https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93153

--- Comment #3 from ZhuoYang <logicyangz at gmail dot com> ---
Created attachment 66180
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=66180&action=edit
GCC 16.2.0 reproducer bundle: full ccp1 dump, preprocessed source, and compiler
version

I reproduced the behavior described in this PR with Homebrew GCC 16.2.0 on
aarch64-apple-darwin25. The second call uses the pointer returned by placement
new, converted to A* when passed to call().

Reduced testcase (repro.cc):

#include <cstdio>
#include <new>
struct A { virtual int func() { return 1; } };
struct B : A { int func() override { return 2; } };
int call(A* a) { return a->func(); }
int main() {
    A a;
    static_assert(sizeof(B) <= sizeof(A) && alignof(B) <= alignof(A));
    int x = call(&a);
    B* b = ::new (&a) B;
    int y = call(b);
    std::printf("%d %d\n", x, y);
}

    g++-16 -std=c++23 -O3 -fdump-tree-ccp1-details=ccp1.txt repro.cc -o repro
    ./repro

Expected output: 1 2
Actual output at -O3: 1 1

GCC 16.2.0 prints 1 2 at -O0 and with
-O3 -fno-devirtualize -fno-devirtualize-speculatively.

In early_objsz, the second call is still an OBJ_TYPE_REF. The ccp1 dump shows:

    Folding statement: _28 = OBJ_TYPE_REF(_27;b_16->0B) (b_16);
    repro.cc:5:32: optimized: folding virtual function call to virtual int
A::func()
    Folded into: _28 = A::func (&a);

The GIMPLE after ccp1 still contains the store of B's vtable before the second
call.

Reply via email to