Comment #0 by snarwin+bugzilla — 2022-12-14T19:02:26Z
As of DMD 2.101.0, the following program compiles and runs successfully to completion:
---
struct S
{
this(ref S)
{
assert(0);
}
}
void main()
{
S[] a = [ S() ];
auto p = a.ptr;
// append until reallocation
while (a.ptr == p)
a ~= S(); // no assert
}
---
This program should assert at runtime when the array is reallocated, but does not, because the copy constructors of the array's elements are not called.
If the copy constructor is changed to a postblit (`this(this)`), the program asserts at runtime, as expected.
See also issue 20879, which affected .dup and .idup.
Comment #1 by schveiguy — 2022-12-14T19:20:15Z
Just to give some clearer picture of how this causes issues:
```d
import std.typecons;
import std.sumtype;
import std.stdio;
import core.memory;
void fun(RefCounted!string foo) {
SumType!( RefCounted!(string) )[] foo_storage;
foo_storage ~= SumType!( RefCounted!(string) )(foo);
foo_storage ~= SumType!( RefCounted!(string) )(foo); // reallocates, but no copy
foo_storage = null;
}
void clobber()
{
int[1000] x = 0x123456;
}
void main(){
RefCounted!( string ) foo = RefCounted!( string )("blah");
fun(foo);
clobber();
GC.collect(); // collect the items already in the GC
writeln(foo);
}
```
What happens here is:
1. An array of one element is created with the sumtype.
2. The array cannot hold a second element, so it's reallocated. However, the copy constructor of SumType is *not* called. Therefore, `foo` has a ref count of 3, but there are afterwards 4 references (the single element array, the double-element array, and the original foo on the stack).
4. clobber makes sure the stack doesn't point at the arrays.
5. GC.collect cleans up 3 foo elements, reducing the count to 0 and deallocating.
6. The writeln is now writing garbage. on my system it just spewed random memory to the screen.
Comment #2 by teodor.dutu — 2022-12-16T17:12:25Z
a ~= S() is not supposed to call either the copy constructor or the postblit because the newly appended element is supposed to be created in-place inside the array. This test ensures this behaviour (only the constructor is called, without the postblit): https://github.com/dlang/dmd/blob/d405b343e301e5c37a90e66131b3f4d588bed2f2/compiler/test/runnable/sdtor.d#L2244-L2245
The issue comes from copying the array during reallocation. The postblit is correctly called while the cpctor is not. The following code prints "loop" 15 times before printing "pblit" 15 times when the array is reallocated:
---
struct S
{
this(this)
{
writeln("pblit");
}
}
void main()
{
S[] a = [ S() ];
auto p = a.ptr;
// append until reallocation
while (a.ptr == p)
{
writeln("loop");
a ~= S(); // no assert
}
}
---
Comment #3 by schveiguy — 2022-12-16T19:47:17Z
Yes, the problem is the copying of the existing data.
I can think of a few ways to fix this, but they aren't pretty. We either have to add more special members to TypeInfo, or change the hook so it can explain to the runtime how to call the copy constructor.
Comment #4 by robert.schadek — 2024-12-07T13:42:18Z