Profiler is inconsistent about biography for GHC's heap

[igloo]

My slightly modified rnBind has this clause (i is an Int):

rnBind sig_fn
       trim
       i
       (L loc (FunBind { fun_id = name, 
                         fun_infix = inf, 
                         fun_matches = matches,
                         -- no pattern FVs
                         bind_fvs = _
                       }))
       -- invariant: no free vars here when it's a FunBind
  = i `seq` setSrcSpan loc $
    do  { let plain_name = unLoc name

    ; (matches', fvs) <- bindSigTyVarsFV (sig_fn plain_name) $
                -- bindSigTyVars tests for Opt_ScopedTyVars
                 rnMatchGroup (FunRhs plain_name inf) matches

    ; checkPrecMatch inf plain_name matches'

    ; return
        (L loc (trace ("Using FunBind for " ++ show i)
                (FunBind { fun_id = name, 
                                  fun_infix = inf, 
                                  fun_matches = matches',
                     bind_fvs = trim fvs, 
                                  fun_co_fn = idHsWrapper, 
                                  fun_tick = Nothing })),
          [plain_name],
          fvs)
      }

If I comment out the trace and run

ghc --make J -fforce-recomp +RTS -p -hcrnBind -hyHsBindLR -hb

then I get notrace.png, all VOID. However, with the trace I get trace.png, LAG turning into DRAG. The two graphs are also a slightly different shape, which is curious.

Also attached is the core for with and without the tracing. I think the only interesting difference is

-      a30_ [ALWAYS Just L] :: NameSet.FreeVars
+      a30_ [ALWAYS Just L] :: HsBinds.HsBindLR Name.Name Name.Name
       [Str: DmdType]
-      a30_ = w1_ ww3_ } in
-    let {
-      a31_ :: HsBinds.HsWrapper
-      []
-      a31_ =
-        __scc {rnBind ghc-6.9.20080517:RnBinds !}
-        __scc {idHsWrapper ghc-6.9.20080517:HsBinds} HsBinds.WpHole } in
-    let {
-      a32_ :: HsBinds.HsWrapper
-      []
-      a32_ =
-        __scc {rnBind ghc-6.9.20080517:RnBinds !}
-        __scc {idHsWrapper ghc-6.9.20080517:HsBinds} HsBinds.WpHole } in
-    let {
-      a33_ :: HsBinds.HsBindLR Name.Name Name.Name
-      [Str: DmdType]
-      a33_ =
-        HsBinds.FunBind
-          @ Name.Name
-          @ Name.Name
-          name_
-          inf_
-          ww2_
-          a32_
-          a30_
-          (Data.Maybe.Nothing @ (GHC.Base.Int, [Name.Name])) } in
+      a30_ =
+        Debug.Trace.trace
+          @ (HsBinds.HsBindLR Name.Name Name.Name)
+          lvl45_
+          (HsBinds.FunBind
+             @ Name.Name
+             @ Name.Name
+             name_
+             inf_
+             ww2_
+             (__scc {rnBind ghc-6.9.20080517:RnBinds !}
+              __scc {idHsWrapper ghc-6.9.20080517:HsBinds} HsBinds.WpHole)
+             (w1_ ww3_)
+             (Data.Maybe.Nothing @ (GHC.Base.Int, [Name.Name]))) } in

[igloo]

In case this doesn't get sorted soon, I think these are all the relevant changes I've made:

hunk ./compiler/rename/RnBinds.lhs 40
-import Digraph         ( SCC(..), stronglyConnComp )
+import Digraph
hunk ./compiler/rename/RnBinds.lhs 45
-import Util            ( filterOut )
-import Monad           ( foldM, unless )
+import Util
+import Monad
+import Control.Exception
+import Debug.Trace
hunk ./compiler/rename/RnBinds.lhs 321
-   binds_w_dus <- mapBagM (rnBind (mkSigTvFn sigs') trim) mbinds
+   xs <- zipWithM (rnBind (mkSigTvFn sigs') trim) [1..] (bagToList mbinds)
+   let binds_w_dus = listToBag xs
hunk ./compiler/rename/RnBinds.lhs 498
+       -> Int
hunk ./compiler/rename/RnBinds.lhs 501
-rnBind _ trim (L loc (PatBind { pat_lhs = pat,
-                                pat_rhs = grhss,
-                                -- pat fvs were stored here while
-                                -- processing the LHS
-                                bind_fvs=pat_fvs }))
+rnBind _ trim i (L loc (PatBind { pat_lhs = pat,
+                                  pat_rhs = grhss,
+                                  -- pat fvs were stored here while
+                                  -- processing the LHS
+                                  bind_fvs=pat_fvs }))
hunk ./compiler/rename/RnBinds.lhs 519
-       trim
+       trim
+       i
hunk ./compiler/rename/RnBinds.lhs 528
-  = setSrcSpan loc $
+  = i `seq` setSrcSpan loc $
hunk ./compiler/rename/RnBinds.lhs 537
-       ; return (L loc (FunBind { fun_id = name,
+       ; return
+        (L loc (trace ("Using FunBind for " ++ show i)
+                (FunBind { fun_id = name,
hunk ./compiler/rename/RnBinds.lhs 544
-                                  fun_tick = Nothing }),
+                                  fun_tick = Nothing })),
hunk ./compiler/rename/RnBinds.lhs 549
-rnBind _ _ b = pprPanic "rnBind" (ppr b)
+rnBind _ _ _ b = pprPanic "rnBind" (ppr b)

[simonmar]

Needed for further profiling of GHC, we should investigate for 6.10.1 if possible.

[simonpj]

Decreasing priority. (High cost-to-benefit ratio!) Simon & Simon

[simonmar]

optimistically assuming this is fixed by the fix for #3039.