[Git][ghc/ghc][wip/sjakobi/T27296-stable-simpl] WIP: Stabilise anonymous float ordering in untidied Core dumps

Simon Jakobi pushed to branch wip/sjakobi/T27296-stable-simpl at Glasgow Haskell Compiler / GHC Commits: b925ad78 by Simon Jakobi at 2026-06-08T20:59:53+02:00 WIP: Stabilise anonymous float ordering in untidied Core dumps Anonymous floats are all built with OccName "lvl" and noSrcSpan (newLvlVar), so the source-span/name sort key is identical for every one of them; sortOn then falls back to the unique-driven input order -- the very churn -dstable-core-dump-order is meant to remove. (Tidied dumps like -ddump-simpl are unaffected, as tidy gives the floats distinct names lvl, lvl1, ...) Add a content-based, unique-independent tie-break (rhsKey): the floated literal, if any, then the RHS size statistics. Add test T27296b pinning the float ordering in an untidied -ddump-float-out dump. It is a makefile_test that seds the dump down to just the bindings (collapsing each pass header to a bare "Float out" separator and dropping the FOS config / size lines), so the six lvl floats are asserted to come out ordered by literal value. Co-Authored-By: Claude Opus 4.7 - - - - - 5 changed files: - compiler/GHC/Core/Ppr.hs - testsuite/tests/simplCore/should_compile/Makefile - + testsuite/tests/simplCore/should_compile/T27296b.hs - + testsuite/tests/simplCore/should_compile/T27296b.stdout - testsuite/tests/simplCore/should_compile/all.T Changes: ===================================== compiler/GHC/Core/Ppr.hs ===================================== @@ -28,10 +28,10 @@ module GHC.Core.Ppr ( import GHC.Prelude import GHC.Core -import GHC.Core.Stats (exprStats) +import GHC.Core.Stats (CoreStats(..), exprStats) import GHC.Types.Fixity (LexicalFixity(..)) -import GHC.Types.Literal( pprLiteral ) -import GHC.Types.Name( pprInfixName, pprPrefixName, getOccString, getSrcSpan ) +import GHC.Types.Literal( Literal, pprLiteral ) +import GHC.Types.Name( getOccString, getSrcSpan, pprInfixName, pprPrefixName ) import GHC.Types.Var import GHC.Types.Id import GHC.Types.Id.Info @@ -46,8 +46,8 @@ import GHC.Types.Basic import GHC.Utils.Misc import GHC.Utils.Outputable import GHC.Utils.Panic (panic) -import GHC.Types.SrcLoc ( SrcSpan(..), srcSpanStartLine, srcSpanStartCol - , pprUserRealSpan ) +import GHC.Types.SrcLoc ( SrcSpan(..), pprUserRealSpan, srcSpanStartCol + , srcSpanStartLine ) import GHC.Types.Tickish import Data.List ( sortOn ) @@ -99,7 +99,21 @@ Uniques*, so two dumps line up across rebuilds. The sort key is: whether the OccName *contains* a '$', which marks a derived binder: a worker is @$wfoo@, but a call-site specialisation is tidied to @bar_$sfoo@ (no leading '$'), so a leading-'$' test would miss it. - 3. the OccName string, as a final lexical, deterministic tie-break. + 3. the OccName string, as a lexical, deterministic tie-break. + 4. a content-based tie-break on the right-hand side ('rhsKey'): the floated + literal, if any, then the RHS size statistics. This matters for the + anonymous floats: 'newLvlVar' builds them all with OccName "lvl" and + noSrcSpan, so keys 1-3 are identical and, without this, their order would + fall back to the Unique-driven input order -- exactly the churn we set out + to remove. (In tidied dumps like -ddump-simpl the floats already have + distinct names lvl, lvl1, ...; this tie-break additionally stabilises the + untidied dumps -ddump-simpl-iterations, -dverbose-core2core etc.) It is a + best-effort tie-break, not a total order: two RHSs that agree on the + floated literal and on every size statistic still fall back to input + order. And it is Unique-independent for the floats we target -- the one + exception is a rubbish literal (LitRubbish), whose 'cmpLit' compares the + underlying type with the Unique-dependent 'nonDetCmpType'; that does not + arise for the numeric CAFs this targets. Recursive groups are never split: a 'Rec' is one 'CoreBind', placed as a unit by its earliest-source member, with its members sorted by the same key. @@ -114,24 +128,36 @@ suffer the cross-module churn this flag addresses. useful for debugging the compiler itself. -} +-- | The sort key for one top-level binder. The trailing 'RhsKey' is a +-- content-based tiebreak, used only when two binders agree on everything +-- before it. See Note [Stable Core dump order]. +type DumpSortKey = + ( Int -- source-span bucket: 0 = real span, 1 = noSrcSpan (sorts last) + , Int -- source-span start line + , Int -- source-span start column + , Int -- $-rank: 0 = derived ($w/$s) binder, 1 = its origin + , String -- the OccName string, a lexical tiebreak + , RhsKey -- content-based tiebreak (see 'rhsKey') + ) + -- | Reorder a 'CoreProgram' into a stable, source-location-driven order for -- dumping. See Note [Stable Core dump order]. Used by 'dumpPassResult' when -- -dstable-core-dump-order is enabled. sortCoreBindingsForDump :: CoreProgram -> CoreProgram sortCoreBindingsForDump = sortOn bindKey . map sortRecMembers where - sortRecMembers (Rec prs) = Rec (sortOn (bndrKey . fst) prs) + sortRecMembers (Rec prs) = Rec (sortOn (uncurry elemKey) prs) sortRecMembers b = b - -- 'sortRecMembers' runs first, so a 'Rec' is already sorted by 'bndrKey' + -- 'sortRecMembers' runs first, so a 'Rec' is already sorted by 'elemKey' -- when 'bindKey' sees it; its first member is therefore the minimum key. - bindKey :: CoreBind -> (Int, Int, Int, Int, String) - bindKey (NonRec b _) = bndrKey b - bindKey (Rec ((b,_):_)) = bndrKey b - bindKey (Rec []) = panic "sortCoreBindingsForDump: empty Rec" + bindKey :: CoreBind -> DumpSortKey + bindKey (NonRec b rhs) = elemKey b rhs + bindKey (Rec ((b,rhs):_)) = elemKey b rhs + bindKey (Rec []) = panic "sortCoreBindingsForDump: empty Rec" - bndrKey :: CoreBndr -> (Int, Int, Int, Int, String) - bndrKey b = (bucket, line, col, dollar_rank, s) + elemKey :: CoreBndr -> CoreExpr -> DumpSortKey + elemKey b rhs = (bucket, line, col, dollar_rank, s, rhsKey rhs) where s = getOccString b (bucket, line, col) = case getSrcSpan b of @@ -145,6 +171,27 @@ sortCoreBindingsForDump = sortOn bindKey . map sortRecMembers dollar_rank | '$' `elem` s = 0 | otherwise = 1 +-- | A cheap tie-break on a binder's right-hand side, used to order binders that +-- are otherwise indistinguishable -- in practice the anonymous floats, which all +-- share OccName \"lvl\" and 'noSrcSpan' (see 'GHC.Core.Opt.SetLevels.newLvlVar'). +-- It pairs the floated literal (if any, looking through the @I#@-style box and +-- casts/ticks) with the RHS size statistics. For its Unique-independence and +-- best-effort caveats see Note [Stable Core dump order]. +type RhsKey = + ( Maybe Literal -- the floated literal, if any (Nothing sorts first) + , (Int, Int, Int, Int, Int) -- exprStats counts: terms, types, coercions, value binds, join binds + ) + +rhsKey :: CoreExpr -> RhsKey +rhsKey rhs = (litOf rhs, statsTuple (exprStats rhs)) + where + statsTuple (CS tm ty co vb jb) = (tm, ty, co, vb, jb) + litOf (Lit l) = Just l + litOf (App f a) = case a of { Lit l -> Just l; _ -> litOf f } + litOf (Cast e _) = litOf e + litOf (Tick _ e) = litOf e + litOf _ = Nothing + instance OutputableBndr b => Outputable (Bind b) where ppr bind = ppr_bind noAnn bind ===================================== testsuite/tests/simplCore/should_compile/Makefile ===================================== @@ -316,3 +316,22 @@ T27296: -dstable-core-dump-order T27296.hs 2> /dev/null \ | sed -nE 's/^(\$$fEqKey|\$$fOrdKey|\$$fOrdKey_\$$ccompare|size|findI_\$$slookupG|lookupG|member|findI|\$$wrotate|rotate|insertG|insertManyI|insertTwoI|weight|balance|ratios|fromAscI)( .*)?$$/\1/p' \ | uniq + +# Companion to T27296 that pins the ordering of *anonymous* top-level floats. +# See Note [Stable Core dump order] in GHC.Core.Ppr and the header of T27296b.hs. +# This is an *untidied* dump (-ddump-float-out), the only place the "lvl" +# collision is observable. The sed collapses each float-out pass header to a +# bare "==================== Float out ====================" separator (dropping +# the noisy FOS config) and drops the "Result size" and "-- RHS size" lines, +# leaving just the bindings of each pass -- the six lvl floats in the first pass +# must come out ordered by literal value (1000..6000). +T27296b: + $(RM) -f T27296b.o T27296b.hi + '$(TEST_HC)' $(TEST_HC_OPTS) -O -c -ddump-float-out -dsuppress-uniques \ + -dsuppress-idinfo -dsuppress-module-prefixes -dno-typeable-binds \ + -dstable-core-dump-order T27296b.hs 2> /dev/null \ + | sed -E \ + -e '/^=+ Float out/,/=+$$/c\==================== Float out ====================' \ + -e '/^Result size of Float out/,/^ = \{terms/d' \ + -e '/^-- RHS size:/d' \ + | cat -s ===================================== testsuite/tests/simplCore/should_compile/T27296b.hs ===================================== @@ -0,0 +1,21 @@ +-- See Note [Stable Core dump order] in GHC.Core.Ppr. +-- +-- Companion to T27296 that pins the ordering of *anonymous* top-level floats. +-- Under -O the boxed Int constants in sel's branches are floated to top level +-- as separate CAFs, all of which the compiler names "lvl" with noSrcSpan (see +-- newLvlVar). Before -dstable-core-dump-order their dump order was the +-- unique-driven processing order; the flag's content-based tie-break (rhsKey) +-- now orders them by literal value -- here 1000..6000, despite the scrambled +-- source order. This dump is intentionally *untidied* (-ddump-float-out), the +-- only place the "lvl" collision is observable; tidied dumps like -ddump-simpl +-- already give the floats distinct names (lvl, lvl1, ...). +module T27296b (sel) where + +{-# NOINLINE sel #-} +sel :: Int -> Int +sel 0 = 5000 +sel 1 = 1000 +sel 2 = 4000 +sel 3 = 2000 +sel 4 = 3000 +sel _ = 6000 ===================================== testsuite/tests/simplCore/should_compile/T27296b.stdout ===================================== @@ -0,0 +1,54 @@ + +==================== Float out ==================== + +sel :: Int -> Int +sel + = \ (ds :: Int) -> + case ds of { I# ds -> + case ds of { + __DEFAULT -> lvl; + 0# -> lvl; + 1# -> lvl; + 2# -> lvl; + 3# -> lvl; + 4# -> lvl + } + } + +lvl :: Int +lvl = I# 1000# + +lvl :: Int +lvl = I# 2000# + +lvl :: Int +lvl = I# 3000# + +lvl :: Int +lvl = I# 4000# + +lvl :: Int +lvl = I# 5000# + +lvl :: Int +lvl = I# 6000# + +==================== Float out ==================== + +$wsel :: Int# -> Int# +$wsel + = \ (ww :: Int#) -> + case ww of { + __DEFAULT -> 6000#; + 0# -> 5000#; + 1# -> 1000#; + 2# -> 4000#; + 3# -> 2000#; + 4# -> 3000# + } + +sel :: Int -> Int +sel + = \ (ds :: Int) -> + case ds of { I# ww -> case $wsel ww of ww { __DEFAULT -> I# ww } } + ===================================== testsuite/tests/simplCore/should_compile/all.T ===================================== @@ -603,3 +603,4 @@ test('T25718c', normal, compile, ['-O -ddump-simpl -dsuppress-uniques -dsuppress test('T19166', normal, compile, ['-O -ddump-simpl -dsuppress-uniques -dsuppress-all -dno-typeable-binds']) test('T27261', [extra_files(['T27261_aux.hs'])], multimod_compile, ['T27261', '-v0 -O']) test('T27296', [], makefile_test, ['T27296']) +test('T27296b', [], makefile_test, ['T27296b']) View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/b925ad78dc1fb2836062e3414a59295b... -- View it on GitLab: https://gitlab.haskell.org/ghc/ghc/-/commit/b925ad78dc1fb2836062e3414a59295b... You're receiving this email because of your account on gitlab.haskell.org.