John Fine <johnsfine@xxxxxxxxxxx> writes:
>> 1. The inline keyword is just a hint. The compiler decides if it uses
>> it or not.
>>
> True. Compared to other C++ compilers, GCC gives the programmer more
> control over whether things are inlined, including more often simply
> obeying the inline keyword. But in the language standard the inline
> keyword is just a hint. Whether GCC obeys that hint depends on a
> variety of switches and other factors.
In C the inline keyword is just a hint. However, in C++, the inline
keyword gives a function internal linkage. If you write
void fn() { }
in a .cc or .h file, and there are no calls to fn, then the object file
will contain a definition of the function fn. If you write
inline void fn() { }
then it will not contain a definition if there are no calls.
In other words, in C++ the inline keyword has a language meaning which
changes the linkage of the function. That said, with regard to whether
the compiler actually inlines a function rather than emitting it out of
line, the inline keyword is just a hint.
>> 4. GCC does inlining not below -O3 or the explizit activation using
>> -finline-functions
>>
> I'm not sure. I think the decision rules are more complex and depend
> on more switches and there would be exceptions to a statement as
> simple as the above.
Yes. For example, -finline-functions-called-once is on by default at
all optimization levels. The algorithm which chooses which functions to
inline is complex, and can be controlled by -finline-limit and by
various tuning parameters.
>> And that also means to me, that there is no way to _not_ inline code if
>> it is given in the header file,
> Using the Intel 10 compiler at a decent level of optimization, it is a
> constant source of frustration that there is no way to tell it to not
> inline things defined in header files. The project I work on has some
> very large templates that must be in header files and shouldn't be
> inlined. The Intel 10 compiler always uses its own judgment and its
> own judgment is usually wrong.
>
> I don't use GCC much in similar situations, but I don't think GCC is
> that flawed on this issue. I think there are practical ways to get it
> to not inline such functions.
Such as __attribute__ ((noinline)). The Intel compiler recognizes a lot
of gcc attributes and may recognize that one.
Ian
