Inference issues with broadcasting

[tshort]

I'm not sure if this is a repeat.

julia> function gbitarray(i)
           a = 1.0:0.1:10.0
           @inbounds i > 3 ? a[(a.>3.0) .& (a .< 5.0)][1] : a[5] 
       end

julia> using Cassette: @context, overdub

julia> using Test

julia> @inferred(gbitarray(4))
3.1

julia> @context Ctx;

julia> @inferred(overdub(Ctx(), gbitarray, 4))
ERROR: return type Float64 does not match inferred return type Any
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] top-level scope at REPL[7]:1

The following function infers fine:

function gbitarray1(i)
    a = 1.0:0.1:10.0
    @inbounds i > 3 ? a[a.>3.0][1] : a[5] 
end

I tried it on Cassette master and v0.2.6. This is on Julia v1.3.0-rc3.0 on Linux. I'm not sure how to debug inference issues.

[tshort]

Defining the following seems to make it infer (I think). That's based on a hint in another inference issue.

@inline Cassette.overdub(ctx::Ctx, f, args...) = f(args...)

[tshort]

But, does that bypass the overdubbing?

[vchuravy]

That certainly bypasses overdubbing, for inference failures I recommend using Cthulhu to understand and isolate the failure.

…

On Fri, Oct 18, 2019, 18:38 Tom Short ***@***.***> wrote: But, does that bypass the overdubbing? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#153>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABDO2WJYCYMLJN7GZTNRRDQPI3IHANCNFSM4JCLOTOQ> .

[tshort]

I think it's a broadcasting operation causing the issue. The example in #56 also fails for me, so there may have been an inference regression on this.

[tshort]

Here's a simpler function where inference fails:

@inferred Cassette.overdub(Ctx(), x -> (x.+x)[1], [8,9,99])

[tshort]

I can follow the path of inference failures with Cthulhu, but I don't know why it's failing.

[tshort]

There is a broadcast test, and it is failing for me.

[tshort]

These seem to infer fine on Julia v1.2.0, so something changed with broadcasting since then.

[vchuravy]

Thanks I will have a look at it.

[tshort]

The following infers with Julia v1.2.0 and v1.4.0-DEV.666 (2019-12-26), but in doesn't infer with Julia v1.3.0.

using Cassette: @context, overdub
using Test
@context Ctx;
@inferred overdub(Ctx(), x -> (x.+x)[1], [8,9,99])

Currently, Cassette's tests don't pass on Julia v1.3.0. Does that mean that "full" Cassette support will skip v1.3.0? I mainly ask because its an unusual situation where Cassette is at the mercy of base-Julia changes.

Edit: this also fails to infer in Julia v1.3.1.

[oxinabox]

Can this issue be retitled sinmce as the MWEs demonstrate this is not restricted to bitarry ?

[oxinabox]

The following infers with Julia v1.2.0 and v1.4.0-DEV.666 (2019-12-26), but in doesn't infer with Julia v1.3.0.

for me it fails to infer in:
Version 1.4.0-DEV.713 (2020-01-22) and in 1.5.0-DEV.139 (2020-01-22)
(as well as 1.3, but it is fine in 1.2)

---

We can even go slightly simpler:

using Cassette: @context, overdub
using Test
@context Ctx;
@inferred overdub(Ctx(), x -> x.+x, [99])

[vchuravy]

We can even go slightly simpler:

That passes for me on (on Cassette master):

Julia Version 1.3.1
Commit 2d5741174c (2019-12-30 21:36 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-7660U CPU @ 2.50GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, skylake)

[oxinabox]

That passes for me on (on Cassette master):
Thats weird. weird. weird.

Here are my version infos:

Julia Version 1.2.1-pre.0
Commit 93929550b6 (2019-08-20 05:33 UTC)

Julia Version 1.3.2-pre.0
Commit 2e6715c045 (2019-12-31 00:49 UTC)

Julia Version 1.4.0-DEV.713
Commit 8852200b83 (2020-01-22 12:38 UTC)

Julia Version 1.5.0-DEV.139
Commit c2abaeedf8* (2020-01-22 06:54 UTC)

Platform Info:
  OS: macOS (x86_64-apple-darwin19.0.0)
  CPU: Intel(R) Core(TM) i7-8559U CPU @ 2.70GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, skylake)

(12) pkg> st --manifest
Status `~/temp/12/Manifest.toml`
  [7057c7e9] Cassette v0.3.1 #master (https://github.com/jrevels/Cassette.jl.git)

Possibilities:

1. Weird Mac OS interaction?
2. Something to do with me building from source?
3. Ghosts.

[oxinabox]

Digging down into it

Something close to metal that fails

using Test
using Cassette: @context, overdub, disablehooks
@context Ctx;
const ctx = disablehooks(Ctx())  # disabling hooks makes code_typed shorter to read

@inferred overdub(ctx, Base.Broadcast.instantiate, Base.broadcasted(+, [99],[99]))

And if we look at the @code_warntype (on Julia 1.5-DEV)
This seems to be where thigns fo wrong

3 ┄ %17 = Cassette.overdub(##overdub_context#253, Core.apply_type, Base.Broadcast.Broadcasted, $(QuoteNode(Base.Broadcast.DefaultArrayStyle{1})))::Core.Compiler.Const(Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1},Axes,F,Args} where Args<:Tuple where F where Axes, false)
│   %18 = Cassette.overdub(##overdub_context#253, Base.getproperty, bc, :f)::Core.Compiler.Const(+, false)
│   %19 = Cassette.overdub(##overdub_context#253, Base.getproperty, bc, :args)::Tuple{Array{Int64,1},Array{Int64,1}}
│   %20 = Cassette.overdub(##overdub_context#253, %17, %18, %19, axes)::Any

Whereas in Julia 1.2
It is:

3 ┄ %19 = Cassette.overdub(##overdub_context#371, Core.apply_type, Base.Broadcast.Broadcasted, $(QuoteNode(Base.Broadcast.DefaultArrayStyle{1})))::Core.Compiler.Const(Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1},Axes,F,Args} where Args<:Tuple where F where Axes, false)
│   %20 = Cassette.overdub(##overdub_context#371, Base.getproperty, bc, :f)::Core.Compiler.Const(+, false)
│   %21 = Cassette.overdub(##overdub_context#371, Base.getproperty, bc, :args)::Tuple{Array{Int64,1},Array{Int64,1}}
│   %22 = Cassette.overdub(##overdub_context#371, %19, %20, %21, axes)::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1},Tuple{Base.OneTo{Int64}},typeof(+),Tuple{Array{Int64,1},Array{Int64,1}}}
└──       return %22

---

[vchuravy]

Digging into that with Cthulhu shows that a failure to infer print_to_string which I have noticed elsewhere leads to a Base.throw being inferred to return Any, thus polluting the return type of bcs. This reminded me that I started looking at it #166

[vchuravy]

@tshort can you test that this helps your use-case?

[tshort]

With Julia v1.3.1 and Cassette master, the following both infer fine now!

• gbitarray() from the first post
• overdub(Ctx(), x -> (x.+x)[1], [8,9,99])

[tshort]

@oxinabox, does that fix your tests, too?

[oxinabox]

Working on 1.3. 🎉
Still fails on 1.4 and 1.5