analysis performance problem

[aviatesk]

analysis performance problem

[aviatesk]

It turns out this is the performance problem; with the following diff (ignoring report cache restoring) the abstract interpretation terminates:

diff --git a/src/reports.jl b/src/reports.jl
index d29b855..8b677dd 100644
--- a/src/reports.jl
+++ b/src/reports.jl
@@ -170,6 +170,8 @@ function restore_cached_report!(cache::InferenceErrorReportCache,
                                 interp#=::JETInterpreter=#,
                                 caller::InferenceState,
                                 )
+    return
+    
     report = restore_cached_report(cache, caller)
     push!(isa(report, UncaughtExceptionReport) ? interp.uncaught_exceptions : interp.reports, report)
     return

jυλια> using JET; @time interp, frame = profile_call(JET.virtual_process!, (String, String, Module, Symbol, JET.JETInterpreter, JET.VirtualProcessResult));
[ Info: Precompiling JET [c3a54625-cd67-489e-a8e7-0a5a0ff4e31b]
 12.957541 seconds (32.86 M allocations: 1.810 GiB, 6.63% gc time, 54.95% compilation time)

jυλια> length(interp.reports)
188

jυλια> length(JET.ANALYZED_LINFOS)
9436

[aviatesk]

With cb6be90, now we can run report_call(JET.virtual_process!, (String, String, Module, Symbol, JET.JETInterpreter, JET.VirtualProcessResult)) in 75 secs.
We still definitely need to improve analysis speed, though.

[aviatesk]

With 7045d02, the self profiling finishes in 26 secs. We can improve yet more !