[backend] do not let caches grow too big

Summary:
This changes the behaviour back to before the "fork once then pipe new tasks
for each source files" era: before that we would fork once per source file,
which had the effect of emptying the caches for every source file. Without this
the caches can grow unchecked.

This is probably suboptimal but should at least be the same behaviour as before.

Reviewed By: jeremydubreil

Differential Revision: D8615124

fbshipit-source-id: 69fc101

infer/src/backend/InferAnalyze.ml

@@ -10,11 +10,19 @@
 open! IStd
 module L = Logging
 
+let clear_caches () =
+  Ondemand.clear_cache () ;
+  Summary.clear_cache () ;
+  Typ.Procname.SQLite.clear_cache ()
+
+
 (** Create tasks to analyze an execution environment *)
 let analyze_source_file : SourceFile.t Tasks.doer =
  fun source_file ->
   let exe_env = Exe_env.mk () in
   L.(debug Analysis Medium) "@\nProcessing '%a'@." SourceFile.pp source_file ;
+  (* clear cache for each source file to avoid it growing unboundedly *)
+  clear_caches () ;
   Callbacks.analyze_file exe_env source_file ;
   if Config.write_html then Printer.write_all_html_files source_file
 
@@ -79,9 +87,7 @@ let main ~changed_files =
     (if Int.equal n_source_files 1 then "" else "s")
     Config.results_dir ;
   (* empty all caches to minimize the process heap to have less work to do when forking *)
-  Summary.clear_cache () ;
-  Typ.Procname.SQLite.clear_cache () ;
-  Random.self_init () ;
+  clear_caches () ;
   if Int.equal Config.jobs 1 then (
     Tasks.run_sequentially ~f:analyze_source_file source_files_to_analyze ;
     L.progress "@\nAnalysis finished in %as@." Pp.elapsed_time () )

infer/src/backend/Tasks.ml

@@ -27,9 +27,13 @@ let run_sequentially ~(f: 'a doer) (tasks: 'a list) : unit =
 
 
 let fork_protect ~f x =
+  (* this is needed whenever a new process is started *)
   EventLogger.prepare () ;
   L.reset_formatters () ;
   ResultsDatabase.new_database_connection () ;
+  (* get different streams of random numbers in each fork, in particular to lessen contention in
+     `Filename.mk_temp` *)
+  Random.self_init () ;
   f x
 
 

infer/src/backend/ondemand.ml

@@ -236,6 +236,8 @@ let analyze_proc_name ?caller_pdesc callee_pname =
       summary_option
 
 
+let clear_cache () = Typ.Procname.Hash.clear (Lazy.force cached_results)
+
 (** Find a proc desc for the procedure, perhaps loading it from disk. *)
 let get_proc_desc callee_pname =
   match !callbacks_ref with Some callbacks -> callbacks.get_proc_desc callee_pname | None -> None

infer/src/backend/ondemand.mli

@@ -31,3 +31,6 @@ val set_callbacks : callbacks -> unit
 
 val unset_callbacks : unit -> unit
 (** Unset the callbacks used to perform on-demand analysis. *)
+
+val clear_cache : unit -> unit
+(** empty the cache of ondemand results *)