Can a c-style cast call an explicit conversion constructor?
Consider the following code:
class Vec3
{
public:
explicit Vec3(float All)
: X(All), Y(All), Z(All)
{
}
Vec3(float InX, float InY, float InZ)
: X(InX), Y(InY), Z(InZ)
{
}
Vec3()
: X(0), Y(0), Z(0)
{
}
float X, Y, Z;
};
void Morph(const Vec3& In, Vec3& Out)
{
Out = In;
}
int main(void)
{
float Array[3] = {1.0f, 2.0f, 3.0f};
Vec3 Morphed;
Morph((const Vec3&)Array[0], Morphed);
}
On Microsoft VS2010's compiler, it converts the line (const Vec3&)Array[0] to a reinterpret_cast and a copy constructor to pass the parameter to the function. Morphed gets a value of [1.0f, 2.0f, 3.0f].
On another compiler from a third party, it converts (const Vec3&)Array[0] to a call to explicit Vec3(float) and then the copy constructor to pass the parameter to the function. Morphed gets a value of [1.0f, 1.0f, 1.0f].
I verified this by looking at the disassembly from both compilers:
VS2010:
483: float Array[3] = {1.0f, 2.0f, 3.0f};
0000000143E68049 movss xmm0,dword ptr [__real@3f800000 (145DA1F18h)]
0000000143E68051 movss dword ptr [rsp+28h],xmm0
0000000143E68057 movss xmm0,dword ptr [__real@40000000 (145DC9704h)]
0000000143E6805F movss dword ptr [rsp+2Ch],xmm0
0000000143E68065 movss xmm0,dword ptr [__real@40400000 (145DC9708h)]
0000000143E6806D movss dword ptr [rsp+30h],xmm0
484: Vec3 Morphed;
0000000143E68073 xorps xmm0,xmm0
0000000143E68076 movss dword ptr [rsp+58h],xmm0
0000000143E6807C xorps xmm0,xmm0
0000000143E6807F movss dword ptr [rsp+5Ch],xmm0
0000000143E68085 xorps xmm0,xmm0
0000000143E68088 movss dword ptr [rsp+60h],xmm0
485: Morph((const Vec3&)Array[0], Morphed);
0000000143E6808E lea rdx,[rsp+58h]
0000000143E68093 lea rcx,[rsp+28h]
0000000143E68098 call Morph (143E67FB0h)
Other compiler:
184: float Array[3] {1.0f, 2.0f, 3.0f};
0000000001E1F3B0 mov rax,qword ptr [000000000432D1B0h]
0000000001E1F3B7 mov qword ptr [rbp-14h],rax
0000000001E1F3BB mov edi,dword ptr [000000000432D1B8h]
0000000001E1F3C1 mov dword ptr [rbp-0Ch],edi
185: Vec3 Morphed;
0000000001E1F3C4 lea rdi,[rbp-130h]
0000000001E1F3CB call Vec3::Vec3() (0000000003C119C0h)
186: Morph((const Vec3&)Array[0], Morphed);
0000000001E1F3D0 lea rdi,[rbp-140h]
0000000001E1F3D7 vmovss xmm0,dword ptr [rbp-14h]
0000000001E1F3DC call Vec3::Vec3(float) (0000000003C119E0h)
0000000001E1F3E1 lea rdi,[rbp-140h]
0000000001E1F3E8 lea rsi,[rbp-130h]
0000000001E1F3EF call Morph(Vec3 const&,Vec3&) (0000000001E1F350h)
Needless to say, this is causing many problems. Which compiler is correct? The third party code relies on Microsoft's implementation, but we are compiling with a different compiler for a different platform.
As far as the C++ standard is concerned, the conversion of a float & into a const Vec3 & is undefined behavior. Both compilers are right because the code is outside of the protection of C++.
If you want this to be a protected conversion, you have to stop the type punning and do a real copy of the float array into a Vec3.