how to use @snoopi

[maartenvd]

Sorry if this is not the right place to post usage questions, but I'm really excited to try out the new snoopi macro (because inference is what I'm having problems with). However, when I try to snoopi on a call that takes 1+seconds to infer, I get the following trace:

 (4.100799560546875e-5, MethodInstance for eltype(::Type{Array{Expr,2}}))                                                                                                                                                                                         
...                                    
 (0.0013608932495117188, MethodInstance for methods(::Any, ::Any))                                                                                                                                                                                                
 (1.1331579685211182, MethodInstance for *(::StridedView{Float64,2,Array{Float64,1},typeof(identity)}, ::StridedView{Float64,2,Array{Float64,1},typeof(identity)}))   

The call I'm curious about is the multiplication, but also in how long all methods called by multiplication take. The problem is that the sum of all timings given by snoopi does not equal the time it takes to infer the multiplication call. There seems to be function calls missing, which perhaps got inlined, but even with --inline=no there is no change.

to be clear, I ran:

using SnoopCompile,Stridedc
a = StridedView(rand(5,5))
vs = @snoopi a*a

[timholy]

Usage questions are fine. It reports only the "top-level" method; if foo(::Int) calls bar(::Int), foo cannot be inferred without first inferring bar. But only the inference of foo gets reported.

It may help to look at how it works: there's nothing to intercept recursive calls, only the entry point.

If you can turn your confusion into improved documentation via a pull request, I'd be very grateful!

[maartenvd]

Oh, I assumed that by calling
jl_set_typeinf_func
inference would from now on use your supplied function to do everything, so therefore I expected to see a recursive trace.

I can take a look at the documentation :)

[timholy]

Oh, I assumed that by calling jl_set_typeinf_func inference would from now on use your supplied function to do everything

It does, but it only intercepts calls to typeinf_ext. That's the "entry point" from Julia's C side; the C code doesn't handle the recursion, that's all done by the Julia side from typeinf_ext. To intercept recursive calls we'd have to intercept abstract_call as well, and that involves more rewriting/overwriting the methods in the compiler. This is the only place I've found that you can observe without substantial disruption.

[timholy]

Fixed by #35