D Language Foundation June 2025 Monthly Meeting Summary

[Mike Parker]

The D Language Foundation's June 2025 monthly meeting took place 
on Friday the 13th and lasted just under an hour and forty 
minutes.

## The Attendees

The following people attended:

* Walter Bright
* Rikki Cattermole
* Jonathan M. Davis
* Timon Gehr
* Martin Kinkelin
* Dennis Korpel
* Mathias Lang
* Átila Neves
* Razvan Nitu
* Mike Parker
* Robert Schadek
* Steven Schveighoffer
* Adam Wilson
* Nicholas Wilson

## The Summary

### run.dlang.org

Rikki said someone had asked him to bring up that 
run.dlang.io/.org needed updated libraries and compilers. He said 
now was a good time to remind everyone that infrastructure was 
important to the users.

I asked if anyone had spoken to Petar recently, as he was the 
maintainer. No one had. I reminded everyone that I had spoken 
with him before about migrating the service to a DLF-owned 
server. He'd first wanted to rearchitect everything to make it 
easier to manage and deploy before migrating it, as right now it 
was a bit of a pain. I didn't know where he was with it.

Steve said someone had reported an issue that the GitHub actions 
weren't running. He'd thought he'd be able to restart them and it 
would deploy, but it appeared not to have worked. The ideal thing 
would be something where it automatically deployed when you ran 
an action. I agreed, and added we needed to get it onto a DLF 
server at that point.

(__UPDATE__: The service has since been updated, but it's not yet 
configured for easy deployment nor on a DLF server.)

### DFA methodology, risk assessment, and goals

Rikki said that before embarking on implementing a Data Flow 
Analysis (DFA) engine, he had done a risk assessment, as it was a 
lot of work with a high probability the implementation would 
fail. His main concern was whether the community would accept 
such an engine attached to the language by default.

He listed three criteria by which we could evaluate it: 
performance, false positives, and the extent of the language it 
modeled. We'd done the false positive approach with DIP 1000. We 
didn't like it. Pretty much no one was willing to give up on 
performance. That just left the extent of language modeled. 
That's what he was sacrificing to keep it fast and minimize false 
positives.

Walter said there was another option. He'd implemented DFA in 
both the front end and the optimizer. It was a lot harder to do 
in the front end. He suggested Rikki investigate adding it to the 
optimizer.

Rikki said he'd looked into how other languages were doing it. 
Clang Analyzer, for example, was in the front end. That included 
Swift, Objective-C, C, and C++, and they were way ahead of where 
we currently were. It was definitely doable.

Walter knew it was doable, it was just a lot more difficult with 
lower odds of success. The problem with doing it in the front end 
was that the ASTs were far more complex there than they were in 
the intermediate code. Covering all those cases meant that every 
time you added something new to the language, you had to go in 
and account for it in the DFA. That made the burden of work much 
higher. He'd rarely needed to add anything to the DFA in the 
intermediate code because it was much simpler and lower level.

Rikki said the big problem with that approach was that it would 
then be compiler specific. Walter said that was true.

Átila said one possibility was to copy what LLVM did and have an 
intermediate layer that could be shared between compilers. Rikki 
said you were basically working with a front-end DFA at that 
point if you weren't organizing the glue layer. That was well 
beyond what he was doing. He was successfully doing a front-end 
DFA straight off the AST. That wasn't really the issue.

Átila asked what the issue was, noting that Rikki was trying to 
figure out whether it would get merged if he implemented it. 
Rikki said it would have to be merged as part of the 
implementation. He wanted to find out if we could turn it on by 
default in the language. Átila said it wasn't in the language. It 
was a compiler implementation detail. Rikki disagreed, saying 
that there would be attributes eventually. It would be replacing 
DIP 1000 and things like that.

Átila noted Rikki had given Clang as an example of what was ahead 
of us. Were they adding attributes to C++? Rikki said yes, they 
were adding type qualifiers for nullability. So you could do 
non-null as a type qualifier, pointers for function parameters, 
and things like that. It was actually pretty advanced and pretty 
cool how they were trying to make C and C++ code safe to work 
with.

Átila asked why this required changing the language. Couldn't we 
just put some attributes into `core.attributes`? That wasn't part 
of the language. Rikki said he wasn't here to talk about 
attributes today. That was a different ballgame. Átila said 
Rikki's question was about turning it on in the language. Asking 
why it had to be a language thing rather than a compiler thing 
was an obvious response to that, and Rikki had brought up 
attributes in response to that question. Átila didn't think 
attributes required language changes.

Rikki said that for now, he was just concerned with whether we 
could even turn on a DFA that tested the language. Whether 
attributes were in `core.attributes` or in the language didn't 
really matter because he wasn't supporting them at the moment. 
His only concern now was whether we could even turn on a DFA that 
tested the language.

Átila asked if it would throw errors, and noted Rikki was saying 
'in the language' again. He still didn't understand how the 
language had to change to accommodate this. Rikki said if the 
engine were reporting errors, then it would make the language 
fail when it was turned on. The idea was that it would be turned 
on by default. It wouldn't be opt-in.

Átila repeated that he still didn't understand how the language 
was being changed. Rikki said it would restrict what you could do 
based on rules that an average person might not understand, and 
he had to eliminate that. Átila thought he understood and asked 
if Rikki meant that he was trying to turn what is valid code 
today into errors. Rikki said yes, and without any attributes. If 
it had to have attributes right now, then that wasn't a good sign 
for it.

Walter said that when he'd implemented `@live`, he deliberately 
did not turn it on by default because of the tremendous increase 
in compile time. A lot of people use D for its fast compilation 
speed. Other languages have very fast compilation, too, and if 
DFA is on by default, it would be a big problem for us. It needed 
to be optional, either turned on with an attribute or a compiler 
flag.

Átila said that if not an attribute or a flag, then it had to 
somehow be super duper fast. Walter said DFA was slow. It 
required a fair amount of data to do the tracking. That couldn't 
be avoided. People used DMD because it was a fast compiler. 
Taking that away would be very costly. It needed to be optional.

He didn't see a big problem with it being optional. You'd throw a 
switch that turned on DFA and it would compile more slowly, but 
it would find more errors. Wasn't finding errors the reason why 
Rikki was proposing a DFA? Rikki said it was also meant for 
enabling language features like isolated and such that we 
couldn't do now.

Martin said that, if he understood correctly, Rikki was mainly 
after getting first test results, whether any code would break. 
He asked if there were any BuildKite results for Rikki's PR. How 
many projects broke?

Rikki said there was no PR yet because it wasn't finished for 
nullability and truthiness. Once he got it to the point where he 
was happy with it, hopefully in another month or two, there would 
be a PR. It would be behind a preview switch so that nothing was 
at risk. Then we could answer some questions. Did it do false 
positives? Did it introduce a slow down compiling real code 
bases? It was running at the same speed as DIP 1000 on his code 
base. He thought that was pretty good.

Martin suggested temporarily merging a PR with it turned on by 
default just to get some results out of BuildKite, then go from 
there. Rikki wanted to do that. He expected projects in the wild 
to throw out a lot of messages. Because he'd worked so 
defensively in his code base, any message it was spitting out was 
basically a bug in the DFA.

Átila thought there was value to this if it was something 
optional. We could talk about the language stuff like isolated 
later. If it was worth doing, why not do it? The worst that could 
happen was that we made it optional and didn't change the 
language. He didn't see how that would be an impediment to doing 
the work.

Rikki said the only bad outcome for him was that he didn't prove 
if we could have a DFA engine to test the language. It didn't 
matter if it were true or if it were false, he just had to prove 
it one way or the other. Átila said Rikki could do that anyway, 
given that there was value to this no matter what. He could still 
do an experiment to see what happened to the language afterwards.

Rikki said he didn't care if it were removed later or rewritten 
or any of that sort of thing. He just wanted to prove whether 
that fact was true or not. Átila said he wasn't entirely sure 
what Rikki needed if he'd already worked on it and it would be 
good to have anyway. Rikki said he just wanted to make sure 
everyone knew that he'd done his risk assessment. He had a 
methodology in place that was very low risk to the rest of the 
compiler, and we could go ahead and merge it to let people have a 
play.

Martin asked if Rikki had a rough lines-of-code estimate. Rikki 
said it would be around 15K.

Walter said he was okay with merging it as an experimental thing 
or as a development thing. He was doing the same with the ARM64 
stuff. As long as it wasn't disruptive to the main code base. 
Once it started rewriting the main code base or extending it 
everywhere, that would be a difficult pill to swallow.

Rikki said it was at one point in the function declaration class. 
It added one or two values for time trace, and it had an entry 
point in Semantic 3. It was all behind a preview switch. Walter 
said in that case he was okay with it.

__UPDATE__: Rikki's PR for his proof of concept [DFA engine for 
nullability and 
truthiness](https://github.com/dlang/dmd/pull/21965) was merged 
in October.

### Library ecosystem

Adam said that he did a lot of library stuff. He thought we ought 
to be able to build a static library, a `.a` or `.lib`, with DMD 
or GDC or LDC and have it work with all of them in the same way 
that C++ libraries built with Clang could be used with GCC.

He said not being able to was a huge ecosystem impediment. He 
wouldn't say it was a blocker, as given enough time some 
enterprising soul could figure out how to make it work. But why 
couldn't he build something with GDC and then use it with LDC?

Steve said he'd run into this when building the new GC. He was 
building with SDC, which used the LLVM backend so was kind of 
equivalent to LDC. There were ABI differences between DMD and 
LDC. The biggest one he'd seen was that the order of arguments 
was reversed. You couldn't just call D functions between the two. 
All the API had to be `extern(C)` to make it work. He thought 
there were a lot of runtime issues there, too. Even though they 
used the same code, each compiler had slight differences.

He wasn't sure what the right answer was, but wondered if C++ 
compilers were really so interoperable. Adam said he'd done some 
searching and apparently people did it all the time.

Martin said that type infos in the C++ world were very simple, so 
it was much easier than in the D world. With ABI compatibility, 
it wasn't just about calling conventions, but also, e.g., the 
exception handling personality function, which was completely 
different on Windows between DMD and LDC. He had no idea about 
GDC, but it presumably had its own little thing in the MinGW 
libraries or something. The precision of `real` was another place 
where they differed.

The biggest bugger, the argument reversal thing, was something 
he'd changed a while back. Calling convention wise, LDC should be 
compatible with GDC, though he wasn't 100% sure, but the 
situation should be much better than before. Also, LDC supported 
the GDC-style syntax for inline ASM.

The runtime would need the most work. We'd need to have a 
streamlined implementation of some low-level stuff, like scanning 
segments of the binary to figure out where all the globals and 
TLS globals were at, etc. That was a complete mess. GDC had its 
own implementation of some things, LDC had its own, and they all 
diverged. They diverged in how the runtime was initialized with 
regard to how to register binaries with a shared runtime. He said 
the list would go on quite a bit. Getting there would be very 
complicated.

Rikki noted that the compilers had differences in the lowerings 
they did, and he didn't see those being converged.

Adam said this was something where we could tell people to solve 
it by using source libraries. But what we were really saying to 
people who didn't want to use source libraries was, 'we can't 
help you'. Martin disagreed that was the case.

I said they could use the same toolchain. Martin said they could 
use dub because it had separate builds for each compiler. So 
whenever you were using different compilers with it, you should 
be fine.

Adam said to imagine getting a pre-built library from somewhere 
else. In library ecosystems, that happened. We couldn't tell the 
whole world they had to ship their source with everything.

I brought up libSDL as an example. They provided Windows binaries 
for MinGW and for MSVC because of the incompatibilities there. 
They weren't 100% compatible. I didn't see the problem with that. 
If you were distributing binaries and you wanted to support a 
wide user base, then you shipped binaries for each compiler.

Rikki said the simplest solution was to designate LDC as the LTS. 
Then people who wanted to ship binaries could build with that. He 
thought that would be our best bet.

Martin said he'd just thought of something. Our situation was 
much worse. It wasn't just three compilers with quite a few 
differences. It was also each version. You'd have to ship 
binaries for version 2.110, 2.111, and so on, for all three 
compilers. I suggested you could just settle on a minimum 
supported version.

Martin said the compiler had this fixed assumption for the 
DRuntime interface. If you were using a shared DRuntime and that 
interface changed... that was really messy. We weren't so stable 
like C++.

Adam said that added more complexity to your build chain and 
development process. The only answer he could see would be to 
slow down our release cadence to something like once a year. He 
wasn't saying we *had* to do anything about this. He was just 
pointing it out as a libraries guy who was going to be dealing 
with this problem in the future.

Steve had seen it suggested quite a few times that GDC should be 
the LTS compiler because of GCC's requirements that things 
couldn't change too much between versions. He didn't know how 
much it changed between point releases. Maybe it could be an 
option to tell people to use GDC for binary releases. He thought 
LDC was breaking just as much as DMD because they were releasing 
quite often.

Martin said we could take every 10th LDC release or so.

Steve said in a binary release ecosystem, you would probably 
specify the compiler version to use with your binary releases. 
Every so often you would bump it to another release. He thought 
it pretty unlikely we'd get to the level of binary compatibility 
that Adam was talking about.

Nicholas asked if this wasn't exactly what we wanted to use 
editions for. Steve said no. Editions were source compatible. 
Nicholas asked if there were any reason why we couldn't lock 
binary compatibility in with the same thing.

Martin shook his head no. He said implementing something like the 
templated hooks that we had now wouldn't be possible in that 
case. If we had an edition every ten releases or so, the last 
compiler release for a specific edition would still need to work 
with the shared runtime library from the first compiler release 
for the same edition. That could be very, very tricky. Nicholas 
said that at least our intention for what binary compatibility 
should be would be defined in that case.

Adam wondered if we could say that we couldn't do it now, but 
that we would work toward it in a future edition. It was going to 
be a major point for people who didn't want to ship their source. 
A lot of the C++ ecosystem was like that. You never shipped the 
source. You shipped a header file and a binary blob. Steve noted 
that C++ header files often included a lot of implementation. 
Adam agreed and said we'd end up doing something similar, but 
from an ecosystem standpoint, it had to be something the user 
decided.

Rikki said GDC would be a less desirable choice for an LTS. 
Windows and macOS were great examples where LDC trumped GDC. But 
the whole idea that you would tie codegen changes and lowerings 
to the edition was absolutely nuts. It meant if someone hit what 
was for them a crucial bug, we couldn't fix it. And it was going 
to happen and would be a problem for us. The end result was that 
we would have to designate a specific version of a specific 
compiler as the LTS. We'd support that for two years or whatever, 
and if you were going to ship binaries, that would be what you 
used.

He noted that Hipreme shipped a compiler toolchain with his game 
engine. That was effectively an LTS. The problem was solved for 
him. That was precisely how we needed to handle this.

Adam said this was going to take a lot of infrastructure work. He 
wondered if there was any benefit to starting the process of 
normalizing the runtime stuff. He recalled Martin had talked 
about upstreaming the LDC runtime stuff.

Martin wasn't sure what Adam was referring to. He said it would 
require a lot of work on DMD for very low benefit. LDC wouldn't 
be improved in any way. DMD's functionality might improve a 
little bit, but it wasn't worth the amount of work to get there.

Adam asked if we should look into slowing down the release 
cadence. He was thinking that if we wanted to get fancy, we could 
maybe have dub do this. The build manager itself would become a 
compiler manager. He didn't know what our install process was, so 
he was shooting in the dark, but why not treat the compiler 
itself as just another dependency? It could download it and 
extract it and give you a library of compilers to choose from.

Steve interjected with a situation he'd encountered with 
raylib-d. It used the DLL version of raylib, so he'd assumed it 
shouldn't be too hard. He'd found it was almost impossible 
because he had to have the exact same MSVC version that the 
binary had been built with. He'd ended up telling people to 
download MSVC and build raylib themselves.

If you were going to have a binary release of something, you had 
to put in the work to say, 'Here are the entry points. You have 
to use them this way.' And everything inside was kind of 
self-contained. He didn't think there was a way you could just 
use whatever version of whatever library you wanted as a 
dependency. With a template-heavy language, we were never going 
to have that kind of binary compatibility. It just wasn't going 
to happen.

Robert didn't think it was a technical thing. His first thought 
about Adam's suggestion to slow down the release cadence was that 
the next time we were here, we'd discuss increasing the release 
cadence because it was too slow. At some point, we had to tell 
people what we were and what we wanted, and we had to say 'no'. 
Leaving people in limbo about things while we talked about it was 
just prolonging the inevitable. At some point, we had to say, 
'Nope. That's not a thing. You can try this or that, but mostly 
you're on your own.'

He said we could do no releases and be binary compatible forever, 
or we could be really quick and annoy some number of people. The 
D community was not a community in lockstep. We weren't here 
because we all liked the same things and were interested in the 
same ideas or had similar tastes. We were here because we were 
all weird and we were all weird in different ways. So he didn't 
think it was a technical thing. It was an organizational thing. A 
people thing.

Razvan pointed out that the release process right now was really 
cumbersome. We had relied on Iain alone for a long time. Now 
Dennis had looked into the script and it was really, really 
complicated. You needed to manually resolve all sorts of 
conflicts. If we wanted to do more releases, then it should be 
entirely automated. Right now it was half-automated, half-manual. 
We shouldn't need a person doing manual stuff. He'd talked to 
Dennis about it and they thought the script could be slimmed down 
a bit because it was doing things it probably shouldn't be doing.

Steve remembered Hipreme struggling because he'd wanted to have 
all of his stuff using DLLs and binary loading. You basically had 
to cut down to no templates in the API to get all that stuff to 
work correctly. That was something you had to accept if you 
wanted binary compatibility.

Martin said if he were shipping a library in binary form, he'd 
probably specify the compiler version he was using, then anyone 
who wanted to use the library would have to use the same version. 
That was the only version guaranteed to work. So you'd probably 
need to stick with one compiler unless you really wanted to put 
out different binaries. Then it was a set of supported compiler 
versions. The main problem here was when you wanted to combine 
multiple such libraries, or when your code wasn't compatible with 
a specific D version because something changed or something 
regressed.

Walter didn't think C gave you binary compatibility across 
compiler versions. Adam said they released maybe every decade.

Martin said that compiling LLVM with clang and the C++ parts of 
LDC with GCC, they worked fine. You could link them together. On 
Windows, the LLVM binaries were compiled with clang and the C++ 
parts of LDC with the MS compiler. That was also binary 
compatible and everything worked. Everything was more or less 
stable in the C++ world. Ignoring the C++ standards for now.

Adam noted that a lot of people who came to D came from the C and 
C++ world. For him, this was almost a kind of marketing thing. It 
was very easy to have a big backdoor of people who just quietly 
left. Most people wouldn't throw big rants on the forum. They'd 
just walk away. That was part of why he was harping on this.

Also, he'd been reading about ecosystems. He viewed this Phobos 3 
project as the core of the library ecosystem because everyone 
could use it. As the Phobos guy, he had to care about the entire 
ecosystem on top of it. He'd started researching it and had hit 
exactly what we were talking about.

Maybe the answer really was to just say which compilers a thing 
supported. He wasn't expecting to get a solution out of the 
meeting, though it was a really instructive discussion. It was 
something he wanted to keep in mind because it was an ecosystem 
and marketing problem. He added that Robert wasn't wrong. At some 
point we had to say that we just couldn't do that, but we had to 
understand who was walking out the backdoor when we said it.

He said that at DConf in Munich, he'd written a build system 
there in the room. He would have to add things like offering a 
selection of compilers to download and figure out. We would have 
to treat the build ecosystem differently than we did right now. 
That was something to think about.

In the meantime, we needed to be mindful of what we were throwing 
in the compiler. What impact did it have across other libraries 
in terms of ABI? Maybe we needed to get this ABI situation sorted 
out. Maybe we had to pick and choose things over time and we 
could start lighting things up.

He said it was an option to identify the stuff what was and 
wasn't binary compatible. I thought that was an actionable item. 
The first step then would be to go through and identify what was 
binary compatible and decide where to go. Maybe put up a page 
describing the incompatibilities or see if there was a way 
forward to reducing them.

Steve said that you'd need buy-in from the language team that 
they wouldn't change those things if you wanted that kind of 
binary compatibility. For instance, when he had done the work to 
enable static initialization of associative arrays, he'd needed 
to add things to the underlying AA type that would totally break 
when used with a different AA layout.

Recently he'd had a problem where AAs were building fake type 
infos and allocating with the GC. In certain cases with the new 
GC, those type infos were being collected before the AA block, 
resulting in crashes. He didn't understand how the compiler guys 
did these things, but Rainer had set it up so that the compiler 
now built the AA type info for the element type, and that worked 
flawlessly. But that required adding a new piece to the AA type 
info, and that was another binary incompatibility.

If we wanted to achieve binary compatibility, we had to commit to 
only changing things like that on a non-regular basis. Adam 
speculated about doing it on an edition basis. Steve said maybe. 
But the AA thing was a huge bug that was crashing things at 
Symmetry, so they'd needed that one fixed ASAP.

Rikki pointed out that if AAs were templated, all those issues 
across compiler versions wouldn't have existed. That had to be 
taken into account. Just because an implementation detail changed 
didn't mean they'd stop being interoperable between versions.

Steve said that would also cause binary incompatibility. If you 
changed that template, then it would no longer be binary 
compatible. Anytime you had a template, it made things really 
sketchy. Things could change without you even thinking about it.

Rikki said it could, but it wasn't always guaranteed. If you were 
careful, like with a templated class for instance, the vtable 
didn't necessarily have to change. And you could always add 
entries to the vtable without breaking, so older compiler 
versions would see the older methods and newer compilers would 
use the newer methods.

Martin brought up a big problem with shipping a precompiled 
static library that was different for each compiler when it came 
to templates. In that case, if the semantics of the template 
changed so that the mangled name remained the same, the linker 
was going to pick one symbol, probably the first one it saw in 
the first object file it pulled out of these libraries. Binary 
compatibility would mean that whenever we had a mangled name, the 
semantics were never going to change. The implementation might 
defer some implementation details, but the semantics must be 
completely the same, with the expected result, the expected 
exceptions, all of that. This was immensely complicated.

Steve said that was why we ought to focus on things that weren't 
templates if we wanted to guarantee any kind of binary 
compatibility. At least in those cases we could see when they 
would change and they wouldn't change because of some dependency 
somewhere.

I suggested that Adam might gather his thoughts, then somewhere 
down the road put a document together on what testing binary 
compatibility might look like. Then we could discuss it at a 
future meeting and see about getting some community feedback and 
participation on it. Adam thought that was a good idea.

### Assign expressions and with statements

Rikki told us that [Dejan had posted in the 
forums](https://forum.dlang.org/post/jzycjbygpwkvujwqbtrj@forum.dlang.org) requesting we add support for assign expressions in with statements. Why didn't we have it?

I noted that Adam Ruppe had just announced it was in OpenD, so we 
could use that implementation. Walter said he had no issue with 
that. Rikki said he had looked at the code base and wasn't going 
to be the one to do it, but he would reply in the forums that it 
had been pre-approved.

Nicholas said it wouldn't be hard to do. He volunteered to do it. 
Steve thought it made a lot of sense given all the other places 
we had it.

As an aside, Steve didn't recall where he'd seen it, but someone 
had asked about with statements and UFCS. Right now, given a 
struct instance `s` with a method `foo`, inside `with(s)` you 
could call `foo` without doing `s.foo`. That didn't work with 
UFCS functions. So with a function `foo(S s)`, you still had to 
use `s.foo`. UFCS made `foo` look like a member, so inside of 
`with` it seemed like an inconsistency that you couldn't just 
call `foo`. He wondered if there was a reason we couldn't do that.

I asked it if was intentional or an oversight. Walter said it was 
might have been an oversight, but it was more likely that it was 
complicated. The name lookup rules were supposed to be trivial, 
but over time all these features had been layered on to it. Now 
we were sometimes left wondering what was actually happening when 
we used these shortcuts. He didn't know the cause of it not 
working, or if it would break other things or be ambiguous if we 
made it work. He hadn't looked at it yet.

I suggested that was something to consider in the future, but it 
looked like we were good for the assign expressions.

Walter said he looked through the front end source code now and 
then and was horrified at how complicated it had become. He'd had 
a couple of PRs merged recently that simplified some things in 
the front end, but it was like mowing the lawn with scissors. He 
found it a bit discouraging. Sometimes he wondered if we weren't 
driving ourselves into a swamp by adding on more and more 
complexity.

Walter said there were times when he didn't feel like working on 
whatever his main project was at the moment, but he still wanted 
to do *something*. So he'd open up a random source file and 
consider how he could make it better or more understandable. A 
few days before the meeting he'd opened the `outbuffer` module. 
It was using a method called `write`, and he wondered why it 
wasn't called `put`. `OutBuffer` was an output range. Why didn't 
it have the output range API? So he submitted the PR to replace 
`write` with `put`. It wasn't anything major, but it was 
something that reduced the cognitive load for anyone reading the 
code. Now you could see it was an output range and it was one 
less thing to learn in the source.

Another example was that he'd noticed that struct parameters were 
initialized in all kinds of random places. So he'd put all the 
initializations into one section. Little things like that made 
the code base nicer and easier to understand.

He urged everyone to do something similar when they got bored 
with working on their main thing but wanted a break from their 
usual stuff. Just pick a file, look at it, and see what you could 
improve.

### Spreading the word

We had gone through our scheduled agenda items at this point. 
Walter wanted to talk about marketing.

He regularly saw articles on Hacker News written by one or more 
Zig enthusiasts. They were always touting various features of Zig 
as being innovative and new when they were actually copied more 
or less from D.

He found this very upsetting because D had those features 20 
years ago, and suddenly Zig announced they had compile-time 
function execution and everyone was going, 'Wow, wow, wow, 
greatest feature ever! I didn't know you guys were so innovative.'

He said if any of us had any ideas for articles, we should write 
them. They didn't have to be very long. Even just a page of text. 
It was worth doing it just on anything and everything.

He brought up a few other Zig features he'd seen people touting, 
including the `restrict` keyword, and how the people touting it 
hadn't realized it was a C feature that nobody really used 
because nobody understood how it worked. Martin said it had a few 
uses for performance code and vectorization. Then we veered off 
into a bit of a discussion about the pros and cons of `restrict` 
and whether it could be reliably analyzed to determine if it was 
used correctly.

Walter reminded us he'd just wanted to mention that we should be 
publishing more stuff about D.

Rikki told us he'd talked about what D did in some comments about 
compiler theory on Reddit. He usually got some likes on comments 
like that.

I said if anyone wanted to put some articles together for the 
blog, they could send them to me and I'd publish them. I didn't 
have time to be chasing people for articles anymore like I used 
to, but if anyone sent me any D articles, I'd publish them.

We then talked a bit about where to share D articles. I said I 
had always shared our articles on Reddit and Twitter, and Walter 
or someone else would push them on HN. Walter thought Reddit had 
gone into a decline from what it used to be. I said what I liked 
about it was that we used to get really trash comments on most of 
the blog posts I shared there, but over time they got better and 
more positive.

Walter said he used Twitter to post a paragraph about something 
he was working on now and then and he'd gotten surprisingly good 
feedback about things like that. It didn't have to be big. Even 
short posts could be effective, so he was regularly using his 
Twitter account to promote D.

Adam noted Walter's Twitter posts about the ARM backend had 
gotten some traction. Walter had noticed his follower count 
increased after he started posting about that. He thought Twitter 
was a major marketing tool. Emails just went to spam folders, but 
people actively signed up for your Twitter messages. Keeping it 
short was effective. He'd seen other people doing it to good 
effect: John Cook, Bartosz Milewski. He encouraged us to do that.

I added that YouTube was by far the biggest place people were 
searching for anything these days. We had a serious absence of 
content there. Mike Shah had his ongoing D series, but we didn't 
have much going on with our channel. We had the conference 
videos, my interviews, and a tutorial series Dennis had 
submitted. I would love to have videos of people going about 
their daily D work, showing how they're fixing bugs or 
implementing compiler features, that sort of thing.

Adam said there was a guy doing short videos about C# features. 
They were a minute and a half, two minutes long. The guy was 
killing it and had started selling C# courses. He wasn't saying 
we had to make a whole thing out of it, but the guy was doing a 
mix of short-form and long-form content.

I said all it had to be was you just sit down and say, 'Okay, I'm 
going to fix a bug today.' Turn on your screen recorder and talk 
about what you're doing as you're doing it. Now we've got a 
compiler bug fixed and we've got a video showing what it's like 
to program in D. It was real-world D programming. There were 
people out there who would watch that. Whether it got a lot of 
views didn't matter. Ultimately it would help us grow our YouTube 
channel because it would give me content to publish and build 
engagement.

If I could have a video to publish every week from people, not 
just Walter, but anybody going about their regular D programming, 
that would be fantastic. It didn't have to be the same people all 
the time. It could be anything about D. I had a series in mind 
that I'd been wanting to do for a while but just couldn't make 
the time for. I wanted talk about fundamental things like object 
files, static vs. dynamic linking, dynamic loading and things 
like that.

Adam said Google Ads was still huge. You didn't hear people 
talking about Facebook ads anymore, or LinkedIn. But you had to 
be on Twitter and you had to be on YouTube.

I said that it didn't cost us anything to upload videos to 
YouTube. That was as much free marketing as we wanted. Walter 
said it didn't cost us anything to post on Twitter either.

I said one of the big benefits about YouTube was that videos on 
an evergreen topic were evergreen videos. They'd show results 
over time. A tutorial about D or showing how you fixed a bug, 
those were evergreen. They would be as valid a year or two years 
from now as they were today, barring language changes. They 
weren't going to disappear into the ether. So when people were 
searching for dlang in the future, those videos would come up in 
some of the search results.

As an example, I had people coming to my personal YouTube channel 
through search results keeping some of my older videos going. And 
new subscribers often went back through my library to view my 
older content. Those older videos were still working for me.

If somebody saw a video of Walter fixing a D bug and thought it 
was pretty cool, they might go back and look at other videos of 
other fixed bugs. It would be a small number, but there were 
definitely people who would be interested in that sort of thing. 
And that would cause them to go into our back catalog, and maybe 
even fix a bug themselves. And at the same time, it was going to 
be generating money for us through the AdSense program.

Walter said it was a win all around and he understood that. He'd 
given Google ads a try years ago and had zero results. I said we 
weren't even going to try that. Andrei and I had done something 
with it a few years back to promote DConf and it had been a waste 
of money.

Rikki said people in Walter's position sometimes did seminars and 
things like that, taking audience questions and recording it all. 
Those videos got uploaded and people watched them. At some point, 
Walter should see about doing a seminar for Mike Shah's students. 
I noted that Steve and Ali had done that. Rikki said the 
difference was that it would be Walter. Steve said, 'Thanks.' 
Rikki said people were always interested in meeting someone with 
a lot of experience, having a chat about some interesting project 
the students were working on, soliciting advice, maybe having 
some sort of presentation.

Adam added that Ben Jones at Utah University was someone else we 
shouldn't forget. Adam had recently passed through Salt Lake City 
and had the opportunity to meetup with Ben and one of our GSoC 
students. Ben was teaching his CS and engineering classes in D. I 
noted that Ben was going to be at DConf.

Steve said that every month we had the online BeerConf. He always 
posted an open invitation for presentations. In the past, he'd 
recorded them and posted them on his YouTube channel. We could 
put them on the DLF channel, too. I said if the speaker was okay 
with it, then I'd be happy to publish them.

### Quick items

I said I had to announce Symmetry Autumn of Code at the end of 
the month or in early July. We needed to start thinking about 
projects. I was going to talk about that with Razvan and Dennis 
soon in one of our team meetings, but asked everyone to please 
let me know if they had any ideas for SAOC projects.

Another thing was shop.dlang.org. Someone had pointed out in the 
Discord a week or so before that it was down. I'd been unable to 
get it back up. I'd done everything possible within my sphere of 
knowledge, then reached out to Vladimir for help. He'd been able 
to determine that it wasn't anything on our end. It was something 
to do with either CloudFlare or Fourthwall, the provider we were 
using for the shop. The next step was to reach out to support at 
both places to see what could be done. In the meantime, I was 
going to disable the custom domain at Fourthwall and just go with 
dlang-shop.fourthwall.com.

(__UPDATE__: My support requests got me nowhere, so I ended up 
going with `store.dlang.org`, which has had no issues so far. If 
you're looking for D swag, that's where you can get it now. 
Proceeds go into our general funding pool along with all 
donations made via PayPal and GitHub.)

### Conclusion

Our next meeting was a quarterly on July 4th. Our next monthly 
meeting was on July 11th.

If you have something you'd like to discuss with us in one of our 
monthly meetings, feel free to reach out and let me know.