Discovered this through some other issue when finishing up the latest frontend merge.


In the following example, I feel that this should not compile at all.

void main()
{
  int[int] a;
  void* b = cast(void*)a;
}


Given that the definition of an AA is (as defined in rt.aaA):

struct AA
{
    void* a;
}


Similarly for this code:

void main()
{
  int[int] a;
  void* b;
  a = cast(int[int]) b;   // fails
  a = cast(int[int]) &b;  // works
}

Correct me if I'm wrong, but both examples should warrant bug reports.


NB: These examples don't compile on GDC because of the safe guards from directly converting between scalar and pointer types.


Regards
Iain

On Friday, 17 August 2012 at 22:15:10 UTC, Iain Buclaw wrote:

> NB: These examples don't compile on GDC because of the safe guards from directly converting between scalar and pointer types.
>

s/scalar/non-scalar/

Iain Buclaw:

> Correct me if I'm wrong, but both examples should warrant bug reports.

Accepting the cast of an AA to void* seems bad.

Bye,
bearophile

fwiw, I raised a bug about this on Orange a while back: https://github.com/jacob-carlborg/orange/issues/17.

I thought it a bit odd that it compiled in DMD at the time, but i don't think i opened a bug about it.

On Friday, 17 August 2012 at 22:15:10 UTC, Iain Buclaw wrote:
> In the following example, I feel that this should not compile at all.
>
> void main()
> {
>   int[int] a;
>   void* b = cast(void*)a;
> }
>
>
> Given that the definition of an AA is (as defined in rt.aaA):
>
> struct AA
> {
>     void* a;
> }

I also once encountered some cases where the DMD frontend would allow conversions between structs and scalar types or other unrelated structs which happened to have the same member layout – which broke LDC, as the LLVM IR is typed.

David