Why does LLVM bitcode have duplicate symbols for constructors? [duplicate] - llvm

Question

Today, I discovered a rather interesting thing about either g++ or nm...constructor definitions appear to have two entries in libraries.
I have a header thing.hpp:
class Thing
{
Thing();
Thing(int x);
void foo();
};
And thing.cpp:
#include "thing.hpp"
Thing::Thing()
{ }
Thing::Thing(int x)
{ }
void Thing::foo()
{ }
I compile this with:
g++ thing.cpp -c -o libthing.a
Then, I run nm on it:
%> nm -gC libthing.a
0000000000000030 T Thing::foo()
0000000000000022 T Thing::Thing(int)
000000000000000a T Thing::Thing()
0000000000000014 T Thing::Thing(int)
0000000000000000 T Thing::Thing()
U __gxx_personality_v0
As you can see, both of the constructors for Thing are listed with two entries in the generated static library. My g++ is 4.4.3, but the same behavior happens in clang, so it isn't just a gcc issue.
This doesn't cause any apparent problems, but I was wondering:
Why are defined constructors listed twice?
Why doesn't this cause "multiple definition of symbol __" problems?
EDIT: For Carl, the output without the C argument:
%> nm -g libthing.a
0000000000000030 T _ZN5Thing3fooEv
0000000000000022 T _ZN5ThingC1Ei
000000000000000a T _ZN5ThingC1Ev
0000000000000014 T _ZN5ThingC2Ei
0000000000000000 T _ZN5ThingC2Ev
U __gxx_personality_v0
As you can see...the same function is generating multiple symbols, which is still quite curious.
And while we're at it, here is a section of generated assembly:
.globl _ZN5ThingC2Ev
.type _ZN5ThingC2Ev, #function
_ZN5ThingC2Ev:
.LFB1:
.cfi_startproc
.cfi_personality 0x3,__gxx_personality_v0
pushq %rbp
.cfi_def_cfa_offset 16
movq %rsp, %rbp
.cfi_offset 6, -16
.cfi_def_cfa_register 6
movq %rdi, -8(%rbp)
leave
ret
.cfi_endproc
.LFE1:
.size _ZN5ThingC2Ev, .-_ZN5ThingC2Ev
.align 2
.globl _ZN5ThingC1Ev
.type _ZN5ThingC1Ev, #function
_ZN5ThingC1Ev:
.LFB2:
.cfi_startproc
.cfi_personality 0x3,__gxx_personality_v0
pushq %rbp
.cfi_def_cfa_offset 16
movq %rsp, %rbp
.cfi_offset 6, -16
.cfi_def_cfa_register 6
movq %rdi, -8(%rbp)
leave
ret
.cfi_endproc
So the generated code is...well...the same.
EDIT: To see what constructor actually gets called, I changed Thing::foo() to this:
void Thing::foo()
{
Thing t;
}
The generated assembly is:
.globl _ZN5Thing3fooEv
.type _ZN5Thing3fooEv, #function
_ZN5Thing3fooEv:
.LFB550:
.cfi_startproc
.cfi_personality 0x3,__gxx_personality_v0
pushq %rbp
.cfi_def_cfa_offset 16
movq %rsp, %rbp
.cfi_offset 6, -16
.cfi_def_cfa_register 6
subq $48, %rsp
movq %rdi, -40(%rbp)
leaq -32(%rbp), %rax
movq %rax, %rdi
call _ZN5ThingC1Ev
leaq -32(%rbp), %rax
movq %rax, %rdi
call _ZN5ThingD1Ev
leave
ret
.cfi_endproc
So it is invoking the complete object constructor.

Answer 1

We'll start by declaring that GCC follows the Itanium C++ ABI.
According to the ABI, the mangled name for your Thing::foo() is easily parsed:
_Z | N | 5Thing | 3foo | E | v
prefix | nested | `Thing` | `foo`| end nested | parameters: `void`
You can read the constructor names similarly, as below. Notice how the constructor "name" isn't given, but instead a C clause:
_Z | N | 5Thing | C1 | E | i
prefix | nested | `Thing` | Constructor | end nested | parameters: `int`
But what's this C1? Your duplicate has C2. What does this mean?
Well, this is quite simple too:
<ctor-dtor-name> ::= C1 # complete object constructor
::= C2 # base object constructor
::= C3 # complete object allocating constructor
::= D0 # deleting destructor
::= D1 # complete object destructor
::= D2 # base object destructor
Wait, why is this simple? This class has no base. Why does it have a "complete object constructor" and a "base object constructor" for each?
This Q&A implies to me that this is simply a by-product of polymorphism support, even though it's not actually required in this case.
Note that c++filt used to include this information in its demangled output, but doesn't any more.
This forum post asks the same question, and the only response doesn't do any better at answering it, except for the implication that GCC could avoid emitting two constructors when polymorphism is not involved, and that this behaviour ought to be improved in the future.
This newsgroup posting describes a problem with setting breakpoints in constructors due to this dual-emission. It's stated again that the root of the issue is support for polymorphism.
In fact, this is listed as a GCC "known issue":
G++ emits two copies of constructors and destructors.
In general there are three types of constructors (and
destructors).
The complete object constructor/destructor.
The base object constructor/destructor.
The allocating constructor/deallocating destructor.
The first two are different, when virtual base classes are
involved.
The meaning of these different constructors seems to be as follows:
The "complete object constructor". It additionally constructs virtual base classes.
The "base object constructor". It creates the object itself, as well as data members and non-virtual base classes.
The "allocating object constructor". It does everything the complete object constructor does, plus it calls operator new to actually allocate the memory... but apparently this is not usually seen.
If you have no virtual base classes, [the first two] are are
identical; GCC will, on sufficient optimization levels, actually alias
the symbols to the same code for both.