.Net Foundation Design Reviews, 28 Aug 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Reviews GitHub Triage

Description

00:00:00 - Approved: Add MemoryExtensions.Contains https://github.com/dotnet/corefx/issues/27526
00:26:42 - Approved: Trim/TrimStart/TrimEnd methods for Memory and Span with trim element https://github.com/dotnet/corefx/issues/31011#issuecomment-416678327
00:38:48 - Approved: Add span-based overloads of String.GetHashCode https://github.com/dotnet/corefx/issues/31302
00:40:37 - Needs Work: Need one step API for creating files on Windows with security https://github.com/dotnet/corefx/issues/31488#issuecomment-416680473
00:44:38 - Approved: File.Move(string,string, bool) overload https://github.com/dotnet/corefx/issues/31511#issuecomment-416681677
00:48:23 - Approved: CustomAttributeData's AttributeType property should be virtual. https://github.com/dotnet/corefx/issues/31614
00:50:56 - Approved: a new GC API for large array allocation https://github.com/dotnet/corefx/issues/31787#issuecomment-416688698
01:09:36 - Approved: Expose the ability to create signature generic instance types. https://github.com/dotnet/corefx/issues/31798
01:15:58 - Approved: Expose additional MidpointRounding modes https://github.com/dotnet/corefx/issues/31902
01:19:36 - Approved: Expose some additional Math and MathF operations https://github.com/dotnet/corefx/issues/31903#issuecomment-416696207
01:32:20 - Needs Work: Add string keyed dictionary ReadOnlySpan(char) lookup support https://github.com/dotnet/corefx/issues/31942#issuecomment-416704108

A

Yes, no, we also have audio right, no idea how the audio quality is, but people will complain if it's not good enough. That is the all right, so it has been a while so I, don't think the backlog I'm actually doing one thing here: somebody get a bit more screen real estate on your hands, just content, alright,.

B

Can you give us time, let's do this.

A

C

I can hear you perfect.

D

A

Hopefully, the audio is also going over the wire. It's not too bad.

A

So, thank you very much for showing up. You have finally a meeting again, but we have enough people to actually have a meeting. Otherwise it would have been very sad. Just me, looking at my monitor, make a random decision which we don't want. So I suggest you just go all the strenuous and then just see what one's out I think this one we have talked to death.

A

It seems it seems, keeps coming back up there now these two already or should we buckle them up with her bundle them up with the hardware review that we do.

C

Neither of them are strictly tied to the hard work and, let's just take a look and see.

A

E

On our comments are pulling on part of it right, we should probably keep them as far the hardware yeah.

C

All of the Armand should be part of the actual meeting. What.

F

About this one's from now Same Same Same Same Same Same we're seeing the first one.

G

Let me just go over the back. Look at this I.

G

Mean it's different for this one give me it was worked on him.

A

You know we, basically, why am I looking at an hour like all it did escalate, so bad.

H

I

Like the picture, though, all right you wait for Steve on this one.

A

Then our area does anybody want us to walk us through this or should be this media top-to-bottom.

C

So, basically, today we expose several types in system. America factor which operate on float, but for certain types of programs float, doesn't provide enough precision for various operations, things like cat programming and the like, where you may need the additional position that double provides. This is basically proposing. We expose new types corresponding to each of the existing types. That would cover double, but we also make it generic so that we could extend it in the future to other types as well, if needed,.

C

Yeah, like half or quad or int.

H

So this point the originals will become go down, but once we have the generics.

C

If they could be potentially become redundant, yes is.

H

There any efficiency Causton in big making generic.

C

Not that I'm, aware of what.

H

Do you mean why it isn't it isn't obviously performance, sensitive type, so I just don't know whether the curve changes when you slide you, when you change something, did you know? Oh, it's using all.

B

If that were a problem, wouldn't it also affect the factor of two, because I would imagine that the jet has highly optimized all of those times or could be, could be changer.

C

Yeah, the only limitation that I know of that exists today is you wouldn't be able to take the address of these vector types, but there's a c-sharp proposal to fix that so I'm not sure that should be blocking.

G

So what is the proposal to fix it.

C

Basically, allow you to take the address of any struct, even a generic struck, provided it meets the constraints that it doesn't contain. A reference type.

B

C

But some people use these types with unsafe code. Today, yeah.

B

But wouldn't we be extending that.

J

C

I was saying: was the existing? Some people use the existing vector two three and four types with unsafe code. If they wanted to also use vector two of T with unsafe code, then the language would need to be extended. So noise.

A

You know people that do unset with the problems in engineering that can really take a dress off right. Yes,.

B

I guess what I was asking is: what about people who don't want to use unsafe? Would there be a translation between the generic and non-generic also for developers, know I can think the.

A

Scenario is you're using unsafe code today, using vector 2 today, you're switching to a vector of double. So today you can write address of that same X right. Well, you on your pointers, they're already there already yes, so people that are already doing unsafe. The question is just they can only do one say, because we don't have more support for you.

K

Abused in the drawback of using generics is, they can no.

B

No I get it I get it I just didn't want to force people who weren't using unsafe just for music as.

C

This work so there's also potentially we've done here where so the current proposal exactly mirrors the exposed API for the existing types.

C

There are some problems with the existing types that have come up before, such as the fields being publicly exposed in the naming of various parameters, which it might be a good opportunity to fix these, so instead of exposing individual fields exposed properties that wrap them, which allows you to make it read only which makes in more efficient, etc.

A

So this feels like a larger review. I didn't, though, because we probably want to go over the whole thing, I guess, and we probably also want to think about what that means for the overlap right with the existing types. These.

H

Are presumably just sounds literation to what we currently have all right? Other names, everything notwithstanding all Tennessee jazz, routine yeah.

A

How do we ask him something like ad with these types we just say you can only use them to float and double, and so we just specialize those two things.

H

If iti added already yeah, that's good, that's typical to me in.

L

Thai, that does not you will never ship.

K

Their intrinsic, when you old.

D

K

Fallbacks right.

A

So but then the question is the software for bugs like. How would you implement that because you can't add two T's right, you would have to say if, if T equals flow, then you know cost to flow to the thing and then, if double did that thing and then L is like oh yeah, you use the wrong time. So.

K

What we do for a factor of T right.

A

Like for vector T, do we support way more than just in just floated doubles where to support them for all numeric types or just float a double.

K

C

Current proposal is just how effective T do float and double, and then other types would be reviewed on their own merit as they came in, because supporting like int here would require additional work in the math library as compared to float or double so there's additional work and review that would need to be done for any other type of tea.

C

K

Know this is an implementation question, but what we also making this API would we also have to change the JIT to start recognizing these, or would we start having Hardware inferences I? Would.

C

Think that Hardware intrinsic would be a better use case here, as it would simplify the implementation required for the JIT, which would be no changes to the JIT required other than maybe this type recognition.

A

So about this, I will table this running and we do a dedicated review for this Berk item, because I feel like people. We should also invite you and some other people are onto that to actually do the whole thing and then because it doesn't seem like a triage should like the level I didn't. Let me just gave it to all right. Then answers in America to have 16-bit floating point number performing to our dad sounds reasonable.

C

So for this one, the proposal is, we expose in interchange, type called system, dot, half which represents the I Triple E binary. Sixteen type there's already hardware support today. This would be required for the hardware intrinsic classes that have already been approved for f-16c in the arm. Half is a, but it may also be required in the future for a certain interrupt with various machine learning. Libraries, the.

M

Commonly referred to as half in this book is Single.

K

K

So you actually misspoke, like it's system, numerix half right, we're not proposing it to put in system yeah.

D

N

C

Proposal was system, dot, half hey, but it yeah. It looks like it. This one says this to numerix interesting.

I

H

Question is when someone types a chime, the default template to this e half yeah.

K

Because other company than 50 other complications with this too, is that there's there's the I Triple E format and then there's the B float16 format, which is kind of little Google proposed it but they're the ones they use it in their machine learning and it's a different.

C

K

C

Float16 in the system don't have have entirely different use cases due to the range of values that they can represent.

C

But core CLR respect to B I Triple E. So if we were to expose one in the system, namespace I would think would be half antenna.

H

Would consciously decided we should wait up to thirty sixteen, or is there any harm in adding it? Because we know we're gonna write it anyway. It.

C

Probably needs to be added as an interchange type in the future for certain types of Interop.

C

But I don't believe, there's been a formal proposal put up for that one yet, and this.

H

Is there a better name in the industry than be float16? It's.

K

Not that I know.

H

K

We had both this would be called half yeah.

A

That also sounds much like a larger program. It's like similar to like charité or you jib eight-string, like it's a new, primitive time and I think we should probably think this to look like well, you know what are the cases where, what's language support for those two, it doesn't need language support that. Can you multiply them in the language right like? How does this work out so.

C

For half in particular, the only hardware support that exists and the real the only use case for them today is as interchange types so, there's no hardware support for adding them or subtracting them, or anything else, there's purely used to save memory space unless you're working with the GPU.

C

So the API proposed here is the bare minimum to be able to write basically convert between half and system, not single.

A

Let's basic checks on it: I just want us to think through like what happens if this becomes a full-featured thing, eventually right, okay, okay, would be very sad if you would have to introduce system that half later because well turns out. We didn't design. The type is to be an exchange type right and that's what I'm saying launch I would I'm.

C

Sure people will request math support, etc in the future. Exactly.

A

And that's why I said like we should probably go over this whole thing and then make sure we're where we list all the concerns that we have introducing and you exchanged up at that level and then see what what works for so, the backing would be somehow packed into a single. Is that right.

C

No, the backing value would be Dubois.

K

Yeah, it's just the conversion. What happened from those two whites to a single man exchanged? Obviously.

H

We care about what the banking, historically right, okay, so.

O

If we added the math support later and right now, the math that math support was terrified. The conversions. It would be a breaking change to add the math support directly later correct if we're changing it, just saying or doing the math or incoming back. That would result in potentially different values. Women. It's.

M

O

M

Would yes, yeah.

C

The conversion is non lossy, first half to single and back it is loss. It can be lossy for single to have so.

K

If you add two of them, yeah do.

C

It them convert.

K

Them both to singles and then head back together now they're hard.

D

K

To 1/2 would be lossy there.

C

Are specific, I Triple E rules for making sure that you compute all your results correctly, which is basically for any operation? You compute to the infinitely precise result and then round and exactly what an operation is varies at the I Triple E spec level.

H

Suffice it say the correct result for every one of these, for those methods is well defined, because we don't need abate that correct.

K

Because we're not proposing the math operations yeah.

A

Either thing with those is adding them later might be impossible.

N

A

The concern I have it's the same. This discussion we had with char a look like you know what kind of language support do we need, because if you don't do it now, we can never do it later.

H

A

The thing is, if you, if you, if you add them, let's say as operators today, let's say you have a you know a concrete operator like a plus or whatever on that can actually define operator. If they don't, you want it to be primitives in the language, you can longer change them, and sometimes you want a written language, because you want specific behavior that you can't get user-defined operators, and so that's the thing where you decide what to do with them. In.

C

The case of half float a little bit of a difference because there is no language to find semantics like checked or uncheck that could change the behavior of the operator. There's only one defined spec for that operation and it's exact. What.

O

You said if we go from single back to half, if we went from half to single, did math and then went from that result about def back to half. That last step is lost, a correct. It.

C

Can't be lossy correct, so.

O

So we could get different values if we add math support later. Therefore, this is a breaking change to do later, correct. It's.

C

Only breaking it, it's only breaking change if you have an implicit conversion, so that was the problem with system math and why we had to add math, and there was an implicit conversion from single to double, which meant that if we added the overloads to system math, anything that was previously passing in like an integer or a float could now return a different result. If there's no implicit conversion, then there's no risk of there being a breaking change by adding the new overloads.

O

Okay, so the only time there would be a breaking the math unless you created your own operators and us replacing those operators, we would take as a breaking change without too much concern.

C

Well, nope, no one else could add extent. It's just system dot, math. So long as we don't define an implicit conversion and.

O

Just won't be able.

C

To do the math, no, they.

M

C

Shoot the math themselves, but they have to do their own librarian, like my username space, stop math, okay,.

A

What I'm saying I think there isn't more than the ones to that that we should probably go over I'm, not sure that the triage meeting is the right way to do that. So that's why I would alter table this way, because it's like a very yeah involved thing I think.

F

She's with any simple request.

E

K

A

Guess that's true, you only left the act once.

O

P

O

Thought through this so I don't know what are the implications to them to the to VB and other languages that don't support extra graph usage outsider.

O

A

Then there's a broader question to that, which is the discussion we have of Matt's a few weeks ago about how do we version expression, trees at all in a compatible fashion, like what's the expectation for new types to show up for existing implementations, for example that are walking the trees and I think that will be part of it? What happens if you know somebody overrides is by reference? That's true: what's the expectation for consumers is they will they blow up today?

A

Do they I supposed to handle that I think my gut feeling this is like it kind of has to be this way. Like I said you need to be able to handle additive stuff, otherwise we can never ever add anything to it. But yes, I mean there's. Certainly a compact implication and I think that was the question of whether we can somehow version them in a somewhat inappropriate way right, but I.

O

Don't think we.

A

Can realistically there's.

O

A

Version for that there's.

O

Better the versioning problem and also whether or not we're going to express something here and off the top of my head I. Think in that category, where Vivi simply can't express the concept right.

I

O

Therefore, if it gets headaches an expression tree, it contains that concept, then it's rather problematic well,.

A

Made it depends on how you use it, but if you inspect an existing expression tree, it might not matter what that you have.

O

You can get.

A

Represented in the language, but you would never get an instance from VB from the letter here.

O

A

Yeah I think the question is like just what I don't even know how you would use them, and also I'm, not sure like I, don't know enough about is reference by RF and expression trees to see whether this is actually the the right API. That's something that the linkage what would take a look at which ones them? Actually, why not? Gladden.

O

M

O

Know is there an example of why somebody would do this because I'm a little confused by why somebody would do this and in.

B

Robably kujan I.

J

B

Marcus is the next yeah.

B

We have a lot of scenarios and dynamic elective and land expressions to generate methods on the fly. Think of things like serialization, that's, this might be an optimization.

B

O

It's just our generation stories, so bad like.

B

My initial reaction of this is I, don't like having Israel rep as a first class we will take on these. Ap is just because that mean like that strongly implies that that's the only attribute you can ever put on that parameter, and what, if we want to add something like is not normal in the future? Do we now have to add another over load of this? That takes is really rough and then is not malleable like it. Just doesn't seem future if it seems like it should be a flag rather than the whole.

O

Flags enum, so it can expand. Yeah.

B

Because I can imagine that as the language matures we're going to want to add multiple difference. Attributes not like a CL are attributed like the the generic concept of an attribute to these parameters. Yeah.

H

A

Yeah I mean that's the thing that me I think needs some design work. So that's I. Think again, that's a work out is that I think we need to take offline it like ask Omar, metas opinion or work. We know what should the direction be for this I think it's fair to say that currently there is a gap here which I don't even know what happens today into shop.

A

If you have in an expression that involves graphs and Rev suite only in the you converted to a land that is it just fail, does it do anything meaningful, I, don't know also.

G

Q

G

Have actual right now, the upon.

Q

Expressions on the link so.

G

Q

Chart is direct person all right. There are two people actually who are supposed to embody. Everything perfect, but Jared wanted to be looking so big in yes,.

D

Q

B

Cs, don't free free, free pound, look an easy API! You wanted more right. Yes, thankfully, the person that once it is here subjected okay, we close the the it's basically a discoverability and performance, optimization, it's faster to do contains than it is to do index up, and if all you want to know is if a memory of T of sanity contains something well sure but contains. Why is it faster? This is because the way that index up works right now is it vector, rises, the search and then once it says this vector contains the match.

B

Now it has to go element by element with a net vector and see what actual item in here matched. You can avoid that last step.

A

I mean just me, usability standpoint, I hate using index of and comparable, you know, or a square that it's larger or equal to zero. We're just the problem with strain. This I'm a little bit missing anything I mean it seems in line with what memory extensions already does so at least, if I combinational over those. Though sorry are you sure, that's the right combination of overloads Spanish, you.

B

Might also want contain setting.

A

So we need the span ones because it's extension methods that don't show up. That's alright, I think it's the same problem. There's this thing is a yeah I variable flavors. It's.

M

A

Pattern we used on all the others yeah, and then you have equality, compare which seems reasonable. Do you want a string specific one as well this evening? Real news man trummer? Well, there's a we don't so the other ones the compared index ones have we'd, only spend off char and string comparison as an argument. Instead of the compare Oh, which is the evening.

B

That seems like something we should probably do. I don't know, one was asking for any output, there's just some consistency with what we already have yeah.

K

You contains this string.

A

Right, it's basically like the same as index off, except that it doesn't do we don't.

B

Know you're already in Excel they're already using contain so takes a string comparison, look.

A

There's one yeah: there's an index off we'd only span of char with string comparison, and there was a contains as well as it contains never mind contain. So there exists a good cook on all right perfect. Then this was as good as good that.

I

A

I didn't know they approve of you guys anymore,.

D

A

Right, that's a definition, get semantics I feel like this. One is yeah that doesn't.

I

This one so yeah I just think.

A

Yeah, this is the other thing Weiss. Did you talk more with them about this after this yeah I think that's something she also was interested.

Q

In this one, yeah.

A

So the problem that we have of this is we need to. We need more nuance here right, so the problem is has just might not be enough, so the problem that we have is there's a proposal to support expression. Trees, I started to support interpretation from Revlimid, so the idea is that you can recommend anything, including types and of this. Hill are completely interpret this now. The question is: do you want to do this? If you know it's interpreted, so how do you? How do you?

A

How do you separate these two things right, because I mean there's this general follow of like well? Is this? Is the right API to ask, or is the other API I would like to say it's like? Well, it's not supported at all, and maybe we should have a different pattern for that, and then this is crossing over what this is not a standard discussion about the future. Well, there's.

M

That yeah there's.

A

The capability API aspect to it and then also I'm, not sure, that's the place. To put it because one-time feature was I mean ignoring the name for a second was designed for the compiler to figure out what features will be available at runtime which might not be Atlantica. So this is the been on a stripper pole.

A

I feel like this one: it's a bit more design work. It's.

D

K

Wouldn't be in runtime, information might.

R

K

R

Also be on remedy PS itself, what we talked about, yeah we're pulling.

A

This for some of these API Livington have just.

R

A

Me real fast, acting sections, remember well, the problem is, it will not fail right if you have an interpreter. So these, if you have a rev limit based interpreter, you may still want to say you know, interpreted ILS still slower than let's say interpreted experience right. So how do you make that trade-off? If you don't know what's happening.

H

All right all right, so this was another yeah. You might end up with his recommend fast instead right, otherwise,.

P

You could ever get what's left for me, yeah is interpreted or something right. I like.

O

L

Made false but is resident fast, yeah, faster, faster.

O

H

H

You could say that there caressed after that work, I.

A

Was able to say who just rented them all that's incredibly hard? Thank you. We're featuring.

M

We talked about in we're trying to figure out how to rationalize it with the rest of the.

S

A

A

Okay, so I think we are in alignment with the wolf, the namespace and the assembly. Location I, don't know where they put it now, who did they say? Look like you said: a new contract, it's just about technology. Well, that's where you know I said no way and how and then FL yawns at the same thing and they said he updated the proposal, but I don't see they maybe just removed the the statement about.

A

I mean frankly, I would expect this independent scholar or something right yeah, so it seems reasonable, we're exposing it publicly yeah. So there was I see in a normal moment now I think the people that I gave them was like. Do you need the public, API I think this is the link time feature.

I

Okay, let me just respond.

F

A

Didn't we talk about.

B

This this sounds great finger there, we top them off, read only memory of child. This is a generic trip.

A

Oh for arbitrary elements, yeah so.

O

E

O

Teach room element you this is everything dropping everything after that? What does the gentleman do? Its.

B

In the particular example given here yeah every every element at the end of the span that matches a trim element is dropped, you're, basically reading from the back the first time, you see an element, that's not the tremendum at you stop, and just whichever one you have one. So it means that the term element can be present in the span.

B

It just has to be in the middle or at the front account, for example, consider but you're trying to turn a period alright from from a piece of text, because maybe you want to normalize nine-point-two just 9. The term element here would be trend. Okay,.

Q

Is it useful to have actually released re or an emeritus of the elements yeah.

T

Q

Was the question Erica.

B

T

Don't know to be told so we have trim on.

H

Anything other than string I've never turned anything like this forever. I think.

P

O

That if you had a large array of integers and you want- and you were filling from the beginning and they were pretty and prefilled with either negative one over zero and then he wanted to ditch the ones you hadn't filled because of his span. And so you might over allocate and then trim so I think.

C

I made that up, okay.

O

C

E

O

Know it's yeah I, don't know why there's a lot of situations in this makes sense, but I wonder if it's for something like that to keep you from having to me or.

H

Isn't it may be that there's different like the equivalent of cream jars, you may want to have more than one value. I guess you just do this repeatedly, because it's cheap.

B

It's actually non-trivial to trade, multiple characters, you mean economies multiple times. Is you call it multiple times in a Lukas entrepreneur? Think if you're walking in awe of everything? Well, because the naive way to do it is say that you want to trim a B and C from the end trim a trim, the trim C. Oh wait. Now. I have to go back to the beginning, because our Mike's DNA.

H

B

H

So I wonder then, if it's worth thinking it more about what the answer whose case is because you might decide, you need trim elements and I, don't even know what that's really is.

G

In this proposed.

H

H

B

For instance, system.string already has over loads of trim start to amend that take an array of characters to try yet I mean.

A

It I mean I would say it's consistent with what we have. If you want to extend it to arbitrary T's, that's what we would have to do. I'm just I've, never seen one entire house, I mean I, I, also kind of recall this that were using and it I also don't think we can add the.

K

Social English fault.

F

K

Right, like the very first one trim, Trinity.

B

K

There's no trim element being passed in on that very first one yeah raising no yeah, which one.

A

I think trimming Horace. That also doesn't do me full of charm if it should just space I, don't.

K

Think, oh yeah.

A

Oh, that yeah so I could think they mean okay. So we can't cut it's.

K

Either make it sweet? Okay, there's.

N

Q

Title stress about memory and span: we are by putting in read almost I'm gonna cover everything.

U

A

Know because the extension method, sylvia's.

V

A

You had a memory now you.

V

Have spam, you don't have a memory anymore seated in cribs memory, and you don't know how the trunk might work. What you get.

N

Back is man get back? Oh, he knows yes,.

V

But if what you need to read only memory, the hell doesn't help you.

B

This is why what we discussed string based trim methods, we put them on read-only memory of charge and also relies.

V

B

I think we would need a use case to.

V

Know why were like? We can come up with some degenerate things, but I don't know that most people are ever going to use them and there any enough for people to write and the complicated cases they should solve for their complicated case. So this seems like superfluous API.

K

If you're using this blankest ring, so we added the one for charm.

V

K

O

K

O

One was we're.

V

Making the string API.

N

K

V

I bet yeah yeah. This is for intz and doing I. Don't know that I want to trim. Do it.

B

I think haffley's case was actually the most common you over-allocated and you just trim off what you do. But to be honest at that point you probably have the index high, which is the last index that you feel that you know at that point you just let yourself at home.

O

And if not you're, in the simple case, where it's a single value, so the yourself is not as bad as if it was because you know what you started.

G

O

There might be some good news cases in machinery walk, but.

N

Right, that's what seen be.

A

Shocked, that's true, but I mean in short, I mean any sort of awful.

N

A

Can be a killer.

V

Talking with general useless I would.

U

W

U

W

E

B

This allows you to get hash codes of sub-strings asp.net needed it and I'll dotnet needed it. We don't expose it as a public API. This is a proposal to your supposes home.

B

Take a look at the three apos. It would be out of to support the scenario. One of the competing proposals was to instead of half get hashcode as a static method on screen, put it as an instance method on span of tea.

B

The problem is spanning tea already defines get hash go to the behaviors throat, which also means we can't make an extension method, and we don't want to change the behavior of span up to get hash code where to use char, because now get hash codes sometimes throws, and sometimes your terms of value, and that would just really yeah.

K

D

K

A better place, to put it, though, then string like I, don't.

B

Know like putting a null string, actually kind of says I'm going to calculate you in the same way. I would have done for if I'm right, which is the contract, that against yeah.

B

Of the other, the other reason that we don't want to just change vanity get hashcode is that it means we also have to remove a data browser number five, which now means it's always going to show up.

B

Does it have to be much alone? Comparing for every other method is but compare info is sealed, so I, don't know why those methods are virtual, that was a discussion. I think they had.

K

Makes specific use case for that? One was I want to put slices of strings into.

B

A dictionary basically a Florida asp.net already has they have a type called out segment, I believe, which is exactly about screen shrink. Second yadda yadda, so.

H

It is this one if I remember right in order to break a reference on a call, as we move the ankle a lot of informations into extension methods, but to do that was reconstructed so because that plane was lost and, as Jeremy points out, that's not ideal security wise because they meet. You have to create an object and then subsequently so he's proposing new methods that don't exist on that effect.

A

So that one I'd like that to table for now for the sole reason that I was just going over the spreadsheets from or you know yes, this one comes up, but there's a bunch of other a clay, P I said come up and every time this comes up on my own did I need something for my I was like.

A

Are there yeah so, like I, think we should think about like how we what we do with the existing API shall be probably best like what patterns should be used to convert them because I mean the other option. Is you just suck it up right? You just put them back where they used to be, and just say, yep. Tough luck. We have echo API, hasn't the call that happened to throw um toughen or don't support a cause. Let's yeah I mean so to me the reason I the way I would decide.

A

This is like how many of these API is do we have, but if this is like a ton of API is in that bucket. That might be the better parent doing it. If it's, if it's not.

D

A lot if it's just the toy like five or six, a.

M

Partial partly but base cool stuff isn't what.

F

Is that stuff lift right now, packages.

M

Bidding on the name.

M

Are the base of this system is secure access control? We.

H

Know from data they're gonna keep hitting this in real conversions. Yes, really! The question is we want to get them to exactly and that's.

R

Why I think to me there's like.

M

Either we have new times we already have new times, there's already existing types exposed most of these things in there's, okay, there's either static methods to replace the Constructors and there's extension methods. For instance, methods might exist, so I think we already have like coverage for all the accolades via what was the thing I was reasonably looking for, it was the one that's I think has been a little bit tricky as the Piper security yeah.

A

I think the one that I looked at was inventing like there's, not even ting like yeah. You know threatening like that, was the event API that many we said event or something like that and that that had a create from existing or something and that doesn't exist anymore, and we don't have to replace rosemary beside that, because you can add extensions to that. Yeah.

G

A

Yeah wait but but you get my point like I think we should look at the all of a Calliope assets show up in there and yeah.

H

The point we send this back and ask Jeremy: they talk to relevant parties there to understand what it would why, whether it's possible to put the back yes I.

O

H

Just keep having this conversation, since you start having a conversation. Sang-Taek your character, uses conversation.

J

There's more landmines for not.

J

That's what it's not that we don't have any hurry. What would it look like today or the missing members? No! Well, if you tried to run an existing buyers, tragical value violators well,.

H

We have so many of those that we've created a tool. The cash.

H

Let's look this back.

G

H

What were they Peru, avionics workers he enters in the last 500 women perfectly? He doesn't.

I

Yeah, that's no I did but like it's a different, it's assigned to him so.

E

Only half this.

E

I swear to god, I've seen this before.

M

H

However, he didn't happen: English, oh I. This is funny I think this came up in the in the PR for max of the basic file system was Tony going on and all.

O

H

Do, inter all because this doesn't exist, we don't have the overview. So I was asking to take me into up off and call the higher functions. I.

A

Don't think you want to write that code.

M

You know I do write on any of you. Love emojis. Only.

I

Every place, and so I was about to say.

H

Neither incoherent.

R

Directory doing you know fine following four dozen others have insulation, potpourri.

V

I, don't know if it should be filed or filing for we're both probably.

M

Both if we're gonna add it, we should.

M

R

Took this here.

A

With a file in fluent directory yeah, what is directory info actually rely on, like.

I

Is it down good question? I was just looking at the move. I know you do, but I think it just close to this API.

O

Good situation so.

I

Oh I misspelled there's a point right now.

B

Where we're introducing a new type of material.

I

Well, that's what happens all right directory enforce all.

F

A

Yeah I think once you answer that I think we can consider this approved. It's just.

M

So the underlying I watched CPS take this. Do.

R

We really need to call delete first, which seems like there's a reason. I know this is the thing on Windows.

A

Move on till they use this, which actually can be pastas flag, but he's saying you can't also do this, but that I don't like the halter know. That is bad. It's like there's like a movie of things that can go wrong.

D

All right, because when you boot Linux.

V

Or Mac, like that, we do that's.

A

Why I asked any question like what caused them to men on the cross-platform thing.

A

F

Right guess: I had to be the type of what you should be virtual.

R

What do you, if you see this before I, haven't seen the request, know what I mean I'm just it.

F

Seems reasonable to me actually is.

A

That part of the animal poor boy.

I

There was probably a bunch of other things.

F

That we already made version for the same reason.

M

It's not the L marketer, so reading photos.

I

Yeah, that's the refreshment for a fictional world.

M

Everything else in reflection is virtual: I, don't have a problem. Yeah.

I

I

G

P

Which is gonna, be versions.

A

Well, we have generally not considered making things virtual the breaking change Ram. It totally is the I'm just talking about the divergence between platforms. Oh.

D

Well, that's a big problem.

A

But I don't want to fix this move on.

T

M

Shouldn't be dealing to be a long.

B

Maybe think we're gonna be quite elegant, wonder if the number of elements in the array is always constrained to 32 boats still any 141 lat.

B

But you can, you can allocate an array of like two billion longs, for instance,.

X

This gets to a red bull stuff.

X

You hear me yes,.

X

So that was the ideas that, like we jump, the bigger ones on the large object team right off the bat and we can get rid of some of our other cashing things that we still have. In course, you are.

I

K

X

B

B

Q

Weird just suggested right that baby will get rid of some bullying. Okay, so.

X

Amber eight we're.

Q

X

B

You know I hear me, mark I thought that we had some areas where we wanted to allocate a raise that the GC wouldn't move around, because they might store secrecy if we put them in the large object heat from the get-go, even if they're small rings that bikes minimize their exposure, we I thought we had talked about this. A few weeks back I know: we've talked about various pathways.

V

Making compaction that.

V

For clear memory, protection practice is only a security funding. When you consider.

X

So yeah this. This is particularly useful for long for long for long term, pins.

A

Sorry I said this last part: we.

X

Have that we have this further, like overlapped, is one of the things that needs that I.

A

X

That we have that we have a really complicated, semi, complicated way of doing it. We keep trying to age, those and so on and so forth. It's very old cache that we have.

B

Why would you remember passing zero for I.

V

B

V

Opening sentence they want to allocate a large object in the gym zero. Why? Because you're about to you're gonna soon as you're done with operation yeah, you want a big array, but.

P

Well, you can choose right. Well, that's a problem right! If you do Dave do I locate something that's bigger than I.

A

Think 65k, you don't get to a kid put it in there or.

B

Y

Like that 65k or something yeah.

X

That's that's obviously, subject to change, but you know the one of the other. The clear memory part of this again goes at the array pool most most of us that comes from discussions that I had with with yon and with with may owning it specifically. You know one of the guarantees of a pools that, like we don't clear.

C

X

I know you don't like that, but it's like from a performance standpoint. It mean if you, if you don't want me, you clear on pool on put back, is the way over a pool worst currently, if you care that it.

B

Clear so we my feeling on this is, we should be introducing api's and allow people to violate type safety like if you, if you really want to allocate a large array.

B

Like the API should be marked as dangerous, but only unsafe type like that part.

X

That part is fine, you know the whole clear memory thing is called unsafe or whatever I mean right now you do that I mean we do that for all of our local or stack stuff right for most of our assemblies anyway,.

X

There's there's eons comments down there: okay.

E

They are saying the tea cannot contain references yeah, if you make which I guess.

F

At least makes it man we save.

Z

Like jisu, simple yeah, yeah.

A

I mean I gather once you have interesting things that the GC might do when he walks to keep yeah but I. Think the question is.

F

Yeah for in centrioles, you may not care.

B

For Vincent, who might walk here in Toledo, which is general truth, was everything so.

E

B

I keep coming back to this point is because, when we had this discussion a few months ago, like I typed, in who I just went on to get home and ask who's calling a rave hold on right and the third map on github was assuming that the data.

B

Seriously well.

A

Which is still the case here, every not giving you might be for something that will be valid. You literally have to say clear memory, true or no.

B

You don't literally just like false, no I'm saying you sure you type in the word false, nothing, nothing about the characters, nothing LSE means I'm, doing something on say. Oh you're, just saying when.

Y

You look at the code said you can tell yeah I. Think that's fair.

Z

So you know having an overload was probably sensible for that. Yeah I mean you put immense.

O

Z

Array and I look at lusher a know.

B

Clear through some things, I.

Z

Did initially unsafe, like.

B

I'm, okay, with the functionality I just want the call site to be very explicit.

B

A

Thing I can I mean I'm, also correct like. Why do they need that? Like do, we know for sure? Is it we're.

X

Clear yeah yeah: it's a performance.

X

Yeah, you know they rape. They were a pool like trims itself. Now that was the feature that I added 2.1, so this becomes more of an issue of when they used all the money used to always retain great forever. You'd only pay the cost once so for long running processes. You know this can add up to more significant, improve the more significant of a notable improvement because you're going to be hitting zero things, you don't if you do stop periodically or an empty array bucket in the ER, a pool is more likely.

Q

Okay, all right so then, to summarize I think we are generally.

G

Sort of full of questions can.

Q

I, can I okay there's a small array of this as well yeah, even though it's what is, it called an allocate large, all right kind of like obfuscation or is the main purpose.

V

Main purpose seems to be to locate something not in the large object, even if they should have triggered a large object allocation. So it's allocate yeah right.

T

V

Kate without the large object so.

Q

There's there so that's the intent.

X

But at the end of the day, this is this something more right so yeah. Well, it's actually counterproductive to you know, go out there. The small arrays I believe are optimized.

B

So sorry, no people they'll start.

Q

Using it for sorry for the clear memory you know, because fulsome that stuff right or for doing random with generations I.

A

Think I think the idea is that you're not supposed to do it for eight element, arrays right, because they're like the benefit of this, is that you basically don't get to clear. But if you only have that elements above it doesn't.

B

So if you keep in mind, if you pass clear memory, false and Jen and say generation zero, like you haven't, done anything everything coming out of transy Rizal ready like it's already cleared by the process of jean sifrin, we literally just bump a pointer, that's out the I'm curious. What's the behavior, if you say, I want to allocate a 50 megabyte array in ten zero.

B

Look, do you expect it to fail, or do you expected to go to the Lord Psalm 24:1? Yes, what happens a.

Q

Period on another thing like what, if you about you, know nonsensical generation and these they're actually way to say LOH or is it gender I mean.

V

Even now we're getting to.

B

And maybe we just say the generations a hint and if we can't fulfill your desire, generational just goes wherever we make it.

B

Which would be fine as long as W heaters, not around I would.

Q

Even throw honestly like if you're doing this, like you know, drinking I actually might have break diploma.

A

That is that yeah, god forbid, we actually change our segment sizes and it sound like o throws where it didn't use to yeah. So if we.

Q

Just like rigorous well, if we regress perf when you're using this, like you know, it's not super guarantee, that's what you're saying I think.

A

It's okay to say: if you know, if you mess with Perth, then you may have to be measure when the new release comes out, because some invariant that's giant right- and this is like super advanced API, so I get.

Q

The flock yeah yeah yeah.

X

Yeah using GC in general is super advanced right.

X

Sorry, I got booted.

B

Yeah, what I'm paranoid about is uh even even innocuous cases. People will say gem zero, clear memory, false, not realize that it's not doing anything and then all of a sudden. We change your behavior when during production and loose my diapers, because what like judge.

T

B

Production environments slightly different than their dev environment, and now things end up with a large object, even selection0. Let me test them and now they're performing some snakes.

Q

Because we decide on the spot, you have you believe. Instead of like you know hanging, you know, if you're building against business is the limit 64, okay and if you were, or whatever like yeah well,.

B

Maybe there we have like an IOT device or something with each other. The zero gives even small.

Q

But again, like you're super optimizing for this API so.

Q

Everyone, unless they might be like you know, hey we can. uh If you ask for Jen zero, we're gonna, try one, you know before we go and roll away is a low. It's actually like free, medically or is in general, it's actually separat something right right, yeah, that's all so we'll do a Bosque flight. One tell await they get one free.

T

B

My generation- oh yeah, the oldest okay,.

Q

Oh, we have juice, it doesn't charge.

B

Is den to effectively oh, it I just lost my eye. That's the best of my knowledge, based on how things used to run okay, I.

X

Believe that's correct now that there is no real, separate LH per se.

X

Should she I believe she merged and I I, just I wouldn't stake my life on it, but I'm pretty confident a lot of my information.

A

Okay, all right I think I wrote up all the questions we had.

B

Other than this something one more question: what's the behavior, if you pass clear memory false and you pass a tea that has references internally, they said that that means that we can never change. Existing struck. Stu have references if they previously doesn't have one that is correct.

S

But that is a bunch of ApS and have that behavior today those.

B

D

E

S

Change the fact that the.

B

Coab will blow up right, yes, but I feel less bad about blowing up on the save code for things like primitives I'm, sorry I feel we're never going to change primitives, obviously, but I wonder if, like we're, potentially breaking up or we're potentially allowing a break for like the ecosystem, si you produce a package. Your struct has three fields intent and now, all of a sudden, you have a fourth field, object right now, whoever's consuming your title break. Okay,.

V

You said you wanted the clear thing to be on an obviously dangerous API anyway. Obviously, dangerous API is Venus I'm.

B

Assuming I'm not getting.

V

B

X

Not sure how you get that broken right.

A

What I mean I'm sure you're like this and.

O

A

Use Levi's package: his truck today has two in cement. I call this API I say don't initialize, then Levi adds an object reference to it and suddenly, when I pass his truck to my to this method, then it blows up because one more object reference it. Oh yeah blows off.

B

Sure so, like the thing is that you've had.

S

X

Thought you were saying it was possible.

B

To actually get through the API I'm, sorry, no, no I'm wondering is it? Is it okay, if that blows up due to a new dependency and.

X

I would say yes, you're.

A

Calling it advanced thing with yeah I'm, not sure that you were actually like I, think there are other cases where you you pass instructions and share things with beta types and if you use something that's already Bob Loblaw of today as well. I'm sorry.

X

S

A

Unmanaged constraint in quite a few places: yeah, it's not the constraint, there is another API. Maybe added I forgot what the name is, but basically the monster you pass it and type, and you say: does this thing contain any of any interior pointers effectively, so I have every any any agency.

O

A

Think that they did I mean the reason method is API is because the constraint doesn't work on all cases right. So the idea was that you can dynamically check those things and then blow up earlier it, and so those are the things where you probably not an unsafe context. You just called an API I pass in a tea and.

X

A

Things will happen so I think that's it's unfortunate, but I think it's okay, because that seems like a corner case scenario anyway. Well can.

B

We at least- and it do you have an overload, can we constrain T to unmanaged, because I can't imagine many scenarios where you're going to pass an arbitrary to Z? Okay, then it's.

D

Normally go to the invite.

B

Or something else right, if we, the other.

D

That's a good one. The other benefit.

D

A

What's the name unmended? Is that actually the thing to show.

B

It's like if, if you're, creating the library and you you have a generic method that is unconstrained and then you call allocate without a zero. We know it's like something's, clearly wrong. I can't.

F

All right I had that. Let's go back here.

D

A

Gcdp eyes are always fun.

Q

Gonna be a first user wind-up Benedum step. Surprise.

A

Alright and I sense back to my Unni, or would she be upset if you do that.

F

E

Oh man, those words.

A

I think we talked about this. This looks when we're there hoping ten days you know, but he has talked about like.

V

Generic method.

Q

Nep from previous version, isn't it a 16 567 think.

A

Yeah, oh yeah I think we talked yeah, that's what we talked about wait a minute. This.

G

A

So basically, what this one was for those that don't know the problem that you have is you want to instantiate a generic method?

A

The problem that you have is you need to find the generic method first and then you have to instantiate it, and so the idea was maybe we should have a way to make the search faster. So we don't have to walk over all the methods of the name, so you can at least you know, search for budgetary parameter count and then the thing that was another one where it was about. You know just one liner, where you can just say just instantiate this one guy, so you can just pass one method and say this one.

A

These are the arguments and then find the corresponding one. I forgot what the actual code symbol was, but.

G

Yeah, this was largest. We had, like you know variables.

G

Representation that.

E

Type of stuff right: yes, something like that.

E

Thank you, nice to meet you.

F

Anyway, we call it a signature. Time is again.

F

As opposed to a normal signature,.

V

I think its first generic argument to first specify the generic type argument. Yes, not grammar in so I'm. Looking for a thing that uses T, 0, comma T, 0, comma T, 1, comma T, 2,.

O

Oh need to be able to.

V

Say I mean I need the placeholder type 0, yes, that one's awesome.

V

Because it doesn't have a type until you instantiate it, but it knows that there is a positional type and so I think. Signature types are titles, positional.

V

A

I think that sounds fine, any concerns of that and we're justice. It's.

V

The number of type I don't understand. Why there's not a number.

V

Unless the signature type is to build a theoretical, closed, instantiate them and I, don't know if the titles, but it makes sense to Eddie. It's probably right.

V

The availa want to understand it well.

B

This is not a sample of using the new. Yes, there was a very I mean. He chose the implementation video now that she was the implementation we had about someone calling this well.

T

He's trying to replace the stuff over there right, if you have you.

V

Know no because the the generic method parameter zero. That I knew was an already existing public API or an existing internal idea.

Q

T

To make during parameter zero, yeah cuz, that's what I thought, but that's what we proved it in the past right: okay, I, think lazy check because that's.

V

What I thought the signature take was I think.

A

The signature type is your: let it I think it's the it's. The generic parameter effectively right, but I think so. This is basically this one here, Mike make generic method parameter, is basically doing the. What is that bang-bang, zero or whatever? No and and then sweet this signature type of work, bang, bang, zero I know it says, will make a generic type that takes this one as an argument, so he needs to basically do this right.

E

A

And though he can say get me the method with that name, but this is a type parameter and then the one with one argument.

G

Yeah, so what are we getting? Ultimately, this.

A

Yes, so he basically needs in order to implement an on runtime implementation. That is just reflection based, but he doesn't do this and call it. So you need to public constructor to do these things, and so that's what he needs to method.

V

Because I still don't know what a sinister Jacob.

I

Welcome to reflection their.

A

G

So we improve it because we just come on all right. Yes,.

A

G

Think, that's basically what it.

D

N

That's mommy just look at ya, okay, so it's making one of those whatever one of those I.

A

Was making one of those say that Turner.

C

Online yeah I'm here.

C

So the proposal here is basically the I triple a spec for a floating point type says that we should that we are required to expose for different rounding directions and we're not exposing three before that are required for binary floating-point types. These are two positive infinity to negative infinity and to zero when you're rounding towards an integer. These are equivalent to truncate ceiling and floor.

C

But when you are rounding to a specific number of digits, we have no operation that can do that today, which can help you account for various floating point errors, depending on your input type of depending on your input data. So my proposal is: we should expose these type, these additional rounding modes, both to be more, I triply, compliant, and to allow users to correctly account for inaccuracies in the floating point. Algorithms.

A

Yeah I mean people will not be happy, but I think just search the internet for midpoint voting and try like all the people that are trying to do rounding, and nobody knows what the these things mean. Like it's a like it, it's not super hard, but it requires.

A

You know a decent amount of understanding of what the is going on like it's. A typical problem of everybody runs into Brown I'm, not doing what you want to do, because you have this running from school. That is one way and then e of the other rounding. So let me add to your morning times, but I think it's I don't see an alternative to that, because if you need the function, you really you really need it and I don't know where else. To put it quite frankly, I.

I

V

Her naive question: if we added the midpoint around the you know the framework to Oh, how has it been 12 years.

V

I

That's true for a reason why we didn't add the.

C

Single precision floating point types until I went in and added them for to. Oh.

N

B

Is there a risk that some customers might have switched.

B

Sorry, what introducing new unit values doesn't like? Isn't that a competitor you.

P

Mean like when somebody wrote a switch statement and I only expect two cases. Nothing stops anybody from passing a method. That's gonna, do a switch on the thing from.

V

Passing in the number, if that max value, no, but it didn't practice, people don't do that, just like their codes already done. Yeah.

T

I thinking them values not working together. Okay I will remember that.

A

Week, yeah I would.

P

Don't you think in my framework.

A

Would I would I would.

Y

Say I would say like.

A

Y

A

Your point, I just don't think my neighbor will write a switch statement against my point. Rounding when.

D

L

A

Somebody else toodaloo, okay, so that's I mean my argument. When.

P

There's they're not my rights, which you're gonna write.

B

P

It's yes, which is why I'm.

V

Checking through the framework.

B

And I see it up in quite a few places, yeah yeah I mean we'll, probably, which I mean we publish my.

M

Statement it is a breaking change, yeah, what we allow it: okay.

N

A

An acceptable breaking change all right. These, the names you want I might.

K

Be missing something: is it really made by rounding those three operations.

C

For the so force for rounding to an integer, they are not exactly midpoint rounding, but for rounding to a specific number of digits, it does impact midpoint rounding so, for example, 4 0 11.5, we'll round to 0, as will 11.6.

K

What about what about it's, not necessarily it's right like if I want to round to one one decimal place right, yeah,.

C

Then the the the second digit impacts the midpoint mounting of the digit, that you're rounding to.

K

Right, but in this case, when I'm saying to infinity yeah, it's not necessarily the midpoint right like if it's point or if it's 0.01 to positive infinity would give me the next right. It's not midpoint, I'm gonna, think about it. A.

A

Bit just single D, here's a little bit no I, know I, think you're, right, I think it's just a you know. Miss miss main, probably should have been just called rounding note or something, but because it's still e, not just midpoint, might already right, because if you call an API, you say round to five elements after the period, then you don't really. You know I guess. My point is the wrong way to look at it. I guess: yeah cuz you're want to go to a precision. You want to wrong right.

K

Where the II knew like two values today, yeah.

J

K

Only four cases when it's the midpoint yeah but.

A

Midpoint learning is the only one that we have in the framework so like whatever. That was the name they picked and.

A

Darn it just watchin SRAM work.

AA

A

2.0 and they edited it will need a good documentation. Well, that's for sure that we already have that problem. Yes, it's like this type is not something people, oh yeah,.

K

The definition is specifies how mathematical grounding should process a number that is midway between two members.

K

These three values are not going to be perfectly right. It's.

N

Pretty Midway slightly less than.

B

Visions, yes, I mean I, guess one event is when you're called math dog round. Intellisense is gonna, kick in and say well, you're, more values. You have to choose from yeah.

K

V

Midway, it's just using a much higher standard deviation.

V

It's statistically the.

W

All right, Hannah you're up again.

C

Okay, so this is similar to the last one where the I Triple E spec, specifies a number of operations that are required to be exposed and we're not exposing all of them. These are the ones we're not exposing for some cases. You can't emulate them using other instructions, because the I Triple E spec, says that these operations must compute to infinite precision and then round according to the current rounding mode.

C

So if you were to do two operations like a multiply and then add, it would be two operations and you drown twice and get a different result. These are also useful in various places for doing floating-point manipulation at the lower level, which allows you to build more complex algorithms.

C

For I partially went through the spec, but people have been heavily requesting if used, multiply, add mostly as a as automatic support in the runtime, but having explicit support would help solve some of the issues that exist today, like the precision differences that could occur by automatically doing that conversion copy sign we already have today, but it's internal, and we only use it internally.

C

Next value in previous value are basically the equivalent of increment and decrement for floating-point operations, which you, which we also use today, but we have to do using big converter plus one and then convert back and then I log to and scale two. We also effectively use today in various places by extracting the exponent using bit masking and shifting.

F

I mean naming is interesting here, I mean Kobe, San Louis is that intended for which one puppies I shouldn't be capitalized. Yeah.

C

K

C

Integer log, which is somewhat different from log two, in that it can return a more precise integer when the integer is not exactly representable I mean.

A

Honestly, I'm not super concerned with calling it I love because I mean like a dog is also already or like you know, 10/10 age right. We kind of used those already so within the mask laws. I think it's, it shouldn't be all capital they mean because it looks like the.

K

Actual, oh, it's.

U

Interesting old.

K

Like because it's a an abbreviation itself,.

B

Q

B

Know any one David.

Q

Coleman is.

U

Q

B

Things that's listed here is that the C API is called log and which might help for discoverability I mean but googling dialogue. That operation is not on the first page is IBM. Has a product called hi, Logan updates, I hope by log B is on the first page, walk I.

A

Honestly, I'm fine with either name okay I can go within lock, I can go over I log I mean it's I. Log is a little bit unfortunate because the eye on the air look basically indistinguishable.

A

D

C

And then I left a comment that next value in previous value might be named better as increment and decrement. But since that's their effective operation but I didn't know. If that would be confusing to people.

B

I don't understand how math is an enumeration. I mean the description in the comments here, makes it clear what it's going to return it just.

K

It's it's not a whole number increment right. It is the next yeah.

AA

It is plus bit value one, so you know next.

V

Next largest value, previous small value and I feel like I should do it in the moment. Supremo.

C

Increment kind of makes sense in that it is incrementing to the next representable value, just like it does, with increments by a scalar.

B

Increments by the.

C

Next representable value, which that just happens to be one integer yeah.

B

But I think I think when people see increment they're going to say how smart to think like yeah.

W

B

W

And next smallest I know what have a bit increment with decrement.

C

Yeah that sounds reasonable to me.

M

C

What's happening in.

M

The implementation and all these, and where does it adding one bit, yeah yeah, it's just adding one baton, removing one bit what the structure looks like, though they've like a few we're gonna have, for example, is that gonna do the same.

M

If we were to add what like, we have to have the same thing for half or, like the other representations of these floating-point numbers, and they all this bit increments. Well,.

C

All the I Triple E types are.

V

So this never moves between the mantissa and exponent. It's always in one of those places. According to the column, yeah.

C

It's always the next representable value, which is effectively bit plus one for binary, floating-point values. Okay,.

J

Then I have no problem.

J

B

And if you call next value on positive infinity, it does.

C

What it returns positive infinity, because there's only one positive infinity.

AA

We would want the docs to say return to the next largest value until you.

V

Achieve positive.

C

Decimal values don't work that way, but we also don't expose I Triple E decimal types.

F

Okay, any other windows. Besides next we didn't prove this video I look. Do we want to call it something else.

M

M

C

The B stands for base and for binary types. That's.

Q

All right, if nobody feels strong, it will just leave it as I log. Do you want to go out or not, or just basically say that I love this one yeah.

M

Q

Nation I'm looking.

M

C

It's not exactly doing a cast under the hood, it would not call math after a log. It would extract the exponent and directly return that as an integer, which can be more precise when the integer is not exactly representable I.

C

Believe it's zero, but I'd have to check the the I Triple E spec. It lists all the special handling in their Suites.

M

They call log two as it I.

C

Think I log2 would be more familiar to people who have existing code, which is we've kind of gone with the C runtime naming convention for a number of the existing math types.

A

Yeah I mean I'm honestly on the opinion. That's probably the right thing to do because they are so specialized. I mean there's, not something. Somebody opens up the internet since on math and tries to make sense of the operations. But usually it's like you know about a corporation. You look for it. The specific technology right, yeah.

B

So so 10 our job, just to confirm, like I, would expect like log 2 that returns double to return, not a number. If I pass in zero is the input you're saying I, Triple E has defined what I log 2 returns as well.

C

Yeah, it's got exact handling it so there it's got exact specification for both the integer and the floating-point return log and they are different in some cases.

Q

All right, so then it is that.

A

It so then just look up a sign and approved yeah. My cookies.

G

F

Generally, when you both went today, ideas approved.

U

So this is basically just a special.

E

U

P

M

U

A

First time you attend the meeting, you have to attend ten millions in a row inside and I. Consider you.

AB

Now, but this is basically users pass a specialized dictionary for Strings to be able to use grid only span of car a pianist, and basically, that is the discussion that is being ongoing. Here, is either adding a new type of implement section area, string, Diwali or you're studying extension methods for like triad, get volume.

A

So what is the problem? The problem is, we need to be able to look up with a spam, and we can't do that on the generic dictionary, because we makes a little sense to do that for an arbitrary T. It only makes sense when the TF is to be straight. Yes,.

K

Span doesn't implement equality.

K

Like string does.

K

Basically, it says: are you pointing to the exact same memory? Not are you planning to wouldn't.

M

K

Compared with it,.

M

You're básico.

J

Better yeah soon.

G

B

Compare of we don't set out. No, because you can't pass out as a team I mean you, you can't have an ie quality. Compare of renal is mathematical because it can't ever be use this generic or but I mean whether young wait.

K

B

About what's done here is a work around this introduced in a new interface.

B

A

Its certificate is very unfortunate, then I.

AB

Think, overall from the discussion, people agree that this could be better as extension methods, because its pros and cons and were for the generic methods we have to cast the compare to see period that the dictionary was created initially to CPS a nice string, equality, compare and then use the rhythm. All this plan compares working.

A

Person, if only the extension properties that were made indexer so much nicer, because you cannot really do an extension index or today, but I would I mean it's a bit unfortunate. The extension methods are basically having to bypass simple, effective and reach into the internals of the type to actually work at all, but having a new type seems a bit over the top, yes yeah. So.

B

I tried as part of the utf-8 work. I was also experimenting with making string. Compare also implements I equality, compare of read-only memory of Charlie and things just blew up because at that point type inference no longer works, because you know implements all equality, compare of teeth and I only compare review, yeah.

B

I

Yeah, this is messy.

A

This warrants more talk, I mean I, would say, I think the scenario is more than valid. We need to somehow do a broad.

A

Allocating for the probing, but using a brand new type seems a bit over the top for that.

B

Well, the brand new type being the different or the interface type. It's cool well,.

A

The first time I think there's nothing. We can do about it right if you, if you want to have any sort of polymorphism for that kind of thing, you that's what you have to do, the same thing where we have span action and we don't miss, span action right like that, if that one I can live with it to me, is more like the having a new collection type seems a bit heavy-handed for that needs a very special last time. It's a it is a very special last time. It should be a.

Q

Specialized it should probably.

AB

Go to specialized.

Q

Days, business.

AB

Is it part of the power collection suffered or we cannot do it in power collections, because we need to have access to internal? What's our plan for power collections to have to have them, like you know, on a side, oh I think that was yes.

Q

Is to cut them.

O

AB

The side and have them their own implementation and put these guys in series.

T

Very nice, like.

AB

We don't want to duplicate the implementation of the visionary, so.

B

If you were really to not take an arbitrary or inequality to her and if you just took a string, compare directly as a constructor parameter, almost all the api's, have you neither available to you? Yes, so.

AB

That was converted like reducing to disband to string and those two. No, you don't need to convert Spanish.

B

Train there, the api's that we already approved today take a string comparison parameter for doing things like hash code calculation. There are already AP eyes on the read: only span of chart all equals that take the shrinking through the best yeah. That's but again it doesn't. It means that you would take my sorry daughter string can pair of the restrain comparison Benigno.

B

That might be a bit strange, but at least it solves a 99% scenario, but.

AB

Yes, but you would have to take that in the extension methods and the dictionary will still be using the compare that was created right, because here is a dictionary under the covers. Yes yeah, you will have to re-implement. So if you, if you use the indexers api's and the dictionary api's that you're not extending it will use the compare that you create it with or the default compartment.

B

Yes, so you would have to wrap your own compare or which you could do it's just more work. Yep.

X

AB

Here, that's that's! Why we're like stuck with adding and you compare that implements read-only.

B

Oh you've, actually some plastic yeah.

AB

That's the first proposal, and then we came up like if you scroll down. We did like a comparison of like as extension methods, oh and where the pros and cons see in between both of them I mean.

A

Yeah like if you do derive type that is actually not quite as bad.

B

Well, I was dictionary ever intended to be so. No it's not. It doesn't like give you a notification. No.

A

But literally design in to oh because ArrayList and hash table were just we had virtual methods on them, so that and then we said it that's too slow. So we basically selling me to lists and dictionary are basically as cheap as humanly possible, and then you have the you know: the object model types like the collection of T and the the key collection of T that just wrap the other ones. They give you notifications, but but this is the recite a different example of the reason why we haven't sealed them.

A

You may just want to add your message so that it was also pre, pre extension methods and, and so I think it's an it's in line with a lot of you habits you at the time. You know what I like about the extension methods, owes it just lights up on existing infrastructure. Right I mean if you don't control the association of the. If you don't control the type, then you can never look it up with this band. That's the that is still a very significant limitation of the design. As.

AB

A

AB

Work everywhere, yeah and I think I also rather make the the internal field, the fields internal rather than private, protected right.

B

So this implies that you would go to a house and vegetarian compare in the dictionary constructor yeah.

T

B

The extension methods, otherwise you would have to convert the string every single time. Yes,.

AB

That's the only upper cut of extension methods that people would need to know that when they create a dictionary, they need to pass this new specialized comparison and that.

B

Also means, if you create the dictionary with a default constructor which uses equality comparator and default I believe that it won't limit all back through the non-optimized tends to be a stream comparison. Maybe that's.

S

Something that should be specialized well wouldn't even handle this in the in the extension methods, because.

AB

In order to use the equals and get haschke the read only span of the car, we would need to have this compared to be busting into the dictionary you could you could special.

B

Case equality, compare of char dog default and say your ordinal right and that's something you could do it around yeah. That's what I'm thinking I mean it.

A

Is complicating the cover? No, it is a but I think that I think, like honestly, if the extension methods still require that somebody instantiate the dictionary with the right. Compare then I think it's on part of the the device type of holder, because it again only works.

A

If you control the Annunciation, which in many cases you don't or at least control the comparator, which I think new practical terms, probably means you control the association of the object, which would make it much much less useful like if this just works on any dictionary instance and does something sensible, I hope will be a huge value in and then I can even live with the fact that they're slightly ugly right, because the stack while using spats so you're already buying into slightly ugly but man, you don't have to look at a string to look up things.

A

Lectionary, look, that's cool, I mean I, don't know that I think will be a reasonable trade out.

B

So you're leaning toward the extension methods interface.

B

A

I have no idea what it means, an implementation that might be messy as hell but yeah, but it's a core type I mean I. Think that's! Okay to say you know, I want.

AB

To have things more perform, I mean plus extension methods, everybody that has a a generalist rant about. You will just get them instead of coming to introduce.

B

In the attack, how expensive is it to pull out the compare and see if it implements this interface and then do an interface dispatcher.

B

As from within the extension method, well they come there and so.

A

Dictionary.Com.

B

Power is public, and yet- and it's given to you as my equality, compare of string in your case right, so what you would want to do is you need to be able to call good hash code and even the new over loads of good hashcode and equals right. So that means your options are trying seeing if that equality, compare also implements an interface or try casting into string, compare and then do a virtual method. Dispatch is either of those acceptable from a performance standpoint.

AB

Well, we go through the extension methods route we're already not carrying about performance right. Why is.

A

That so because we're.

AB

Not guaranteed that the comparator that we were getting is the one that we that are going to use the performance API so comparison, because we we still have to check what type of comparable yeah.

B

That was true. Yes, so you could end up having an allegation is.

AB

B

With this location,.

AB

Yeah, because we were gonna say, is this: compare up type ice cream equality, comparing it.

B

Has two, and if the area is your own custom comparator rather than using string, compare one of the built-in ones? Yes you'll in current allocation every single time you call anybody. It.

D

B

True, we've got it's over there, you go over the Drivetime economy, no derive tight forces you to pass in unknown equality, yeah.

A

Sure but I would just say you have the same. One of scenarios work right. They would say the scenarios they were performing. They were performing on both sides. Yes, the thing just handles more scenarios that you made a great performance yeah, but.

B

It will not blow up to be fair. How commonly is someone gonna pass in not a string, compare exam, I think.

A

Most people will just I mean the vast majority of people will just use the default. Compare we just tell them. If that hurts. Don't do that, and the thing is that I would say that if the thing doesn't work at all for different comparative will be very sad because that's that's I, think the the vast majority of cases right and then yeah, it's drinkin, Paris or the other ones right. That Gordon, like in our case, I, think, is what we, the most popular ones, will do my guess.

A

So if you handle both of them in a meaningful way, I think you're. Yes, you have an if check hast, but I mean I mean that you would have regardless Agnes unless you are willing to blow their.

A

You know dictionary.

F

So is there more design work than needed know to figure out whether that's doable.

E

G

A

And do people buying my argument that it's better to work on any dictionary incidence that it is to only work on the ones that you control.

B

It makes the type it makes the API is easier to use when you don't have to travel as much code. Yes,.

I

Which I think is.

G

This room is available.

W

Why would we need this guy here at all in the second proposal.

G

Where would you pass this evening.

AB

Because you need that compared to that interface to implement high quality comparison passive to the dictionary URI in IEP is right. Oh because.

A

The stud assumes that you have that guy I see. So if we work on arbitrary ones, then it wouldn't be a problem with the interface that we could just extend string compared with these methods. Here, yeah call it a day.

A

I mean default is effectively automotive, correct.

B

Yep I have not seen how the extension methods would work. To be honest, look maybe it's something: that's called an implementation, but if you're an extension method, the dictionary is going to give your the way that you look up an entry in the dictionary is by saying: do you have this particular key I think he is typed of string that you're given to read on this panel charm? How do you turn that into a string some of the dictionaries public API scan working on? They wouldn't write?

B

Well, if it's an extension method on dictionary, you have to call Terry's public API and there's no public API. That takes a Tiki like Olivia. At least we would have to make some fields internal.

AB

B

G

B

We're changing dictionaries.

G

B

Both in both approaches.

A

Now this has other problems. If you I mean you could do the following right, you could make these internal methods on dictionary right. Then you just fall with them one to one but no you're below a dictionary right, because you know I have what five more method that the JIT has to always get.

B

And it only makes sense of two: you forget.

A

B

A

Is unfortunate, but if you make the internal is public or internal, if you make the privates internal well, then you expose a lot of to all in the school which.

G

Might not be what you want.

A

So I guess these are the things. So let me just summarize what I just said.

B

So University Extension methods would work when you don't control the dictionary, but are there scenarios where you really don't control the dictionary and the hot pad submissive? Are you like I would think that the scenarios that are being talked about here are going to be very custom scenarios where I.

F

Mean that's what I'm saying I mean that that would be my gut feel to this that what I said it's true, but yeah, I've, zero data on that right, I think I.

F

Don't know I've seen plenty of cases where people have dictionaries rather than like you.

A

But they probably expose it as I dictionary, right or as key collection most of the time so we'll be extension, methods.

B

A

They would, if you could don't cast my video. We know what you want.

AB

Also, we get a scoop of read only the back part. We printed it strengths up education, well,.

B

It if you want an implementation on the comparer itself, that also makes sense, in certain cases.

B

Looking in the in the API that we could approved earlier, I put it on the text info they're beyond compare it goes. Our string compare, makes sense to you in certain cases.

AB

Because of us extension method, indexer will take care a strain still.

AB

We potentially be enough like to be missing some data when we call get value of freedom only span of power, because their hash value.

B

Is different, regardless of which implementation you end up, going with I think there's still value, and maybe considering this as a separate proposal, just like you have here.

K

Is there a reason, there's no set value.

K

Or was it just not a complete yeah.

AB

These were just the extension methods. I came up from discussing like the primary methods. Okay,.

G

Both this for now.

I

F

Guess I should lock that.

I

I

Makeshift fixes here.

K

And we did you read the comment all.

N

Those for a canary it's directly with the.

B

We'll, concurrently with is an adoption area right same.

N

Time it takes know, but.

B

If the extension methods works, the interface.

AB

I think I could rather the concurrent extension methods that might lead to nerve extension methods.

AB

It's real place where we need this, not really. It was just a member proposal and for the power collections have worked, but at the end like, if we want to put it in our collections, we would need to replicate all the right.

K

No, what's that one, but.

AB

This would be real good for, like the street internee.

F

So I don't think that I'm, you think that I'm in a place where I'm strongly leaning over the other I think they both have drawbacks like no. No, the solution seems to be things that we, like so I, think they.

A

Have another way to reason: is they both have merit yet? So the question is like how much of the merit outline and you know, overcomes the limitations that it has right. So I would like to hear some other.

F

Opinions from my custom and other people like do, we really think we should.

E

Complicate our collection types for that in that way, because I mean I could totally imagine well here we just add a new.

A

Collection type into the coral collections, library for those people and that's good enough right, it's not in the core. Nobody cares. It's a it's a target to thing. I, don't know it seems.

F

Like this expression of how many wants to accept an enzyme before it feels heavy or complicated and like there's a few.

A

Things here, like you ever, do you have the different interface that you have to add is to implement it on the existing home Paris. You cannot really understand she ate the dictionary ability of the was a default comparative about losing features, or you have to have a very complicated interaction between the extension methods and the dictionary which might cause code bloat at runtime, and you did more things that you need.

A

So the question is like is this: is this whole package worth it right, I think because there seemed to be no designer, as that yep you get the best of all worlds with one version, one of the downsides and that's that's the thing I mean everybody who uses we'd only spend off charge today and uses dictionaries, complaints about the fact that they cannot index with that they may have to create a string for indexing, but I, don't know they all seem to be like the same sort of people that are just playing or we the only spend off char, do a typical.

A

Let me parse a text file and convert scanned of exercises like okay, but how many real-world workloads to be half this problem that aren't already web servers, but they already had their own implementations of.

A

So what's the scenario there scenario.

K

There was ml done. Deadhead today has its own string segment tag called DB tax right.

J

K

Basically wraps a backing string and then, as two things right yeah, we started and we're trying to replace that DB text type with read-only memory today of golf car. And there are those cases where we're putting DB text into dictionaries in the library. And yes,.

K

So I mean that was that wasn't a plane right. That was like we're actively trying to remove a type that solves this problem. Today, that's duplicated with the framework takedown, so I do have.

B

I I'm gonna be meritless, I do have in the UTM a string, branch or inspect the type that is intentionally used to be able it's. It can be used as a tiki inside there right, like that's the entire reason.

K

Which is all like reasons for like solving this is a framework great, because peanut has one people write, their own ml might has their own. The.

A

Team, but this is the, but this is followed by it so like it seems, like I, have the same impression that everybody has their own version of the string segment as soon as you have at this problem disappears because you can just say: yep I, just override get hash code and do whatever I want to do it.

K

Disappears but now you want to be used to of these libraries together well,.

A

That's one but but but you could add it at the fix of this, as we had a string segment, we don't have a charge which, which would be a different problem than this ball right, because this is basically saying if the only thing you have is we only spend off char. How would you change the expiry to do that?

A

But then dictionary is just one of your symptoms, but there may be other symptoms that you still have right, in which case, if you add the string segment time or whatever the case might be from what we know. We actually named this, and this problem disappears, and we don't have to do the avenues here and you do the same thing right. So that's why I'm saying I'm understand.

O

You have a lot of other, are myself.

A

So yeah I agree with that too, which is why I'm saying that to me, there's not like this doesn't feel like it's one of those things you just say. Oh, let's just go for it. It's really.

K

It's what I mean there'd be implicit, conversions.

L

Right, like string to string signal, is important yeah. The problem.

A

Is usually they're only going one way, though, which is like, depending on who's, controlling what you may find yourself in a spa, be again have to allocate because you don't you don't agree on the time.

K

Mr. excitement claim wait. We for read-only memory. We have.

B

Exposed that how to get through the backing string, there's train segment I have also does out, because it's for compatibility with API to take shrinky intent, the s-parameters and it's the.

A

Your problem that you also have existing for my are: where we'd only spend do we have a method phone? We can only spend off char, no, but you can say to I get string. No, because spam doesn't can spend never references.

B

K

D

That's right all right.

D

Both Lia go fetch.

K

The work is decide with one or even like. Are these other competing designs right so.

L

I think the part of the design work is that convincing ourselves that this is the path they want to go down.

L

U

This could lead to like.

AB

Adding extension methods like I, said no.

U

L

Dictionary it just feels like it's the first, it's the key right.

F

Because, then, because it requires.

L

You to have the behavior that it these two things equal right. It is yeah.

A

Yeah one argument is it's. You know you only fix this one. Then you are done right because it's the most important one. The other argument is like once you start doing this, you chase the long trail. You you put the same kind of ugliness on all the collection types at which point it's like. Oh my god do. I want this right.

A

It's only the would have added to a segment.

K

Why didn't we when we had to race agent, all.

Y

F