GitLab

Ben Gamari pushed to branch wip/tick-prefer-anf at Glasgow Haskell Compiler / GHC

Commits:

5db84ad6

by Ben Gamari at 2026-06-21T10:16:00-04:00

Core: Preserve ANF when ticking a tick-headed expression (#27415)

`mkTickCpe` promises not to push a tick onto the arguments of an
application, so that CorePrep's ANF invariant is maintained: an argument
must be trivial and, in particular, tick-free (see the `arg`
non-terminal in Note [CorePrep invariants] and Note [mkTick breaks
ANF]). `getStgArgFromTrivialArg` relies on this.

However, its handling of a tick-headed expression delegated to
`tickTickedExpr`, which, when commuting the new tick past the existing
tick stack and pushing it inwards, called `mkTick` (i.e. `mk_tick
False`) directly -- silently dropping ANF preservation and re-enabling
PSCC3.

With `-prof -fprof-auto -finfo-table-map`, a definition such as

    foo = Just id

has its body wrapped in an SCC by `-fprof-auto`. The SCC is commuted past the
source notes heading the `Just` application and, with `preserve_anf`
lost, pushed onto the (function-valued) argument as `(scc<foo>
breakpoint) @a`. That argument is no longer trivial, so
`getStgArgFromTrivialArg` panicked:

    getStgArgFromTrivialArg
      (scc<foo> id) @a_aiNd

Fixes #27415.

4 changed files:

compiler/GHC/Core/Utils.hs
compiler/GHC/CoreToStg/Prep.hs
+ testsuite/tests/profiling/should_compile/T27415.hs
testsuite/tests/profiling/should_compile/all.T

Changes:

compiler/GHC/Core/Utils.hs

@@ -343,7 +343,7 @@ mk_tick preserve_anf t orig_expr = mkTick' orig_expr
    mkTick' expr
      -- Deal with ticking a expression headed by one or more ticks.
      | Just (ts, e) <- tickedExpr_maybe expr
 -    = tickTickedExpr t ts e
 +    = tickTickedExpr preserve_anf t ts e
    mkTick' expr = case expr of
      Lam x e
@@ -413,11 +413,13 @@ mk_tick preserve_anf t orig_expr = mkTick' orig_expr
  -- | Apply a tick to an expression headed by ticks.
  tickTickedExpr
 -  :: CoreTickish             -- ^ tick to add
 +  :: Bool                    -- ^ preserve ANF? (See 'mkTickCpe' and
 +                             --   Note [mkTick breaks ANF] in GHC.CoreToStg.Prep)
 +  -> CoreTickish             -- ^ tick to add
    -> NE.NonEmpty CoreTickish -- ^ existing stack of ticks
    -> CoreExpr                -- ^ inner core expression
    -> CoreExpr
 -tickTickedExpr t1 t2s e
 +tickTickedExpr preserve_anf t1 t2s e
    -- Case 1: common up 't1' with a tick in the stack.
    --
@@ -431,8 +433,24 @@ tickTickedExpr t1 t2s e
    -- Case 2: 't1' can be commuted past all the ticks in the stack, e.g. because
    -- it has tighter placement properties than all the ticks in the stack.
    -- Push it inwards to expose cancellation opportunities.
 +  --
 +  -- We must thread 'preserve_anf' through here.  Otherwise we may turn the
 +  -- argument
 +  --
 +  --     scc<f> (src<l> (Just x))
 +  --
 +  -- into the ill-formed argument
 +  --
 +  --     src<l> (Just (scc<f> x))   -- SCC on an argument: not a CpeArg!
 +  --
 +  -- whereas with preserve_anf we keep the SCC outside the application,
 +  -- respecting ANF:
 +  --
 +  --     src<l> (scc<f> (Just breakpoint))
 +  --
 +  -- See Note [mkTick breaks ANF] in GHC.CoreToStg.Prep and #27415.
    | all (tickishCommutable t1) t2s
 -  = apply_ticks t2s $ mkTick t1 e
 +  = apply_ticks t2s $ mk_tick preserve_anf t1 e
    -- Fallback: keep the new tick on the outside.
    | otherwise
@@ -455,7 +473,7 @@ tickTickedExpr t1 t2s e
  {- Note [Pushing SCCs inwards]
  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 -Amongst all ticks, SCCs have the laxest placement properties (PlaceCostCentre,
 +Amongst all ticks, SCCs have the most lax placement properties (PlaceCostCentre,
  as described in Note [Tickish placement] GHC.Types.Tickish):
    (PSCC1) SCCs around non-function variables can be eliminated.

compiler/GHC/CoreToStg/Prep.hs

@@ -918,10 +918,14 @@ mkTick will push the SCC into the constructor application, resulting in:
    \ (eta :: Char -> Bool) -> BindP (p :: Int) (scc<oneM> eta)
 +i.e. the SCC ends up on the argument `eta`, which is no longer a valid `arg`.
++
  To avoid this problem (at least until 'mkTick' is more thoroughly reworked to
  avoid this infelicity, see #27141), we define a variant of 'mkTick', called
  'mkTickCpe', which does not push ticks into constructor applications (this is
 -the only optimisation done by 'mkTick' that can break ANF).
 +the only optimisation done by 'mkTick' that can break ANF). This is the concrete
 +meaning of the 'preserve_anf' flag threaded through 'GHC.Core.Utils.mk_tick':
 +when set, 'mk_tick' will not push a tick onto an application's arguments (PSCC3).
  We prefer using a small variant of 'mkTick' rather than using the 'Tick'
  constructor, as the latter can slightly degrade profiling reports by failing to

testsuite/tests/profiling/should_compile/T27415.hs

 +{-# LANGUAGE ExistentialQuantification #-}
++
 +module T27415 where
++
 +import Control.Concurrent.MVar
++
 +data ResultVar b = forall a . ResultVar (a -> b) (MVar (Maybe a))
++
 +mkResultVar :: MVar (Maybe a) -> ResultVar a
 +mkResultVar = ResultVar id

testsuite/tests/profiling/should_compile/all.T

@@ -23,3 +23,4 @@ test('T20938', [test_opts], compile, ['-O -prof'])
  test('T26056', [test_opts], compile, ['-O -prof'])
  test('T27121', [test_opts, extra_files(['T27121_aux.hs'])], multimod_compile, ['T27121', '-v0 -O -prof -fprof-auto'])
  test('T27182', [test_opts], compile, ['-O -prof -fprof-late'])
 +test('T27415', [test_opts], compile, ['-O -prof -fprof-auto -finfo-table-map'])