Multiple Inheritance vs. Interface

[Pavel]

I still don't agree. If you use single-inheritance under C++, there is no
performance penalty compared to any vtable-based programming language which
would result from the fact that C++ provides multiple-inheritance. Since the
vtables and the data members of classes can be laid out by the compiler in such
a fashion that any "this" pointer never has to be adjusted in a
single-inheritance hierarchy, C++ is as effecient as it ever gets.

[snip]

Because the compiler does not know whether the base is the first
base in the particular most-derived class, ... [snip]

If the compiler wants to do anything with a class, it needs to know the complete
definition of the class and all its base classes.

No, it does not need to know "the complete definition". I am not sure about
"anything", but specifically for generating code that calls a virtual function
all the compiler needs to know is how to calculate the address of a particular
subobject of the object of the most-derived class pointed by the pointer given
to compiler. That particular subobject is uniquely defined by two requirements:
a) it contains the sub-subobject pointed to the given pointer as one of its bases
b) of all such sub-objects it is of the most derived class that overrode the
virtual function being called

such subobject can be calculated by
- subtracting an offset stored in the virtual table pointed by given pointer,
from that given pointer (that offset can be stored in the table in more than one
way)
- or by creating multiple virtual tables where some entries (those allocated by
the classes where the pointer's type is their first base) point to the actual
functions and the others point to pointer-adjusting thunks that forward control
to that original function after the adjustment.

None of the two ways is in general faster than the other (although I admit the
"thunk" way avoids direct penalties for calling virtual by pointers on first base).

So it knows whether a certain

base class is the first base class of another class or not.

No, in general it does not know it at the call site. In the example you snipped
form my previous post for no reason this fact was clearly demonstrated:

When compiler is to generate the code for the virtual call

bPtr->getBOffset() or bPtr2->getBOffset()

it does not have access to the definition of the most-derived class of the
object pointed to by bPtr.

The call site is in client.cpp. bPtr and bPtr2 point to an object of either
class D (that has B as the first base) or D2 (that has it as the second base)
but the compiler does not have definition of either of them in view because
these definitions are in d.h and d2.h, respectively, and client.cpp does not
include either of these two headers, directly or indirectly. Thus, compiler has
no way of knowing whether bPtr and bPtr2 point to the first base of their most
derived classes or not.


What you probably

tried to say is that the compiler, receiving a pointer to Base* cannot know
whether the real type of the object is Base or Derived,

It is a true statement by itself, but it is not what I am saying.

so a "this"-pointer

adjustment may have to be performed when any of Derived's methods should be
invoked.

Note that this adjustment is most likely done inside a "thunk"-method which
simply adjusts the "this" pointer and invokes Derived's implementation method
that does not need to adjust the "this" pointer.

Correct. Extra Invocation is not free though (even though it is not a full
invocation but just a jump on most architectures; but jump is not free either).
Simple adjustment will work better for non-first bases; thus some
implementations still use it; and some classes will work faster with it.

Like so:

class Base1 {
int b1;
virtual void print1 ();
};
class Base2 {
int b2;
virtual void print2 ();
};
class Derived : Base1, Base2 {
int derived;
virtual void print1 ();
virtual void print2 ();
};

,____________________ ___________ ,________________________
+0 | |vtable |->. print1 | |void Derived:rint2 { |
|Base1 |----------| | print2 |->| std::cout << derived;|
+4 | |int b1; | |_________| |//&derived == this+16; |
|________|__________| ___________ |} |
+8 | |vtable |->| print2 | |_______________________|
|Base2 |----------| |_________| ,________________________
+12 | |int b2; | | |__asm { |
|________|__________| |---->| this -= 8; |
+16 |int derived; | | call Derived:rint2 |
|___________________| |} |
|_______________________|

And even that is an implementation detail: A compiler writer may choose to
generate Derived:rint multiple times,

They can by that creating the code bloating issue (for a change, unrelated to
templates) with all its accompanying instructions-cache-use-efficiency issues.

Also, think of all the the nice linkage issues: what will be external names (or,
for that matter, addresses) for versions one and two of the Derived:rint2? The
definitive address matters because in C++ you need to provide the capability of
calling-by-pointer-to-member, for the 1st and non-1st bases.

BTW such calls via pointer-to-members, even via the first base-pointer cannot be
made as efficient as they could be in the languages with the single inheritance
(such as C++ version ARM1 ). It's all IMHO, of course, I have to be careful
now. This is because at the time pointer-to-member is initialized it is not yet
known whether it will be used with the pointer-to-the-first base or
pointer-to-the-higher-than-first base.

Long story short -- you cannot receive something (in our case classes that can
be interchangeably used as bases for both single and multiple inheritance) for
nothing (in our case no performance cost). (although I would agree it's worth to
die trying . )

each version using its own set of

offsets. This would eliminate the need for thunking code completely at the cost
of larger executables:
,____________________ ___________ ,________________________
+0 | |vtable |->. print1 | |void Derived:rint2 { |
|Base1 |----------| | print2 |->| std::cout << derived;|
+4 | |int b1; | |_________| |//&derived == this+16; |
|________|__________| ___________ |} |
+8 | |vtable |->| print2 | |_______________________|
|Base2 |----------| |_________|
+12 | |int b2; | | ,________________________
|________|__________| |---->|void Derived:rint2 { |
+16 |int derived; | | std::cout << derived;|
|___________________| |//&derived == this+8; |
|} |
|_______________________|

Regards,
Stuart

-Pavel

 

[Stuart]

No, it does not need to know "the complete definition".

You got me there. I meant to write complete declaration. Sorry.

I am not sure
about "anything", but specifically for generating code that calls a
virtual function all the compiler needs to know is how to calculate the
address of a particular subobject of the object of the most-derived
class pointed by the pointer given to compiler.

This is an implementation detail of the compiler. "Thunking" compilers
do not have to do any additional work when they invoke a virtual function:
Base* ptr = ...;
ptr->foo ();

will result in the following (pseudo-) ASM:
push ptr;
call *[ptr + &foo]

where &foo gives the slot of the vtable where foo's address is stored.
In order to be able to generate such code, the compiler only has to know
the complete declaration of Base and Base's base classes (or else it
could not figure out the slot number of foo), even if ptr is initialized
with an instance of the hitherto unknown class Derived.

[snip]

[...] I
admit the "thunk" way avoids direct penalties for calling virtual by
pointers on first base).

Here we go. If I may cite you from up-thread:

I think the above is not accurate. C++ code does suffer performance
penalties from using multiple inheritance. Moreover, and what's
especially frustrating, even the code that does not use multiple
inheritance (in fact, any code using virtual functions) suffers from
at least one performance penalty imposed by the way C++ supports
multiple inheritance: the necessity to read the offset of the call
target [...]

Apparently you proved yourself wrong: It is not an inherent limitation
of the C++ language that it _must_ be slower for single-inheritance
hierarchies because it also supports multiple-inheritance. I think it
would be more correct to say that C++ can be as efficient as any
single-inheritance vtable-based language, provided the implementation of
the C++ compiler is good (IOW, uses thunks).

Regards,
Stuart

 

[88888 Dihedral]

Pavelæ–¼ 2012å¹´10月8日星期一UTC+8上åˆ4時57分53秒寫é“：

I think the above is not accurate. C++ code does suffer performance penalties

from using multiple inheritance. Moreover, and what's especially frustrating,

even the code that does not use multiple inheritance (in fact, any code using

virtual functions) suffers from at least one performance penalty imposed by the

way C++ supports multiple inheritance: the necessity to read the offset of the

call target within the object of the most-derived class overriding the virtual

method and subtracting this offset from the passed pointer to let the virtual

function implementation access to the object it expects.



Languages with single inheritance can assign a single offset from the start of

the virtual table of the most-derived class of an object to the start of the

slice of that class' virtual table correspondent to the virtual table of any of

its bases. This effectively means that any class in such a language can have a

single virtual table and the objects of the most derived class and the

correspondent objects of all its base classes can have a single address.



Complicating C++ specification by introducing special kind of classes that would

be forbidden from being used as base classes in multiple inheritance (or,even

more complex but more rewarding as well -- the classes that cannot be used as

other-than-first base classes in multiple inheritance) could eliminate this

penalty for the language users who do not use multiple inheritance.



C++ does not have a chance of assigning any virtual function once definedin a

class a single offset in a virtual table; therefore, it has to have multiple

virtual tables. As it is, that is without the complication mentioned above, C++

can not let its compiler know at a virtual call site that the call is on the

object that is the first base of the most-derived class; hence the necessity to

always read and apply the offset at run-time.



Languages with single inheritance and interfaces still have same performance

advantages as languages without interfaces for the classes that do not implement

interfaces (that is, they live to the promise of not imposing cost of a feature

on the code that does not use it better than C++ does). For a class that does

implement interfaces, the implementations have a choice of either avoiding space

cost but making calls by interface significantly more expensive or adding

pointers to individual "interface virtual tables" to the layout of an object of

such a class and having calls by an interface only one indirection more

expensive than regular virtual calls. I believe Java 1.0 took the first path and

Java 1.1 and all its further versions took the second.



-Pavel

I think testing equivalent source programs in objetive c and c++
under the same os can tell the differences.


linux or bsd can be

 

[Richard Damon]

True but thunk approach comes at higher cost for calling at non-first
base of extra jump as compared with 'classic' approach. Extra jump in
chunk is equivalent to at least one extra memory read (only to
instruction instead of data cache) and some instruction decoding. The
approach also somewhat increases memory usage with the second virtual
table and thunks. But you are right about zero-cost if calls by
non-first base are not used -- I completely forgot about thunk approach.

-Pavel

Actually, most of the jumps can be removed, by placing the thunk
directly in front of the code for the subroutine.

In fact, if we define that when calling a virtual function, the this
pointer is always set to point to the sub object of the type where the
function was first declared as virtual, then the function receiving the
call knows exactly how to adjust the this pointer to point to the now
full object (of the type the function belongs to). and can place that
adjustment as a thunk just before the normal code for the function if we
want to allow it to be called "normally" in a non-virtual manner without
needed to pre-adjust the this pointer for those calls.

The only case where these adjustments become non-trivial would be in
cases of virtual inheritance where the pointer adjustment will need to
do a look up rather than a simple offset.

The simple procedure here is:

When calling a virtual function that was first defined in a non-first
base class (or a non-virtual function that is last defined in a
non-first base class), before making the call (this will be a simple
addition to the base pointer if we have non-virtual inheritance).

At the function side, if the function is a virtual function of a type
that has the initial definition in a non-first base class, include code
to adjust the this pointer from that base class to the actual class just
before the normal entry point, and have the vtable point there. The
regular entry point can be used for non-virtual calls, with the caller
passing the "normal" this pointer.

This is probably minimal overhead (works best if the this pointer is
passed in a register so the address offset is quick and easy). The only
case where you get "extra" adjustments is if you make the call through a
pointer to a class derived from one is non-first base derived from the
class that first defined the virtual function to a function defined in
class similarly after the multiple derivation, as you will adjust to the
base and then back.

 

[Pavel]

You got me there. I meant to write complete declaration. Sorry.


This is an implementation detail of the compiler.

I doubt it. I think the necessity to know the address of a particular subobject
follows from the necessity for the virtual functions to be able to access the
members of that particular sub-object (e.g. its data members); therefor this
necessity is hardly an implementation detail but a generic corollary from the
language definition.

"Thunking" compilers do not

have to do any additional work when they invoke a virtual function:
Base* ptr = ...;
ptr->foo ();

will result in the following (pseudo-) ASM:
push ptr;
call *[ptr + &foo]

this works as long as *[ptr + &foo] evaluates the different address for every
different ptr at which foo() can be called for the object of the class (and
there can be plenty of these for a single definition of foo()). Which leaves the
implementation with choice to either use thunks with jumps (here is your extra
work) for all but one such different addresses or have multiple foo()s with the
ptr adjustment code inlined in each but one of them (there is no extra jump but
the duplication of the whole code of foo, again for every different ptr at which
the foo() can be called for the object of the class.

where &foo gives the slot of the vtable where foo's address is stored. In order
to be able to generate such code, the compiler only has to know the complete
declaration of Base and Base's base classes (or else it could not figure out the
slot number of foo), even if ptr is initialized with an instance of the hitherto
unknown class Derived. agree.

[snip]
[...] I
admit the "thunk" way avoids direct penalties for calling virtual by
pointers on first base).

Here we go. If I may cite you from up-thread:

I think the above is not accurate. C++ code does suffer performance
penalties from using multiple inheritance. Moreover, and what's
especially frustrating, even the code that does not use multiple
inheritance (in fact, any code using virtual functions) suffers from > at

least one performance penalty imposed by the way C++ supports

multiple inheritance: the necessity to read the offset of the call
target [...]

Apparently you proved yourself wrong:

I have already admitted that the above statement was not accurate; I forgot
about the thunk approach; but not every implementation chooses it, and for
reason: as we saw above, thunks penalize calls by non-first bases by an extra
jump. As for the alternative of copying the entire function, it would penalize
the program by unconditionally bloating the code size.

It is not an inherent limitation of the

C++ language that it _must_ be slower for single-inheritance hierarchies because
it also supports multiple-inheritance. I think it would be more correct to say
that C++ can be as efficient as any single-inheritance vtable-based language,
provided the implementation of the C++ compiler is good (IOW, uses thunks).

As I hope I have shown above, the thunk approach has its own penalties as
compared to the direct pointer adjustment at the call site; thus I would not
unconditionally qualify it as "good". Different programs may perform better or
worse with this approach than with the direct adjustment.

Regards,
Stuart

-Pavel

 

[Pavel]

Actually, most of the jumps can be removed, by placing the thunk
directly in front of the code for the subroutine.

Not unless you generate multiple copies of the subroutine because there can be
more than one offset for the same function (D::f() can be called by either of
its the bases that declare virtual f() and there can be more than one such base,
some inheriting from the other).

One problem with generating multiple functions is that one function can be
relatively long (in code terms) and *all of them has to be generated* regardless
of whether even one of them is used (as opposed to the code-bloat caused by
templates that only generates really used functions). For example, if you have
this hierarchy: class B { public: virtual int f(); };
class B1 {...}; class B2 {...}; class B3 {...};
class D1: public B1, public B {...}; // no f() override!
class D2: public B3, public D1 {...}; // no f() override!
class D3: public B2, public D2 {... virtual int f(); };,

you will have to generate 4 virtual tables and 4 different functions fully
copying the body of function D3::f(): to call it by B*, D1*, D2* and D3*; of
these, the former will not have adjustment and the latter 3 will.

In fact, if we define that when calling a virtual function, the this
pointer is always set to point to the sub object of the type where the
function was first declared as virtual,

How can you define 'this' to be always set this way? This same pointer can be
used for other purposes, too (for example, for accessing non-virtual members of
an object of its 'compile-time' type. It *has* to point to a particular
sub-object (which one, in general depends on the way how it was created; but if
a class is not derived from the same base class more than once, it has to be to
that only sub-object that is of the class matching the pointer type).

then the function receiving the

call knows exactly how to adjust the this pointer to point to the now
full object (of the type the function belongs to). and can place that
adjustment as a thunk just before the normal code for the function if we
want to allow it to be called "normally" in a non-virtual manner without
needed to pre-adjust the this pointer for those calls.

The only case where these adjustments become non-trivial would be in
cases of virtual inheritance where the pointer adjustment will need to
do a look up rather than a simple offset.

The simple procedure here is:

When calling a virtual function that was first defined in a non-first
base class (or a non-virtual function that is last defined in a
non-first base class), before making the call (this will be a simple
addition to the base pointer if we have non-virtual inheritance).

This does not seem to be complete statement and seems to be talking about two
different things with two different costs:

- To call a non-virtual function on a non-first base the function of the base
class or its ancestor will be called. If it is an ancestor and its offset in the
"compile-time" class of the pointer is not zero, you might need to offset the
pointer but that offset is known in the compile-time so it is less of an issue
(no extra memory read or a jump)

- To call a virtual function by a pointer, compiler *cannot tell* whether the
pointer's compile time is the first base or not (and even whether it is the
"proper" base or the most-derived class itself). This information will only be
known at run-time and can differ at that same call site from one call to
another. "Thunk" approach only works because it, too, uses the run-time value of
the pointer; compiler cannot generate any addition "before making the call"
unless the generated code contains the memory read from the area (obviously,
indirectly) pointed to by the pointer by which the call is.

At the function side, if the function is a virtual function of a type
that has the initial definition in a non-first base class, include code
to adjust the this pointer from that base class to the actual class just
before the normal entry point, and have the vtable point there.

Yes, but without jump it creates the whole new function for every such base
whether it (the function) is going to be called by that base anywhere at all or
not. I think that's why thunks with jumps are used in practice but I am unsure
about the multiple versions of the functions (for trivial functions, it may not
be bad idea though; but I expect complications with the name and cross-compiler
linking. I am not sure if ABI supports such things though (simply don't know);
otherwise, each compiler will name its "additional" virtual functions for the
same 'user-visible' name differently).

The
regular entry point can be used for non-virtual calls, with the caller
passing the "normal" this pointer.

true, but probably besides the point.

This is probably minimal overhead

It probably is -- in terms of instructions involved; but code bloating seems to
be nasty; and it does lead to measurable performance penalty for sizable programs.

(works best if the this pointer is

passed in a register so the address offset is quick and easy). The only
case where you get "extra" adjustments is if you make the call through a
pointer to a class derived from one is non-first base derived from the
class that first defined the virtual function to a function defined in
class similarly after the multiple derivation, as you will adjust to the
base and then back.

-Pavel

 

[Stuart]

On 10/12/12 Pavel wrote:
[snip]

As I hope I have shown above, the thunk approach has its own penalties
as compared to the direct pointer adjustment at the call site; thus I
would not unconditionally qualify it as "good". Different programs may
perform better or worse with this approach than with the direct adjustment.

Agreed. So this is the lesson to be learned: If one uses a
single-inheritance hierarchy, it is wizer to choose a C++ compiler that
does do thunking. Likewise, if one uses multiple inheritance a lot, one
should rather use a call-site adjustment compiler (or ensure that the
most used interface/base class is the first base class).

Of course, one only has to care about this if the slow-down is really
due to the additional thunking code or due to the pointer adjustment.
Most of the time, this will be the least problem.

Regards,
Stuart

PS: This thread turned out to be more interesting than I had initially
thought. I have never heard about the pointer-adjustment technique
before. Thanks for sharing, Pavel.

 

[Richard Damon]

Not unless you generate multiple copies of the subroutine because there
can be more than one offset for the same function (D::f() can be called
by either of its the bases that declare virtual f() and there can be
more than one such base, some inheriting from the other).

One problem with generating multiple functions is that one function can
be relatively long (in code terms) and *all of them has to be generated*
regardless of whether even one of them is used (as opposed to the
code-bloat caused by templates that only generates really used
functions). For example, if you have this hierarchy: class B { public:
virtual int f(); };
class B1 {...}; class B2 {...}; class B3 {...};
class D1: public B1, public B {...}; // no f() override!
class D2: public B3, public D1 {...}; // no f() override!
class D3: public B2, public D2 {... virtual int f(); };,

you will have to generate 4 virtual tables and 4 different functions
fully copying the body of function D3::f(): to call it by B*, D1*, D2*
and D3*; of these, the former will not have adjustment and the latter 3
will.

As I have been thinking about it, supporting pointer to member functions
basically will require any implementation to have a virtual table for
every distinct base class (That defines virtual functions) excepting
single derivation. So the the 4 virtual tables are basically required.

There does NOT need to be 4 different functions though. There may need
to be 4 entry points, but the whole body doesn't need to be duplicated,
the entries which need to adjust the this pointer, can adjust the
pointer parameter and then fall into/jump into the main subroutine. You
may not be able to write the equivalent to this in C++, but the complier
doesn't need to. If you were to attempt to write it is C++ what you
would need to do is something like:

D3::f(D* d) { ... } // This parameter being shown explisitly

D3::f(B* b) { return f(static_cast<D*>(b)); }

and then let the compiler use tail recursion like elimination to replace
the call with a jump.

The next step is to realize that we can eliminate the jump by placing
the code directly in from of the original function. In the case of
multiple layers, they can be stacked B* -> D1* -> D2* -> D3* to
eliminate the jumps, or the compiler can decide when the cost of the
additional conversion exceeds the cost of the jump, and just add the
jump. Thus the cost will never be higher than an adjustment followed by
a jump, will be 0 cost if no adjustment is needed, and only the cost of
a single adjustment if called from a single level of non-first base
class (no jump needed).

For many systems, if we are careful with are ABI design, so that the
this pointer is passed via a register, the adjustment can be as simple
as a single instruction in the basic case, a add or subtract immediate.

Thus a possible layout of the entry to the function would be the
equivalent to

D3::f__thunk_B: this <- this - offsetof B in D1
D3::f__thunk_D1: this <- this - offsetof D1 in D2
D3::f__thunk_D2: this <- this - offsetof D2 in D3
D3::f start of normal code for D3::f

If a jump is cheaper than 2 subtracts then it
D3::f__thunk_B: this <- this - offsetof B in D3
jump B3::f
D3::f__thunk_D1: this <- this - offsetof D1 in D2
D3::f__thunk_D2: this <- this - offsetof D2 in D3
D3::f start of normal code for D3::f

This makes the thunking cost for a virtual function on par with the
fixed adjustment cost for calling any non-first base member function,
with at most the addition of a jump instruction (and this additional
cost ONLY occurs if there are multiple non-first base derivations
happening).

How can you define 'this' to be always set this way? This same pointer
can be used for other purposes, too (for example, for accessing
non-virtual members of an object of its 'compile-time' type. It *has* to
point to a particular sub-object (which one, in general depends on the
way how it was created; but if a class is not derived from the same base
class more than once, it has to be to that only sub-object that is of
the class matching the pointer type).

The key feature is that if you call using a pointer in the "B"
sub-objects vtable, than you need to call it with a B* pointer as this.
If the function is actually a D3 member function, then IT will adjust
the this pointer as described above. For a direct virtual call (not via
a member-function pointer), the compiler can choose the virtual table
for the most derived class that it knows about that defines an override.
If the object isn't of a more derived class that defines an override
(say D4), then the code will just execute the adjustment at the call
site and go directly into the function. If there is a subsequent
override, then the call will go into the above thunk entry code, correct
the this pointer. The compiler also has the option of using the most
derived class's vtable and not adjust the pointer, knowing that this
will go into a thunk that will adjust the pointer. If for example we
call a g() that was declared virtual in B and D2, it will need to go
into a thunk that adds the offset of D2 in D3 and then jump into
D2::g(). This is a trade off of gaining space (using the common thunk,
instead of adjusting at each call) at the expense of time (the cost of a
jump). If jumps are really expensive, it may be possible with proper
linker instruction to build the preamble for g to be:

D2::g__thunk_D3: this <- this + offsetof B in D3
D2::g__thunk_B: this <- this - offsetof B in D1
D2::g__thunk_D1: this <- this - offsetof D1 in D2
D2::g start of normal code for D2::g

(and further add later thunks in front until the cost of the extra
adjustments exceed the cost of the jump).

Note again, the main body of the member function has the this pointer
pointing to the (sub)object of the type of the member function, the
possible different type of this is only for the input to the thunk,
which effectively casts the pointer to the needed type.

then the function receiving the

This does not seem to be complete statement and seems to be talking
about two different things with two different costs:

- To call a non-virtual function on a non-first base the function of the
base class or its ancestor will be called. If it is an ancestor and its
offset in the "compile-time" class of the pointer is not zero, you might
need to offset the pointer but that offset is known in the compile-time
so it is less of an issue (no extra memory read or a jump)

- To call a virtual function by a pointer, compiler *cannot tell*
whether the pointer's compile time is the first base or not (and even
whether it is the "proper" base or the most-derived class itself). This
information will only be known at run-time and can differ at that same
call site from one call to another. "Thunk" approach only works because
it, too, uses the run-time value of the pointer; compiler cannot
generate any addition "before making the call" unless the generated code
contains the memory read from the area (obviously, indirectly) pointed
to by the pointer by which the call is.

Perhaps I wasn't as clear on what I was describing here. In the case of
a call via a pointer to object -> name of member function call, the
compiler knows a lot of the layout, and can use this to choose how it
wants to code the call among its options. It can validly choose any of
the vtables as long as it uses the proper object base for the call.
Since with the structure I have described, "down casting" thunks are
more expensive, it might make sense to down cast at the call site and
then make a call that, at worse, will only need to up cast.

Yes, but without jump it creates the whole new function for every such
base whether it (the function) is going to be called by that base
anywhere at all or not. I think that's why thunks with jumps are used in
practice but I am unsure about the multiple versions of the functions
(for trivial functions, it may not be bad idea though; but I expect
complications with the name and cross-compiler linking. I am not sure if
ABI supports such things though (simply don't know); otherwise, each
compiler will name its "additional" virtual functions for the same
'user-visible' name differently).

As I have pointed out, these entries do NOT need "whole new functions"
but a bit of adjusting code at the entry point to make the call with the
"wrong" type of this now correct. From the call side, they may act like
different functions, since they all called with different types of
parameters, but the body (after the thunk) is likely to all be the same
(not just identical looking code, but being the same code at the same
addresses). Of course, if the function is so simple that the code for a
version with the different base is "better" than doing the adjustment,
the compiler if free to make them separate functions.

true, but probably besides the point.

It probably is -- in terms of instructions involved; but code bloating
seems to be nasty; and it does lead to measurable performance penalty
for sizable programs.

It seems to be minimal, and only imposed when needed. As opposed to the
other method presented which places in the vtable an offset FOR ALL
FUNCTIONS and the offset is applied AT THE CALL SITE for ALL CALLS.

This will add more space and time for most programs (since it needs the
offset even for 1st base entries, and adds code to every call) as
opposed to only when needed.

 

[Pavel]

As I have been thinking about it, supporting pointer to member functions
basically will require any implementation to have a virtual table for
every distinct base class (That defines virtual functions) excepting
single derivation. So the the 4 virtual tables are basically required.

There does NOT need to be 4 different functions though. There may need
to be 4 entry points, but the whole body doesn't need to be duplicated,
the entries which need to adjust the this pointer, can adjust the
pointer parameter and then fall into/jump into the main subroutine. You
may not be able to write the equivalent to this in C++, but the complier
doesn't need to. If you were to attempt to write it is C++ what you
would need to do is something like:

D3::f(D* d) { ... } // This parameter being shown explisitly

D3::f(B* b) { return f(static_cast<D*>(b)); }

and then let the compiler use tail recursion like elimination to replace
the call with a jump.

The next step is to realize that we can eliminate the jump by placing
the code directly in from of the original function. In the case of
multiple layers, they can be stacked B* -> D1* -> D2* -> D3* to
eliminate the jumps, or the compiler can decide when the cost of the
additional conversion exceeds the cost of the jump, and just add the
jump. Thus the cost will never be higher than an adjustment followed by
a jump, will be 0 cost if no adjustment is needed, and only the cost of
a single adjustment if called from a single level of non-first base
class (no jump needed).

For many systems, if we are careful with are ABI design, so that the
this pointer is passed via a register, the adjustment can be as simple
as a single instruction in the basic case, a add or subtract immediate.

Thus a possible layout of the entry to the function would be the
equivalent to

D3::f__thunk_B: this <- this - offsetof B in D1
D3::f__thunk_D1: this <- this - offsetof D1 in D2
D3::f__thunk_D2: this <- this - offsetof D2 in D3
D3::f start of normal code for D3::f

If a jump is cheaper than 2 subtracts then it
D3::f__thunk_B: this <- this - offsetof B in D3
jump B3::f
D3::f__thunk_D1: this <- this - offsetof D1 in D2
D3::f__thunk_D2: this <- this - offsetof D2 in D3
D3::f start of normal code for D3::f

This makes the thunking cost for a virtual function on par with the
fixed adjustment cost for calling any non-first base member function,
with at most the addition of a jump instruction (and this additional
cost ONLY occurs if there are multiple non-first base derivations
happening).

Agree. This is good design.

The key feature is that if you call using a pointer in the "B"
sub-objects vtable, than you need to call it with a B* pointer as this.
If the function is actually a D3 member function, then IT will adjust
the this pointer as described above. For a direct virtual call (not via
a member-function pointer), the compiler can choose the virtual table
for the most derived class that it knows about that defines an override.
If the object isn't of a more derived class that defines an override
(say D4), then the code will just execute the adjustment at the call
site and go directly into the function. If there is a subsequent
override, then the call will go into the above thunk entry code, correct
the this pointer. The compiler also has the option of using the most
derived class's vtable and not adjust the pointer, knowing that this
will go into a thunk that will adjust the pointer. If for example we
call a g() that was declared virtual in B and D2, it will need to go
into a thunk that adds the offset of D2 in D3 and then jump into
D2::g(). This is a trade off of gaining space (using the common thunk,
instead of adjusting at each call) at the expense of time (the cost of a
jump). If jumps are really expensive, it may be possible with proper
linker instruction to build the preamble for g to be:

D2::g__thunk_D3: this <- this + offsetof B in D3
D2::g__thunk_B: this <- this - offsetof B in D1
D2::g__thunk_D1: this <- this - offsetof D1 in D2
D2::g start of normal code for D2::g

(and further add later thunks in front until the cost of the extra
adjustments exceed the cost of the jump).

Note again, the main body of the member function has the this pointer
pointing to the (sub)object of the type of the member function, the
possible different type of this is only for the input to the thunk,
which effectively casts the pointer to the needed type.



Perhaps I wasn't as clear on what I was describing here. In the case of
a call via a pointer to object -> name of member function call, the
compiler knows a lot of the layout, and can use this to choose how it
wants to code the call among its options. It can validly choose any of
the vtables as long as it uses the proper object base for the call.
Since with the structure I have described, "down casting" thunks are
more expensive, it might make sense to down cast at the call site and
then make a call that, at worse, will only need to up cast.

As I have pointed out, these entries do NOT need "whole new functions"
but a bit of adjusting code at the entry point to make the call with the
"wrong" type of this now correct. From the call side, they may act like
different functions, since they all called with different types of
parameters, but the body (after the thunk) is likely to all be the same
(not just identical looking code, but being the same code at the same
addresses). Of course, if the function is so simple that the code for a
version with the different base is "better" than doing the adjustment,
the compiler if free to make them separate functions.

It seems to be minimal, and only imposed when needed. As opposed to the
other method presented which places in the vtable an offset FOR ALL
FUNCTIONS and the offset is applied AT THE CALL SITE for ALL CALLS.

This will add more space and time for most programs (since it needs the
offset even for 1st base entries, and adds code to every call) as
opposed to only when needed.

-Pavel

 

[Pavel]

On 10/12/12 Pavel wrote:
[snip]

As I hope I have shown above, the thunk approach has its own penalties
as compared to the direct pointer adjustment at the call site; thus I
would not unconditionally qualify it as "good". Different programs may
perform better or worse with this approach than with the direct adjustment.

Agreed. So this is the lesson to be learned: If one uses a single-inheritance
hierarchy, it is wizer to choose a C++ compiler that does do thunking. Likewise,
if one uses multiple inheritance a lot, one should rather use a call-site
adjustment compiler (or ensure that the most used interface/base class is the
first base class).

Of course, one only has to care about this if the slow-down is really due to the
additional thunking code or due to the pointer adjustment. Most of the time,
this will be the least problem.

Regards,
Stuart

PS: This thread turned out to be more interesting than I had initially thought.
I have never heard about the pointer-adjustment technique before. Thanks for
sharing, Pavel.

I think the really useful person on this thread was Richard. See his last post
-- it is a good design and seems to be best of two worlds. The code generator
will be able to select the most optimal thunk; unless the depth is too high, it
probably will be the non-jumping one (adding a constant to a register does not
affect CPU pipelines and imposes only absolutely necessary cache traffic). I
guess I lost my point but I do not feel bad -- it *was* productive.


-Pavel