And that breaks a _massive_ amount of kernel code, including such core functionality like SPIN_LOCK_UNLOCKED and a host of others. There are all sorts of macros that use member initialization of that form.
This does not break the code at run time, this breaks the code at compile time, and should be less painful.
So breaking 90% of the source code at compile time is ok? I think not. The kernel relies really _really_ heavily on such structure initializers, and breaking them would effectively break the world as far as the kernel is concerned.
I agree, it would be a bad idea to compile the existing C code by g++. The good idea is to be able to produce new C++ modules etc.
No, this is a reason why C++ modules are _not_ a good idea. If you could write the module in C or C++, but in C++ it compiled 100-200% slower, then you would write it in C. Why? A simple matter of numbers:
Say it takes you 100 hours to write and debug the module in C++, and 140 to write and debug it in C. I estimate that at least 200,000 people would download and compile a single version of the kernel with your module (not an unreasonable estimate). Note that I'm not even including the people who do repeated regression testing of versions, or people who download and compile multiple versions of the kernel. If the source file takes an average of 1.0 seconds to compile in C and 2.0 seconds to compile in C++, then:
(2.0 sec - 1.0 sec) * 200,000 = 200,000 seconds = 55.6 hours
140 hours - 100 hours = 40 hours
40 hours < 55.6 hours
So for a single version of the kernel your module, you've already wasted 15.6 hours of time across people using it. Over time that number is just going to grow, _especially_ if people start writing more and more modules in C++ because they can. If you want to build C++ in the kernel, write a compiler that does not include all the problematic C++ features that add so much parsing time (overloaded operators, etc).
You mentioned a bad example. The struct list_head has [almost?] all "members" inlined. If they were not, one could simply make a base class having [some] members outlined, and which class does not enforce type safety and is for inheritance only. The template class would then inherit the base one enforcing type safety by having inline members. This technique is well known, trust me. If you need real life examples, tell me.
Ok, help me understand here: Instead of helping using one sensible data structure and generating optimized code for that, the language actively _encourages_ you to duplicate classes and interfaces, providing even _more_ work for the compiler, making the code harder to debug, and probably introducing inefficiencies as well. If C++ doesn't work properly for a simple and clean example like struct list_head, why should we assume that it's going to work any better for more complicated examples in the rest of the kernel? Whether or not some arbitrary function is inlined should be totally orthogonal to adding type-checking.
For #defines core_initcall() ... late_initcall() I would type something like this:
class foo_t { foo_t(); ~foo_t(); }
static char foo_storage[sizeof(foo_t)];
static foo_t& foo=*reinterpret_cast<foo_t*>(foo_storage);
static void __init foo_init() { new(foo_storage) foo_t; }
core_initcall(foo_init);
This ugly-looking code can be nicely wrapped into a template, which, depending on the type (foo_t in this case), at compile time, picks the proper stage for initialization.
You proved my point. Static constructors can't work. You can add silly wrapper initcall functions which create objects in static memory at various times, but the language-defined static constructors are yet another C++ feature that doesn't work by default and has to be hacked around. C++ gives us no advantage over C here either. Plus this would break things like static spinlock initialization. How would you make this work sanely for this static declaration:
spinlock_t foo_lock = SPIN_LOCK_UNLOCKED;
Under C that turns into (depending on config options):
spinlock_t foo_lock = { .value = 0, .owner = NULL, (...) };
How could that possibly work in C++ given what you've said? Anything that breaks code that simple is an automatic nonstarter for the kernel. Also remember that spinlocks are defined preinitialized at the very earliest stages of init. Of course I probably don't have to say that anything that tries to run a function to iterate over all statically-allocated spinlocks during init would be rejected out of hand.