.Net Foundation Design Reviews, 10 Jul 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: GitHub Quick Reviews

Description

00:00:00 - Needs Work: Mark Windows-specific APIs https://github.com/dotnet/designs/pull/138
01:42:25 - Approved: Add static hash helper methods https://github.com/dotnet/runtime/issues/17590#issuecomment-656829666
01:42:35 - Approved: Expose a ValueRef property on LinkedListNode(T) https://github.com/dotnet/runtime/issues/37681#issuecomment-656831777
01:46:48 - Approved: Make IncrementalHash.CreateHMAC ReadOnlySpan(byte) key public https://github.com/dotnet/runtime/issues/35869#issuecomment-656833490

A

Let's do this alright, so I wrote this proposal last night, so it's effectively something we talked about before, which is we have these new fancy attributes that allow us to mark api's as being platform-specific, so what I did is I used the the data file that we used for the platform, compat analyzer that we shipped a few years ago as a starting point, and so what I did? Is they basically just intersected dad with the assembly information from dotnet, core and I?

A

Basically, just first started with a bunch of assemblies that I think should be just marked wholesale as being window specific, as opposed to putting attributes on specific types. I would put them on just the entire assembly, which makes it a bit easier for us to reason about, and then the other thing I thought about was okay. What version annotation? Are we actually going to use because you know fundamentally, what you would normally do?

A

Is you would say: okay, I call in a West Pacific API, which means you know my feature that uses this API is only available when the underlying API is available, and you know for witness that would mean we would hunt down the P invokes. You would find the Windows version that introduced set. P invoke and I would just basically forward that information inside the attribute, since Donnan core doesn't support anything lower than Windows seven and pretty much everything that we use was introduced before Windows 7 or in Windows 7.

A

In other words, we don't have any usage really of api's that are newer than that. I would just suggest. We just use Windows 7 instead of hunting down the version number, because it basically does the same thing for us, and so basically what I want to do in this meeting is I want to review effectively the assemblies that I proposed. We are marking and also some of the types and methods that I am proposing, marking any questions so far. I.

B

Just have a general comment on the windows. 7 thing I realized that Windows 7 is the only thing still in support and I realize that dotnet core in particular doesn't support. But if we're defining the API surface- and the attribute is part of that, then I think it would be better to formally define that as the correct thing.

B

A

Would you say, there's someone.

B

Else someone else could implement, for example, win32 registry and have it support and actually run on Windows XP.

A

Sure, but like how would that work when we don't support them on time in Windows, XP.

B

They're people do all kinds of crazy things, including compiling core RT, to run on UEFI, so you can run bare metal, dotnet.

A

So I think my stance would be if somebody is willing to make port or runtime to Windows XP, then they're more than welcome to back for the attributes as well. I would not optimized for the ok I mean if somebody wants to change the version numbers I'm not going to block that I. Think if somebody says oh actually, this was introduced in I. Don't know 3.5 then sure, like I'm I mean I'm much I'm not see I.

A

Don't think I see much value in this, because most of these API is are basically as far as we are concerned. You know available forever and it's really more for things where the mainstream user will see different versions in practice right versus. In our case, we can pretty much destroy that ninety-nine point: nine, nine, nine, nine, nine, nine, nine nine nine percent of our customers will only ever see two in seven or up because we don't. We don't support anything lower than that.

A

A

Think, what's more you're seeing is WinRT because there's some things he n-double.

C

To go because we don't have a moniker to acknowledge the split between NT and whatever the other mode was 9x.

A

C

5 because that's when the you.

A

C

Old line had died.

A

Yeah, and that gets most of the stuff is probably from that era. I.

D

Wholly agree with the emo there's, not a lot of value or in spending any more time on anything less than seven yeah.

C

It's just not worth hunting down. If somebody else wants to do it for correctness later, then sure we would look at the PR yeah.

A

The parlament practices why it's hard it's. Basically, you have to hunt down the actual P in vogue and then intersect that with the Windows version information- and that's- maybe not it's.

B

Actually, automatically from the windows API Docs I'm, using it for P invoke generation that I do, for example, yeah.

A

The problem is Emeka is like so the sir just to give some insight into how this works.

A

So when I, when I roll platform compared what I did is I basically wrote a tool that that looks for a platform, not supported exception in the code base and then effectively bubbles dead up to the public entry points, so I'm not actually looking at peen books at all, so like the the database, doesn't know which P invokes this thing called, which, if you wouldn't it so basically when I do that, then you would have to effectively redo the analysis to a certain extent and that's not trivial, because the analysis was also you know, hand authored and hand cherry-picked, because the tool would over report or under report, and then we just manually edit or remove things from that set.

A

So like I think right now it would be very manually, but somebody would actually have to go through, let's say registry and say I, recalling any registry API server. You know in version X or version X, plus one right and like I mean we can probably do this if you really cared but like, as I said, it seems I.

B

Was just saying given where the base frame work, it would generally be good to be more correct when possible right.

E

B

Have that or at least allow people to go and fix it if they felt they wanted to spend the time themselves. I.

F

D

I view this really as a we support this on this OS version, miss or greater, and that's exactly what we're saying when we probe in seven this.

G

Morning, yeah, where.

D

Is this works I'm supported here and hop, and we do we don't support this on XP.

B

Right, but that brings in the question of when Windows 7 goes out of life in two years and we no longer ship a a runtime for it. Are we going to go and update all these to be whatever the latest is we could yeah.

A

We could I mean, on the other hand, old same argument applies, but that seems like a lot of work for very little gaming.

D

Right, it's it's busy work for no gain, but there would be nothing wrong with us doing so. Yeah.

A

I would agree with that too. Yeah.

D

If I were writing an application targeting a runtime that runs on an unsanctioned platform, I would just disable the analyzers which are performing these Chuck's.

A

Yeah, although that argument is suspect in general, right like if you say like oh I, used this feature and then my fixes that turn off the feature. Then it's not necessarily and the right thing to do, but yeah I.

D

Just I wouldn't guard my own chucks I would just call the API is because I happen to know that the environment, the apps running in, would work and if there's an analyzer saying hey, you didn't guard, this I would suppress the analyzer right.

C

A

Right so, let's, let's just go over the list, because there's lots of stuff here so for the assemblies so basically I did is I took the obvious ones that I knew about so basically the registry registry access control system events should be pretty much uncontroversial, I think event, lock has also Windows only because it's tied to the Windows Event log mechanism. So a perf concert.

D

There been a bunch of comments on the issue itself which aren't showing up in your view here. Yeah.

A

Because that's that's github for you! So I can open another tab for this. Let me actually find the PR for.

D

This so feels duplicative.

A

Actually, I can probably do this right. I can probably go here right click and then control click on that guy right, yeah, look, I, know how to browser and all right, so so directory services protocols is now supported on Linux, so I will take this out.

A

So that means directory services and directory service account. Management is still not correct.

D

That's my understanding.

C

I account management is, is not supported anywhere other than windows in this release. I, don't know what is in system directory services versus system directory services protocols, so we may need to just follow up on system directory services study a lot.

C

But Joe made one of them work.

D

Yeah we made the one that contains LDAP, which is the system directory services protocols. We made that one work that.

A

Makes sense to me.

D

To be clear, anymore, I was just suggesting, rather than looking at this view. I was suggesting and looking at the changes view and then you'll see the full list, plus the comments interspersed yeah.

A

We can do that, I mean as long as you look at markdown, but yeah did polyworks.

A

Yeah because I have taste.

A

We wait until github. Finally, when is this thing: okay, yeah somebody was asking: why did I check in the excel file? Funny yours, because I just had the same problem for magnets and a blog post, where I did not check in my previous excel file and when I come back six months later, it's nice! If I, have the analysis still for this, and this is the design sweeper, so I personally don't care, but if somebody feels strongly about it, I don't have to check it in there.

A

So let's is so director says you just handled.

A

Yep, that's the same here.

A

So I don't know what drive infer showed up but like a drive in for apparently had at least significant ApS that didn't work.

D

That no should be broken up because in general it's.

A

A

Then that's easy yeah, so Mina should be handler. We don't ship a net five version.

A

They will probably prevent us from annotating it because they attribute on exists in at five I.

C

Mean if we think it's enough value, then we can publish a new version of or we can Rev.

D

The package the attributes need to be public, can it can they just be added to the assembly as internal like we do it knowable and.

A

Premium principle: yes, because we own the analyzer, and so if you don't have to look at the visibility of their tree, I mean in general, like you have to go out of your way to not make it work anyway, because I mean unless he was love, symbols, or something like that. So you might already know.

D

We should be on the same plan as we are with nullable, as we are with the link or attributes which is he doesn't care what assembly they come from. It just cares about namespace and name yeah.

A

That's fair and.

D

Then it's just compiled and assembly I mean.

A

Is the any reason leadership in that five version like? Why are we back porting this guy.

D

We just targeting that standard. Oh, oh I,.

A

C

At this point, it's possibly just harvested.

A

D

Change this to it, we still build it that standard, oh so, say.

A

Yeah, so we should have net standard compliant to use for this I guess! That's! That's the bottom line here so yeah.

C

I mean it may actually take some work in the analyzer to support locally defined attributes, because if it starts off with get typed by name, then it needs to accept that there'll be multiple with those multiple types with that name.

A

Well, the way we would do it is similar to inaudible right, but I see what we would do is you would basically define the attributes as internal and then basically just have them in their library defined by itself right, and so they don't pollute the namespace for consumers, but when the unless.

C

A

The the methods will still find them.

C

Who has been working on if it starts off with the grab the type symbol and then compared.

E

With attribute.

C

Declarations against that type symbol, that'll need to be rewritten to ask each attribute what its name is. Yes, it has to do the slow check instead of the fast jack.

C

So good thing we talked about that because it changes your spec.

A

So it's this and one time in terms of Windows runtime, you know I might ship this guy interesting.

D

That's the change for five. We just stopped doing that.

A

So I guess, then we don't need it.

C

Anymore, is that.

A

A

Is it just because we don't support what is it called now when you I yet, or is it so I'll be going to do it later.

D

There's there's a whole new infrastructure and it all moved to a separate repo watch.

E

A

So basically we should. We should look into this and then decide what every do it over there or we don't do it at all.

C

So their concept still exists, but ownership and code base moved well. Color base is entirely different concept, moved.

E

So, and that would make this one out of scope for this particular PR then, and we fork another or create another ratio over there right.

A

Which particular thing the Windows runtime.

E

Or the Windows runtime just make that.

A

Yes, I basically just deleted it. Our macro just deleted and say, looks like we don't ship it anymore. I. Think, honestly, everything that requires what that is tied to Windows runtime would probably need different and attributes anyway, because when seven wouldn't be super useful, there.

A

So that sounds yeah this one didn't show up, because this one was probably added after we did sorry I'm looking at the cryptography one open, SSL I think this one was probably done after platform compared did this pin so I only know the ones did well.

C

D

C

Mean what would we call it for Linux right? There is no Linux version, it's also not Linux. Only it's not Windows it'll work on FreeBSD currently is it works, and it's supported on Mac, though I intend to do the paperwork to make that not be the case. Insects, yeah,.

A

F

A

So let me quickly talk about this, because that's the other thing we just brought up with blaze or the other day, so pretty in principle. The way the attributes work right now they assume that if something isn't marked it's cross-platform and if something is marked, then it's only supporting the ones that are being marked right.

A

So because that's why I said earlier, this is only really meant for things that are clearly platform, specific things that are partially portable, like reflection amid or something else that is portable too many operating systems that doesn't really work super nicely with these attributes, because then the the set of attributes or the set of checks that we expect the caller to do would vary over time. Right versus registry will always have the same check.

A

So this the check itself is stable right as soon as you promote fragile checks, where you have to compare for this operating system or that operating system on that operating system, you're, basically back to where the problems were with brows, that checks in JavaScript right, because it's a moving base. Basically this.

D

One just everything other than Windows.

A

Yeah, so that's.

D

What I'm saying.

A

So we don't have that capability today. So for Blaser, the same thing came up because Blaser has a similar problem, where you don't really want to mark everything as supported on Windows. You know Mac and basically everything but.

D

A

That's what I'm saying so like basically for Blaser, where we are talking about doing that, because you really want to execute, not include the challenge with that is, though, that it's still somewhat unclear what the check would be. We would basically just have a you know if not this operating system check, which would probably work reasonably well. We don't have that yet so, like that, that's why? For and also the nice thing is, this is really an analyzer realistically.

A

So it would be something that we can do effectively every time the SDK ships we can update their analyzer. So that's something that we can do if we don't fit it into five or we can do it post five. Four, that's why I'm focusing here on the windows only once, because you know they're, easier and they're kind of stable as well. I guess.

D

I'm not I'm, not understanding we can. We can update the SDK anytime, the SDK ships, but right.

E

D

It requires attributes in the libraries. How do we update that until a year from now? Oh.

A

Yeah, so that's so that's the thing we talked about the blazer guys. So the way that this would work is, you would save between five and six. So all, let me say differently so today the problem is we. Yes, we can up-to-date the kids, often as we want. We kind of the de-reference assembles every time you want, because we don't have a servicing story for those right. The only way we can update them as when we ship a new version of.net, which basically means donate $5 at 607.

A

So the way we would do it is, you would say similar to have that from competitor burger, but just say the you know. Until we can actually physically put an attribute in the reference assembly said, we would ship a sidecar file for the analyzer. That just is a list of API as it knows about, and then we would churn that one and then you know next time we can actually bet the reference the same miss you would convert them into actual attribution into the references that we said.

A

That's how that would work and.

D

Each ship, the sidecar file as part of the SDK yeah.

A

You would shiver literally with the analyzer, so you would still not ship out-of-band like a nuclear package, but you would ship effectively it as part of the SDK, and so we can touch it as often as we can touch the SDK and.

D

Presumably, the design would be that sidecar file always wins. It overrides anything. That's in libraries, yes,.

A

Because it would basically be the you know the thing that versions more frequently, so it's assumed to be later so.

D

I just ask that we have that design like fully buttoned down before we ship. That's part.

A

In what way to basically your concern is that, because the design is not locked down, we may not be able to.

C

A

Yes, so like that's what I'm saying but yeah.

E

A

Mean we can talk offline but like in my understanding, is that it wouldn't change anything we do in five all right. It will just change the way the analyzer other the sidecar file would work which none of them would ship for five all right. My.

D

Point is we're saying we're gonna punt on some things right now for dotnet 5, because we don't have a way of doing it and the way we'll address that until dotnet 6 is via doesn't yet to be designed mechanism that we have a hand-wavy plan for and.

E

I'm sewing it'd, be.

D

Great to have it more than hand-wavy, because we're basing not doing something done at 5. On that thing, existing fare.

A

Like I'm yeah I just want to what the fire drill right because it so the thing is for for things like this one. Here we now effectively. This doesn't for this only works on sorry. This doesn't work on Windows. Those things are easier like for blaze in particular. The problem is that the set of API is they know about is, is is still churning right.

A

So, like my confidence that we can just a note ate them in five four, even if we said let's, let's spend resources on this is not very high that we get them right anyways. So that's why I'm saying to me they're, getting it done by five I kind of assumed that we don't have the time to do that. So no matter what we do and that's why I don't think, there's a fire roll, that's necessary, because we can do it later.

A

I agree with you that the design should probably be more fake than hand-wavy, but given the experience of platform compound I think I have a pretty good design in head. That will work for that. So I'm not very concerned about that, but yeah I can I can totally write it up. This was my goal for next week. Anyway, all right, so then for the OpenSSL one. Let me just record the fact that it would be nice to have a negative marker in there somewhere.

A

That was under.

F

C

Mean it does also raise the potential problem of anything that has version it needs this kind of version attribute on a Linux, because what does Linux mean in that case we certainly or I assume we don't want to paint versions of every distro ever then. Maybe the answer is we just hope we don't run into it. I guess.

A

It depends on all you see this right and the way I would do it is I would basically say we what we would annotate those things. We know that are black and white right, so we know for a fact. This thing doesn't work on Windows right. We don't know the actual Linux distros red all works on, so we would basically not flag it and then just you know, potentially get a runtime error for those right and then basically we would just you know, see who plays I might just mark.

A

This thing doesn't work in Blaser and if something else, ships like watch that doesn't support them either, then the similar thing would apply. Yeah.

D

Just to promote something that was mentioned earlier, like the the fact that it doesn't work on Windows now it doesn't mean that it won't work on Windows next release, for instance, like things, can always get added in right.

A

That's kind of the the thing with the negative attribute is that the negative in tribution would probably be version less because it would be somewhat. Maybe it would be and maybe would have a version when you would say I know I know for a fact. It doesn't work on this version of Windows.

A

The question is, then: how people.

D

Will just be the opposite, probably be this version or earlier, rather than this version or later yeah yeah, I guess what I'm really getting at is if Windows has been moving to a model where functionality is acquired via the store, rather than via some particular operating system version, we don't want to box ourselves into a corner where I can go acquire something like this from the store on an existing Windows version. That's already in market sure that makes sense right.

A

All right, so let me scroll down further.

A

So why did we stop shipping protected data? Take.

C

The API surface isn't changing and it's in a new package, so we we version locked it. So the package doesn't grow. I, see.

A

So it's also nice than a tool yeah.

C

Thank unless it's net core app three but but yeah, it's it's locked to a TFM, that's older than five! So.

A

That means you just have to be more formal about or don't level attribution, yeah.

C

I assume that I'll I think that's probably the same thing. That's true on all the ones that Adams pointing out. We don't ship a five-o package of that we just we don't have a net five asset in the package, because we've watched to a lower TFM, yeah.

A

There's probably a decent one of them that are like this and pretty sure that I mean a good chunk of these things here from the compact pack, so it wouldn't be shocked.

C

A

Ol net standard actually or look good chunk of them on that standard, it's.

C

All gonna be are all of our herbs and some herbs yeah.

A

I think, once we have this lock down, like it just rinse and repeat for the other ones, so let's find in it all right, specific API. This is where it gets more controversial.

A

So, as I said earlier, like one of the concerns are have is I, don't want to be like I, don't want to have a world where we have like a lot of platforms, attribution all over the place, so I still want to have a kind of like a model where it's like. This obviously never works on winner. So this obviously only works on windows, because that's much easier to reason about them.

A

These moving targets, but it seems the console API is in particular, are basically never going to work on anything but windows because of the way the windows terminal works. So are there any objections with marking these.

D

Just to confirm, if I'm running inside of, like OS 10, for instance and I open a terminal window, these api's will not work.

A

So the M mark is, they will throw an exception so that I assume that's tool.

D

And you try and use these I.

B

Think the question was with the new Windows terminal, which is completely different, so so it's not necessarily. They won't work on other platforms. It's we've not made them work on other platforms, yet they.

D

Explicitly throw platform not supported exception in our UNIX implementation and consolate yeah right.

B

It's but it's not, we couldn't ever support them. It's we may in its we don't today we.

D

Don't today, and for almost all these.

C

We don't know how to support the modules and if we did add support for them and in six we would just remove the attribute yeah.

A

Like to be clear, like I, think what tena is asking is these attributes? Also these these api's, take in the new terminal window. I think there's an assumption that they're more unix-like, because they support the NZ escapes but I still think as a bunch of API is like, for example, buffer manipulation and with no manipulation that doesn't have an NC escape sequence equivalent. So there's no way to make them work on pure UNIX.

A

So, in that sense, yeah I think they will probably never work anywhere else and even if I mean basically the model, what that would be. We just remove the attribute or we basically.

A

Never Poe just remove the attribute at that point.

C

Because we'd be back, we don't know how to describe that. It works on any UNIX distro that contains lib whatever, which is mostly the things from 2020 and yeah and we'd, be like um yeah works everywhere. Asterisk yeah.

A

Alright, so there gets ten console when people are fine with that, but this time here the both methods are marked. The type doesn't do much else, so I'm almost inclined to just mark the entire type, which is kind of the way I would think about that like. If they see the work host methods of a type don't work, it would mark the type seems fine, like this little value and being able to instantiate a type. Then it can't actually do any very useful work. I mean.

C

We have the same problem if if, in fact, I think it is polymorphic and if the encrypt and decrypt are overrides, then really the thing that you want to flag is the person who said new of your you're building a thing: that's that's going to blow up somewhere else granted the fact. If you return one, you have no idea if the caller is going to be expecting it or not, but it gives you the the warning of like you thought you were doing something with platforms, and this isn't gonna do what you want. That's.

A

A good point: polymorphic things might actually not work today. For that reason right, because both I think the call set is through the base type usually and yeah wasn't probably doesn't help with that I mean we've data flow you should you should be able to tell that it's the other instance, but there would be extra work.

C

Maybe I mean like if, if you call a method that returns a protected configuration provider, then you call provider dot decrypt. It has no idea of whose decrypt you're gonna end up. That sure.

A

Like I was just thinking even the naive kolevatov rx equals new DP API, protect the configuration provide, and then you say, X dot, decrypt I. Think when you look at the call sort of extra decrypt I think because the wave c-sharp works I think it would see through the if it's a polymorphic function through the base type, normally I. Don't think it will tell you that this is from the DP API, protect the configuration provider and then.

C

So within them you get the wrong attribute, set yeah.

A

So like within the I think within the method, you would be able to use data flow for this buddy across across method boundaries. You're kind of out of luck- I mean they have this but like I, would not use it because it's extremely expensive to run and, as you said, I think in those cases we're just better off just saying sorry, the constructor itself is perform specific, so you don't get to new it up.

A

A

Yeah, so this is getting into this territory, where we don't have a good way to describe this with an attribution when certain combinations of parameters will throw. So my personal opinion is, if that's the case, we should not put an attribute, and we can then decide whether we have a dedicated analyzer that detects these particular combinations of parameters or whether we don't and I would drive that. Basically, by on how popular is the API, is it actually worth it.

D

But I mean these: if you're passing in like a literal period or something to refer to the local machine, why not just call the overload that doesn't take it because sometimes you're not passing a literal, it's passed in as a variable from somewhere up higher and set it. You know earlier in your program sure, but at that point now you actually want to be aware of the fact this API might not work.

D

F

Way for distinguish.

D

Yeah no, but if, if it's not the thing is, if it's not a constant value, then it could fail right and that's the purpose of the analyzer. That's the purpose of this attribute to tell you like hey. This method is not really supported, like I guess.

C

I don't see what, let's take a better example.

D

Scroll down to the memory met files.

D

There's there's literally no way to use memory in that file without passing in a string and if you pass in an on Milstein acelino string, it works. If you pass it on null string, it doesn't.

D

So we're not going to say memory, not files and support. Yes,.

A

I think I think to Levis appointment. You could imagine a world where we have basically dedicated api's for that, but my concern with that is that it's very noisy right, because if these API is exists, which is against the case, maybe father existence would be I. Believe, like these things are super widely used right and then you basically have on of false positives.

A

You ask everybody to change their code, in which point the analyzer also loses its value when it just flex too much aggressively right, so I think in those cases we are better off just saying: well, let's just not have the attribute instead of a dedicated analyzer. That can basically say sorry, you put you I, think you're passing a normal string here that won't work, but and but the problem with that is these arcane or handwritten analyzers for every single case right, unless we find a way to formalize it, which might be very hard.

C

Because the only thing you would do, if you accepted it as a parameter- and you were the one calling it is, suppress it because you probably don't want to mark your method as oh I- might only work on Windows right.

A

Yeah I mean I would opt for being conservative here, like I mean that's the general stands. We take analyzers, it's just reducing numbers of false positives, even if that means you're, risking false negatives, because I mean we live in a world right now, where basically only have false negatives. So if we can reduce the number of false negatives, that's still a huge win. But if we draw on everybody in warnings, then we're not maybe not helping anybody.

A

So then, if we scroll up here like I, would just say everything that is, that is not always failing. We should not mark that would be. My stance. Has anybody a problem with that? So.

D

If there's any combination of parameters where it can work, cross-platform do not attribute the API well.

A

I guess maybe, if I phrase it like this, it's a bit harsh I would say if, if I mean, unless it's a super remote case right, then maybe not but like. If, if there is very common cases of the API that will work cross-platform, that enough people use, then I wouldn't mark it.

D

Interestingly, until I read emails, doc, I hadn't considered a particular case, but in stock may be realized as possible.

D

He had annotated some inu values rather than you know, a whole type or the method that take those new values and that actually works well for us- and there are these situations like I, talked about this later in the doc like on socket. There are several api's that you know only work. If specific I only work on Linux, it's a specific value was passed in, but as long as, if we annotate the enum values that don't work instead of the method, then we can probably still catch 99% of the misuse right.

C

Yeah, but where it's string, that's a harder yeah with string I, don't have.

D

A good option, other than I I mean I've, seen a bunch of call sites that you know, as Levi was saying. Just to you know, open call period, close quote and I feel like the noise, for that will just be just annoying yeah yeah.

C

I mean I I think that if the method body is anything other than just throw, P n SC or maybe has some compat of argument- validation followed by an unconditional pn SE that we should not annotate it at this point, I agree so.

A

That means basically for these ones here get processed by ID works. If the second argument is a period.

A

D

A

So that's a good example right, like I mean I could imagine some attribute. We put it on the parameter and say this must be X to be supported but like if it's not even a literal, then we're screwed right.

A

So get processes string the same thing. It will work when you pass in bot or the current machine name.

A

I guess the second parameter has always yeah alright. So then we have this one. Nobody has any objections to the other ones.

A

Yeah, so graphics I was not sure, but isn't graphics based the DC, the the ones are called device context thing and when 32 like, why would be about? Would that work anywhere else? Well,.

D

There's the live, GDI + Lib that we use for compatibility on 1x Emmet, look at the code for this there's an explicit comment that basically says this works on links. huh You.

B

Know we explicitly support system drawing on UNIX based operating systems? Not all operations necessarily work the same for work, but they are implemented as people request them or fixed, where possible. Gotcha.

A

Well, in that case, ok, if the type is generally supported, then the type is supported, but not all methods, in which case you should just mark these methods. I mean again, these are the ones we know well, but.

C

Yeah, like Steve's comment here, the copy from screens only fail on Mac. They work on Linux according to the UNIX implementation here, and then they will work on Mac if you set an environment variable. So these that it's back to the the method body has logic that only has an else pn se. So.

E

C

Not attribute these.

C

It sometimes works on every OS I.

D

Mean, though, maybe maybe the analyzer in this, this gets beyond the current design, but you know: we've talked about confidence levels for analyzers in the past, you're like being able to have some the warning and some things be an ID suggestion, or some things be about warning, but like at a I'm, only 25 percent sure this you know, don't trust me on it. Maybe we need something like that in some of these cases.

A

Yeah I'm yeah I mean that's something that we should probably play with, but like I'm.

D

A

Not saying you're gonna five.

D

I'm just saying like we're taking baby steps here, yeah good.

E

For a baby step, could we start with just some pretty extensive remarks and the API Doc's I mean.

A

We already have that today in many cases, but the problem is that the dogs are us or not actually always up to date or not always read, and it's actually kind of hard to do that. But I was thinking of in the past was saying. Instead of generating a warning, we basically just changed the colorization of the identifier that the ID displays, which kind of maybe make it orange, or something to indicate that there might be an issue here, but we're not causing a lot of noise that one we haven't really explored any further.

A

But like the problem of any sort of like warning or a suggestion, is that I'm really concerned you just ground the people when in actuality what you really want is when somebody writes code, you want them to consider the impact that they have from their cross platform thing right as they're writing it once they once they've written it once it's tested. The video also goes down.

A

Because you'd, then you just back to square one where it's like. Well, what does it mean if you 25% Shore right or 31% Shore, like whether you do the cut of line.

D

At twenty seven point, two I.

A

Mean what might be useful is basically providing this information on demand right. If somebody right clicks their library and says, tell me everything there might have a problem beyond what you what you normally think or something but yeah.

C

Alright, but now we're back to needing some other kind of attribute of maybe sometimes fails. Yeah.

A

Maybe platform-specific alright man, we met files, so we talked about those. We should not mark any of the create methods and.

D

I'm, pretty certain to create or open and I know the open existing will just not work like so the ones below my comment, I believe, should all get.

C

Creator, open wouldn't work if you pass null.

D

It'll, well, yes, and no it'll I think it won't work. I'm 90% sure it won't work.

C

That's a buck. Twenty seven point: two right now: source thought that net is the UNIX implementation, haha victory I.

D

Feel like it has an extremely convoluted I'm, not afraid I can hear right now, but memory service there's an extremely convoluted set of logic, split across the PAL, where every possible path leads to a P, NSC or an argument. No, but that could be well yeah.

C

So Creator the Creator open methods prohibit the name to be null or empty. So if null is the only thing that works for create from file, then then yeah we can go ahead and mark create or open and open. Existing is Windows only yeah.

A

C

Open existing definitely doesn't work because you need.

D

To give it a name.

C

Open the null file, obviously.

A

I'm sure that's very useful. All.

E

A

Name pipes: nobody, sorry, the regular pipes, one run as client works on UNIX you're, saying.

C

All of them I, don't know, I was just looking at the one that was highlighted. Well, I mean like. Is it because presumably Moe got this from a tool? Maybe he ran it on, like you know, net core too, but is this that, like it fails on Mac but works on Linux and the.

D

I think it was probably Mike if I memory serves like even the tool I'm like one one and be implemented into low or something like that. Yeah.

A

I think most of the heavy lifting was done in one one: okay,.

A

Which is also why we probably have things that are changed at this point, but, like I,.

C

Mean so just if you did the analysis in one one, but we did the massive API bring back for two, oh and a lot of the api's we left out of the of one. Oh and one one were things that were window specific. Are we missing a lot of data here? No.

A

Because I think I so I know I, don't think so because I added definitely we did it on a Senate push. I ran it again because we I remember now that we added the comp at pack, which was also new tool, time frame and that's where a lot of the window specific things came in to begin with, so it might have been actually too old, I might have not been too much. I mean yeah, I think it was probably too or not the moment. Yeah.

C

So maybe we made it work into one and you analyze to over yeah.

A

Something like this yeah.

C

As long as we have the two Oh API bring back wave, then word yeah well,.

A

Most of these api's didn't exist in one one: yeah, that's actually I'm thinking about it anyway. So right, pipe stream.

A

Okay, so it means gonna be probably want to mark the enum. Then right.

D

Yeah I hadn't noticed my comments initially that two lines down you were suggesting marking.

A

Okay, so, basically would not mark pipe stream set mode. We would mark just being a member I.

D

Think that's I think that's the right call yeah that should catch the majority. I mean unless someone was like casting the integer value to the at which point.

C

They're asking.

D

C

Regarding enums, just devil's advocate in here, if this means that we would give you a warning that you have platform-specific code, if you wrote, if pipe stream, dot read mode, equal, equal, pipe transmission mode, that message.

F

A

C

I don't know if we can special case equality and switch statements and because those are equality but I'm just pointing out that would be. The concern is where somebody's using it as a their probation, Garden yeah that they're using it as a guard instead of they're using it as an input value. That's fair I mean.

A

Let me think about this for a moment, I think we can probably special case that the inlaws it would be my guess, because the thing is we know, of course, what you do with the value. So, if I think for an enum, probably the the most useful thing to do is block assignments, and everything else is probably fine right.

C

I mean there's gonna, be yeah.

A

Simon's method calls right like things where your pasta value effectively to something else, yeah.

C

I, don't know how easy it would be for all the flow stuff, but doing like what what the compiler does for stack I'll expand. Essentially, you would get the like. Okay, you assigned this to a caution thing but like.

C

If now, if you only use it in the guard, I'm not gonna, warn you, but if you passed it as something else now, I have to warn you that, like some paths may be weird so I'm, just this one I think is gonna, be the biggest source of us introducing noise, and it may be that we keep refining it over time but I'm bringing it to to consciousness right now. That's good! My guess is. It won't be too bad because.

D

We you know to do the guard checks personally, this your are you're gonna have to walk up from the expression to see. If there's you know what you block your inside of and to do that. You're gonna, pass through assignments and you're gonna, pass through arguments and whatnot, so yeah I.

C

Can see I can imagine code that has a local.

D

C

Lot likely a bug hard, but yeah yeah well, I can see you assigned a local and you're using that in an if later you didn't pass it as an input parameter to something, and so it's like well, it's not all assignments are bad. It's non-local assignments are bad, but then we need to paint the local as a as a bad value itself. So that's the complex flow things, and maybe we don't need that and wave one, but I mean.

A

What's also worthwhile is just actually seeing what happens when we turn it on right, I mean like I would not be shocked with the vast majority of usages is just in um passing being passed around and then, if the, if check is extremely low, then also suppressing it locally is fine too, in my opinion like if it's.

E

A

That you'd use it, then maybe we should specialize it in the analyzer, but.

C

Just making sure that we're thinking about all the potential problem, spaces and but I yeah. So as long as we have considered it, and especially there's the wonderful annotate, these things run the analyzer and see how many warnings we get. And then that tells us what things we forgot to consider. Yeah.

A

A

Mean I think right now, the way the analyzer works, anyways I, think I'm not sure how much money generalized it out, but I believe last I checked there was to punt. There was still plumbing the flow analysis to the generic framework through anyway, which means it does flow analysis for data as well, so I think at the time we run. We can basically ask flow analysis for data as well, so it wouldn't make it more expensive.

A

But my concern similar to like Stephens, is like the more convoluted you make, analyzer the light more likely that you have bugs right and that's yeah that we should avoid if we can help it yeah.

C

Like I'm not saying, we need to fix it right now, just that it it's if we thought about it and we think it's the right thing to do. That's great I, just don't want us to ya.

A

C

Caught by surprise on things, yeah.

E

All right so hope help me make sure I got the right takeaway on that. So are we going to leave the attribute on that method? Then nobody.

A

D

Know we would do it so.

E

D

E

But then that would mean we, the analyzer would need to check arguments to the method. Look at the values to see if they're in ohms and see if the e-news had or see. If the Union value has the attribute right, yeah.

D

I think it's a I think it's a sliding scale here we start by adding it to the EMU member. We see you know either by observation about running the analyzer report ever code based we have available to us how much noise there is and we start adding special cases. For you know we only flag, it's an assignment or an argument and then the only flag, if it's blah blah blah I mean add in the data flow analysis, and you know we kind of go down the rabbit hole right.

A

Okay, I can think right now, like we're shouldn't. Do anything basically like, depending on how the thing is written I believe she already will find new members anyway, because it just appears effectively. You know, field reads right, static field, weeds, and so they should effectively show up as anything else.

A

So the only thing we then can do is what you said is basically saying: we only complain about the usage of an enum if it's stored or passed as a parameter rather than on every read so to avoid effectively the case we say: X double equals some enum value, but I think that's something as Steven said.

A

I would honestly just wait for actually running it on a real codebase to see what patterns we see, because my gut feel is it's probably for in ohms that are very frequently used in configuration, probably relatively rare, to read these values outside of actually creating something.

D

Race is a good good point, though, if it's not already like, we should definitely have it in our plan to pick real applications that we have access to compact code bases like it's a code bases whatever and like use those those are gonna, be much more informative, been saved on runtime yeah about the value of the nyo certain, whether it's noisy I mean.

A

My default assumption is, we run them on on runtime and rosilyn already right or that's I think one of the things we do usually so I would start with that.

D

It's don't think.

A

D

I mean we should definitely do not like that's a min bar, but I feel like we need to on some actual applications as well. We.

E

Are trying to say something.

F

A

Yeah and we'll check whether we can find an app suite somewhere so.

C

I, thank you Jeff, that what we have is the iterative plan of step one is we don't special case anything. So it's talking about the Cena member is a warning, and then we say: okay, we identify a class which is it's in a it's being used in a in equality or a quality check. Let's ignore that one, because that's a that's a guard and then and then we keep finding places of it's not inappropriate to use this enum member and and then we suppress the warnings in those cases.

C

Sorry, we don't raise the warnings in those cases so that people don't have to write the suppression, but we start naive and then work our way up, based on data and.

D

Then, if we find that's, that's just an efficient or insufficient, we switch to basically you know, any use of the Inu is okay. Unless it's in one of these known places where it's not like being assigned or stored in them, yeah.

F

Okay, thank you.

E

A

Right so then that's cool further, so port, so yeah so I said, should be marked this entire thing, because the list of API is were very high, but I also remember that we made part of it available. Cross-Platform, like is that, like it's the data here outdated? Or is it really so that most of the things don't work?

A

C

Looking at this later here, it's.

A

Suggested type doesn't work at all. It.

C

Was implemented about okay.

A

So probably the data is outdated here. Yeah.

C

Because, like serial error received event args that event is just to get set, get private set. So there there's no pn SE and in reading that property is.

A

There anything that Phyllis pop from not supported in supported this one was everything supported. Yes,.

D

My guess is, you ran this over an auto-generated PLLC assembly, yeah.

C

The fact that it's painting, like everything, means it was as Eric likes to call them a turret assembly.

A

A

So should we then just start marking it or she'll be missing. It.

D

We should probably investigate it, but it's gonna need to be rather than I.

D

Also, don't know I'd have to check I, don't know if there's if the implementation, this works on Mac I didn't check that.

A

Well, it probably I, don't think, makes and serial ports right so.

C

You might be able to do an rs-232 to USB or USB.

E

C

Rs-232 extension breakout board, they Windows machines can support rs-232 over.

D

Usb I assume that count as well.

A

That is a bold assumption, and this it's unless it's, uh what does it fire I'm.

C

Sure it can, if it's Apple, certified hardware, whatever.

A

C

Magic negotiate thing is.

A

F

A

C

Sir I think ports or whatever assembly this is wouldn't have to go back to owner for assessment.

A

C

Unless you can easily just rescan assemblies and conclude, the il body is just throw P. Nse I mean.

A

I can, which is probably the easiest thing to do, and it's use it as a starting point. The tricky thing usually is to find the right implementation right. So what I did for the platform? I just downloaded the the run time and didn't run it over the run time and then found public entry points in the run time but like in when you get packages, it's harder, because you have to basically pick the correct binary which right.

C

And so I guess your data may also be missing if it's the public entry point calls a helper thing which eventually concludes wait. We didn't write this on on Mac and gets a p NS either and.

A

No because it like so I need to know I'm not just naively checking the route, because that would not find almost anything in the PCL. What I did is I said up to five levels, deep, which also is one of the reasons why I have lots of false positives, because if you do that well, there's a lot of code now between you and the platform or supported exception. So there might be a lot of conditions under which it throws, rather than it doesn't work at all right.

A

But, like that's yeah I, don't think we are very likely to miss those things. We are very likely to over index.

A

But I think we already did the initial pass. I think Steve, like three years ago, the initial pass and removed a ton of stuff from it.

A

A

Basically, what's even a saying, none of these things are platform-specific, then be.

D

Used to be listener, timeout manager, one is but the IP endpoint in socket address ones or not.

C

D

C

Say I just checked the timeout manager and all those sets are through.

C

Assuming they work on yes, because I'm in timeout manage or not manage, does yes, so make sense. Possibly this means we have a Windows version and that they work there. All.

A

Right then, all the suckered ones, I assume also not love from specific. Then almost.

D

All of them are not so the set IP protection level, one at the bottom, is window specific and then the IO control ones in the middle. There are this work for some input values based like some of the control codes, work and some of them don't is.

A

That you know what small there.

D

Is yeah and I if you scroll down just the tad.

D

And then there's a very rare path and accept async, that's not worth talking about other than to say. If you have the socket async args configured in a very specific way, it may not work.

C

But using the all-or-nothing heuristic, that means we wouldn't paint it correct.

D

And we definitely should not the API super widely used, either directly or usually indirectly, on UNIX, you.

A

Mean the IO control one or no.

D

No, the accept async one I see.

A

So even IO control, though we probably wouldn't mark this one right. We just mark the enum and then so. We should just part.

D

Basically, all the union members that are not the three that I listed yeah.

A

But you know means.

D

C

A

But the means and fires methods go. The only method on socket we would mark is set at P protection level. I believe.

D

That's the case, yeah.

A

Okay, so- and let me just record the these guys here,.

G

There's also the socket information constructor, which I believe only works on Windows. So oh.

D

Right those were added after you Ranger to Lima, which one.

A

D

Just brought those back, we just brought those back for done in five right: Corey I.

G

Think it's been there it just it. It didn't work all.

D

Right did work on Windows.

D

D

Duplicate method, or something like that.

G

Yeah, there's duplicate and close, which I did we make that one work on UNIX I.

D

Honestly, don't remember: I can't.

G

I know Anton did something: I can follow up on that at.

A

These two measures duplicates.

G

A

Close or is there one method called duplicate in close there's.

C

One method called duplicate in close I.

A

C

Which source thought that that that says is pn SE on unix, okay,.

A

All right, let.

C

Matt, don't overwork some windows yeah.

D

So it's the two things: there's a constructor that takes a socket information type that one throws PN SC and then the duplicating close method throws pianist II.

D

Good memory, Corey.

A

All right, then so I'll good task, extensions, we cleave them all as well, so unmarked those.

A

A

And the transmit file options so.

D

That's an e new and the first two options do work on on UNIX, the less the rest. Don't I.

A

See so that means Momo.

D

I can't remember right now, if we just ignored them or if they actually throw an exception. I couldn't tell you I.

A

See so by these two, you meant it was just the two members. It didn't mean the two types oh yeah.

D

Correct yeah to teach me listener not allowed to natural bristle that that throws okay.

A

So I should I'm in France.

D

But file options not disconnect and transmit file options, don't reuse socket are both respected. So I should have been more clear. Okay,.

A

So that means we unmarked disconnecting reused socket, but we use, but we keep TCP listener allow net reversal. We keep transit file, options, use currently PC use system thread in right behind cool and then same UDP, client, right, yeah, yeah.

C

I mean the TCP listener allowed, NAT traversal looks like it would have, it would get a bubble up. It would need the attribute anyways because it calls set IP protection level. I.

D

Almost caught IP protection inside no.

C

Yeah I mean it looks like whatever makes server socket be null, it will succeed because it just sets a field, but then probably something will blow up later.

A

Yes, interrupt like I think they all they all look very common to me, which yeah I'd be shocked. If any.

D

Of these work on you, next is probably an accident yeah.

A

All right, Jeremy, you know this better than anybody.

C

Yeah I mean I would just mark honestly. I'd marked the whole CSP key container info.

A

So that means for you unlock all of this. My whole time, yeah.

C

So CSP key container info and CSP parameters marking both of those will cover the one like that'll, then make them it'll bubble up to the appropriate constructors and properties on the RSA and DSA crypto service provider types and then crypt derive K is correct and import CSP love is correct. I think.

A

Okay, so sorry, you're, basically saying if your mark sees people container info, we don't have to market property on the service provider. What.

C

We we will, because the analyzer would yell at you that you're a cross-platform member that calls a platform specific API right. That.

A

C

A

Sure that it does that.

C

Well, we would get the warning in our own build that we call the platform specific API from an on attributed context. Oh.

A

Yeah that one, but yes, yes, so.

C

Then we would add this.

A

C

The property or add the attribute so get CSP key container info on both RSA and DSA CSP. It should be Windows only the whole return type from that should be Windows only instead of the properties individually, and then you marked on an RSA crypto service provider, the two constructors that take a CSP parameters. Those are both good. You I think probably missed an equivalent one on DSA crypto service provider, but we should just mark the whole CSP parameters type as platform-specific.

C

Technically, it's just a struct, so it would work on unix, but there's nothing. You can do with it. So I would over paint that one because it makes life happier. Okay,.

A

So that means we will add a new member here. That is basically CSP parameters. The type should be marked yeah and then basically, what about the fallout of that is yeah and the other ones that just talks with vitalize you've said they were right.

C

It looks like we made that those work, so you can, you can remove the import CSP blonde. Okay,.

A

Imports this you blow I, remove from both the DSA and the RSA.

A

And then what about the middle of one, the Crypt arrived key.

C

That one is definitely windows only okay.

A

C

Feel like there should be a second hit of that, but maybe it's in a different assembly. It got complicated, Lee moved for layering for net core before we tried completely removing it because it was not gonna work on non windows, so it got duplicated in the last there right as we were starting the project before I join the team.

A

So, in other words, it's not your fault and then the x5 will not in shame constructor.

C

Correct that only work, someone does the one that takes it. Better only does cool.

A

Yeah, so serve is mortal I feel like say, might just even like I. Don't think anybody here has love expertise to say that.

A

So, let's just go over the possible of this. If.

C

If the string comma exception, constructor really throws PN se, that's really sad. So.

A

Which one? But the.

C

First thing under WCF was an exception type, which said that its message, comma enter exception, was platform-specific yeah.

A

That's possible that it's also picking was picking up a turret assembly, but.

C

Fair enough, alright.

A

Threading as long as the string is not so that means we should probably mark them as unmarked them for the same argument.

F

C

Only the Constructors right, Steve.

D

Yeah the the open existing and try open existing will work.

D

A

So we'll work on won't work.

D

I will not yeah just the Constructors are fine, but the two following those need a string and.

A

Then I guess semaphore same thing: right: yeah.

A

Okay, so thing lined wait. One will also be unmarked. Then.

C

Thread set apartment, state.

D

Tri-State apartment state will work and always return to Falls, but then set apartments that I believe will throw, because it just says it: try sent apartment state returns, false, throw I.

C

Think Oh yep I got it.

A

A

Cool, so that's actually decent number of API, as you marked so what's harder for us to do, is checking that we didn't forget anything but like I mean one thing I could do is run the tool again. It's.

D

Gonna say we could do one of two things. We can run the tool again or we could just search for throw new platform, not support of the exception and see if you know eyeballing it, it seems like a little bubble up to anything else. In most cases, I would imagine it'll be just at the top level yeah.

C

A

When some of the.

C

Pn SES will be just doesn't work in core, but like that, we just never that it's a member. We brought back for net standard to alignment, but we didn't actually bother with an implementation. I can't think of any off the top of my head, but I believe we would have some yeah.

A

Let me do this. Let me just go run the tool again, because I think we were good people back. Then we added a manual inclusion list and a manual exclusion list, so we should not like. If I run it again, we should be able to still subtract and add the ones we've did manually and then everything that is net new. We can then review.

A

So that should be doable.

D

There definitely seemed to be some that are missed, like I. Just did a search, and one of the first things that comes up is our generator, which I didn't see in the list here.

A

Yeah argued writer I think we edit later right, I think there was a lot of discussion on not adding it and then adding it eventually. I.

A

Think in general like if it's run time, throwing at guy I can't see that.

A

So with an internal runtime : out of luck.

C

Well, but like our getter Raider has public override in to get hashcode, throw new pn SE.

A

Fair enough yeah, alright, I will just run this thing again and then see what falls out anything else. So the only other thing that I've heard as far as Merkle's here is the exclusion addition to the spec, which I will probably drive more from a move from a blazer standpoint, and then the other part was the analyzer on enums.

A

The DEM is anything else. I.

A

Had the next, an ER 201.

D

A

D

The overarching find anything we missed.

A

Yeah, so basically it's two things so I, just like the the four things, but it's the spec work for NASA. No, oh, it's the spec work for the enum analyzer. It's the removal or the exclusion attribute, rather than the inclusion one, and then the run the analysis again.

C

Yeah and with the exclusion attribute I would we want to warn if you called that from a unattributed context.

A

Yeah, that's the kind of thing we need to think. Oh probably knows.

C

A this, it is a you are not as cross-platform, you think you are, but yes,.

A

So there so I haven't really thought.

C

About that positive and negative or infectious yeah.

A

So the the way we thought about this in a better context is so so blazer has two problems. One of them is it's really cross-cutting concerns that that don't work in laser that are like not really platform-specific at all and then there's the other problem that it's very widely used. Api's, where, like file, I/o networking, you know threading, so I, don't think it's desirable to say.

A

We need to file new class library that we suddenly warn you for everything that might not work on blazer, because the vast majority of class libraries don't care whether they would work in blazer right because basis. This somewhat narrow scenario for the net developers right versus not running on Mac or not running on Linux, you could argue all they're kind of the same computer, so you probably should be optimizing for that, but it is a very new environment, like you know, an Apple watch, for example. That is super crunch constrained as well.

A

It probably would be like blazer, where it's like. Okay, we have an analyzer for that, but we don't care about blazer issues by default.

A

You have to opt into that because otherwise we just draw on people and warnings again for platforms in, but if they may never care about- and you know if the scenario really is- that the vast majority of new class libraries will not target this environment, then you know we should have a way to opt these things in manually or opt out manually or whatever the process for that is, and they will probably also publish to the context. Sorry, no.

D

A

Know that was it.

D

Do we, how are we thinking about API? Is that through a platform not supported exception, regardless of OS.

A

C

The ones whose documentation says platform not supported, exception, dotnet, core : in all cases, yeah.

D

F

D

A

They should abort it just.

D

Throws regardless we.

A

Should mark those as obsolete and just say, you should never call these api's.

C

In the in the net, 5 TFM or.

F

C

What well I guess in whatever, wherever place, they're throwing I guess it'd be weird if it's net standard and it shims to framework. But that would probably be a well.

A

So that's the ball movement standard right like so that's why originally the the analyzer was very heavily designed to be out of pan, because the assumption was we can never annotate that standard after the fact right, but now that we converged and a lot at 5:00 timeframe, I think we can just assume that we can touch the reference assemblies in general right. But it's yeah, it's more like a servicing problem, I just like!

A

Well, we can't serve as the existing ones, so you only get them the next time you get to ref the assembly set, which is done at 5:00 done at 6:00 kind of thing right and so I think for the for the basis scenario like I will formalize the you know, sidecar information, somehow in the SDK and I think we can decide that we generalize this for observation as well. If we really cared enough but I think it's totally fine to say be I'm. Sorry, like you know, we will do it and done at 6:00.

A

If you forget something about at 5:00 like it's, not terrible, it's terrible for blazer, because there's lots of stuff that doesn't work on blazer and people who want to build blazer apps, don't want to have land mines everywhere, but I think for for esoteric api's Devi decided to make go everywhere. You know that we will not have that many on them right and so I think we can probably scan for them relatively straightforwardly and.

A

Yeah I guess in my case that would just be the it's not an exception on all three platforms right, that's kind of how you would find them today and then we would just mark them as obsolete.

A

Besides task thought was its threat or the board is anything else that we have in that category, that we know of.

D

A

D

Of the get object, data is related to a runtime serialization 90%.

C

Of an app domain so I mean those are things that, like the runtime dropped and then I feel that there are things that we brought signatures back for net standard. Oh, that there are things that we didn't make work, but I can't think of any of those I'll stop my head. Yeah.

A

I think four yeah they do is no masking is because then I can. Actually. You know check that my tool finds them. I guess it'll it'll, lift through OPN is easy. It's easy. If they literally do. If it's an internal runtime callin death throes. Somehow, then we out of luck but like abdomens, published Wilson managed code would be my guess.

A

Probably some of the remoting stuff probably also throws yeah yeah. Ok, let me let me search for the Olson and make sure.

C

A

Covered Tanner.

C

Since you would be the most likely to know, is there anything other than intrinsics that we have architecture based limitations on and do we need to be concerned about those like no.

B

C

On x86 64, no.

B

That I'm aware of off the top of my head there might exist something though I think for the most part. We we try and limit that type of stuff to just to the intrinsics, though, and everything else we try and make sure has some kind of software fallback yeah.

A

So one thing I was thinking about this is another spec I wrote, which is in the design sleep when anybody cares, which is capability, API checks, where the assumptions are a good example here would be reflection amid and also the intrinsic swear you expected to check. You know a thought is supported property somewhere and have any analyzer. They were able to do that.

A

The challenge with these analyzers is that because the api's are in open-ended said, it's been hard to formalize this, but I think I have a mechanism that that could work in those cases, and so we could imagine a case where we would basically have an attribute that you apply on a on something that says well. This is specific.

A

This is basic guarded by capability check, so the call side should be under a capability check or I should for what its requirement to its callers with a capability check, and so it would work similar to him the minute wise platform attribution, but instead of requiring effectively an if check against, is always platform or earlier, sir, or later you would basically say for the capability API that this thing marks and that works reasonably well.

A

For those kind of cases, it might be tricky for things like hardware intrinsics, because we got fancy with inheritance, where it might not be able to tell that it's okay to check against the derived one and then cover the base one, but we may be able to formalize this.

C

I'm, just like we have, you, know, OS and then architecture is the only other thing that I can think of. This ran on my box, but doesn't run over here. Yeah.

A

I wouldn't want like so to be clear, like I, don't think what we should do is we should not formalize rids or architecture similar to as platforms I would just say. We have OSS because they're super common and then we have effectively capability checks and that's for everything else, because I mean, for example, we know ref admit you don't want to really have people saying for like. Oh, you know, there's like three different things. You can check for right and it doesn't seem worth it.

A

We should just say this community ApS with these things and you check those.

C

Yeah and that that's why I'm just I'm making sure that, while we're talking about the notion of OS specific that there's not another big sweeping category, that would cover things and architectures, the only thing I could think of that would do that. Yeah.

A

The one thing I will say those like it so native is probably so the bigger problem is, is people have been repping native api's and shipping them themselves and and that's a bigger problem, but I think you could argue that they have to do it themselves. I mean.

C

For the analyzer spec, you could decide that you wanted to paint anything that used it extern, but that's probably just super noisy yeah.

A

I mean if you call an OS platform API, we can probably check those so I think it would not be unreasonable for us to find, pin box to Windows and and then tell you that all you calling a windows API, because we know about it and even suggest the correct attribution, because we know what version they were introduced in right. So that wouldn't be unreasonable. But, like I, think that would be some non-trivial amount of work for us to actually do that. Yep.

F

A

I think we generally don't care I think the idea is, if you write a teen work yourself, like you're, a platform extender so like you have to do some, where heavy lifting yourself yep.

C

Fails on the cost value analysis, I think.

A

That's fair all right, so then we have to 30 minutes left. Unless somebody has something related to this. I would just open up the backlog.

E

Got something tangentially related to this and hope I think you were kind of looking at the the other PR for adding other platforms to the iOS platform list as well. I was confused on where we were landing on the Mac, OS and OSX platform. Comparisons is that appropriate to bring.

A

That up here, Oh totally.

E

So the issue I should look at or yeah I'm trying to out grab a link here, so I'm pull 39 I was 64.

E

We had talked during the API review, 39 I was 64, I was pull I'm Statia yeah.

C

A

C

A

E

We had talked about trying to hiding OSX, making sure OS, X and Mac OS are equal right, but Mary had pointed out that that is a breaking change. The way we had implemented it so trying to figure out the right way to merge those things together. So.

A

As to comment, oh there.

A

Yeah this way we do it the other around, but yeah I would basically just say: Mac OS is an alias for OS X and we just hide. It was X because who cares what the underlying string is.

C

Yeah, we do explicitly in framework guidelines, say, never depend on the value of two string, except for the members whose whole purpose is to build a string. But two string is supposed to be for diagnostic and display purposes and not for run time. Evaluation, yeah.

A

I mean we say that, but like the.

C

Safer change is probably to special case the is OS platform on on Mac, since we already have a split build there.

E

Then the underlying string is forever OS X yeah.

A

I mean it, it seems fine I, don't think I mean. If somebody has an aesthetic issue with it, then that's unfortunate but I think that's it. Let's see, I mean if we really cared what people see in the debugger write it. We can easily have it. What is it deeper debugger display attribute where you say you have a two debugger display method and that one basically doesn't render Mac doesn't render. Was the exit Windows Mac OS right like if he really cared enough but like I.

B

A

With Merrick the problem is, it was accessed around for so long that people very likely have written code against it and I don't dare to reason about what code they have written. Yep.

E

Okay, so then, just having.

B

E

Os return the value of OS, X and, and then the way we were trying to do that is whenever you're calling constructor paestum in a string, we would just coerce the string. So if you passed in the string, Mac o Mac OS, we would coerce the string to OS X and what everything else flow through thereafter I think that works well in that approach.

A

So I didn't get the cores and part of your base is saying.

E

You would change the constructor.

A

Os platform to say, if you passed in OS X. So if you posted make.

E

Them to pass in the string, Mac OS we we would set the value of the platform name to be OS X. So then the rest of the logic for equality and everything lines up. Jamie yeah.

C

I, don't know that I guess you know, I think that it just on the Mac OS build. We would make the equals method, have a variant build that accepts either of the two names and we would make the is OS platform use. That method I mean.

A

The option is we just don't do that at all? We.

C

Consider it a culture or a context, specific comparison. The two strings are the same: if and only if, you're running on a Mac I.

A

Mean you yeah, you could do that or you could literally just say we don't care, but you could just say you literally have two strings. You have mac, OS, OS, X, they're, both valid strings and then the only party that has to handle that they are equivalent is the implementation of is OS platform.

A

And, of course, the analyzer right, because our analyzer doesn't see that yeah bit but our nice to do that, anyways, because the unless it never sees the literal string, no matter what I do analyzer has to know that it was X, a micro, a sec, : yeah.

C

Because we don't have, we believe we explicitly didn't make an API that tells you what your current platform is, so you couldn't have written it as an equal, equal sensibly yourself right.

E

C

You're, comparing a variable to some other variable.

C

Right but but where did you get one of them that would be you're, comparing a literal against, so we can make the equals work, but I, don't think that we need to care about the two strings and because you well, you can have written Marik's example. Here. Someone somewhere gave you the OS platform. The system did not give it out to you.

A

Yeah, that's the.

C

A

Original design was always that, like we can have kind of families of OSS and specific OSS, both identified by OS platform, which means the idea was always.

A

That is why, as platform can accept, where can return true for multiple distinct values right so, for example, you could imagine that you know we have one two entry there and if you ask a bun to IU Linux, which it says yes, and if you ask you a bun to it, also says yes, right and I think that's kind of why we never expose an API that told you what do as it is, so you don't need, so we never have to compare them yourself.

A

You just literally take the thing and pass it on to is West platform, and that's the only thing you can you you get to do so. You can only basically ask the currently running OS. You cannot compare when the masses, so in that sense, I, don't think we have to line the strings.

E

You know, as far as this type goes, OSX and Mac OS are different and then in the implementation error in the is OS platform and is OS platform. You know minute whatever the minimum earlier I can have a lot of those names yeah those those checks would have to handle that they're equivalent to platforms which.

A

I, don't think, is terrible right, because it's kind of this you know the Ubuntu model right where you say like well. The party that is running knows, which you know that's the party that can actually tell you which names are equivalent in which ones aren't.

A

The only downside to this model is that basically the analyzer has to replicate that. But that's always.

E

A

Because we don't have literal strings right, we only have enums here. So as soon as you have different identifiers, the thing has to know what it is.

E

Do maintain the equivalence check and the analyzer and method.

A

I mean if that becomes a concern. One thing I thought about is an attribute that we can put on OS X. That basically says alias for something like that, in which case the analyzer can just read that metadata to actually know that.

A

But right now we only have the case for Mac, OS or and if check seems more appropriate, I guess.

E

A

I'm, just to be clear, I think that's going to be a problem, no matter what we do because of the way OSS version and work right. So another instance where I was brought up was imagine we want especially eight different browsers and wisdom, so you have know what Chrome and then you have a GM or whatever we will call it right and then you have so does edge. Iam, not say it's Chrome, probably not right.

A

So, like no matter what we do like there will always be these cases where one thing is multiple things, depending on how you ask so I really don't see, there's a there's a.

E

Way around that.

A

E

Should there be time info to do the equivalence check in one spot, but what's.

C

Like local dentists,.

E

C

A

E

But is platform checks what your current platform is not check and see if two things are equivalent, but.

C

I would why would you want to know that chrome and edge iam are sometimes isomorphic, like the only the whole purpose of this type? Was you never get to ask anything about anything other than am I this okay.

E

C

There's no comparison, there's no! You get to ask what am I and you write your own switch, a fall that has a bad else because we've added new platforms. It's you ask, am I this and you react to that. Yes, no question. It's the the 20 questions. Version of what am I a.

B

Reason why you might want to know is because, even though both are chrome and they both reference line, there is difference in what extensions are exposed in some, in some extension, specific api's that may be available. It's just like if you're writing a graphics, a graphics engine you're going to write against Vulcan or DirectX, but you might also want to expose specific functionality if you know you're running on an AMD card or an Nvidia card or an Intel, and so you may want to query for expect for specific extensions beyond that base level and.

A

That's fine I mean, if you can imagine doing this in various ways. One of them is, we literally have I mean it's. Maybe a bad example here that we have west platform entries for extensions. Then you can ask for those- or we literally just have a different way of doing this, but I think the idea would be the problem that we basically have is that Wes's get forked in version over time.

A

So at a given point in time you know five years ago, for example, there was only chromium, there was nothing else, and then you know Microsoft build their own browser based off of chromium, and now we basically have chrome and edge both based off of chromium, but there probably are chrome, specific things in over chromium and there's probably Microsoft Pacific things over over chromium as well, and so the way this would work is, you would say we would have just started with chromium.

A

Then we would have added chrome and edge later and then the person who writes to check knows that edge is a superset of chrome, okay, so a chromium and but that's kind of like domain knowledge. At the point you write the code, but but the person wrote the code a few years earlier, only knew about chromium. So that's the only check they have and that's why it's bad.

A

If you expose an enumeration because then the problem is when we add new things to it now you have bad else code in your in your in your code, but that person from two years ago also would not have had the need to ask for an edge specific extension because it didn't exist at the time he wrote the code. All he cared about was something that was in chrome, right and I.

A

Think that's kind of the the model Bobby want to have this a bit more open-ended and a bit more flexible so that the implementation can can be a bit more forgiving. But that doesn't mean that you can't specifically ask for specific extensions. I think that's that's reasonable! It's just that! We didn't have the need yet, but it would probably model follow the same model me, but you don't get to numerate the extension you just get to ask: do you have extension X and that thing can return true or false.

C

Yeah, but that would be a browser, API specific thing.

A

Yeah they're encouraged.

C

To follow the same model, yeah.

E

All right so Mac OS will be Mac. Os OS X will be OS X. The LS platform type will have no equivalent comparison logic at all right. It's just super simple and then runtime information is OS platform. Yeah.

B

E

Handle the equivalence checks and, if needed, the analyzer would duplicate the equivalence checks and.

A

Do be clear like even if you were to make a west platform having equivalence checks, it wouldn't help you analyzer yeah, unless I can't call it right, because analyzer runs on full favor right.

C

And the analyzer is gonna have to do it anyway, because we're adding Mac OS, but the the rid is OS X. So the string that someone is gonna have using this new.

E

C

Were just discussing is going to be the old one, because that's what it.

E

Is the only reason why I was thinking having runtime information to have a helper method to check equivalence between two platforms was so that the analyzer could also call into that, and then this platform would call into it the same way. But I don't know that was worth making that a public API yeah.

C

And that analyzer can be running on something older than that five, so the API wouldn't exist a call so.

E

A

A reason why it's called aesthetic analysis but ya know it's a common problem. I mean like that's, that's another reason why we wanted some of this TFM data to be more static and msbuild as data, so you can create for multiple parties, but like it's actually hard to do that Sofi to this day, don't have that.

E

Just kind of realm since last night to that you and I talked about trying to do support like Windows one. You know Windows eight one without a dot and having that get inferred. The way nougat does of inferring the dot between the eight and the one yeah.

A

E

That brings back the ambiguity, and you know what does 14 mean? Does 14 mean 14 and does 14 in 1.4, so we were headed down the path, we're just not trying to infer dots always requiring at least one doc, and the versions of iOS 14.0 would be required in iOS 14. It's not supported so.

A

What happens if I don't put one this analyzer give me an arrow for that or yeah.

E

A

E

Let's see, how would that surface.

C

E

C

You applied the minimum of OS, whatever attribute, and it says like this doesn't match my regular expression. So this is garbage and it purports an error. I.

A

Think I'm! Okay, with that, like think, if we, if we, if you want silent behavior for those, then I would have said we should match what nuke it does, because then you have at least the same problem. Yeah.

E

And that's where that's what I've started the wind toward last night, but iOS 14 feels like we would be leading people into a pit of failure. They're both everywhere.

A

E

Right I was 14 assume that means 14, and we would interpret as one point for that.

A

I mean it's the same problem, but we said like in the TFM, because we can't undo it when you get right because right, let's, fortunately ship behavior but like we said like because we both have dotnet 10 soon.

E

A

The current versioning speed right. We basically said that's why we burn in the point Oh in the end up end up in the project files, and we encourage everybody to put a point zero, but it is still equal to not have the point zero and that five still means net 5.0, but also net 10 will mean net 1.0, which is not useful.

E

A

That works for me all right. So then, all.

C

Right, can you have an issue from yesterday that, after the meeting got some feedback that I want to put before people so one seven, five: nine, zero, five, nine zero five.

A

C

Yeah so the thing that sailed through super easy I used hash as the method name, but there's already a property named hash turns out. That's not legal!

C

A

I can see then so.

C

The I think the best suggestion so far on alternative names is hash data as the verb, so hash data, hash data, hash data and try catch data.

C

The instance method is called compute hash, the other one would be get hash, but if you'll get and compute together feel very weird, so oh.

A

My god does, it know mean I can finally compute a hash code in one line of code. Yeah.

C

A

I, wasn't here like yesterday, know what there I guess: I didn't pay attention. Well, maybe.

C

Yeah, it was crypto, no one cared so yeah, so the proposal here and then just replace all hashes with hash data is I. Think what I am happiest with the other one would be stateless, hash or static hash. But that feels more like we put a word here because we have to I mean honestly, that's what we're doing, but I.

A

D

Clerk is special.

A

D

What I'm doing is I'm just looking at all the static members that are already on things like sha-2 food, sex.

C

Beata, the hash algorithm based class to find a property called hash which I forgot about and I mean.

A

What about a hash at all.

D

Well, so here's the thing like when I type, sha-256 dot I, see create equals and reference equals. So, honestly, anything that has the word hash would be clearly the thing I want to use here: yeah.

A

D

Have hash tag, though, seems just fine books.

A

D

Yeah one, what again, when I hit Don I'm going to see four things once this hash great yeah.

C

All right so I'll update it to say hash data and then I guess you need me to do bounce it through ready for review and approved. So it gets picked up.

C

Gets picked up for four, what Emo's video scrubber yeah.

A

Because yeah, it's nice to have notes, as it turns out by the way for the designs I, also edit labels. Now, so we can also mark the designs.

A

Which, of course, I didn't do, which makes you makes.

C

A

Slightly useless but I can we met.

C

At the feature, so good job.

A

It's under a poet fest, obviously.

C

It sounded meaner than I intended, but you.

A

Know that's, it was supposed.

C

To be hard, sarcastic, polite sarcasm, that means our guests. Listen to my intention.

A

All right, so it's no more blocking, we said API needs work.

F

All right, good enough.

G

C

C

Anybody have a 10-minute issue.

A

As a picker vendôme, one.

D

Are there any marked is blocking no.

A

I checked earlier this morning, I mean unless somebody marked it as a meeting started.

D

Give me one sec.

A

Okay, I've launched my viewer app. It just takes a while, which is very sad.

B

I've actually got one that that I think is really small and would be useful to get in in the next five days. Okay,.

A

B

A three seven, six, eight one.

D

Four hours later.

C

And then I have a one-minute one on standby. If we finish this one on time,.

B

So, basically, for a linked list: node we have an ability to expose the under. We expose the underlying value, however, there's no ability to expose that as a ref, which would be incredibly useful, especially for large trucks, to avoid a copy. This is completely safe to do, unlike exposing the underlying underlying array for list of for for an actual list.

B

So the proposals just expose public ref, t-value, ref, ref item and.

A

There's already a favorite property right, yeah.

B

And value ref matches the name that we've used elsewhere and.

D

Where is this end up getting used where the perp is important? Yeah.

A

B

So I actually there's a native memory allocator that I'm porting from native to manage that uses a linked list over a value type. And this is going to be for managing GPU memory, which obviously can't be managed by the GC and.

D

Linked list is the right data structure, but.

B

That's what they use in C++ to.

A

Yeah I mean I will say, though, like I'm kind of Stephen, like that I would yeah, maybe I'm sceptical, but like because I mean you're right that in C, both of us use linked lists but like linked lists in C++ are usually better perf wise. Then this one is there's a lot of interaction going on and I'll be curious if you're. Actually, you might.

E

Be better off for, like.

A

With your own linked list, basically that's my point: it's not so much better use linked lists over arrays, it's more like if you define your own linked list, you'll be likely to have better proof than having this interaction through another major type in the middle, our own.

B

Linked list is actually fairly efficient. It's just a simple class wrap that the nodes are just a simple class wrapper over a value that points to the next node. There's.

D

Not what simpler.

B

You can get than that. One.

D

Nice thing about this, too, is, if you're allocating a bunch of these nodes, one after another, you're going to get a cache locality as well. Just due to how the GC works, that we do have a lot of data structures that have T's as Get Set properties. I, wonder if not just here, but we should take a look through some of those other data structures and start exposing more refs throughout the framework. I looked.

B

At some of the other data collection types- and there didn't look to be anything that could have obvious- that could obviously be done. That was also safe to do. Okay,.

D

B

D

For high performance scenarios, like one argument to be made is hey, maybe actually we should have been using a field set of properties here, because then you don't have to worry about having a value rough now, I.

C

Guess the well I guess a ref can't be re-written to be null or can it you'd have to have a ref to a ref.

B

To make that retinol, yeah.

C

So this, so even if the setter is guarding null, if this wouldn't bypass that, but it would make mutable struck work.

B

So there is no setter value.

A

B

There is a Center for value, but it doesn't guard against no lit just assigns whatever you give it fair.

C

B

So if T was a class like an object, you could assign the underlying value to be null, but you couldn't make the ref itself null.

D

Yeah this this API isn't exposing anything you couldn't already do just like Tanner said it's just allowing you to do it in a more efficient manner. Yeah.

C

D

C

Like list, it seems weird because with lists, if you got a ref to a the position, 5 remove, like a bunch of things, can make position 5 no longer part of the list or now part of a different concept of the list. But this the node would have been removed and now it's removed so you're modifying dead memory. At that point, which is fine.

B

Outside of an actual you, you can have a node, that's outside of the actual link list, because it's a separate entity right and you can insert it into a new separate list independently. All.

C

Right this is what I'm saying this one seems safe to me, because you're talking.

B

C

You're talking about the thing in this box, as opposed to with accessing a ref to an array which is what list of T, would give back. That's you're accessing a chunk of memory that may change its semantic, meaning right.

G

C

Operations on the bigger type right and.

D

It looks like we've used. We actually do use value ref as the method name for other concepts along these lines.

A

So well, is you a quick one, Jeremy hold.

E

Up before we jump over to that one, there was a question on the set of cache helper methods on YouTube that we missed Fredrick had asked what about hash bytes.

A

Yeah to me a hedge by it's has all hash data. They all seem fine, I, don't have strong preferences, I.

C

Feel like I know, we've used data elsewhere, I feel I. Don't know that we've used bytes but I can I can look and make a final note later.

D

Yeah generally, when we include a type name like bytes and an API, it's because we want to clearly call out the fact that it's dealing with bytes as opposed to any other type of T in.

C

This case, there's.

D

That distinction is not necessary. Yeah.

C

Data feels better to me, but I'll, look for similar prior or mo just wrote me a blank check so.

A

Yeah, the only thing we care I care about I think so much Alleva is like make sure the name hash is somewhere in there. It's a little bit I, don't care.

C

A

C

Eight six, nine.

A

Sorry: three: five, six, eight nine, eight six.

C

A

Why do I do this.

C

I would blame German number orderings, but you're not miss you're, not flipping the last two. So.

A

No, like that, doesn't work for me.

C

This is just span overload for a thing. We have an array for.

A

Well, then, that seems fine I.

C

Told you was easy, I.

D

I will bring up the issue that came up last time, which is: do we have to handle mole or empty in a special matter?.

C

The array one handles null if you've turned it into a span already. That's your fault. All right sounds good. That's that is literally our answer everywhere, except for where it isn't so doing something other than that is the exception and I I think I own 90% of that special case code, which is all in things like the KDF for opening PF X's.

C

Because that spec is weird.

A

All right, then, I think we've done a few useful things today.

A

All right, then I guess I, guess see you guys next week yeah or some of you, maybe even later today, Thanks. Thank you.