Marge Bot pushed to branch master at Glasgow Haskell Compiler / GHC

Commits:

14 changed files:

Changes:

  • compiler/GHC/Core/Opt/ConstantFold.hs
    ... ... @@ -2173,6 +2173,43 @@ builtinRules
    2173 2173 
     -- there is no benefit to inlining these yet, despite this, GHC produces
    2174 2174 
     -- unfoldings for this regardless since the floated list entries look small.
    2175 2175
    2176 
    +
    2177 
    +
    2178 
    +{- Note [Built-in bignum rules]
    2179 
    +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    2180 
    +We have some built-in rules for operations on bignum types (Integer, Natural,
    2181 
    +BigNat). These rules implement the same kind of constant folding as we have for
    2182 
    +Int#/Word#/etc. primops. See builtinBignumRules.
    2183 
    +
    2184 
    +These rules are built-in because they can't be expressed as regular rules for
    2185 
    +now. The reason is that due to the let-can-float invariant (see Note [Core
    2186 
    +let-can-float invariant] in GHC.Core), GHC is too conservative with some bignum
    2187 
    +operations and they don't match rules. For example:
    2188 
    +
    2189 
    +  case integerAdd 1 x of r { _ -> integerAdd 1 r }
    2190 
    +
    2191 
    +doesn't constant-fold into `integerAdd 2 x` with a regular rule. That's because
    2192 
    +GHC never floats in `integerAdd 1 x` to form `integerAdd 1 (integerAdd 1 x)`
    2193 
    +because of the let-can-float invariant (it doesn't know if `integerAdd`
    2194 
    +terminates).
    2195 
    +
    2196 
    +In the built-in rule for `integerAdd` we can access the unfolding of `r` and we
    2197 
    +can perform the appropriate substitution.
    2198 
    +
    2199 
    +To support constant-folding, Bignum operations are not allowed to inline. As a
    2200 
    +consequence, some codes that would benefit from inlining of bignum operations
    2201 
    +don't. An idea to fix this was to only have built-in rules for BigNat#
    2202 
    +operations and to allow Integer and Natural operations to inline. However these
    2203 
    +operations are often too big to inline and we end up with broken
    2204 
    +constant-folding and still no inlining. This issue is tracked in #20361
    2205 
    +
    2206 
    +Bignum built-in rules can be disabled independently of other built-in rules by
    2207 
    +passing the -dno-bignum-rules flag or programmatically with the `roBignumRules` field of
    2208 
    +RuleOpts.
    2209 
    +
    2210 
    +-}
    2211 
    +
    2212 
    +-- | Built-in bignum rules (see Note [Built-in bignum rules])
    2176 2213 
     builtinBignumRules :: [CoreRule]
    2177 2214 
     builtinBignumRules =
    2178 2215 
       [ -- conversions

  • compiler/GHC/Core/Rules.hs
    ... ... @@ -65,6 +65,7 @@ import GHC.Core.Map.Expr ( eqCoreExpr )
    65 65 
     import GHC.Core.Opt.Arity( etaExpandToJoinPointRule )
    66 66 
     import GHC.Core.Make     ( mkCoreLams )
    67 67 
     import GHC.Core.Opt.OccurAnal( occurAnalyseExpr )
    68 
    +import GHC.Core.Rules.Config (roBuiltinRules)
    68 69
    69 70 
     import GHC.Tc.Utils.TcType  ( tcSplitTyConApp_maybe )
    70 71 
     import GHC.Builtin.Types    ( anyTypeOfKind )
    ... ... @@ -708,10 +709,8 @@ matchRule :: HasDebugCallStack
    708 709
    709 710 
     matchRule opts rule_env _is_active fn args _rough_args
    710 711 
               (BuiltinRule { ru_try = match_fn })
    711 
    --- Built-in rules can't be switched off, it seems
    712 
    -  = case match_fn opts rule_env fn args of
    713 
    -        Nothing   -> Nothing
    714 
    -        Just expr -> Just expr
    712 
    +  | not (roBuiltinRules opts) = Nothing
    713 
    +  | otherwise                 = match_fn opts rule_env fn args
    715 714
    716 715 
     matchRule _ rule_env is_active _ args rough_args
    717 716 
               (Rule { ru_name = rule_name, ru_act = act, ru_rough = tpl_tops

  • compiler/GHC/Core/Rules/Config.hs
    ... ... @@ -14,6 +14,10 @@ data RuleOpts = RuleOpts
    14 14 
        , roExcessRationalPrecision :: !Bool
    15 15 
          -- ^ Cut down precision of Rational values to that of Float/Double if disabled
    16 16 
        , roBignumRules             :: !Bool
    17 
    -     -- ^ Enable rules for bignums
    17 
    +     -- ^ Enable built-in bignum rules (requires roBuiltinRules to be True too)
    18 
    +     --
    19 
    +     -- See Note [Built-in bignum rules] in GHC.Core.Opt.ConstantFold
    20 
    +   , roBuiltinRules            :: !Bool
    21 
    +     -- ^ Enable or disable all builtin rules (including bignum rules)
    18 22 
        }
    19 23

  • compiler/GHC/Driver/Config/Core/Rules.hs
    ... ... @@ -15,5 +15,6 @@ initRuleOpts dflags = RuleOpts
    15 15 
       { roPlatform                = targetPlatform dflags
    16 16 
       , roNumConstantFolding      = gopt Opt_NumConstantFolding dflags
    17 17 
       , roExcessRationalPrecision = gopt Opt_ExcessPrecision dflags
    18 
    -  , roBignumRules             = True
    18 
    +  , roBignumRules             = not (gopt Opt_NoBignumRules dflags)
    19 
    +  , roBuiltinRules            = not (gopt Opt_NoBuiltinRules dflags)
    19 20 
       }

  • compiler/GHC/Driver/Flags.hs
    ... ... @@ -592,6 +592,8 @@ data GeneralFlag
    592 592 
        | Opt_NoLlvmMangler                  -- hidden flag
    593 593 
        | Opt_FastLlvm                       -- hidden flag
    594 594 
        | Opt_NoTypeableBinds
    595 
    +   | Opt_NoBuiltinRules
    596 
    +   | Opt_NoBignumRules
    595 597
    596 598 
        | Opt_DistinctConstructorTables
    597 599 
        | Opt_InfoTableMap
    ... ... @@ -973,6 +975,8 @@ codeGenFlags = EnumSet.fromList
    973 975 
        , Opt_ExposeAllUnfoldings
    974 976 
        , Opt_ExposeOverloadedUnfoldings
    975 977 
        , Opt_NoTypeableBinds
    978 
    +   , Opt_NoBuiltinRules
    979 
    +   , Opt_NoBignumRules
    976 980 
        , Opt_ObjectDeterminism
    977 981 
        , Opt_Haddock
    978 982

  • compiler/GHC/Driver/Session.hs
    ... ... @@ -1660,6 +1660,10 @@ dynamic_flags_deps = [
    1660 1660 
             (NoArg (setGeneralFlag Opt_NoLlvmMangler)) -- hidden flag
    1661 1661 
       , make_ord_flag defGhcFlag "dno-typeable-binds"
    1662 1662 
             (NoArg (setGeneralFlag Opt_NoTypeableBinds))
    1663 
    +  , make_ord_flag defGhcFlag "dno-builtin-rules"
    1664 
    +        (NoArg (setGeneralFlag Opt_NoBuiltinRules))
    1665 
    +  , make_ord_flag defGhcFlag "dno-bignum-rules"
    1666 
    +        (NoArg (setGeneralFlag Opt_NoBignumRules))
    1663 1667 
       , make_ord_flag defGhcFlag "ddump-debug"
    1664 1668 
             (setDumpFlag Opt_D_dump_debug)
    1665 1669 
       , make_dep_flag defGhcFlag "ddump-json"

  • docs/users_guide/debugging.rst
    ... ... @@ -1225,6 +1225,18 @@ Other
    1225 1225 
         compiler will panic if you try to use Typeable instances of things that you
    1226 1226 
         built with this flag.
    1227 1227
    1228 
    +.. ghc-flag:: -dno-builtin-rules
    1229 
    +    :shortdesc: Disable all built-in rewrite rules
    1230 
    +    :type: dynamic
    1231 
    +
    1232 
    +    This disables all the built-in rewrite rules. Mostly useful for debugging.
    1233 
    +
    1234 
    +.. ghc-flag:: -dno-bignum-rules
    1235 
    +    :shortdesc: Disable bignum built-in rewrite rules
    1236 
    +    :type: dynamic
    1237 
    +
    1238 
    +    This disables bignum built-in rewrite rules. Mostly useful for debugging.
    1239 
    +
    1228 1240 
     .. ghc-flag:: -dtag-inference-checks
    1229 1241 
         :shortdesc: Affirm tag inference results are correct at runtime.
    1230 1242 
         :type: dynamic

  • testsuite/tests/codeGen/should_compile/T20298a.hs
    1 
    +module T20298a where
    2 
    +
    3 
    +foo :: Integer
    4 
    +foo = 10 + 20

  • testsuite/tests/codeGen/should_compile/T20298a.stderr
    1 
    +
    2 
    +==================== Tidy Core ====================
    3 
    +Result size of Tidy Core
    4 
    +  = {terms: 3, types: 1, coercions: 0, joins: 0/0}
    5 
    +
    6 
    +foo = IS 30#
    7 
    +
    8 
    +
    9 
    +

  • testsuite/tests/codeGen/should_compile/T20298b.hs
    1 
    +module T20298b where
    2 
    +
    3 
    +foo :: Integer
    4 
    +foo = 10 + 20

  • testsuite/tests/codeGen/should_compile/T20298b.stderr
    1 
    +
    2 
    +==================== Tidy Core ====================
    3 
    +Result size of Tidy Core
    4 
    +  = {terms: 10, types: 3, coercions: 0, joins: 0/0}
    5 
    +
    6 
    +foo2 = IS 10#
    7 
    +
    8 
    +foo1 = IS 20#
    9 
    +
    10 
    +foo = integerAdd foo2 foo1
    11 
    +
    12 
    +
    13 
    +

  • testsuite/tests/codeGen/should_compile/T20298c.hs
    1 
    +module T20298c where
    2 
    +
    3 
    +foo :: Integer
    4 
    +foo = 10 + 20

  • testsuite/tests/codeGen/should_compile/T20298c.stderr
    1 
    +
    2 
    +==================== Tidy Core ====================
    3 
    +Result size of Tidy Core
    4 
    +  = {terms: 11, types: 4, coercions: 0, joins: 0/0}
    5 
    +
    6 
    +foo2 = IS 10#
    7 
    +
    8 
    +foo1 = IS 20#
    9 
    +
    10 
    +foo = + $fNumInteger foo2 foo1
    11 
    +
    12 
    +
    13 
    +

  • testsuite/tests/codeGen/should_compile/all.T
    ... ... @@ -145,3 +145,7 @@ test('T25166', [req_cmm], makefile_test, [])
    145 145 
     test('T25177', normal, compile, ['-O2 -dno-typeable-binds -ddump-simpl -dsuppress-all -dsuppress-uniques -v0'])
    146 146
    147 147 
     test('T16351', normal, compile, ['-O2 -ddump-simpl -dno-typeable-binds -dsuppress-all -dsuppress-uniques'])
    148 
    +
    149 
    +test('T20298a', normal, compile, ['-O2 -ddump-simpl -dno-typeable-binds -dsuppress-all -dsuppress-uniques'])
    150 
    +test('T20298b', normal, compile, ['-O2 -dno-bignum-rules -ddump-simpl -dno-typeable-binds -dsuppress-all -dsuppress-uniques'])
    151 
    +test('T20298c', normal, compile, ['-O2 -dno-builtin-rules -ddump-simpl -dno-typeable-binds -dsuppress-all -dsuppress-uniques'])

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