Rust Programming Language Incremental Compilation Working Group, 22 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: incr-comp meeting #3 (2020-09-22)

Description

Third meeting, at a new time slot. Here is the agenda/notes we used during the meeting: https://hackmd.io/3B-HlrkKQ3KIsv9JjYENPA

A

Okay, cool um all right, so I guess the first thing on the agenda is that um I added some work items to our ongoing projects tracking issue. It looks like everybody's aware of that, because people started to pick work off of that, but if there's anything on there you're interested in um either, I can probably answer some questions about it or felix can, or we can certainly point you to somebody who can so um definitely get in touch, and let us know if you want to work on any of that stuff.

A

um Santiago. Did you want to talk about your pr or.

B

Talk about that, I don't know like if you want, would be easier, probably for you to explain this.

A

Okay, um so santiago's been working on one of the linked issues to reduce how much in line how many inline function, instantiations we're getting in debug builds because we don't actually run the inliner or any of the optimizations that need multiple copies in the same crate like per cgu, um so we're just we're just adding compilation time there. So um the overall issue is actually to do some additional work beyond.

A

What's been done now to um to generate a copy of the inline function in the upstream crate, even in release mode so that we can link against it potentially in a downstream crate. um That hasn't been done yet. But what santiago has done is push up a pr that changes um our instantiation strategy, basically for inline functions and debug mode. So now we get just one copy per crate. Instead of potentially one copy per cgu that uses it um and we did a perf run on that and the results were a little mixed.

A

um There were some regressions up to like almost six percent on um some of the incremental tests, but we also got a 35 improvement on full non-incremental builds for regex and I think a 13 improvement on the full build for encoding in debug mode, which uh seemed like pretty significant wins to me. um So I'm feeling like we should probably land this, because this seems like the correct strategy. Even if there's some crates or some situations that this doesn't.

B

Help so, as far as I can tell this, this only changes when, when in line is always the rest, is kind of the same behavior we had before. The only thing that I was wondering is uh related to like, if, if you expand a bit the code above of of the changes that are in the in the part that it says in line always on all that we were, we were using, um I mean we are using this in line. You know gg use option.

B

uh I guess if some, if somebody pass past that flag as true or something like that, we should always because this is what what the user is is really suggesting right. It's asking you to do so in that case like um well. If someone sets that to true like we would be, I mean what what I don't know.

B

This is like how the code previous previously behaved, but um I don't know why, like um in the in the top part of the of the code when right before the mono item fn arm like we are, we are doing some checks there and returning instantial mode globally, shared regardless. If that is true or not so I I I mean, I don't know, if that's what the code should be doing, but uh it's a preexisting thing. What I'm mentioning.

A

B

But but it's in some cases it seems to me that that we would we would want if in line you know, c use is true, just like returning sensation mode like a copy or something like. That is what what a good guess we should be doing, but.

B

Or even just switch like the, I don't know it's it's hard to explain what code I'm referring to like in the air, but.

A

B

I think what you're saying.

A

Is maybe we could um reorganize that branch there? Because I think, if generating something.

B

A

Copies is true. We probably want to.

B

Do local copy all right, so so what I'm saying is to move the line, the code we have from line 95 and switch it to to what we have in line 105 like to switch those things. So we can just return quickly if generates like uh if in line in all cgs is true, because it seems that that's what we want, but uh I don't know like. Maybe anyway, we can continue with this. I think.

A

Yeah, that might be a good cleanup, um I'm a little curious. What generates cgu internal copies on the def. Does I'm not sure it's the same as our variable up top generate cgu internal copies?

A

um It may have something to do with like what annotations are applied to the specific item or what kind of item it is or something. So I'm not I'm not sure offhand if we can do that, but that does seem like a reasonable cleanup.

C

B

David, if I understood you correctly, when we were um talking about follow-up uh things to do with this, what we really want is like right now um like right now, when we, when we don't when we are not going to inline our code, we want just to like reuse or refer in some way to what what another trade have generated right, but um what I wonder in the actual code, what we want is is to not generate this instantiation mode copies.

B

um It's just that and referring some other way what what the other crate had or.

A

Yeah, so um I think what we want to do to actually close out this ticket entirely is in debug mode um well, so backing up for the upstream crate. We need to actually generate a copy of that function in that crate, so that we can link against it.

B

A

And then downstream in debug mode, instead of generating any copies of that function in our crate in any of our cgus, we just want to link against the upstream copy, but right now, there's no upstream copy at all, because it's marked in line so uh felix we're talking about the pr santiago opened for 76896.

A

And the tl, dr, is that basically it changes the instantiation mode for inline always functions in debug mode. It's so that, instead of generating a copy per cg that uses it, we generate one copy in our crate and reuse it across all cg use because, obviously we're not running optimizations in llvm. So it doesn't matter.

A

um We did a perf run. There's some regressions, mostly on different incremental performance tests up to six percent. In some cases, there's also a 35 improvement on the regex debug, build in the full case and a 13 improvement in the encoding debug full case.

C

As well so when you say inline all so, you might always I'm assuming there's no even no language guarantee there that we actually in mind, but I don't know for sure, even though I just remember offhand, that's what we're talking about right. It's like actually treating it like it's, not truly, always yeah, okay, okay,.

A

I I'm not sure if there's documentation to this effect, what I've always heard is that all of the inline attributes or suggestions not guaranteed.

C

Simply without always, certainly without always it's a suggestion, that's just you know, what's the way up the compiler, but I'm curious, but always now okay. This is interesting.

A

um So my feeling was that we should probably maybe there's some tweaks we can do or something to this somewhere to recoup some of the lost performance on some of the incremental tests, um or maybe we should dig into those some more and figure out exactly what's going on there. But um a 35 improvement on an actual crate in non-incremental seems pretty significant and the change makes sense. Logically, so we should probably look at taking it.

B

So um wesley, like the like high level, like I understood what you said uh like what we want to accomplish, but um it seemed to me like I, I I'm not sure, because I don't like know exactly the code what's going on, um but uh it seemed to me from what I from what I have test that uh just by by doing this change, we are like. uh We are really um doing what what what you are saying like I mean um from the from the main crate.

B

We are making a copy and we are referring from the other one, but I I don't know I may be a bit confused but um like it seemed to me that we are really with this. We are really doing what what we are, what the the ticket like asked to do. So I I'm not really sure um like. Maybe I.

A

Think we've got, I think, we've got a good chunk of it, um so part of it is just generating fewer copies and part of it is generating copies upstream so that we don't have to generate any copies in our crate at all, because right now we're still generating one copy that we're sharing across all of our cgus.

A

um And if we look at the ticket.

A

uh Michael said, we could probably make use of this dash share generics, infrastructure for making inline functions available downstream um and then the the comment that actually spawned this ticket is kind of talking about that same idea where, if we actually generate a function in the upstream crate, we can link against it the downstream crate without any duplication at all there.

A

um So this is definitely like a big chunk of it, but there's there's probably still more work. That needs to be done there to close it out.

B

Okay, yeah: we can continue discussing a sync and with concrete like more concrete.

C

B

To the code and all that, like yeah.

A

um Okay, uh david, do you want to talk about um what you've been working on.

D

Sure I don't know, there's a massive amount to say, but I've been looking at the ticket to add split, dwarf support.

D

I expect to be a reasonable chunk of work to do there because it's different across platforms, but what I've done so far is look into what clang is doing so I've kind of written up all klang stuff, and I can share that in the the soul upstream um and see what it's doing finding out. What he's asking for split dwarf makes to cause lvm so that I can replicate that and what russia does.

D

So I've figured that out. I think I've now been looking at the our wrapper functions um through lvm. I've modified those, so that'll pass along information that we'll need, and I'm now adding a flag. So I can try and wire it all up and I'll see if that.

A

Works, um do you know like how how well supported is this like llvm? Do you think we're gonna get much platform support out of the box, or are we gonna have to like special case.

D

um I'm not sure um there was a couple.

D

um Kline was special casing it so that it would only do it when the target output, elf or wasm, um and then there was also some special casing for gnu and ming w, but that seemed slightly unrelated, even though it did use the same flags. So there's some. I want to look into it a bit more, but I think we'll get linux at least.

A

Okay, are we I use windows, but I don't really use windows for rust stuff, so I I'm speaking out of ignorance here, but at least in.net world split, like just out of the box from dlls.

A

Does anyone know if, like that's already happening on windows, or do we have to do something special to get split, debug to work on windows, regular windows, msvc,.

A

Okay, I'd be kind of interested if that's like already happening, because I that sort.

C

Of feels like the.

A

Platform default, but you just mean.

C

That the debug, the debug metadata, is in a separate file. Yeah, I'm pretty sure, that's what happens but yeah. We should.

A

Check all right cool, um so that's everything on the agenda. uh Is there anything else anybody wants to talk about.

B

C

Okay, we, oh obviously this will be in the meeting, but we there is one thing I was thinking about. um We probably I probably didn't I I know I didn't do enough to try to elicit um new volunteers, but something we probably should think about. Perhaps because we did see some attrition. um Obviously.

C

C

But there's nothing. I've been thinking about too hard. Obviously,.

A

um So now that I'm thinking about it, actually, I think, there's still a pr that we got cc'd on uh that. We should look at from uh check my.

A

A

Okay, yeah the lazy decoding of death path table during incremental compilation. Pr, I think, is waiting on us or somebody from the incremental compilation group to do a review. um I will post a link in.

D

D

A

Is there anything else anybody wants to talk about.

B

A

Okay, um so we are trying to hold scheduled meetings so assuming this time works for everyone, uh I guess we'll see everybody back here again in two weeks.

D

A

C

Okay, all right.

C