##DEBUG_VALUE: operator*=<int, int>:rhs <- [RSI+0]
    ##DEBUG_VALUE: operator*<int, int>:rhs <- [RSI+0]
    ##DEBUG_VALUE: main:b <- [RSI+0]
    xorl    %eax, %eax
    movq    %rbx, %rdi
Ltmp6:
    callq   _scanf
    .loc    6 34 17                 ## main.cpp:34:17
    leaq    -20(%rbp), %rsi
Ltmp7:
    xorl    %eax, %eax
    .loc    6 34 2 is_stmt 0        ## main.cpp:34:2
    movq    %rbx, %rdi
    callq   _scanf
    .loc    6 35 17 is_stmt 1       ## main.cpp:35:17
    leaq    -16(%rbp), %rsi
    xorl    %eax, %eax
    .loc    6 35 2 is_stmt 0        ## main.cpp:35:2
    movq    %rbx, %rdi
    callq   _scanf
    .loc    6 22 18 is_stmt 1       ## main.cpp:22:18
Ltmp8:
    movl    -24(%rbp), %esi
    .loc    6 23 18                 ## main.cpp:23:18
    movl    -20(%rbp), %edx
    .loc    6 23 11 is_stmt 0       ## main.cpp:23:11
    addl    %edx, %edx
    .loc    6 24 11 is_stmt 1       ## main.cpp:24:11
    imull   $3, -16(%rbp), %ecx
Ltmp9:
    ##DEBUG_VALUE: main:c [bit_piece offset=64 size=32] <- ECX
    .loc    6 39 2                  ## main.cpp:39:2
    leaq    L_.str1(%rip), %rdi
    xorl    %eax, %eax
    callq   _printf
    xorl    %eax, %eax
    .loc    6 41 5                  ## main.cpp:41:5
    addq    $24, %rsp
    popq    %rbx
    popq    %rbp
    retq
Ltmp10:
Lfunc_end0:
    .cfi_endproc

For this code, the assembly at -O2, -O3, -Os, & -Ofast all look identical.