Sort out frule API

[willtebbutt]

Implements option 2 from #128 . Adds API regression tests to ensure that the considerations made there don't slip through the cracks in any future API revisions.

Currently got a weird type instability thing @oxinabox any thoughts?

[oxinabox]

Tell me more of your type-stability thing?

[willtebbutt]

Tell me more of your type-stability thing?

see e.g. here

The inference test here is failing to properly infer, I think. Unclear to me why it should be doing so.

For example:

using ChainRulesCore, StaticArrays
sx = @SVector [1, 2]
sy = @SVector [3, 4]

very_nice(x, y) = x + y
@scalar_rule(very_nice(x, y), (One(), One()))

@code_warntype frule((Zero(), sx, sy), very_nice, 1, 2)

Variables
  #self#::Core.Compiler.Const(ChainRulesCore.frule, false)
  @_2::Tuple{Zero,SArray{Tuple{2},Int64,1,2},SArray{Tuple{2},Int64,1,2}}
  #unused#::Core.Compiler.Const(very_nice, false)
  x::Int64
  y::Int64
  @_6::Int64
  Ω::Int64

Body::Tuple{Int64,Any}
1 ─ %1  = Base.indexed_iterate(@_2, 1)::Core.Compiler.Const((Zero(), 2), false)
│   %2  = Core.getfield(%1, 1)::Core.Compiler.Const(Zero(), false)
│         (ChainRulesCore._ = %2)
│         (@_6 = Core.getfield(%1, 2))
│   %5  = Base.indexed_iterate(@_2, 2, @_6::Core.Compiler.Const(2, false))::Core.Compiler.PartialStruct(Tuple{SArray{Tuple{2},Int64,1,2},Int64}, Any[SArray{Tuple{2},Int64,1,2}, Core.Compiler.Const(3, false)])
│   %6  = Core.getfield(%5, 1)::SArray{Tuple{2},Int64,1,2}
│         (ChainRulesCore.Δ1 = %6)
│         (@_6 = Core.getfield(%5, 2))
│   %9  = Base.indexed_iterate(@_2, 3, @_6::Core.Compiler.Const(3, false))::Core.Compiler.PartialStruct(Tuple{SArray{Tuple{2},Int64,1,2},Int64}, Any[SArray{Tuple{2},Int64,1,2}, Core.Compiler.Const(4, false)])
│   %10 = Core.getfield(%9, 1)::SArray{Tuple{2},Int64,1,2}
│         (ChainRulesCore.Δ2 = %10)
│         (Ω = Main.very_nice(x, y))
│   %13 = Ω::Int64
│   %14 = Base.broadcasted(Main.One)::Core.Compiler.Const(One(), false)
│   %15 = Base.broadcasted(Main.One)::Core.Compiler.Const(One(), false)
│   %16 = Base.broadcasted(ChainRulesCore.:*, %15, ChainRulesCore.Δ1)::Any
│   %17 = Base.broadcasted(muladd, %14, ChainRulesCore.Δ2, %16)::Any
│   %18 = Base.materialize(%17)::Any
│   %19 = Core.tuple(%13, %18)::Tuple{Int64,Any}
└──       return %19

[willtebbutt]

This is generally passing now. @oxinabox @nickrobinson251 what would you like to happen with the nightly failures?

[simeonschaub]

There's something weird going on with @allocated since JuliaLang/julia#33717. We might need to adjust our tests eventually, but I think it's okay to ignore failures on nightly for now

[willtebbutt]

@shashi @YingboMa does this change work for you?

[oxinabox]

Main documentation in /docs/src need to updated

[willtebbutt]

@nickrobinson251 @oxinabox shall we merge?

[nickrobinson251]

Would be good to see the partner PR to ChainRules

[nickrobinson251]

Seems good to me -- we just need to decide if we're releasing this as 0.7 (i don't seee why not) or keeping it at 0.7-DEV until the macros added too

Also, tests need to pass before merge

[willtebbutt]

Have tweaked the CI slightly to require that tests pass on 1.3. Will merge provided that everything seems fine.

[willtebbutt]

Before merging, @oxinabox are you happy with just calling this merge 0.7.0? Adding the convenience macros isn't breaking, so 🤷‍♂

[shashi]

Yeah it should work for ForwardDiff2. cc @YingboMa

[oxinabox]

Just fix to not use import and then LGTM