https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115413
Bug ID: 115413
Summary: Missing optimization: devirtualize the call in
"if(typeid(*a)==typeid(A)) a->f();" structure
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: user202729 at protonmail dot com
Target Milestone: ---
Created attachment 58397
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58397&action=edit
Code to reproduce the example where gcc devirtualizes the call.
As mentioned in the title.
I think code of the form
if(typeid(*a)==typeid(A)) a->f();
where `f` is a virtual function in `A` can be optimized to eliminate the
virtualization.
(Of course in practice it won't gain much performance because `type_info`
comparison involves a string comparison which is slow, but if the `else` branch
is `__builtin_unreachable()` it could potentially result in a performance
gain.)
As pointed out in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115410 ,
however, the address of `a->f` may not be the same as `A::f` if `a` comes from
a shared library, however gcc already does that optimization in some cases e.g.
A a = create_object();
a.f();
where `create_object` is loaded with `dlopen()`, then the address of `a.f` will
actually be different from `A::f`, nevertheless gcc optimizes it to call `A::f`
anyway. Because of ODR, the behavior must be identical.
Example: in the linked file, compile with
g++ -fPIC -shared -o libshared.so shared_lib.cpp -Wno-pmf-conversions
g++ test4.cpp -ldl -Wno-pmf-conversions
./a.out
then the 4 addresses will not be identical. (Note that the program technically
invokes undefined behavior by referencing the address of
`(void*)(&MyClass::hello)` which I believe violate ODR; nevertheless, it just
serves to show which copy of `hello()` actually gets called.)
