--- import std.stdio; class C { int n; this(int n) { writefln(" this(%s) at %s", this.n = n, cast(void*)this); } ~this() { writefln("~this(%s) at %s", n, cast(void*)this); } } void main() { int i; writefln("Stack is at %s", &i); writefln("Heap is at %s", (new void[1]).ptr); { C cHeap = new C(0); // will be destroyed on scope exit scope C c0 = cHeap; // C(1)-C(4) will be allocated in heap // C(1), C(2), and C(4) will be destroyed on scope exit // C(3) will be destroyed on garbage collection scope C c1 = cast(C)cast(void*)new C(1); scope C c2 = true ? new C(2) : null; scope C c3 = (new C(3), new C(4)); } writefln("after scope"); } --- As a result even if `C` is a `scope class` the program will compile without `cHeap` and `c0`, but every `C` instance will be allocated in heap and C(3) will be destroyed on garbage collection.
(In reply to Denis Shelomovskii from comment #0) > --- > import std.stdio; > > class C { > int n; > this(int n) { writefln(" this(%s) at %s", this.n = n, cast(void*)this); } > ~this() { writefln("~this(%s) at %s", n, cast(void*)this); } > } > > void main() { > int i; > writefln("Stack is at %s", &i); > writefln("Heap is at %s", (new void[1]).ptr); > { > C cHeap = new C(0); // will be destroyed on scope exit > scope C c0 = cHeap; > > // C(1)-C(4) will be allocated in heap > // C(1), C(2), and C(4) will be destroyed on scope exit > // C(3) will be destroyed on garbage collection > scope C c1 = cast(C)cast(void*)new C(1); > scope C c2 = true ? new C(2) : null; > scope C c3 = (new C(3), new C(4)); > } > writefln("after scope"); > } > --- > As a result even if `C` is a `scope class` the program will compile without > `cHeap` and `c0`, but every `C` instance will be allocated in heap and C(3) > will be destroyed on garbage collection. The code now allocates C(0) and C(1) on the heap and C(2) on the stack. I could not test C(3) and C(4) since comma expressions are no longer supported. I think that this behavior is fine since a scope class does not necessarily mean that it will be allocated on the stack. The spec specifically tells "When a class is constructed with new and assigned to a local scope variable, it may be allocated on the stack and permitted in a @nogc context." [1]. Since the compiler does not do any dataflow analysis, it is hard to understand whether rhs directly calls new or it does other shenanigans (like the casts in the c1 initialization). Therefore, I think that this bug report is invalid at this point. [1] https://dlang.org/spec/attribute.html#scope-class-var