.Net Foundation Design Reviews, 27 Aug 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Review: GitHub Quick Reviews

Description

00:00:00 - Needs Work: API Change Request: Expression support for ref and readonly ref types https://github.com/dotnet/corefx/issues/26772#issuecomment-525417923
00:05:44 - Needs Work: Add a ValueStringBuilder https://github.com/dotnet/corefx/issues/28379#issuecomment-525424732
00:24:08 - Approved: Add Path.RemoveRelativeSegments Api https://github.com/dotnet/corefx/issues/30701#issuecomment-525429074
00:34:56 - Needs Work: List(T).AsSpan() https://github.com/dotnet/corefx/issues/19814#issuecomment-525432693

A

No, we are live well, in principle, barely love in practice, what we also recorded so all right any preferences. Otherwise they would just go talk to whatever thing the top to top the top three. Actually, what we don't want to talk about.

B

Let's go with these guys.

A

Screamo does not help you, who Marcus is ready for review.

A

I suppose the bottle severed AP landing expressions. I would table that him because we just yesterday in the language meaning we had also talked about extending expression, trees and there seems to be a general.

A

Let's just a concern on how to do that in general in a safe way, because the concern is that the prime consumers of life expression, trees, a link providers and it's a bit unclear like whether they will do any work support whatever new trance we're supporting where the concern is we're breaking pretty much everybody who does think providers today if we were to generate you know also.

C

Just augmenting the types and the framework for something like this doesn't really matter when the language to get generate those expression, trees, I, agree.

A

And I think as a general like I, don't know, proposal I, guess how the language can generally bind to them by name or something where they basically get met from the you know. The syntax representation involves in but I think it's too early, so I think I would fund the whole thing for now and then just come back and created exactly I'd.

D

Remove the ready for review. That's somewhat lot waiting for the language, the ground.

B

Yeah I guess you could mark it as block and then we can just say.

B

B

B

B

A

This looks extremely familiar, like didn't, mean already add those things.

E

And often be given time, oh you.

A

Have to snow the end of some generic API experiments where we have parse.

E

And tryparse and that's a.

F

Really so we don't have to get videos.

G

And the runtime team added some optimizations, the existing ones, to avoid.

E

Oxygen via heavy tasks, so like house flag, for instance, is one of the things that NRS mentioned, but there's no generic version, which is still right, I, wonder.

G

I thought the reason we didn't add generic overloads before was because the explosion issue, which is why they often for the intrinsic approach, energy.

A

I mean this would be generic methods right, but I wouldn't be that bad.

G

Right but generic methods online on generic, oh.

D

Because there is a right.

A

So you're saying all the other than you mieta generic ones are all handled by the JIT, intrinsically no.

G

I was saying like, for example, has flag. They opted to make that intrinsic, rather than adding a generic.

E

Overload so if you look at the il there's still a box instruction in the il, but the job first away and says I know you're in.

A

Me hands flag, wouldn't add much anyway right because I mean these returns of all right. I. Think the problem of this kind of API is that you have to cost yourself right. You basically get back in there an instance of system got array which is, you know far from being great, and you may say, I have to call time off, like you really wanted to say the type of the enum only ones right, but this is what you would have to write today and yeah.

A

That's ugly now is that the biggest following in the world, probably not but I, knew that the second one reads much nicer and feels more intrinsic to what we would have to do. How.

H

Often, are people actually.

A

Yeah, that's a fair thing. um She check well, 45% of apps are doing and imported API to demonstrate now. I grant you build. This is I. Think I don't actually look at the data right now, but.

A

I think a good chunk of it is is, is not especially stuff like logging or something like that right, I think I, like just very simplistic, so I think the.

G

Idea behind it, it's fine provided that the JIT team, like y'all, for example, doesn't think that there's going to be horrible downsides to doing so. Yeah because I know he's raised concerns over generic methods on all generic types. Absolutely.

E

So it sounds like you would want. Instead of approving this, we would want someone there trying another thread saying yeah.

G

Like it like, if yon says, there's no problem with it and I think it's fine, but he says there is then we might want to address that concern. Okay,.

A

So they didn't scroll down further yeah like I, would say so. Actually the white.

A

Oh I'm, sorry, yeah and somebody else like this one I think is borderline insane I guess then it seems way to watch for what amounts to a simple convenience overload.

F

Yeah they're suggesting adding a lot of other api's the handle me boxing on this from them and they said it's not that doesn't even work.

B

At least one on this time, but you needed something else: Oh.

A

Yeah I think the little bit over the top.

E

That comment also relies on a.

E

G

Is fine basic when they edit it yeah if you'd be edited already brilliant yeah, we've got, you know a delegate, but they both have to be the capital type name rather than the keyword totally.

I

Okay, iö has been able to express the identity, yeah.

E

I'm very much Jim I'm, looking through the doc right now,.

E

Back in 2018.

G

You know the the reason it's not come up for any API proposals is because it's a horse-breaking tap.

G

A

I guess this should be actually constrained to being a mother.

A

I

I

Looking at bigger sets of data, 50% immunity, application.

I

Yes, Father just people do that we'll start out pretty valuable that I've designed to do.

H

Including the panties.

I

I know how to read bar charts, but.

H

B

What's the question the.

H

50% of the apps that including indices or is it.

A

Well, the way it works is you basically have the ABS in different floors and consider each folder effectively an AB. So it is in principle possible that you have one library that is used by 100 apps and then we would see I need usages of the API, but it's the same library like in the past. We try to separate that out, basically to measure how much cost is there for the ecosystem, if you will, but it gets more complicated right and so like I, basically just yeah, yeah, sure sure and I I have a it's.

A

Similar thing usually makes it easier to understand the data more.

E

Complicated but that number almost matches new to soft all Jason's prevalence in apps. It was.

J

A

That's very possible at the same time, they're like very often when you look into the data, it's much more. So if you have enough absolutely the API for telemetry last time, I checked the N was like 60,000 reports or something but like wanna see once you reach that level of penetration like if you have a decent mix of of what people actually do in the wild.

A

James James I was probably responsible for some of it, but probably not for all of it. Yeah no I'm.

G

Sure Carter, that's also like command line API, where they want to do validation of is the entry I parse, something matching an existing value and I want to display help. Let me show what all.

K

The possible entries are yeah, we tell you, you know, there's a value to us, the attributes you could mix it with the description attribute on it.

I

And I think that yeah.

A

All right, so then, let's see what he on says something we can maybe improve it next time or good Jack, depending on where that goes.

I

Without language key words, enforcing that it can be will be used correctly. This should never be public I'm. Just dirty now,.

L

I see this is basically its preamble, that is this direct.

E

The the one of the issues with this is because it's a restaurant that also does things like the rate pool management under the covers like, if you do anything like unintentionally, make a copy of it. You've now corrupted memory application.

B

But that operated correct so.

E

This is this is a value type right as a strike, so if you make an implicit copy of it, such as passing it to a helper method and not using math or even just saying, like VAR b equals a you've now made a copy of this, which means now, I fir cated the state of how this thinks.

E

The array pole is actually being managed, and once you end up with that anywhere in your application, you've now effectively corrupted the array pool for the entire application, which could manifest itself as random errors and completely unrelated comments right.

G

And if I remember correctly, Jared either had put up always planning to put up a proposal around this kind of thing um and then there's also the proposal around adding the I numerable of span. The full frame span, support.

K

H

I didn't part of the game. Is there another proposal and copying of there's something that I discussed was yeah like I know you and I talked about it before hello.

I

Harry pay and been working with farad and maybe, like their team figuring out. What we can do to make one and I should say, but we should not do this at all until that feature is back.

E

Yeah I'm a degree and saluted so one thing that could actually make this better is I know that the language team is considering modifications to the formattable string type that will actually use this under the covers safely, and the idea is that, if you're using string interpolation, you basically get the benefits of this, including a single allocation under the covers. Without you, the developer ever having to touch this title, I, don't know their current status on that I just know it was something they were investigating. So.

A

What does that doesn't help you Gentiles, when you try to construct a string right?

A

C

Levi, dear cousin, if we were to use my thoughts based on proposal that I'd made to the language team and their soul iterating on, but it would still require most likely a type like this to be public, even if it was buried in you know, system run time, compiler services or something does how prominent the type is impact. Your concerns around security and rape, well, I would.

I

Say if they're gonna do that it has to be a enticement they just split into every assembly. This is a lot.

C

Of code to blue without.

I

Language support for making sure that the structure doesn't ever get copied and I don't want this public at all. So.

A

I mean they're completely by the array pool, although my because you effectively rely on collaboration between the consumers right and if you copy and then one guy to another, why I have to their pets to it, then you have problems right. So if we would have this feature, baby basically say there is a struck that can only be passed by reference. Would the concern disappear or would they be still concerns at all this time? Well,.

I

A person could have a meaningful discussion, yeah.

A

I think that I'm just trying to make it says stuff like what the action item would be for dairy all right. If you would just say well before you get any sort of like by ref feature in, we should basically Auto reject all features that reliably doubles drugs right because they're, almost all of the same problem I mean we just removed one from from the Jason 80s.

I

Right so this is exactly the same taste right here.

E

Because if immutable struck by themselves like how they have their own problems with regard to data consistency, but when you start having mutable strokes, the.

H

Post is called the state.

E

Of the application, yeah like it's without language, you know guardrails, it's just too hard. There is, you know, and.

G

If we did have any language feature, allowing essentially ownership semantics of a struct there's a lot of other low-level door unlock that people have been asking for so, but everything you mean by raffle. You need something else: well, ownership, meaning you have a single owner of the type, and so everything has to pass by wrapped until.

H

You explicitly.

G

Transfer ownership or explicitly make a copy via a specialized API in there, it's essentially unique pointer and shared pointer in sequence, possible, I'm.

A

Just asking because I mean, like ownership ARCA, you know like like Russell, will never come by a register.

D

Metallurgic yeah so maybe like if.

E

We can't have that maybe we could at least have a concept of move all the types the.

A

Thing that they probably the fairest marker here would be below.

C

I mean there are all the alternatives. We, you don't need language support to, validate that you can do it in analyzers, and we could say we have a you know, an analyzer that results in errors. When you pass one of these values string builders by something other than by reference. Someone could of course, turn it off, but then someone could also violate compiler rules in other ways. So we could strike a middle ground was something like that.

E

A

Think to me that the action here would be, we need to have a set of emeralds that are on metaphor. Yes, I! Think if people turn it off, I'm, okay, with saying: okay, you deliberately off it into a world of pain at that point, I'm fine with that. But if it's basically yes to opt-in to get correct code, then we already know that only types of customers would do it. Another way, 95% left box right that so like without this being hidden by default. I, don't think this is viable.

E

But I think that is. That is probably an interesting compromise to consider, though, because well yes, language support can come in many forms and, as Steve just said, one of them is analyzers. I mean.

A

The one thing we have to consider is that, no matter what you do, whether you make it an analyzer and make it a seashell language feature there. It's just it's no runtime enforcement, but you can still have a customer that doesn't know about this. Let's say f-sharp doesn't get the feature and in time and then you still have two people who used to type. You know in security right, but I think there are yeah.

C

And someone can also call or a pooled shared return or a pooled on shared return like yeah.

A

But I think it's fair to say like if we handled c-sharp a handle like the water, 85 percent, 95 percent case and then Oh sure, if you, if you go off the rails, that bad things might happen. But.

M

F

F

A

A

I mean the question: there is also I, think Jared brothers, never earlier that if you just enforced by ref, he's still missing out on all the other places negatives for you make X rental copies, which is not necessarily sebulba's, know.

I

Like whatever like, this type, would need to be 100% prevented for making copies yeah.

A

This is why I think Jared like in the past has pushed back analyzer features for things where well, you have to understand the language semantics really in-depth, and it's not just oh yeah, I forced every method call to be passed by refu also have to do a lot of things like Oh, some property returns it the complainant makes a defensive copy. Like you, don't necessarily see this involves an API. Is the compiler doesn't expose that knowledge.

H

A

So it would be hard to do this. Man lies are correct there. Yes, that's types.

E

A

B

E

B

Get rid of generics.

E

And add accurate templates and then we'll rename all of our files, dot CPP.

H

Can't wait for the game. We have to explain this to everyone. Yeah.

E

But after this I walk out of the room and you all jumping.

E

Actually, kinda dish.

A

So you didn't see once we detected off-screen, but here it is.

F

F

J

You said rookies first I was good. You can't sleep.

H

H

As it just removes redundant.

E

Slight bass, don't you already have a normalize, API I? Don't.

B

Think the explosion on alliteration it happens when we don't you're right, okay, I thought.

G

We also determined normalizing paths is the wrong thing to do, because the API sigh they're going to succeed or fail based on a lot of scenarios and we removed most of that validation objects up the.

A

Thing will be said earlier was: if we really care about this iñárritu. Basically, no were there. You know to positive the same file yeah. We would have to effectively provide an API for that and then call an operating system ache. They will tell you whether both things, basically, it is also the same thing and then basically normal consumers would just rely on spending quality right.

A

I mean it gets a different.

H

Thing like you've, been doing a good path service shape.

H

A

I mean this email is still available from the point of view that, if you use pass to just render to the UI, it just looks better right, I mean nobody likes dot.

E

Nothin man I got. What does it actually point now? I like this would be especially important in things like I must build scenarios where our menu.

H

E

Where, when you concatenate a bunch of these, you end up with a bunch of like you go to the parent directory, you.

A

Know the project system, like other things that are in the perhaps I'll about creating tree nodes, for example, it makes a lot more sense.

E

But wondering why this movie must be a generic normalize.

E

L

Like doesn't get full path effective in you know it.

E

Does yeah so one of the things that get full path does is on Windows. It also expands short file names to full file names which.

B

Were like what ones toward funding every way he's.

G

Also calling out that it will normalize it relative to the current directory, which he.

F

A

Oh, he wants to walk, but in that case I'd rather had an overload to get poop of the text. The path you wanted to be related to or.

G

A way to say: don't, try, I, I, guess what he wants is a way to say normalize without getting the full path, because you may want to keep it relative to an unknown to occur me unknown directory, but still do the rest of the logic.

A

Oh, oh I see no matter what you tell plug. He won't to get it here, a lot of guys have it back as.

E

They they gave a scenario for why they use get full house. Then.

B

A

um I'm not sure I, like the names addresses, but I also don't have to be better at anything.

A

What wouldn't dare me comment on at all? He was the one mockingly ready for you. He suggested to try remove relative seconds mm-hmm. He wanted a path. Experinece 1:22, thanks girl. Again it could.

B

Yeah, that's the one common.

A

Man was referring to and then it's very boring Molly says what looks fine immaculee to you.

E

Would collapse be a better name I.

B

E

Oh, it's not actually doing normalization yeah.

G

It's, for example, I. Don't think we would want to swap backlashes for forward slashes, etc, because you never know if it's a network path or anything like that, I think really. What it is is we're collapsing unnecessary segments like dot dot, flash.

A

Yeah I think lies to me. Something could learn to get full power like you won't effectively something back. You can use for comparison time versus this one really. The scenario is I. Think collapse makes sense, because all you're saying is like I: don't expand, dot, thoughts, right, yeah and it's a question what they want to call the collapse: rx bond I! Guess you can go either way, but so what is the purpose.

G

Of using the API evolving, the segments, rather than just keeping whatever the user employ like.

A

I'm thinking, for example, like you know, take take one API review where that's right, if you, if you drop in 50 assemblies into a peer reviewer, I, try to effectively compute what the common pathless and use that as to display string for you know your left side, your right hand, side when I don't want to see God thoughts in there cuz. It doesn't have anybody right right.

G

But in that case by the time you're displaying the string, why couldn't you get full path from some working directory? Oh I could.

A

But then I have a full Eclipse I've ever taught it on one day. I wanted to be sure.

N

Also, as we say that misspelled also, we need to minimize the length of the birthdays. So if they have a lot of this pattern collapsing, it will help just shorten the.

A

Thinking of like things like, for example, let's say you have something like premiere or Visual, Studio right but effectively, you have the context of a project. Whenever you render a path he wanted to base its tag relative to the project file. No, you want to collapse, thoughts if that makes sense, but if they go outside, then you feels.

H

Like this might make sense of your concatenated, multiple relative paths about to resolve that.

G

Right, so it's basically only for display purposes.

G

One sink is if it's for display books well, I mean if you're using it functionally, there's no need to do anything. The path will just be used correctly. So it's really only you all of.

N

The concern as I mentioned in msbuild is the length soldier of the path I mean like. If you have like very long. One is that if you collapse it it'll fit with a limit would benefit them has been yeah. We got some immiscible videos before because of that I mean, like you know, beautiful, adding some stock, which is it exceeds the max parent. We just at that time.

L

People should yeah.

L

A

Keeping things relative is I, love, always right. You have to be very careful at me, KPI you're, calling because, depending what API you're calling there, my interpretive Alice is the current directory, which is not necessarily what you want. They can still have.

H

I'm necessarily a fully relative yeah, I think so really even yeah.

A

That's what gets full path does I get off at minus ending is actually will remove the dots as well yeah, but it will make it relative absolute, given what you pass into it.

A

But you would normally do and then it would preserve the dog's office in order to express the relationship right.

A

Yeah I, don't like it seems a little bit like about us getting in this dozen hours. It doesn't seem everything to be in the Navy actually seems.

H

A little niche, but also someone is implementing it themselves that implementation platform. We support, Eastham argument for putting it.

G

I, don't think we even have proper for every platform probably got some weird scenarios.

N

Yeah your question inclusion. You know it was kind of onion to many other cases like well.

E

If you're using it for just file system and make sense, but earlier and the thread they talked about using it for URLs and I'm like wait, a minute query strings in fragments.

E

E

So if we do make this API public like we would want to explicitly call out, it should never be used on me, or else.

M

M

There's also some times for the web browsers.

M

B

If I'm pissed, because if your leads work.

B

Fiducial relative- that is another word. No did it work.

F

We discussed this one of that lot recently for the week before.

F

Oh yeah, that's right would be useful.

F

G

I mean bid on the fungicide, but every minutes or maybe no we discuss more well, we discussed a similar one and then I said ban has proposed to other API. Is the ban also has another one? That's almost like this, which I think it's also marked ready for review.

E

Because yeah, because one of the things that we had talked about last week was maybe we actually have a now spam method, hanging off of list of T that allows you to pass in a delegate and within that delegate party on the span have fun yeah.

E

To return this man from olicity like now, you have problems.

G

Yeah, it was the the move, one that we discussed last.

C

So you have problems, even if you just get the span and then case a list on had or something like that, that might invalidate it. Yeah.

G

1981 for that one's a guy needs work, but it's basically, this API list of TS man.

M

G

1981 for Tim Feeny proposed API is like who lines different because I, don't think is my ass been finalized, yet is.

E

It silly question: would we ever under any circumstance, change the backing store behind list of tea from anything other than to your right? I? Don't think we can without it being a breakage, it would be a serialization break and change which we can always like paper over, but do we expose the actual back into your a and any other API I'm.

G

Not sure if we expose it but I think we document that it's a linear buffer of memory.

E

Okay, so effectively a tra, so something that we could actually project as a spanner inventory right.

G

I think, theoretically, we could, you know, expose it as unsafe memory if we ever want to support over two billion elements, but then you're getting into crazy land.

H

I don't like hanging has been. This went, save me by putting it in a separate class onto pacings easily something on it.

E

But they much what Steve mentioned on the phone lucky you get miss pan back, you add an element of the list: Nell ahem.

G

Right but that's why we had said we should expose the delegate instead, which returns the underlying spand, while incrementing the version count, so that you can make any modifications to it yeah so.

E

We could, for instance, block calls to list of t dot out while the delegates being invoked, but now we're potentially adding code to list of T dot ad, which is already a hyper optimized code, that the.

A

I think the others that work, though, like with you, if you have a delegating idea, and now you want to access other local variables on the outside, that may not work right.

G

I think the biggest concern would be. Is there some scenario where you have the span where you could using safe code, break some semantics where, for example, one API one thread pulls the span? It's got the view. Another API, possibly another thread, modifies the list, removing elements and then trims the capacity you now have a span to an array which is no longer tracked by the actual list. That.

E

Should be okay because list is not thread safe, so if you have as long as we don't do any nonsense with the array pool inside of list which we don't do, but it's okay for those views to get out of sync right.

G

But it's a slightly different scenario because you actually have access to the underlying data, which is now only tracked by that span that you have well.

A

To be fair, you can like what you just described, though sometime multiple thread, so you can just do injustice level the this manacled of the different method. That does what he just said this and let me exit this band again and any of that data. Yeah.

G

Yeah, so if you get the span and then on the same thread, all remove trim, your span now no longer puts points into the list. Yes enforced arbitrarily burkhart. Yes, you could call it.

E

G

E

String, it doesn't point, are discriminatory points to it points to just an array that has garbage data, and if that the GC will eventually.

A

In that sense, I mean I'm, honestly, okay of saying, and maybe the one in the left is the entity I, don't lecture the delicate approach works, though, because how would you capture the elders, though? Yes in a culture yeah? But if the thing is like, if you, if the out of stuff that he requires that expands, then what you can do that other expands for what purpose other.

E

Expense other buffers that you want to copy data from let's say then there should be a copy to nothing. Probably why.

A

They're complicated because the one on the Left already does everything you want.

E

B, if we introduce methods for the scenarios of people wants, rather than just giving them primitives, and we can at least make sure that those scenarios work correctly well.

G

We could also, rather than add a yet we could return a span to the underlying array, just be a a span and was.

H

G

That the count is incremented before you do that. Basically anytime, you call as span increment account and if you call as read-only span, you can just return. This underline leave only span. We.

H

Increment the version number, if you just sent a single value with a noticed so.

E

We never check the version number when we modify the list. We only check the version that we're believing right great.

G

Yeah, but basically you have to if you're modifying the list, then it's supposed to kill the enumeration, because the enumeration might no longer be valid, so we have to assume that the user is modifying the list. If we return a mutable span, you.

H

I also think it that will get some performance.

H

Here's something very specific yeah. You may not want yeah.

A

My concern goes that Goods right, so either do what Levi suggested, which is basically on the other priv said. Maybe this pattern is the alternative. We give people more specialized api's, which is an ever-growing that's right. Thank you that one that the person once was, for example, bought copies about moves right and either we do that or we just say: okay screw. You only give you one API, but then it that API better be the escape ad for everything else. Yeah, and that's why I'm on the fella? Do they delegate because there's performance implications?

A

This may not have access to the AutoStick frame and then it will come back to us and still want the other API. So we might as well just either give people this API and then be done with it and then say: okay, if things become more mainline, maybe just add more specialized operations. We will don't have to do funky stuff, like you know, adding using stew with glasses right, but.

H

Also think that the people using this type of a are really becoming more and more comfortable with moderately unsafe api's, so adding one more to the mix, don't modify the list while.

K

You're using this well, it's safe to modify it as long as we increment version.

G

Yeah yeah, so so I think it would be fine for us to expose what Ben has there as long as we also have as read-only spanning and as read-only memory and.

A

What should work this by zero because they could? Let's say you increment this thing like that, let's say I called s-pen linking of diversion, then I mean humor waiting and it's a generation. I still do stuff to the to this man or I. Guess you cannot grow it right. Yeah.

G

Yeah, using only growing and shrinking you, you can give it a span. The only action you can do that would be that would modify. It is removing elements yeah.

N

And can the list of t get collected before before the end of usage of the span that return? Yes again, the list listed.

E

A list of T can't do anything is listening to use just a wrapper around in the rain right after gets collected, the array still reachable so which he doesn't have a finalized reinforcements.

O

That's explain.

O

E

The language team ever going to add that say about your new future, where you can only do this inside of an unsafe block or something like that. I think would be too big with a breaking change and all the API as we would want.

G

To do it, who, in the framework we can't do so I mean we would do it here, for instance, right, you can only do it on new API moving forward, but this is also not.

E

On safe, well, it's it's unsafe from the perspective of it gives you access to something that the list of T API is never intended for you to have access to.

G

Unless we now say that your intended have access.

E

G

That point honestly.

E

We should just say list of T actually has an API that returns the underlying to your rightful stop, like we just say, list of T's, always back by DT right, well, see the problem.

G

With that is that's the T was gone, then it returns for a second well, no well I! Guess we could provided we increment. The version count and it's no different from returning spam. No.

E

We don't even increment the version count we say like if, if list of T, if list of T has a way to get at the underlying backing store- and we say like this is this- is something that we just trust: people who access it to do the right thing with, like literally just making a field or a property, or something like a race segment of T like there. You go, modify.

G

It I think there's a difference between trusting people who did the right thing and like if we were going to make it an unsafe API that gave the raw array without implementing the version count. I wouldn't want it on list itself. If we're exposing on list which arbitrary people are going to use and people, newcomers to C, sharp and dotnet are going to be like.

G

Oh look, I can just access the array they're going to start doing the wrong thing yeah, and so, if, if we're going to expose it on list of T directly I want it to be the correct, safe thing, and if someone really wants to get the backing without incrementing version count as we don't expand memory, Marshall spen.

E

The same there's no there's no difference between returning a memory of T and the right segment of team like there's. No real difference between well.

A

I would say boy go perfect, I would say like in our a segment that's harder to use than you spend no I'm talking memory. T at this point ignore spam yeah. We need to be memory of T or or Spanish identic. On the point of view that you can't accidentally go out of the brain.

E

Well same thing, with returning your waste right.

A

Depends on what, if he has you're calling, because almost nothing takes nor I said when you always have to be crazy. Partying in the underlying right.

G

I think if we were to do this right and we were to do it, then we would want a span which increments the version count and returns a span of the underlying array, as read only span which does not increment and just return to read, only span of it. And then we should consider the memory overloads for people who want to store store it in.

G

Tore past the discovery.

G

Their problem probably be some concern there around lifetimes with memory of teasing, but I think that one, you could probably trust people to do the right.

E

Thing if you document to it, if, as long as this is backed by a regular T array, then there are no real lifetime issues.

A

So what are the thoughts then yeah you and they yeah yeah.

G

P

E

Yeah yeah I I, don't like having these api's for the sake of adding them like I like having our facial areas. Well,.

A

The scenario is, for example, the the bug corporation. We talked about right, yeah,.

E

But we could add an API for that. Why.

A

But it's not just one API right, I, think I. Think the idea from at least for me, would be that this is your catch-all like whatever things we don't have specialized in API Explorer. Well, that's that you really need to squeeze out one more thing. Well, then, you go along this path. The.

G

Other thing that it gives access to is a list. You can pass it to any API which was previously taking a buffer without needing to copy. So everywhere where we have a byte array, you could have an existing list of fight and then pass it through without needing a copy so pass through the Riddler's phantom to directly yeah. So you now have the ability to use system collections, api's, essentially anywhere, where we were previously just taking it to you right.

E

That to me, that's not a very strong argument because they are intended to be separate data structures, so I'm.

C

Also, having trouble coming up with a scenario where I would use a list of T to build up a byte array that I then use in api's to take a memory of bike. Well,.

G

If you're using UI data structures, then you probably have a list, for example, and then you may want to pass that list to some other API, which was only taking a tear a web.

D

Wouldn't you guys this wouldn't help it a TIA right, yeah.

E

And there's there's no circumstance ever in which at any point in the future on this planet, we would ever want list of T to be backed by anything other than a contiguous I'm.

G

Pretty sure it's documented under TR 84 as contiguous memory list.

A

Of T yeah, a list of T is meant to be like the equivalent of the stringbuilder right said for me very hard to imagine something cheaper than an array plus account I mean a little rain.

C

Wouldn't accept stressful week list yeah.

A

Sure all the things that logically it's different from how you construct them right that guy's, you know like a it's treatable, that you don't necessarily want the same semantics as if you don't you don't constitute. As my livable of chart right I mean you could but like it's not this is this one. You pass it on, as the collection usually well.

E

It to to the point that was just settlement on like this would this would prohibit us from making less turkey a collection of a linked list of a raisin or the cover selection areas like it. We are permanently forbidding the ability to do that. If we have these eight I'm.

G

Think we're already forbidden based on everyone, depending on it being a one access time. It would still be over one access. One know: if it's a linked list, then you have to do additional chance to to determine the correct array that you're indexing into which could be duck. Is the amortized over.

H

I mean so you're thinking something like plus plus.

A

Assumable that had the same problem right Billy, like I, wasn't excited to char, so I didn't turn char based on an index and we turn into linked lists. um All right, so I mean climb to believe that yeah. If we believe that that's an optimization, you want to make them that's a bad API, but I think the sizes are different. Widest preamble, there probably has more than 8 elements on average versus this is.

N

A

N

A

N

Somebody I mean like subclass list of T and.

A

Give you any virtual, so you have no, you basic can only provide additional methods.

E

You kind messages, so we would never be able to do this on collection or to, for instance, it was a someone of people are intended to subclass.

A

I mean list of tea was designed to be a more performant irregular. The radius was very heavy because it was designed to be wrapped. It was designed to be overwritten, and so it was a very heavy thing and this one was basically yeah. If you want to construct some all the race, that's the type to use right. So in that scenario, like I, don't know like this, that I did anybody ever. You know think that something this thing would make sense for this stuff. Tea, like I'm, I, I, thought.

A

That's why we exposed link list of tea, feel.

H

Like you would have a lot of people, this was changed. Being the high-performance focused thing that it is right now. Oh yeah.

G

Yeah yeah I think at this point, 20 years after we made the type, we cannot change it ever off of being linear membrane, we're increasing the size, doubles the length and then copies and.

N

Nails that way you can achieve the same result. Any yeah. You basically.

A

Call get type get field and then I'm going to score array and then reflection you caught that guy over but yeah. We.

E

Break with one of the few hours that I actually come out right now, so what you name the private field and in this stuff you know I change the type of the kind of fuel. Oh, when you change it to invariant array of G doesn't.

C

That break serialization.

E

There's! No because you can control the serialization api's to make out correctly I. Think rapper.

C

Currently it doesn't, it doesn't implement I, serializable you'd have to add the interface now. Yes,.

A

Yeah we'd like to breaks it breaks up edition either.

J

But is not a breaking trail.

A

Yeah, so what are your thoughts? Music, yeah? You know awfully quiet, yeah, I,.

C

I understand some scenarios for why someone might want it at the same time, it makes me nervous. It's super easy to screw up it locks us in, although I don't think that's as big a concern.

C

I'd love the smoking gun scenario that says you know this is a critical path in asp net or in windows forms or wherever, and because we lack this a span as memory we're 20% slower. You know because to me it just at the moment it feels like micro, optimizations here and there and I don't know it's worth the risk for that.

G

C

Gets back to a figure who asked who said it, but having like the key use cases, not the, maybe not that we might want to do that like what is the key thing that this is enabling that you know it is, is worth it.

H

It might be worth creating a separate type that is in the buffer. That gives you this lower level access.

G

What what been commented, because Dan asked do you have a motivating scenario, was lists of T is a very convenient data structure for creating an array of own unknown size. However, to use it as an array, you need to copy the entire contents again with two array: copy, etc yeah.

G

So it's he's wanting it explicitly to just use a global array and then treat it as an array for some api's that only take arrays, which also makes it simpler to expose api's in the concept of today, you have to support both T array and I list and I collection. If you want to support the varying amounts- and this would allow you to much like we have with span- allow you to only expose span, make everything else. A simple, singular implementation that boards.

N

What it said before we name name it like unsafe or something like this. They will.

A

Understand it by moving into the system and an interrupt service in space.

E

Right, you would have to very easy, but I to get know like how one percentage of our developers honestly understand that the Marshall.

A

E

A

I think putting a memory marshal will be a mistake, so it's usually nothing.

E

A

Yeah put it in the unsafe: that's.

E

Not the right place for you.

A

I always think what kind of job services class collection Marshall is enough, with the speed bump like I'm, more concerned of api's that are on by the youth types, because it's like well what what? What the indication do we have that you shouldn't call this a teen I bet, if you have to that, given that when job services not used by default, you have to buddy using for it. Almost anything under staying is low level and I'd say to various degrees.

A

I think that's as much as there's the thing you have, because you have Marshall in there in the type name, but just if.

E

It's not an extension other than that or cannot becomes for me. No.

A

I would not make an extension method that would be bad. I would leave it as if edits effect here, which is basically a you, have to call the API directly. Let's just say, collection marker god, s back.

E

Other election types we would want it out in there. So that's an actually a nice segue over.

A

Yeah there's another proposal here which young.

A

Which yeah? Because to me the same thing like adding a type of two methods seems a bit strange, but this one seems like a thing we have also talked about in the past.

H

This returns, oh.

E

H

A referential value.

E

In yeah, this assumes that you have a dictionary where T value is a large value type well, I, think part of it is the cost of.

A

Double hashing right, the only hash month can read and write to this thing without ever having to hash again line by the key. Don't have to look up the bucket.

E

That implies that calling this first API would always add an entry to the dictionary that doesn't exist by the way. That's how C++ is map type of person yeah. It's.

B

A good one that does it just wouldn't work for correct.

B

E

Curren dictionary by definitions meant to be used by multiple threats: concurrently, coma fighting what-what-what. Does the rough p-value pointy without like the ref T value can be invalidated as soon as this method returns, because a dictionary could have had all of its values really.

E

H

If we create new nodes.

E

Well so, for example, you call you call value rough on some key and then another thread immediately deletes some key from the dictionary and then ready now kicks back in and returns a ref to value. But what's it pointing to oh.

P

Yeah, it's definitely it's a simple yeah.

E

The current dictionary was like quite clearly, should just be removed and conceded.

A

By Mary and Vanessa note that.

E

The index was said a block on dictionary. Yes,.

A

E

Like that, that's fine, because the indexer can be treated as an atomic operation, because the indexer doesn't have indirection like the Stars this. This is giving you a reference to something as an atomic operation, but the reference is no longer guarantee valid once that that's on the cooperation heads, which means that dereferencing, the value has to be included as part of that operation. So.

A

Let's say I mean the only thing I understand is like evaded could be larger than what you can atomic to read a ride right. No, it's it's.

E

Not just that the um the the reference itself is going to be pointing into an array somewhere, but the elements in that array are going to be continually being overwritten and shuffled by concurrent dictionary as other threads access attention.

A

Oh, here's a missing in the.

E

Storage location that you're returning my general particularly.

A

It could be changing there, it could be reassigned.

E

To another key, even before this method returns to well.

N

This can have an even not everyone attended the reference.

E

Somebody's will on the right leg, but for for other data types like list of T dictionary of T, whatever, like I, don't really care all that much because if you have another thread, touching the data structure, okay, yeah, it's okay to corrupt that data structure. If it's not thread safe and current dictionary is explicitly documented as being bright, safe and we're. Now returning a corrupted reference from an API that is theoretically be in common with Rett same type,.

N

What we, as we considering with this, is answer so unsafe.

E

If you well, this doesn't actually have to be unsafe. It's ungodly no.

A

I mean the first one wouldn't be an se, but it's just performance, optimization.

I

That you could say guys and then wire using perfect right.

A

I mean the second one, I've never heard somebody asking, but the first one I've seen plenty of people having talked about this in the past, so.

E

Trying to figure out why you would want the rough T value instead of the T value yourself, like are people really in the habit of storing monstrously large drops inside dictionaries.

A

Well, it's more the case where you positive only want to be read to it and then based on whatever it is. Maybe you want to write it again, which kind of what's the double access right, not.

M

Having to react.

E

Plus the two of us alone, the traversals generally a loved one once the hashes finish only have a bucket they do. But if the hash is well done, then the bucket lookup will be on average one element: it depends on the state of the chain.

A

I mean this is I mean again. This is the last mile API all right. It's like, if all fails like you, go down that path right. You don't write this thing in that code.

E

This place well, what's the scenario flow together, like is: is someone actually getting a perfect benefit out of it.

H

If you consider the other way, super poseable, some sort of artsy type of hash in it.

H

Like an iterator yeah.

G

I mean most of the C++ standard types return or have the ability to return a threat already with T ampersand.

H

Not directly opposed to it, but I think of any times where I would have waited.

N

But theevans a hashing can't change right. It's not for the queue like he's were supposed to be immutable. Well,.

E

If you decide the dictionary, you are free to everything. Yes,.

N

The buckets will be reassigned about.

E

E

Basically, if you, if you use the same I'm, sorry if you take the rules that C++ mandates- and you apply those same rules here, while you're holding RF to that T value, you are not allowed to make any changes whatsoever. Did it aside for maybe not dating you can't add you can delete I guess you could look at their stuff.

E

Actually, you can't even look up other stuff, because even a look-up on a dictionary from our implementation could end up rehashing Andry, bucketing yeah the account, especially to use a string as we have a concept of we have a concept of detecting weather or there are too many collisions and if so, will actually switch to a different hash algorithm if TQ string and that redistribution can happen even on one up but a practice. It doesn't.

H

But okay, yes, for maybe some specific performance benchmarks and use cases yeah.

A

Ask for scenarios say one thing: the other one I think the problem of all of his api's will be that Illinois been. He doesn't have to think too hard to come up with micro benchmark right. The question is like: do we buy the scenario right like do? We think that's compelling enough to have an API okay?

A

Well, maybe for.

P

Sure a process request because he wants.

A

To call the API, alright so I'm sure.

H

There's a scenario.

A

It's just a question of like this.

N

Is your big scenario yeah many people.

A

N

Because something is possible.

A

Doesn't mean we should expose it right, because you can. You can make this argument for any data structure for any abstraction all the time. But then the problem is eventually you run out of you know. Api is to bypass right and then the other problem is also like. How do you like it seems to me like? There is a there's, a line somewhere where you say yes, certainly, the abstractions are generally not violated and that's real reason about large-scale.

A

But if everybody all the time, like my part, is the key abstractions then well now, suddenly you can only trust anyone in mind off cause the list of T, but you know you possibly some method. Is it something trunking and suddenly your array, it's like completely messed up, I mean I'm, not sure this whole. So these API is, but there is a line somewhere. I think we still try to find where it is because you spend the whole to 1 to 2 and then 3 or to certain extent, speeding up the API.

A

It's right, but well.

G

I think I think a lot of these are previously. We had no way to avoid the copies and things like that. Now we have a way to avoid copies to pass data around, and things like that and people want to be able to do that and so they're suggesting places where any this is a core data type that's frequently used. Let's basically allow you to unlock all the perf. In any scenario you need given assumptions on the type that I don't think we're going to be able to buy later in future.

A

Yeah, the problem is that that is a by definition. An open-ended said right, at which point you kind of sign up to add. You know: 700 api's, on court types, all.

G

Right and then the question is, can we simplify those to a set of API is like a span or ref returns where.

G

You essentially expose that and then people can build their own api's off that without us needing to add hundreds of individual api's that are more specialized.

A

Well, by definition, Macaulay ever beat us of what star and then you can extract all of that, but that's not necessarily the best way to do with.

G

You, but it's like how many eight guys could be appointed, how many copy two additions that you know a copy to span. Could we have avoided it? I've just exposed and that's.

A

Honestly, I don't have a good answer, but the only case I know over other than with ApS understand you. If that's the only API is we have to add that might be better than the s-pen API right. If that's just the first of like twelve different sets of API, then eventually we have to draw a line in circa. Enough is enough.

A

No, you have to go down on the ends like that, like I, think that's, that's kind of where I'm at and I mean no dictionary, and all this domination has come up before critical paths, but by this slightly more than remove one, but I mean we are pushing the envelope right. So we are in the one point: five, you know: percent.

E

Scenario: okay: I am curious about your idea of having a surrogate which is like I've, already computed the hash of the same.

G

That's all for today.

A

G

Jim, can you generally just do that by instead of taking string of teeth take in or instead of having a string here, whatever have an inky and then hash the hash somewhere, because you can always just call X docket hash code and.

E

Yeah, but it's for certain key keys such as tree, that's faster to let dictionary do its own thing rather than they call it get house cover yourself.

E

This dictionary can actually special case a normal circuit in terms.

A

Thanks guys online- and we see you next week.