[project @ 1997-10-19 21:51:43 by sof]

Bargain multiplier opt_UnfoldingKeenessFactor added to discount computation

diff --git a/ghc/compiler/coreSyn/CoreUnfold.lhs b/ghc/compiler/coreSyn/CoreUnfold.lhs
index 36a6746..e254958 100644 (file)
--- a/ghc/compiler/coreSyn/CoreUnfold.lhs
+++ b/ghc/compiler/coreSyn/CoreUnfold.lhs
@@ -45,7 +45,8 @@ import Bag            ( emptyBag, unitBag, unionBags, Bag )
 
 import CmdLineOpts     ( opt_UnfoldingCreationThreshold,
                          opt_UnfoldingUseThreshold,
-                         opt_UnfoldingConDiscount
+                         opt_UnfoldingConDiscount,
+                         opt_UnfoldingKeenessFactor
                        )
 import Constants       ( uNFOLDING_CHEAP_OP_COST,
                          uNFOLDING_DEAR_OP_COST,
@@ -482,13 +483,21 @@ smallEnoughToInline _ _ UnfoldNever  = False
 smallEnoughToInline arg_is_evald_s result_is_scruted
              (UnfoldIfGoodArgs m_tys_wanted n_vals_wanted discount_vec size scrut_discount)
   = enough_args n_vals_wanted arg_is_evald_s &&
-    discounted_size <= opt_UnfoldingUseThreshold
+    size - discount <= opt_UnfoldingUseThreshold
   where
 
     enough_args n [] | n > 0 = False   -- A function with no value args => don't unfold
     enough_args _ _         = True     -- Otherwise it's ok to try
 
-    discounted_size = (size - args_discount) - result_discount
+       -- We multiple the raw discounts (args_discount and result_discount)
+       -- ty opt_UnfoldingKeenessFactor because the former have to do with
+       -- *size* whereas the discounts imply that there's some extra *efficiency*
+       -- to be gained (e.g. beta reductions, case reductions) by inlining.
+    discount :: Int
+    discount = round (
+                     opt_UnfoldingKeenessFactor * 
+                     fromInt (args_discount + result_discount)
+                    )
 
     args_discount = sum (zipWith arg_discount discount_vec arg_is_evald_s)
     result_discount | result_is_scruted = scrut_discount