Permalink
Reply
On Tuesday, 9 April 2024 at 00:14:01 UTC, Walter Bright wrote:
>On 4/6/2024 9:04 AM, Carl Sturtivant wrote:
>Don't straw-man this: none of this says that you HAVE to use GC in D. There is no reason to oppose an ace GC for D on those grounds.
Consider the performance deficit from write-barriers.
Assignments to a pointer probably should call DRuntime handle. As code example, any time one assigns to a pointer (including dynamic arrays, class references, delegate contexts and so on):
int* ptr;
ptr = new int(35);
...it would be treated as (without doing the safety checks):
*cast(size_t*) &ptr = __ptrWrite!int(cast(size_t) ptr, cast(size_t) new int(35))
The DRuntime definition for present behaviour would be simply
pragma(inline, true)
size_t __ptrAssign(Pointee)(size_t oldVal, size_t newVal)
=> newVal;
and this would remain an option, but alternatively __ptrWrite can have a write gate. If the write gate would just ignore any pointers not registered to the GC it would probably continue to work with existing code with no changes other than a slight slowdown.
This would enable a tri-color GC, and it's better performance could easily win far more than the write gate overhead costs in most cases. The existing mark-and-sweep collector can still remain an option for those who don't want write gates.
Also even when the write gates are used you could still forgo them explicitly in unsafe code to by writing *cast(size_t*) &ptr = cast(size_t) newValue. Or probably use a less ugly library function to do the same.
