Simon Jakobi pushed to branch wip/sjakobi/T27296-stable-simpl at Glasgow Haskell Compiler / GHC

Commits:

5 changed files:

Changes:

  • compiler/GHC/Core/Ppr.hs
    ... ... @@ -28,10 +28,10 @@ module GHC.Core.Ppr (
    28 28 
     import GHC.Prelude
    29 29
    30 30 
     import GHC.Core
    31 
    -import GHC.Core.Stats (exprStats)
    31 
    +import GHC.Core.Stats (CoreStats(..), exprStats)
    32 32 
     import GHC.Types.Fixity (LexicalFixity(..))
    33 
    -import GHC.Types.Literal( pprLiteral )
    34 
    -import GHC.Types.Name( pprInfixName, pprPrefixName, getOccString, getSrcSpan )
    33 
    +import GHC.Types.Literal( Literal, pprLiteral )
    34 
    +import GHC.Types.Name( getOccString, getSrcSpan, pprInfixName, pprPrefixName )
    35 35 
     import GHC.Types.Var
    36 36 
     import GHC.Types.Id
    37 37 
     import GHC.Types.Id.Info
    ... ... @@ -46,8 +46,8 @@ import GHC.Types.Basic
    46 46 
     import GHC.Utils.Misc
    47 47 
     import GHC.Utils.Outputable
    48 48 
     import GHC.Utils.Panic (panic)
    49 
    -import GHC.Types.SrcLoc ( SrcSpan(..), srcSpanStartLine, srcSpanStartCol
    50 
    -                        , pprUserRealSpan )
    49 
    +import GHC.Types.SrcLoc ( SrcSpan(..), pprUserRealSpan, srcSpanStartCol
    50 
    +                        , srcSpanStartLine )
    51 51 
     import GHC.Types.Tickish
    52 52
    53 53 
     import Data.List ( sortOn )
    ... ... @@ -99,7 +99,18 @@ Uniques*, so two dumps line up across rebuilds. The sort key is:
    99 99 
          whether the OccName *contains* a '$', which marks a derived binder: a worker
    100 100 
          is @$wfoo@, but a call-site specialisation is tidied to @bar_$sfoo@ (no
    101 101 
          leading '$'), so a leading-'$' test would miss it.
    102 
    -  3. the OccName string, as a final lexical, deterministic tie-break.
    102 
    +  3. the OccName string, as a lexical, deterministic tie-break.
    103 
    +  4. a content-based tie-break on the right-hand side ('rhsKey'): the floated
    104 
    +     literal, if any, then the RHS size statistics. This matters for the
    105 
    +     anonymous floats: 'newLvlVar' builds them all with OccName "lvl" and
    106 
    +     noSrcSpan, so keys 1-3 are identical and without it their order would fall
    107 
    +     back to the Unique-driven input order -- the churn we set out to remove.
    108 
    +     (Tidied dumps like -ddump-simpl give the floats distinct names lvl,
    109 
    +     lvl1, ...; this additionally stabilises untidied dumps such as
    110 
    +     -ddump-simpl-iterations.) It is only a best-effort tie-break -- RHSs
    111 
    +     agreeing on both components keep their input order -- and Unique-independent
    112 
    +     for the numeric CAFs we target (a rubbish literal is the exception: its
    113 
    +     'cmpLit' falls back to the Unique-dependent 'nonDetCmpType').
    103 114
    104 115 
     Recursive groups are never split: a 'Rec' is one 'CoreBind', placed as a unit by
    105 116 
     its earliest-source member, with its members sorted by the same key.
    ... ... @@ -114,24 +125,36 @@ suffer the cross-module churn this flag addresses.
    114 125 
     useful for debugging the compiler itself.
    115 126 
     -}
    116 127
    128 
    +-- | The sort key for one top-level binder. The trailing 'RhsKey' is a
    129 
    +-- content-based tiebreak, used only when two binders agree on everything
    130 
    +-- before it. See Note [Stable Core dump order].
    131 
    +type DumpSortKey =
    132 
    +  ( Int     -- source-span bucket: 0 = real span, 1 = noSrcSpan (sorts last)
    133 
    +  , Int     -- source-span start line
    134 
    +  , Int     -- source-span start column
    135 
    +  , Int     -- dollar-rank: 0 = derived ($w/$s) binder, 1 = its origin
    136 
    +  , String  -- the OccName string, a lexical tiebreak
    137 
    +  , RhsKey  -- content-based tiebreak (see 'rhsKey')
    138 
    +  )
    139 
    +
    117 140 
     -- | Reorder a 'CoreProgram' into a stable, source-location-driven order for
    118 141 
     -- dumping. See Note [Stable Core dump order]. Used by 'dumpPassResult' when
    119 142 
     -- -dstable-core-dump-order is enabled.
    120 143 
     sortCoreBindingsForDump :: CoreProgram -> CoreProgram
    121 144 
     sortCoreBindingsForDump = sortOn bindKey . map sortRecMembers
    122 145 
       where
    123 
    -    sortRecMembers (Rec prs) = Rec (sortOn (bndrKey . fst) prs)
    146 
    +    sortRecMembers (Rec prs) = Rec (sortOn (uncurry elemKey) prs)
    124 147 
         sortRecMembers b         = b
    125 148
    126 
    -    -- 'sortRecMembers' runs first, so a 'Rec' is already sorted by 'bndrKey'
    149 
    +    -- 'sortRecMembers' runs first, so a 'Rec' is already sorted by 'elemKey'
    127 150 
         -- when 'bindKey' sees it; its first member is therefore the minimum key.
    128 
    -    bindKey :: CoreBind -> (Int, Int, Int, Int, String)
    129 
    -    bindKey (NonRec b _)     = bndrKey b
    130 
    -    bindKey (Rec ((b,_):_))  = bndrKey b
    131 
    -    bindKey (Rec [])         = panic "sortCoreBindingsForDump: empty Rec"
    151 
    +    bindKey :: CoreBind -> DumpSortKey
    152 
    +    bindKey (NonRec b rhs)     = elemKey b rhs
    153 
    +    bindKey (Rec ((b,rhs):_))  = elemKey b rhs
    154 
    +    bindKey (Rec [])           = panic "sortCoreBindingsForDump: empty Rec"
    132 155
    133 
    -    bndrKey :: CoreBndr -> (Int, Int, Int, Int, String)
    134 
    -    bndrKey b = (bucket, line, col, dollar_rank, s)
    156 
    +    elemKey :: CoreBndr -> CoreExpr -> DumpSortKey
    157 
    +    elemKey b rhs = (bucket, line, col, dollar_rank, s, rhsKey rhs)
    135 158 
           where
    136 159 
             s = getOccString b
    137 160 
             (bucket, line, col) = case getSrcSpan b of
    ... ... @@ -145,6 +168,23 @@ sortCoreBindingsForDump = sortOn bindKey . map sortRecMembers
    145 168 
             dollar_rank | '$' `elem` s = 0
    146 169 
                         | otherwise    = 1
    147 170
    171 
    +-- | A content-based tie-break on a binder's right-hand side: see point 4 of
    172 
    +-- Note [Stable Core dump order].
    173 
    +type RhsKey =
    174 
    +  ( Maybe Literal              -- the floated literal, if any (Nothing sorts first)
    175 
    +  , (Int, Int, Int, Int, Int)  -- exprStats counts: terms, types, coercions, value binds, join binds
    176 
    +  )
    177 
    +
    178 
    +rhsKey :: CoreExpr -> RhsKey
    179 
    +rhsKey rhs = (litOf rhs, statsTuple (exprStats rhs))
    180 
    +  where
    181 
    +    statsTuple (CS tm ty co vb jb) = (tm, ty, co, vb, jb)
    182 
    +    litOf (Lit l)    = Just l
    183 
    +    litOf (App f a)  = case a of { Lit l -> Just l; _ -> litOf f }
    184 
    +    litOf (Cast e _) = litOf e
    185 
    +    litOf (Tick _ e) = litOf e
    186 
    +    litOf _          = Nothing
    187 
    +
    148 188 
     instance OutputableBndr b => Outputable (Bind b) where
    149 189 
         ppr bind = ppr_bind noAnn bind
    150 190

  • testsuite/tests/simplCore/should_compile/Makefile
    ... ... @@ -316,3 +316,13 @@ T27296:
    316 316 
     	    -dstable-core-dump-order T27296.hs 2> /dev/null \
    317 317 
     	    | sed -nE 's/^(\$$fEqKey|\$$fOrdKey|\$$fOrdKey_\$$ccompare|size|findI_\$$slookupG|lookupG|member|findI|\$$wrotate|rotate|insertG|insertManyI|insertTwoI|weight|balance|ratios|fromAscI)( .*)?$$/\1/p' \
    318 318 
     	    | uniq
    319 
    +
    320 
    +# See T27296b.hs for what this pins and why. The six floated "lvl" constants
    321 
    +# are scrambled in source order; grep them out and the dump coming out
    322 
    +# 1000..6000 confirms the stable, value-ordered float ordering.
    323 
    +T27296b:
    324 
    +	$(RM) -f T27296b.o T27296b.hi
    325 
    +	'$(TEST_HC)' $(TEST_HC_OPTS) -O -c -ddump-float-out -dsuppress-uniques \
    326 
    +	    -dsuppress-idinfo -dsuppress-module-prefixes -dno-typeable-binds \
    327 
    +	    -dstable-core-dump-order T27296b.hs 2> /dev/null \
    328 
    +	    | grep '^lvl = I#'

  • testsuite/tests/simplCore/should_compile/T27296b.hs
    1 
    +-- See Note [Stable Core dump order] in GHC.Core.Ppr.
    2 
    +--
    3 
    +-- Companion to T27296 that pins the ordering of *anonymous* top-level floats.
    4 
    +-- Under -O the boxed Int constants in sel's branches are floated to top level
    5 
    +-- as separate CAFs, all of which the compiler names "lvl" with noSrcSpan (see
    6 
    +-- newLvlVar). Before -dstable-core-dump-order their dump order was the
    7 
    +-- unique-driven processing order; the flag's content-based tie-break (rhsKey)
    8 
    +-- now orders them by literal value -- here 1000..6000, despite the scrambled
    9 
    +-- source order. This dump is intentionally *untidied* (-ddump-float-out), the
    10 
    +-- only place the "lvl" collision is observable; tidied dumps like -ddump-simpl
    11 
    +-- already give the floats distinct names (lvl, lvl1, ...).
    12 
    +module T27296b (sel) where
    13 
    +
    14 
    +{-# NOINLINE sel #-}
    15 
    +sel :: Int -> Int
    16 
    +sel 0 = 5000
    17 
    +sel 1 = 1000
    18 
    +sel 2 = 4000
    19 
    +sel 3 = 2000
    20 
    +sel 4 = 3000
    21 
    +sel _ = 6000

  • testsuite/tests/simplCore/should_compile/T27296b.stdout
    1 
    +lvl = I# 1000#
    2 
    +lvl = I# 2000#
    3 
    +lvl = I# 3000#
    4 
    +lvl = I# 4000#
    5 
    +lvl = I# 5000#
    6 
    +lvl = I# 6000#

  • testsuite/tests/simplCore/should_compile/all.T
    ... ... @@ -603,3 +603,4 @@ test('T25718c', normal, compile, ['-O -ddump-simpl -dsuppress-uniques -dsuppress
    603 603 
     test('T19166', normal, compile, ['-O -ddump-simpl -dsuppress-uniques -dsuppress-all -dno-typeable-binds'])
    604 604 
     test('T27261', [extra_files(['T27261_aux.hs'])], multimod_compile, ['T27261', '-v0 -O'])
    605 605 
     test('T27296', [], makefile_test, ['T27296'])
    606 
    +test('T27296b', [], makefile_test, ['T27296b'])

    • —
    View it on GitLab.
    You're receiving this email because of your account on gitlab.haskell.org. Manage all notifications · Help