Ticky: Gate tag-inference dummy ticky-counters behind a flag.wip/andreask/infer_ticks

Tag inference included a way to collect stats about avoided tag-checks.
This was dony by emitting "dummy" ticky entries with counts corresponding
to predicted/unpredicated tag checks.

This behaviour for ticky is now gated behind -fticky-tag-checks.

I also documented ticky-LNE in the process.

8 files changed, 39 insertions, 7 deletions

diff --git a/compiler/GHC/Driver/Config/StgToCmm.hs b/compiler/GHC/Driver/Config/StgToCmm.hs
index ab64076a79..4b0e126f98 100644
--- a/compiler/GHC/Driver/Config/StgToCmm.hs
+++ b/compiler/GHC/Driver/Config/StgToCmm.hs
@@ -29,6 +29,7 @@ initStgToCmmConfig dflags mod = StgToCmmConfig
   , stgToCmmTickyAllocd   = gopt Opt_Ticky_Allocd          dflags
   , stgToCmmTickyLNE      = gopt Opt_Ticky_LNE             dflags
   , stgToCmmTickyDynThunk = gopt Opt_Ticky_Dyn_Thunk       dflags
+  , stgToCmmTickyTag      = gopt Opt_Ticky_Tag             dflags
   -- flags
   , stgToCmmLoopification = gopt Opt_Loopification         dflags
   , stgToCmmAlignCheck    = gopt Opt_AlignmentSanitisation dflags
diff --git a/compiler/GHC/Driver/Flags.hs b/compiler/GHC/Driver/Flags.hs
index 759d8b48c0..ca5e5a0548 100644
--- a/compiler/GHC/Driver/Flags.hs
+++ b/compiler/GHC/Driver/Flags.hs
@@ -295,6 +295,7 @@ data GeneralFlag
    | Opt_Ticky_Allocd
    | Opt_Ticky_LNE
    | Opt_Ticky_Dyn_Thunk
+   | Opt_Ticky_Tag
    | Opt_RPath
    | Opt_RelativeDynlibPaths
    | Opt_CompactUnwind               -- ^ @-fcompact-unwind@
diff --git a/compiler/GHC/Driver/Session.hs b/compiler/GHC/Driver/Session.hs
index 34d34815ce..2de16a6e1e 100644
--- a/compiler/GHC/Driver/Session.hs
+++ b/compiler/GHC/Driver/Session.hs
@@ -2333,6 +2333,8 @@ dynamic_flags_deps = [
         (NoArg (setGeneralFlag Opt_Ticky_LNE))
   , make_ord_flag defGhcFlag "ticky-dyn-thunk"
         (NoArg (setGeneralFlag Opt_Ticky_Dyn_Thunk))
+  , make_ord_flag defGhcFlag "ticky-tag-checks"
+        (NoArg (setGeneralFlag Opt_Ticky_Tag))
         ------- recompilation checker --------------------------------------
   , make_dep_flag defGhcFlag "recomp"
         (NoArg $ unSetGeneralFlag Opt_ForceRecomp)
diff --git a/compiler/GHC/Iface/Recomp/Flags.hs b/compiler/GHC/Iface/Recomp/Flags.hs
index 90f3afebbc..b1f6c66e79 100644
--- a/compiler/GHC/Iface/Recomp/Flags.hs
+++ b/compiler/GHC/Iface/Recomp/Flags.hs
@@ -66,7 +66,7 @@ fingerprintDynFlags hsc_env this_mod nameio =
 
         -- Ticky
         ticky =
-          map (`gopt` dflags) [Opt_Ticky, Opt_Ticky_Allocd, Opt_Ticky_LNE, Opt_Ticky_Dyn_Thunk]
+          map (`gopt` dflags) [Opt_Ticky, Opt_Ticky_Allocd, Opt_Ticky_LNE, Opt_Ticky_Dyn_Thunk, Opt_Ticky_Tag]
 
         flags = ((mainis, safeHs, lang, cpp), (paths, prof, ticky, debugLevel, callerCcFilters))
 
diff --git a/compiler/GHC/StgToCmm/Config.hs b/compiler/GHC/StgToCmm/Config.hs
index 28c11e1321..623a7b8f0a 100644
--- a/compiler/GHC/StgToCmm/Config.hs
+++ b/compiler/GHC/StgToCmm/Config.hs
@@ -35,6 +35,7 @@ data StgToCmmConfig = StgToCmmConfig
                                                  -- join-points (let-no-escape)
   , stgToCmmTickyDynThunk  :: !Bool              -- ^ True indicates ticky uses name-specific counters for
                                                  -- dynamic thunks
+  , stgToCmmTickyTag       :: !Bool              -- ^ True indicates ticky will count number of avoided tag checks by tag inference.
   ---------------------------------- Flags --------------------------------------
   , stgToCmmLoopification  :: !Bool              -- ^ Loopification enabled (cf @-floopification@)
   , stgToCmmAlignCheck     :: !Bool              -- ^ Insert alignment check (cf @-falignment-sanitisation@)
diff --git a/compiler/GHC/StgToCmm/Ticky.hs b/compiler/GHC/StgToCmm/Ticky.hs
index 57bf1a97b6..4aec817412 100644
--- a/compiler/GHC/StgToCmm/Ticky.hs
+++ b/compiler/GHC/StgToCmm/Ticky.hs
@@ -319,7 +319,7 @@ emitTickyCounterTag unique (NonVoid id) =
   let name = idName id
       ctr_lbl = mkTagHitLabel name unique in
     (>> return ctr_lbl) $
-    ifTicky $ do
+    ifTickyTag $ do
         { platform <- getPlatform
         ; parent <- getTickyCtrLabel
         ; mod_name <- getModuleName
@@ -664,18 +664,18 @@ tickyStackCheck = ifTicky $ bumpTickyCounter (fsLit "STK_CHK_ctr")
 
 -- | Predicted a pointer would be tagged correctly (GHC will crash if not so no miss case)
 tickyTagged :: FCode ()
-tickyTagged         = ifTicky $ bumpTickyCounter (fsLit "TAG_TAGGED_pred")
+tickyTagged         = ifTickyTag $ bumpTickyCounter (fsLit "TAG_TAGGED_pred")
 
 -- | Pass a boolean expr indicating if tag was present.
 tickyUntagged :: CmmExpr -> FCode ()
 tickyUntagged e     = do
-    ifTicky $ bumpTickyCounter (fsLit "TAG_UNTAGGED_pred")
-    ifTicky $ bumpTickyCounterByE (fsLit "TAG_UNTAGGED_miss") e
+    ifTickyTag $ bumpTickyCounter (fsLit "TAG_UNTAGGED_pred")
+    ifTickyTag $ bumpTickyCounterByE (fsLit "TAG_UNTAGGED_miss") e
 
 -- | Called when for `case v of ...` we can avoid entering v based on
 -- tag inference information.
 tickyTagSkip :: Unique -> Id -> FCode ()
-tickyTagSkip unique id = ifTicky $ do
+tickyTagSkip unique id = ifTickyTag $ do
   let ctr_lbl = mkTagHitLabel (idName id) unique
   registerTickyCtr ctr_lbl
   bumpTickyTagSkip ctr_lbl
@@ -692,6 +692,9 @@ runIfFlag f = whenM (f <$> getStgToCmmConfig)
 ifTicky :: FCode () -> FCode ()
 ifTicky = runIfFlag stgToCmmDoTicky
 
+ifTickyTag :: FCode () -> FCode ()
+ifTickyTag = runIfFlag stgToCmmTickyTag
+
 ifTickyAllocd :: FCode () -> FCode ()
 ifTickyAllocd = runIfFlag stgToCmmTickyAllocd
 
diff --git a/docs/users_guide/expected-undocumented-flags.txt b/docs/users_guide/expected-undocumented-flags.txt
index 7af6d42080..33958b1578 100644
--- a/docs/users_guide/expected-undocumented-flags.txt
+++ b/docs/users_guide/expected-undocumented-flags.txt
@@ -97,4 +97,3 @@
 -rtsopts=some
 -syslib
 -this-component-id
--ticky-LNE
diff --git a/docs/users_guide/profiling.rst b/docs/users_guide/profiling.rst
index d68c056946..dcfd42b84e 100644
--- a/docs/users_guide/profiling.rst
+++ b/docs/users_guide/profiling.rst
@@ -1740,6 +1740,31 @@ Using “ticky-ticky” profiling (for implementors)
 
     Track allocations of dynamic thunks.
 
+.. ghc-flag:: -ticky-LNE
+    :shortdesc: Treat join point binders similar to thunks/functions.
+    :type: dynamic
+    :category:
+
+    These are not allocated, and can be very performance sensitive so we usually don't
+    want to run ticky counters for these to avoid even worse performance for tickied builds.
+
+    But sometimes having information about these binders is critical. So we have a flag to ticky them
+    anyway.
+
+.. ghc-flag:: -ticky-tag-checks
+    :shortdesc: Emit dummy ticky counters to record how many tag-inference checks tag inference avoided.
+    :type: dynamic
+    :category:
+
+    These dummy counters contain:
+    * The number of avoided tag checks in the entry count.
+    * "infer" as the argument string to distinguish them from regular counters.
+    * The name of the variable we are casing on, as well as a unique to represent the inspection site as one variable might be cased on multiple times.
+      The unique comes first , with the variable coming at the end. Like this: `u10_s98c (Main) at nofib/spectral/simple/Main.hs:677:1 in u10`
+      where `u10` is the variable and `u10_s98c` the unique associated with the inspection site.
+
+    Note that these counters are currently not processed well be eventlog2html. So if you want to check them you will have to use the text based interface.
+
 GHC's ticky-ticky profiler provides a low-level facility for tracking
 entry and allocation counts of particular individual closures.
 Because ticky-ticky profiling requires a certain familiarity with GHC