.Net Foundation Design Reviews, 9 Apr 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Review: GitHub Quick Reviews

Description

00:00:00 - Approved: Making the info GC.GetMemoryInfo provides public https://github.com/dotnet/corefx/issues/34631#issuecomment-481358549
00:37:15 - Approved: Revisit Index/Range API requirements https://github.com/dotnet/corefx/issues/35972
00:49:27 - Approved: HWIntrinsics API Proposal: VPMOVZXBD and friends need pointer-based overloads https://github.com/dotnet/corefx/issues/35768
01:09:05 - Approved: Take another look at the `COMISS` and `UCOMISS` hardware intrinisics https://github.com/dotnet/corefx/issues/34881#issuecomment-481371966
01:13:43 - Approved: Clarity and control on a JsonDocument lifetime https://github.com/dotnet/corefx/issues/36152#issuecomment-481375975

A

um Should be like now, everything works out to plan, so yeah I applied this this blocking partner label so that we can separate eyes or three other things, because there's tend to be too much stuff in it, and people suffer priority salvation where we just have people in the room, they only care about their stuff and then I mean look there. That's the that's. The thing Mumbai only suffered like I, think for like three months, I think so we should finally take a look at this VR.

A

That was very like racing that we have for a wife.

A

I just copy the URL, so I can be on my own screen as well.

A

Why do you want to walk us.

B

C

um So new screen down but I just want to see you going to the other yeah and so there's two maybe eyes, and when I open the the issue there I only post, 180 I but ii, any piano is requested by some another issue thread. So the DC memory info one is basically just something that we that I exposed to be CL. It's a internal API before, but you know obviously being useful to. Everybody else, gives you a list of things that they're all dirty. So, first of all, it's a read-only thing.

C

First, and some of them are things or like, for example, the the first. Alright, you can scroll up a little bit can see the top yeah, so the high memory lowest threshold. That's something that right now as soon as the constant starts, it does not change, but we we we might consider providing an option to change it. You know, while the processes and the other ones are the numbers the last GC saw so, for example, memory load.

C

You know last memory last GC, summary Moto's, fifty percent no no available memory again whatever the last is, is long same with heap size. Information are.

D

These measurements generally invites like total of their memory, so.

C

Total available memory size, the fermentation all invites memory Lotus and percentage zero to a hundred percent. Oh.

E

I see the comments are, should we include bytes, okay, yeah and then the threshold is the same. It's also good.

C

F

C

The reason sorry about it, it's an ant, it's because normally we get these from global memory, satellites which returns memory load as a zero to augment empty words.

G

What does Franco for what is it mean for fragmentation to be in bytes? What is that number? Actually, returning from.

C

Mentation is the sum of all three objects on the feet. That's what we call.

E

It from the animation and.

D

This was, as of the last collecting the correct visits, while.

C

You better, let me go alright. Last you see whatever the last piece is all so.

G

Fragmentation is the sum of all objects on.

C

Three objects: a free.

G

Objects: yeah, okay, that seems like a an internal nomenclature leaking out like I I, wouldn't expect that.

H

Does it not include things where the entire page or segment or whatever you work with, is considered free? So it's a of active GC areas. Oh so the.

C

Reason why there's this concept is because you know if we don't talk about pinning for a second right, if there's no pins at all, when you compact the heap, these are the the free. Let's say: no, everything will survived every non free object on the heap survived these. These are this numbers? How much you can get back?

C

That's? Why isn't it so.

D

But this is only if the last collection was a gem zero collection, this that.

C

D

Include dead objects in gelatin right.

C

Right, you can then choose to say like no, if you think so the example usage would be like. Oh, the memory load is quite high, but probably not high enough to the high threshold. You know, but then you told GC right. So you know what, if you know you want to do something you want to react before that, then you could say, like hey, there's no total available memory, and this is my heat. You know total heap size and looks like to be a significant fermentation of the you know.

C

They use the memory and then you can say like the phone. If fermentation is also a significant percentage of that, I may choose to do a compacting Gentoo, so I can get that much space back.

D

Is a sprayer speed on.

C

uh Right now they are one of the people that are nothing for us. Yes, it's.

D

There a scenario in which the GC might change in some future version, where these properties are no longer populated.

C

ah It's very unlikely because these are all things like GC calculates and all we do is like know at the end of the GC would just say like it: let's put these numbers there and you can get it if you want okay,.

E

Then the Colton give GC memory info gifts, that's what the last GC saw right, yeah right. So if you hold, if you hold this struct and keep reading it over time, you get no yeah, no, no Jesus exactly yeah.

I

Or today there is no another way you can do this.

C

mmm So memory load and you can get sort of a if you're, using hosting API, which almost no $1 of I know I, do know of one team that you start but yeah almost no one does keep science, you can not not those heap size. You can call it that total memory. Our juicy class has another method called internal memory, but that's sort of that's weren't heat, sighs, fun meditation, so you can get that, but it would be nicer if you can get both heap size and composition.

C

You know what the I shall keep my heart large. It is, and then you can say like. Oh, if I did I think I can Jen. Do you see? This is how much I can get back and.

J

Also, to clarify your price, the GC membrane for is not up taking in place whenever energy happens. So it's going to do that again.

H

Cuz, it could be a it could be a destructed representing a cursor. It's a pointed toe.

D

So, just to clarify the fragmentation is from the free objects. Only in the generations that we collected and the heap size is the total keep size across no.

C

This fragmentation actually total front many on the heat, just because GC is not collecting. That generation doesn't mean GC can't know.

C

I

The question is: do we need to have a new struct for adjusting? No more info can be.

A

What's nice, if making it instructors, it's clear hold useful day, but if you make them study with only then as soon as you cure them, you get them, and you know they no longer snaps over this one unit right there. No you get like one data pointed one. Second, I did it for another. Second, they don't longer correlate to each other and I. Don't have to know Tom making me.

D

To thinking how expensive or do you think it would be to actually populate the struct like you? Do you already have all the information right there and it's just a literally none copy in the fields yeah.

J

These are requests implementing that his feelings. The end of that three.

A

And so like in terms of like comments like the comments that are in the code, I I, don't agree, like I, think percentages in our case that usually loads between zero and one, because that's basically, if you want to format them the spain format. Like that's the factor, the the debate you range, we expect to basically put it into percentage, quite frankly, whether it be suffix them with percent or not. I, don't think makes a huge difference because special then load and a percentage in nature, but.

K

The other ones are.

A

Honestly sale with suffix levels by it's because especially electric notation to me, sounds more like a percentage. That sounds like a number of pie. It's like, if you say, I, don't know like total available memory by it's totally. You know a heap.

C

A

It's a technically.

C

I'm calling fermentation ratio if I want a percentage of.

A

C

My you know also another things: I mean I, have absolutely no objection. If you don't think there should be bite at the end of that, but normally um like imperfect, for example, where we deal with numbers all the time, I think if you don't they bite in the in spite. But if you have like megabytes on for yours, you say: oh my go guys. Yeah yeah.

G

I think that I think the clarification is just whether it's a unit of memory versus non unit of memory like it sounds more clarifying bytes versus kilobytes, it's ever clarifying bytes versus percent, or vice versa, something else yeah.

H

Since I did meet percentage or ratio because certainly when I first looked at the word threshold, I thought it was an absolute number.

D

Yeah yeah, but if we change it to a floating-point number like the the return type presumably indicates for something for Rachel. No boats are bigger than it.

L

Can be bigger than name but they're, not necessarily precise.

D

Well, it's not necessarily that they have a range bigger than int. It's the slight! If you're using a float. That's because you wanted to return something that could have yeah I.

H

Still think that adding ratio.

D

Were presented, okay is how.

A

Would you call the first woman I memory, though threshold percent.

G

Can you define exactly what this concept is? Oh.

C

It's just love saying you have you know eight years memory you've used four games in physical memory. That means your your memory load would be 50%. The high memory load threshold is a threshold that GC considers to now. You've entered high memory load situation, so normally this is about 90%, but you can use a config to change it. So.

A

C

A

Destiny GC all right. That's really exaggerated, more aggressively GC.

C

That's when we start to more aggressively consider doing gen to see this number could exceed no, it's on percent max then.

F

You're not using more than.

C

F

Me know, then close up of a high load threshold.

H

That wouldn't make your if you're in high mode, you just pretend 100 I mean.

C

If you specify a number that's more than under or just happen, I.

A

Know so it makes it so basically what this indicates is like what is the percentage of memory usage that did you see, considers high memory load is that 80% is in 90% and I'm 95%. So this.

H

Was how? But how close we were to hitting the high memory state.

C

Threshold usually means like a number that you consider something else: yeah and.

L

I think the normal for our existing done it at the air just to prefer double over float.

D

Never use forward exactly yeah if, if we do name the property %, we just need to normalize it to 0 to 100. That's one I.

L

Think it should be normalized between zero and one because for precision purposes. Oh.

D

I agree with that: we just shouldn't call it 470 I mean the the term percent literally means baseline between 0 and 100 yeah. So you guys you, okay,.

M

Enough, like you, can't call it a factor, then no that go well, although I.

L

Think a lot of people who work with doubles and floats, regardless of what percent normally mean, will expect a value normalized between 0 and 1, and we could document. This is normalized between 0 &, 1,.

A

Well means to be clear: like I mean in the VCL, if you string.format, something is a performer, that's percent. We expect the weighted between 0 & 1. We don't expect available in 1, but.

D

It's all knotted as a percentage yeah. We multiply it by a hundred new place, showing it all.

A

You call them online later, then, if you don't close the percentage right.

L

So so 4 percent we consider one to be 100 for formatting purposes, yeah, yeah, yeah and for other purposes and I think like system drawing works, the same way yeah.

A

I guess my question is: what do you call the input to the screen format? Then that'd be four minutes percent vector.

K

D

A

D

H

D

H

Would be is the thing over another thing, so this is a usage over the limit. That would be a right there. That would be a ratio yeah.

G

H

With Marshall so I mean I, guess once it's wanted to double, and especially if we put like subjects on the other things going, yeah Tory.

A

One double afloat.

N

Who's me yeah, that's one.

A

And so what's the name then high memory Lord by.

C

The way this one, this might create some confusion for people who actually use that config, because in the config we simply say this number is best. Hydrogen Sarah.

H

Each time they are not Jimmy anti-memory load setting minutes.

J

Go see the high memory.

H

Load setting like just telling you what the DC was configured at so.

D

You're really just reflecting this number- that's breadcrumb config back out to the user, then so.

C

If you don't specify or come thing, we have defaults right. If you do, then it's not number, you specified what.

A

We call it and do it in the country I.

C

Mean threshold, something like that: cuz.

A

I mean maybe we should just have the same name. They already have my big thing and the config has to be an integer.

B

C

An int there's.

J

Another plan to actually.

L

J

L

They can figures in it, then I would say just return in int here as.

D

Well, just with what exactly looking for the thing.

J

Well, the I remember no special, there's another API proposal that proposed to be able to change this so well.

L

In which case I would expect people would be inputting in it there as well.

C

Actually think it's called GC Hyneman.

N

B

Even see this later even fight this.

C

Oh I I asked Ron to document us and he said, don't be Oh.

L

Yeah and if we're going to support setting in the future and point values, don't actually do anything because we're not going to like say 61.5% and then actually treated at 61.5% well, three days like sixty two or something then I think just in in between as well. For that. So.

A

It's called high memory present. You said right.

C

Gc hi Menma present I guess.

A

High memory and we should probably split OH, yeah yeah and then okay, we say it's high memory percent of me said it's an image: okay,.

D

Memory load seems like that properly yeah.

L

D

We were dividing to launch together.

J

Seems a bit confusing that it would be a high memory load persons, which means either 100, and then you have a very low comparable like the same and a verbal value yeah, but he's a double you can do you.

D

Could you could make the retinue argument reasonably that one of them is literally just being reflected from config and one of them is being dynamically to get it sir, but yeah.

G

Except that the config is so buried and so non discoverable compared to an API. That's so prominent right.

J

D

Fear this problem I mean.

L

Come on zero to 100, for this is probably fine, because I don't think people are going to be wanting more than like three digits of precision anyways, in which case the the floating-point.

M

Rounding error is not going to matter so what what how you would use that number, though, presumably you would use.

A

It in a computation which gets the very first thing you would do is divided by 100. The.

J

Way, it's already used in context is by doing comparing, because you keep memory no than the high memory on threshold and saying that the high memory, especially kind of an absolute value that you use- and you say if I am at 80% of this- a memorandum threshold. Then I'm your high memory pressure, for example, I, guess I'm going to reduce the caching. So you really use memory nodes and compare it to other. Like ok,.

A

Ambitions really wants to be the same battery otherwise like well.

L

I mean it yeah, you want them both 0 to 100, but I. Don't think you need both flow or both int, because if you've got one in the other double, then the amp will off implicitly convert up to double right.

A

It just seems odd, I mean like honestly they're like I, don't know because we could say the radio is low, it still or double so.

C

In this stage,.

C

But making the memory load as double, because when you specify something you probably don't need to say like wow like I, have to in some 90 I want 90.1, it's fine to say just 90. But if you have a lot of memory by hundred twenty eight gigs one percent is a lot, then you could wait for small processes. You couldn't want this to be a double that so you can get when you know very precise numbers.

C

I mean I can, and this also allows us to change it, to compute precisely.

A

Yeah I, don't have strong preferences like I would generally say, like I would try to make ap eyes. Look like similarity. I mean there seems to be no downside to making it afloat. I mean like yes, it's a bit questionable whether you ever returned the video, but is this seem to all of you? Look at the API one as an in one is the flow, but they're. Both percentages is just mean. That's alright.

A

I mean why not just return the absolute number instead of a ratio, you could I mean think that's the other thing I, because it's a memory, though it is always a name, but there's also seem to be it right. No.

D

No I mean a long, it's literally the number of bytes that have support for me and.

F

L

What you can't win against total available mean yourself yeah. What about this is the case for you of hot pluggable memory in your system. It will.

H

L

H

And what do you consider these yeah.

D

Because this is a immutable struck, you wouldn't use a struct to go set properties on a few signal and maybe if we have a way to modify the DC at runtime, then that will take a double.

H

Yeah, so whatever we want for this top guy in the memory would be the long and then put Mike something into everything accepted.

D

When we load which formula many bytes as well but normalize, literally everything bites good, say like when, when we detect this many bytes or more on you, so we trigger reducing when we. This was the the memory load in bytes when the last you see at home in his home, and then you have all the information there available to you. If you want to do your own ratios that way, we also don't have to worry up what versatile.

N

Even though double spoiler.

A

So, okay, so it means everything is long. Everything in suffix of bytes.

D

Okay, let me write this down. I mean GC team. Does that some reasonable just expect normalizing, we literally everything the bytes and expose no ratios, or this truck.

C

The only thing would be if people did specify them config they. My thinkin.

H

Well, but config says percent in this doesn't yeah. So this is your interpretation of the components. This absolute number I.

C

H

If everything on here is long and bytes I think it's easy for somebody to go back and get the percentage if they want and and it's better than, if we have it percentages compared with double percentages compared with long total or absolute values. Also.

C

So you I think that you right I think you guys think that API should be more like prominent in our and how people use a framework, but I think people on the runtime side are really trying to push for more configs, well more complex. That makes sense anyway. The configs won't be I. Imagine like documented, you know, talked about, and and so.

G

Never changing I.

C

Heard from rich lender that we sort of the compatibility bar on poor doesn't seem to be the same as on desktop. Well, that's.

G

Different for you guys, that's the thing right. So it's one thing to say that you know the compatibility bar, isn't a big deal so from a you know, a config perspective. We might change configs from release to release, but from an api perspective we ship an api and that's it.

C

Okay, so you're, suggesting that we are.

G

Suggesting there's a reason: we're thoughtful about adding new AP is and what we name them. And what not. Because oh.

C

I completely get that I'm just also I. Have you know I completely understand. We need to be thoughtful into adding new API I'm. Just saying that complaints are also sort of you know, probably not up to the API standard, but configs are not like we're at least trying to make them not like completely hard to discover. We are trying to make them. You know to the you know, stand in a place where that's this garbled.

G

Yeah discoverability there, but what I thought were the path I thought you were going done was well we're, making our configs more prominent and we're documenting them. They're Rivlin api is for all the men, I.

C

Highly doubt, that's gonna be the case.

J

You make the high Millennial threshold of bytes yeah. How do we compute it back to a ratio to a person thing like? How do we compare it so.

C

We never official available memory so.

N

C

Be the percentage of total available in there yeah.

L

And I imagine that if you exposed an API that allows you to change the threshold at runtime, that would probably be a percent that what the user wouldn't have to know the current memory available, because you're changing config at that point, yeah yeah.

D

So bites all the way down the bullet done. It.

A

Was that good, so the other thing is DC memory, and first of all, do you might be the only right. Yeah would be my guess.

G

Unless we cheat I heard someone say that we might have the ability to change the threshold, and that may be.

D

L

Yeah, this would be a method on DC that would allow you to change it. Maybe.

D

We should call.

L

The snapshot or something no, because.

D

It is important, I mean I, mean.

A

Honestly, the API gives you that already or anything, you've always method you get back the destructor, then yeah.

K

A

Don't like it I can find.

K

O

G

If I'm, the only one, then I'll stop mentioning it, but I still have trouble with the word. Fragmentation for the concept actually conveys. What.

A

Would you name it soon? It.

G

Sounded like it was free bytes.

D

It's not a memory. That's immediately Frio, correct right.

G

C

E

C

Crack, though, because there could be other fear, more bias, we just haven't discomfort, though.

F

C

Very different, so if I have a heap that you know, that's like one gig right, I have 500 Meg's of 3d objects right now and next time. If I do a compact change into DC, I will likely get more than 500 Meg's back because of site when the 500 Meg's I still have other objects that wouldn't survive. I didn't.

G

I'm at the last did you see with that girl right sorry.

C

G

If you could atomically, if you could have atomically computed the number it would have been accurate at the time of the last JC. Is that true or false, but.

C

But last GG may not be Oh for GC, so I couldn't have computed. That number. All this is saying is that this is the matter.

E

This is the minimum.

C

Of stains that you can reclaim, if you were to do compacting them to DC, is.

D

Free objects, the the name that's used like within the DC and.

C

We're going to documentation is car keys.

D

C

I, don't I, don't know em.

G

Used an SOS, for example, right like that's, what's printed out and in the heap dump and whatnot yes,.

C

D

We should use the same from all of you sure, yeah three.

C

Object: bytes, yeah, but confusion, but.

G

It's also, it's also a min right. It's it's I thought that also need to be in the name, but.

C

Three objects: you couldn't. Some people could interpret that us. You know like sort of three of us stuff didn't survive. That could be really.

G

Fragmentation could also could have a whole lot of other meetings as well. My mind goes to things like fragmentation is the percentage of of how much of the the heap is has gaps in it or.

C

Bring objects or gaps, if you understood it that way, that means you'd be correct. So.

B

Would it be been dismantled their lights, then.

E

I mean you can.

C

Call it from Asian buyers, I'm, sorry.

H

If I have some some objects that have become dead in generation, zero, but I call GC collect a million times so that you know it actually does things? Are the bytes in they're counted as fragments that are not given that there's now nothing allocated you conceal if.

C

You only have general nothing yeah and you have three objects so.

H

I had some objects, they got unrooted yeah.

F

I called GC collect a couple times right is fragmentation, zero or some other? No there's no.

H

To me, and that would be.

F

Hero lightweight okay, so if the entire, not counting the entire generation heap, has become a completely.

H

Unused fragmentation of zero: it's only about objects that heaven.

F

H

The middle, so it's.

C

Not that's why I'm saying it's not good to call it free, a bowl, because now you, you just implied something to do with you know: lifetime I.

L

Think the question was: let's say that you have objects, a B and C and the heap, and then you've got a bunch of memory. That's free never been allocated on the rest of it. Yeah.

C

That's not powerful.

L

C

Yeah, that's not part of fragmentation yeah! It's only like if you get the gaps. Yeah like that, so.

L

If you free B, then B would count as the fragmentation since a and C are still alive right.

C

So if you have something on the close right, so this is my free.

C

So fermentation what means us yeah and you can, if you did a GC, you could come to this situation because we may not have done on compacting GC. But if you were to do a contingency which you can induce via one of our API, then these would be become like this.

C

L

Part and then the rest of its free memory, but not fragmented, exactly.

C

Right so this it on our segment right. This is our thermo safe. So this part, that's not part.

J

Of our innovation yet, and so on, you're on the first thing, what is memory note? Is it the four slots or if you remember,.

C

Low does kind of different memory limit. Is the physical memory loads which mean? Let's say you touched all the pages that your objects are on, then that would be this divided by available memory right, but of course you could be in a situation where you have a very large object. You haven't touched to all pages and that it would only be touch pages because it's a physical memory load right. It's not minute. Bytes yeah,.

J

G

It what about reclaim abow or something like that, is that at all accurate I.

C

Think that's just the cosine fermentation there's a very accurate meaning into it. We climb the bone could be I, don't know, I, don't even know how people were interpreted.

E

Bytes sounds like the fragmented.

H

E

H

Quicker you autograph, yes, I guess.

G

I'll drop it for me and I think your fragmented in it to me. It sounds like something about the bytes themselves are split.

I

C

Think if you talk about memory, that's in general, that's not really how people.

J

Stephanie fenders done correctly, what you mean is memory that is not usable, because it's just long to be like the gap, is just not to be reused. Oh yeah.

G

That or I guess part of my problems, I think it's like I hear fragmentation and I think 59%. You know because it's like how how fragmented am I, you know what I don't know I'm having trouble edges. I'll drop it if other people have no problem with it like to.

A

Be honest, it seems like something we're documentation would help, but because it seems like there's something where you kind of handled it do. What my only just did right like I agree, and you have.

L

Yeah I, don't think Stephen could see the picture I'm like.

F

More Charlie, but.

H

The picture that may only here on the board is exactly what I think of when I think of fragment, like fragmented.

F

Bite, yes, how much.

H

Memory is in the middle of other stuff, and if you like that, you wish these were squished together, so that you could, if you want to.

C

Understand what.

F

C

Bytes here, they're totally consider for allegations. Yeah right I mean we're not just saying oh, because these are free objects, we're just gonna, not care about them anymore. Until next time we try to like squeeze them together, mm-hm, but these are like no next time. If this is Gen 0, you can do something. We would totally see you like and we fit it here. If we can we'll do that before we expend the energy first.

E

Of all yogic ate an object which takes like one.

J

And a half of this right.

C

So this would be made into one Ryoga: okay,.

J

So you may still have from ninety days. You have okay turn object, which is a bit bigger than the one like there's three one, your hydrophone, so.

C

The biggest out there you can fit here is a.

J

Slightly smaller.

C

Okay yeah, so then we repeat that like, if we're so in the allocation code path and you're, looking at another object that can fit here, yeah so.

J

You can still get to a point where you have only one word of memory between newly allocated object and the next one, so that would be in the sense fragmented, because you could interrogate, even though that.

C

Would be a central fronting, then you are describing.

A

So then I think that's just look at the last method you have here, which is the total bytes allocated right, I.

C

Actually didn't yesterday, I didn't know, remember, there's a precise argument that his eyes tells us if you wanted to suspend the yeah. So basically this yes, you know so far how many bytes have.

E

Been alakina in this process you know: do you think that people who are calling this in general are going to pass all.

C

I so there's a long thread where people are discussing this they're, like yeah I, believe some the stronger side was saying like they would call it with his eyes. So true they wouldn't want they they didn't mind just instantly. This is use in some sort of like benchmarking thing was. It was even from that Bank over phone guy, or maybe it was Adam yeah.

N

So, like they're.

C

Saying, like okay, wait before we tear down the process, we just call this, since that thingy I mean we already spent me and some co-packing shut down anyway, so they want to get like. Oh did this. You know if I change to a newer version of framework did the total allocate advice, although.

G

You know in a benchmarking scenario, with this I'm trying to say with this even matter if you, if you call it before you start any work associated with the thing that you're measuring and then you definitively know that the thing you're measuring has ceased to run anywhere, and you call this again. You're gonna get a fairly accurate number right if nothing else was happening in the background like you're, not doing any other background work.

C

It's that's true, I think. If, unless you you have some background work, I guess that's what people are saying right.

G

But if you had background work in your benchmarking, you're doing something wrong: yeah.

H

You know, sizing true is: has a at least potential significant performance penalty. I think that I would say that we should go ahead and have it default to false. So somebody who doesn't understand the consequence gets the less impactful thing like if they're just like, oh I, want to go like blog. What this no I.

C

Mean they're, you can't imagine this use in non shutdown. Obviously, right now, if I have like different like if I were to write a compiler like I'm I'm, a compiler benchmark, it's very possible I want to at each phase of the compiler I want to sing like a how. What is that, like advice, so.

H

But I think having the default defaults means. If you don't know what you're doing you get the thing. That's nicer on the rest of the runtime and I think you know what you're doing then you can bastard yeah, so I would be both a couple. Yeah.

H

It just signals which way we expect someone who doesn't know what they're doing which, without a visit.

G

And just a check is this: is this expensive or is just just like maybe precise, is false? Is this just fetching a value from somewhere, or is it actually.

C

Every every manage throughout the house, this, the oldest tours on the project so.

G

This is iterating through all the threads in something.

C

And you need to keep track of a tenth run.

A

All right so I just of this alright, based on what we just said, everybody agrees with that.

C

So should we like edit the first comment to reflect this or what is that so.

A

Normally, like so you're free to do that, I mean usually like after the feedback has been submitted. I don't expect it to update the issue to match that you.

C

A

Do not like, if.

C

A

You update your source code to match that that's fine I mean I mean if this is the last comment on this word I think that's my Lois.

E

All right all right, thank you.

A

All right so one down so no I think Eric wants to look at index range because we want to get this one done spoon thanks. Perry thank.

N

A

Went on a plan for this one, so I think Jeremy knows that. So basically, what we have done is we have changed the way the teashop compiler binds to index in range, meaning originally, the goal was that index and range is just you know, enormous struct that the compiler binds to stood next to. So, if you want to index into a collection that uses the range, then the collection physically has to provide an index of the tags range and then the same way you want to use.

A

They had character which binds to index you need for an indexer that has that that takes index. Unfortunately, there's a couple of downside, the biggest one being that sometimes you just can't add members like, for example, let's say you want to use our bedont interfaces, mother and father yeah. They only have one indexer, but the compiler team basic another way where, instead of doing that, the compiler effectively emits the code that takes the length, slash, count, property and effectively. Just does the right thing.

A

So now that means we normally need to expose that many indexes to take range or or our index. So what this patient reflexes? You know the proposal that I hacked up yesterday of given the entire API service that we, if we did for an index, let's go with that and decide whether there's a few things you want to remove, and so this is basically that I think there was some simple form from Taric already on the thread. So I think this proposal is no longer.

A

You know the most up-to-date one, but I think that's we can I could probably edit this.

G

The comments all basically came down to do: we need those remaining slice index methods that you left it. Yes,.

A

The only reason I left them in is basically so. My understanding is that the compiler- let me actually get this right so yeah, so the so the compiler will not do the lowering for index on arbitrary and right. If you want to slice using index by itself, then you just have no way of doing it unless relate the indexing. Well,.

G

Now you do you suit rate, you know: ma'am open, open bracket, hat 35, dot close bracket. Oh.

A

Yeah: okay, your basic using range bother the index. Yes right, that's fair, yeah I, don't have a strong opinion like I mean if that, if I think that that bed works for me.

A

So that would mean registry could do this. You would well. This is not inconvenient at all.

A

A

A

All right, I just updated it to say, was it doing alright, so we kill them all. Remember extensions I left in, but following your rationalist, even we could probably get rid of the aspirin whether it takes index and only we.

G

Could I don't I, don't have a strong feeling about it? I I think it's actually I mean yeah I, don't have a strong opinion as long as we're I think there's a difference between going from zero to more than zero, but once we're adding one adding two doesn't seem like a huge deal. Yeah.

A

Then so string, basically I killed the one that takes up string following original. You could do it or we.

G

Would also kill the one that takes range, no.

A

I thought that's the one we need, because the compiler will bind to that one.

I

Fine Angela, what who engine straining engine yeah.

G

And if I, let me already have oh.

A

I think yeah. That was the thing that in academics.

A

Three, so let me actually don't contain those.

I

Only many issues that Stephen raised about midsummer rain, yeah.

A

So that's so honestly: I have not a strong opinion. I think. The reason why we put it under compiler services was who didn't expect anybody to do that directly and having that method seemed weird on an array, but given that we now have this generalized method, you actually having an aesthetic slice method.

A

Inspired to me, I, don't.

G

Have a strong opinion either I just wanted to make sure we were visited because I realized it kind of slipped through without much attention paid for it, because it was, you know, just a little helper for the compiler to use so I just wanted to make sure we talked about it, but I don't actually have a strong opinion. We.

A

Can't make an extension method because yeah.

I

Also is a problem: if we change it, so it will be a breaking change for somebody using mix and match between the compiler version and.

A

We fear that's true for this entire change, right, yeah,.

G

I mean we would add a new one and then once the compiler money.

P

G

Propagated being used guilt, we'd remove the old one.

P

Change how we feel about having range index. You know.

D

They would still be needed on types that don't have a slice, nothing, except for.

M

Straying away well for ones.

L

That need a complex slice method right like tensor types, yeah well,.

A

Basically, eight, so this there's two things right: the the biggest problem was that requiring the indexer is very unfortunate in cases where you have interface. It's because you, you fundamentally can't have an extension index array so binding to slice, allows you to say well, we could add sliced methods or interfaces in the VCL and then you can index into them just fine, a thing in case we have a country type like a list of T.

A

The idea was well it's nice if we don't actually have to add them to begin with, because we, you know it's just nice if indexing with the index and leaves just works, fine and then for range. The problem that the compiler can really decide what to do for you, because you you have to decide what type to create right. It's.

K

A

Somebody has to provide that from action again, whether you do it on the type or whether you do it as an extension method is actually up for the concrete time right.

D

Well, I guess on one hand, I have a little bit of heartburn with putting extension methods like slice on list of T unless those extension methods have, unless those extension methods offer a projection over that list of T, without making a duplicate. Because normally, when you see the word slice, you think people are free, Norman and.

G

Unfortunately, we're already down that rabbit hole. We.

D

G

D

But we don't have slice methods there right. We have something well.

G

But they're gonna show up in to come home from a usability perspective. They're gonna show it the same way: array right, yeah,.

D

I'm saying for somebody who's not using this compiler syntax like if they just see a slice method off of list of T like rent they're, probably reasonably going to think this is a review over I I.

G

Would think the same thing about the compiler syntax? But that's just me: I guess: okay,.

A

Honestly said, the compiler syntax I would even work with it, because it's an indexer, you expect. Historically, the properties are cheap and indexing is all one yeah.

K

A

You know when I talked with Matz about it. Basically, the conclusion was well Ranger. You know, ranging against this. Word is new in the in this thing and just have to educate people that well, if you get an e first of all, you have to allocate something usually right, even if it's a view, so it's not like super cheap like an index above an in, would be, and then the questions is whether you get a view or not depends on the type. So you have to check and.

G

Excuse me, for me at least I've used this example a few times, but the thing I think we're gonna have to be on the lookout for and, of course, you'll are in core FX and we'll probably even want to write some kind of analyzer, for it is if someone writes like span of char s, equals array sub dot yeah.

G

That will allocate copy an array, but it's not just like particularly visible in the source and.

D

I imagine like it's it's off the table to make this particular syntax return a span as opposed to a string or the right return. Whatever it's a nothing. Yes,.

A

I think for array, I think I I can go one way or the other I. Don't have super strong feelings on array: yeah, balanced strings I, don't think you get to a transparent like like they're, so not usable for people and extreme is just so much more useful that I think in common code, where people just don't care, because things are relatively short, most of the time.

A

The copies, don't matter that much and people do something today like left and right and like you just give them a way to do it in a more compact, syntax and.

L

If people want to have a span instead, if we're going to provide an analyzer for catching places, where we're doing this in the framework, we could probably provide an equal analyzer that says and sort of spend some yeah but put as span and handle it that way. Yeah yeah.

D

Isn't the point that people want to use a nice syntax on yeah.

G

You get that well, there's two pieces to the nice index. You can still do a span and use range like that's what those memory extensions were, so you could say as span Open bracket close bracket, it's not work. It's you. The argument for returning the same type, which is pretty much locked, is whatever, however, I feel about it, which is I. Don't like it. But it's you know it's whenever you start with you end up with yeah. If.

J

You want something different or.

G

Performance reasons: well then, you know what you're doing and you write these new extra characters. Sure.

D

I mean if that arguments already been beaten to death, and it's already been yeah I mean.

A

This will be Leonard, win and LD am I mean. To be honest, this is one feature everybody have seen a lot of custom made in one way or the other right so I, don't doubt the either side can you know say that, based on customer data, that's the right or wrong decision, but I mean internally in an ABC. We have just very different requirements, have been all that right.

A

So, like some things that we like a syntactic sugar for the platform may not make you wait for the customer- and vice versa like this so anyway, so I think then this is the one I so to be filled out. This list here, I got from I, don't know from where I think was an issue that I think tarik or somebody else open.

A

So I assume this is complete, but we should probably do another pass and making sure that we don't have any other API is injuring around, but other than that I think we established the principle years, so it should be ins and repeater, for you go down the other side. So.

G

Forget subarray did we want to leave it where it is, and it's also.

Q

G

Poor the compiler uses, or do we want to move it? I.

A

Yeah, let me XD copy and paste this game and make that clear.

G

Question sorry I was going to say, I, don't know, we never discussed this with the compiler team, but we need to figure out whether it actually takes to ins or whether it takes a range.

I

Gets alright, the special kids a.

A

String anyway, right I think so following Jared, he was the one that said earlier, that he doesn't mind specially casing, string an array and I mean clearly. The compiler today has to do work to special case that gets up a rate method somehow and whether the special cases for this location, another one I think- is largely irrelevant. Well, what about things like array, segments, I, guess.

D

We were just out of sight spotted that, oh, it already has a slice, nothing never mind on.

L

If I think run time, helpers should largely be reserved for things that you can't do in regular C sharp code a lot of times, they're, currently things that are actually implemented in the run time with special case support, and but this sounds like it's just creating a new array and copying the existing sub array into it. So that's a trivial c-sharp code. There.

D

There are some considerations there, especially because arrays have weird pseudo covariance, so and I think Steve actually just check this and last night, so I'll. Let him talk about it. Yeah.

G

I, just didn't, read and prove some stuff that Eric had done previously, but it does it. It tries to ensure that if you give it, you know if T here is, this is object, but the actual array is a string array. It still allocates a string array rather than an object array. Things like that. Yeah.

L

But that seems like things that string the string class already does so or the the array class already does. Rather.

G

Yeah, it's certainly writable in c-sharp right.

L

G

Just it's just it's not just new teeth right.

L

D

It's not intuitive one.

R

Thing I heard was, you could do a span and then bracket daughter bracket, but in that counter to our guidelines that don't use a spend or slice just do a span and pass and the slice arguments, because.

O

Faster, that is, a span open for an open square bracket, stuff close square bracket, close bread.

D

Oh because you arranged to add stuff yeah yeah.

R

H

Yeah so I wonder if, instead of slice, we would call this, like you know copy segment and then that way, we're avoiding the notion that the word slice is a deep projection. We're saying the word copy it's visible and the compiler can special cases you're talking to your a case, specifically yeah, the string already has it at some string yeah. We already understand the substring makes a copy. It's the thing it does. Yes, number ready would be the equivalent come here, but we've never had a separate yeah.

L

And sub-arrays sound like my first intuition, if I heard get sub-array would be it's basically doing a slice rather than incurring a copy, it might be doing interior point dramatic or it's going. He write it right. There break the break into T bracket bracket, oh yeah, so.

A

My fear, though, like even in the method slice your argument does not get better yeah, so.

H

That's why I said copy.

I

But how do I painful description for the Cabana, because I hear that this would be very.

A

G

Way, you say breaking change, I mean this is new in this release and we can mitigate it. Just by saying we at the new one that we wait a while and then we remove the old one yeah yeah.

H

We don't have to move it in the same, commit we just want to remove the existing than before we ship yeah.

G

H

We don't want people running off of beta things with, for.

D

The preview builds and then mixing them with our vehicle. We also but I mean we've done stuff like this before, where we just haven't, exist, side by side photos, and so here's.

A

What we should bet this decision with the compiler team for sure, and they may have strong opinions on that, but I think that I think it's in the spirit of what the feature is. Writing to me. The fact that it's named that it's not an instance member, but it's aesthetic, member, especially cased, in the same way that you know on string.

A

It's not slice at substring right leggings, it's both our core types and it kind of makes sense, but I think the end result feels a little bit more self consistent, because array is full of static helpers. There are no instance, vendors on SZ arrays. Realistically, so let all the waiters buy.

H

My comic, we should call it shopping segment, sort of sense.

B

Should we call it copy to.

A

Match the other ones, so I wouldn't do that, because the argument that you're bringing up holds for everything we do- and you know, and all the kids to be accepted- that it can be a copy right. So how's. This method special because.

H

It means that array, dot space and array dot as and that slice don't have like arrays are implicitly convertible to see. That's.

A

H

A

H

A

It's not like what.

H

I'm sure he isn't called slice, it's called substrate yeah.

L

I'm saying we should not.

H

Use the same word: we use on the spam that it's implicitly convertible too, especially because what someone's using bar, you really can't tell what the codes do sure. So,.

L

H

Mean granted, this is a static method, so you have to call it using a different syntax, but I still think that we should not use the word slice, given that this is implicitly convertibles back like.

D

If only somebody would write it, design guideline which says slice of shall be on a location for anything. What.

A

Can be a location frame that could be a view. I think that I think this. This is more like the question of like how do you think about the compiler feature in general right, because the compiler feature just says: I bind to method called slice. We have no choice but calling it slice in all other cases. That's why why is there a special because array and spring are given up? Are you basically saying that any range in excess should always be a view? It's really that's not the case. Well,.

D

It's no matter what you're calling we control the api's that developers use when they interact with the framework in the runtime correct like if, if we create an API, where it's likely for developers to do something that has performance implications that they weren't expecting, then that means that we presumably could not on our top scorer at play. Right.

A

But I mean it's not to be clear: I, don't think anybody called slice. People will just take an array, Open bracket and then index range at.

H

Which point it doesn't matter what the matter? Yes,.

A

Correct, that's what I'm saying so like I think the the eyes I mean the people have to be aware of it. I think is already the case, no matter what we don't.

H

Know so that means, when you see the word slice, it should be the good word and not may be the bad word that we should not use the good word on a thing that I want this.

D

Is our opportunity to say the word slice and the framework is all one. No.

A

Because I'm saying you can't do that unless you're willing to say an arranged based index, she'll always produce of you know how.

D

Would you do the I read only interface isn't so what we're talking about is just the word slice should answer.

A

Its ancient saying, if you want to provide an indexer yeah, it takes range yeah. You can't add the member, because it's an interface, you must provide a method called slice. You have no choice, you need, must name it's nice, because that's how the compiler of my argument is so.

K

A

Basically means, even if you said slideshow mean it's always the view that means for framework types, ignoring string. It was a special case. It always must be a view. It's effectively. What you're saying slice.

H

I'm not saying.

A

H

Should be saying, that's my oh one view on things that are implicitly convertible to spam.

G

That's too much point that is no longer what it is.

B

G

Fact that the compiler requires the use of the word slice on.

O

G

Type, you want to slice, it's no longer that.

A

You could you could to talk to compiler and say how most different words and slice it would I give this and to come up with one, but.

G

At that point, we also didn't have to go and add that thing to span a raised segments memory string. Well, you know all the other ones, but.

D

That's a good point accident should we at least make all the word.

Q

Please hold on.

D

Sorry here we go work.

Q

That does both like it prefers slice and then falls back on the thing that actually allocates.

D

If we did call these method slice, could we at least mark them in advance whenever, because the compiler is going to find them, either way sure this like get hittable rattler get minimal reference? We made all of those methods of it about no matter which.

G

One are we going to make animal grounds? Remember anything.

D

That's called slice that isn't an O of one production.

G

We haven't added any of those yet right, but presumably.

D

They would be outed, I read only one, so we're about to add one right right. Wow.

G

Static, no yeah.

A

I mean to be honest, like I could go either way like I think putting a little balls on it. Doesn't save you that much my opinion, because I serve to explain. What's it to someone we're.

G

Gonna, if we're gonna put a browser on, we may as well, not move it off a front. I'm helpers, yeah.

L

You know I think maybe just keeping slice is fine and if people really want to differentiate in the future, we could have something like get view, because that's what you would need to do for something like list of T anyways, because returning a span of a list of T doesn't make sense, but returning a list view or a list of T does so maybe slice is just fine and no that's reasonable.

L

A

You'll still have to provide the slice enough that any of it would not have to move through things well,.

Q

I think it would be that hard to provide to mean a different to have them seek for to like. Can you have copy and slice I.

L

Think the hard part is coming up with a name for it. That said,.

H

It seeks for two that means: there's a breaking chain. There's a bloomin breaking change, that's very hard to reason about because, if you add a method that has the higher binding than now, it's a compile break where everybody hey.

B

So also to be clear like if you should look factory worried, but you have that already with a bunch of times. Yes,.

A

But it's never grade like I mean that they they try to limit these cases right. X lies today, for example, it's not really a thing anybody uses well, you use it if you're coming from other languages, basically.

B

A

I think the I think at some point like Jerry, did the scanner and github and XE I mean who actually used slice and don't let go there? Wasn't that many yeah I think the question is just if you look back what was the motivation for this feature? The motivation for this feature was to fall right like some minor perf canes. But, more importantly, we don't have to sprinkle. A bunch of AP is with the framework right.

A

So if, if we land on design, that kind of requires us to add two methods in sync that seems pretty bad yeah I would say that that kind of defeats. The point of this feature, which was let's have fewer api's and mister compiler, just bind to them they're already there so I, don't.

K

Like slicing, that's fine.

A

But then to see the spawn, you would have to change the omit, the api's on span, which I don't think we want other.

Q

Than suggesting adding both they're suggesting having either/or so like having them do the binding, but I can see the argument about getting precedents. Bumped you have both of you actually have both of them implemented.

B

How about this so I can take.

A

The section out of here and we filed a separate action item for this, but then we want to move the method or not, because I think it's largely fallen well to this discussion here. The.

D

Ones, I publish nothing.

A

Yeah and then we can, we can have the conversation with the compiler guys again, if you want to do they want to do that. You, let me just quickly file an issue.

A

K

D

Didn't guess if, if there were critical mass to accept asleep eligible a alarm today, like I, wouldn't.

H

I just think it's thinking, rapid thieves.

P

Oh, it is yeah.

L

But we're gonna have some api's that are like that, and we've got plenty today and I I think the answer is we need to make our existing analyzer repo, more public, more available and encourage people to actually write analyzers for cases when they hit things like this yeah.

D

But if, if the code that you're, just naturally inclined to write, is flagged by the analyzer as you're doing it wrong what not just stuff.

G

Well, so speaking for someone who cares a lot about perf, the only types that I really think that's going to be an issue for is array and string, everything else isn't implicitly convert mean most thing. Other things aren't implicitly convertible to span anyway, and the whole design for c-sharp has basically come down to you get out what you put in, and so anyone using this feature on their own type. Regardless of you know what type is they're going to be using slice to mean what you you get out? What you put in you.

L

G

Generally, won't be convertible to span, not.

H

So my my reason of not wanting to use the word slice on array is, if we're already special facing straight with a special face array and those are essentially the two types that are implicitly convertible to really span I. Don't.

G

Care what name we use on a right, yeah.

H

That's why we're trying to avoid sliced there, because then, and it's when you're working with a thing that looks like an array, it's obvious what slice means, and it's much more contextual of what's nice to me, given that it's static gets less relevant but I'm trying to be slightly curious, but not to the point that I'm carrying order well.

L

You already have the case with string and array that, if you just call like string dot, something it's going to operate and return a string, you have to do as span to get in span semantics. So maybe that's consistent with the existing logic there. Well.

D

And but somebody may be able to have brought up earlier, like read-only Spanish, our food equals Life Drain breath in that was to do anything. Yeah yeah, like that it turned out. We absolutely need to admit, um um have an analyzer, the truck know what I mean.

H

If we were designing this and span at the same time, I would say we should not have any implicit, certainly.

D

The solution to overall problems in inter shall not be box without breaking the customers.

H

You for two-thirds of one more thing and.

A

So it's not the issue, I think the other one we can consider API approve. So that gives us there is something to work off of and of course, a.

A

Positions on the screen here.

N

N

Did you use fast.

D

L

The hardware intrinsics have generally been fast because we've said they're hardware in forensics go ahead.

A

We have Friday, they have thing to know Wednesday. We have a Jason one over idea, so in fact Cindy move zero extended wait. Does he work.

D

L

So for these api's we encountered two subtle issues with it. The first one is that the return type we decided for zero, extend to return unsigned and for signed, extend to return the sign type.

L

What that doesn't match what the native signature does and it doesn't match the normal see sharp semantics for 0 inside extension, so we said we should convert these two always return the sign type to both match the underlying native intrinsic and the normal c-sharp convention for 0 inside extension, and then additionally, we said we need to break out the memory specific, the memory load versions into their own methods, because they have different semantics than the implicit signature that takes a vector.

A

Or the first one I understand the second one, always so.

L

So so, basically for most api's, we don't have explicit overloads that take a pointer. We just have them, take the vector type and then internally, the JIT will say if this vector is coming from memory, then emit the more efficient memory encoding. The problem with this API in particular, is that it doesn't read 128 bits, so you can't do like SSE load, vector 128 and expect the more efficient encoding to happen.

L

Instead, this operates on the lower half of the register, based on whatever size, you're working in so the lower half of elements, so the lower 8 bytes or the lower 4 shorts, lower, two floats, etc, and so we need to have an explicit, separate overload that takes a pointer to make it clear. The semantics differ and be able to emit the efficient memory encoding.

A

For one right there all what we all seem to forward to another overload straight, no.

L

They are the recursive, because I tied the JIT implements in forensics only.

D

If you call that way, if you call them pure reflection, we can replace the.

L

Agents, it's fancy magic to say, I, don't.

A

Want to know that this goes back to what virtually I wish. It would have lied to me all right, so I yeah I mean similar to equip to I. Don't have many objections. That seems reasonable to me. The names seem to be in line with what we decided earlier, so unless some glasses Apollo. That I would consider this manner. So.

D

The the only changes is were changing, unsigned assigned, in which one type and breaking.

L

Out the overload so you've got no, no short, so am I useful them again yeah and this one is a bit more of.

M

A mouthful here.

L

D

It's the same pattern right, yeah and.

L

Then, breaking on to explicit T start overloads going a little bit of a play right. So it's.

D

Yeah- and these are all unaligned- correct, uh correct, okay, cuz! Normally we put the aligned, suffix I, think right right: okay, Steven.

G

I'm sorry I was muted my eyes glaze over in all these names anyway care backers. Well, that's.

A

What I thought all right that means API approve if you ready for viewing move blocking partner.

B

Hopefully should have added API at pool.

A

All right this one going second one, this.

L

One should hopefully be easy as well. We already reviewed and approved this, but I ended up finding. There was a thing that we didn't cover so so this issue was basically the ordering of the existing api's are confusing. They're currently compare equal, ordered scalar, and that can imply that we are doing that. These scalars themselves are ordered versus us doing an ordered comparison, and we said we should reorder those to be that's. My help doesn't.

A

How big, if they'd like them, you mean swap the votes right. Yes,.

L

A

L

Help I'm realizing that there's also an enum that can an enum for the 256 bit overloads, because they've got significantly more complexity, where they're similar word ordering constraints and I just want to basically wanting to double check. We're fine with also reordering these. The the one difference is that there's some additional contextual keywords here so like non signalling and signalling versions, and so we basically just need to decide where the order, if it's still going to be ordered equal non signalling for the same order. We had said previously.

L

A

Okay, so last time we said, compare ordered, scaler, accept sex and compare unordered, scaler xxx right and now you're proposing right.

L

So we've got an enum which doesn't have the compare prefix, but it still has all the other keywords. So it's got the xxx keyword, it's got the ordered, keep ordering under order keyword and then it's got signalling and non signalling as well.

L

And want to determine if we were going to keep those ones the same or if we were going to reorder it to be.

L

Ordered signaling less than or ordered less than signaling I mean.

A

In general, I would try to make these consistent, so that mean you get the rule that they apply uniformly and it seems like lashing. You have to make up your mind on like what the rule is right and so so I think the, but.

M

You feel, like you're, probably better suited to make the damn hat.

L

I think the most logical ordering that's consistent with how we ordered the other ones would be ordered, xxx and then signaling. So you might have unordered equal signalling and unordered equal non signalling. Right generally, don't we put the verb at the very beginning for.

D

The verb is in or the operation.

L

Right but we don't have a communal hair verb here, because it's an enum that has to the compare method: yeah, yeah, okay, yes,.

A

They see that now, not the under method, name right, yeah that works for me. Okay, then that's I, don't know exactly what I'm approving yeah, but.

L

A

An extra proposal, but right.

L

So the the ordering proposed would be, we match, ordered xxx signaling. You.

A

Write this up somewhere or oh.

D

It's at the bottom of the Muppet as well, so in.

L

D

It's in the very long I forgot to put the actual proposal, no, but I think at the bottom of the issue. You have an example. That's.

L

The example of how they're perfectly other, currently what you want yeah and we want to change it.

L

Ordered xxx signaling.

M

A

Damn I was just about to come in and say anything all right, so.

A

M

M

A

You Bobby even half time left. I can't believe.

P

R

I, go on to that question. Fisher I didn't follow up question on the GCE issue: oh yeah, we have get allocated pipes for card thread and we I just added get total bytes allocated should be rename it to get total allocated vise to match fights as the unit it should get.

H

A

There, which, which is serious now.

H

A

No I meant, like the one that you just said, the other one that we approved before yeah. We just approved it. It's.

H

Me you mean: when did they get allocated bytes for Fred coming yeah.

A

That's the one we need to change her.

H

R

L

A

L

A

L

One you know that bites to everything, except we didn't cover the another ordering yeah.

R

H

The prior art, aside from the fact that the unit makes more sense in the end, require art, was that okay.

A

So I put that's going I'm, you can see it but like we have get allocated bytes for current man should get told about it.

R

You just approved this API I, just swapped the bytes and allocated match. The theme is a you.

A

G

And many many already agree well, this is it yeah.

J

J

N

H

H

You okay, so we started talking about this previously meeting. We ran out of time so based on talking with Christophe duena who's, the only stake holder right now outside the base of never itself.

H

He thinks that what what we had was too many knobs and switches and too complicated, and that we should pare it down, which is fine, so he in fact recommended using the board alright. So the scenario is you have a Jason element. Jason elements come from a JSON document. Jason document is disposable. That means Jason elements are transitively, disposable and there's nothing. You can do about it. If you have one and you would like it to live longer, you need to be able to make it live longer.

H

So so this is where this new method is coming in originally proposed attached. Yeah that has some does, that is that a destructive verb or not so based on Christoph's recommendation clone yep, and so that's the easy part of this proposal. The guys.

G

Could question on the easy part of this proposal sure when does Jason element expose a parent yeah.

H

G

So, there's no way to get back from a Jason element to the document that it contains that it that contains it. You.

H

Can't get from an element to the document you you can't get from an element. That's a member of larger objects to the things that it was contained in okay,.

L

H

Can only go, let's call it down, yep got it, so clone in particular will clone the current sub element. So if you're not, if it's not the entirety of the document, it will make a copy of the smaller thing for GC reasons, and then the boolean is or to expose. Is this already in a state where you can let it live for a long time?

H

It won't indicate it if it's minimal, but it will indicate that it can't be disposed and the best answer I came up with is persistable once we didn't have the verb be called attack, but I don't know that I like this, but basically persistable like how would I use it because basically could always call clone in worst case it does nothing right, so you can always call clone at worst case it copies.

H

Q

This case, it does nothing.

H

So you're saying that's cheaper than so close as always, if the output of clone you can always save indefinitely, it's now GC semantics, so especially.

R

Let's say if you want stored addition, I never go to chain, you say if it's possible store it and stored it alone, the what annoys us always.

G

I think I think you would just call clone any store. The resulting.

H

Good, but if you knew that you already had the outer objects, you would know that you don't need to call clone on the inner objects, because that would actually bring it copying. A bunch of segments over and over again well, I would I.

G

Wouldn't go and just return itself it is persistable is true. If.

H

You had a hundred Meg payload and you only call clone on like on the word true. Did we really want to hold a hundred mega rounds as the as the GC reference.

H

Or do we have eyes anyway, I? Don't.

G

Understand I would have assumed that clone would have changed the internal structure to make it more of a tree.

H

Coppy, so clone is copy all of the bikes from the original payload, because we maintain them as they were because of we have the method get raw Jason, which means that we need to keep it in the original form, and so it needs to keep its if minimum its sub portion of the object alive. The question is: does it also keep any outer portion alive if anybody's already called clone.

H

And both answers are valid and it's hard to decide it's hard to ask someone: did they mean be the smallest possible clone or simply don't be disposable.

L

Isn't this bum we normally have with clone as well like? Is it shallow? Is it the eye? Cloneable is terrible, never use it. Well,.

H

This isn't that going. Oh yeah- and this is this- is deep. It's just a question of yeah.

G

I think I think clone it does. A name is problematic. We're calling it is persistable. It should just be persist.

A

Yeah, the question is like to the original thing that crystal brought up. Those is that a destructive operation in the original instance.

H

A

This is kind of implies, decided.

H

We can get rid of the bully and we can say if you want to store it call clone and be smart about. When you want to do some things, you just can't ever prove there. The thing is it's something of another thing: yeah.

A

I still understand how the property helps. You move the move, the problem, though, because I mean like: if it's, if it is disposable, then it means I have to copy 100 megabytes worst case, which it may not need right. If.

H

It is disposable and you have you're looking at a thing, and you want to maintain this state. You have to copy whatever this thing right.

A

So that means so, it seems like the problem is that clone may copy more than you needed, but you can always run into because of whether you create a property or not. If.

R

Somebody possibly is relevant. Yes, it's already in an expositional state by what I call true, no clone should be no of that to just return itself.

A

And this thing's internal ID and then clones implementation says they persisted. We've returned this otherwise good copy. We.

H

Can do that! It just means that that if you wanted to shake off all the outer portions of an object, because you now scope down into a smaller review that you have to do a lot more work to do that. That's fine sure.

G

What how do you, how do you know if there are too many states here to be represented by a single boolean.

H

G

So there's always going to be, but what if you wanted to like I mean once you call close like if you're operating on a cloned thing is persistable is going to return? True correct. Yes,.

N

G

So but you don't know at that point whether it's because you are part of a larger clone thing or because you were your own cloned thing right.

H

G

True, actually, so, then, what it doesn't really help you yeah.

H

So we could just do clone and the answer is: if you want to store something called clone or have knowledge that it's okay for you just yes,.

A

Or just make all renewal, then the ID other feature later well.

H

That's that, if we can say the clone is no op. If it's yet and then we can do it like, we can make a minimal clone later. First off, like yeah.

P

Oh, so someone will not to copy the whole thing well copy. If it's, if pretend we have this.

H

Persistable, if this persistable is false, clone will copy exactly what it needs to. If it's persistable is true, it says, return this so multiple objects from the same hundred biked a load I'll keep the same hundred Bank a little odd. If you only took one thing, I know how to make Talos it's a very good, but why are you persistently? We should only persist a ring clearer than that.

H

H

And then there's still solves the. What I understand the serializers need is. It wants to be able to create the document return, an element, it's not bound by the story, so.

R

Why not on this call clone and always cream I might have a space to store each other.

A

Like if you have evidence for um that, you need this, then you can edit later. If we have to for now, I would just leave it at home and then do trim later. If it actually shows up, I mean it's a general problem that we have like if you, if you store an object reference, and that is a potentially large tweeted, your whole alive, like I, mean it said. This is not specific to this.

R

Because I spa, the problem is, if, let's say, you're, storing multiple objects, multiple gesture elements. Well, you have multiple copies around yeah each clone, but I well,.

H

P

Only told each of them I.

H

Think he was recommendation was clone, says if it's already detached from the disposed. It returns this and if we have mean for the small version later, then you had a numerical trim. Yes, okay, I think I'm calling for him and you call what everyone you care about. I.

A

Think, like I mean there was a similar problem like that, like I, think it was the job of whole numbers doing stop stringing by basically doing expand use. It's like you, keep a four character string around, but you leave division from a flow gig document.

A

What the quality document is not a lot for ever I'm like it's off to swallow, my saying: well, maybe we shouldn't do that in the game, but you could also imagine the smarter system that just a text said and then basically dynamically compacts the string and remove the other one, but if nobody else holds it alive but like in our case like we yeah, you could imagine that in this case, I would insert most of the time somebody up the chain knows what's going on and you carve out like a subtree from the document first and then then you just say: yeah each of the notes of Philidor.

A

Can we can only keep the darking in their own, then that's fine, and then you just keep that guy around the.

H

Biggest internal use that we have was always going to be cloning on the development and so realize you're doing it some subpar space, not the existing thing called scrolling and amidst that is the I didn't know what to do with this yeah.

A

I mean this: this.

G

Replaces all the detach stuff that was put up as PR previously right? Yes,.

H

Okay, the PR will get a lot smaller now.

H

Actually, probably keep all the stuff that I had, it was just all switched in to internal and then then the test still get small.

A

We did well alright. Thank you guys. Then I see you tomorrow.