2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[ConFold]{Constant Folder}
6 Conceptually, constant folding should be parameterized with the kind
7 of target machine to get identical behaviour during compilation time
8 and runtime. We cheat a little bit here...
11 check boundaries before folding, e.g. we can fold the Float addition
12 (i1 + i2) only if it results in a valid Float.
16 {-# OPTIONS -optc-DNON_POSIX_SOURCE #-}
19 -- The above warning supression flag is a temporary kludge.
20 -- While working on this module you are encouraged to remove it and fix
21 -- any warnings in the module. See
22 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
25 module PrelRules ( primOpRules, builtinRules ) where
27 #include "HsVersions.h"
30 import Id ( mkWildId, idUnfolding )
31 import Literal ( Literal(..), mkMachInt, mkMachWord
33 , word2IntLit, int2WordLit
34 , narrow8IntLit, narrow16IntLit, narrow32IntLit
35 , narrow8WordLit, narrow16WordLit, narrow32WordLit
36 , char2IntLit, int2CharLit
37 , float2IntLit, int2FloatLit, double2IntLit, int2DoubleLit
38 , float2DoubleLit, double2FloatLit, litFitsInChar
40 import PrimOp ( PrimOp(..), tagToEnumKey )
41 import TysWiredIn ( boolTy, trueDataConId, falseDataConId )
42 import TyCon ( tyConDataCons_maybe, isEnumerationTyCon, isNewTyCon )
43 import DataCon ( dataConTag, dataConTyCon, dataConWorkId, fIRST_TAG )
44 import CoreUtils ( cheapEqExpr, exprIsConApp_maybe )
45 import Type ( tyConAppTyCon, coreEqType )
46 import OccName ( occNameFS )
47 import PrelNames ( unpackCStringFoldrName, unpackCStringFoldrIdKey, hasKey,
48 eqStringName, unpackCStringIdKey, inlineIdName )
49 import Maybes ( orElse )
50 import Name ( Name, nameOccName )
53 import StaticFlags ( opt_SimplExcessPrecision )
54 import Data.Bits as Bits
55 import Data.Word ( Word )
59 Note [Constant folding]
60 ~~~~~~~~~~~~~~~~~~~~~~~
61 primOpRules generates the rewrite rules for each primop
62 These rules do what is often called "constant folding"
63 E.g. the rules for +# might say
65 Well, of course you'd need a lot of rules if you did it
66 like that, so we use a BuiltinRule instead, so that we
67 can match in any two literal values. So the rule is really
69 (Lit 4) +# (Lit y) = Lit (x+#y)
70 where the (+#) on the rhs is done at compile time
72 That is why these rules are built in here. Other rules
73 which don't need to be built in are in GHC.Base. For
79 primOpRules :: PrimOp -> Name -> [CoreRule]
80 primOpRules op op_name = primop_rule op
83 one_lit = oneLit op_name
84 two_lits = twoLits op_name
85 relop cmp = two_lits (cmpOp (\ord -> ord `cmp` EQ))
86 -- Cunning. cmpOp compares the values to give an Ordering.
87 -- It applies its argument to that ordering value to turn
88 -- the ordering into a boolean value. (`cmp` EQ) is just the job.
90 -- ToDo: something for integer-shift ops?
93 primop_rule TagToEnumOp = mkBasicRule op_name 2 tagToEnumRule
94 primop_rule DataToTagOp = mkBasicRule op_name 2 dataToTagRule
97 primop_rule IntAddOp = two_lits (intOp2 (+))
98 primop_rule IntSubOp = two_lits (intOp2 (-))
99 primop_rule IntMulOp = two_lits (intOp2 (*))
100 primop_rule IntQuotOp = two_lits (intOp2Z quot)
101 primop_rule IntRemOp = two_lits (intOp2Z rem)
102 primop_rule IntNegOp = one_lit negOp
103 primop_rule ISllOp = two_lits (intShiftOp2 Bits.shiftL)
104 primop_rule ISraOp = two_lits (intShiftOp2 Bits.shiftR)
105 primop_rule ISrlOp = two_lits (intShiftOp2 shiftRightLogical)
108 primop_rule WordAddOp = two_lits (wordOp2 (+))
109 primop_rule WordSubOp = two_lits (wordOp2 (-))
110 primop_rule WordMulOp = two_lits (wordOp2 (*))
111 primop_rule WordQuotOp = two_lits (wordOp2Z quot)
112 primop_rule WordRemOp = two_lits (wordOp2Z rem)
113 primop_rule AndOp = two_lits (wordBitOp2 (.&.))
114 primop_rule OrOp = two_lits (wordBitOp2 (.|.))
115 primop_rule XorOp = two_lits (wordBitOp2 xor)
116 primop_rule SllOp = two_lits (wordShiftOp2 Bits.shiftL)
117 primop_rule SrlOp = two_lits (wordShiftOp2 shiftRightLogical)
120 primop_rule Word2IntOp = one_lit (litCoerce word2IntLit)
121 primop_rule Int2WordOp = one_lit (litCoerce int2WordLit)
122 primop_rule Narrow8IntOp = one_lit (litCoerce narrow8IntLit)
123 primop_rule Narrow16IntOp = one_lit (litCoerce narrow16IntLit)
124 primop_rule Narrow32IntOp = one_lit (litCoerce narrow32IntLit)
125 primop_rule Narrow8WordOp = one_lit (litCoerce narrow8WordLit)
126 primop_rule Narrow16WordOp = one_lit (litCoerce narrow16WordLit)
127 primop_rule Narrow32WordOp = one_lit (litCoerce narrow32WordLit)
128 primop_rule OrdOp = one_lit (litCoerce char2IntLit)
129 primop_rule ChrOp = one_lit (predLitCoerce litFitsInChar int2CharLit)
130 primop_rule Float2IntOp = one_lit (litCoerce float2IntLit)
131 primop_rule Int2FloatOp = one_lit (litCoerce int2FloatLit)
132 primop_rule Double2IntOp = one_lit (litCoerce double2IntLit)
133 primop_rule Int2DoubleOp = one_lit (litCoerce int2DoubleLit)
134 -- SUP: Not sure what the standard says about precision in the following 2 cases
135 primop_rule Float2DoubleOp = one_lit (litCoerce float2DoubleLit)
136 primop_rule Double2FloatOp = one_lit (litCoerce double2FloatLit)
139 primop_rule FloatAddOp = two_lits (floatOp2 (+))
140 primop_rule FloatSubOp = two_lits (floatOp2 (-))
141 primop_rule FloatMulOp = two_lits (floatOp2 (*))
142 primop_rule FloatDivOp = two_lits (floatOp2Z (/))
143 primop_rule FloatNegOp = one_lit negOp
146 primop_rule DoubleAddOp = two_lits (doubleOp2 (+))
147 primop_rule DoubleSubOp = two_lits (doubleOp2 (-))
148 primop_rule DoubleMulOp = two_lits (doubleOp2 (*))
149 primop_rule DoubleDivOp = two_lits (doubleOp2Z (/))
150 primop_rule DoubleNegOp = one_lit negOp
152 -- Relational operators
153 primop_rule IntEqOp = relop (==) ++ litEq op_name True
154 primop_rule IntNeOp = relop (/=) ++ litEq op_name False
155 primop_rule CharEqOp = relop (==) ++ litEq op_name True
156 primop_rule CharNeOp = relop (/=) ++ litEq op_name False
158 primop_rule IntGtOp = relop (>)
159 primop_rule IntGeOp = relop (>=)
160 primop_rule IntLeOp = relop (<=)
161 primop_rule IntLtOp = relop (<)
163 primop_rule CharGtOp = relop (>)
164 primop_rule CharGeOp = relop (>=)
165 primop_rule CharLeOp = relop (<=)
166 primop_rule CharLtOp = relop (<)
168 primop_rule FloatGtOp = relop (>)
169 primop_rule FloatGeOp = relop (>=)
170 primop_rule FloatLeOp = relop (<=)
171 primop_rule FloatLtOp = relop (<)
172 primop_rule FloatEqOp = relop (==)
173 primop_rule FloatNeOp = relop (/=)
175 primop_rule DoubleGtOp = relop (>)
176 primop_rule DoubleGeOp = relop (>=)
177 primop_rule DoubleLeOp = relop (<=)
178 primop_rule DoubleLtOp = relop (<)
179 primop_rule DoubleEqOp = relop (==)
180 primop_rule DoubleNeOp = relop (/=)
182 primop_rule WordGtOp = relop (>)
183 primop_rule WordGeOp = relop (>=)
184 primop_rule WordLeOp = relop (<=)
185 primop_rule WordLtOp = relop (<)
186 primop_rule WordEqOp = relop (==)
187 primop_rule WordNeOp = relop (/=)
189 primop_rule other = []
194 %************************************************************************
196 \subsection{Doing the business}
198 %************************************************************************
200 ToDo: the reason these all return Nothing is because there used to be
201 the possibility of an argument being a litlit. Litlits are now gone,
202 so this could be cleaned up.
205 --------------------------
206 litCoerce :: (Literal -> Literal) -> Literal -> Maybe CoreExpr
207 litCoerce fn lit = Just (Lit (fn lit))
209 predLitCoerce :: (Literal -> Bool) -> (Literal -> Literal) -> Literal -> Maybe CoreExpr
210 predLitCoerce p fn lit
211 | p lit = Just (Lit (fn lit))
212 | otherwise = Nothing
214 --------------------------
215 cmpOp :: (Ordering -> Bool) -> Literal -> Literal -> Maybe CoreExpr
219 done res | cmp res = Just trueVal
220 | otherwise = Just falseVal
222 -- These compares are at different types
223 go (MachChar i1) (MachChar i2) = done (i1 `compare` i2)
224 go (MachInt i1) (MachInt i2) = done (i1 `compare` i2)
225 go (MachInt64 i1) (MachInt64 i2) = done (i1 `compare` i2)
226 go (MachWord i1) (MachWord i2) = done (i1 `compare` i2)
227 go (MachWord64 i1) (MachWord64 i2) = done (i1 `compare` i2)
228 go (MachFloat i1) (MachFloat i2) = done (i1 `compare` i2)
229 go (MachDouble i1) (MachDouble i2) = done (i1 `compare` i2)
232 --------------------------
234 negOp :: Literal -> Maybe CoreExpr -- Negate
235 negOp (MachFloat 0.0) = Nothing -- can't represent -0.0 as a Rational
236 negOp (MachFloat f) = Just (mkFloatVal (-f))
237 negOp (MachDouble 0.0) = Nothing
238 negOp (MachDouble d) = Just (mkDoubleVal (-d))
239 negOp (MachInt i) = intResult (-i)
242 --------------------------
243 intOp2 :: (Integer->Integer->Integer) -> Literal -> Literal -> Maybe CoreExpr
244 intOp2 op (MachInt i1) (MachInt i2) = intResult (i1 `op` i2)
245 intOp2 op l1 l2 = Nothing -- Could find LitLit
247 intOp2Z :: (Integer->Integer->Integer) -> Literal -> Literal -> Maybe CoreExpr
248 -- Like intOp2, but Nothing if i2=0
249 intOp2Z op (MachInt i1) (MachInt i2)
250 | i2 /= 0 = intResult (i1 `op` i2)
251 intOp2Z op l1 l2 = Nothing -- LitLit or zero dividend
253 intShiftOp2 :: (Integer->Int->Integer) -> Literal -> Literal -> Maybe CoreExpr
254 -- Shifts take an Int; hence second arg of op is Int
255 intShiftOp2 op (MachInt i1) (MachInt i2) = intResult (i1 `op` fromInteger i2)
256 intShiftOp2 op l1 l2 = Nothing
258 shiftRightLogical :: Integer -> Int -> Integer
259 -- Shift right, putting zeros in rather than sign-propagating as Bits.shiftR would do
260 -- Do this by converting to Word and back. Obviously this won't work for big
261 -- values, but its ok as we use it here
262 shiftRightLogical x n = fromIntegral (fromInteger x `shiftR` n :: Word)
265 --------------------------
266 wordOp2 :: (Integer->Integer->Integer) -> Literal -> Literal -> Maybe CoreExpr
267 wordOp2 op (MachWord w1) (MachWord w2)
268 = wordResult (w1 `op` w2)
269 wordOp2 op l1 l2 = Nothing -- Could find LitLit
271 wordOp2Z :: (Integer->Integer->Integer) -> Literal -> Literal -> Maybe CoreExpr
272 wordOp2Z op (MachWord w1) (MachWord w2)
273 | w2 /= 0 = wordResult (w1 `op` w2)
274 wordOp2Z op l1 l2 = Nothing -- LitLit or zero dividend
276 wordBitOp2 op l1@(MachWord w1) l2@(MachWord w2)
277 = wordResult (w1 `op` w2)
278 wordBitOp2 op l1 l2 = Nothing -- Could find LitLit
280 wordShiftOp2 :: (Integer->Int->Integer) -> Literal -> Literal -> Maybe CoreExpr
281 -- Shifts take an Int; hence second arg of op is Int
282 wordShiftOp2 op (MachWord x) (MachInt n)
283 = wordResult (x `op` fromInteger n)
284 -- Do the shift at type Integer
285 wordShiftOp2 op l1 l2 = Nothing
287 --------------------------
288 floatOp2 op (MachFloat f1) (MachFloat f2)
289 = Just (mkFloatVal (f1 `op` f2))
290 floatOp2 op l1 l2 = Nothing
292 floatOp2Z op (MachFloat f1) (MachFloat f2)
293 | f2 /= 0 = Just (mkFloatVal (f1 `op` f2))
294 floatOp2Z op l1 l2 = Nothing
296 --------------------------
297 doubleOp2 op (MachDouble f1) (MachDouble f2)
298 = Just (mkDoubleVal (f1 `op` f2))
299 doubleOp2 op l1 l2 = Nothing
301 doubleOp2Z op (MachDouble f1) (MachDouble f2)
302 | f2 /= 0 = Just (mkDoubleVal (f1 `op` f2))
303 doubleOp2Z op l1 l2 = Nothing
306 --------------------------
314 -- This is a Good Thing, because it allows case-of case things
315 -- to happen, and case-default absorption to happen. For
318 -- if (n ==# 3#) || (n ==# 4#) then e1 else e2
324 -- (modulo the usual precautions to avoid duplicating e1)
327 -> Bool -- True <=> equality, False <=> inequality
330 = [BuiltinRule { ru_name = occNameFS (nameOccName op_name)
331 `appendFS` FSLIT("->case"),
333 ru_nargs = 2, ru_try = rule_fn }]
335 rule_fn [Lit lit, expr] = do_lit_eq lit expr
336 rule_fn [expr, Lit lit] = do_lit_eq lit expr
337 rule_fn other = Nothing
340 = Just (Case expr (mkWildId (literalType lit)) boolTy
341 [(DEFAULT, [], val_if_neq),
342 (LitAlt lit, [], val_if_eq)])
343 val_if_eq | is_eq = trueVal
344 | otherwise = falseVal
345 val_if_neq | is_eq = falseVal
346 | otherwise = trueVal
348 -- Note that we *don't* warn the user about overflow. It's not done at
349 -- runtime either, and compilation of completely harmless things like
350 -- ((124076834 :: Word32) + (2147483647 :: Word32))
351 -- would yield a warning. Instead we simply squash the value into the
352 -- Int range, but not in a way suitable for cross-compiling... :-(
353 intResult :: Integer -> Maybe CoreExpr
355 = Just (mkIntVal (toInteger (fromInteger result :: Int)))
357 wordResult :: Integer -> Maybe CoreExpr
359 = Just (mkWordVal (toInteger (fromInteger result :: Word)))
363 %************************************************************************
365 \subsection{Vaguely generic functions
367 %************************************************************************
370 mkBasicRule :: Name -> Int -> ([CoreExpr] -> Maybe CoreExpr) -> [CoreRule]
371 -- Gives the Rule the same name as the primop itself
372 mkBasicRule op_name n_args rule_fn
373 = [BuiltinRule { ru_name = occNameFS (nameOccName op_name),
375 ru_nargs = n_args, ru_try = rule_fn }]
377 oneLit :: Name -> (Literal -> Maybe CoreExpr)
380 = mkBasicRule op_name 1 rule_fn
382 rule_fn [Lit l1] = test (convFloating l1)
385 twoLits :: Name -> (Literal -> Literal -> Maybe CoreExpr)
388 = mkBasicRule op_name 2 rule_fn
390 rule_fn [Lit l1, Lit l2] = test (convFloating l1) (convFloating l2)
393 -- When excess precision is not requested, cut down the precision of the
394 -- Rational value to that of Float/Double. We confuse host architecture
395 -- and target architecture here, but it's convenient (and wrong :-).
396 convFloating :: Literal -> Literal
397 convFloating (MachFloat f) | not opt_SimplExcessPrecision =
398 MachFloat (toRational ((fromRational f) :: Float ))
399 convFloating (MachDouble d) | not opt_SimplExcessPrecision =
400 MachDouble (toRational ((fromRational d) :: Double))
403 trueVal = Var trueDataConId
404 falseVal = Var falseDataConId
405 mkIntVal i = Lit (mkMachInt i)
406 mkWordVal w = Lit (mkMachWord w)
407 mkFloatVal f = Lit (convFloating (MachFloat f))
408 mkDoubleVal d = Lit (convFloating (MachDouble d))
412 %************************************************************************
414 \subsection{Special rules for seq, tagToEnum, dataToTag}
416 %************************************************************************
419 tagToEnumRule [Type ty, Lit (MachInt i)]
420 = ASSERT( isEnumerationTyCon tycon )
421 case filter correct_tag (tyConDataCons_maybe tycon `orElse` []) of
424 [] -> Nothing -- Abstract type
425 (dc:rest) -> ASSERT( null rest )
426 Just (Var (dataConWorkId dc))
428 correct_tag dc = (dataConTag dc - fIRST_TAG) == tag
430 tycon = tyConAppTyCon ty
432 tagToEnumRule other = Nothing
435 For dataToTag#, we can reduce if either
437 (a) the argument is a constructor
438 (b) the argument is a variable whose unfolding is a known constructor
441 dataToTagRule [Type ty1, Var tag_to_enum `App` Type ty2 `App` tag]
442 | tag_to_enum `hasKey` tagToEnumKey
443 , ty1 `coreEqType` ty2
444 = Just tag -- dataToTag (tagToEnum x) ==> x
446 dataToTagRule [_, val_arg]
447 | Just (dc,_) <- exprIsConApp_maybe val_arg
448 = ASSERT( not (isNewTyCon (dataConTyCon dc)) )
449 Just (mkIntVal (toInteger (dataConTag dc - fIRST_TAG)))
451 dataToTagRule other = Nothing
454 %************************************************************************
456 \subsection{Built in rules}
458 %************************************************************************
460 Note [Scoping for Builtin rules]
461 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
462 When compiling a (base-package) module that defines one of the
463 functions mentioned in the RHS of a built-in rule, there's a danger
466 f = ...(eq String x)....
468 ....and lower down...
472 Then a rewrite would give
474 f = ...(eqString x)...
475 ....and lower down...
478 and lo, eqString is not in scope. This only really matters when we get to code
479 generation. With -O we do a GlomBinds step that does a new SCC analysis on the whole
480 set of bindings, which sorts out the dependency. Without -O we don't do any rule
481 rewriting so again we are fine.
483 (This whole thing doesn't show up for non-built-in rules because their dependencies
488 builtinRules :: [CoreRule]
489 -- Rules for non-primops that can't be expressed using a RULE pragma
491 = [ BuiltinRule { ru_name = FSLIT("AppendLitString"), ru_fn = unpackCStringFoldrName,
492 ru_nargs = 4, ru_try = match_append_lit },
493 BuiltinRule { ru_name = FSLIT("EqString"), ru_fn = eqStringName,
494 ru_nargs = 2, ru_try = match_eq_string },
495 BuiltinRule { ru_name = FSLIT("Inline"), ru_fn = inlineIdName,
496 ru_nargs = 2, ru_try = match_inline }
500 ---------------------------------------------------
502 -- unpackFoldrCString# "foo" c (unpackFoldrCString# "baz" c n) = unpackFoldrCString# "foobaz" c n
504 match_append_lit [Type ty1,
507 Var unpk `App` Type ty2
508 `App` Lit (MachStr s2)
512 | unpk `hasKey` unpackCStringFoldrIdKey &&
514 = ASSERT( ty1 `coreEqType` ty2 )
515 Just (Var unpk `App` Type ty1
516 `App` Lit (MachStr (s1 `appendFS` s2))
520 match_append_lit other = Nothing
522 ---------------------------------------------------
524 -- eqString (unpackCString# (Lit s1)) (unpackCString# (Lit s2) = s1==s2
526 match_eq_string [Var unpk1 `App` Lit (MachStr s1),
527 Var unpk2 `App` Lit (MachStr s2)]
528 | unpk1 `hasKey` unpackCStringIdKey,
529 unpk2 `hasKey` unpackCStringIdKey
530 = Just (if s1 == s2 then trueVal else falseVal)
532 match_eq_string other = Nothing
535 ---------------------------------------------------
537 -- inline f_ty (f a b c) = <f's unfolding> a b c
538 -- (if f has an unfolding)
540 -- It's important to allow the argument to 'inline' to have args itself
541 -- (a) because its more forgiving to allow the programmer to write
543 -- or inline (f a b c)
544 -- (b) because a polymorphic f wll get a type argument that the
545 -- programmer can't avoid
547 -- Also, don't forget about 'inline's type argument!
548 match_inline (Type _ : e : _)
549 | (Var f, args1) <- collectArgs e,
550 Just unf <- maybeUnfoldingTemplate (idUnfolding f)
551 = Just (mkApps unf args1)
553 match_inline other = Nothing