.Net Foundation Design Reviews, 25 Sep 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Reviews: GitHub Triage

Description

00:-41:-28 - Approved: Add 'split' support for ReadOnlySpan(char) similar to string https://github.com/dotnet/corefx/issues/26528#issuecomment-424438136
00:40:55 - Needs Work: BoundedConcurrentQueue(T) https://github.com/dotnet/corefx/issues/24365
00:46:33 - Approved: Capability APIs for runtime code generation https://github.com/dotnet/corefx/issues/29258#issuecomment-424442400
00:53:49 - Approved: Vector(T) should have a ctor that accepts ReadOnlySpan(T) https://github.com/dotnet/corefx/issues/30968#issuecomment-424448085
01:09:40 - Approved: Add IPEndPoint.Parse() & .TryParse() https://github.com/dotnet/corefx/issues/31291#issuecomment-424450665
01:17:49 - Approved: Make internal directory separator helpers public https://github.com/dotnet/corefx/issues/31570#issuecomment-424461314

A

All right we are back, and hopefully everything got connected correctly. So if you don't have audio, please let us know in the chat thing we have audio, don't these according to more audio meters, all right, so we will do another part of our open design issues on you, Tom Dayton, zoom out little bits. We actually had a chance of seeing more than two items at once any conferences for order, otherwise that would just sort them all the stimulus and then again just yeah most comments.

A

Let's start with the most controversial, firstly, make sure we don't get anything.

B

A

All right bond the concurrent view- let's finish this, I I've seen this too many times. I.

B

Mean that one walked on Steam see.

A

C

See you but I can psychically envision what you're looking at.

D

A

It seems like you, picked up with a neat set of skills, while you were on leave.

A

C

Right very useful I would.

A

Totally take that all right, so let me just scroll down to the last time. I commented on this because.

C

So you asked whether it could be combined into a single type and into the existing concurring queue and I shared. What that looked like, including a commit that would show the implementation and then I commented on various as part of that various API is currently on a concurrent you that, wouldn't when you constructed it kind of we the bound that would not work, and you asked why wouldn't some of them work I further explained.

A

So what would you prefer I was doing then I.

C

Personally, don't have a strong preference. I would like to expose the functionality I think it's valuable. There are already a few places in core FX, at least I think at least to where we're taking advantage of the code being there, and we just sort of include that the type as you as internal, also you can use it directly.

C

I don't have a strong preference with how it's exposed, whether it's part of the existing type or whether it's standalone a few people that have commented seem to prefer the standalone, because they think that whether certain API is function are not based on a new constructor parameter seems flawed I'm playing with that I, don't I, don't do the same way but I'm. Okay with that, so I think people agree. We want the functionality.

C

We just need to decide whether we want it as basically a new constructor and do a few methods and then some methods on the type of not working when that constructor is used or as the new type. So.

E

There's like what my books would not work so.

C

If you scroll up to my comment on July 6th I, don't know if you guys have I am, but it's this comment that I just posted I've, just pasted into the I, am how.

A

Do you have it, but I can click on it from what is it July 6 it's at 60%.

F

Of the way down the page, we see yeah, oh.

E

So that just bounded capacity would not be that against the owner. I think we're looking for a different time. I, don't really say.

F

No, this is not going to the bottom right.

E

E

So we already have triad correct correct, and what does it? When does the current Falls today and.

C

Never returns Falls today about.

A

The interface it's it's.

C

It's part of an interface where some other implementations of the interface will return to us, but this simple notation the interface, because concurrent queue is unbounded. It will always return true, a little NQ and return true. But with this new bound thing, if you reach the bound, obviously than triad won't have anywhere to play with return points, Oh.

G

So in queue would just either have to spin or throw.

C

Either spinning so spaces available or throw if we decide that we want that to be two semantics.

E

To array would throw in the valid operation exactly one.

C

So this is what I explained in my long comment below, but I'll summarize so the way this is implemented to be very efficient concurrent to is itself made up of it's basically a linked list of circular queues, the circular rates and when you end queue- and one of this in a second- is full. We then created a new segments, but if it's not what we just in queue to the current segment, so as long as there's room in the current segment in which is circular, we just keep using the current segment.

C

Now there are certain operations, however, that Oren just takes something out and puts up again. Some of them are look at something that's currently in the queue like peek, for example, or to array which needs to copy the data, and there are some significant race conditions, corruption, possibilities. If we were to allow looking at the data, while it was also also being removed, basically- and so.

H

What how we deal with that.

C

Is if you perform any operation on concurrent queue that causes us to need to look at the data but not remove it? We basically mark that segment as no longer being valid for any cues, because what could happen is, let's say, let's say vacuous. Let's say the second is full and then someone DQ. So there's. Let's say you know one space left and then someone says I want to look at the data in this segment, and so it.

C

Well, basically, if he, because it's circular, we could end up going around the end at the end of the segment and overwriting something while it's currently being looked at, in which case you end up with torn data, and so we deal with that by saying the moment, you look at something in the segment without removing it. We just prevent future and hugh's into that segment. We create a whole new segment and future all future and queues go into that new segment.

C

And then you know the current one is drained by DQ's and eventually it's garbage collected, and if you you know, then look at the stuff in the new segment while we create yet another segment and so on, and this all just works because concurring queue is already unbounded, and so it supports any number of segments. But if we go with the approach of just having this bounded capacity, that's based on the fact that you just construct concurrent queue with a single segment. You can never add additional sections.

E

How many items are in the queue or how many slots we have in the queue cuz if the bound means how many items we could still do it I understand why we cannot say you know it's. The the bounded capacity is for the slots, so.

C

The this is an implementation detail will be a choice right now. The implementation is incredibly efficient because the T answers to your two questions are both. Yes, one of the same the number of slots equals the number of, and therefore we don't have to do any additional interlocks, who don't have any additional checks? It's just is there space available if, yes, that I'm under the bound, if know that I'm at the mound?

C

If you want to start introducing the idea that the number of the balance could be different from the number of slots, then we need to track those separately, which any means that will be more cost for the bounded tracking, because we would need to do interlocked operations and the like, but I expect it would also mean there would be more cost even for the existing cases, because at the very least, we need to check whether we are bounded or not and then decide which path to go down. Do we do those two?

C

Each of that counting or not so.

A

I have another question. So that's okay! If we, if we were to go with a new type right, bound queue.

E

Or something, how would that be different, because then trying to accept this inefficiency that we just talked about basically I, think that a new type would say: well, it's bounded and it would pay a bit more cost for having a separate count for a number of items from you know, new.

G

Pieces to a desk Christopher is just true. It is.

C

True, however, if we made a certain choice, the other choice is which what I initially proposed is.

C

Basically, we just expose the segment type with a different name and how is a different one? You don't have to worry about all the operations that would look at the data, while they're still in the queue, because we don't expose those api's.

C

We simply don't expose to array copy to get a numerator anything that.

E

These are operations that, like people, expect to have on from the queue.

I

Couldn't you have the bounded queue and then have explicit like freeze date, operation which would allow you to do copying enumeration, but.

E

Then we want to have the checks in every single member. Am I frozen yeah.

I

But that's slightly cheaper than the cost of checking length versus count etcetera, so you mean, like you, would have a freeze and an unfreeze. Basically I mean that's what we do in like WPF for all of the data types where we have to have concurrent access across threads. But then you need to freeze it when it's actually being like ready.

I

C

So you're saying that you will basically switch it back and forth between a concurrent mode and.

I

A non concurrent, basically, and if so, if your operator, if you're doing anything that doesn't require enumeration, it can just be in concurrent mode. But if you have to enumerate for any reason or you want to use link on it etc, then you would have to freeze it to do that. But there's a change.

C

That state there's no reason you even need those methods. If you agreed that these methods were never used when it was being operated on concurrently, then you could certainly implement them. You wouldn't actually have to switch its mode. You just have to agree that don't know race conditions, because no one's raising.

A

Sure, but how will you do that I mean the basics? If you would not throw is that you're saying you just say, be careful I mean.

I

C

Saying that I'm saying that, if you have the methods, the methods would effectively be no offs because you couldn't guarantee. How is this you expose the methods, but in order to use the method safely, you couldn't be using the collection concurrently. Therefore, the methods themselves don't really have value.

C

The only thing they could possibly do would be to sort of set a flag and then, if you did not hit the race condition- and we happen to do and we check the flag some of the time, we would be able to tell you you were doing it wrong, but there are race conditions, so we wouldn't always be able to tell you right. I.

A

Think they can practice the what ends up happening with these methods. Is that you're setting in the debugger? So nothing is happening right now, I just want to inspect what the curse set is I. Didn't then, that's where two array of six super useful right or just enumerable and generous for you or they just want to inspect with the current state, is true of your mother's thread still running and they excluding you while you're saying in the debugger, you might have other I think.

E

We have to like, if these methods don't work, I, wouldn't even call it concurrent queue, because there's such an bounded concurrent queue because they, it would be such a strong expectation, but would create such a strong expectation that, like it's just like the concurrent view, accepting bounded so.

C

I think I think the other there's a sliding scale here. I think we have a choice to make one ballot choice would be close. This issue, don't do anything. We continue doing what we're doing internally and people can do whatever they want externally. Another would be well. We expose the type and we cared more about API surface area than we do about utmost and performance I, don't know what the trade-off would be. Let's say you know, 80% of the performance.

J

And all we care about.

C

And we do that, we expose a new type. We can do whatever we want to make, it thread safe and we don't the best part, but we get decent perch and we expose all these eight guys. So.

E

By the way, coincidentally or not, I was actually looking yesterday. Ml dotnet has a type called objects and they actually use concurrent queue as an implementation. Yeah and I actually noticed that they have a problem that it's unbounded. So, if you keep, you know putting returning objects to the queue close in there, it kind of grows indefinitely it. It will never be collected and whatnot and I was kind of looking at the code and I was thinking.

E

It would be nice because for pools you kind of wants to you know as long as I kept hundred of those, and if somebody you know if there are zero objects in the pool, I create new one. If there are hundred I kind of like garbage collect if somebody little neck is that every all the excess is to you object pool, so they could have some info what counter and make their own decisions for me. Well, you they kind of wanted to use.

E

They actually are using concurrent camp because it is actually I also measured sample, and it's actually quite fast, despite the fact that some people complain about calculate view. I kind of found it fast and the implementation is very being, but it doesn't have this bound, which would be super useful for implement. Not going was if you're.

K

Accessing it all in the same place like I, am, and you can just have a camp yeah, they.

A

Don't expose the confirm here by the rapid: is it intimate eight meter of object, pool so yeah I think this.

C

Is where I was getting back to the comment about you know, Chris Powell's, coming out, you can maintain a separate challenge from the number of slots but like that adds cost and that adds cost, whether it's in the implementation or out of the implementation. So if you you know, maybe maybe they don't need the utmost in purp and it wouldn't matter for them. But I just happened to put that into like big fat fundamentally into a type of you want to expose by there.

C

We could- and it's just a question of how you know do we want the.

K

Flexibility of the API versus the best possible for do you have a bluff freedom. The kind of course I mean I've, never seen into lot operation show up on your mind as it's obviously IDs. It's good place.

C

So right now, concurrent queue entails effectively: one interlock operation, one interlocked /, & q or DQ.

C

If you were to add another one, will it double the number and when you talk about the sort of this low-level, if you're really looking for throughput interlock, there's kind of the things that you start counting to look at sort of the rough cost of things put in perspective, a walk in dotnet is effectively to internet to interlock operations. I, wonder dammit. If our experience is also colored by the fact that we tend to look at x86 and x64 like duh Steve.

C

Do you happen to know what interlocked operations are more expensive, an arm Hardware, given their different memory models? Yeah, basically figure everything on arms a little and interlocking armor are significantly more expensive.

E

There in the real world also, it so depends on whether there's contention or not correctly, right.

C

Well, that's the other thing, that's one other one of the things that concurrent you provides so the. If there's more than one item in the concurrent queue, then NQ errs and DQ errs actually synchronize on different things, so the interlocked operations aren't on the same memory, location they're on two different locations. If you had your own count, if you were maintaining it, you, basically all in cures and all T cures would have to interlocked on that same location.

G

So you get more contention, but the interesting thing is when I implemented a pool and.

E

I want a bound, it doesn't have to be super precise. You know like for all the pools, I kind of want around hundred items in the guild, but I don't care whether it's 99 or 105. So.

C

That's another one of these one of these limitations, the other one of the reasons this is able to so he takes the back eye. But what I propose doing here is a very, very, very constrained. Api, based on limitations of the implementation, and one of those constraints, is that the sahte, the bounded capacity, is limited to being a power of two, which is what lets us by by having a power of two and have that be the number of slots based the size of the array.

C

It allows very efficient indexing into the array, because we don't have to do things like mod operations. We can, just you know, knock off the top bits and get the index into the in the array.

C

It's, and so you know if we want their trade-offs like if we're okay, with the capacity being a power of two and as Christoph says, just sort of being approximately in the vicinity of the bound that you would want, then we can. You know encode, that limitation into the implementation and get performance benefits if instead, we want to be able to support any bound and have that be exact.

C

That's going to make the invitation to be more expensive and we would probably end up doing something like having a single count that we interact in a lot to increment and decrement on. I. Would.

I

Think that, if we're going to expose this as a highly specialized perk, oriented concurrent queue, then we shouldn't have it under system collections concurrent, because people will expect most of the stuff under system collections to work which Ling etc, because all the other types do today, I would think that maybe a new namespace that specifically says like this is a purporting to type it to stop right. Full functionality might be a better place for this and potential other collections that would operate in the same manner. What.

F

About this, giving whistle coach specialized.

C

C

Resolution on this I think.

L

That what core FX once is the hyperlink type and that we probably don't want that in the public API and until we can decide if we want to make something mean or something generally usable. This is probably a not actionable item. Okay,.

E

But I would say that and tell me if I'm wrong. We all agree that kind of we don't want a public type. There is so butchered that it doesn't look like a concurrent queue. All right like it needs to have some like needs to be able. You need to be able to enumerate it and to array copy code to need to work through. You know, I, don't.

B

K

E

A public type called bounded concurrent few that doesn't have I would, as long as it doesn't define.

L

Those methods, I would I would be willing or I would think it would be reasonable chef I, don't think it would be willing to it that it would be reasonable to add a new type that has those methods on they just work. Oh.

J

L

Say you have the.

J

Metrics I mean.

L

It's races. We.

J

Could change the name of it if the queue name is? What's bothering you if I make it more constrained.

J

E

J

E

Know purposes like being you know, eternal, then it's fine. We can call it whatever you want, but, like I, think a public type that is called bounded. Concurrent view needs to look like a concurrent. You I would.

I

Think at least that that's true it's in the VCL, yes,.

M

But that's good, though, because it means people see the value and right now we're just haggling over naming.

L

Yeah so I think we need scenarios to know what thing we want to shed from whether we make this thing with more concurrent queue and make the public and then don't use it internally or if we make a thing that looks like the internal one and we rename it because what people wanted so.

E

Actually, what's up missing disease, this issue, like Stevie, opened the proposal. What was what was the reason.

C

Because we're currently using this in turn, this type is in cortex, codebase I separated this type out, so it could be used on its own as internal we're currently using it in one or two places as an object pool. Oh as.

E

Enough so by the way we actually started to work on kind of you know, maybe there are some types in a brunette that you should realize, and one of the things that we as I mentioned I started to look into it's the object pool, so the work items seem to be related and use of it. So.

A

Then maybe we should wait until we actually have an object, pool proposal and then they can look at this and say what's the supporting time to make that as another full, you wouldn't.

F

Care about a memory potentially no, not.

B

C

Not care about ordering less than you know. You're thinking maybe bring this tipe in at the same time that they use scenario from someone yeah.

A

Because this jeremy said read additional.

B

And never you have like.

A

Contention between go X or Y like you need to have some scenario to say like what would they do to my son I right I mean you. Don't we base is what you want to avoid. Is this worst of all worlds outcome where people they care, but proof don't use the time because it's so slow, maybe with a care about usability, also don't use the type, because it's actually not usable, and then you have the type of no customer, and that's we have done this a few times in the past.

A

That would be bad, so Nexus, let's do this, it's just Marcus's needs work and I will just be caught. The fact that we will come back to it once we have an objective proposal, so it doesn't show up everything it.

E

Almost I think any other scenario, Steve or basically object. Pooling is scenario for this addition.

C

There are other scenarios they're, just not as prominent. So if you were, you have two producer-consumer situation and you know you were handing off data from the producer to consumer and you were okay, throwing away Jada if the producer got too far ahead of the consumer. You could use something like this, but there are other policies in such a situation that are viable as well. You could, for example, throttle the producer or you might want to get rid of the newest instead of the oldest or overwrite or whatever.

C

So it ends up being one policy in a producer consumer scenario that may or may not be viable. So to that particular point signal are actually has a very similar internal type as one of their core algorithms. They might be able to remove that internal type and use something like this, which means that you know that the type would obviously get more testing like it would maybe be more specialized and so on, but.

E

Is it used for under cooling and sitting.

C

Or in a constantly.

E

Producing it's a consumer.

C

Producer scenario, what they do is because I know the scenario, because I actually wrote this code for them, but you have a producer writing data as quickly as I can to each consumer, and the consumer only cares about like the 200 ish most recent elements right like we, we have a concept of being able to specify a capacity like you said, but the capacity is just a suggestion and it could get fudged one way or the other if the internal type decides it's more efficient to do that, but it this sounds, like you know, an ideal use case for this time.

E

So, whatever consumers producer scenarios where you can lose today done and.

G

Not give in signal right, yeah.

C

Signaler has a concept of things like a stock, ticker scenario and stuff like that or I.

C

Only care about, like the the most recent 50 moves in the game that I'm playing stuff like that, like it is a mode the signal, art and runnin, where you only care about fresh data, not scale data, you can imagine artists and even just what we're doing right now, video chat right you if things get backed up it's better, to lose frames than to like buffer everything, yeah yeah.

C

If you always want to be up-to-date, that's the most important since.

L

We're designing scenarios, that's definitely not an API review. So that's probably yeah.

C

But it's a team. We should contact, though, when thinking more about one.

E

That's interesting, I was mainly asking because I was wondering doing it bounded, concurrent pure. We just have optical. If there are interesting scenarios where you can, you know you want a cue, you don't want object pool. Yet you are okay with losing data, then maybe we need kind of both types.

A

All right, well I'm of the list.

I

We push this one back, yeah.

I

C

Know, I still can't see the scream and you do when you switch to you. Is you just tell me which one you're looking at yeah.

A

So what I did is go to go to the issue, tab, clutter down to everything that this market is ready for review and just sort by oldest to newest, and then we just walked in top to bottom.

C

Okay, thanks for looking at the system, numeric directors, home yeah,.

A

And we just said we pushed this off last time because it was too more too involved and so I think same is true of the proposal for half I think that's another one. We should look into, but it's a bit more involved all right. Here's one from us on the read-only char, similar to string.

A

We only spent up charming.

E

So it's basically split. It is played it's equivalent to string that split or span of the charm and.

A

We have been done this in the past because we have to allocate right now.

N

I think this is yeah I. Think we, why not just what I'm gonna write I mean.

E

You're allocating anyway on a real person, I think. The reason is that if you return an array well right for which part, so it's an array of arrays, so it's being read as.

M

You're returning I read only list a real little yet charm. Why not just return an array of readable memory? Chara I mean.

C

You're you're allocating anyone a more general question here and that is, we've talked in the past and I still think we should use some sort of split operation. That's allocation free write, some sort of enumeration of the results, and that would be what you were using if you were using spans and cared about the allocation of memory. You're not going to use this API. If you care about allocation and perf, and if you don't care about allocation of park, you could do span to string dot, split I.

E

Completely do this is the API that emo is talking about in the middle. It's less convenient to use this trick, that split and not as performant as the one that we actually want. Mary.

K

It's more useful than I video analyst isn't like you can change the elements yeah.

C

So it's your suggestion, then Steve to have a custom, enumerator type them yeah I mean I, suggest we have I thought you. You suggested something like this for the utf-8 string as well, maybe I'm misremembering, but some sort of some sort of non ienumerator. But so you know move necks. Current Saints, the region yeah turns read-only spans like a beacon any allocation freedom. This is the. There were actually two interesting things about that.

C

The first is, you might have to define a different enumerator type for each one of these methods, or at least you might have to define different enumerators, because maybe the enumerators behave differently under the covers, and that's okay right, that's just work on our end, the other thing that becomes a bit weird and.

J

It was type 2, no experience.

C

I just expose ienumerable; no, you actually have to expose the type, because if you excited through ienumerable, then it becomes Fox since I, which need a different type. If there were differences in behavior for different overloads, why wouldn't? Why? Couldn't you just make the single enumerator you know switch based on a mode or sometimes.

J

C

Could the reason that I was at least avoiding that, in my own prototype is I, didn't want to have a lot of branching within the enumerator itself, but maybe that doesn't matter in the end. So.

E

Before we talk about implementation, there's actually I wanted to run an idea there. Actually, there is actually a different way. It was completely different way to design it. I was thinking about one choice. We have, we return the numerator and whether we have one or many and I mean that's. You know performance, maybe some performance optimizations, you could do another one is to use the the pattern that we use in other span. Id is where you have to pass in a buffer.

E

So basically it's a split method where you have to provide a span to put the segment's into it. Stand-Ups fans in this place, where it could be span of something it could even.

O

Sound like too full of ins yeah, you.

E

Know like we're, basically because the the issue of enumerator is that it doesn't have random access. So you kind of it's it's not as convenient as today. Split returns, an array and you have random access and you can kind of access the first element and the third one and then go back to the first one or second one so I think the most equivalent API is one where we don't return an array, but we take in a buffer and if you provide the buffer, that is too small. We just give you the first.

E

You know buffer that length. But how often do.

A

You need random access. I mean like pretty much all the time. That I need is things that, if I didn't care about correctness, sure I just index into the thing and just say you better give me something with these three elements, so I think for correcto. You need to do bounds, checking anywhere and at that point I would argue just to follow. You have your four variables that holds that four pieces you care about, and then you just go from there. So.

E

The mod I don't know most of the time when I was writing code that needed to where I needed to split something I need it like a third, and you know third and eighth icon.

E

It's so convenient to do it when you have a pattern and it's so inconvenient when you have an enumerator and especially that these enumerators would have to be structs. So you cannot like loot, use. Mink like there are many things that you know they would it's just well he's a bull, my god, that's the thing. Fio.

C

L

Will wonder why to perform? Explain it like yeah, because it couldn't get part if it's gonna open spend the only way they can do. That is, if it cheats and turns into a pointer in and so.

A

L

A

To offset like, in order to do this, would you suggest where you? Basically, if you have a try pattern where you don't know the number of items up but I want a third and an eighth.

E

One so I pass a buffer of eight, and you know the split splits up to the eight 500 now I, just access the third indirect one, really nice Ian screaming she.

A

Basically, just say so, but in this case you wouldn't ask your buffer, you basically the same if the buffer isn't that enough to bad. Just give me as much as you have, which is different from most of our API, will just tell you false when you have to be possible, but.

E

I mean we have those API is that they return. You know like they returned how many they filled in and whether they are done or not. This could be the same so like if I, if I care about the third and an eighth one I pass about matter of size, eight and if I don't know what I care about I passed some buffer and then the API tells me whether it had more. Would you do both or would you just use the most? So that's what I was thinking. I, don't have a very strong opinion.

E

It's just I was actually prototyped in the past, the enumerator solution, and this one it's not I, don't have yet. As I said, slam dunk opinion on which one but the enumerator was not suited. It was very inconvenient in the scenarios that I had.

L

The GC under cousin I was span, and we understand that the point is pointing to a managed object and that keeps it alive if it goes into memory. As a as a white star, for example, we lose the object tracking right so.

B

L

An array of like yeah, sorry, it's a strain so make it right. Are you the last reference to the array is not captured in a span? You call split. If we're returning a memory, you can.

L

Only works if you first makes a copy right, because it's the if we took a scanner.

E

You're right and turn it into a pointer, then.

L

Now the GC says the object is attractive right.

E

So, okay, but I would say as this there are some comments below here. I will send it back to us on and say: first like we don't want. The API in the middle is the conclusion. Yes,.

J

We don't want the.

E

We don't want an API that is not as performant as it gets, because we already have one of those, its string, that split and.

C

It I could imagine, though, that there is a scenario where, like I, have I just have a bunch of comma separated values that I just wanted for each arm alone, and that's fine.

L

That's a different proposal, which is the split iterator I'm using Java's word for this. Is that error that Nets? Because we reasons previously discussed it can't make it a real winner yeah. So we would just have the the thing that meets the numerator pattern that isn't that enumerable.

C

And then sure you could for you, but we're not we're not rejecting that us know below proposal, though it sounds like right.

L

We're saying this proposal does not is not approved. It needs too much structural work, the rest of the talk period, edits punt it back to the author for cooking.

E

Said there cool, like on paper what you just said sounds completely reasonable yeah there are scenarios where you just innovate, I have to say: I could not run into a case where that would be enough for you, like yeah,.

L

I just didn't write, I think I mean I, think that we just want the next word and though this gets weird because once semantics come in but like Alexa right, give me the next token. All it is is a token iterator. Just give me a token Chomp Chomp Chomp Chomp Chomp. The problem is the definition of a token changes. Once you pass like a less then or yes,.

E

Prototypes such API, because then you realize.

L

Oh shoot, I need a side if.

E

You were writing a language that has those contextual tokenization, so.

L

You know you're writing markdown for the first time and then then you can do it. Then that would just be a give me. The next word, but word becomes very split. Can't do word because going into an open square brackets are different than the square bracket, and yet okay.

E

L

E

Leaving from the issue, it seems, like awesome, just opened the issue, but the proposal is done by somebody else. Correct yeah,.

L

E

Joe but for environment.

M

Right, I personally think that allocate low allocation would be the only really real good reason for this.

L

Whether that be an iterator or build my buffer or you.

M

Know your size to bring.

L

A bug bring a buffer.

M

Read or two is like.

L

Might be some things I.

M

L

Hold your array of separators and you use the I believe we wrote something or index.

J

L

Then you split at that plus one and we wanna like this is an easy user extension. Well. Okay, it's not.

E

Super easy, for example, use vectorized operations, which you can probably use. You know if you implemented if.

L

You gave every single separator instead, another way of separating.

F

Know you can do.

E

A live in the rain by them you can much both ways depending which, which batter is longer the one that you splitting all the bat-ropes things that you want to spiff. You.

L

Know just tries about things and tells you which one works.

A

Yeah yeah Ben just mentioned it another way to do. It would be like a callback, but you would have a span action or something ready which you could do as well, and if you have a local function, you market static, so you don't allocate the closer that might be not entirely horrible either, depending on what state you need to action, but for the most part you would probably need some state somewhere. So I'm not sure you can practice all but I. Look in so.

E

It makes it so difficult to know whether my clothes, steady.

A

Keyword local function is setting in front of it.

L

If you do open, paren, close paren arrow, who open paren, close paren that doesn't allocate, but if you just write the word food that does, you know, see guys c-sharp language specifications.

A

All right so I think then this one I think we can say closed right because we definitely don't want those api's and then whatever a little puzzle there is, it would be new code. We.

E

Don't close when we don't like the API but I, don't like the scenario, cuz I think we like the scenario we should solve this problem. It's a very common scenario. Just don't like the API you're.

A

Saying the okay, pretty.

E

A

Then in marketers.

L

Events to close it and open it, there.

F

L

F

Me did you post on your comments. I did it I know.

A

You were waiting to get calls again and I.

F

Really hope they didn't.

A

Need my comment: no, it didn't alright. The next one expression support for ref and with only ref types, I think this one. We had a chat with the so yeah I talked with Jared about this, so we don't really have a really thoughtful story about how we will deal with expression, trees and the new language semantics. So he's basically saying this is not an API problem.

A

I really have to think how this whole thing composed meaningful way and they haven't touched it in a long time and so don't think they would do what they are suggesting here, because it wouldn't even work the way you want it to anyways. The only way you could make these ones work as if you do it. Basically, you create these expressions by hand, and you just use the culture and portion of it when we say compared to lambda at the very end.

A

Well, you best uses a glorified way to do Revlimid, which I believe is not. You know bad, but then again, like these things are also designed to be mapped from source, so you would have to design a little bit more. So.

E

Somebody Ari was working on the reflection for virus types. Where did we like? Is the teacher working? It was handed off to someone I can get you the name, because.

F

I think it's very related well.

L

B

L

Reflection is different than the reflection on what the by reflection should have been done months ago, yeah and check. In finish,.

B

And so now, I can't it off. It's just done. Okay,.

C

Or got killed, yeah.

E

C

Certainly didn't.

E

You kill me, isn't it super elated like if you don't have reflection for by riptides? This feature is useless yeah, you say not dehydrate the biotypes on the other side, right, I, just yeah, I, think.

L

That the PI reported ID was doing is done completely and because the only thing he was working on.

C

But I still see them as it's different, though, because from this you can still create delegates that take rent or rep read-only, as their parameters like this is separate from reflection. I see.

E

You saying you can you, can we can create types that just use right of types and signatures? They don't actually eat yeah, because.

C

I mean this is really just.

E

Coach I mean right: okay,.

E

Yeah I can I can have an expression that uses span in their signatures, but doesn't actually create or modify this done. Yes, it sounded. You know like.

L

You were saying that if we took this, it's essentially putting a patch on a sinking ship that, like somebody's ego, think about what yeah we need to do to get the new language semantics. In this whole thing,.

A

Jared, you basically listen said like we shouldn't. Do it at all period and I'm like well, we probably want to happen, is I should do it at all? It didn't seem to be believing that business doable in general well breaking compound because use the other problem on every scene. The time you had an expression tree node now the question is what happens for people like yeah right, because you need somebody else to actually accept those expressions.

A

Isn't do something meaningful with that right if the only target is very mature, we can you know this is just libraries feature we can do that, but every single person that actually uses expression trees today, like the other problem, is when the compiler turns your source code into an expression tree. They don't really have a spec for what tree you're getting back right. So people made this. You know assumptions are wrong.

A

If you do this thing, the compiler will have an object cast in there somewhere, so I, but now that I can cast this node to a cast node and then I access, the inner expression and then every time they change. These things called breaks so they're very afraid to change the expression trees at all, because then not every time, I did it like the world came to an end, and so I think that if we add new nodes, it's the same problem now you have. You know somebody has to that.

A

What's the visitor has to handle that note now and I'll deal with that, since I mean it.

E

Doesn't mean that whole feature of expressions is kind of yeah I mean I. Guess it just.

C

Cannot it's like adding a value to an enum right? Technically, it's legal but you're, probably gonna break every color cuz or every consumer cuz you're, not even known to do it well.

A

If they, if they're likely to do a switch, if they only pass it around, then it's fine but that's a form of expression trees. But you expect the consumer to walk over like hey, and that is the fundamental problem. I think on.net framework I think it's basically a non-starter because it's in place update you break the world potentially, but it well I. Guess in that case, not because you would.

E

Especially it's just when one producer of the tree.

A

Suddenly starts.

B

Producing assuming.

A

That the consumer.

E

Cannot consume and.

A

I think he basically just said your work with people like you have to make sure the retinue knows that there's a meaningful consumption story. The other thing we talked about was the whole idea that the compiler can lower things. Whether you can you can determine like how far do you want this thing to be? Do you want the actual syntax, or do you want the lower thing? Because you don't care about this right and they don't have a story for that either.

A

So, there's a bunch of stuff that needs to be thought out and I think this one is not in the shape to do that. Yes,.

L

It sounds like somebody needs to own, you know, system, weak expressions and they need to understand what the future that's supposed to be before we start taking any I kind.

E

Of don't like the exposition that you described, I see it as a bit of a cop-out. It's like we have a teacher. There are problems with it and we're not gonna work on it. We're gonna just not evolve. It seems to.

J

The pole I like what Jeremy.

E

Said, which is we need to have a down there that looks at this holistically? Not just like hey, let's just slap, this one explained no additional expression and to.

A

Be fair, it was overly like communicating his point, but he basically said: let's start with no and then let's find the snares that are broken and I think about how we would fix them right, because it's not a slam-dunk to just add new things to it, right which I think is it there and then.

L

The answer some sort of the code gen won't ever produce these things, but we had API. So you can make the trees manually and if that's, what the owner of the teacher decides right, then we can evaluate that. But right now we're in a middle state and let's keep let's keep getting rid of things. That.

E

L

E

A

J

E

J

I believe so engineer, don't.

E

You write like the main scenario for expression, trees. What is this yeah, what you say.

E

This year, the main tsunami, you know sorry.

A

For what part for.

E

Expressions well.

A

There's I think there's two primary sonars: one of them is the I just use it as a way to the referee MIT, so I have construct note by hand and I just call rectum it at the end and the other scenario is you write code in C, sharp and then rely on the compiler into an expression tree? Is that for link bike variable and other things right? So it's another.

E

And the last one is basically.

A

Like any of our memory F, some of the mvvm frameworks do the same thing in order to access things from the view model, yeah.

F

A

Yeah one-time feature has we have talked about this multiple times this.

E

Property returns, true, I think.

A

It's something that I had someone had a list of things. We need to go over and believe it's in this check, but I believe it is yes, one.

L

Time future the thing that doesn't do what anyone thinks it does, because it's really a compiler features much right to.

J

A

You guys are all there. Let me read it: it's not my only fault, just pointing it out, no I think so the name makes sense of it. It's just it's. It's basically aesthetic this capability of which features a runtime. Your target will definitely have right.

E

A

Comes to mind, and then maybe easy does not enough, because they're working on I'll interpretation. So if you do things above optimization, you might want to know whether your il will be interpreted whether they actually going to be executed. So, yes, you can ref omit. However, it's not an actual good running a train, so there's a yeah, so I think we need to have the API I. Just don't know where to put it I, don't think it should be here. I know that for a fact,.

I

J

Did you not add comment about this, because this isn't connected to the next version of the standard we're talking about having light-up feature for recommended, particularly yes,.

A

That's the thing we need to design actually, but.

L

Theoretically, as you know, said rep them, it could be functional, but not jaded or we could be in a so. We may be only like.

J

Sometimes I said that we don't have.

L

Recommend it because, like device card, maybe couldn't create new things out of them and then, if we're in device very, can we get even yes, we come okay. What.

A

Is the my skirt is that it's.

L

The windows, the.

A

L

Things be happy if you wrote your stuff and see and make things be a lot of question Bart if you wrote them in c-sharp. Basically,.

C

Every every executable page needs to be signed so that you can't run any any unauthorized turn on the machine. So.

J

But I think even with it has a JIT I, think it's a little too general I think we get tied to expression, tree compiler, other places or cogeneration wanted to determine that it's interpreted or not should be on next to the feature, as opposed to, because each feature could still be.

E

Independently done so, I was looking at the comment, and it says it lists the runtimes that this feature is intended for I'm, the only one that really is a product. It is a marine I think it would be super useful to understand like how would we use it? In summary,.

I

Because I, don't I think most people are wanting this, for, like can I dynamically generate code yeah. That's me asking API.

J

You want to get and they.

I

Don't care if it's interpreted or JIT it they just care that they can dynamically create code and then run it yeah. So.

L

Then that would be ref emit. Do you work and then rep emit the huge it as a separate question. Yeah.

O

E

I understand they also like in you know, current easily karate and dotnet native unlisted as well and I understand that they want to move. For example, karate that's a bit forward, but it would be like we've had this problem in the past, we added some ap is for stuff that hasn't yet shaped or was not only track to shape, and then, by the time we first added the ID I, when the platform book was about to shape and they discovered that there's a problem with the api like, for example.

E

This is not the right question that we want to ask. There's another question that in what answer, and now we have an IDI that nobody uses yeah. Okay,.

A

Instant legacy yeah.

A

Yes, I think this one just needs to design I think that's I think we need to get together at some point in just talk about a lot in the meeting and never need work. I believe that's probably well.

E

This one is even different because needs work. It's we understood what the API is and we don't like it. So, let's redo it, this one is like I, don't have enough background like well.

A

I mean that gets more feature.

E

Time it just happens with.

A

Us I think it needs work is just up here. The scenario is varied. Right, I, don't even know that this one, maybe you guys know.

E

A

I'm pretty confident it is.

C

I can I can envision environments where it's not supported, but I just don't know the Windows.

N

Store application, no I.

C

N

Cannot did, you know, is very by polish yeah. Okay,.

C

So store applications in Windows iOS everything else and we go yeah.

E

As we said so, I understand that I was just more like. Is this the question that you? What.

A

So I was okay. I think that goes back to the API right. I would I would say, there's a scenario where you need to detect somehow with their coach and is support and the questions. How would you do that? What API is that affected? What capability question do you want to ask and that's the part that needs work more I. Think the general scenario is the thing about it.

C

Portably, you need to be able to do something at the answer's. No, no, no is.

E

Different than so, it is different than a cogeneration. Jit is very implicit and I would want to understand why compiler would care no.

A

It's not for the compiler, like it's basically think of this way right. So we are about to address in mid to donor center at London : 41. Well, it will throw performance support on iOS, so you need to be able to say up front. Will this drawer can I call it a because you don't want it yeah, that's the engine G! It.

E

Could be doing interpretation so you do you use ref emit and it basically gives you a blob and then you call the blob and it does. You know, interpretation and doesn't do any exact.

A

Something I think this rocket am I. Just link is actually talking about these two things like there is the does it work period and then well it works, but is that what you want it because you use benefit for performance optimization? You probably don't want to interpreted version of that, because it's probably not fast enough like take regex write it like regex. The only reason why you compile to RL is for per like action trumpeting. The regex is probably faster than trumpeting the ILO that represents the regex, and so like it's the owner.

A

You know right and then you have to make the decision whether you want to do that. The questions. Do we have an API that answers that today? The answer is no you, don't you just call the API and pray right and I. Think that's why I'm saying I think the scenario is valid. It's just. Where does the API go and what kind of question does allow you to answer? I mean yeah I think has Judas not the right one. That's what I was.

E

Yeah, but this question is what people would last and with like even tears. Jade like.

E

Initially, well maybe I don't know what else.

I

Would you try said I think the actual question here is: can I dynamically generate executable code because it doesn't matter most people don't care if it's interpreted or if it's jaded or if they're literally creating a function pointer emitting raw assembly bytes and then calling into that.

F

I

Care can I generate code dynamically in an execute. Yes,.

F

That is probably also fair. All.

A

Right so we talked to today vector of teacher, never contacted except we'd. Only spent of te is that one of those things no I don't think there was one of what I know. There was not.

F

Levi's, you Father, you better know what this is about.

C

There was basically I was running into problems where I had a read-only stand. A bite, for instance, and I was trying to get a vector, a bite from it, and I kept falling back to memory Marshall and I'm like I. This can't be our answer to go to the lemon.

C

It's a pretty slice and just make this. Is there a reason why we added the vector of T from span, rather than just having it from we'd? Only spin.

C

It's presumably you would want both well I mean you said that vector currently has a vector constructor that takes the span of T. Oh.

J

C

Because I'm, because if I'm reading data, for instance like I, say that I want to vectorize a string, operation or vectorize a search or something like that, it would be great if I could slice and then and that data's coming in to me as a read-only, spammy I. Think.

L

The question is, why does it have a writable or why did it take write a post, I'm saying.

C

Why do we have the existing one to begin with? Why didn't we just die one that took a read-only span because.

E

If you want to modify okay because I think isn't, doesn't it does it.

B

D

E

It does it create copy or not. Yes, it does, because it stores in register.

C

Span a little bit, it may interest and oversight during the initial API review. I don't know, but.

E

I thought so you can reinterpret past using memory, Marshall, yes, and that it doesn't create a copy. It does so it. It depends on.

C

It depends on whether your vector, sorry, if you have a wrap vector, then it will. The register will point to memory which is no hockey. If you have a regular vector instead of a rep vector, then everything is going to end up in the x mm or Y and M register, which is a copy, is.

E

Okay, you can create so many vectors. You can create many vectors, they.

J

I

E

In registers they are not, the data is not in registers. No.

I

Their stored in stack space, exciting.

E

I

On the heap, wherever it is in it exactly.

E

So it's not always in registers. Therefore, if you interpret fast, you can basically use the same space. Instead, you can pool you can't reinterpret.

C

Cast of stead of light as a vector you.

E

Can most you can do.

C

Is reinterpret cast a spell of a ref of that is maybe a ref of vector ER, but you can't reinterpret cast of span itself.

I

As anything that you have to do, a load and the ditch should be responsible for, if it's already in register and lighting that and if it's not register, then then causing it to load efficiently into register.

C

hmm Because a reinterpret cast takes a TNA. You, as generic arguments right and span, can never be a generic argument. So what you would have to do is you would have to pull out the say breath my span sub 0, so that you have a reference to the byte pointed to by the spam. Then you run that through reinterpret cast and you're like. Oh, this rough bite is now a vector or something went out, but keep in mind it might be on the line so that you really don't want to do that.

C

Casting like hey I swear. We got everything.

L

Good everything that I see in this class is it's doing a read and a copy in the new memory. So he knew this stuff took in everything in vector like if the vector that takes just a T value. It says, like Oh, a star assigned star the place I'm going to write equal, like you can't.

I

Necessarily rely on the source code, there half of the implementations in the day, okay yeah, but.

L

As long if the source code is is a representative, then all of the numeric values are copied. The the vector doesn't just inherit the pointer to the memory of where it was going.

L

So the vector constructors are all copying into a new structure. If you were to take the pointer of that vector and start overwriting, its memory you're.

I

Not overriding the original value, the general assumption of vectorized algorithms is that your span or Ray is a pointer to memory, and when you get your vector, you are doing a load from memory into a register yeah. So that's the general assumption. So this constructor would be saying: read this from memory into a local register for immediate consumption, so.

L

I believe that the question that Steve asked is correct, and why did this to expand and said if we don't expand? The answer is someone forgot eight letters.

L

It's an obvious mistake: it treats, it is a read-only, expand. It copies the value out of it again according to E, according to the il it does. This equal, unsafe, read on a line as vector of that guys.

E

I swear, I was reinterpret casting spans two vectors and vice versa. I just need to I think it may be when we had reinterpret cast as an instance member on I.

C

Think he what you were turning a stand, a fight into a standard vector by possibly.

E

Yes still, as you say, as you observed, it means that it cannot be just the reinterpret cosmetic now we now have, because those are generic parameters which won't accept span as arguments, but I think we used to have an instance method on stram to do that. Yes, and then it worked, and it was just we interpret constant, was not copying and then maybe we just move this API here and kind of didn't change. The signature. Well, let's see where Levi you remember correct, like I, was I.

C

Mean we were all together on this code at some point, I was involved with the reinterpret cow stuff, but I wasn't involved with this. So.

E

I, don't know the history of Mississippi I think this is possibility that we basically had it such that it was not copying when we moved it to this constructor and then yes, it requires copy down, and then somebody forgot to change it from spam to eat. Understand why they.

C

Changed the code to copy it's possible that, because we do a lot of unsafe stuff internally taking a span was all that was really necessary for our own internal steering's being part, of course, elearning. Horn honks, so we may be just never ran into it before I mean.

G

B

G

C

Into it before, if I just open this, for you I.

E

Think so, what's the conclusion should we add it I think we need to add it. I mean now it's kind of too limiting we make okay, we need to can be obsoleted, almost sure that we make a copy on your cell. You check all right, so I mean we've done. Three percent make a copy, which we have to add this. Even if it's not hardware accelerated, we make a copy.

D

What did photographer I can't see that.

D

F

I think it just doesn't load in the register so which is effectively the same as copying.

A

You ready just do a load of the pointer that this pilot isn't I. Think that's all the good.

E

So, let's assume the thing, that's you, what do we do? Then? We have been looking certain.

A

No I mean like okay yeah, but movie yeah will be obsolete. The other one will be hided yeah. So the thing that would be a good idea, like you really don't want to spend one night building I mean.

L

It doesn't hurt. If you have a span, then you have a span. It.

I

Just avoids the implicit cast, we don't do Han.

E

What about the other members somebody proposed the monument it needs to happen.

C

The the other members that were proposed were on it, assign students.

F

Which does it gotta be a ring? Legging.

C

It back out interest will be like one one thing that I actually didn't write in here that I wonder if we need is a constructor overloaded also just takes readily stand a fight as a personal read, almost manatee just.

B

A specialized them yeah common enough that we want to avoid the type of waterhole yeah.

C

That way it it saves the developer from having to do their own reinterpret. Houston, don't want to edit right here, so you would so the the addition would be public. Dr. Reid only stand and fight in addition to readable sanity and then copy to stand by to try coffee to stand by, because nectar can only be used for the marrow types. So there's no worry about, like exposing private feel to a went down. I, don't know what will happen now when you have Becker MIT, but.

E

Is this bandit you pass or this pan that you pass in this needs to be right, science grade? Well it just it needs to be I. I would assume at least as big as size of echo.

E

What, if it's like, you know, let's say my vector- is of ain't and I pass use, read only span of seven bytes and.

C

If the total length of the span is less than 32 bytes assume unit, 296, bed factor I would assume that's what room? Who would one I've seen this? Well, if you pass it a scale s to 20 Oh, what I said.

E

C

Is polluted two mites and if it's too long I think it would just truncate I, think that's what vector does.

I

Yeah, so what's the minute is the bite or mode since the JIT will already alive the typo I.

L

Think this is normal span of bite to vector of team. Yes, it's standby. Do dr.

I

Genius basically.

L

We're not taking.

I

We're not exposing the real nice, you were just exposing a sway.

C

In a fight no so the the initial initial proposal was to work with sanity. I was wondering if we should, in addition, have overloads to take standby time having I'm having trouble good scenario.

E

It's basically Spanish bite in his like in this mental mental model. Is it's any Magnum? It's a memory yeah, it's little.

I

Release of memory yeah, so so you're saying expose these. In addition to tea. That way, you don't have to do TDOT at fight cat fight or whatever I'm.

C

Basically, trying to avoid people going to remember a marshal at all, but impossible, especially sometime you say, it's kind of a dangerous class is as bytes in memory motion.

C

We I don't see where else it could be, because we can't do it for any generic span. I thought we made it an extension.

C

Do we believe that it's made of slime eccentric, nothing, it's a static method on memory marshal there's no best, and we don't wanted this as well, because now it would allow you to take an arbitrary T and just start bidding of which we want to disallow. In the general case, yeah.

B

I just editing it yeah. It's.

C

Public to us alone,.

B

Extension method: okay,.

F

So then we believe divided a little table for them. I.

C

The thing is: I didn't test: what happens if you add these, and then you have vector a bite. What happens now when you have read-only span of T, where T is bite and renal, it's kind of like yeah. The thing I remember.

E

The discussion is, if we build little span of by its dangers, but to read on the span of bite, it seems to be very correct.

C

It's not chopped start like you, you now have access to private and internal peons and the developer the struct might not have intended you for you to have access there. So.

A

Modulo this the PI thing. We all agreement that these overload series Nova correct, giving me API shame. But then let's approve this I didn't know that you should look into the biting and then we can't decide what to do.

E

So the only thing so we ended this span of a load and we say it's a mistake: do we continue adding Spano? Oh, it's copy perseverance.

E

Redoubling area I started.

B

E

Going with their let's now have two overloads now they.

M

E

Taking we have the table is longer than the other. Why.

O

Well, because you're.

A

So I can make you I'm gonna do some conventional dessert. What.

J

Is it can make you relatively.

A

B

A

Right, the next one, Nick Craver, proposed that we had IP endpoint parson tryparse I can be looking to do dinner. Then we decide that the value of law now.

E

We decided seems reasonable.

E

We seem to have never discussed it. That's what Steve said. None Steve said seems reasonable and I.

E

A

I was just talking to Carol about that I.

A

F

C

There a well oh well stuff might happen for what is one for me. What you mean so IP address like it's. You know, aw, don't CD or the ipv6 format. What's the format for life, but.

A

We don't have any partners of these active ability. The first ones as well I, see I thought you guys adding the overloads for spirit, I.

L

Mean I assume that parse is just anti to string, and presumably the part of the to string for ipv6 have to put it in square brackets. The same way that URI does because all the way the port separator is ambiguous.

A

Yeah I mean pricing, a PLC seems sensible.

L

Yes, if it's in the two strings, as if it's ipv6 it puts the address in square by Z and.

C

Then : port at the very end yep, so.

L

Yeah, what that mean this is make a parse which can read too soon, okay, and if it turns out that there's 38 different rfcs that we also need to support later. Well, then those become other things and.

E

Then we have their way of parse exactly so one of the implications is maybe then we need to add. You know some Paris methods. You can specify, like the format's, that you are willing to accept, but.

L

Usually those are this are similar or similarly there's an overloaded to string of what format should have printed. This only has one format that it prints them.

E

Sure but people you added these other other C's.

L

Yeah, it would.

E

L

Give it weight at that point we would, we would say this is what two string does. This is what parts does and then here are the overloads for where it turns out everybody and their little brother has an opinion on how you should format. One of these things yeah and.

E

The only negative is the ones that take out parameters. If we add an additional parameter, it would be a bit strange, but what.

L

It's fine, it is not just them. You get cars and tryparse, which is the default format, and you parse the dotnet comment. When you get parsing tryparse of like passing in the formatting, they look like you know: RFC I'm, totally making up a number, so I'm, not gonna, say one probably so do we do.

E

We have parts methods that expand on IP address.

F

E

Because another thing that maybe would be worth looking is why just IP end point like do we have its IP address, has parse every dollar spent of charge? Well, we already print those things.

C

E

C

Yes, as well yeah.

L

Yes, IP address has it, but the address : portent does not and I, don't know what IP input does begin, the DNS name, but that's another one right. Presumably it says it's not ipv6, in which case it doesn't put its properties and that's fine, because the NS says you can't have a call in your name right, IP.

C

N points constructor doesn't even do dns resolution so environment, oh.

H

It takes IP address so.

C

You, okay, there's a separate in SN Point class that takes a string.

L

So do we need to add force on DNS in point.

B

There's no issue for that. So no.

L

That's a joke. If we have IP in point for tracking, where it's pre resolved in DNS in point for where it may not be resolved and we're fixing one, we should probably fix the other.

A

All right, so you objections them to the API. Otherwise improved is is, do we want to add DNS in point to it.

C

Do you know some point has I'm looking at the implementation of your screen? It has oh you're. The format I've never really seen before so.

L

What is the MSM funding? It's like IP endpoint, but for a sharing instead of an IP address. Oh, like.

A

You know foo.com Colonia, yeah, I,.

C

Think I think we should be open to it, but there's no need to combine them and I'm with the Levi's. The two string is a little funky. Okay, it's got this weird address family, slash thing: okay,.

L

We considered and said that's a networking issue.

A

But is the only two derivatives from endpoint.

K

C

We added in support unix. What is it means yeah.

A

Because if it is, then maybe we should make it a virtual on endpoint know.

N

Try well, these are static methods. Oh sorry, nevermind yeah,.

L

Soon, as you admit, static, virtual and figure out how not to break people.

L

So Steve does UNIX domain socket need a parse right now, or does it also need thinking I, don't.

C

Think so it's just a file path. I.

C

Think I think we should be okay with just doing IP n by right now: okay, yeah, it's worth having talked about and then we'll be ready anyway, no position in the future.

A

All right make interim director separator help us public.

A

Three one: five: seven zero make internal directory separator help us public. So.

M

This is a P ends in and trim ABI, so we use a lot as Emma spilt been going through. There Collier realized they need to use it.

L

M

The time to trim that.

L

If we take span, we also need to take memory.

M

Be that was brought up the question I had down in. There was slate, we don't have any of the other api's taking memory that modify the paths and it seems weird to add one, and it would be something we should do like coalesced eclis.

L

It turns off the end and you have a memory you pass it in. If the link to come back comes back less, then you slice yourself from zero to that plane. Yeah. If it's trimming something off the beginning, you know how to do it. It's only when you're, when you're trimming off both ends, you don't know what it so.

C

I can work around. We tell.

L

C

Right now exactly just trim. One then trim the other. That way you can do all the math fix it yourself, yeah, but inspect I.

L

Feel that there was something that we did where we said to start adding memory overloads, because because people couldn't.

C

Tell what modifications we had done? Well, there is a there was a an accepted proposal for like read-only memory, internal trim as an extension method. So presumably you could use that you could just call that directly. Assuming you know what the top separator character is and then you don't need any of these helping others I.

C

I

Jeremy actually did what you lens down.

E

But it's a bit concerning that like we would be taking, you know like we, don't need span of try and read only memories giant string of they all they all kind of have this redundant span of charge as a tough die. Yet we keep having all the loads for well.

L

It's like it's, it's just a question of when we're, if we're trimming off on both ends so like we have the trim, whitespace thing which moves removes whitespace off both ends of a string. If you had a memory and we give you back a span because that's the yeah understand that you don't know where that fat was and.

E

I'm good I understood bad. We never discuss it. I'm. Just saying it's kind of unfortunate that we would. We are in the world where we have 23 overloads, but since this has a.

L

Implicit direction, you don't need it, the user can just do it themselves. If they had the memory, they don't have to obviously have slices and songs, but they need to find.

E

Indices where they want these dice, but you.

L

Would say my memory thing equal my memory thing that slice 0 to call this method. That way, but again, like the.

C

Trim method exists already now, as.

M

Of very recently, but is there are two characters.

M

Two characters and you might need to trim yeah there's the alternate directory separator in the curtain, though, will only just double-check. It.

M

Yeah we check both and trim, doesn't take I, don't know about that which one it is call them particular to happen, because I I'm wondering that this is just built. It's.

B

All it's awkward to build it.

M

Up to like yeah, we do maybe allocating, in that case, write an.

L

Extension method, I'd read-only span of char that can trim based on a read-only span of chart, which is a collection of characters, dependent.

M

So, which? Which would.

L

J

That we, that we'd be adding and wanting.

M

To add an API, the path that gives you back of the lobby array and it's weird actually to there's another real thing with this. Is that like it's suboptimal, because on UNIX, there is only one.

C

Specific scenario: no.

M

L

Is pretty common with just doing I? Guess the problem, the Aleutians in order to do it with them least allocations. We would hold the static array and use that array, but we can't give you the array, because then you get right to it. We could.

C

Give you your rails.

L

Or read on them stone, but you know you're now.

J

Going to understand.

L

And now your basic methods get angry at you, because fans leaped across flow and yeah. So having the method, encapsulate I'm talking about this array and you never get right access to it is and there's.

M

In these cases to there's like no array again in UNIX I mean we could build the array and have it have just one member on units sure sure, but the code is just a little bit more optimized to their kids for yeah. So.

E

I, don't know so, would it be acceptable if this API just gave you and how much you need to trim? Basically, it takes. We only have overloads the pics man when the returns you a bull, the other one. We can see whatever you, you call the API, you get any int and then, depending on what your data is stored in you slice it using the same. Why would we keep the.

H

Bull one, then we would just say the infirmary times that I get it.

L

Number, if it doesn't end in a thing or.

E

It returns the link, no, no, the pool I meant ends in directory separator. It always returns a bool I. Don't understand why you have to open loads here because it doesn't actually, we.

L

Have to because of.

E

L

Here we're not allowed to depend on implicit test; we don't then, okay, we.

M

We actually started with the other path api's with no stress. Oh.

O

You were some chemical at that time. Yeah.

M

O

Okay, fine, yes,.

L

So, yes, we make the string it. It calls internally to read, only span one even it just it makes it easier for F, sharp and VB and cobalt yeah.

E

Okay, so just for the versatile methods we could return in int and it kind of solves all the problems except it's not this convenient yeah, but.

L

We did the tricky word can return an end. Then people can write the absentia methods themselves, take the negative number and turn that in the pulse or you know they just write. Yet that's no wait. What would negative number mean when do we return it? So if it didn't internet separator, we.

E

Would rather return input that link or we would her negative one.

L

Yeah we should we.

E

Should return an int that they can just plug it in to slice, and it gives them what they expect right, which.

L

Is zero negative number, but I feel that the only reason anyone would ask end is: if ends in then call trend, and it's like you know you did this wrong. This is like asking contains key, followed by dipping into the dictionary like how about we. We make it easier for you and you just have to get this answer, and if you wanted to bully and say you can make your own less than you.

E

Can always do it? If we return the length, then you just check if it's the length I don't need to trim. Yes, if we return the trend length, then.

L

If it's equal to the input, then it exactly you don't need to do it so that.

E

You can do it, it's gonna just give you the same thing. So it's like this is one method start two methods, because we need to string in the interviewed only spent of I.

L

Don't know if it's I think it's a reasonable amount of efficiency versus like it expresses the intent if it solves the span versus memory problem and I, don't I I can't imagine a case where somebody's using the boolean to not then just called a trim. What is this, what is the span versus memory problem still following there? Yes, if I have a read-only memory of charge and I called the trim, it will get implicitly cast down to read only span of char. Give me a span, but that's useless to me, because I need the memory.

L

So now I have to go back and called slice myself based on the length of the span that was returned. So what I wanted to believe sure. Therefore, why don't we just sleep down you the link right, because if you have a span.

J

You were surely, if you're a server you.

L

Can call slice and you didn't waste the time copying. The extra register look at me at our partners for sure.

E

A

Twice as fast I, don't understand why the why the boolean that is not enough for.

C

You well boolean, probably is any Anderson to our furry separator yeah, because you might have multiple separators, so you're gonna have to call it until we.

M

Don't we don't really trim the one oh nevermind done because they're sort of meaningless unless you've normalized.

I

The courage of communication says: if ends in directory separator and is not a root path. Then return path, dot, slate slice, zero to path, dot, length, minus one yeah. Sorry I'm, like.

L

Paging this over again, because that's why.

K

The really it's not enough, because.

L

You need that extra question of. Does this ad colon backslash, in which case you don't remove it listen. So you know what the trim and tell me what the length is after the trend and instead of me, does it end in this, because then we need to change the definition of busy in business to say it could, but some that's important so.

A

Basically, because I would argue then than these API is I'm, not meaning to ignore this payments for a second, let's only talk about the spring line right, so we have true ending directory separator and we have ends in directory separator. So when I have C colon backslash one returns false iczer one returns true, but the other one still doesn't trim it. So like they're, not really correct.

M

So we can, since I call it it's just that's a very common bug too, by the way right. So we need when you treat it like. If you trim C colon backslash, like you turn it into a file right.

L

So so, I think that the primitive in this operation is trend, and so we should only have the trim one, and someone can now write the boolean correctly in terms of that, unless what they really wanted to know is whether they end in a back slash, slash well,.

M

They do because there's another the other case two. This thing is that you actually have to add them, because looking that way, if you're, comparing I actually just tweeted on this yesterday, like if you don't have a directory separator on both, you can't compare them successfully, like notably like, but they know what I do. This was the issue with zipping and so, but.

A

I would I get the primitive for that, so basically there's two primitives, then one of them is remove it. If it's there, the other one is additive. It's not there. So.

N

You would never need one fool, then well remove it, it's remove it if it's safe to remove it right and edit. If it's actually thinking.

E

That the primitive is index of ending directory separator once you have that all the other ones you can implement them by comparing lengths or calling slots.

M

And you'd have to call, is you know its path, but you get to pair the route length and some sort so I think it's like this also happens and like we, we put this bug in multiple times. I think.

L

The thing that would be needed here, given that we've seen a lot of bugs with that, is find the one method in terms of whatever you want and we can solve the overloads later. Finally, one method that will not than someone writing a bug and if we can't write it in one method, then maybe we shouldn't do it yeah there.

I

Are big people who are joining paths will just use Pathak, combiner paths that join, and so this one is really for. I want to logically compare two paths to see if they're equal, the the ends in.

M

I

But the trim one is well: if you trim both and you can you compare them and they're equal, then it's the same as saying if, if either ends in didn't trim to make them eat one well.

M

I

L

Treatment starts with is.

M

The other one great, so you can't trim so.

A

It's a bigger the thing where you needed to know what it is and there's.

M

A certain starts like if you're gonna say starts with like to see if something's nested, you have to have an ending separator.

M

Alright, it's a pretty complicated logic right well,.

A

Or if you have to format it right, if you have to come, if you have to construct a command an argument, for example, you need to end up with quotes. Well then, if an intimate ended, directory, separator I have to put a space in between otherwise I get. What was behavior on the command line. Then.

I

I

F

And the things out the tool windows where you can have as many as you want yeah.

M

Like true for the purposes of path, but it's always you're supposed the pattern you're supposed to be following- is to always normalize it. So you call get full path and.

N

That always takes away the extra separators so, but it's trying to resolve it so like I, would agree with Jeremy.

A

That can we can we find the BPI b1 and then we can talk about which flavors you want them in like span versus memory, but I'm no I think even if you can want a string, what they give you just focus in the spring. When do we think that the current overloads are the white ones, I.

M

Believe so, just because of the complexity of the trim.

E

M

E

One, the the ease of the timing was talking about if your source is memory of charge, this is a joke. You have to do more work to.

M

Do more work and we we can add the memory of char and it was actually asked I just like it's. There are other api's in paths that manipulate strings, that take really span and probably should take memory.

M

That was my that was my my feedback on that I'm fine with adding it. It just seemed weird to add one and then sort of dangle the other ones off sort.

L

Of wanted to deal with that as a separate issue.

M

We're at appropriate memory of car to path it overloads right now and just do the whole thing holistically. It just feels like that.

L

No one knows what ends and directory separator means. I think.

I

It means ends in path, doc, directory separator character, which is the property we expose. That exposes the array of all directory separators. Yes,.

L

But I think the people, but then also think that it very validly means if I remove. This am I talking about the same path and that's not true on UNIX. If your path is front, slash and it's not true on Windows, if it's simple, a backslash and so there's the question of like does the is rooted need like there are weird states once you're talking about those the dinner at paths and and and if Windows you just wrote, backslash backslash question mark backslash like I, don't even know what that means, and it means nothing.

E

Maybe I missed it. If I pass is C colon, backslash and zine directory separator, it tends Journal, it returns. True. No, it is so I thought we basically said we will make it smart, no.

L

Travis, no, this to true.

E

L

This into the directory separator and you go great, so trim it and it goes okay, no.

O

People would cease to methods, they have something in common right. That's where I think this is setting up date. They have two different.

L

Meanings but they use the same word to describe them and I think that's a pit of failure. Well,.

M

I'm I'm happy for suggestions on naming it's just sort of awkward I'm.

L

Just trying to understand who calls the boolean one and who calls the trim one, and can you get them on a common primitive that has a universal.

I

Meaning right now in source, fountain and it looks like ends and directory separator- is only called by crit.

M

Well, no there's its shared code, so it's all over the place right.

L

Because it, if trim, is the user space, primitive, I called trim and it didn't change so I need to add a separator or I called trim, and it did change so it was normalized at man. I want the old string so.

I

Internally, we've got both trim and ensure yes and ensure is the one that you say if I wanted to have an ending separator I call that and if you want it without the new Pomfret, maybe that's the right thing to expose these.

L

Are just annoying with the span because I need to give it a destination right and if it was fine and you need to write to it and it's like yeah.

L

Trims easy, so you can constraint over the exact same memory, I.

M

Yeah, we can say, trim nan, Rudy, non-rooted, cheering any directory, just trimmed ending separator past root, or something like that. So.

E

I understand that they are still shared within our code, but can we just not expose the ends in like they seem to be basically very tricky to use properly? It's not enough dirty separator yeah. So we basically just.

O

E

We add one more overload taking memory of char and when we go further.

E

Why do all the public ends in directory separator? What why do we need ends in directory separator to be public good.

M

The same reason, we use it internally all over the places to know that, like when we're comparing paths, I.

E

Just died just.

M

Me change to zip to actually call this and then add it. It.

L

Starts with- and you want to ask- is this directory that directory you want to normalize the directory, which is make sure that there's a slash on the end of both of them yeah. Well. So if you want.

M

L

If you're, comparing.

M

Two plans: you trim it right if you're like for equality but you're if you're checking for starts with you have to actually add so.

E

What I was think there are scenarios like comparing? We should just get archives.

I

E

Having IP is in the middle there and either perform it now, convene a convenient. The same thing, I think is similar here. It's like I think we should either have high level IPS that correspond to scenarios or very low level primitives that you can compose this.

B

One seems to look like.

E

In the middle it has.

D

Some hormones, but not.

E

E

M

A really I've been thinking about this, like ever since I started on the team three years ago about how to write, compares for paths, and it's really really tough to do because there's when you're talking about equivalency, you know what does it mean like when you get something? That's like C colon backslash, and then you get one. That's back. You know, backslash backslash question mark back, slash, C, colon backslash, and then you get another one. That's like represents the device syntax it doesn't.

M

He put the C colon in it, but it actually points to the same location.

M

It's really really hard to like represent that in an API that people would not fall over themselves, but.

E

People try to use this ends in will run into many of the same problems, so it's kind of like we are just passing the bat correct right. Well,.

M

The the the most likely API that I would probably it's just like you have to puts it really verb extreme for Abidjan ever it is it like saying like the because again you say like is nested right. You know it's like well, that's only from a spring perspective right. You know you can still represent that same nested path in a different way.

M

L

Of them is a path problem, which is the string of letters that you type and another one is a filesystem problem, so the device syntax and C colon backslash. They are a file system, the same that they are not path missing, crimped, because they're different strings, so the path class would be about string, manipulation, not about logical, go resolve, send links, go find, am I going to touch the same block on desk. That's that's the fun yeah.

M

L

That people walk down that path pretty quickly indeed, but I. Think of it. So if we make things that are very clear about the level of sports they're doing for you, then we can answer a certain set of questions and they're, probably the question people are trying to answer and they're writing their own bugs, because even if we expose these like one of them is a stick of dynamite and the other one will it match and they can't put them together in a safe way.

E

But do you think that this ends in directory separator is a string operation or a file system operation? The answer? That's a string. Well, I guess, but you see it implies that it's a file system operation, because director is a file system concept, not a string concept.

E

So I agree with you. If you get the method remove from strings or you know, that's the string and what that's.

L

Where it comes back to it does it should have showed it such a public method, return true or false for C colon backslash.

I

Do you want to compare normalized strings, in which case you keep the same like path, route and everything, and but you give it a consistent syntax or do you want to both? Or do you want to also resolve that to a consistent location is the way I would like you that difference.

L

And then there's the whole windows versus POSIX windows versus UNIX problem on capitalization, so vs. is writing to a fat system yeah. So.

M

L

M

That people want- or is it's hard it's hard to represent and trying to be very, very careful about how I expose additionally TI's here, because you know the is rooted when people get wrong all the time combined. Obviously, people get wrong as well. That's why we had a join, but so in any case Li. Certainly, the trim line is like super important to like provide now the name and I'm totally open to coming up with something very explicit is that's a super common bug right.

M

It's trimming, trimming a root directory.

E

So the only issues with this memory of char but I don't actually have a problem with the first two methods because they are super useful and and then at the high level they are, they are the extreme. They are the high level they do. What the caller, who wants to do, yeah the two other ones. They seem to be like trying to be primitives, but they they kind of, makes a bit of I'm a primitive that doesn't know anything about file system but primitive.

M

That doesn't know anything about the file system. The answer.

O

L

Today, if I can answer them, yeah that you didn't that you as the caller need to also understand a good path. Past that point the end zone or you're gonna trim well, if you're gonna do in something you.

M

Might need to understand where it has no I mean only if you trim it's typically that one is like went again to add.

M

Right so maybe the right answer there is just put in sure and.

I

If we hadn't sure not know Joanne's.

M

I

We hadn't sure, maybe we could just document that if there is not if it's like, if it already ends in the directory separator and therefore to know off, then the span we return is the span. That was input rather than a copy of it. That way, it's still efficient. Wouldn't you don't return a span right to the SEM.

L

Well, so we wouldn't well either we take the spam with you and we youtaite it or we take an input span on a netflix fan and we have to copulate.

M

So anyway, I can see. I could see that I can save the argument of like not doing ends in and putting insurers, because that bling yeah.

L

I just want to make sure these guys don't have a this. Is the right API to call unless and.

L

M

Not there's nothing wrong with calling heads and directory separator it's just what you would done unless you're writing your own version of trim and it's a really bad.

M

I could see that people would miss the fact that there is already a trim and calling in themselves building their own, so fine I'm. So do we want to change this to trim and ensures ensure.

E

But then answer has this problem of the new things in the queue right here. It needs to take like 2 buffers in yeah.

L

We would normally have to take one span and oh, we need you couldn't do that, because we need to know that we would have a trip we would have to have to try and shorter yeah, so so in shorter span would need to copy yeah.

I

Unless we expose the end that way, people can generate.

E

It so I still do we have more scenarios than just trim where somebody would want to call this engine, and you mention you say really, but compared is so complicated that I, like don't believe that anybody will write it correctly. Sized implements that end. The scenario no.

L

One should see their.

M

Own that they should you join yes, so it's the compare. One is really what that's for, but we just said: even you cannot implement the compareto correctly again. I can't implement it correctly. It's just that people take it to mean something that it doesn't mean right. So that's that's the tricky thing like coming up with the right burbage for that, so that they don't like.

M

So it's like the the old security stuff that we have right per pass on desktop. So right.

E

When you try it, when you try to restrict.

M

Somebody to a particular location.

E

Let's say somebody tries to implement, compare this ends in directory, separator the biggest problem that they had like.

F

E

Of code that is super complicated, and this one is just like: hey open code.

M

The biggest problem is that they're like this is three well, that's. That's the easiest mistake to make. It's just not add, like just assume that something is underneath, but it's.

M

It's it's sort of the the negative logic on that right so like if you are going back to the security restriction right, it's easy to block access to a given string, right to say like or to allow access, rather so, like we can say, like you, can access this string of a path right, but we want to block that thing. There's so many other ways to get into it right that, like it was impossible to really implement correctly.

M

So this, the the same issue comes in here when you're, like oh I'm, trying to provide a compare path is making it clear that, like what that means is that that thing is definitely underneath or the answer that comes back outside of that is not false. It's maybe right.

M

Okay, she'll be so sort of like the file exists problem that I also tweeted, like it.

A

Seems very kind of like in the weeds are discussing and designing the API. Do we think there's that? Because we only have five minutes left either? Rather sorry do.

M

We want do we want to come back with a we. We what.

E

I would do is I would approve the first two methods, I. Think there's some controversy correct. We all agree. It's it's useful.

L

You want do you want the spam trim to return the int of the length or do you want to the span or the charge recited expand in the memory overloads or just do the span overload and let memory people understand that they have to do extra work. I.

E

I see the negatives of memory. I would just add memory because it's high level API for the average developer, who just wants to freaking, Shrimant and good and move on the int. One is kind of not convenient enough. That I will just add it. Yes, we had three overloads, but at least the method is very useful and it's simple to use, and it does exactly what you want. That's the latter I don't feel super strongly people in the room thing that we should add them.

E

I don't object, but I would I will discuss it more because I think it would be useful to understand deeply what are all the scenarios and what you really would have to do to implement the scenarios called the correctly because we I.

M

Think we should come back with it with the comparison with in with insurers for the second part of it.

I

The msbuild team, so we can just go, get me hijack comment back on how they're, using it and yeah.

L

The bullying one concerns me of you: it's a you really need to know what you're doing API on a on a type that is notorious for people really not knowing what they're doing.

M

Yeah, so yeah I'm totally fine, with coming back with that one separately with a discussion on ensure right because I, presumably that's really what we want is we want to ensure API yeah all right, because you're.

N

Either going you're either going to trim or you're going to add great, then an accurate.

M

Yeah right read-only memory right.

A

Really this and then I think we're done for today.

A

There's no way you can remove more api's all right. So then this is this one then I say thanks to the people online me try to figure out my life disconnect this for you now.