• Teodor Dutu (29/29) Nov 29 2022 Hi,
• rikki cattermole (6/6) Nov 29 2022 Hopefully this work will mean that some code will be removed from rt.
• Teodor Dutu (6/8) Nov 29 2022 Absolutely. All these PRs add new template hooks and also remove

Teodor Dutu <teodor.dutu gmail.com> writes:

Hi,

These weeks I fixed the final bugs of the new 
[`_d_arraycatnTX`](https://github.com/dlang/dmd/pull/14550) hook. 
When cleaning up the code, Razvan and I took a closer look at the 
[extra lines produced by 
`-profile-gc`](https://github.com/dlang/dmd/blob/9c6e6b3f95a9cc9b873a4d61b58a5e6f0e8101e3/druntime/test/profile/myprofilegc.l
g.linux.64.exp#L2). They are referring to the new `_d_arraycatnTX` and seemed
to indicate that it allocates some extra memory. In [this forum
post](https://forum.dlang.org/thread/uyjolkupyabdilczdain forum.dlang.org) I
asked how to deal with them. Following Max's advice, I did some digging and 2
things turned out:

1. the [old 
hook](https://github.com/dlang/dmd/blob/a90ed729205e6faeb2803fbc70733d5ec6d58701/druntime/src/rt/lif
time.d#L2339-L2373) uses a private function called `__arrayAlloc()` to allocate
the concatenated array, which is not instrumented, whereas the new one simply
sets the `.length` of the resulting array. This expression is further lowered
to `_d_arraysetlengthT` and this hook [**is**
instrumented](https://github.com/dlang/dmd/blob/a90ed729205e6faeb2803fbc70733d5ec6d58701/druntime/src/core/internal/array/
apacity.d#L57-L64). It is possible that this is the cause of the differences.

1. My implementation of `_d_arraycatnTXTrace()` was buggy. This 
caused some unnecessary bytes to be allocated. After fixing this, 
I'll be able to better look into point 1 from above and be more 
certain regarding the newly added lines when `-profile-gc` is 
enabled.

Regarding `_d_newitemU`, we decided it would be more rational to 
move more of the lowering logic to the hook, as DRuntime already 
handles [`struct` and context pointer initialisation in 
`copyEmplace()`](https://github.com/dlang/dmd/blob/a90ed729205e6faeb2803fbc70733d5ec6d58701/druntime/src/core/life
ime.d#L1274-L1280). I am going to use the same logic in the new `_d_newitemT`
(instead of `_d_newitemU`) hook, together with calling the `struct`'s ctor.
What's problematic with this approach is that I need to make the compiler
create a "dummy" object before calling `_d_newitemT`, so the lowering would
look like this:

```d
// original expression:
S* s = new S(args);

// lowering:
S* s = (void[0] ctxPointerTmp, _d_newitemT!S(ctxPointerTmp))
```

I would like to be able to obtain the context pointer in the 
frontend of the compiler without having to allocate any other 
objects. I asked about this in 
[this](https://forum.dlang.org/thread/mdybepujnpglwqyztnya forum.dlang.org)
forum. I am looking forward that the community will help me find a cleaner
solution.

Thanks,
Teodor

rikki cattermole <rikki cattermole.co.nz> writes:

Hopefully this work will mean that some code will be removed from rt.

That bunch of code is a real problem.

Right now no one can implement a GC without compiling it into druntime 
or reimplementing stuff in rt. Not a great situation!

It also contains my most disliked set of hooks, memset as they could be 
very easily removed entirely.

Teodor Dutu <teodor.dutu gmail.com> writes:

On Tuesday, 29 November 2022 at 21:04:22 UTC, rikki cattermole 
wrote:
 Hopefully this work will mean that some code will be removed 
 from rt.

Absolutely. All these PRs add new template hooks and also remove 
the old ones.
And most of the old hooks are in rt. For example, all the old 
hooks I am working to replace now are in rt/lifetime.d.