-
Notifications
You must be signed in to change notification settings - Fork 1
Meeting weekly 2012 03 06
Attending:
-
T: tjc
-
G: graydon
-
M: marijn
-
N: nmatsakis
-
B: brson
-
P: pcwalton
-
G: Should we schedule 0.2 release around features or by time
- a lot of detail work in the bug list
- e.g., hold it for regions, unwinding, etc?
-
P: I don't think we need to hold it on anything but if new features are done, let's take them.
-
G: I am seeing a fair number of people having trouble with FFI being underpowered
- inability to pass structures by value
- inability to do interior allocation
- inability to teach cargo to build C code (the ultimate glue)
-
P: We should fix FFI bugs having ability to build C is good anyhow
-
G: Trying to decide how to prioritize, the two issues seem independent
-
G: Giving rustc the ability to process C files... does that horrify anyone?
- basically by passing C files through to gcc
- T: that's what ghc does
-
B: There are open bugs and someone was interested in implementing it
-
G: Speed-related things queued up, in theory compiler should start to go faster
- in time-passes, main bottleneck is the LLVM code generation
- only way to attack it, basically, is to generate less code
- will try and instrument and see if we can easily generate less
- P: I think a lot is cleanups
- G: and maybe ref counting
- G: but maybe there is some low-hanging fruit
-
G: rustbot automation is crappy
- is it worth spending time improving it? adding features?
- N: going faster would be good
- G: partly that's just the cycle time of the compiler
- G: main thing I wanted to change:
- messaging structure is garbage, winds up doing multiple builds
- no prioritization
- P: try/chooser syntax. e.g., turn off valgrind, be more selective.
-
M: with regard to bug triage, if each of us is spending a day, question is where do we start
- last guy will have started at top of the bug list
- should we keep some information about where to look?
- is 5 days a wek overkill?
- G: it's possible. I've been starting at random in the list.
- P: I did not think you had to spend 8 hours doing bug triage, just enough to keep it under control
- G: No, but there is "bug-related" work that needs to get done
- there are plenty of actually legitimate bugs
- goal is to make sure that number of open bugs does not grow endlessly
- N: is this just a problem of not knowing what's not assigned?
- T,G: no, there are other scenarios too (duplicates, things that became easy)
- T: so shall we make a wiki page to coordinate this?
- G: there exists one: "bug-trackers"
-
T: issue #1906, suggestion of adding an infinite loop construct
- general consensus about "loop" as a keyword for infinite loop
-
M: the resolve pass might require a complete rewrite
- shall we open an issue to rewrite resolve?
- P: sounds good
- G: it's sufficiently complicated that it's hard to tell what it does
- P: also slow
- G: might be worth combining changes to export control (per-item
pub
) with rewrite- or is that too much bundling?
- concern is that a rewrite might yield a structure that doesn't fit?
- M: we should probably combine them if we decide to go forward with the bug
- P: what was final decision? default is crate-protected?
- G: some concern over how that interacts with tags?
- N: what about "descendants"?
- G: I was thinking priv would be "this module + its children"
- G: minor tidbits that need resolution:
- enum variants
- should enum variants inherit the visibility of their parent?
- e.g.,
pub enum foo { bar() priv qed() }
- P: that makes sense to me
- G: do we then generalize that rule? what about modules?
- T: if you don't want that you can use explicit exports
- N: I think not... enums are different because they project into the outer namespace
- G: do we need a keyword for crate-level visibility?
-
crate
as the keyword - P: could use
protected
.- misc: carries undesired heritage, long
- G:
local
?- misc:
local
is not very clear
- misc:
-
- enum variants