Adds interpolation syntax to `Threads.@spawn`, to evaluate arguments immediately #33119

NHDaly · 2019-08-30T04:13:26Z

Adds $-interpolation syntax to @async and Threads.@spawn, to evaluate arguments immediately.

Add the ability to evaluate some parts of a @spawn/@async
immediately, in the current thread context.

This prevents variables being "boxed" in order to capture them in the
closure, exactly the same as wrapping them in a let-block locally.

For example, $x expands like this:

julia> @macroexpand @async $x + 2
quote
    #= task.jl:361 =#
    let var"##454" = x
        #= task.jl:362 =#
        local var"#9#task" = Base.Task((()->begin
                            #= task.jl:358 =#
                            var"##454" + 2
                        end))
        #= task.jl:363 =#
        if $(Expr(:islocal, Symbol("##sync#95")))
            #= task.jl:364 =#
            Base.push!(var"##sync#95", var"#9#task")
        end
        #= task.jl:366 =#
        Base.schedule(var"#9#task")
        #= task.jl:367 =#
        var"#9#task"
    end
end

Fixes #30896

Add the ability to evaluate some parts of a @spawn immediately, in the current thread context.

~~Adds a new macro, Threads.@spawncall, which evaluates its arguments immediately, then @spawns the closure with the argument values.~~

This acts more like a "parallel function call", where the arguments are evaluated immediately, in the current thread, and the function is then invoked in any available thread with the stored argument values.

EDIT: Updated to add $-interpolation syntax support to @spawn to allow you to manually specify which parts of the expression to evaluate early.

This prevents variables being "boxed" when capturing them in the closure.

~~This mirrors the behavior of golangs go command, which requires it be used with a function call.~~

Fixes #30896

NHDaly · 2019-08-30T04:19:38Z

(If we move forward with this PR, we'll probably want to add an equivalent for @async. I don't love the name @asynccall because the consecutive 'c's make it hard to read. There's @async_call but it's a bit ugly...)

vchuravy · 2019-08-30T07:35:35Z

I am not the biggest fan of adding a new name/verb with subtlety different behavior.

I liked Jeff's idea of passing arguments to be evaluated, which seems similar to let

NHDaly · 2019-08-30T14:58:24Z

I am not the biggest fan of adding a new name/verb with subtlety different behavior.

Yeah, i agree, i think that's a reasonable concern. Thanks.

I liked Jeff's idea of passing arguments to be evaluated, which seems similar to let

My concerns with passing arguments to be evaluated are two-fold:

If we require users to specify which arguments to evaluate, it's not functionally much different than just having users use a let block. My main motivation for introducing something like this is to enable users to super easily write multithreaded function calls (as easy as it is in go) without having to understand the intricacies of scoping and closure capture. I think that if we had a mechanism that was specific to function calls, and we encourage its use, then people will just know "oh i call functions with this verb, and I spawn arbitrary blocks of code with this other verb." And then their code will just work as expected without having to understand why sometimes variables are boxed.
I think the syntax is strange and non-obvious. If I just read @async i j k println(i, j, k), i would definitely not understand what the first few arguments were doing there, and i wouldn't know when I was supposed to put them there or not.
- Maybe this could be improved if it was a more explicit kwarg? Something like @async capture=[i,j,k] println(i, j, k) or something?

But yeah, i do agree that it could be non-ideal to have these two verbs that are pretty similar yet subtly different...

What if this also had a different syntax, not just a different name? That could help distinguish them better?
Maybe something silly like:

function spawnfunc(f)
    (args...; kwargs...) -> @spawn f(args...; kwargs...)
end

julia> spawnfunc(println)(1)
1
Task (done) @0x0000000114d958d0

julia> fetch(spawnfunc(sort)([3 2; 1 0], dims=2))
2×2 Array{Int64,2}:
 2  3
 0  1

EDIT: One problem with doing it this way, I think, is that enclosing @sync blocks stop working with the @spawn, because the ##sync#72 variable is out of scope.. It would probably work if spawnfunc were a macro as well, something like @spawnfunc(f)(args...) but i'd need to think about it more.

NHDaly · 2019-08-30T15:03:14Z

Maybe this could be improved if it was a more explicit kwarg? Something like @async capture=[i,j,k] println(i, j, k) or something?

... Or, as another idea, maybe we should just use $ for this? It's more or less consistent with the concept of interpolation, and basically matches its use in other contexts...?

So it'd be something like @async println($i, $j, $k). Which actually seems not terrible to me? Anything being interpolated is evaluated in the current Task, and anything not interpolated is evaluated as part of the closure, in the spawned Task.

quinnj · 2019-08-30T15:11:00Z

I love the interpolation idea.

NHDaly · 2019-08-30T15:20:59Z

It's growing on me as well.

Oh, haha, i just realized i already suggested it back in January as well, in the original issue... haha it's funny how our brains are little automata that respond the same way to the same situations... 😅😬

#30896 (comment)

Here's what I said then:

We could consider allowing @async to use the $ interpolation syntax within its expression body to indicate pass-by-value vs pass-by-reference semantics. This would be kind of like what @btime does.
In my example, that would become:
    @async println($i)
Sadly this might be breaking, though, for any expressions that already contained quoting and interpolation inside the @async body. 😦 (e.g. @async println(:($x + 5))

I'm less worried about that caveat now than I was then.. I think all the cases where this would start to take affect would currently be syntax: "$" expression outside quote errors, so this should be available syntax! :) I think it would be fairly straightforward to parse the $ -- it'd be the same as is done for @eval, and I don't think there'd be any ambiguous cases.

While i'm here, i'll copy over the other references from the original issue:

@KristofferC erC's issue about this from 2016: Explicit capture of variables in a closure. #14959 (comment)
This discourse thread explaining why the boxing happens: https://discourse.julialang.org/t/surprising-capture-boxing-behavior-in-closure/20254

vchuravy · 2019-08-30T22:56:39Z

IIUC the semantics of Cilk and Go agree here, they both use spawn/go on function calls and evaluate the arguments of the function call before spawning. https://www.cilkplus.org/sites/default/files/open_specifications/Intel_Cilk_plus_lang_spec_1.2.htm#keywd.spawn

The general closure syntax of @async/@spawn makes this harder, but I rather like the pun upon the interpolation syntax (similar to BenchmarkTools).

NHDaly · 2019-09-10T17:46:16Z

Since there haven't been any further comments on this PR, i've gone ahead and implemented the $-interpolation approach.

It seems pretty good to me! And i am 95% sure it is a non-breaking change, since it's only weakening the pre-conditions on valid input. :)

The original example now works, via this interpolation:

    begin
        outs = zeros(5)
        @sync begin
            local i = 1
            while i <= 5
                Threads.@spawn setindex!(outs, $i, $i)
                i += 1
            end
        end
        @test outs == 1:5
    end

If we're happy with this, we can move the implementation to a shared common location, and use it in @async as well. (and clean up the implementation.) :)

NHDaly · 2019-09-10T17:53:03Z

I guess one very real problem with the current approach is that the $ has a slightly different meaning than it has in some other macro contexts, since we are simply moving the variables to be evaluated earlier, rather than actually evaling the interpolated expression and splicing the result in, in-place:

julia> @eval $(Expr(:call, :+, x, x))
4

julia> fetch(@spawn $(Expr(:call, :+, x, x)))
:(2 + 2)

julia> @btime $(Expr(:call, :+, x, x))
  0.036 ns (0 allocations: 0 bytes)
4

I'm afraid this might be confusing. EDIT: Though maybe not... See below:

That said, after thinking about it more, I guess actually the behavior is consistent with the meaning of $-interpolation.. It's just that both @eval and @btime promise to eval the resultant expression, whereas @spawn does not, which is why the Expression persists unchanged. This is the same as you'd get when interpolating into a string or a quote, for example:

julia> "$(Expr(:call, :+, x, x))"
"2 + 2"

julia> :($(Expr(:call, :+, x, x)))
:(2 + 2)

So maybe there's nothing wrong here after-all? It does seem a bit mysterious though, because people might be used to interpolating into an @eval call.

StefanKarpinski · 2019-09-30T22:01:48Z

Bump, this seems like a really nice idea. cc @JeffBezanson, @vtjnash

vchuravy

Looks good to me! Just a bit of polish and adding news and I think this is good to go.

vchuravy · 2019-10-01T17:37:45Z

base/threadingconstructs.jl

@@ -103,22 +103,28 @@ Create and run a [`Task`](@ref) on any available thread. To wait for the task to
 finish, call [`wait`](@ref) on the result of this macro, or call [`fetch`](@ref)
 to wait and then obtain its return value.

+(Values can be interpolated into `@spawn` via `\$` to evaluate them in the current task.)


Should be a !!! compat note

Done, thanks. Is it okay that I added the compat note below like this, or should it go above the section on $-interpolation? Should I move that paragraph below the compat notes?

base/task.jl

base/expr.jl

test/threads_exec.jl

NHDaly

Thanks for the review, @vchuravy, and sorry for the delay! :)

EDIT: also added NEWS. (Unclear why the macos test is failing, but seems unrelated?)

base/expr.jl

NHDaly · 2019-10-23T13:45:33Z

base/threadingconstructs.jl

@@ -103,22 +103,28 @@ Create and run a [`Task`](@ref) on any available thread. To wait for the task to
 finish, call [`wait`](@ref) on the result of this macro, or call [`fetch`](@ref)
 to wait and then obtain its return value.

+(Values can be interpolated into `@spawn` via `\$` to evaluate them in the current task.)


Done, thanks. Is it okay that I added the compat note below like this, or should it go above the section on $-interpolation? Should I move that paragraph below the compat notes?

base/task.jl

test/threads_exec.jl

NHDaly · 2019-12-15T03:30:34Z

Friendly ping. I think this was mostly ready to go? :)

vchuravy · 2019-12-15T19:17:29Z

Can you squash and rebase?

JeffBezanson · 2019-12-16T19:51:58Z

Yes, let's get this rebased and merged.

This acts more like a "parallel function call", where the arguments are evaluated immediately, in the current thread, and the function is then invoked in any available thread with the stored argument values. This prevents variables being "boxed" in order to capture them in the closure.

This acts more like a "parallel function call", where the arguments are evaluated immediately, in the current thread, and the function is then invoked in any available thread with the stored argument values. This prevents variables being "boxed" in order to capture them in the closure. Fix at-spawncall to work w/ kwargs; add unit tests Rough implementation of `$`-interpolation in at-spawn

This acts more like a "parallel function call", where the arguments are evaluated immediately, in the current thread, and the function is then invoked in any available thread with the stored argument values. This prevents variables being "boxed" in order to capture them in the closure. Fix at-spawncall to work w/ kwargs; add unit tests Rough implementation of `$`-interpolation in at-spawn Update spawncall tests to exercise the interpolation Fix my basic "lift $" function for $ inside :() Add string-interp test Escape the `$` reference in at-spawn docstring Oops, remove spurious "end" at end of file Factor out `$`-lifting; share b/w `@async` & `@spawn`

NHDaly · 2019-12-18T05:40:49Z

I'm on vacation in Kerala, India (it's gorgeous here!! 😁) But the internet is pretty crappy AND I only brought my crappy Chromebook. I just rebased and pushed. It's building locally now, but it takes about 20 min to download and build 😅, and we just had to leave to go shopping. I'll check again when I'm back. Thanks for looking at this! Hope it can make it into 1.4

…

On Tue, Dec 17, 2019 at 1:22 AM Jeff Bezanson ***@***.***> wrote: Yes, let's get this rebased and merged. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#33119?email_source=notifications&email_token=AAMCIEPMRW3TWY4P6Q3ZE2DQY7L6BA5CNFSM4ISJBH42YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEG74JTQ#issuecomment-566215886>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAMCIEL2XMEEIFVMUXXFK7LQY7L6BANCNFSM4ISJBH4Q> .

@macroexpand

…uate arguments immediately. Add the ability to evaluate some parts of a `@spawn`/`@async` immediately, in the current thread context. This prevents variables being "boxed" in order to capture them in the closure, exactly the same as wrapping them in a let-block locally. For example, `$x` expands like this: ```julia julia> @macroexpand @async $x + 2 quote #= task.jl:361 =# let var"#JuliaLang#454" = x #= task.jl:362 =# local var"JuliaLang#9#task" = Base.Task((()->begin #= task.jl:358 =# var"#JuliaLang#454" + 2 end)) #= task.jl:363 =# if $(Expr(:islocal, Symbol("##sync#95"))) #= task.jl:364 =# Base.push!(var"##sync#95", var"JuliaLang#9#task") end #= task.jl:366 =# Base.schedule(var"JuliaLang#9#task") #= task.jl:367 =# var"JuliaLang#9#task" end end ```

NHDaly · 2019-12-18T13:49:31Z

Okay looks good! :) It looks like the windows test failure is a flake? Thanks for the review

Introduce synchronization (via a `Channel()`) to force spawned tasks to run after the local variables are updated, showcasing the problem and the solution with `$`-interpolation.

@macroexpand

…immediately (#33119) Adds $-interpolation syntax to `@async` and `Threads.@spawn`, to evaluate arguments immediately. Add the ability to evaluate some parts of a `@spawn`/`@async` immediately, in the current thread context. This prevents variables being "boxed" in order to capture them in the closure, exactly the same as wrapping them in a let-block locally. For example, `$x` expands like this: ```julia julia> @macroexpand @async $x + 2 quote #= task.jl:361 =# let var"##454" = x #= task.jl:362 =# local var"#9#task" = Base.Task((()->begin #= task.jl:358 =# var"##454" + 2 end)) #= task.jl:363 =# if $(Expr(:islocal, Symbol("##sync#95"))) #= task.jl:364 =# Base.push!(var"##sync#95", var"#9#task") end #= task.jl:366 =# Base.schedule(var"#9#task") #= task.jl:367 =# var"#9#task" end end ```

Introduce synchronization (via a `Channel()`) to force spawned tasks to run after the local variables are updated, showcasing the problem and the solution with `$`-interpolation.

bramtayl · 2020-10-19T22:11:06Z

I wonder if this could be made more general? Perhaps marking any closure with $ could essentially "freeze" it? It would be a much nicer syntax than the let-block trick currently used to get around #15276

bramtayl · 2020-10-19T22:15:46Z

As a bonus, $ currently errors at top-level, so it could be added pre-2.0

NHDaly mentioned this pull request Aug 30, 2019

Boxed variables in closures passed to a Task can cause race condition #30896

Closed

vchuravy changed the title ~~Adds Threads.@spawncall, which evaluates arguments immediately~~ Adds interpolation syntax to Threads.@spawn, to evaluate arguments immediately Sep 11, 2019

fredrikekre added the domain:multithreading Base.Threads and related functionality label Oct 1, 2019

vchuravy reviewed Oct 1, 2019

View reviewed changes

fredrikekre added the needs news A NEWS entry is required for this change label Oct 2, 2019

NHDaly commented Oct 23, 2019

View reviewed changes

JeffBezanson approved these changes Dec 16, 2019

View reviewed changes

JeffBezanson removed the needs news A NEWS entry is required for this change label Dec 16, 2019

NHDaly added 3 commits December 18, 2019 10:02

NHDaly force-pushed the spawncall branch from 8c3a488 to 036ef18 Compare December 18, 2019 05:31

NHDaly force-pushed the spawncall branch from 036ef18 to f275297 Compare December 18, 2019 08:42

JeffBezanson merged commit 0df3fe7 into JuliaLang:master Dec 18, 2019

tkf mentioned this pull request Dec 18, 2019

Interpolation syntax in Threads.@spawn breaks $-aware macros #34138

Closed

NHDaly deleted the spawncall branch December 19, 2019 06:00

NHDaly mentioned this pull request Dec 19, 2019

Race condition caused by variable scope getting lifted from a multithreaded context #14948

Open

Keno mentioned this pull request Dec 19, 2019

Race condition in new threading test #34141

Closed

NHDaly mentioned this pull request Dec 19, 2019

Fix race condition in test introduced in #33119 (oops). #34142

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adds interpolation syntax to `Threads.@spawn`, to evaluate arguments immediately #33119

Adds interpolation syntax to `Threads.@spawn`, to evaluate arguments immediately #33119

NHDaly commented Aug 30, 2019 •

edited

Loading

NHDaly commented Aug 30, 2019

vchuravy commented Aug 30, 2019

NHDaly commented Aug 30, 2019 •

edited

Loading

NHDaly commented Aug 30, 2019

quinnj commented Aug 30, 2019

NHDaly commented Aug 30, 2019

vchuravy commented Aug 30, 2019

NHDaly commented Sep 10, 2019

NHDaly commented Sep 10, 2019 •

edited

Loading

StefanKarpinski commented Sep 30, 2019

vchuravy left a comment

vchuravy Oct 1, 2019

NHDaly Oct 23, 2019

NHDaly left a comment •

edited

Loading

NHDaly Oct 23, 2019

NHDaly commented Dec 15, 2019

vchuravy commented Dec 15, 2019

JeffBezanson commented Dec 16, 2019

NHDaly commented Dec 18, 2019 via email

NHDaly commented Dec 18, 2019

bramtayl commented Oct 19, 2020

bramtayl commented Oct 19, 2020

Adds interpolation syntax to Threads.@spawn, to evaluate arguments immediately #33119

Adds interpolation syntax to Threads.@spawn, to evaluate arguments immediately #33119

Conversation

NHDaly commented Aug 30, 2019 • edited Loading

NHDaly commented Aug 30, 2019

vchuravy commented Aug 30, 2019

NHDaly commented Aug 30, 2019 • edited Loading

NHDaly commented Aug 30, 2019

quinnj commented Aug 30, 2019

NHDaly commented Aug 30, 2019

vchuravy commented Aug 30, 2019

NHDaly commented Sep 10, 2019

NHDaly commented Sep 10, 2019 • edited Loading

StefanKarpinski commented Sep 30, 2019

vchuravy left a comment

Choose a reason for hiding this comment

vchuravy Oct 1, 2019

Choose a reason for hiding this comment

NHDaly Oct 23, 2019

Choose a reason for hiding this comment

NHDaly left a comment • edited Loading

Choose a reason for hiding this comment

NHDaly Oct 23, 2019

Choose a reason for hiding this comment

NHDaly commented Dec 15, 2019

vchuravy commented Dec 15, 2019

JeffBezanson commented Dec 16, 2019

NHDaly commented Dec 18, 2019 via email

NHDaly commented Dec 18, 2019

bramtayl commented Oct 19, 2020

bramtayl commented Oct 19, 2020

Adds interpolation syntax to `Threads.@spawn`, to evaluate arguments immediately #33119

Adds interpolation syntax to `Threads.@spawn`, to evaluate arguments immediately #33119

NHDaly commented Aug 30, 2019 •

edited

Loading

NHDaly commented Aug 30, 2019 •

edited

Loading

NHDaly commented Sep 10, 2019 •

edited

Loading

NHDaly left a comment •

edited

Loading