.Net Foundation Design Reviews, 17 Dec 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Review: GitHub Quick Reviews

Description

00:00:00 - Approved: MemoryMarshal.GetRawArrayData https://github.com/dotnet/corefx/issues/36133#issuecomment-566723424
01:49:09 - Approved: API request : CompressionLevel.Highest https://github.com/dotnet/corefx/issues/36348#issuecomment-566725990

A

All right so so this table, which ironic is not a table, is basically trying to list the different kind of signatures that we have and what rules we have for raise span and memory right, and you said earlier, like it's managed memory versus unmanaged auditory manually, managed memory right, but I don't think that's quite the the same.

A

What this table tries to convey, because even if you ignore span and memory entirely, if you just look at erase, we already have rules for race who basically owns the contents of the array at any given point in time. Right. That is clearly a GC managed object but like if you give somebody an array after the.

B

It's more about this, how much premiership by quantity or not.

C

You know documentation every time you want to use an API that takes an array or memory or span you have to eat. Dogs and I said no. No to me, it should be like anything else. In our framework, which is there are default rules, you don't need to read right for those and then there are exceptions or you know many like level performance-oriented api's and they should have a warning. Sometimes we had the warning to the net method. Name, we say dangerous or sometimes.

D

C

Don't do it, it's just implied it's a very low level, ID I too bad. You kind of need to you know, or we don't expect users to just kind of like hey. Let me just use this API. We don't expect them to use it. So we have API is that we need.

E

To be a bit more aware, what's going on last week, I thought you were saying everything.

F

Should be done, I would say, including the ground.

C

Maybe you miss I, didn't say everything should be in the presentation. What I was saying there will be some api's. Like you know, pipelines you cannot use pipeline to about two eating dogs. Who has a memory when in which I di I mean look at like you.

F

F

Synchronous set method where you're giving something as an input parameter. Like classically, we've always said, we expect that junior developer, collars or just people who figured out the way the things they're already working are going to modify the array after they pass it in and that's probably not what they meant. So most of our things have taken array as a form of a setter. We said.

F

We said if you are saving the reference, as opposed to saving a copy, take the memory instead of as family, so we we have on that first highlight, we have very explicitly said, take a memory if you want to keep the reference instead of a copy, that's where we start deviating between the lot of Rheims for doing and what never you.

E

So we are talking about the synchronous set, not synchronous, nonsense. Synchronous concept says treated the same as I, so it's I like it.

B

Synchronous set, says, memory means.

F

It's to reference: if you have a method called like set buffer and it takes memory, then you are going to keep a reference to this buffer. Really nice set buffer and you take a span. You can't keep it legit.

D

F

You had to make a copy if it was an array. It is our default leader or I. Think.

E

That took already is make a difference.

F

Of copy some majors.

E

Might have been the type system.

F

That conveyed that ownership, so can you give me an example of a synchronous signal, makes a copy over.

E

Aes key not from poverty, not from cryptography namespace, because.

G

Have a stutter and copy.

H

An array of a bastard immutable.

A

B

A

B

Ok, so sock is when you close it it's a buffer. It stores records, doesn't copy for a rare.

G

B

Already makes a copy yeah.

C

B

A guy having trouble even thinking about the.

I

Picture damn of her.

B

Suck it, you think it varies. Yeah.

I

B

Gives it something.

I

Something like sent off, or just like from the method name, it's clear that the API.

F

Intends to take ownership and.

I

F

Case I'm also including properties that are arrayed properties and what we do in the centers of those which again, we've really hit on everything. That is, a property that is returning an array. It almost certainly shouldn't have been a property because, most of them, their gifts, are returning copies, which means they're not extending.

A

To collections for so I couldn't start trust for you.

F

A

I mean constructors very often you do that right. You take an array of data, you call the copies. Yeah I mean that's, that's pretty much the by far the norm.

B

We took the memory, we would be saying we're holding the reference.

F

If we took a span, we would say we're taking the data. Really, that's fine, I guess the memory is well again in a constructor. If you could have taken, if you were making a copy, you should have taken spam. If you're taking a memory, then the type system you were using that to convey I am holding a reference to your left interesting.

F

That's the guidelines that we removed form upper.

E

80S so I mean my mental model. Is.

C

Even discussed in education I think some api's do make a copy, but it's because the backers are small and it like. Nobody was thinking about birth. That I can vote here and like hey. Let's just make a copy, it's probably safer. I would say the main expectation I think the father's is that they're elected they.

J

Are actually not coffee can be a possible.

F

But the question is it like who owns free memory? The question is: is that is the caller allowed to make a modification to the array after they call the method? And if you know, is that or in when they do make it such a modification? Does that modify the state of the object so and I think in most places that we have set data? So maybe it depends on the type of array, but I think anything is a byte array. If it's set data yeah, we say we're taking the data out of so.

B

Now every example set whatever set buffers all this towards a reference doesn't copy anywhere. Okay,.

F

Whatever sockets, what about properties that are arrayed properties with a sense we happen.

B

To be happy, I also.

A

B

From Purdue, where Joe has a lot.

F

Of I'm in terrible.

B

F

Hate them like.

B

F

A key a so symmetric algorithm that.

B

Key symmetric, algorithm.

F

B

A little byte arrays so to me.

C

Said they're like a proper.

B

G

The constructor does the same thing.

C

It basically means somebody will be storing the reference for later all right like to the commit when you connect instructor in person already, if I don't like it's very obvious to the color, that this array will be copied.

C

Will be used for a very long time, the.

K

L

C

Long time, right, yeah.

F

C

I don't understand we won't copy because they're.

B

Easier to copy defensively I know.

C

Berets like do we want this.

B

No, like, like I, don't mean that wafer.

C

Yeah, if I pass an array to the constructor, it's very clear that the object will be using this data or the lifetime of this order. Yep.

M

Well, is it using the.

C

C

It will be using yes same thing with property when you call the property setter and you pass an array. The data will be used, crabapples for the lifetime of this off. No one ever writes in that array. Again, everything is clear, since the color of the constructor of the second pass, the array for the constructor for the property is not the screen, but for the constructor, the caller owns of the array and the.

B

C

The object I don't understand why we wants to code. It's basically.

F

One on their own stores: now they are.

E

Trying to say whether.

F

We want to, or not it's in my experience, the propensity of things based on net make a copy for the facto standard in that MIT is to make a copy, which means everything that doesn't is an exception. I.

D

Think I don't think.

F

We can change the de facto, because now we violate rule number one of the guidelines well,.

C

So I think that it's what you said that it's a always copy and it's a de facto standard I'm, not I'm, questioning the debate, I think for very small payloads that, like represented you know some training or something you know like very small one. Yes, we copy, because it's kind of opportunistic and we like whatever, because not just copy it small, do we have many API is just truly operating on backwards. They like they operate on backers and they copy I will be sure. This is such a stupid thing to do.

C

The buffers can be very, very large, I. Don't so give me examples of things that like then, they accept buffers that are potentially large and they come I would say. That's a.

F

Broken idea, I mean I, can't think we would have a sort of buffer set with anything bigger than like I'm.

C

C

Think coffee for, like things, okay, I mean distracting the list of tea and you know I pass an array we copy, because the expectation is.

C

B

C

Don't need some like buffers and they yeah.

B

So memory stream I might remember stream I'll take like well what Henry's are pointers I.

A

Think it's just separate T bracket bracket from byte bracket bracket, I think that might change this.

C

Maybe that's maybe a good. There are some set of buffers and there are things that are collections. Yeah.

F

C

I, don't think we copy buffers cuz.

F

I mean like right: I was gonna, say I, think we always copy buffers. No.

B

Memory stream caught it doesn't copy, give.

F

It a friend right, but the memory.

B

The thing has the least trying to get a ready to.

F

If you gave me, the array in the constructor I can give it back to the right like it has. A very special dedicated I am transferring it mode. Yeah ray segment also is a very dedicated. I am I, am transporting or I am explicitly wrapping or buffer. So.

C

Maybe this is the the difference and it would satisfy what you know. What prompted me to have this said like we in our Jersey can have lots of things that we.

E

Need to be PI drivers.

C

To copy this thing's completely, don't.

F

Make sense but like again it's cryptography because that's we have like a SN encoded data type raw data, which is a binary property. It's all by the stupid football fan.

B

F

And it closed on get a.

B

F

Weird cones on set, but not yet.

B

F

You access to the property most.

B

Of the page, most of the things in crypto.

F

Most things that return a raise at least byte arrays that I've seen do return copies of the array and.

I

F

I

So they shouldn't different properties, yeah I.

F

I

F

I

Had we had some incidents a long while back in the framework where we had api's have returned properties that could literally just returned the back and filled with the static field, and then the caller could change the values and now cause we're as Cirelli. Okay,.

C

It's only mostly.

I

I never expected the array to be change. Yes,.

C

So we have this guideline and this one I stand behind it. If it's a static thing, yes, you can, you cannot beat any static array factor without code. When we can't decide you got, you don't have actually one that was like it's sensitive. It was like. Oh.

C

B

Half dot; okay, but yeah.

C

Thing is yeah, so that's an extreme case. We definitely cannot eat this. I'm gonna, say yeah. If you, basically, if there is, it should.

E

Be fine, you cannot intervene all.

F

Right we effectively multiple things.

E

F

Are not common right like from properties, we have the guidance that don't return. You mutable privacy right good, so like that's where all these, like, oh the key being used by this, you know a subject. Well, that's part of its private state, so I can't return the field, so I have to make the copy so Jeremy. What do.

C

You think about this. Do you think now me? That's a new realization can rewrite the guidelines that we separate basically collections of things from PI, and then we have rules for buffers that don't force people to copy and you it and.

F

I think that would work X. Then we can I guess mostly what I'm talking about here I'm every array that I'm thinking of is about so when I'm looking at.

E

Different yeah, so I so I, then in this case I'm still I'm, basically be cooking. Byte byte buffers just like again like we.

F

Started talking about buffer.

E

F

X.509 certificate to constructor we take by the way we pass it into Windows Windows read the data, so it's constructor could have been other than Spanish, so it reads the data it stores it in some state. You cannot like it's done. You cannot change that anyway, there's a way that you can get that data back, but it's Windows like we are now giving you our private copy, because you cannot modify this one so that the object that's nicely different I mean.

N

That's we're passing.

G

O

F

Like we could have taken it as a stream or a span and.

F

So I mean like the question on you know: if we take a thing as input in a action that not a setter just like go, do some work, we don't think none of them. We take copies on and we shouldn't have taken the memory in these because and then the question is: when we look at anything that is a set is, do we believe that for a raise they have arrays and then byte buffers?

F

Do we think that the answer is that, by default across the framework that we're already making copies as our fan, behavior, and so that is what it is, you can take? It's.

B

Not it seems like step centers they're, all.

F

F

B

I think are iffy.

F

All right, the instructors are set methods, yeah.

D

F

So now we have satellites and constructors and that there and again I was probably thinking of constructors. They said property and the constructor is the same case which will kept on different cases, but according to David, the despot checking says constructors copy and set methods that they set up. That's.

C

Very strange because I would almost say it's more valuable to copying the set property than me in a constructor. Exactly I passed their a and write it on the object I own, both of them. Why they feel cool I like make a comment, doesn't make any sense whatsoever efficient.

B

So Levi in PDF a string we copy the bytes right. Yes, you have to why.

F

I

Guarantees on well form into something and immutability. If we allow you to keep having a reference to the original buffer, you could mutate it out from under us, say the of.

B

It is immutable if that's not not amusing, to copy. You have two.

I

B

C

I'm fine with that tell.

B

Me, the only ones that one is ever.

C

Sure have anything copy, I would.

F

C

I think it'll.

F

Come down to if we, so we have a guideline that says, if you take spanner memory, you can also take array for people who don't understand what span of memory are, and there are working in variant languages of the right they're, not working.

I

Yeah, it makes.

F

It easier practical and.

I

F

What we have is, what does the other over there hold.

B

On hold on odds positively a year ago, yeah, if.

F

B

A memory shut or unavailable to it. Well, you should have a ready, accepting variants.

F

That take like an all-stars like super low low. Well, no, you should figure out what you're gonna think and but it's it's. You need to make something work for somebody, who's, working in little dotnet and doesn't and if they can't talk about spam. What's.

A

Wrong with that.

I

Yeah, keep in mind, I believe that guideline was talking about EAP. Is he wants to make approachable to a wider developer audience I think if we're talking about something, that's like very low level like something that only a small handful of people will ever use. I. Think Jeremy wasn't our guideline. That's something like standard memory by itself would be appropriate there. If you're yeah.

F

It's if you're super low level API then just deal with whichever a spanning memory you need, but if you're high level API that can ever be expected to be used by a freshman or sophomore college student than you need to think. Oh, how.

B

Can we judge that it's not a good I, think.

A

I think I never a lot of it's, not the high order. It's more about. Do you have an alternative right, because I mean some of our API is that we ever introduced recently are effectively we implementation of API, as we had since v1 I mean had encoding API sensitive one. They all take arrays they're, all not performance. We have new API. Is that don't do that and they only take span and that's fine, because if you only be able to raise, you can still call the old encoding. Api is right same the streams there.

A

Many API has to take streams and that's fine like you, don't have to have all of them, but I think they're. The fundamental problem. There's high level versus low level. You can't have it on an API baby, PI boundary right, because then you have this like. Oh no, we have on a new encoding API, as you have to have array alternatives, which makes no sense, because it's very hard to relative some of the stuff right yeah, but it, but the one mine is for a given scenario.

A

A person that uses the language that cannot deal with spans have to be able to write code. That does that. Now, if that port is not fast, that's fine, but at least you can write the code right like we had one example where we had the new path to join a piace behave differently for how they combine paths, they're, basically different way from you know. We have path, don't combine, which is the old API, and we had a path toward join that has the better behavior that one only took span through your Lee.

A

So we just added it over the only text that also takes drink, so you can just call it, and so like that slices the idea, like sometimes you have an API by API to valancy, and sometimes it's just yep. This tag is not for you use the other take over there, yep that take a slower, but your language sucks. Oh I'm, sorry.

F

Yeah I figured, we said we're just generalizing it to e f-sharp has made some equivalence rules. It can work with this now oops dad. You need to have an array based.

A

F shop added support fast and slow, so it's really the more VP in COBOL and whatever the quit language. Yes,.

F

A new hobby language somebody makes work because.

B

People don't let me think about passing Occident I'm count to APS I. Think I read, which forces like you to make a sub array by copying push the card of coffee in the relevant passage.

I

Which pompously most people are perfectly okay with that, to be honest,.

B

Again, it's not great, but it makes the scenario most people who got data confirms Lee was.

A

B

A

That's partially I mean like how many people do you know they care about to the performance of like Castro, and they don't write in C sharp.

C

Like I will say.

D

C

Was under the impression that this was not always effective, well,.

D

Just remember: I was considered okay.

C

I'll tell you how I do they consider it's basically and it's related to because I think it's for their language. It's like information, evn Cabo needs, like the CLS rule that we have. If you don't have any other alternative, you should provide. A key is that all basically frame of should have a key is that all languages can call so, let's say, don't have encoding the things byte arrays? Yes, we do awesome.

C

Do we care about the navigating and coding being called by all languages? Maybe not that's. How I.

E

Write and that's why we changed.

F

It- and it doesn't say overall.

F

You can work with is erase the scenario should be a comment, pretending you can't say new span, because you can't pretend you can say like pretend you, your language or the universe, you're in says, I, understand the mod wreck for hreflang types in that I understand what my graph means. But I don't understand this one. So I can't talk about this. Then it shouldn't be a case where you know.

E

C

Who uses some CLR language calls us and says: ok, I was trying to use this API to do this thing. I cannot do it. We should always have an answer. Well, maybe not this API there's some other API. You can write interesting and.

A

I mean that's kind of like what we have done. The amended, and the other thing I would say is that you know people that do we might have to farm in code. Ask yourself how much time on a day you're spending I mean you post pictures on Twitter of that application for for right.

A

So, like the the amount of applications that we created, I, just throw away application to just test, something is staggeringly high right and if every single time you I provide 20 lines of code to just call this one method, it takes a span like. Why would you bother- and you don't have to because, like.

B

I think the problem is like these to be.

B

What the difference is like, if you can't do it at all, because your language.

A

Is like short, Fresno's, but I think my point is like it's not just the fact that that takes panel array. It's also the oh there's, a method that returns me an array versus there's a method that takes a span and I have to do the wrong for thinking with the.

P

A

Even if I Super Sonic is not that hard to use still I read file, read all text in my my throw back concepts all the time, because it's a one-liner versus three lines like I mean I, think the if it we talk about productivity, I, think these things also start to matter and like I, think the idea that we can only give people high perform on AP is and will force them to do the right thing and just don't think flat.

B

I think the issues we've taken high-performance means heart use and that's the that is the fundamental flaw with our thinking like it doesn't I use. Well, it laughing is not hurt using it just performed better, but.

A

It's a continuum I like to lick more rule. Do you have right the more complicated it gets right? Leg, dad's fundamental mean the the span API is by and large are harder to use than the older neurons right, and that is something you can't completely abstract away right, I mean some things are just faster because we made it faster. That's that's that's goodness, but by and large I think, if you, if you want to have code that has fewer locations, it's not going to be as straightforward.

D

Transformation.

C

Routines that can transform. You know, change some factors. Well, they have hard to use, no matter whether they use pants or a rice or anything. It's.

D

C

In comparison to read all text and it's like copies all the destination.

F

Writing API! Yes, but.

A

In fairness, like spend, us tend to change the shape I like they're, not they're, not completely as independent as we make them to be either, but.

C

I think it's because since we had spam, we took opportunity to focus on collection, so we created API use that I efficient in battle manipulation, and then they use those. You know we're techniques that make them hard, sometimes yeah, something.

A

C

G

C

If I instantiate.

C

E

Yes, it sometimes makes.

F

E

From a static method, it's like what the hell.

C

I think like there's some things that, like almost no, it makes sense to me I'm thinking. We should have something, but then I read them I try to apply them. I like this doesn't make any sense whatsoever. I just instantiate. This object, I passed my right. The guidelines tell me that I have to keep copy today. That doesn't make any sense to.

A

Be fair, the captain doesn't said they get under says. That's an exception. Has a document that carefully right, I think I think also there's like more every API security code.

C

You have to document the cases and he said no, we don't want people to be reading documentation all the time. Well,.

A

I think that I think what he was saying is that we need to have a set of default rules that apply often enough that you don't have to read documentation.

A

That's that I think that's it I think that's a that's a separate issue, but I think the the the matter concern was right, like we do have a set of like defaults that give them an API shape. We can reason about those or what happens if it comes I think that you all agree with that now. The question is: what are those rules but I think the other thing we have to keep in mind is that the API shape and the names also matter right. If you have an API call the right segment.

A

Well, of course, the properties that aren't going to copy the array that will be done right. The whole point of this type is to hold on you're right right. So, let's elected something like okay. What expectation does a person have using that API right, memoir stream, maybe on the fringe, but again same thing, you logically right into something in memory, so it kind of makes sense that you don't expect it to copy the array right, but I think they're boobs cases well.

A

C

That this is the.

F

F

Take some byte array of the data and then closings array, and then it has a public method called set input that similarly copies the buffer they're, both divided array and many code either right. That was the first one.

F

The collection, which was an encoding array and.

E

The byte array being the buffer and with it can be seen like when there are charts of XML parsers, which ones I.

F

Had to switch back to framework following the chain of method calls for this thing in core eventually ended up with be NSE, so this is obviously something we broke up with. This is the first example that I found so I'm just like we. That is an example of a non crypto thing that took a buffer that made a copy.

A

What you smash your phone, the table that stop stirring and.

F

It's not because there's the bottom on top okay.

A

That's one way to do it, don't think that's what.

F

Do I have in mind but with which direction.

Q

B

Arrays yoni is the only word case where you can copy or not, based on.

C

It doesn't seem like we have a lot of consistency right.

E

Again, that's that's! For.

B

Arrays I think from for memory not.

E

Feels like you.

B

Don't papi ever it seems like that's, not what the government says, but what's it when you come here, wait what? Yes, it don't hug memory right.

F

Yeah we say: if we're taking a memory, we don't make a copy of ever. That's that's what memory same.

R

For RS I want I want.

F

Your buffer, because, aside from AC method, it means I, want your mother. If it's a if it's a if you could have taken a span or a memory span as I, want your data memory as I want your bi-stable.

E

Doesn't have everything so you missed. Maybe what.

C

Prompted this table, but now because we don't copy, we have to copy every time we return remember.

C

That was nothing in.

F

The same object: well, this is what does a property, an array returning property good. So what.

A

B

The arrays weather event properties, yeah memory.

C

Is the key? We have lots of things we're like a bathroom with some payload arrives. That's a you know, event house yep, you get a new event. Oh yes, we have a property. What is the payload? Oh.

D

He learned many others day.

C

So the payload we wanted to return. We don't remember this time. It's probably memory me and Jeremy says no. You cannot do it like.

D

C

Just like I've just got it from data. That's. Why would I make here's another copy I already copied that five times before it even got to the user is.

B

That against the rules, it's.

F

Pretty to remember, that's read-only I mean so. The real question is because I think what prompted all that was: Matt Ellis, writing an event handler that received one of these events, and this question was: if I save this event args after the event, fires can I reliably be disability.

F

Welcome to who owns the buffer I'll, give it a rest after they're.

D

Baked cuttle magnetometers here it.

F

Has that, if that, if that is a thin projection over so variable or any other recycled memory, it, you can't make sense of what it meant to really.

E

Propose no actually I said.

C

E

C

C

Property returns, read-only memory, the lifetime.

E

Is tied to the object that has this property? That's reason.

F

So then you would, if you're returning the read-only memory, you've already made a copy of the data or it's a.

E

Non recycled, I, don't care all basic.

C

Needs for every mutable I know the you know like you can do whatever you want with it. This object please. Okay,.

A

But like how does that help your case because event, arc snow, Morales obsession the event, arts, yeah.

C

Then he would be.

A

C

The rule would say: I now have a reference to event arcs. Nobody prevented me from storing a reference for later. Therefore, I know that the memory is valid because it has the same lifetime guarantees.

A

Okay, that's the assumption that the person is making, but is that actually valid for your case, because I'm pretty sure that the event axe is not the thing that is disposable and is port? So how do.

B

You that's fine, it's fine. So the question is.

B

C

You in this other case, then you in the Dangerous Ideas. If we ever need those who we will do and then they are documented right. Okay, they use a memory.

D

The rules are very complicated. You don't.

C

Have to read documentation if it's very clear from the lab, like the level of the API, that the Jaypee one of those you don't have to name it will be, and if it's like, you can imagine that some people would get confused like they are using MEDLINE scenario, ap ice and suddenly run into them. You call them dangerous or something like this. Basically, somehow we try to communicate to the developer that they need to read or start this night yeah.

F

The example that I gave thread is, if this event arc has a has a private field of we have finished the event callback. This is in a dead state now and they I'm called the memory that it throws. Yeah of like the process.

D

You could have already saved out the.

F

Memory, and so so with memory you actually just like you can't have a like implicit. Where did this go because, as it starts flowing back, you've lost the dangerous you've lost that, like you, you have created C memory problems in net, which we try to avoid. That's pretty fine! So really it's. If you return something as a memory, you can't fool the memory you can't, but like it's gone.

C

Guidelines and usage of the dangerous guys don't take memory out from dangerous api's and keep passing it. The passing it to AP is that may not know that this memory is show.

A

By like it like percentage-wise the same as the horse right because, like 99% guidances, don't do that? Yes,.

D

A

Like if you're on this case dangerous case, sure that, but that's true for all bad lady.

A

Specific case of like you, cannot return every the only memory and pool memory at the same time like that got and still holds for hundreds of cases like.

C

A

But that's true for like evening either that says, do or do not right, like there's always exceptions right like so now. The question becomes like in your case, though, like literally allocating.

F

Again for span and memory.

B

You know if forced no, no, so first of all you could have happened, you could have used every segment all right to avoid the and then you expose the entire array anyway. Yeah.

D

B

The actual full array by default- and you can you- can use us an infinite paths- the default, but like.

I

Developers just doing their natural thing call any of the unsafe or memory marshal a key eyes like that's, not a road. We want to start going down.

B

These you don't have to, and they only really have to if we don't suppose it bears that T memory and span, that's where you have to like use APA no.

I

They don't they don't have to ever like we should. We we shouldn't say developers have to call these api's to fulfill certain scenarios like we, we should say developers never have to call these api's ever and right now they never have to. They can literally call talk to array. X is.

B

I

B

Really whoa, oh I, see you me so I say so force a copy. Okay,.

I

Yeah, that's to a coffee, yeah master.

B

S

I would return.

R

B

Instead and then to array, they want to get a copy, yeah.

F

All right, so what were said so that, if you have so, if you are an instance, property returning, a read-only memory, mm-hmm that that is fine as long as the buffer is tied to the lifetime of the object that the property is defined on that'll, be the state on that. That's it so.

B

If there's something socially this in practice, it has to be in a right, yeah.

I

I mean you can no finalizes working I. Believe you like I,.

C

C

Basically, now say: okay, this is one of the something and say you know more and say yeah. Yes,.

D

C

Make it disposable and you can say the contract, is the memory started to do disposable.

I

How do you feel that everybody can we don't? We don't want to expose those types of API to developers by default unless, in.

B

That interesting so again, the problem is that.

F

If you're looking at any particular method, you can't tell it once we start doing that. You cannot tell looking at this method if it's safe, to read the memory because of each read the memory, they call a helper method, passing argument in it. You don't know if that disposed. The often you now can't reason about the memory, because you've left the state escaped a little better, so no code reviewer can ever review this for accurate. They need too much. They need the entire universe of data. Yes, so I think we want to avoid that.

F

So, like the answer is you've.

D

E

F

Lives with the other, it's not disposable, it can't go back.

C

T

C

In the memory and I'm sexy, a Teta object then gave me memory well, I know they like them about their look their times. Don't pass the object to anybody else. I know now: they're like yeah.

F

Marie, if there are times you can prove it, but in the general.

C

D

Yes, sir, in the general.

F

Case you can't say it can be proven. Therefore, it's bad when we're talking about memory, management.

B

It needs to be obvious. It's right, I mean having a network the same problem, yes, and if it was because a different trip that had this fellow something that's almost in the president, using.

F

B

Know it's super hard to notice, but yeah like if you, if you want to be disposable, you have to be spent action.

F

You have to make it captured context. You can look at the data, you have left the context, no.

B

I took a copy if you want I, think it's fair to say that they're on today, yeah he's like what we're saying but I'm. There should be super ears off okay, yeah, I,.

I

Would also be fine with that, but again, like my, my requirement would be that the developer has to opt in to the unsafe, 'no sand. They have to know that they're off even too unsafe nose and then.

B

There has to be.

I

A safe alternative that there let you by.

B

Default and the issues mostly object. Reuse. This puzzle is not that bad. If you dispose.

I

Doctor for you, isn't it so no keep in mind that memory is an obstruction right memory can point to random native memory as well, so you could actually just end up a be in the process. I mean it. Can video can't.

C

Really you know like that: yeah I would say: memory should not be backed by native memory unless.

B

You are, you are design failure.

D

C

Are basically in this unsafe space? You are a district documentary API, oh yeah, every.

I

C

Forget, like we've.

D

I

You can't tell the developer the consumer of the memory of T hey by the way you can always make the assumption that it's backed by an array or some other GC thing like you. You can't say that if what you're giving them is an obstruction right, it was intended to support all these different types of snare. I.

B

Didn't we do a bunch of work and lay before we they make it super hard to do that? No, no, no, we didn't know did right was all shot down. No, you, okay,.

C

If you, if you don't know anything about the memories that you got right, it's very unsafe and you don't know I would say the guidance in my opinion should be. When you expose memory, it should be safe memory. It should not be back, but by native memory and therefore the rules that we gonna write out, write down the Graduate.

C

E

You do not return memory unless you know it's backed by a right.

C

E

C

Very simple: guy back: there are some cases that it guys we want to use.

C

During this bucket of very crispy, you have to document that this API is super super dangerous. Now now we're talking about super.

G

Believe I find it in of those dangerous saying you can't have a finalizer, that's the only rule by piranha. We.

B

Made it so before memory in the very beginning, I had a voiced, our constructor, that's sucking a pointer. We got rid of it completely right, I'll be I, did the oh, no I.

I

Don't think that was over the case.

B

R

I

R

You can answer him designated, but we're not shipping tensor stable yet, but that was one of the principles behind tensors yeah memory. Degree semantics over overtop of a native pointer.

I

It's intended to be able to allow this like it's, it's not necessarily a good idea, but it is intended to allow knowledge.

B

In the beginning that we literally had a white star like pointer, and let me remove that made it really hard to do by default.

H

B

The owner I mean you can still do it, nothing you can. But it's way less obvious. Don't worry, I'm way hurt it actually exposed I mean.

I

About the education, oh, my gosh Haitian can have whatever you know. The actual type can have whatever implementation of once as a consumer, because all I see is the abstraction. I can't make any assumptions about how the implementation works. I.

B

Understand that, as you talk.

I

A different obstruction, I get it.

C

Yeah at the moment use this very advanced technique with memory, and you return it to people. You are the newest api's. They need to know that when.

D

I see my memory from here. They.

C

Are in dangerous state, they shouldn't be passing it to AP. Is that.

C

B

C

B

Bamboo even now the unsafe, but the owner has to understand that like but I think you were saying it's super rare, that's using our four phase guiding.

F

Sets do not it's very advanced it's if you're again, if someone was doing this in the raw data API review and get some hotel, that's what they were doing. We would have like a ten minute discussion of why. Why do you think that's the right answer.

C

F

Basically, you're saying if somebody.

C

Comes to the nuances and find a way, it's not good to be 90 eyes, but remember it. We will basically.

C

B

Would say you wouldn't do it in a fair, but it's fine doing it in a closed system where you own the interaction.

F

B

Fine in a closed system, no something's.

F

Public is oh, no, okay, well close.

B

As in the app owner also also almost all the pieces of the thought use the entire thing, if you know the legs on your and you only entire thing, that's your you.

C

Mentioned that I developed a library where I have video card buffer, that's memory, anybody can write to it. It just changes, you know pixels on the screen and it's a static memory. It never goes away like you start your up and we allocate the bathroom now you can change pixels, totally fine yeah.

H

We've got if we want to implement Rio it likes to walk.

C

C

B

That'll be fairly first case, we've kind of really can't allocate memory, you, okay, might remain, you pin it forever and then that's dustier in a pointer you're still like my memory, so that I mean.

F

If you're pending it, if you call pin- and you have no call to unpin.

P

Accelerated never ever ever.

F

Ever ever go away, that's the same as an array that the problem is when that, if the address can be used for different semantic purpose later, it will reuse right, never use yeah, so that's using them or if you're, calling OS free or you like, Alec and right now,.

B

You're in the middle of some other object.

F

B

Thing isn't just a new memory, it is literally it's reused or things going away because of other things in the same system all right now. This was.

I

Actually- and this was a question that actually just came up in chat- something that I don't think our customer base really understands all that much like Stan of T and memory of T literally can be backed by effectively a raw pointer like once the underlying object gets disposed. The pointer is still a pointer like if you try to use it now, things will happen because we don't really have the ability to track that access and say.

F

I

Throw a nicer exception, rather than rush the process. That's way up, your state I get.

F

Very iffy about having span as a return on something because it's the same argument of you capture the span here you want this. The span factory leave your sight for a minute. Now you have no idea if that span is safe or not, and and if we tried writing an analyzer for that, it's going to be false, positive after false positive. Thank you. It's even the same issue with the rare pool sort.

M

F

And the guidance is if you return an array, it's not cool, if you, if you give it in a callback and it's obvious that it should not be called then like you, can try your luck.

C

Probably should not be sent to that race. Yeah.

F

A wreck, full duration, really don't be vocal implementation. They never leave you. They never leave your library.

D

B

Your imagine you read from a stream parse the payload, but you ready from a pool before positive stream, because you want to reuse memory, pool we don't care so.

C

There are some api, so did girls back to. Yes, we could have opted yeah.

D

Who's, this memory.

C

Yeah then we would have to document. We would be basting down same pattern, a packet we would have to document the api's people who use german have to be aware of many things what's going on. We just don't want to do it, so we don't pull this memory, but we want to slice it. We want to communicate that it's read-only that there's no reason to modify it and so on its own, but.

A

Would say it also depends on the f model right exactly.

C

Yeah in David's.

A

World I think it is fair to say that, given the way the website is built or given, given the way about pipeline works, you would have to do really unnatural acts of our let that lifetime, like you I, mean at that point, I would say. Well, if you stash in a random object into a static variable X. Is it formatted from some other thread that runs simultaneously yeah, okay, but that? But at that.

S

Point you can't round.

R

A

B

Think at that point of view doesn't requesting you I think that that's really.

I

Common enough, an ASP not full framework where we have the offenses against stuff yeah by.

C

The way all disposable objects he doesn't train was you kind of, cannot share them. If you, if your barn for safety is something will go wrong with the program because not everybody knows I cannot review. Is that you were you giving this argument? Oh I'm, reviewing the code. I cannot tell whether it's correct or not, therefore, I rejected. Well, that's true for all disposable objects that are shared yep.

R

I

Necessarily true, but the bar that we've been holding ourselves to is, if you misuse an instance API on a type that you could end up with corruption on that pipe or on that particular instance. Okay, whatever we haven't really wanted to open up the can of worms, which is we misuse this, for instance, of the type.

C

When you said corruption, you mean the case when the memory school- yes, yeah memory,.

I

For if it tips, people yeah memory for.

C

Yeah, no, no, then, when I already said many times, I totally the api's the moment you return to memory to users, you have pipelines.

C

You know it's not it's not full memory, it's just I shared it between the who knows. When can you know when can I read? When can I intake this memory, but it's not globally shirt. It's just shirt between you know, I pass it around is.

F

A probable I, don't I, don't think that right now we have disposable things that return buffers Rob. If they are disposable and they're returning a buffer. It was it happy yeah.

P

The first time you accessed it or whatever.

F

They made it like, maybe it's a one time defensive copy and then the property returns the same thing over and over again, but that they don't give you the thing that actually got disposed. It's always on.

I

F

And even obviously,.

I

Like over the past year or so I've become, you know a little less I, don't know militant in my position here.

D

I

It doesn't make sense for certain scenarios, as you pointed out, for people to be able to do the unsafe, high performance things. But what we've seen, unfortunately, especially with regard to span and memory, is that people naturally gravitate towards those unsafe things, because they don't realize that they're unsafe.

F

I

R

I

The problem you know of allowing allowing developers to opt into these unsafe, high performance things, but having them explicitly acknowledge that they're unsafe that they know what they're doing. That seems to be the best balance, I think. But how do you native issues, but.

A

How do you do this in a fair manner? Guess Pinet, where the framework is supposed to do the unsafe stuff, not the user, but then.

I

The access to like, if some of them are particularly annotated with like hey this, this represents cool data, for instance, if you misuse this, you can end up corrupting cross request state stuff like that.

D

I

Know how to do that we view we tend to use things like API names, checking things out, yeah.

B

C

I, the way I would do it is the span is the way to expose in touch memory to users because span has very clear lifetime and therefore each.

C

Operates on memory to users, unless you communicate this super excited yes today, but then what to it? Why? You should already know you did it when.

B

C

Expose memory juicers.

B

C

Hdp like Heather's and they have strings, and they can be spans that point to this unsafe memory, but the user can only like get the span and they cannot store a stabilizer. Therefore, you shouldn't be deleting this manner till the frame pops. The.

B

Promise vanish synchronously, you can't you couldn't copy to a blob storage API in a control, so you can do.

C

Something yeah, then it's unsafe and frankly, I, don't know what the API actually is called, but that maybe we just call it unsafe, because it's like one IP I did sunset in a sea of IP is that are meant to be for mainland scenarios and safe. If it's like a feature area, it's unsafe, well, unsafe, yeah, okay,.

A

Let me ask you: if aggression keep happy wrapping this up or do we need to continue the discussion.

C

I wonder whether we can kind of write the guidelines.

C

C

But at the same time we don't have to do things like memories.

F

All the time so I mean the interesting question is all the erase stuff that I wrote down here. That's my guess, based on experience being on the team, what it is when we make copies among them, which is effectively, we always make I. Think.

C

The realization that you in this analysis is that there are a race of items and there are areas of fights and I think what you said is more true for array of byte items, array of bytes yeah. There are some examples where you need copies, but I wonder whether it.

F

Is the only things that I can really think of that operate on bytes other than really crypto soccer? So this is straight networking and it takes it as input doesn't return it up. That's.

C

One we don't have a problem, you can say it says: if it's a backer of bytes, then the rules are don't copy. Correct copying, buffers of bytes literally is like well, then I don't want to have to say B I was speckled. What was the point.

B

He's really big buffers also like whether you ever want to copy a giant buffer and.

F

Then, but like, and that's easy to convey in your choice of taking them like if you're a set and you're taking the memory that says I'm keeping me I'm keeping this buffer, the one that's weird is asynchronous methods. They have to take any way they can to this family, but what they're saying is I'm keeping your buffer until the.

B

Total CPI yeah once.

F

The task exits, the running state. That task cannot talk about it anymore. The owners gonna defer to the top left up, King returns memory. That's.

B

S

F

The the input the input memory is authority is the term to the caller, and that was I think the rule that started this entire thread is well. We passed this thing underneath think callback. Are we allowed to touch the back.

F

Because it was an input to an async amendment, so do I have to release it. But if it's a property of a Naga.

D

It's a property is different again, so the one having this and so.

F

We said if we say it's tied to the lifetime of the object that it lived on, it wasn't an input parameter. It was a property on the thing. Then it then it belongs to the event, arms and if you hold the event herbs, it's still yours. Yes, yes,.

B

So buffer hints, um you saw so a buffer path to it. Isn't nothing goes away earth if you're not it has its finished right, but if it's an arbiter has a buffer, it's fine to stash the object and I think just soon as a submit is the circle markets it's technically. It's.

C

A metal, it's a come back up again whenever.

B

F

C

You owned it usually.

F

Did we gave it to ya, but our guidance, the guidance that Levi wrote, that Matt was looking at and trying to reason about whether you can save this memory is really for if you've passed a memory into that call directly and then the promise on that one is you won't let go of it when the task finishes. You won't be saving that one and that's where I have a synchronous set method. This confuses me and I refuse to have any opinion of what that does.

B

Oh so I guess one different cities. You don't normally pass in a read-only memory right. You know, you know pass that memory, that's pray only! So you don't leave the.

F

Filling if it's like stream dot right, you pass it.

B

Over for me, don't worry because the stream.

D

F

Not going to modify the content, so it uses reloading their rate to convey I will not.

D

Modify your data under.

F

Any circumstances and stream that right should not hold.

D

F

That memory and use it beyond the tasks X ready right Natalie takes a really spammy.

F

B

To be sculpted leg out on summer summer summer, global has some are like four different. Some years of, we.

A

Should have something like a like a check o sword like a borrow check or awesome for.

B

Now he is do not copy. The memory is insane yeah.

A

C

The issue is cancellation, token and timeouts. We often have to tell we have to tell be able to tell the difference. We feel like our customers need to be able to tell the difference whether something can send an operation or it was canceled because of time out. You know today we just throw in or say something about cancer. Now you know tasks cancelled and it was cancelled either because of time out or somebody cancelled. There is a.

D

You can write you.

C

Can write codes to kind of detect it, but it's super, you know, did this dance that I I mean the conservation talk in the back? She passed was cancelled, acknowledged well.

E

Hi Tracy and it's racy.

C

And developed I would want to add to the agenda a discussion about how we improve maybe fix this.

E

Issue this I'm aware of that you can.

F

Be yeah we could change the task engine to throw time out exception.

C

I was discussing today something with father and we were brainstorming. Maybe you can have new exception. Basically, that inherits from.

A

C

F

Timeout, instead of opera, yes,.

C

So then the issue is now: we have to timeout the exceptions, so we were thinking women. You have an interface.

T

Multiple inheritance, mmm what.

A

Tosca titled the other one to turn the other one into a task time except few yeah.

C

We don't have to discuss it today.

C

A

Probably won't see me would be good to have a proposal or like at least the code sample that outlandish I. Don't know why you care.

D

C

We know whether something timed out and you should retry or the user X. You can search anything.

F

C

We don't want to keep you, try it and thank you. Cancer right.

F

Now, for that, is you created the time out half of the cancellation tokens if you combine with their input to.

B

F

We cancel because this cancellation tokens had.

B

That same problem is.

F

Ignore kind of you I'm.

B

Like the clovers, it's disgusting yeah, if check this yeah.

F

So I mean which this was.

B

Reinforced, hopefully,.

F

Derived from operation cancellation further reinforced, they do not catch test, canceled exception, which means maybe Steve will just make that booty. We can take the breaking change and we can make it not public in the next version.

F

Yeah, so we should drive it from upper season task except task. Cancel exception was a mistake which I've never been. Yes, it was intended as an implementation detail with the engine, but the.

H

Problem is how many users are.

F

H

F

H

The exception right now, yeah.

A

That was already.

F

Sorted because some things throw operation, canceled attempted in the context where you would think of it to a task council exception. That's.

D

F

Class and that's the one in the v3 guidance I specifically said: do not catch tax, councilors, actually differently than operation. So.

A

Just to be clear, I've canceled, the the next to review slots because they're basically and the time time everybody's taking time off so the next book.

B

A

I feel sorry for you.

U

Q

V

A

Then that's so. Do you have an issue yeah? That's what I thought, because they walked in the reason I suppose, which one is it.

D

A

For one, why do I not see this visit mark leave here, but if already.

D

Approved is a memory much oh.

A

Yeah, let me hold on hold on hold on dotnet runtime.

D

What did you say, nine, seven, one, nine four one: nine.

V

D

A

Right, let me throw this over so guys. I will kick you out of this meeting and drawing the other one just for consistency, reasons Oh.

H

A

Very confusing I mean it's already confusing, but might as well just do the right thing.

A

A

It's the it's funny all right, I like it, some echo echo, is always good all right. So we approve this. Why are you here? Yeah.

D

This ACL changes a bit different to the other ones, because we don't have enough a public API surface area to to resolve the problem. So the only way we could figure out to make the correct calls. What the main issue here is that there we're involving two different assemblies and defining the new classes in sister Maya file system access control, but calling it from system are your memory, mapped files.

D

So because they are in different assemblies, I am unable to call a constructor that is private and find in memory. Matt Farah, so I want to make it public and you can go to comments stuff. So you can see the one I'm talking about go to one more right.

W

D

It exists. It's just private and I left an explanation of why this problem being arise in the previous ratio changes that we made. This is the only one that special in this.

H

So what scenario is this enabling.

F

um Are we looking or if you were copying it as close object, it's even worse than others,.

D

F

When it's private, we can do whatever we want to the implementation. If you make it public, you have to define the behavior okay.

I

F

Right now, it's implemented as certain tables not null. So now it's going to do a check or argument no attached and it's asserting it's not closed. It's asserting it's not valid, but but you have to decide now if it becomes public. Is this going to internally call duplicate handle, make its own instance of the same memory back to 500 and the caller can call this place without the invalidating this object?

F

So you you you're, leaking the handle in a way that we never did before I think, or rather the caller can influence the handle vehicle object. So you have to define that to be I. Think.

A

What you want, though, is you would probably say: I want to transfer the ownership to that object. Because that's the object you pass around afterward, that's opera Jay gets disposed, so you don't want to leak. The handle.

H

F

Our stream already has.

A

F

Thing, that's taking us a handle and yeah I.

A

Mean I mean three duplication sounds bad because I really key in the handle who owns the other handle you just duplicated, isn't.

D

This related to what Levi answered in a no-stress, he said we should make sure to add the option of linking it open. So.

I

The leave out the parameter is for the file stream, not for the final handle.

R

I

F

I

Send it by email is we we need to either documents what the semantics should be, as is Jeremy get called out, or we need to add whatever necessary parameter can make that happen right. So it should be effectively.

F

It's it should be consistent with something and if, if file straining Marty has taking a safe file handle, we should do the same thing and if that's different than storing the handle, then now you need. You probably want to account for the semantic changes of whatever is calling this now, whether it's you have another private constructor that has private memory, mapped file safe to handle comma bool. Am I inside this assembly I, don't care what this value is, but you've called a private constructor now, so we're happy like that.

F

You just you can't just flip this from private to public. There are consequences, and you have to you- have to followed it as.

D

Far as the API.

F

Is concerned, it's just. You need the implementation to be consistent with other things that they say examples so.

H

It looks like five stream will dispose your handle so.

F

H

Takes ownership it takes.

F

Uses are no option use and not.

H

Take ownership and it doesn't increment the reference count. It assumes that it's taking that reference that you pass it already right.

F

So so it's doing ownership transference, yep, okay, so that kind of which.

D

Is what that methods.

F

Doing right now, so you just need to change. The debugger starts arguing, okay, good.

D

So that's one API and the other two are classes that weren't defined in the main description and I'm mentioning them in the next comment.

D

All the security changes that I've made before already have classes like this design and they allow modifying the security properties. Are you defining it exactly as it's finding about the framework? That's the thing they did not exist in framework, so we have I, have I, am defining the security of it for memory map file for the first time, and the class needs to have two properties defined.

F

So this is, this is completely new stuff for a completely new feature. This isn't yeah. Okay, that's the.

D

Thing that means.

F

We actually have to look at uh-huh well, I come wait for that. One I mean.

D

The number one question is.

F

Can we use an existing access? Rule type? Has a memory map file access was different than a file access rule I.

D

Don't know what the consequences of doing that would be. Okay, then the ones approved.

F

And this is not okay, so the first is: can you avoid defining a new type without having any weirdness? If not, then sorry.

V

Like why do you need new types so.

D

Whenever we create a security object, we are, we can modify the access, control and the right via these rules. We can also retrieve them from the security object and get a list of access rules and outlet rules right and since I'm, defining the security class for the first time for memory, mapped files well, I needed access. Rules are now the truth.

F

Right, so it's really just it like. If we, let's start with, can we use file access, rule and file on it rule and okay there's if they are somehow incompatible in a way that doesn't make sense for just an argument exception of like you can't create an access rule, remember, wrap file that specifies synchronizer.

F

Then then we would need them, but like cuz, I guess you're not defining on here at the memory-mapped file rights whatever that is. Is that a type that already.

S

F

S

Last property in.

F

Your type device well.

V

Like this guy here right, one.

D

F

Anything that's together that one exists, I, think yeah all right! That's.

D

F

We have the it needs to be different because it needs to take a different template type thing type exams than that. That's a B, so.

A

One FM, it doesn't have these type, so how does it work and on airfare work if the if these types don't exist, the.

D

Classes were different, they were ending, their security classes were different. The structure which.

A

One, the XS will be ordered or.

D

Another memory mapped file security class.

A

We didn't want to take those because.

D

A

I mean like be exposing the echo concept. The dotnet framework has like the only reason we don't put them on directly on memory. Mapped file is because we want to keep a close on top of memory mapped file. So what I'm asking is like what? What in our architecture requires us to define new ties? Why can't we just take whatever the shape? Dotnet framework has and basically just take it, as is that's the part I don't understand, because for the most part we try to basically take the occurred concept.

A

On.Net framework just planted on core in the only cases where we haven't done this or tried to avoid. That is if we, if we have crazy dependencies, because right now the echoed sits nicely on top and the windows compact pack, and we don't want to put this in the core.

A

So originally, we also talked about whether we can just expose the Constructors rather than the extension methods right in the answer was what we could, but that will bring in like six new assemblies to to to call it, which we didn't want to do so like that. That's why I'm not sure why we need new types. If the conv already exists and got infirm or whatever api shape, they have, it would just keep yeah.

F

Framework has memory mapped file security unless the reason so so, if you need a new type in court, but that type is we're copying from framework because we're copying a framework content.

I

F

Making it just work on core in the way that's.

I

Fine said we were copying.

F

This type exactly is this framework who agency sure, and we might even look at what, except to see if you were to be.

X

F

If so, the question is, why aren't you using memory-mapped all security.

D

F

I not, why are you not eating? Why are you not just porting the type from dotnet framework called memory, mapped file, secure well,.

D

I did it just it's in there PR and what I was strolling. The email thread is that, when we're reporting, we don't need a make. K approval right.

A

Like I mean, if you port, an existing API is, is the only thing we really care about. Is the layering and thank you like you, don't want to add this thing to call up and then transitively pulling 12 new zombies right that something we don't we we try to avoid right, but as far as the API ship goes, the reason we add these API is from don't and framework.

A

So we get compact moved on if I'm like when you're bringing an existing code right so changing the API shape is well at that point like what's the point, if you want to be compatible, you just have to take it, as is anyway, so there's no.

D

Point in the video, but.

A

D

Could you buy you the finally, thanks to again the what important for memory mapped file security and because I have to make sure that it it's compatible with framework right so having these main effect that.

F

New feature is not the same as yeah.

D

I think we already approved.

F

Was porting the concept from core framework to completely new item which should be tracked differently? Yeah.

O

Okay, I mean one difference: I see that this thing is implemented. Access to dinars rules, where's Mary, Mac, define security, implements object, security sure, but if we didn't need it and print it, why do we need it.

V

Yeah that bit yeah.

F

You're not porting code at that yeah. It's a new feature. We should treat it as such, whereas the rest of this issue is about porting. The concepts from.

O

F

Question would be if.

O

You leave newly this new type by word. Yes,.

D

We can discuss it next time, I'll make sure yeah.

Y

O

That I can use a file.

D

Security, sorry, the pilot rules and you suggest- or if we don't need them at all well,.

F

Like is it like what I remember of this thing? As we talked about the new excite, we had to create to move effectively constructors and move the properties to extension properties.

F

And so like that, all got approved on how we're handling the concept port from frame up decor. If these types don't exist in framework work, either adding a new concept or we were missing something in there in that port, and it sounds like we're adding a completely new concept and that that's not this issue. The API approved that this is having that you're requesting an amendment to is just taking the security concept that we had in frameworks moving into for. If this is a new feature, that's different, so a different API proposal, a different. We.

D

F

This we're adding.

A

Well bump the API back to needs work then, because they think that sunny, we should probably talk about Plus, unfortunately, for the notes, I need an N API changed somehow. Otherwise the notes don't pick it up. If you really want to.

D

F

This should be moved to a different issue completely if you want to market as needed, change.

A

That was more information emanated, but that's ok.

A

I'll goddamn it oh boy.

A

Well, let's just say: qualified names are not great when you review a POS.

A

Here we go, that is more reasonable.

L

I

Did everyone else go to the bathroom no.

O

We were reading on you go to finish up. Writing his own yeah.

A

I'm my brain might have gotten to the bathroom, though.

I

Okay, now the charm you.

O

Here to help review and other ideas.

A

Alright, so done so next issue, then otherwise it just go through the list.

A

Nobody has anything all right and all the way back to the beginning.

P

Mr. pitchforks, your apt, oh.

I

Right, so this was basically um what people want to do here is they want to give it an array, get a reference to where the element at index zero would have been in that array. Interesting aspect to this is: if the array is empty, it will return a reference to basically just passing into the array which is totally valid in il and the CLR.

H

I

H

I

Performance; optimisation for people who need to be able to avoid those Chuck's and really hot cut outs.

I

There was some discussion on this and folks like yada and others more intimately familiar with the runtime did come up with scenarios where it would be useful. We already have this as an internal level that we make you so very frequently in our hot code. Hats, but because it's internal means no one else.

K

I

K

I assume that if they're very simply the key would be better to garbage okay correct, you can.

I

Get you can get a reference to it, but you should not be one person, as in the the ref T will point to a meaningful address and you can even compare it against other refuse because it is fully GC trackable. But if you try to dereference in it, you'll get undefined behavior.

F

I

You get is totally fine like you can you can lose the unsafe API for that to a certain extent as well things like on what does it like get? What is it's panicky? Don't get raw get reference whatever it is that extension method. We added that will do the same thing. It will give you a reference for an empty span, the points after the end of the span- and it's totally legal to do that.

I

So this is what this method does is, if a raise of all it more reps like it, doesn't throw an argument all up something I've actually norms, because we turn off all our checking. If array is not null, it returns, a reference to where element zero would have been inside of the aware.

C

It's just all the alternative API skip tracks, so let's not just do okay, I'm gonna for each items based on the length you have to like hey. If I mean the zero K, then I cannot actually so. This is bypassed the index elevators.

I

Know there, so there are three traps that are performed by the runtime. The first is a typical null check, which is still going to be performed regardless. We can't lie to us out. The second is a link check which we now are bypassing, and the third is in an array variance check which we're also bypassing.

I

So, if you have a t array, that's actually of type u array, for instance, and you try getting a ref of the array sub-zero, it will also perform that variance, checking or behalf using this API. That variance track will be skipped.

A

So how is this API different from the existing as ref api's? There.

I

Is no API to do this right now today. Well,.

A

There is one that takes a read-only spin and there's one that takes a spin yeah.

I

But there aren't any to take a race sure.

A

But any array you can convert implicitly to a span or we'd only spend that.

I

Conversion might fail.

I

Yes, that's why so.

A

Which conversion? Which fails the.

I

Array variance check could fail at runtime. So if I have a string array which I hit the downcast to an object array of Cass, County I, don't know which I could cast to an object array and then I tried. Turning that into a span of object, the span of object, constructor will actually throw saying hey. This array is actually of type.

D

I

Can't turn it into a span of object on your behalf. Okay,.

A

Then ask we're.

I

From these API, this API would avoid that check as well. So.

A

Any reason, then, you can't add an overloaded texty array and call it esra I.

I

I, don't necessarily care what the API name is I think there was some debate on whether there should go on unsafe or memory Marshall there in the thread there was some discussion that unsafe doesn't really make sense, since there aren't any other array asked. Api is own unsafe, so memory marshals, probably the best place for I.

A

Mean quite frankly that, given that memory, Marshall dot as Rev on spans that between ref T versus read-only ref teeth, it seems like exactly the same thing just for rates. Well,.

O

The runtime Clippers like, why did we decide to not put it there, but that's what was initially a.

I

Runtime helpers still do perform. The public API is on. There still do perform the sanity checks where, as the proportional API is generally do not, but again, if you, if you scroll down like toward the end, there's discussion on particular names, particular behaviors and the scenarios and enables.

O

What happens if I pass in another a here you'll get a.

I

No reference obstruction, not an argument, no exception and or something in.

O

The medical itself or after I started using the ref within the nautical.

I

Okay, because what the way that this method is implemented is that it tries to dereference the array and then and then add some and add a fixed, offset to the thing that it dereferenced that do you, reference will result in nora. Existing is Ray.

A

Just so that this bytes to T, which is not what we want either so I mean I, don't know like at that point.

A

It should pull, have a different name, at which point it might also be in a different time.

I

This is obviously carried on the same like Jo, don't call it unless you produce actor, though so.

C

Memory a times I, my inversion is: it has API with my new memory. Well,.

A

They have a lot of ApS that don't deal with memory at all.

O

There's one create from pinned area that takes arrays yeah.

C

O

Clears the memory of T? Yes,.

C

Basically, also memory marshal station memory always find somebody's signature yeah. This one is now. So that's that's that's the thing that gives me a bit of a ball sure.

A

But where else will you put it that we don't have a home for unsafe api? Is for race really.

I

Yeah, so it's we have discussed, maybe making.

D

As you type for.

I

This to contain like unsafe or a APRs look whatever there is see here. There was actually discussion on that. Let me see if I can find it I.

A

Mean unsaved Ezra F is pretty close. Yes,.

Y

A

It does pointers to tease, but.

I

Yeah, in fact, I have a comment for March 18th say you know, run time, helpers isn't historically considered unsafe because most of its methods, most of its methods, do argument. Checking but unsafe is.

F

As unsafe as you can get without, you know, taking a pointer.

I

Unsafe as rough actually would break because it would conflict with the inti argument.

A

Yeah, that's a fair point. Well, the existing one takes a 50, so maybe there's.

I

No memory Marshall has rough footwork or memory Marshall die.

F

Sorry unsafe die dead.

V

Roll red, yeah.

E

V

Wallrath get.

L

I

To make this number, this is an API that you're going to have to read, but all memory Marshall would on say tape. You guys are like that as well. So.

O

Like if you name it, get draw a ref and put it on an unsafe class that then we avoid the issue with method overloading right, yeah.

I

If we put it out on site, we'll just have to call Ezra sure yeah.

O

If you're already calling it error, a guitar or a deaf forget, already run unsafe costumes, their face for sure.

X

I

Fine by me can.

X

We I don't really.

I

Have a strong preference system which it goes to your lives on.

A

Can we put refs I'm in the name that is.

V

Getting audited good ball or a ref would welcome you here.

Z

F

Get first ref unchecked.

A

That's all y'all I'm safe, how much more, how much more unsavoury but.

W

It's tell you why you want.

W

Do we have something like yes.

I

There there already is something like this for span. It's unsafe dog Ezra.

V

Q

Son really fat.

H

Ezra, you would have to get.

I

To know yourself, sorry not unsafe, not as right, unsafe, dog getting rough, maybe.

O

Must know get reference every.

I

Martial get reference that memory.

O

Mercial get reference or there that's the one that we even without connect. Why.

I

You know what that's what there should be there should be. There should be an overload of member markl, get where points mm-hmm. That's really what I should be call them, because.

A

I

Behaves exactly what those two maybe you're.

F

Just you're come intercepting the spam conversion. One case you.

O

I

Know is not valid to passing for.

F

Scattered yeah, because if no one gets been ready to expand its converted.

F

A

F

You write if you write like things that has spanned: it's like the only extension that doesn't throw 100.

F

Because we just said, oh call news fan of the input. What about string is no that's a static. Look at the negative.

M

A

It was simple away before extension. With this very thing, it was a two thousand five methods.

C

No, you cannot enforce it. You cannot.

A

Unfortunately, no because the compiler says extension methods can only live on static types.

C

A

Luck: I. Try it many times.

C

A

You know why they don't do it. I know. It's proof.

A

Well, somebody look up as a father man that gets that's what the compiler spends most of its time after reading characters.

O

So Levi, you know how there's this related issue or not to say add era yeah. How are they related.

I

So this is specifically for rice, because for a raise being that there are contiguous sequences of elements, it's meaningful to ask things about like how are these two addresses related to each other for regular objects? It's not really meaningful to ask those questions occasionally, if you're getting the raw data for an object. It's probably because you wanted to reinterpret trust. That object is something else, and at that point we really should provide specific api's for what you're trying to do.

A

Okay, so I mean to me that should be a memory. Marsha get reference, no.

V

I

It just is, it has.

V

I

Different behavior than the other other ones.

A

So we go back to unsaved, get water.

F

Because memory Marshall get reference with an array and it's implicitly getting converted to a span, and that will let me null array turn into the empty span, and it will give you reps zero.

F

This one will throw in a reference exception. That's.

O

Right so for passing another.

F

O

Dip some IDs are different between those Libyans well.

A

But you would have to pass in a span of buy it and then detention, ative of a tea.

H

F

To span right now- and you won't have an exception- attorney- get references just oh.

H

F

Exception or index exception, bypassing ref things of zero yeah.

I

That's really all it is which is was like chips and it bypasses various trucks. It still doesn't know. Chuck and.

F

Then so moving this as an overload means it will it'll bind tighter to an array or a very.

D

Another array and.

F

The null array will now go from having previously worked to throw an exception. Does.

A

It even come out today. Yes,.

F

Because all because already is implicitly ghost fans and the implicit conversion operator for array takes the null array to the span.

A

So we called the one that takes spend on do anything on this bed right. We don't have Envisat conversion from the rates who would only spend.

F

A

If you have Bolden, I won't compile correct so then, today.

F

You can't write the code plus an arena span, which is part o. We have both kinds of spans I'm.

A

Saying if you have the say, I, don't think it composite. If a raise both this is convertible to we'd only span and span, then you wouldn't be able to call this API unless you do credit or nestings.

H

We have a equivalent version for Strings. Yes,.

I

String dog compatible reference.

F

Okay, your emo, your your your right since we already both have both get reference of read-only span of T and get references span of T. An array cannot be passed implicitly. So if you're doing it, you were already getting.

H

F

Recalling the add yourself or using a cast or whatever so so, we can add all.

M

I

Calling strings on compatible references always safe, even on an empty string, because in the worst case, it's just going to give you a retinal terminator, which you can dereference no problem. So in this particular case, you'll get a reference to God knows why.

O

I forget other reasons not to put it.

F

In wash up, but in this case we're just overworked where we are doing a thing that is just a specialized. Oh.

F

Oh, that's gross not more statistically no.

I

Memory marshal give reference sounds fine to me. Yeah.

O

Let's go get there first, another we're.

I

Just overloading.

O

These two minutes.

A

Dad we don't have any of those.

M

Oh, you do have them a lot. It's.

A

Just that, for the 99% time.

F

Where we had rated spanning in core to that, the array ones will throw argument no exception with this bad ones.

F

A

Right so we have five minutes or do you want to talk to that.

A

Eric you're still on hold.

M

A

It says on hold in the AAP I, don't know what that means. Maybe that means he's not actually in the call all right rejected by.

A

Anybody all has content, otherwise you can just look at the next one.

I

V

Could still consider it even without.

V

Discussion yeah: it only has five comments. I think.

H

V

H

Useful, if you were packaging, something ahead of time, that would be downloaded by many people.

H

F

Emo, when you're you know converting raw to ping, do you say smallest or fastest I do smallest.

C

Y

Best compression.

S

I

Think most people honestly would call it best rather than highest because like if you ask a hundred developers, what does it mean to be the best compression scheme? Ninety-Nine and then we're going to say results in the smallest product.

C

P

A

N

Has time trade-offs as well? It says I.

X

P

N

P

F

Again best is best seems like it's competing with optimal for, like it's, the best choice, yeah.

Q

F

You know declaratively fastest as fast as no compression is nope Russian and compression level. Smallest output is like yes,.

Y

Keep it going until you find the smallest.

T

Smallest size, yeah yeah, we.

A

Really call the smaller size, not just smallest, well,.

F

What's the smallest compression level, but it is if, like JPEGs, compress zero to two nine smallest review zero, which is not compressed. How.

A

Does he want highest I think highest.

F

Makes no sense tonight I.

A

F

Why is the smallest output is unambiguous or smallest sizes on ambiguous? It's it's clear that you're not related to any particular respective did I count, I, think iesson.

E

And smallest is if a contiguous.

C

Because people who know that these eighty ice underneath actually take is we think that it has something to do with. You know as Matthew's to see them value to the instance. We probably will well except the highest, may not be the highest end. Right like we should express the intent the highest might be actually the lowest in.

F

E

F

Well, for like JPEG compression and zip compression, bigger number is spend more time on it, but like.

D

F

Other algorithm could come along, then we use compression level for that. Zero is ICD as small as you can and five is like trial. Try.

O

A little bit yeah I was thinking.

O

D

We approve a name or discuss a name for a new API. Do we keep in mind that maybe people won't look at the intelligence, the documentation? We assume people? Never? Okay.

D

A

Don't have to assume a QB makes me know we have data, don't do that in fact, I will do this in an hour from now. Well, I am participating in the WinForms API use rate to study, so I will not be the new dogs. I will just sum up my way to intellisense and then maybe something will happen we'll find out.

C

A

I, don't know why they? Why didn't.

C

A

This was just little just added a new API that they're testing.

Q

All right so then,.

D

Q

A

And what is that? A lot of the compression there? One right, yeah.

F

But I recommend the smallest size or smallest output for some things. Motor size high is my vote. Highest sounds like it can be misconstrued depending on an algorithm I.

O

Think in the case of broccoli, optimal is the same as highest. Our goal.

F

Is whatever the implementation.

O

Wants to do with it yeah optimal.

F

Is you choose optimal, is I, have no opinion here. Optimal is do something good for that's all. Yeah do something good, so.

O

I, don't I don't want to have an opinion for context for party it says optimal is as the absolute best profession without regard to cooperation. Time.

F

M

I think optimal.

F

Is something that you'll have to look at her things like deflate it? Maybe halfway in the spectrum, I think basically, maybe.

E

C

Don't optimize to excel in the party? Okay, yes,.

O

F

Forget for, like JPEG compression you'd, probably say.

X

O

Now it becomes writers right now the documentation bottommost says the compression operation should be optimally compressed, even if the operation takes along a tactical view. So.

I

What we would do that is, we could probably change the documentation to say like it's. It's really a flag passed into the algorithm, saying that the algorithm should choose what it, what it perceives to be an appropriate hole. What it thinks it's optimal, basically,.

O

But then, our this algorithm, how the cuts customer who looks at compression level be able to reason about what optimal.

F

I, don't care just.

O

Compress it yeah. Yes,.

F

Has nothing in the past, especially minion, no has an opinion. Optimal is.

I

Just what did you send.

F

I

F

I

So to your question, to us and I think on I think we would probably just document if you don't know what to pass in passing optimal, and we would also encourage that to be used. Its default parameter everywhere, I see so.

O

No value so now optimal and highest in broccoli mean the exact same thing.

F

Turns out fast, it I mean I, guess fastest usually implies it's the fastest thing it's like for no compression, but no compression is the fastest.

O

How do we get the balanced in barley? Well,.

A

I think the idea of body was that it was highly optimized for basically just being super fast, so there is only one one particular table whatever they have. They.

O

Have like 11 or something you know there is a bell you could. We could evaluate a balance, but nowadays well.

A

The underlying algorithm has it has a gear, then yeah. We can choose that I. Think now.

F

A

Would allow you don't.

F

Right so we could change our bratli implementation in five to say. Optimal now means it's a more optimal balances, Basin hyman! If you.

A

F

Smallest didn't change the smallest so.

A

We can't change the innovator, but we can't change what was what underlying in that maps to we can say in it five we say: I don't have to 11 maps to 10, but.

O

What is more, each one.

A

The time with an.

O

A

A behavior, it's a behavior breaking change mine early, but like I mean what are you welcome? But if you really want maximum you, you specify Mexico.

I

V

I

Good point like this, we would probably want to change the Broadway constructor to allow you to specify the actual integer.

F

E

C

It's selected list, anything else seems to be like I need to think like what does it actually mean? Is it the smallest number or.

E

The size or ya.

F

Really need the idea is to separate the notion of most of these compression algorithms have an integer of compression lesson, and we don't want to say the highest number is that it's the most compressive or smallest size, gays.

D

F

I mean we can call it most impressive, like I, don't ever gonna start having Jeremy isms into where.

I

You have that that's what I should just.

N

Invest because I think people from.

I

Well, the majority of people will know what best to.

N

Be the best is ambiguous. You have to look it up. Yeah, it's Lois.

N

A

Let's, let's have lunch because.

F

I think a lot of people on Stack, Overflow I would say what's the best compression level and you would and then you're like not best.

F

So so, objective answers, like smallest maximum that are the highest, is ambiguous, supplied.

H

P