.Net Foundation Design Reviews, 20 Feb 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Review: Quick Reviews

Description

00:00:00 - Approved: Add CancellationToken overloads to Socket.ConnectAsync and Socket.AcceptAsync https://github.com/dotnet/runtime/issues/921#issuecomment-589234342
00:11:07 - Approved: Socket.SetRawSocketOption https://github.com/dotnet/runtime/issues/865#issuecomment-589239349
00:22:22 - Approved: Update Encoder, EncoderParameterValueType to match GDI+ https://github.com/dotnet/runtime/issues/30543#issuecomment-589242229
00:29:12 - Rejected: Add ParallelOptions constructor with MaxDegreeOfParallelism parameter https://github.com/dotnet/runtime/issues/30652#issuecomment-589246889
00:46:31 - Approved: SequenceReader(T).TryRead overloads to read a specified number of elements https://github.com/dotnet/runtime/issues/30778#issuecomment-589257975
01:05:23 - Needs Work: SequenceReader(T).TryPeek overloads to read a specified number of elements from any position https://github.com/dotnet/runtime/issues/30779#issuecomment-589267066
01:24:38 - Approved: Add generic overloads for MethodInfo.CreateDelegate https://github.com/dotnet/runtime/issues/30800#issuecomment-589273550
01:39:39 - Approved: Add TaskCompletionSource.SetCanceled(CancellationToken) https://github.com/dotnet/runtime/issues/30862#issuecomment-589274248
01:41:12 - Approved: Sse2.StoreScalar(int*) https://github.com/dotnet/runtime/issues/31179#issuecomment-589275674
01:44:34 - Approved: long Math.BigMul(long, long, out long) https://github.com/dotnet/runtime/issues/31184#issuecomment-589278166
01:50:05 - Approved: Avx.CompareXXX convenience methods for float/double https://github.com/dotnet/runtime/issues/31193#issuecomment-589280411

A

I also can't see what Evo can see because emos not looking at what do you, milk inside.

B

So we know live so I can tell you that, um while looking at it I know that's the case whoops, let me look at it again. Oh it's delayed there like doesn't.

C

Matter we never really make use of that. Like thirty. Second of some of you felt we're doing well,.

B

What would you do in those 30 seconds? Do.

A

We have it, take your mouth.

A

B

It's just this thing of you late from YouTube, but that even if something happens like what's the worst, they can do like I mean how bad of a profanity can you issue that we would shut down the stream I? Don't speak German.

D

So I don't know.

B

All right, so, let's look at this guy again, because we last time we tried to finish it. We didn't sure polish, also hans schindler.

B

Let me zoom out a little bit. Maybe so we have a chance through all the members of this room, just not able to see the okay I have.

E

A 20/20 vision, okay, I'm.

F

Close to the screen.

F

B

Think I will order some binoculars for review.

B

While we were in the middle and then we said, we need more time, I.

B

Think last time we discussed why we should return major task over tasks and I. Think the conclusion was it's okay, but thought of it doesn't do much yeah.

G

I think it's okay.

G

Under the covers, we have a pool of socket a sink of that darks and we can utilize those to be the backing store, the value tasks. So, whether we do it occasionally blue text or not I, think it's it's both reasonable here that we would optimize it and the 99.9% use cases. It's just a weakness.

G

So now the usability impacts, oh, not a huge deal, yeah.

B

Because they waited all the time right so.

E

Sure the constellation Dawkins be optional. I thought last time we.

H

Said they couldn't because yeah.

A

Because they, if you don't have it than the the shorter signature, would win, because the exact match always be something that was inferred from the people's.

H

Right so they're functionally not optional, there's no point in making them optional, because.

I

H

The other overload yeah.

I

Cuz you either explicitly come to the house.

E

Yes, so we could make them adjust. There's no point: that's the conclusion. Yeah.

C

Yeah and and Steve there's no there's no weirdness with these overloads of returning value to house when the other ones returning tasks, whatever Thanks.

G

Yeah we've talked about this in the past and it's one of those gray areas that the concern would be. Someone was using the return value in a way that works, fine with tasks and isn't no nois.

H

Value tasks and.

G

They go and you know we would analyze. You know you should really pass in your cancellation token here and now. They've changed their semantics because they were using the return value in an appropriate way based on the overload that was called and it doesn't show up anywhere in their code, especially if they're just dotting off of it or they're, using with our which won't even cite you don't like lar.

G

So it's an issue in the in the sort of you know weighing the balance. I think value test probably still makes sense, especially for accept, because it's pretty common for accept to return synchronously. You know you've already got a connection waiting and you're listening backlog, and so you accept it. It's just right there and see you avoid the task. In that case,.

C

And your analyzer should also help improve the health of the ecosystem or in that house right so I.

G

It's not doing it's, which reminds me I need to actually get that checked in other.

B

Acing methods and Socratic stuff extensions: what about those.

G

G

But I don't remember it I know: buddy Tamar should open up an issue around some of the UDP focused api's I. Don't know if that Plus. This is the is everything or, if they're still again,.

B

But there's no reason that we wouldn't add overloads with cancellation token for the others. It's just additional work, but there's nothing wrong with it. I suppose, right, yeah.

G

I mean we need to look at the individual methods to see if there's anything, that's not actually cancelable. Like me in the discussion, it's more about whether you should have a disconnect, async I, don't know big I, don't know yes or no, whether translator you talk to make sense on disconnect async.

G

Despite you, we could consider on a case-by-case basis, risk question right.

B

Well then, what about the other ones I.

B

Guess so, can we don't do anything which convinced also that cancel except they think is.

B

J

K

That's the new API ordered to do the other thing or.

B

K

G

It's a weird pattern, but it's following the existing.

L

And there's anything above what was this different from this one? On middle word,.

B

Oh sorry, again, let's see.

G

Then both the TCP client in TCP, listen, I, think those are just thin wrappers around I. Think in general, all the methods just delegate to the underlying socket.

M

So there's not much work well done. Yep all right, I.

G

Do think it wouldn't be worth subsequently someone taking the action item to go through and make sure that I even have an issue or we open an issue covering all of the remaining socket methods so that we stop doing this piecemeal yeah.

A

Volunteering yeah.

L

Just basically.

L

All right, so let me just paste this back in.

B

hmm Let me check Cory on this one because he was just volunteered.

B

What's your github thingy scalable.

D

Corey, oh that's right, but do you scale.

C

You can just start talking the person's name, names yeah.

B

It turns out some names: they're not helpful, like my name- probably works, but if you say Jeff, Joe, Mike, Cory, there's too many that dick suggests an autocomplete before in the last four walls. Maybe it's I.

L

B

L

H

It doesn't based on please in the org no and who's actively on the issue and.

B

What move phase them I think.

H

L

Gives you what about Corey all right? What's context on this one Stephen, you have context on this one sure so.

G

B

I think I don't know like I'd I, don't think there is my and any way of doing that.

G

I'll, try and talk through it so on song can be supposed to set Sokka gumption method, but because of how things work- and there are the underlying values on windows- are different from the underlying values on Linux are different, they're lying values on Mac OS. So we have to do a translation layer between the actual values that are specified and what we passed the underlying OS. But that means that if you pass something, that's not in our translation mapping, we fail.

G

We just and there's no way the public surface area for you to say, but I want to use XYZ option on Linux or X. You know ABC option on Mac, OS and kind of say: okay, if I'm running I'm going to accept this or if I'm running on Mac OS set this. So this would basically be a way to skip our translation layer just pass directly to the OS, rather than you having to do that. You invoke yourself and you would be responsible to for knowing the right codes that for the OS that you're talking.

G

Between cities in the state patches, we don't provide the option doesn't prevent you from using something. What.

C

We provide a list of ones that we know about if.

G

We knew about it, it would encode it in our and then you call the other one right. Yeah.

A

What's get raw socket options supposed to do if option value span isn't big enough broke whatever the current.

G

One does I guess: well, the current.

A

One allocates.

G

Okay, then yeah I guess it would grow well.

J

The current one you specified length, so it will have some behavior. Your length is not enough.

J

B

We have said the existing one. What do you mean by existing one to comment that does come existing it gets signature? Oh oh.

E

N

Size of the byte array or device change over time like, depending on the option you pass in the data returned, can be different size but like, for example, when something changes into your s. Like.

E

Windows, would it change this size I can't.

J

Think of hand of any any options that are like that, because.

N

J

N

E

What even, is completely you know like if it's not a constant size, I.

N

Think we said, oh well, sorry, there would still be backward.

J

Compatibility like binary compatibility with older, older sizes like I. Imagine there would be some sort of like like a lot of px. Struts have size and.

E

The point is like: okay right can I call this method and know that it's not.

G

My concern I, hear you saying my concern with the try. Is it may be? Well, first, I I'm with Corey I, came to give any case where, for a given option, the size of its value would change, certainly between options, but for a given option. I can't think of any reason it would my concern with the try is when looking at it. I would actually think that it was talking about.

G

The specific option is specifically because this is about you know getting the value for a specific option on a particular, less I can imagine thinking the tribe mentor well I'm, trying to get the value for this number and if the number doesn't exist that don't worry about it maybe I mean maybe that's not the case, but that would be my concern. I think it would be clear what the tribe, what the one thing is. Try was referring to.

A

Is the so in the current? What is option link the thing that you have to give it the right answer for the option? Yes, so this option value not length, is you either have the right answer and it works or you have the wrong answer and it gives you like whatever the unsupported yeah cuz I, give you called like I, don't think that that address is an option, but if you felt like you know, get madder and it's like well I'm! Sorry, you moved from the ipv4 to ipv6. The answer is bigger.

A

Now you told if I get a tour. I have four bytes and it's like nope yeah, I. Think.

J

The cases worse, you would use a try here would be I can't really think of anything specifically around the byte buffer being too small, and it's.

J

Great and the business.

M

J

Really is being added as a method that lets people like they know specifically what platform they're on, and they want to do something that we just don't support right.

A

But my point really is: if these, if the span length, is effectively an input into the underlying thing- and you have to get it right to get the right answer, then we don't need to try because there's no Chrome pattern. Well so.

E

Just so I is the method throwing if this man is too large for one when it's too small I.

G

Think only when it's too small, because I think.

E

Jeremy, so we don't have to have exactly the right size.

A

Well, like a lot of win32 stuff, is you give it a size of thingy? If you give it a thing it doesn't know about, it tells you like I, don't know what version of the future you're from like I can't fill in your structure, I.

E

Mean it's such a advanced idea. It almost doesn't matter what we do here. Yeah we're really just adding.

J

It so people don't have to clean, go, get themselves.

L

Yes, you should, you should should be back actually there's nowhere for the wonderpana go because.

L

A

A

Okay, that's the thing screens can do now. Welcome to 2020.

O

A

Apparently, today is just full of 2s and zeroes, so.

B

You should be seeing the screen now, of course, I don't get it bacon. Oh yeah, that's.

A

Meaning YouTube can't see.

F

H

P

P

A

Maintain the status.

K

Even know what I just did I.

A

Think now we're projecting and recording for three seconds of the future.

B

Okay, this one works. They like.

B

Well, I guess: I cannot close, think I'll. Okay, it's like! If you immediately go to we're not washed up due to structure or something, then that's not going to be like any better than the side of the train right. So.

J

Actually, if you're using cork you're doing this quite a bit more than what.

A

Yeah you, your probably get there just that.

J

Yeah with quark you, you unset it and then said it.

J

You are optimizing or packing size.

K

B

B

B

Somebody how we are.

C

So, who here no system call, you know what that sounded, like I have contacts as.

A

One of the things that we changed the guidance for it in this version of the book. Perhaps these static read-only should be lazy. Initialized get only properties.

J

Their existing.

A

Do we need? Do we really want to make the the CC tour do more work.

H

In line, then it's not the other C C's were.

A

The pre CC tour.

C

I mean they're only initialized if you actually use this type yeah.

A

But if you use any part of the pipe well, no we're not using all of the encoders.

H

Looks like these are just options: yeah news, new entries.

Q

Comes down to how much we care about consistency with the existing ten options, I'm fine, either way, I just brought it up for this I love.

F

The name of this.

C

Know this is an 15 drawing for some reason: it's not being a captured by source format. All because.

L

It's pulling the different people, all right, they're, not being the victims.

A

Before not all right, source stuff doesn't maybe having one of the days where who packages don't get scanned or yeah.

I

J

A

I

J

For this EP, what happens if you're, on an older version, I.

C

Mean these, so these literally are just magic strings for lack of a better term. The encoder class literally just wraps a quid it's kind of like oh ID, but whatever it's kind of like oh ID button immutable. What is this time? The system don't run?

C

That's being quarter live. Oh sorry, I'm late I'm because it couldn't do in space here, a system-wide imaging, oh, that might be like.

B

Okay, no I can find it tight. Okay,.

C

That makes me so yeah, it's literally yeah, okay, it's just a bunch of Quinns. No, it's a strong, the tech. Do it yeah, which is pretty normal for.

I

O

B

Windows, for that matter,.

M

L

B

L

R

A

Want you to know about my CPU.

R

Is going clear, these are just accelerators for those.

A

Sitting, presumably when you try and use one on a system that doesn't understand the magic answer, you get a some sort of error that is documented from the places to take one of these things. Well,.

C

I would imagine it would be the same as have you passed just a random like would not knew what enough yeah, which one ready has to happen to find behavior right, yeah, I.

A

Guess it's I guess the real question is: is the encoder constructor validating that the ID is known or.

H

Religious, what is it whenever you.

A

Pass it into something else, it's just a yeah.

H

Why is it a class you'd.

C

Like to see because it was because.

B

We defaulted to class I think that guy might have been written by the person sitting to your left. No, no.

L

That's what my default views classes on this prove another thing. Yeah.

C

So it's it basically is it. It's held TDI how to interpret the Act so imagine that this is a parameter type and there's a block of data. That's passed in with it, which is the actual value, and this tell TDI and how to interpret the byte buffer. That's being asked him.

A

Are you still with us Steve yeah, okay I got a weird there re notification that the meeting had started.

A

Thank you miss being weird.

B

Alright, so then I mean this seems.

D

J

Straightforward I feel like I'm, ok with it, as have we verified these things. Oh that.

E

Would be part of testing like that, I mean, and what do we do again if you on the OS that doesn't have these features whatever.

B

Gdrs behavior already is, and I mean they mean we have in the past at some point, dropped for supported versions of Windows right and then I. Think you just get platform unsupported or you know, they're kind of have to attack, ok, loves there or anything.

D

This is alright, so then, one down again.

B

Even I think you're up again because you're a bully the closest person on this one.

G

This is fine, this power, so that this class is a just a grab bag of options and the idea was you constructed it and use you know the initialization, syntax or properties to set the options. This is just asking to promote the most common one to a constructor parameter on a new constructor.

C

G

C

Did this initially.

G

Thinking like there'd be a whole bunch of other options. They ended up being three, and so we hadn't added constructors, because if you're here, I will use this initialization pattern. But.

C

What what about, if you ever want to add another int property to this class in the future.

G

As with any other constructor that has an int, it would have a very specific meeting and you know you could easily add the property and then the question of adding another constructor. If you wanted to, and if you ended up with multiple hints in your constructor, you need to know what they meant. But what.

C

G

It's not like property.

C

Here, because what I was trying to figure out is if, if in practice, people would end up using named parameters in order to differentiate the different instantly constructor and if so, why not just use the object, initializer, sometimes right.

G

At that point, you would just use the initialization subtypes so.

C

G

Request is really about saying: new parent little options instead of new parallel options, open brace.

E

G

Great parallelism so.

E

Would it make sense to go even further and have a implicit cast from the introductions? Oh.

G

It does, it casts the allocated well.

E

G

Allows to me that seems scary for.

O

G

Never mind the allocation, we think the same parallel options. Options equals eight like what the hell is. Eight.

D

A

Thank you do not be cute with.

H

Well, I I mentioned.

E

That because Steve said we added this time, there are many. You know there are three options, but actually most you know I understood. Most people just won the games. It would be an easy way.

J

To let people just pass in aids to all those without any.

H

Yeah or maybe a default way to say, the max degree of parallelism is your logical processor, which is probably the most common scenario. I.

B

Mean they're kind of I would say at that point: you probably want to just change over lots of for each right or whatever other common methods take parallel options today, like.

C

The thing is worth more.

B

C

Only saving someone from writing like 20 characters by.

D

The NSA I mean.

B

Well, I think part of the problem is that you write for each loops in line right, and you want to like. Logically, you want to make just because you make it parallel. You would like to have the for each or the for loop still look very close to what you would normally write, so we do producing characters there. I can just make it a little more similar right. Then I then argument that by a lot more than just did I want to save some characters.

B

G

But in that case you can also just create one of these instances. You store it in the static somewhere. You pass it everywhere. Gonna.

A

Be just like saying, you know that's parallel with it's as parallel options: yeah I! Guess it's fair that keep.

B

Pull we don't one, the only thing you may want to pause it. Every single time is the conservation token right.

G

Yeah I mean if there are three properties I remember correctly. There are three properties on here: there's the parallelism, that's the translation, token and there's a scheduler right and they sort of an order of usage. It's that order. You know most usages, specifying the parallelism. The second most especially in token in distant third, is specifying a scheduler right.

B

Yes, I mean I, think insulation talk is probably the only one where I could actually see as promoting it as a as an argument than two for each other. If you cared because that one seems to be, you know, variable versus the other. One seem to be mostly invariant to what you're doing I.

C

Mean you could always just make the constructor take two parameters and make the cancellations we're kind of default.

C

That way, you just have the constructor there for anyone who needs a horrible.

J

K

Haunted cases, but you may want to do that right.

A

So if you had it, if you made the constructor where it took the parallelism, comma, a defaulted, scheduler and comma, a default cancellation, token yeah or whatever order, we think makes sense, then that's the you have the go, build this constructor. In addition to the creates that call.

G

So the only problem with that I'm looking at the type just yeah so more, better or worse and probably worse when you construct a default parallel options. Instance like no parameters, obviously there's a camp, a sedated, a scheduler defaults to default. But if you explicitly set the task scheduler property to null, then we will treat that null as meaning use the current schedule, or rather the default.

G

So just we have yep.

B

So should we then add the optional default or slightly optional, translation, children or not I mean.

C

If, if it's common enough for people to want to specify it, we may as well make a convenient if we're already making this, how.

J

Common is it because this is a blocking call rate; it doesn't return a task.

G

Right it just pretty, it basically causes the loop to stop processing further iterations and then throws they.

C

Think of it, like you, have a UI based application, which is doing work on a background thread. Cpu work on a background thread. You get a cancelled on me. Yeah like you want to. Yes, we'll come with it. You know.

J

You see anything wrong with.

A

Having it just, oh, it's a deal yeah, given the task scheduler default versus null being current distinction. It seems weird that if we that the we can explain the int one in terms of the existing default constructor, but it feels weird like it feels weird to me, given that you didn't get to specify the value in it, but it has a value and it's not known like, but.

B

That's true: we gotta some what we do yeah, even if it don't add a constructor, that's the behavior. We currently have right. So that seems.

A

But it feels a little different of like this is the this. Is the default state versus like this? Is the state and you gave me an int and like I, also made up some other things? It just feels a little wrong. I'm, not saying like no I'm pointing at a you know a little wrong I mean.

B

To me like saying, scheduler is rare, so you only ever said it as a property seems fair, so we don't have a constructor and then I think it's also fair to say: yeah. There is a default that is not now, but that it's because it's a property right I think it would be weird if we had an overload that takes a scheduler and then the one you called it doesn't have one. Then it's not now. That would be weird I. Think the.

E

Only thing that I worry about this I I feel like the concurrency level. You would probably want to create this basically static, store it and use it all. The time cancellation token you kind of pass different one every single time, so.

A

You probably just pass null or the the non the non token every time, but maybe maybe imagine you're passing you.

G

Can also just you can also just capture whatever you're using in the body of your loop, which is what a lot of you said this of doing those naming. You know the downside to that. You know. That's basically,.

E

What I was trying to make I would probably not use the cancellation token, because the concurrency level I would just have a static. Therefore, I don't know if it makes sense to and if there's a constructor parameter, I.

G

Think if it were up to me, I think I would rather just leave let sleeping dogs lie or just constructor with the integer, but not go further than that.

B

A

Think, since what we're really saying, is you either you're like that you need to make the local or the static or I guess an instance field I, don't know that there's any reason for change, but if we want to add you know one, that's funny yeah.

A

E

Only this thing was destruct yeah. Maybe then we could cut any visit cast.

F

M

A

You know it's like you see the.

O

Surpassing eight you're.

A

Like take me to this method, it's like one that takes a parallel options: you're, like no I passed, eight an.

C

Attorney attorney the scenario on his head: if the, if the real answer is that people are trying to specify a max degree of parallelism and the knew enough of parallel options is the only way for them to do it. Should we actually just have an accelerator straight on parallel block for each listen area and ignore parallel, ignore changing the parallel options. Class itself.

J

H

Also, parallel for we're, just eight is going to be confusing in relation to the length, but.

C

I guess what I was getting at? If, if, if there's like one or two overloads that we could add to parallel off for each that, would get 90% of the way there. That seems more reasonable than making the change in parallel options, because it would just make the overall call site more.

G

This was the example here was written using for each but I. Don't know that the predominant.

H

Okay, I would have actually imagine that for is more predominant because, like if you're doing some over in array you're not going to for each over the entire array, because that's incredibly inefficient, you're going to allocate a like sack allocate in array. That's eight! Why? Because that's how many cores you have you're going to segment the array compute so that you don't have to do any kind of interlocked operations and then you'll just do a quick four over the final eight resort has to be the final something.

J

I'm kind of with Jim Jeremy.

B

Little I mean, if you don't want to do it, then you can also Jake I.

H

Personally, having used parallel for a number of times and not liking, the default behavior of I scale indefinitely according to the task scheduler I would enjoy. Having an overload that allowed me to specify trivially the max parallelism, an overload of water parallel parables options, okay, because because that just applies to everything we don't have to consider about the like twenty overload to four and forty.

E

Something I mean by trivially: you think this constructor is really specifying it or not. Yeah just call in parallel options, because kids and you'll end up in line when you become free.

Q

E

J

E

I am a bit confused. You basically are saying you don't want the implicit cars because it allocates well. This allocates to an asset allocation is a minor.

G

The most part, the work you're doing here like allocation, doesn't like you, and only if you're, using this you're doing a lot of work by initial, like oh, that allocates was sort of a set, a separate thing. Okay, the.

N

G

Concern with the implicit cast is like I. Imagine that first parallel for line and it's a zero 108. Yes and like what so this one is not me yeah. The separate concern is like if I have powerful options, options equals eight like it's, it's very confusing. What is what is that mean like there's no concept of number going to in hobbies, so.

E

So, let's, let's one by one so you're saying at the coal side, there's gonna be a bunch of integers and it's confusing what they are. Yes. Well, we have this. Don't we have this problem all over the place like this is a problem with source code correct like if you really want to make it explicit what it means.

E

You know we some people use a code binary, otherwise here coincide. Samsung sometimes are not very transparent. What's going on they're, a bunch of you know.

G

H

Like we add one constructor or we have take 40 methods, so with a look there.

E

Is no need for confusion. It's the confusing you just created parallel options and you pass the age. I I have no idea what that means. Well, I thought it was an enum value at first, so yeah. So.

Q

E

Is still confusion, I, don't know what we, if it really curves.

G

Then then the existing thing is the best. There's there's no ambiguity with new parallel options. Max degree of parallelism equals eight yeah.

C

We should just keep them it. It's seriously not like this. Ati is not allowing any scenarios that can't already exist even on a single line of government. I mean I'm. Okay, with.

B

The API is goals because its inner said, if you're white, as a lot I, can see it being slightly better.

H

K

B

R

B

You're saving two characters: yeah.

E

So anyway, my stance is, you know: I buy with I I buy the argument. This is kind of very, very boss, and it would be nice if I could just pass the parallelism, but then I'm saying we neither go all the way out and yeah I know I realize many people don't like it, and if we don't go all the way out, then I don't know how much it improves it's still ambiguous. It saves you four characters.

H

Constructors are the default way to initialize objects and initialize your lists or additional sugar on top of that, and so every time I come across an instance where the constructor does nothing and then I have to go and call them assures it. Bugs me good at.

A

H

A

H

A

I would I would agree, should have the constructor if we could do it for all the parameters, but given that we can't encode the default behavior of the task, scheduler I think we should just leave it as it is.

I

S

No implicit conversions I, don't like that.

A

S

No, this is just a very.

A

Old create set call yeah.

B

I'm I'm confused on my API to will do after this I. Don't think approving it and they're not approving it, but.

M

It handles very well, but you don't see, let's trust after more discussion.

L

B

Sequence, reader of T try to read over those with a specific Betsy for a number of the limited. Didn't you look at this before it looks familiar.

C

So this this API very much confused me when it first came through because on most buffer read ad is the span that you pass in is the maximum buffer that you're able to read, and then the API returns the number of elements that had actually populated into the buffer. These api's turn that pattern on its hand. These API say you must fill the entire span with data, otherwise fail and.

A

They're not filling a span they're producing a span they're.

C

Related to yeah, they're, they're related to other api's that are that are being used in hmm so.

J

C

They lying slice.

J

The underlying story- yeah.

C

So if you can't so, if you can't read three bytes, for instance, then it won't add anything at all. Try read will return false I believe I.

E

See and there's this subtle hint with they come even past in yes, yeah I.

A

Mean if, if we think this is valuable, I think it should be trying me to that yeah because usually account is somewhere in the minimum or maximum range, and this one is both yeah but I feel like oh try to meet seven like that's, not to return five right, yeah.

C

Jeremy kind of comment on what I was sending them like we, we need to frankly whether this API is try to read exactly why I read up, because most of the other API is on sequence. Reader are exact, not up. Hey.

E

Why do we like what's the scenario for this like what? Why do we want exact number.

C

Of body, because it's mainly used for things like networking protocols where I need to be able to read I need to be able to read a 64-bit integer off the wire I, get to read exactly eight bytes in order to have meaningful data, because otherwise you have to do the dance that.

B

You read, for we did another fool and you stitch it together, and then you actually can do something about it. So if the data is seem.

E

Different fathers, it will allocate Oh.

C

Is there a common I? Think, that's I think it might just put it on. You know, afford it. We.

A

Talked about that before was that this or something else I know we have talked about it before, and it's in a thing that makes me not happy with understanding.

C

A

Are lifetimes of things.

E

Well, they are safe unless you say otherwise, so I assumed.

H

A

E

Allocate a smaller area copy the data and give you a span over the Ray, which you don't care about the lifetime. But.

A

I think if it fits in one segment, it's gonna give you a span from that one segment and then it can refill it later after you called consume yeah. So sometimes you own the memory and can't do anything about it, and sometimes you don't don't it then shouldn't do anything about it.

E

J

Has the same lifetime of the span being worst, and if we give you an array, you can't tell you have to treat it as it was so.

C

It looks like the behavior is, if it has that many bytes, are they like looking at fellas comment? If it has that many bytes available, it will return a slice of a single span if it's able to otherwise it will allocate a new array. I can County two strands together and then help that to you. So this is an allocating API. Potentially, that leaves based on.

E

Colors columns, so we we do it because we believe that stitching, like passing in buffers in stitching is nuts. There is already one in.

C

Place in bars because.

E

In case the date I started, did that yeah I understand that, but.

C

The thing is this: API will return true if the data is already there, but in multiple discontinuous buffers. The problem is in that scenario. It would have been cheaper for you to provide your own buffer to help and with this API it's now saying like I'm going to do the expensive work for you, the expensive work you were trying to avoid it well.

E

The thing is that if, if the account is usually very small yeah, you know it's the optimization that you said when you pass in your own daughter probably would be useful, but if it's very large than it wouldn't yeah well,.

J

Actually, we it's kind of the reverse you're expected to have fairly large segments in your sequence. So the idea that this would allocate is for, like very rare scenarios, where you're crossing that boundary and the larger thing you're trying to pull out the more who you are to be crossing that boundary yeah. Well,.

C

The very particular scenario list of here is where the length isn't felling I've compiled tagging, so I got a real-time and then and.

A

Then, as when you get big enough that your LOH, that means the the GC is not gonna, reclaim it. If it was us, if you asked for 8 and it happened to segment, then you're gonna get a gin, 0 pickup on your thing, if you read the data and then discarded it, so there is your so it's.

H

A

Number you're asking for the less you want to use this, but the more you want to use this because it makes your work easier. So it does feel that this is a performance trap for the title that you're using, because you think that your performances.

C

Yeah this this is meant to be a high performance API and, like Jeremy said you can't you can't reason about the performance of it. I I think this has to go back through. We designed.

E

So the alternative would be you pass your small stack allocated. Let's say you want to get a name. You want to get four bytes yeah. You would prefer the value to be from the actual buffer, but you pass a separate buffer as well yeah you pass any slip span and.

A

If it I mean I, don't know that I like that for a public signature. But if it was, you know some signature that you gave it a span of TIMP and then it will give you back either. I filled temp because I had to or I gave you a slice, so it's it will copy when it needs to, but that would have been the exact same work that you had to do. Yes,.

C

So an alternative proposal could be, for instance, try read contiguous where it only it only returns. True, if it's able to give you a span from the underlying sequence reader, otherwise it returns false, and at that point you either now have an option of telling it fine allocate and give the data to me or let me pass in a buffer in the Newton Philip. If.

E

We if we went all the way and kind of optimized the API as much as we can I think this is much better even than just every time asking the user to stock allocate and parsing and then, as you said, most of the time, the you know the stock allocation or the is everything the space on the stock is completely yeah and you and you can have you- can have a greater.

A

Calculator for that as well, it is a little because you do have two conditions that you're gonna return. False one is I, don't have that much data and the other is I. Don't have that much data in this span yeah, but presumably you can ask it how much total day is remaining, so you can just ambigú ate them.

C

Well, the second overload should never allocate other right because it will give you back what remains if it's available. Otherwise it should return. False it'll give you back what no.

A

Is it read only sequence as a struct? No, no! No, but if I, if you have to try, read 75 yeah, presumably false means there's less than 75 total bytes, yes and true means here's the sequence over. Then yes, yeah yeah, so it's not it'll. Give you what's left no.

C

It dont it'll if there are 75 or more bytes and will return and appropriately sliced sequence allocation for you so yeah. So then that that try read is I.

O

Take that back, it's not a location.

C

For your, like.

C

The sequences of value, but it's a.

A

Yes, yes, you're right, it'll do work to figure out the sequence position, it'll call slice, it'll out the slice and then so it'll it'll wrap the original buffer without making a copy. Here we have about the rest yeah, so the.

C

A

One seems fine, the first one. Yes, he was.

C

Wrong yeah from the second one you could say: are you single segment? If so, Joseph give me this bad, but the call work to do, though I.

A

Mean if we added a tribe we contiguous that produced the span, presumably the implementation would do better than doing the sequence slice followed by asking, like it wouldn't start with yeah current segment or current span, and then ask the question.

B

And so one was there I think one of you suggested. The idea would be passed in the span right then, but what still means we would have to allocate the tree. It just means you have better than allocating.

A

Right well, I mean so if we did the version that took the number, you want the temporary if it needed to to move beyond boundaries, and then it told you here's what the answer is that you then you as the caller would pre-allocate, and then you would use the same one over and over again on subsequent calls. No, no.

E

And that's one, but that's not what anybody was saying. What Levi was saying: we have an API that only gives you the value if it has it in contiguous in buffer. If it doesn't, you have a fallback code BAM where you call this API a way where you cook into your own stack, allocated memory right there, but.

N

C

Point is an allocation ever.

N

Made right so we were talking about two different.

E

But why is that better? It's better, because the first one, the API that I give you a buffer if I have it otherwise I return false! It's that it's the most efficient API. You can have well.

B

Why is that better from let's say, taking a ref span? Yes, where the transom is tagging? Okay, that's bad, and if it's a contiguous buffer, it just override that with basically.

A

The because, in the event that you only wanted to do, they give me than one shot otherwise I'm going to do something completely different. You didn't have to stack, allocate just to pass the links, check or true.

B

What I mean that's pretty key number on the fact that, in your else, you don't always allocate somehow.

A

And then baby, but you're you may have the. If you can give me the one shot in a span. I want to do that else. I'm gonna call the sequence one and then walk over the spans myself.

A

But is that very likely you're using a very complicated type, with very complicated I want you never to allocate anything semantics so.

C

A

C

Another comment in this thread is there: is an existing try, be to method the houseless fan and its behavior? Is it allocates in their honor, yeah and I think that we were said about that at the time? Yeah. Sorry.

M

But is it what you want is that why we to that.

C

Sequence, reader doc. Try sorry to try read to a sequence. Reader doctor I agree to some.

A

C

Maximum is this: one is counted to be exact. It reads to a delimiter, it reads to, for instance, a comma, and then it returns everything up to, but not including mouth, okay or optional. You can include it if you have cerulean true, but it else a read only spam and that read only spam if it's from a discontinuous buffer, maybe a bite or a base point just news up in a way right.

E

On this one also I wonder if we made the parameter left whether there are implications for like content, because users would stack, allocate a span, a spy ref to this method. Now, when you get it back, it may be either the stack allocated span or some internal span in the buffer and I wonder whether our JIT is not super confused at that point was supposed.

Q

To be doing well, this stack unwind doesn't.

A

It'll, never look at the span to figure out how to do the stack on white because that okay, that's it saved the stack size at the beginning of the function, says.

E

Partner there, maybe it's worth just measuring yeah I, wonder maybe they're very enough. So you know by ref it's not as bad as I initially thought. Maybe it would solve the value, but I think I would measure, because this this API will attend. You want optimize as much as we can I mean.

A

What's funny about it, because this is a registrant, you can't stack out like you can't use the span snack out like and then pass that by ref. You have to use the unsafe created from the pointer because the compiler would be like, but but sequence reader could totally have saved your static spam.

A

It's like yes, it could have, but it's not going to what's confusing with you, you can't pass a stack I'll expand into as a writable or actually, as even as a read-only span up into or as a method or instance argument on a restaurant yeah. You have done. You have to go through complicated, like turned into a pointer first, okay. Well then, yes, that is even.

M

A

If we took the writable span and then how did here's got, the right answer is, but.

C

One of the things that's confusing me is the this issue, links to the issue for try, peek and then within that issue. There's a comment saying this is not an appropriate method for reading a large range of data, but the very particular scenario given here is potentially reading a large range of data. So I there's a mismatch there somewhere.

E

All right now totally agree. The first overload is needs to be changed, I mean there are several options we could go to either by Rev, for the assumption that Levi was saying two different ideas yeah, but the one as designed right now is really not the right. One, yeah I think try, read contiguous or.

A

Whatever we want to call the disambiguate Iran it for the span, one would be okay and then try read Eldin the sequence, because that's that's essentially that's slice unless it would be out of towns. Yeah well.

E

It's yeah slice in advance, also and I assume that most people would not use this API indirectly correct. They would use something like a utf-8 reader.

C

There's a proposal to have xx.

E

A reader work over a sequence for either a bite yeah okay proposal, but that's what I assume that you would pass it, and this is very low-level API that readers will use pops. Why select the code to use? It is so working well.

C

It yeah it gets a bit hectic there, because the utf-8 sequence, reader or whatever would need to start going into future buffers, because if you're parsing a double, for instance, it's like you can't just assume that oh I got 20 bytes. Let me stop parsing. The twenty first bite might be I didn't hear. So at that point, there's potentially a talking.

I

World yeah, but after 15, even Tanner's, math, more yeah computing semantics with hopefully more so. Yes, it does yeah by.

H

Consider only one 768 digits longest represented a double for it to correctly round-trip and summers to the exact results. Oh.

A

I'm, only thinking of time span where I already fractional seconds, because that's that we're we're below our kick count. I'm like I, know that I got some.

C

Approve the read-only sequence, elements and read-only span back back to the designer yeah and so.

B

Then we also rename it to try. We get exact, correct.

A

Like I think it's good to have the word: I mean we could make the parameter name. The exact count yeah, but but I do feel that well for.

H

A

If they I feel that my instinct, if I look at this, is try, read count as a maximum yeah.

E

But for this fan it makes sense exact for the sequence it doesn't know, but the point is that it never it never failed. Like we said the first overload, it could fail if it doesn't have enough data. The second one can not have enough data is false, because there's no other condition.

E

There is no other error condition. The first one has two error conditions. The second one only has one yeah I guess it's.

A

Just the does someone find out by reading the documentation or by intuitively understanding what the method does that try read. Six will never return. Five yeah I liked I read exact.

E

So there's the thing: I don't understand what even men make honest one. So you passed a you know six! Well, what do you think the method will do you said to the unit six, you all most other read methods.

C

Like string not read, read out to you sex yeah, it's.

D

A learning that data is okay,.

C

Yeah, like I, got burned by this in actually.

A

In sequence, we were a few weeks ago: yeah, you screamed, I read you say and well read up to a hundred and it says I read three: do we put exact.

E

In the metal name or in the parameter name.

A

Yeah I think it's better in the method name right.

C

Because you could also have a method called try read up to, for instance, and that would say: oh you passing thirty I'm going to return potentially less than thirty it'll return, three. If I can and then you would do stitching yourself. Alright,.

D

A

This really just because screen read is read out to, and this one is give this it like. The produced value will always have length equal to the county. You gave it it's different, it's.

Q

A

Most people think string that mean does a.

A

Lot of people write the code of as soon as read returns a number other than the size. You asked for that. Obviously that meant into file and then they get sad when they use SSL Street.

K

N

Another one yep.

H

So this is just thing read, but without advancing yeah.

B

Don't you already have like a gazillion ways to do this? This.

C

Is interesting because it's a people to skip okay I've, never really seen.

A

H

Emily, we're going to say.

A

No to the span one and.

A

Like try copy two seems fine.

C

The skips weird and I don't know if it advanced the reader or not so yeah. Why wouldn't you just slice the reader before calling a coffee too, like we don't have try coffee to skip for Spanish? We don't coffee too. We just expect you to slice it. First, maybe could the reader doesn't like.

C

You were saying hunter.

H

I was going to ask if there.

C

Was an advanced method on sequence reader, which mute mutates, the existing sequence reader, because it's immutable extra.

A

Yeah, but you can do try to copy too, with the tribe. Even if try peek doesn't get the skip, you try peek I've countenance lice and yeah, so it's a little less efficient. This bit is this skip applied.

E

Before you come here after who's, your bubbly before yeah.

C

Because otherwise, you would slice the destination event, yeah well,.

A

I would say this nation if you wanted to start writing five bytes in you would just slice your destination span by buying before passing it of the destinations, but.

E

No, no, no I, think one interpretation of the skip is it copies to my spam and then advanced is basically but because.

A

He grouped it with all these TriPeaks I think it's not advancing I think it's trying peek into a destination span, I.

B

Mean if you read the title of it, that's kind of what he's trying to do right. We do specified number of elements from any position.

H

Is that the span returned isn't too big? Do we ever need to allocate for this fan that we return goal.

C

Of this is similar to the try. Read that we have just talked about and span would allocate in.

H

Which case the skip avoids a to large allocation when you're meeting will.

C

Skip really tells you well, because you might need to go into a different chunk. Skip, will help you identify, which one I bet.

E

That there is inferred escape is that you peek before you did something with the data, but because you eat not read it didn't advance so now, you're doing another eat now, you're thinking everything.

A

And you basically want to advance the.

E

Previous thing, that you're in have yeah.

A

But you're still not.

E

Committed to reading it, you know exactly.

C

But you should just be able to slice the sequence we her yeah, like sequence like, if this would this would make sense and follow every other pattern we had in the framework and we're buffers. If we can sequence reader, don't slice skin, you don't copy to society, I think if you're eating very.

E

Small values: nice would be expensive on sequencing, good sequence.

C

Reader is, as far as I can tell a very large mutable restaurant. So yes, having sliced return of instance, would be expensive yeah. But if you yeah.

A

And then I guess it depends on how big your skip is cuz, because if you didn't the tribe, so we had tribe peak count which is again from possibly what they want is exact here and it gave you the sequence and then you wanted to skip something you can just now. You have the sequence instead of the sequence reader, and now you can slice that. But if, if, if you're doing this a lot, then you're returning a complicated thing to get a simple and but but again, I think we're gonna do the same.

A

Presumably we want this never to be allocating in the span version, so we're gonna pump that one back as figure out how to deal with the fact that you want this, only work on a contiguous thing and describe the rest of your behaviors and then I think the try peak with skip and count again, I like it I, can understand what it how it makes sense. This.

C

Was previously approved on September 10th and then we had approved it I don't know, because during the copy all of the labels got messed up and the history behind us I mean labels, call telling them what they should be good. There's a comment from September 10, basically saying me a princess. We.

E

Maybe it's already sick.

L

Buddy, you see.

C

They just look for a comment. The first comment dated September 10th edited by you in battle.

B

That was one rail wasn't here. So that's my dividend. Enormous.

C

Candidate there, but it looks like potentially I know that when we moved issues from koreff extra long time, the history behind what labels got assigned- Wendall W done during the moon. So that might be confusing us where I don't know why. It's marked ready for review.

B

Right, there's no history for labor changes not anywhere much I'm not go there.

H

B

A good lost in the transfer yeah so then I have no idea. I could.

M

Almost go watch them video on video. We could but would not be finding fishing right now.

B

Well, I, don't know how people feel about it. Should we just defer this, or should we make a decision now looks.

E

Like we're all ready for them, but we should we should talk to David, because I think the API would be better for what I understand it's you use if we avoided this implicit allocation, I.

E

Can imagine how the implicit allocation is reasonable if we think that he I won't be called by mmm-hmm. You know directly by lots of developers, but if you like, it's fine, it is so complicated. Then it's just a loan of an API for us to using, but we wouldn't.

C

Use it, though, because it has a potential for allocation, we would try to work around it service with passing exam exactly so.

B

It seems to me when it was approved we unapproved too, because discussion, because basically it wasn't proofed and then there was a lot of discussion about allocation, as we can imagine then I, apparently, which also do not remember at all when a tool overall, of course, X and asp net to figure out how many of the API is. We have that return spans that actually allocate the conclusion was virtually none of them. Do then Stephen said what about things like to array that are called by the implementation?

B

Then I said that's unfortunate, so I, don't think you made a decision or if you made one, we I think change your mind.

B

So I think this is still up in the air and it's good to see that we have the same concerns back there.

R

So, what's ruling go back and say, give me an API that doesn't allocate or makes it in such a way that it's predictable I mean.

A

I think that we wouldn't say the same for the trip' counting the span. We would say the same thing as we just did for try read, which is yeah come up with them. We don't want the allocating version of this come up with a name that makes it clear that it's not allocating and what you're supposed to do to deal with that, etc, yeah and then the try peek. When do we already have a try, peek and they're, just adding an overload.

C

So, instead struggling with the first sentence, you even.

A

C

I pick of an of the next element, not the.

A

Next sequence: relevance! So: okay, so we don't have a count and we don't have a skip and account yeah. Okay, so people you sit about they're out in the span. You see.

R

A

Are outing spans? Yes, we're saying this is pumped back to the author for the same as we did for try read yeah if we don't like that, they're sometimes allocating their kind of a performance trap.

A

Then the try copy two is.

A

Weird I don't know if you've really talked about that one. So.

A

Then we kill me okay, so we do have an existing try copy and what they want. But they're wanting to do now is add the skin, but they've violated the parameter order for overloads.

C

Right, but because it's not actually checked into that, we can change whatever we do. Oh I see the try copy I see any of you guys have done that the first one yeah yeah.

A

The one that just takes us span destination, which makes sense all right and I'm, looking to see if it's advancing or.

B

So for the Tri County so does not.

A

Advance the reader, okay yeah so for try copy I think we would put skip as the second parameter, because it's overloading an existing method well,.

B

Hold on: let's first talk about the try peak one you would just swap skip and count as opposed rend.

A

Because they don't align anymore, oh you want to talk about the thing that's doing sequences instead of it yeah.

B

A

So I mean looking at what they want here. My instinct would be that what they want is tri peak count out value, skip default to the zero, but maybe there's a reason they don't want to default here. I mean.

B

Yeah, my concern is just that no, the end is like it's changing positions between two calls right, but I can also see why you won't skip before you pass in the count, because it kind of makes more sense this way, but maybe we just don't have the second one. If you just say, if you don't want to skip just passing zero, because then it's very clear what you're doing without even defaulting it I mean similar to homie hair. Don't have overloads that, don't you know you that they don't specify the start index.

B

You have to say even if it's zero right.

A

They're definitely the two there because there's an int in the same position, but it changes meanings. Yeah.

H

A

That's a bad overload set so either skip would need to move to the end or a skip would need to move. It would need to be count skip, which I think is weird or it would need to be counted with skip defaulted. So it's at the end, yeah in.

B

Which point I would probably just to be the second overload and then just have only skip comm video because it seems easier.

I

So why can't we have the slice thing on sequence, eater, so sequence: reader is a mutable.

C

Rough struct, we could actually make it a regular struck because it doesn't have any fields that are rough. It has quite a few fields inside of it. The size of the struct is probably going to be on the order of three doesn't bite yourself, so Colleen slice and returning a new struct would involve a copy of 36, bytes ish, and then you would call try peek on that, like in my mind, that's not the end of the world, because struct copies are not all that expensive yeah.

C

You want to need 4 bytes, but you wouldn't call these methods as you run anyway, you wouldn't call a method that has skip. You would just call method that takes account directly and avoid the overhead of calculating the Skip entirely well,.

A

It depends on how how much work you've already done and I know that you're trying to know we didn't into it's 86 bytes, in from where the readers currently pointed. That's a scenario for that weird.

E

Complicated processing and you're not sure if you're done military, you know reading that the first line I know HTTP response basically reading. You know talking by token and they ask but.

C

The the entire point of having a struct is that they're coffee by value like if you, if you need a pointer back to the beginning of the header line, for instance, you have your sequence, reader, the points there and then, as you're progressing. You have a new sequence reader that just says here, advance advance advance and you just have to sequence readers on the staff. Eventually you just pop one of them off like that's. How, though you typed I, don't think.

E

We are speculating, we should I, remember Peggy how, despite misuse and I I know that it was like it's optimizing, a lot of code paths down to the mental. We should just measure design, pontus issue with allocation. Today they didn't, we don't come across it even.

A

Gets my plan right soon so apparently, now we're saying tribe peak is hard because they want to skip Reming here. Well,.

C

What they, what they really should take, and then, if you want to minimize the amount of work being done, is to take in a sequence position not an in scope, because, presumably you already know exactly where you have to skip to and having a sequence position avoid you having to go having to do that. Linear walked.

E

Over the next job.

C

E

I I mean this is used in so optimized called paths. The sequence position is good luck.

C

It has long that's just like that, doesn't bode well for the design of the type of large battle. Well,.

B

That might be true, but that I.

E

Mean it's not just long. It's a sequence position like it's! Actually, too long. It's a it's! An internet.

C

Option long in an object, yeah, okay, so long at all,.

H

At least on Linux, that's going to be passed and registered both I mean.

E

H

E

It's not just passing is then: how do you interpret those two values? You? Basically it's a the object is to skip to a linked list, node and then it's it's so much more expensive than just if I want to eat four bytes up that difference for the second for in this pan, because in.

E

The object, I think it's actually baby, I'm, not I. Now, don't remember it probably is not the box, it's just that account. Still you following a linked list and and in for try.

A

Copy to if we want to talk about the last thing on the list that we haven't get sent back to the author, because we have an existing try copy to you that writes the destination which is documented as not advancing, which is why it's logically routine big, the skip I think our normal rules would say that the skip needs to be the second parameter, but it's not the same type, so we're a little fungible. On that point, it's.

C

Not changing the semantics.

A

Of the thing in the in position, zero normally.

C

A

That submissions go again, though, as well. Well, we say sources are the first parameters destinations and the second parameters, and then other things are the other parameters. Yeah well.

C

Here are sources, those destinations as found.

A

We offset into this horses. Yes, so it's skip is sort of related to source because it's modifying a source instead of my I, don't actually like this yeah.

B

I don't mind it either, like I mean it different, overload. I think it's fine of parameters are not in the exact same position, because you also had take other different arguments, so it seems reasonable within the same envelope that she so I'll be ok with the semantics, though, the.

A

Try copy 2 with a skip I think at this point. We would actually punt it because it's gonna be if the answer for try, because it use the sequence position instead of a skip count, then we would want the same thing for try copy. So it's logically okay, but we don't want to add it now, as a tri peak is going to change significantly I.

C

Also want to know what it would look like for someone like s. We never guys.

C

Cuz, these a few guys that we just talked about I, don't see us using to come with in utf-8 sequence, we're there now.

B

I think they are mostly motivated, but what.

C

Yeah but I want to know I want to know if Castrol would also write their own extension methods on top of these, because even these wouldn't be high performance enough for 180.

B

Alright, then, let me just put this back to API needs work.

B

Alright, so this one.

D

Might have been unapproved.

E

Could you scroll to solder propose.

B

I mean that seems pretty straightforward. If you ask me yeah the.

E

Thing is we not gonna, do any validation on I mean I mean like it's.

B

Like this is the implementation right, if you just.

E

You guys because mediate.

B

So that seems, hey I agree. You don't want to specify the T twice right. Yeah seems the same reasonable.

H

Yeah, it's the same thing that we have I.

C

Know I wanted this one as well.

L

J

Somebody says like okay,.

H

So elsewhere we have said we don't want to constraint where T delegate, because it breaks existing generic code that are wanting to use things they. Basically, you can't call this unless it so, if you're, in an already generic context, you can't call this unless your own T is constrained. Generic delegate.

C

I'm not too worried about that, because if you find yourself in that situation, call the existing overloads.

H

Yeah, but it's so different from every other method in the framework, including some other new ones, where we've just said: we're not going to do it because of generic context.

H

The where T is delegate doesn't buy much.

B

So it seems, generally speaking, the conversation between Levi, Jeremy and Yann says like this is fine. It's it seems there was some concerns about ambiguity which Jeremy discussed with Yan Yan said: I, don't know what I was thinking doesn't seem to be any and yeah.

C

Yeah we were talking about like technical aspects of the implementation. If there could be hidden, offense I mean in the answer is no the.

B

Only thing I can see is, maybe we need some other overloads, but I mean it seems also not problematic, to add these two and then add another place, the depending on feedback right. So.

C

Per 10 or suggestion remove the delegate constraint. I would be fine with that.

A

That's why it looks like something that I wouldn't pretend and something I thought that I wrote because I wrote.

I

It but the edit.

A

History is lost, it.

H

Because you wish to.

A

H

Basically, where tea delegate is a new c-sharp constraint that allows you yeah it's new, it wasn't allowed nothing till. You know capitals.

S

H

Okay, it came in at the same time. You know it's.

C

Not a new constraint, it's a recent constraint, the new constraint.

A

A

The big D delegate told.

H

Us was a good job.

H

But in anyways it basically forces the constraint to be a delegate type. The issue is that, because of the way, generic constraints work, if you're already in a generic context, your own team must also be constrained to delegate for you to then call this method and nowhere else in the framework partially, because it's new uses the where T delegate constraint, because of that yeah.

S

I mean we've said we can't.

H

Change existing ones over it and it doesn't actually by that much yeah.

C

The the counter-argument for that is, if you find yourself in that situation, you could always just call the original overload and then cast hold on why it can't be change the existing ones, because the veered runtime fail.

H

Anyway, then, that seems fine, because.

I

H

Me, because everyone who has this code that compiles and works then has to go change their own constraints to also be 20. It's like using adding.

I

Const to a function in C after.

A

B

But I'm saying like it's like, you could Cooley eight years to the constraint it's so it's still able to comply. You know, but like I've been you they give you. If you wouldn't do it like you like, if you actually pass and he that is not a delegate the clearly you wouldn't have worked right. So yes, it's a source, breaking change but and you.

H

Say we have an unconstrained generic, where you have multiple paths depending on what the T is, one of which is to use a delegate logically,.

E

H

E

A real issue, my question would be how often or ever it will this way. There's.

H

No way to satisfy a compiled, a generic constraint via the compiler yeah.

C

Leave the only way around it would be to just call the existing overload, which is what people are already doing today. Yeah and.

A

Since this is a new method, it's not a breaking change. Adding the adding the constraint on existing things is a source breaking change, but but as a new method, we can okay, you want to use the interactions.

E

Do we believe this, like any significant amount of code where you know he is a delegate, and it's just now, I want to call this method from this other method. I mean that's, you know it's. Is it a theoretical issue? We should answer.

C

It would be nicer for developer audience if white terror suggested we just allow it to work in a wider variety of situations. Yeah.

E

C

Just drop it into your existing code, pads to.

E

Me the main advantage of having that constrained. It's document self documents, the API.

B

I agree, but also I would say if you are method in for create delegate yeah, you should I mean at that point. I get free office what they taught.

E

Me to do runtime checks.

B

E

Maybe it's a you know the title of the return type of the delegate, not the whole day.

E

Should attend so fun good times with you.

B

Can name it P delegate that might be a compromise right, I.

E

Don't I don't feel strongly that way, yeah I. Are you wonder whether it's not just a theoretical, usually yeah,.

N

I would also say: I have.

E

Never seen code that actually, where T is a delegate and it's a it's.

H

An actual issue for for existing code calling existing methods, it's a theoretical issue for this case and I, don't know of any code off my head. That would actually hit it, but it's something that we've brought up previously, so I thought I would bring it up here as well. Yeah.

C

Like it's one of the reasons that we didn't add constraints to api's like unsafe, don't read unsafe dawn right like normally. We would constrain them to be where T is unmanaged, but that would make it extremely difficult to use as api's and existing compounds, because your existing code paths aren't already annotated at your own call sites with where cia's on that- and you may have done.

A

So we do runtime checks.

C

A

T is in there if yeah.

C

A

T, equal, equal type of and.

C

E

A

E

Now I know: I can do this yeah but I think but I. Don't think that really is.

A

Elegance yeah, exactly because.

E

This one is like, frankly this this generic parameter. Here, it's not a legal permit. It's like you know to to make the syntax yeah yeah more convenient. It's not something that floats like the T is not lonely. It's.

A

An inline and I have trouble imagining like I can see it for a bunch they're doing runtime specialization of generics for primitive types like enums till they're back in the fields of my fun. But it feels weird that anyone would do it for delegate so.

E

I'm not like I.

A

Know what you added the constraint here and another thing we should consider is: we can.

E

Always remove this constraint if we can find people complaining, I mean.

B

The argument of that thing of tariffs, thinking, though, is that that basically means you will never ever use the constraint, because their argument can always be made. They give you an existing generative master context. You always, which basically means adding new constrains, is kind of business. Well,.

C

There might be an argument for leaving the actual constraint off, but having like an analyzer constraint or something like that, that says: hey, you didn't actually validate that he is a delegate before calling this API. It might final at one time, I.

B

Would I wouldn't know I would honestly say that maybe there is a fundamental problem of constraints, because you can't count them the way they gifted, because if you could cause them away, that becomes a limitation of that method, which is fine well.

E

Because, in a lot of cases, you don't want to cast them away here we don't care, because the only thing that we do is return from create delegates. Do it.

B

Only existing codes that are described by tag, if you know UT, is a delegate, but for some reason you didn't it didn't annotate yourself of that right for.

H

Example, you shipped a library that uses this under the covers yeah, where you have a tea in its own constraint as you shift it before the constraint exists. You now can never change to use the generic version right because doing so means that you have to change yourself to also be where to delegate which breaks anyone else on the street. But.

C

To be perfectly honest, so yeah, if you don't change anything whatsoever about your code, your code will continue to work because it'll call the existing of them. Yeah.

H

It's still just one of those kids who look like. If we drop the constraint, then it works just like every other create or marshal delegate API. We have oh yeah and you you can allow users to make their code more readable and maintainable.

H

Without necessarily hurting anything you're, just in the same world that every other method.

C

These, the the only other counter-argument I still have to that is classes like I know, on classes like unsafe, we left off the constraint in order to allow it to be used in a wide, a variety of code, but in cases like that, we were actually enabling brand new scenarios that never could have happened in the past. This the scenario clearly already works, so there's no need to drop the constraint in order to make existing code work. Yeah, that's the only counter-argument I can keep I can I can make against that I mean.

A

The only thing that the constraint is providing, which is but is positive value, but like the positive value from the constraint, is that while you're writing the code, you get the red squiggle telling you you pass the wrong type here. Instead of, and you finally hit the code at one time, it tells you this did not make any sense. Yeah.

N

A

All of the existing things that we have on a gnome from where we added generic stuff to it. We don't have constraints because they wouldn't any any existing things that are delegate types and all the things with a type of you find out, link so lemonade. Something barks.

E

And the cost of this I mean the theoretical scenario that I actually think is so theoretical I cannot imagine people having already existing code that it's like. What do you do with a code? You would call the existing.

C

Ideas and then cast.

H

Existing code generally doesn't use where T is delegate. It.

S

Doesn't use T.

H

Whose is that type? They already know what type of Heligan they want? It so hard coded type of know that they use they made themselves, have a T and they just use type of T yeah.

C

They passed type of T and as a parameter to create delegate and then on the way out. They have parens T, parens object and then caster they're.

H

Basically, implementing the friendly thing themselves: yeah.

C

Which is sure I I guess my vote would be keep the delegate for now. It's only like we're still in preview, one timeframe or what I heard and see what falls out over the next few months as we try to light up code around it. If.

S

It ends up being.

C

A problem within the libraries and that's really good evidence. We should properly train and.

B

So we keep the constraint and then we don't for did. Somebody complains well.

C

If we ourselves complain too, because we have a lot of code that would benefit from this. If.

O

C

Trouble scaffolding, our own code against us that tells us other people as well.

C

Yes, it would be good data proving Tara's.

A

Argument, our general meta point of if I think, is good and but can be undone later for different goodness yeah. That's that's better than picking the one.

H

That's not undoable yeah. It might be worth noting that we've also in other places explicitly called it t delegate, not T.

C

What other places for.

H

Example in Marshall mmm it I think part of that is because it existed before the constraint yeah, but it still makes it clear that this is not a meaty. It's a tea constraint to, in particular, good.

C

Five pointer for wo, you know I'm fine, with to delegate to delegate.

C

All right, because again we're going to be doing the check at run time anyway. So it's not like it's nice for costume, okay, I mean Kree.

S

C

Expensive, anyway, okay.

E

I

It's mine, I, don't know who it was before, but now it's mine.

C

Apparently we're all very hungry we should have had doughnuts here. Yes,.

C

M

B

Posted it it's approved, let's do one more at least one more, but I know you can do much more than one one and not very fast today. For some reason they are like debate and.

C

We are a smoke for one. I should just get a fog machine.

E

So it's when you want to cancel cancellation. This.

C

Is a throwing version of twice I cancelled.

B

Well, we have it, try said canceling that takes the cancellation Coker a.

A

Try pattern says you should have an ngon tribe or you should have a throwing version of your tribe yeah sure why this.

C

A

Less slower Jackson, which.

C

He's fine, okay, Steve already said it's reasonable. All.

H

Right, one exception is going to be thrown what.

C

Everybody's operations in droves, yeah yeah, it's basically an exception, saying you've already you've already transitioned the test to a final state.

C

Whatever is something that is yeah.

A

Yeah, whatever the relationship is between the zero arguments that cancel tri-set canceled, would be the same relationship between the one. Our darling sounds good, okay,.

C

D

Last night, what happened? I'm, not hungry and I, was hungry last night before went to bed. Maybe.

E

B

Called breakfast yeah you didn't have, did nobody have breakfast here? Hey I, don't know what you could, but.

S

He's only added 4 a.m. or something right.

H

C

Know just sleeps in a fridge or something, and we just.

S

C

So many eggs then.

E

I do what did you have Theory interval? This got the income. It's like.

B

E

D

That's a post one like snack.

H

B

Venue and Kevin, you should not have a.

I

B

Proposed API, okay, this one do we want to review. This is what Stephen, because I think Stephen button. Why you? So, let's just keep this one yeah.

H

B

Drop off at home, sad well, you would have said otherwise. Do you even go tanner? What are you thoughts in this one.

H

This is we're missing a couple methods. We should suppose them so.

B

Here to you and severe in support of whatever they yeah.

H

I wouldn't have marked it ready for we're missing them specifically for intend. You went and it's not the same as convert to and thirty-two or convert 1030 to work and work to you and thirty-two I.

B

Mean in principle I've, nothing against that. We just look at these things. I believe.

H

They straight up return and end, rather than allowing you to explicitly swore to a memory. Look you.

O

H

It allows an optimization when, for example, you know if you've got an out parameter and you don't want to encourage double copy for the return. So.

C

To be fair, the CH it actually does optimize this today, not always it I've seen at least in the code pad so I've analyzed, it's pretty good about optimizing it for int, but pretty bad about optimizing it for long yeah. Don't know why that is so.

B

We this ones, you have a double end and you n64, so nobody, no short, no, nothing.

H

No because there is no instructor, there's no instruction for explicitly storing, so the smallest one is in yeah in this case, all.

C

Right, no, no word the ste.

L

Wasn't even in that case, that seems.

B

Like a fine had to explain.

A

What's funny is when I searched for the name of the function they have here? The first hit is that this function is missing from one of this C standard or one of somebody's run times. Header set even like this is in the is a, but it wasn't to matter so.

H

Apparently, this is the thing everyone misses or we generated our is based off of rethinking the bad header. There was two parts there was me manually going through the the C intrinsics and also going through the architecture manual, and we got a number that the arc that the C intrinsic Smith that architecture manual exposes. This just happens to be one all.

B

Right he's another one: ten are you own, you know up again pretty.

H

Straightforward, basically, there's no function that allows you to multiply two Long's in return, a 128-bit result. This can be accelerated at the hardware level on basically every modern, 64-bit computer, and so the suggestion is just to expose it. There's equivalent functionality exposed and, for example, CBA. So.

B

This one would be basically, we would, we would use intrinsic, but itself is not an intrinsic because we can implement it everywhere. Right.

H

Unlike most humans X, this is considered baseline support across all hardware, and so would.

C

It make more sense to return a you long high you long low tuple as opposed to out and return.

H

That's not how we've done any of the other API so far, okay, and that's also not how, for example, these C++ API for their intrinsic sure.

S

Yeah I mean yeah, so.

H

So I think, ideally, if we ever add the in 128 type, we'd add a separate overload, which was so.

C

You're basing this off of like different, basically yeah, basically, okay, so.

B

What do we think about? The name is that is the next inner big.

A

Ball appears to be the name of the thing that we already have on math yeah.

R

It takes two instant produces already have that name.

H

R

The name is good. There's me, I was pretty and yes.

A

I was gonna, say I, don't like his name but I started by looking at.

C

The the math class uses abbreviations all over the place. I'm surprised we actually wrote out ceiling and floor as.

B

Opposed to so holy a bug that.

C

M

Bubbling that intention.

H

Well, partially, and because that's not how, for example, big mole, the existing board works, it just returns a long, but it doesn't out anything yeah, but also the intrinsic API is, for example, they return.

H

They were as high out.

C

Low I mean if this is going to be marked as an intrinsic, we can make the API surface whatever the most usable thing is, and just have. The jet do magic to make it as often as possible. Right is out low and return high. The most optimal thing. Yes,.

H

And there's actually specialized support in the hardware for just computing, the high bits, that's the more common operation, for example, if you're doing a floating-point manually under the covers, you would do a big mole and only take the upper bits so.

C

You can imagine some scenarios where you simply discard below and then we might even optimize that even further you see you never use. It was right. You.

E

Might just want a more high, but even if it was an alpha emitter, you know the high was yeah.

H

But that also makes the API less usable, maybe comments the most common scenario for big multiplication. When there's two scenarios either you want a full hundred twenty that result or you just want the upper base. There's not really a case where you just want the lower, because.

C

H

Convince no more yeah.

A

Well, yeah modulo negative numbers, no.

C

A

The negative would have been in the high half, not in the low. This one is, you know, getting the low half of a 128 bit multiplied number yeah, the sign-extension message'. For you long, it's fine! It's! The low half is just do what the what our star operator does. Okay,.

H

So why do we think temple is bad, because there's no or I think there would be more requirements to optimize it in the jail yeah.

C

And this follows different as well so yeah and then.

E

You cannot ignore it at all. Well, you can ignore, but you didn't know the j would have a hard time. Reading about that.

C

Boys are different from me. Discarding the out the j can watch that yeah we can. We can see that it gets written to a single register even in.

H

Was ever used in one case, you're working with two primitive types, one of which is address taken and you can easily check for lost keys in the other one. You have a struct where which has two fields and where you have to track and read raishin and whether it's not used you basically have to promote it to two fields instead / to locals. Instead of being two fields, no, you have to track it differently.

C

Overall, it's less work for the job, because I was less in direction. If.

A

We believe that there's a chance that in the next 10 years we will add 128-bit numeric type, then long long. So that would be the right thing and since we can't Oh, since we can't overload by return value yeah, then yeah I think that also the.

H

A

In their door, Christoph's alternate point is instead of out plus return. What if it was void out out and I, don't do them a strong opinion on them, but I think that I think that the people want 128-bit numeric type, so we should reserve them the better signature for that one is a reasonable transition by.

E

A

This, in some cases,.

E

You just want the high things there's actually specialized instructions for that on most platforms.

H

All right, one more so for this one for the compare intrinsics, we followed what the see api's did and on the SSE they only support eight different overloads, and so we just exposed those directly rather than taking an enum, whereas with the AVX version they extended the support to be I, think thirty, two different cases, and so we didn't explode. We just take an enum his input. This is asking that we expose the existing ones of convenience methods on AVX as well just to make the transition more convenient.

B

So I did not follow the condition from what position for you from.

H

Sse to 280x, oh.

C

And it looks like the overloads here: have you have different method groups for nano special versus known as unordered right, the other one, so I didn't know what we're talking about yeah. You would.

H

Yeah, it's basically just asking that we expose all the same ones that exist under sse2, also on AVX, for parody and in the remaining 24 cases. That avx2 also exposes those will only be accessible. So.

K

Hold on leggy AVX inherit some sse4 point to where.

H

K

H

Is vector 128 versus 256 and.

A

You were saying the vector 128 over words already exists: yes, okay, so this is just vector 256 in things that we already have vector. 128, I'll roll it across static, inherence effect again.

C

Are there other new methods being introduced as for proposal that do not have 128 bit over ones, I do.

H

Not believe so, let me open and double check. I do.

A

See that the bottom of the screen right now says the justification, for these is questionable. So what's the question? Oh.

C

The UH the enum that you pass in to those. Basically, if you, if you call the overloaded, takes the enum, the enum, the enum wallets verbose, is descriptive and that's probably a good thing, whereas these are helper accelerators. Just a passing the writing of value for you, I.

P

B

Mean that seems, uh I mean I, don't know, but I would say they are already not super usable, but so.

H

Basically, for the last for the lot, the names between the enum and the names for these methods here differ slightly because the AVX form, because it exposes an additional 24 possible capabilities. It expands the names a little bit to make them more clear in that context, compare ordered and compared unordered or ones where the name has no difference when you're looking at means.

B

If you, if you just want to get the bigger vectors, that's not the super straightforward transition, basically right right.

H

For that, one is obvious: where is the transition for the other ones? Maybe not obvious yeah.

B

I would say honestly if I mean the argument would hold more water, that it's questionable is if we were to add all the over those four things you already have innovators, but even then we already have the smaller vector sizes as overloads right, I I'd. Rather it yeah sighs, extend those overloads because I think that's easier to reason about them. Oh yeah, this one, you know to call a different method with veneno like that seems a bit more and.

A

B

Know what you know it's yeah see I, don't even care I. Just remember that! Wasn't! Oh.

A

Wow somehow I was expecting the enum was describing like I have a tea and I'm gonna tell you which tea it is via II know I'm in like so that it was a void, star collapse or I'm. A witch compare Mogens now.

B

That would be cool le as though I'm a void star collapsing when.

M

You could have named all.

B

C

So no it's no way I'm trying! No it's lunchtime because all of us were a critical mass of officers. Yes and.

E

It's also the donut Thursday did or anything. So we need the time we need to go first to lunch, then for donuts.

B

Let me first end this call here, because.

D

It's getting sad.

C

So radar is long, I think what people murder me right now.

H

C

Null or empty to be similar string as warranty I'm kidding, I'm, not hearing anything right now by the.

B

People on line.