Important performance wibble to callSiteInline (the n_vals_wanted > 0 thing)

See Note [Inlining in ArgCtxt].  This very small change gives quite a
big performance win. Just showing the bigger ones:

        Program           Size    Allocs   Runtime
--------------------------------------------------------------------------------
           anna          -0.7%     -4.3%      0.15
       cichelli          -0.6%     -6.4%      0.15
         fulsom          -0.4%    -18.5%     -8.1%
            gcd          -0.6%    -12.0%      0.06
        integer          -0.6%    -16.2%     -8.4%
          power          -0.7%    -19.3%     -4.8%
--------------------------------------------------------------------------------
            Min          -0.7%    -19.3%    -15.7%
            Max          -0.1%     +0.1%     +5.7%
 Geometric Mean          -0.6%     -1.9%     -4.3%

The original change was to improve a case that Roman found (see test
eyeball/inline1) but that seems to work ok now anyway.

diff --git a/compiler/coreSyn/CoreUnfold.lhs b/compiler/coreSyn/CoreUnfold.lhs
index 5797cba..c630277 100644 (file)
--- a/compiler/coreSyn/CoreUnfold.lhs
+++ b/compiler/coreSyn/CoreUnfold.lhs
@@ -595,8 +595,8 @@ callSiteInline dflags active_inline id lone_variable arg_infos cont_info
                        = case cont_info of
                            BoringCtxt -> not is_top && n_vals_wanted > 0       -- Note [Nested functions] 
                            CaseCtxt   -> not lone_variable || not is_value     -- Note [Lone variables]
                        = case cont_info of
                            BoringCtxt -> not is_top && n_vals_wanted > 0       -- Note [Nested functions] 
                            CaseCtxt   -> not lone_variable || not is_value     -- Note [Lone variables]

-                           ArgCtxt {} -> True
-                               -- Was: n_vals_wanted > 0; but see test eyeball/inline1.hs
+                           ArgCtxt {} -> n_vals_wanted > 0 
+                               -- See Note [Inlining in ArgCtxt]

 
                    small_enough = (size - discount) <= opt_UF_UseThreshold
                    discount = computeDiscount n_vals_wanted arg_discounts 
 
                    small_enough = (size - discount) <= opt_UF_UseThreshold
                    discount = computeDiscount n_vals_wanted arg_discounts 

@@ -640,6 +640,19 @@ branches.  Then inlining it doesn't increase allocation, but it does
 increase the chance that the constructor won't be allocated at all in
 the branches that don't use it.
 
 increase the chance that the constructor won't be allocated at all in
 the branches that don't use it.
 

+Note [Inlining in ArgCtxt]
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+The condition (n_vals_wanted > 0) here is very important, because otherwise
+we end up inlining top-level stuff into useless places; eg
+   x = I# 3#
+   f = \y.  g x
+This can make a very big difference: it adds 16% to nofib 'integer' allocs,
+and 20% to 'power'.
+
+At one stage I replaced this condition by 'True' (leading to the above 
+slow-down).  The motivation was test eyeball/inline1.hs; but that seems
+to work ok now.
+

 Note [Lone variables]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 The "lone-variable" case is important.  I spent ages messing about
 Note [Lone variables]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 The "lone-variable" case is important.  I spent ages messing about