A noinline inline function? What sorcery is this?

Raymond Chen

May 21st, 20202 0

You can declare a noinline inline function.

void g();

// gcc
__attribute__((noinline)) inline void f()
{
    g();
}

// MSVC
__declspec(noinline) inline void f()
{
    g();
}

void tryme()
{
    f();
    f();
}

What sorcery is this, a function that is both inline and not-inline?

The two keywords are not contradictory because they describe different senses of the word “inline”.

The C++ language keyword inline means “can be defined in multiple translation units without triggering an ODR violation.” In other words, it lets you put the function definition in a header file that is included by multiple C++ files.

The function attribute/declaration specifier noinline means “do not inline this function during code generation.” It is a directive to the optimizer not to perform inline substitution during code generation.

Historically, the inline C++ keyword was originally an optimizer hint, but optimizers were given permission to ignore it and make their own decisions about inline substitution during code generation. Nowadays, compilers pretty much ignore the optimization aspect of the inline keyword. The only thing that remains of the inline keyword is the ability to have multiple definitions without violating ODR.

You could say that the modern sense of the C++ keyword inline is “defined right here.” It’s a statement about the source code, not the object code.

In the above example, the function f is a noinline inline function. The inline keyword allows the definition of f to go into a header file that is consumed by multiple translation units. The noinline attribute/declaration specifier tells the optimizer to emit code for f and call it, rather than embedding the body of f into its call sites. The function tryme will call the function f twice, instead of optimizing out the call and just calling g twice.

Raymond Chen

Read next

2 comments

Discussion is closed. Login to edit/delete existing comments.

  • Stuart Dootson 0

    And that’s why they used inline for variables defined in headers with the same ODR dodging property. Makes sense when you explain the meaning of inline like that…

    Let’s just not try and factor inline namespaces into the same explanation.

  • Joshua Hudson 0

    I remember these early C++ books that had all their classes in .h files and all the methods in the classes themselves, kind of like C# and Java have you do it. Any member function in a class was automatically inline. It seemed like a great idea at the time until you saw what mess it compiled to. At the time, this yielded a copy of every inline function in the resulting executable for every C++ file that called it, which was frequently a lot, and it totally broke RTTI because the same class had multiple vtables and RTTI and dynamic downcast compared vtable pointers.

    But today C++ doesn’t have to work with traditional linkers and the compiler emits mergable definitions and the linker knows how to do it, so you only pay for this in compile costs. We can actually revert and start writing it like the entry level books suggested it be written, and expect it to work well.

Feedback usabilla icon