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.
15 module PrelRules ( primOpRule, builtinRules ) where
17 #include "HsVersions.h"
20 import Id ( mkWildId )
21 import Literal ( Literal(..), isLitLitLit, mkMachInt, mkMachWord
23 , word2IntLit, int2WordLit
24 , intToInt8Lit, intToInt16Lit, intToInt32Lit
25 , wordToWord8Lit, wordToWord16Lit, wordToWord32Lit
26 , char2IntLit, int2CharLit
27 , float2IntLit, int2FloatLit, double2IntLit, int2DoubleLit
28 , addr2IntLit, int2AddrLit, float2DoubleLit, double2FloatLit
30 import PrimOp ( PrimOp(..), primOpOcc )
31 import TysWiredIn ( trueDataConId, falseDataConId )
32 import TyCon ( tyConDataConsIfAvailable, isEnumerationTyCon, isNewTyCon )
33 import DataCon ( dataConTag, dataConTyCon, dataConId, fIRST_TAG )
34 import CoreUtils ( exprIsValue, cheapEqExpr, exprIsConApp_maybe )
35 import Type ( tyConAppTyCon )
36 import OccName ( occNameUserString)
37 import PrelNames ( unpackCStringFoldrName, unpackCStringFoldrIdKey, hasKey )
39 import Bits ( Bits(..) )
40 #if __GLASGOW_HASKELL__ >= 500
43 import Word ( Word64 )
46 import CmdLineOpts ( opt_SimplExcessPrecision )
51 primOpRule :: PrimOp -> CoreRule
53 = BuiltinRule (primop_rule op)
55 op_name = _PK_ (occNameUserString (primOpOcc op))
56 op_name_case = op_name _APPEND_ SLIT("->case")
58 -- ToDo: something for integer-shift ops?
61 primop_rule SeqOp = seqRule
62 primop_rule TagToEnumOp = tagToEnumRule
63 primop_rule DataToTagOp = dataToTagRule
66 primop_rule IntAddOp = twoLits (intOp2 (+) op_name)
67 primop_rule IntSubOp = twoLits (intOp2 (-) op_name)
68 primop_rule IntMulOp = twoLits (intOp2 (*) op_name)
69 primop_rule IntQuotOp = twoLits (intOp2Z quot op_name)
70 primop_rule IntRemOp = twoLits (intOp2Z rem op_name)
71 primop_rule IntNegOp = oneLit (negOp op_name)
74 #if __GLASGOW_HASKELL__ >= 500
75 primop_rule WordAddOp = twoLits (wordOp2 (+) op_name)
76 primop_rule WordSubOp = twoLits (wordOp2 (-) op_name)
77 primop_rule WordMulOp = twoLits (wordOp2 (*) op_name)
79 primop_rule WordQuotOp = twoLits (wordOp2Z quot op_name)
80 primop_rule WordRemOp = twoLits (wordOp2Z rem op_name)
81 #if __GLASGOW_HASKELL__ >= 407
82 primop_rule AndOp = twoLits (wordBitOp2 (.&.) op_name)
83 primop_rule OrOp = twoLits (wordBitOp2 (.|.) op_name)
84 primop_rule XorOp = twoLits (wordBitOp2 xor op_name)
88 primop_rule Word2IntOp = oneLit (litCoerce word2IntLit op_name)
89 primop_rule Int2WordOp = oneLit (litCoerce int2WordLit op_name)
90 primop_rule IntToInt8Op = oneLit (litCoerce intToInt8Lit op_name)
91 primop_rule IntToInt16Op = oneLit (litCoerce intToInt16Lit op_name)
92 primop_rule IntToInt32Op = oneLit (litCoerce intToInt32Lit op_name)
93 primop_rule WordToWord8Op = oneLit (litCoerce wordToWord8Lit op_name)
94 primop_rule WordToWord16Op = oneLit (litCoerce wordToWord16Lit op_name)
95 primop_rule WordToWord32Op = oneLit (litCoerce wordToWord32Lit op_name)
96 primop_rule OrdOp = oneLit (litCoerce char2IntLit op_name)
97 primop_rule ChrOp = oneLit (litCoerce int2CharLit op_name)
98 primop_rule Float2IntOp = oneLit (litCoerce float2IntLit op_name)
99 primop_rule Int2FloatOp = oneLit (litCoerce int2FloatLit op_name)
100 primop_rule Double2IntOp = oneLit (litCoerce double2IntLit op_name)
101 primop_rule Int2DoubleOp = oneLit (litCoerce int2DoubleLit op_name)
102 primop_rule Addr2IntOp = oneLit (litCoerce addr2IntLit op_name)
103 primop_rule Int2AddrOp = oneLit (litCoerce int2AddrLit op_name)
104 -- SUP: Not sure what the standard says about precision in the following 2 cases
105 primop_rule Float2DoubleOp = oneLit (litCoerce float2DoubleLit op_name)
106 primop_rule Double2FloatOp = oneLit (litCoerce double2FloatLit op_name)
109 primop_rule FloatAddOp = twoLits (floatOp2 (+) op_name)
110 primop_rule FloatSubOp = twoLits (floatOp2 (-) op_name)
111 primop_rule FloatMulOp = twoLits (floatOp2 (*) op_name)
112 primop_rule FloatDivOp = twoLits (floatOp2Z (/) op_name)
113 primop_rule FloatNegOp = oneLit (negOp op_name)
116 primop_rule DoubleAddOp = twoLits (doubleOp2 (+) op_name)
117 primop_rule DoubleSubOp = twoLits (doubleOp2 (-) op_name)
118 primop_rule DoubleMulOp = twoLits (doubleOp2 (*) op_name)
119 primop_rule DoubleDivOp = twoLits (doubleOp2Z (/) op_name)
120 primop_rule DoubleNegOp = oneLit (negOp op_name)
122 -- Relational operators
123 primop_rule IntEqOp = relop (==) `or_rule` litEq True op_name_case
124 primop_rule IntNeOp = relop (/=) `or_rule` litEq False op_name_case
125 primop_rule CharEqOp = relop (==) `or_rule` litEq True op_name_case
126 primop_rule CharNeOp = relop (/=) `or_rule` litEq False op_name_case
128 primop_rule IntGtOp = relop (>)
129 primop_rule IntGeOp = relop (>=)
130 primop_rule IntLeOp = relop (<=)
131 primop_rule IntLtOp = relop (<)
133 primop_rule CharGtOp = relop (>)
134 primop_rule CharGeOp = relop (>=)
135 primop_rule CharLeOp = relop (<=)
136 primop_rule CharLtOp = relop (<)
138 primop_rule FloatGtOp = relop (>)
139 primop_rule FloatGeOp = relop (>=)
140 primop_rule FloatLeOp = relop (<=)
141 primop_rule FloatLtOp = relop (<)
142 primop_rule FloatEqOp = relop (==)
143 primop_rule FloatNeOp = relop (/=)
145 primop_rule DoubleGtOp = relop (>)
146 primop_rule DoubleGeOp = relop (>=)
147 primop_rule DoubleLeOp = relop (<=)
148 primop_rule DoubleLtOp = relop (<)
149 primop_rule DoubleEqOp = relop (==)
150 primop_rule DoubleNeOp = relop (/=)
152 primop_rule WordGtOp = relop (>)
153 primop_rule WordGeOp = relop (>=)
154 primop_rule WordLeOp = relop (<=)
155 primop_rule WordLtOp = relop (<)
156 primop_rule WordEqOp = relop (==)
157 primop_rule WordNeOp = relop (/=)
159 primop_rule other = \args -> Nothing
162 relop cmp = twoLits (cmpOp (\ord -> ord `cmp` EQ) op_name)
163 -- Cunning. cmpOp compares the values to give an Ordering.
164 -- It applies its argument to that ordering value to turn
165 -- the ordering into a boolean value. (`cmp` EQ) is just the job.
168 %************************************************************************
170 \subsection{Doing the business}
172 %************************************************************************
176 In all these operations we might find a LitLit as an operand; that's
177 why we have the catch-all Nothing case.
180 --------------------------
181 litCoerce :: (Literal -> Literal) -> RuleName -> Literal -> Maybe (RuleName, CoreExpr)
182 litCoerce fn name lit | isLitLitLit lit = Nothing
183 | otherwise = Just (name, Lit (fn lit))
185 --------------------------
186 cmpOp :: (Ordering -> Bool) -> FAST_STRING -> Literal -> Literal -> Maybe (RuleName, CoreExpr)
190 done res | cmp res = Just (name, trueVal)
191 | otherwise = Just (name, falseVal)
193 -- These compares are at different types
194 go (MachChar i1) (MachChar i2) = done (i1 `compare` i2)
195 go (MachInt i1) (MachInt i2) = done (i1 `compare` i2)
196 go (MachInt64 i1) (MachInt64 i2) = done (i1 `compare` i2)
197 go (MachWord i1) (MachWord i2) = done (i1 `compare` i2)
198 go (MachWord64 i1) (MachWord64 i2) = done (i1 `compare` i2)
199 go (MachFloat i1) (MachFloat i2) = done (i1 `compare` i2)
200 go (MachDouble i1) (MachDouble i2) = done (i1 `compare` i2)
203 --------------------------
205 negOp name (MachFloat f) = Just (name, mkFloatVal (-f))
206 negOp name (MachDouble d) = Just (name, mkDoubleVal (-d))
207 negOp name (MachInt i) = intResult name (-i)
208 negOp name l = Nothing
210 --------------------------
211 intOp2 op name (MachInt i1) (MachInt i2)
212 = intResult name (i1 `op` i2)
213 intOp2 op name l1 l2 = Nothing -- Could find LitLit
215 intOp2Z op name (MachInt i1) (MachInt i2)
216 | i2 /= 0 = Just (name, mkIntVal (i1 `op` i2))
217 intOp2Z op name l1 l2 = Nothing -- LitLit or zero dividend
219 --------------------------
220 #if __GLASGOW_HASKELL__ >= 500
221 wordOp2 op name (MachWord w1) (MachWord w2)
222 = wordResult name (w1 `op` w2)
223 wordOp2 op name l1 l2 = Nothing -- Could find LitLit
226 wordOp2Z op name (MachWord w1) (MachWord w2)
227 | w2 /= 0 = Just (name, mkWordVal (w1 `op` w2))
228 wordOp2Z op name l1 l2 = Nothing -- LitLit or zero dividend
230 #if __GLASGOW_HASKELL__ >= 500
231 wordBitOp2 op name l1@(MachWord w1) l2@(MachWord w2)
232 = Just (name, mkWordVal (w1 `op` w2))
234 -- Integer is not an instance of Bits, so we operate on Word64
235 wordBitOp2 op name l1@(MachWord w1) l2@(MachWord w2)
236 = Just (name, mkWordVal ((fromIntegral::Word64->Integer) (fromIntegral w1 `op` fromIntegral w2)))
238 wordBitOp2 op name l1 l2 = Nothing -- Could find LitLit
240 --------------------------
241 floatOp2 op name (MachFloat f1) (MachFloat f2)
242 = Just (name, mkFloatVal (f1 `op` f2))
243 floatOp2 op name l1 l2 = Nothing
245 floatOp2Z op name (MachFloat f1) (MachFloat f2)
246 | f2 /= 0 = Just (name, mkFloatVal (f1 `op` f2))
247 floatOp2Z op name l1 l2 = Nothing
249 --------------------------
250 doubleOp2 op name (MachDouble f1) (MachDouble f2)
251 = Just (name, mkDoubleVal (f1 `op` f2))
252 doubleOp2 op name l1 l2 = Nothing
254 doubleOp2Z op name (MachDouble f1) (MachDouble f2)
255 | f2 /= 0 = Just (name, mkDoubleVal (f1 `op` f2))
256 doubleOp2Z op name l1 l2 = Nothing
259 --------------------------
267 -- This is a Good Thing, because it allows case-of case things
268 -- to happen, and case-default absorption to happen. For
271 -- if (n ==# 3#) || (n ==# 4#) then e1 else e2
277 -- (modulo the usual precautions to avoid duplicating e1)
279 litEq :: Bool -- True <=> equality, False <=> inequality
282 litEq is_eq name [Lit lit, expr] = do_lit_eq is_eq name lit expr
283 litEq is_eq name [expr, Lit lit] = do_lit_eq is_eq name lit expr
284 litEq is_eq name other = Nothing
286 do_lit_eq is_eq name lit expr
287 = Just (name, Case expr (mkWildId (literalType lit))
288 [(LitAlt lit, [], val_if_eq),
289 (DEFAULT, [], val_if_neq)])
291 val_if_eq | is_eq = trueVal
292 | otherwise = falseVal
293 val_if_neq | is_eq = falseVal
294 | otherwise = trueVal
296 -- Note that we *don't* warn the user about overflow. It's not done at
297 -- runtime either, and compilation of completely harmless things like
298 -- ((124076834 :: Word32) + (2147483647 :: Word32))
299 -- would yield a warning. Instead we simply squash the value into the
300 -- Int range, but not in a way suitable for cross-compiling... :-(
301 intResult :: RuleName -> Integer -> Maybe (RuleName, CoreExpr)
302 intResult name result
303 = Just (name, mkIntVal (toInteger (fromInteger result :: Int)))
305 #if __GLASGOW_HASKELL__ >= 500
306 wordResult :: RuleName -> Integer -> Maybe (RuleName, CoreExpr)
307 wordResult name result
308 = Just (name, mkWordVal (toInteger (fromInteger result :: Word)))
313 %************************************************************************
315 \subsection{Vaguely generic functions
317 %************************************************************************
320 type RuleFun = [CoreExpr] -> Maybe (RuleName, CoreExpr)
322 or_rule :: RuleFun -> RuleFun -> RuleFun
323 or_rule r1 r2 args = maybe (r2 args) Just (r1 args) -- i.e.: r1 args `mplus` r2 args
325 twoLits :: (Literal -> Literal -> Maybe (RuleName, CoreExpr)) -> RuleFun
326 twoLits rule [Lit l1, Lit l2] = rule (convFloating l1) (convFloating l2)
327 twoLits rule other = Nothing
329 oneLit :: (Literal -> Maybe (RuleName, CoreExpr)) -> RuleFun
330 oneLit rule [Lit l1] = rule (convFloating l1)
331 oneLit rule other = Nothing
333 -- When excess precision is not requested, cut down the precision of the
334 -- Rational value to that of Float/Double. We confuse host architecture
335 -- and target architecture here, but it's convenient (and wrong :-).
336 convFloating :: Literal -> Literal
337 convFloating (MachFloat f) | not opt_SimplExcessPrecision =
338 MachFloat (toRational ((fromRational f) :: Float ))
339 convFloating (MachDouble d) | not opt_SimplExcessPrecision =
340 MachDouble (toRational ((fromRational d) :: Double))
344 trueVal = Var trueDataConId
345 falseVal = Var falseDataConId
346 mkIntVal i = Lit (mkMachInt i)
347 mkWordVal w = Lit (mkMachWord w)
348 mkFloatVal f = Lit (convFloating (MachFloat f))
349 mkDoubleVal d = Lit (convFloating (MachDouble d))
353 %************************************************************************
355 \subsection{Special rules for seq, tagToEnum, dataToTag}
357 %************************************************************************
359 In the parallel world, we use _seq_ to control the order in which
360 certain expressions will be evaluated. Operationally, the expression
361 ``_seq_ a b'' evaluates a and then evaluates b. We have an inlining
362 for _seq_ which translates _seq_ to:
364 _seq_ = /\ a b -> \ x::a y::b -> case seq# x of { 0# -> parError#; _ -> y }
366 Now, we know that the seq# primitive will never return 0#, but we
367 don't let the simplifier know that. We also use a special error
368 value, parError#, which is *not* a bottoming Id, so as far as the
369 simplifier is concerned, we have to evaluate seq# a before we know
370 whether or not y will be evaluated.
372 If we didn't have the extra case, then after inlining the compiler might
374 f p q = case seq# p of { _ -> p+q }
376 If it sees that, it can see that f is strict in q, and hence it might
377 evaluate q before p! The "0# ->" case prevents this happening.
378 By having the parError# branch we make sure that anything in the
379 other branch stays there!
381 This is fine, but we'd like to get rid of the extraneous code. Hence,
382 we *do* let the simplifier know that seq# is strict in its argument.
383 As a result, we hope that `a' will be evaluated before seq# is called.
384 At this point, we have a very special and magical simpification which
385 says that ``seq# a'' can be immediately simplified to `1#' if we
386 know that `a' is already evaluated.
388 NB: If we ever do case-floating, we have an extra worry:
391 a' -> let b' = case seq# a of { True -> b; False -> parError# }
397 a' -> let b' = case True of { True -> b; False -> parError# }
411 The second case must never be floated outside of the first!
414 seqRule [Type ty, arg] | exprIsValue arg = Just (SLIT("Seq"), mkIntVal 1)
415 seqRule other = Nothing
420 tagToEnumRule [Type ty, Lit (MachInt i)]
421 = ASSERT( isEnumerationTyCon tycon )
422 case filter correct_tag (tyConDataConsIfAvailable tycon) of
425 [] -> Nothing -- Abstract type
426 (dc:rest) -> ASSERT( null rest )
427 Just (SLIT("TagToEnum"), Var (dataConId dc))
429 correct_tag dc = (dataConTag dc - fIRST_TAG) == tag
431 tycon = tyConAppTyCon ty
433 tagToEnumRule other = Nothing
436 For dataToTag#, we can reduce if either
438 (a) the argument is a constructor
439 (b) the argument is a variable whose unfolding is a known constructor
442 dataToTagRule [_, val_arg]
443 = case exprIsConApp_maybe val_arg of
444 Just (dc,_) -> ASSERT( not (isNewTyCon (dataConTyCon dc)) )
445 Just (SLIT("DataToTag"),
446 mkIntVal (toInteger (dataConTag dc - fIRST_TAG)))
450 dataToTagRule other = Nothing
453 %************************************************************************
455 \subsection{Built in rules}
457 %************************************************************************
460 builtinRules :: [(Name, CoreRule)]
461 -- Rules for non-primops that can't be expressed using a RULE pragma
463 = [ (unpackCStringFoldrName, BuiltinRule match_append_lit_str)
468 -- unpackFoldrCString# "foo" c (unpackFoldrCString# "baz" c n) = unpackFoldrCString# "foobaz" c n
470 match_append_lit_str [Type ty1,
473 Var unpk `App` Type ty2
474 `App` Lit (MachStr s2)
478 | unpk `hasKey` unpackCStringFoldrIdKey &&
480 = ASSERT( ty1 == ty2 )
481 Just (SLIT("AppendLitString"),
482 Var unpk `App` Type ty1
483 `App` Lit (MachStr (s1 _APPEND_ s2))
487 match_append_lit_str other = Nothing