Call Arity: Trade precision for performance in large mutually recursive groups
authorJoachim Breitner <mail@joachim-breitner.de>
Wed, 15 Apr 2015 14:25:13 +0000 (16:25 +0200)
committerJoachim Breitner <mail@joachim-breitner.de>
Thu, 16 Apr 2015 12:06:44 +0000 (14:06 +0200)
Sometimes (especial with derived Data instances, it seems), one can have
very large mutually recursive bindings. Calculating the Call Arity
analysis result with full precision is an expensive operation in these
case. So above a certain threshold (25, for no good reason besides
intuition), skip this calculation and assume the recursion is not
linear, which is a conservative result.

With this, the Call Arity analysis accounts for 3.7% of the compile time
of haskell-src-exts. Fixes #10293

Differential Revision: https://phabricator.haskell.org/D843

compiler/simplCore/CallArity.hs

index 4e4555c..c2a5ad0 100644 (file)
@@ -630,6 +630,9 @@ callArityRecEnv any_boring ae_rhss ae_body
     cross_calls
         -- See Note [Taking boring variables into account]
         | any_boring          = completeGraph (domRes ae_combined)
+        -- Also, calculating cross_calls is expensive. Simply be conservative
+        -- if the mutually recursive group becomes too large.
+        | length ae_rhss > 25 = completeGraph (domRes ae_combined)
         | otherwise           = unionUnVarGraphs $ map cross_call ae_rhss
     cross_call (v, ae_rhs) = completeBipartiteGraph called_by_v called_with_v
       where