2 -- The above warning supression flag is a temporary kludge.
3 -- While working on this module you are encouraged to remove it and fix
4 -- any warnings in the module. See
5 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
8 -----------------------------------------------------------------------------
10 -- Code generator utilities; mostly monadic
12 -- (c) The University of Glasgow 2004-2006
14 -----------------------------------------------------------------------------
19 emitDataLits, mkDataLits,
20 emitRODataLits, mkRODataLits,
21 emitIf, emitIfThenElse,
22 emitRtsCall, emitRtsCallWithVols, emitRtsCallWithResult,
23 assignNonPtrTemp, newNonPtrTemp,
24 assignPtrTemp, newPtrTemp,
26 emitSwitch, emitLitSwitch,
29 callerSaveVolatileRegs, get_GlobalReg_addr,
31 cmmAndWord, cmmOrWord, cmmNegate, cmmEqWord, cmmNeWord,
33 cmmOffsetExprW, cmmOffsetExprB,
34 cmmRegOffW, cmmRegOffB,
35 cmmLabelOffW, cmmLabelOffB,
36 cmmOffsetW, cmmOffsetB,
37 cmmOffsetLitW, cmmOffsetLitB,
39 cmmConstrTag, cmmConstrTag1,
41 tagForCon, tagCons, isSmallFamily,
42 cmmUntag, cmmIsTagged, cmmGetTag,
46 mkStringCLit, mkByteStringCLit,
53 #include "HsVersions.h"
54 #include "../includes/MachRegs.h"
63 import PprCmm ( {- instances -} )
70 import StgSyn (SRT(..))
85 -------------------------------------------------------------------------
87 -- Random small functions
89 -------------------------------------------------------------------------
91 addIdReps :: [Id] -> [(CgRep, Id)]
92 addIdReps ids = [(idCgRep id, id) | id <- ids]
94 -------------------------------------------------------------------------
98 -------------------------------------------------------------------------
100 cgLit :: Literal -> FCode CmmLit
101 cgLit (MachStr s) = mkByteStringCLit (bytesFS s)
102 -- not unpackFS; we want the UTF-8 byte stream.
103 cgLit other_lit = return (mkSimpleLit other_lit)
105 mkSimpleLit :: Literal -> CmmLit
106 mkSimpleLit (MachChar c) = CmmInt (fromIntegral (ord c)) wordRep
107 mkSimpleLit MachNullAddr = zeroCLit
108 mkSimpleLit (MachInt i) = CmmInt i wordRep
109 mkSimpleLit (MachInt64 i) = CmmInt i I64
110 mkSimpleLit (MachWord i) = CmmInt i wordRep
111 mkSimpleLit (MachWord64 i) = CmmInt i I64
112 mkSimpleLit (MachFloat r) = CmmFloat r F32
113 mkSimpleLit (MachDouble r) = CmmFloat r F64
114 mkSimpleLit (MachLabel fs ms) = CmmLabel (mkForeignLabel fs ms is_dyn)
116 is_dyn = False -- ToDo: fix me
118 mkLtOp :: Literal -> MachOp
119 -- On signed literals we must do a signed comparison
120 mkLtOp (MachInt _) = MO_S_Lt wordRep
121 mkLtOp (MachFloat _) = MO_S_Lt F32
122 mkLtOp (MachDouble _) = MO_S_Lt F64
123 mkLtOp lit = MO_U_Lt (cmmLitRep (mkSimpleLit lit))
126 ---------------------------------------------------
128 -- Cmm data type functions
130 ---------------------------------------------------
132 -----------------------
133 -- The "B" variants take byte offsets
134 cmmRegOffB :: CmmReg -> ByteOff -> CmmExpr
135 cmmRegOffB = cmmRegOff
137 cmmOffsetB :: CmmExpr -> ByteOff -> CmmExpr
138 cmmOffsetB = cmmOffset
140 cmmOffsetExprB :: CmmExpr -> CmmExpr -> CmmExpr
141 cmmOffsetExprB = cmmOffsetExpr
143 cmmLabelOffB :: CLabel -> ByteOff -> CmmLit
144 cmmLabelOffB = cmmLabelOff
146 cmmOffsetLitB :: CmmLit -> ByteOff -> CmmLit
147 cmmOffsetLitB = cmmOffsetLit
149 -----------------------
150 -- The "W" variants take word offsets
151 cmmOffsetExprW :: CmmExpr -> CmmExpr -> CmmExpr
152 -- The second arg is a *word* offset; need to change it to bytes
153 cmmOffsetExprW e (CmmLit (CmmInt n _)) = cmmOffsetW e (fromInteger n)
154 cmmOffsetExprW e wd_off = cmmIndexExpr wordRep e wd_off
156 cmmOffsetW :: CmmExpr -> WordOff -> CmmExpr
157 cmmOffsetW e n = cmmOffsetB e (wORD_SIZE * n)
159 cmmRegOffW :: CmmReg -> WordOff -> CmmExpr
160 cmmRegOffW reg wd_off = cmmRegOffB reg (wd_off * wORD_SIZE)
162 cmmOffsetLitW :: CmmLit -> WordOff -> CmmLit
163 cmmOffsetLitW lit wd_off = cmmOffsetLitB lit (wORD_SIZE * wd_off)
165 cmmLabelOffW :: CLabel -> WordOff -> CmmLit
166 cmmLabelOffW lbl wd_off = cmmLabelOffB lbl (wORD_SIZE * wd_off)
168 cmmLoadIndexW :: CmmExpr -> Int -> CmmExpr
169 cmmLoadIndexW base off
170 = CmmLoad (cmmOffsetW base off) wordRep
172 -----------------------
173 cmmNeWord, cmmEqWord, cmmOrWord, cmmAndWord :: CmmExpr -> CmmExpr -> CmmExpr
174 cmmOrWord e1 e2 = CmmMachOp mo_wordOr [e1, e2]
175 cmmAndWord e1 e2 = CmmMachOp mo_wordAnd [e1, e2]
176 cmmNeWord e1 e2 = CmmMachOp mo_wordNe [e1, e2]
177 cmmEqWord e1 e2 = CmmMachOp mo_wordEq [e1, e2]
178 cmmULtWord e1 e2 = CmmMachOp mo_wordULt [e1, e2]
179 cmmUGeWord e1 e2 = CmmMachOp mo_wordUGe [e1, e2]
180 cmmUGtWord e1 e2 = CmmMachOp mo_wordUGt [e1, e2]
181 --cmmShlWord e1 e2 = CmmMachOp mo_wordShl [e1, e2]
182 --cmmUShrWord e1 e2 = CmmMachOp mo_wordUShr [e1, e2]
183 cmmSubWord e1 e2 = CmmMachOp mo_wordSub [e1, e2]
185 cmmNegate :: CmmExpr -> CmmExpr
186 cmmNegate (CmmLit (CmmInt n rep)) = CmmLit (CmmInt (-n) rep)
187 cmmNegate e = CmmMachOp (MO_S_Neg (cmmExprRep e)) [e]
189 blankWord :: CmmStatic
190 blankWord = CmmUninitialised wORD_SIZE
194 --cmmTagBits = CmmLit (mkIntCLit tAG_BITS)
195 cmmTagMask = CmmLit (mkIntCLit tAG_MASK)
196 cmmPointerMask = CmmLit (mkIntCLit (complement tAG_MASK))
198 -- Used to untag a possibly tagged pointer
199 -- A static label need not be untagged
200 cmmUntag e@(CmmLit (CmmLabel _)) = e
202 cmmUntag e = (e `cmmAndWord` cmmPointerMask)
204 cmmGetTag e = (e `cmmAndWord` cmmTagMask)
206 -- Test if a closure pointer is untagged
207 cmmIsTagged e = (e `cmmAndWord` cmmTagMask)
208 `cmmNeWord` CmmLit zeroCLit
210 cmmConstrTag e = (e `cmmAndWord` cmmTagMask) `cmmSubWord` (CmmLit (mkIntCLit 1))
211 -- Get constructor tag, but one based.
212 cmmConstrTag1 e = e `cmmAndWord` cmmTagMask
215 The family size of a data type (the number of constructors)
217 * small, if the family size < 2**tag_bits
220 Small families can have the constructor tag in the tag
222 Big families only use the tag value 1 to represent
225 isSmallFamily fam_size = fam_size <= mAX_PTR_TAG
229 con_tag = dataConTagZ con
230 fam_size = tyConFamilySize (dataConTyCon con)
231 tag | isSmallFamily fam_size = con_tag + 1
234 --Tag an expression, to do: refactor, this appears in some other module.
235 tagCons con expr = cmmOffsetB expr (tagForCon con)
237 -- Copied from CgInfoTbls.hs
238 -- We keep the *zero-indexed* tag in the srt_len field of the info
239 -- table of a data constructor.
240 dataConTagZ :: DataCon -> ConTagZ
241 dataConTagZ con = dataConTag con - fIRST_TAG
243 -----------------------
246 mkWordCLit :: StgWord -> CmmLit
247 mkWordCLit wd = CmmInt (fromIntegral wd) wordRep
249 packHalfWordsCLit :: (Integral a, Integral b) => a -> b -> CmmLit
250 -- Make a single word literal in which the lower_half_word is
251 -- at the lower address, and the upper_half_word is at the
253 -- ToDo: consider using half-word lits instead
254 -- but be careful: that's vulnerable when reversed
255 packHalfWordsCLit lower_half_word upper_half_word
256 #ifdef WORDS_BIGENDIAN
257 = mkWordCLit ((fromIntegral lower_half_word `shiftL` hALF_WORD_SIZE_IN_BITS)
258 .|. fromIntegral upper_half_word)
260 = mkWordCLit ((fromIntegral lower_half_word)
261 .|. (fromIntegral upper_half_word `shiftL` hALF_WORD_SIZE_IN_BITS))
264 --------------------------------------------------------------------------
266 -- Incrementing a memory location
268 --------------------------------------------------------------------------
270 addToMem :: MachRep -- rep of the counter
271 -> CmmExpr -- Address
272 -> Int -- What to add (a word)
274 addToMem rep ptr n = addToMemE rep ptr (CmmLit (CmmInt (toInteger n) rep))
276 addToMemE :: MachRep -- rep of the counter
277 -> CmmExpr -- Address
278 -> CmmExpr -- What to add (a word-typed expression)
281 = CmmStore ptr (CmmMachOp (MO_Add rep) [CmmLoad ptr rep, n])
283 -------------------------------------------------------------------------
285 -- Converting a closure tag to a closure for enumeration types
286 -- (this is the implementation of tagToEnum#).
288 -------------------------------------------------------------------------
290 tagToClosure :: TyCon -> CmmExpr -> CmmExpr
291 tagToClosure tycon tag
292 = CmmLoad (cmmOffsetExprW closure_tbl tag) wordRep
293 where closure_tbl = CmmLit (CmmLabel lbl)
294 lbl = mkClosureTableLabel (tyConName tycon)
296 -------------------------------------------------------------------------
298 -- Conditionals and rts calls
300 -------------------------------------------------------------------------
302 emitIf :: CmmExpr -- Boolean
305 -- Emit (if e then x)
306 -- ToDo: reverse the condition to avoid the extra branch instruction if possible
307 -- (some conditionals aren't reversible. eg. floating point comparisons cannot
308 -- be inverted because there exist some values for which both comparisons
309 -- return False, such as NaN.)
310 emitIf cond then_part
311 = do { then_id <- newLabelC
312 ; join_id <- newLabelC
313 ; stmtC (CmmCondBranch cond then_id)
314 ; stmtC (CmmBranch join_id)
320 emitIfThenElse :: CmmExpr -- Boolean
324 -- Emit (if e then x else y)
325 emitIfThenElse cond then_part else_part
326 = do { then_id <- newLabelC
327 ; else_id <- newLabelC
328 ; join_id <- newLabelC
329 ; stmtC (CmmCondBranch cond then_id)
331 ; stmtC (CmmBranch join_id)
337 emitRtsCall :: LitString -> [CmmKinded CmmExpr] -> Bool -> Code
338 emitRtsCall fun args safe = emitRtsCall' [] fun args Nothing safe
339 -- The 'Nothing' says "save all global registers"
341 emitRtsCallWithVols :: LitString -> [CmmKinded CmmExpr] -> [GlobalReg] -> Bool -> Code
342 emitRtsCallWithVols fun args vols safe
343 = emitRtsCall' [] fun args (Just vols) safe
345 emitRtsCallWithResult :: LocalReg -> MachHint -> LitString
346 -> [CmmKinded CmmExpr] -> Bool -> Code
347 emitRtsCallWithResult res hint fun args safe
348 = emitRtsCall' [CmmKinded res hint] fun args Nothing safe
350 -- Make a call to an RTS C procedure
354 -> [CmmKinded CmmExpr]
356 -> Bool -- True <=> CmmSafe call
358 emitRtsCall' res fun args vols safe = do
360 then getSRTInfo >>= (return . CmmSafe)
361 else return CmmUnsafe
363 stmtC (CmmCall target res args safety CmmMayReturn)
366 (caller_save, caller_load) = callerSaveVolatileRegs vols
367 target = CmmCallee fun_expr CCallConv
368 fun_expr = mkLblExpr (mkRtsCodeLabel fun)
370 -----------------------------------------------------------------------------
372 -- Caller-Save Registers
374 -----------------------------------------------------------------------------
376 -- Here we generate the sequence of saves/restores required around a
377 -- foreign call instruction.
379 -- TODO: reconcile with includes/Regs.h
380 -- * Regs.h claims that BaseReg should be saved last and loaded first
381 -- * This might not have been tickled before since BaseReg is callee save
382 -- * Regs.h saves SparkHd, ParkT1, SparkBase and SparkLim
383 callerSaveVolatileRegs :: Maybe [GlobalReg] -> ([CmmStmt], [CmmStmt])
384 callerSaveVolatileRegs vols = (caller_save, caller_load)
386 caller_save = foldr ($!) [] (map callerSaveGlobalReg regs_to_save)
387 caller_load = foldr ($!) [] (map callerRestoreGlobalReg regs_to_save)
389 system_regs = [Sp,SpLim,Hp,HpLim,CurrentTSO,CurrentNursery,
390 {-SparkHd,SparkTl,SparkBase,SparkLim,-}BaseReg ]
392 regs_to_save = system_regs ++ vol_list
394 vol_list = case vols of Nothing -> all_of_em; Just regs -> regs
396 all_of_em = [ VanillaReg n | n <- [0..mAX_Vanilla_REG] ]
397 ++ [ FloatReg n | n <- [0..mAX_Float_REG] ]
398 ++ [ DoubleReg n | n <- [0..mAX_Double_REG] ]
399 ++ [ LongReg n | n <- [0..mAX_Long_REG] ]
401 callerSaveGlobalReg reg next
403 CmmStore (get_GlobalReg_addr reg)
404 (CmmReg (CmmGlobal reg)) : next
407 callerRestoreGlobalReg reg next
409 CmmAssign (CmmGlobal reg)
410 (CmmLoad (get_GlobalReg_addr reg) (globalRegRep reg))
414 -- -----------------------------------------------------------------------------
417 -- We map STG registers onto appropriate CmmExprs. Either they map
418 -- to real machine registers or stored as offsets from BaseReg. Given
419 -- a GlobalReg, get_GlobalReg_addr always produces the
420 -- register table address for it.
421 -- (See also get_GlobalReg_reg_or_addr in MachRegs)
423 get_GlobalReg_addr :: GlobalReg -> CmmExpr
424 get_GlobalReg_addr BaseReg = regTableOffset 0
425 get_GlobalReg_addr mid = get_Regtable_addr_from_offset
426 (globalRegRep mid) (baseRegOffset mid)
428 -- Calculate a literal representing an offset into the register table.
429 -- Used when we don't have an actual BaseReg to offset from.
431 CmmLit (CmmLabelOff mkMainCapabilityLabel (oFFSET_Capability_r + n))
433 get_Regtable_addr_from_offset :: MachRep -> Int -> CmmExpr
434 get_Regtable_addr_from_offset rep offset =
436 CmmRegOff (CmmGlobal BaseReg) offset
438 regTableOffset offset
442 -- | Returns @True@ if this global register is stored in a caller-saves
445 callerSaves :: GlobalReg -> Bool
447 #ifdef CALLER_SAVES_Base
448 callerSaves BaseReg = True
450 #ifdef CALLER_SAVES_R1
451 callerSaves (VanillaReg 1) = True
453 #ifdef CALLER_SAVES_R2
454 callerSaves (VanillaReg 2) = True
456 #ifdef CALLER_SAVES_R3
457 callerSaves (VanillaReg 3) = True
459 #ifdef CALLER_SAVES_R4
460 callerSaves (VanillaReg 4) = True
462 #ifdef CALLER_SAVES_R5
463 callerSaves (VanillaReg 5) = True
465 #ifdef CALLER_SAVES_R6
466 callerSaves (VanillaReg 6) = True
468 #ifdef CALLER_SAVES_R7
469 callerSaves (VanillaReg 7) = True
471 #ifdef CALLER_SAVES_R8
472 callerSaves (VanillaReg 8) = True
474 #ifdef CALLER_SAVES_F1
475 callerSaves (FloatReg 1) = True
477 #ifdef CALLER_SAVES_F2
478 callerSaves (FloatReg 2) = True
480 #ifdef CALLER_SAVES_F3
481 callerSaves (FloatReg 3) = True
483 #ifdef CALLER_SAVES_F4
484 callerSaves (FloatReg 4) = True
486 #ifdef CALLER_SAVES_D1
487 callerSaves (DoubleReg 1) = True
489 #ifdef CALLER_SAVES_D2
490 callerSaves (DoubleReg 2) = True
492 #ifdef CALLER_SAVES_L1
493 callerSaves (LongReg 1) = True
495 #ifdef CALLER_SAVES_Sp
496 callerSaves Sp = True
498 #ifdef CALLER_SAVES_SpLim
499 callerSaves SpLim = True
501 #ifdef CALLER_SAVES_Hp
502 callerSaves Hp = True
504 #ifdef CALLER_SAVES_HpLim
505 callerSaves HpLim = True
507 #ifdef CALLER_SAVES_CurrentTSO
508 callerSaves CurrentTSO = True
510 #ifdef CALLER_SAVES_CurrentNursery
511 callerSaves CurrentNursery = True
513 callerSaves _ = False
516 -- -----------------------------------------------------------------------------
517 -- Information about global registers
519 baseRegOffset :: GlobalReg -> Int
521 baseRegOffset (VanillaReg 1) = oFFSET_StgRegTable_rR1
522 baseRegOffset (VanillaReg 2) = oFFSET_StgRegTable_rR2
523 baseRegOffset (VanillaReg 3) = oFFSET_StgRegTable_rR3
524 baseRegOffset (VanillaReg 4) = oFFSET_StgRegTable_rR4
525 baseRegOffset (VanillaReg 5) = oFFSET_StgRegTable_rR5
526 baseRegOffset (VanillaReg 6) = oFFSET_StgRegTable_rR6
527 baseRegOffset (VanillaReg 7) = oFFSET_StgRegTable_rR7
528 baseRegOffset (VanillaReg 8) = oFFSET_StgRegTable_rR8
529 baseRegOffset (VanillaReg 9) = oFFSET_StgRegTable_rR9
530 baseRegOffset (VanillaReg 10) = oFFSET_StgRegTable_rR10
531 baseRegOffset (FloatReg 1) = oFFSET_StgRegTable_rF1
532 baseRegOffset (FloatReg 2) = oFFSET_StgRegTable_rF2
533 baseRegOffset (FloatReg 3) = oFFSET_StgRegTable_rF3
534 baseRegOffset (FloatReg 4) = oFFSET_StgRegTable_rF4
535 baseRegOffset (DoubleReg 1) = oFFSET_StgRegTable_rD1
536 baseRegOffset (DoubleReg 2) = oFFSET_StgRegTable_rD2
537 baseRegOffset Sp = oFFSET_StgRegTable_rSp
538 baseRegOffset SpLim = oFFSET_StgRegTable_rSpLim
539 baseRegOffset (LongReg 1) = oFFSET_StgRegTable_rL1
540 baseRegOffset Hp = oFFSET_StgRegTable_rHp
541 baseRegOffset HpLim = oFFSET_StgRegTable_rHpLim
542 baseRegOffset CurrentTSO = oFFSET_StgRegTable_rCurrentTSO
543 baseRegOffset CurrentNursery = oFFSET_StgRegTable_rCurrentNursery
544 baseRegOffset HpAlloc = oFFSET_StgRegTable_rHpAlloc
545 baseRegOffset EagerBlackholeInfo = oFFSET_stgEagerBlackholeInfo
546 baseRegOffset GCEnter1 = oFFSET_stgGCEnter1
547 baseRegOffset GCFun = oFFSET_stgGCFun
548 baseRegOffset BaseReg = panic "baseRegOffset:BaseReg"
549 baseRegOffset _ = panic "baseRegOffset:other"
552 -------------------------------------------------------------------------
554 -- Strings generate a top-level data block
556 -------------------------------------------------------------------------
558 emitDataLits :: CLabel -> [CmmLit] -> Code
559 -- Emit a data-segment data block
560 emitDataLits lbl lits
561 = emitData Data (CmmDataLabel lbl : map CmmStaticLit lits)
563 mkDataLits :: CLabel -> [CmmLit] -> GenCmmTop CmmStatic info graph
564 -- Emit a data-segment data block
566 = CmmData Data (CmmDataLabel lbl : map CmmStaticLit lits)
568 emitRODataLits :: CLabel -> [CmmLit] -> Code
569 -- Emit a read-only data block
570 emitRODataLits lbl lits
571 = emitData section (CmmDataLabel lbl : map CmmStaticLit lits)
572 where section | any needsRelocation lits = RelocatableReadOnlyData
573 | otherwise = ReadOnlyData
574 needsRelocation (CmmLabel _) = True
575 needsRelocation (CmmLabelOff _ _) = True
576 needsRelocation _ = False
578 mkRODataLits :: CLabel -> [CmmLit] -> GenCmmTop CmmStatic info graph
579 mkRODataLits lbl lits
580 = CmmData section (CmmDataLabel lbl : map CmmStaticLit lits)
581 where section | any needsRelocation lits = RelocatableReadOnlyData
582 | otherwise = ReadOnlyData
583 needsRelocation (CmmLabel _) = True
584 needsRelocation (CmmLabelOff _ _) = True
585 needsRelocation _ = False
587 mkStringCLit :: String -> FCode CmmLit
588 -- Make a global definition for the string,
589 -- and return its label
590 mkStringCLit str = mkByteStringCLit (map (fromIntegral.ord) str)
592 mkByteStringCLit :: [Word8] -> FCode CmmLit
593 mkByteStringCLit bytes
594 = do { uniq <- newUnique
595 ; let lbl = mkStringLitLabel uniq
596 ; emitData ReadOnlyData [CmmDataLabel lbl, CmmString bytes]
597 ; return (CmmLabel lbl) }
599 -------------------------------------------------------------------------
601 -- Assigning expressions to temporaries
603 -------------------------------------------------------------------------
605 assignNonPtrTemp :: CmmExpr -> FCode CmmExpr
606 -- For a non-trivial expression, e, create a local
607 -- variable and assign the expression to it
609 | isTrivialCmmExpr e = return e
610 | otherwise = do { reg <- newNonPtrTemp (cmmExprRep e)
611 ; stmtC (CmmAssign (CmmLocal reg) e)
612 ; return (CmmReg (CmmLocal reg)) }
614 assignPtrTemp :: CmmExpr -> FCode CmmExpr
615 -- For a non-trivial expression, e, create a local
616 -- variable and assign the expression to it
618 | isTrivialCmmExpr e = return e
619 | otherwise = do { reg <- newPtrTemp (cmmExprRep e)
620 ; stmtC (CmmAssign (CmmLocal reg) e)
621 ; return (CmmReg (CmmLocal reg)) }
623 newNonPtrTemp :: MachRep -> FCode LocalReg
624 newNonPtrTemp rep = do { uniq <- newUnique; return (LocalReg uniq rep GCKindNonPtr) }
626 newPtrTemp :: MachRep -> FCode LocalReg
627 newPtrTemp rep = do { uniq <- newUnique; return (LocalReg uniq rep GCKindPtr) }
630 -------------------------------------------------------------------------
632 -- Building case analysis
634 -------------------------------------------------------------------------
637 :: CmmExpr -- Tag to switch on
638 -> [(ConTagZ, CgStmts)] -- Tagged branches
639 -> Maybe CgStmts -- Default branch (if any)
640 -> ConTagZ -> ConTagZ -- Min and Max possible values; behaviour
641 -- outside this range is undefined
644 -- ONLY A DEFAULT BRANCH: no case analysis to do
645 emitSwitch tag_expr [] (Just stmts) _ _
649 emitSwitch tag_expr branches mb_deflt lo_tag hi_tag
650 = -- Just sort the branches before calling mk_sritch
653 Nothing -> return Nothing
654 Just stmts -> do id <- forkCgStmts stmts; return (Just id)
656 ; dflags <- getDynFlags
657 ; let via_C | HscC <- hscTarget dflags = True
660 ; stmts <- mk_switch tag_expr (sortLe le branches)
661 mb_deflt_id lo_tag hi_tag via_C
665 (t1,_) `le` (t2,_) = t1 <= t2
668 mk_switch :: CmmExpr -> [(ConTagZ, CgStmts)]
669 -> Maybe BlockId -> ConTagZ -> ConTagZ -> Bool
672 -- SINGLETON TAG RANGE: no case analysis to do
673 mk_switch tag_expr [(tag,stmts)] _ lo_tag hi_tag via_C
675 = ASSERT( tag == lo_tag )
678 -- SINGLETON BRANCH, NO DEFUALT: no case analysis to do
679 mk_switch tag_expr [(tag,stmts)] Nothing lo_tag hi_tag via_C
681 -- The simplifier might have eliminated a case
682 -- so we may have e.g. case xs of
684 -- In that situation we can be sure the (:) case
685 -- can't happen, so no need to test
687 -- SINGLETON BRANCH: one equality check to do
688 mk_switch tag_expr [(tag,stmts)] (Just deflt) lo_tag hi_tag via_C
689 = return (CmmCondBranch cond deflt `consCgStmt` stmts)
691 cond = cmmNeWord tag_expr (CmmLit (mkIntCLit tag))
692 -- We have lo_tag < hi_tag, but there's only one branch,
693 -- so there must be a default
695 -- ToDo: we might want to check for the two branch case, where one of
696 -- the branches is the tag 0, because comparing '== 0' is likely to be
697 -- more efficient than other kinds of comparison.
699 -- DENSE TAG RANGE: use a switch statment.
701 -- We also use a switch uncoditionally when compiling via C, because
702 -- this will get emitted as a C switch statement and the C compiler
703 -- should do a good job of optimising it. Also, older GCC versions
704 -- (2.95 in particular) have problems compiling the complicated
705 -- if-trees generated by this code, so compiling to a switch every
706 -- time works around that problem.
708 mk_switch tag_expr branches mb_deflt lo_tag hi_tag via_C
709 | use_switch -- Use a switch
710 = do { branch_ids <- mapM forkCgStmts (map snd branches)
712 tagged_blk_ids = zip (map fst branches) (map Just branch_ids)
714 find_branch :: ConTagZ -> Maybe BlockId
715 find_branch i = assocDefault mb_deflt tagged_blk_ids i
717 -- NB. we have eliminated impossible branches at
718 -- either end of the range (see below), so the first
719 -- tag of a real branch is real_lo_tag (not lo_tag).
720 arms = [ find_branch i | i <- [real_lo_tag..real_hi_tag]]
722 switch_stmt = CmmSwitch (cmmOffset tag_expr (- real_lo_tag)) arms
724 ; ASSERT(not (all isNothing arms))
725 return (oneCgStmt switch_stmt)
728 -- if we can knock off a bunch of default cases with one if, then do so
729 | Just deflt <- mb_deflt, (lowest_branch - lo_tag) >= n_branches
730 = do { (assign_tag, tag_expr') <- assignNonPtrTemp' tag_expr
731 ; let cond = cmmULtWord tag_expr' (CmmLit (mkIntCLit lowest_branch))
732 branch = CmmCondBranch cond deflt
733 ; stmts <- mk_switch tag_expr' branches mb_deflt
734 lowest_branch hi_tag via_C
735 ; return (assign_tag `consCgStmt` (branch `consCgStmt` stmts))
738 | Just deflt <- mb_deflt, (hi_tag - highest_branch) >= n_branches
739 = do { (assign_tag, tag_expr') <- assignNonPtrTemp' tag_expr
740 ; let cond = cmmUGtWord tag_expr' (CmmLit (mkIntCLit highest_branch))
741 branch = CmmCondBranch cond deflt
742 ; stmts <- mk_switch tag_expr' branches mb_deflt
743 lo_tag highest_branch via_C
744 ; return (assign_tag `consCgStmt` (branch `consCgStmt` stmts))
747 | otherwise -- Use an if-tree
748 = do { (assign_tag, tag_expr') <- assignNonPtrTemp' tag_expr
749 -- To avoid duplication
750 ; lo_stmts <- mk_switch tag_expr' lo_branches mb_deflt
751 lo_tag (mid_tag-1) via_C
752 ; hi_stmts <- mk_switch tag_expr' hi_branches mb_deflt
754 ; hi_id <- forkCgStmts hi_stmts
755 ; let cond = cmmUGeWord tag_expr' (CmmLit (mkIntCLit mid_tag))
756 branch_stmt = CmmCondBranch cond hi_id
757 ; return (assign_tag `consCgStmt` (branch_stmt `consCgStmt` lo_stmts))
759 -- we test (e >= mid_tag) rather than (e < mid_tag), because
760 -- the former works better when e is a comparison, and there
761 -- are two tags 0 & 1 (mid_tag == 1). In this case, the code
762 -- generator can reduce the condition to e itself without
763 -- having to reverse the sense of the comparison: comparisons
764 -- can't always be easily reversed (eg. floating
767 use_switch = {- pprTrace "mk_switch" (
768 ppr tag_expr <+> text "n_tags:" <+> int n_tags <+>
769 text "branches:" <+> ppr (map fst branches) <+>
770 text "n_branches:" <+> int n_branches <+>
771 text "lo_tag:" <+> int lo_tag <+>
772 text "hi_tag:" <+> int hi_tag <+>
773 text "real_lo_tag:" <+> int real_lo_tag <+>
774 text "real_hi_tag:" <+> int real_hi_tag) $ -}
775 ASSERT( n_branches > 1 && n_tags > 1 )
776 n_tags > 2 && (via_C || (dense && big_enough))
777 -- up to 4 branches we use a decision tree, otherwise
778 -- a switch (== jump table in the NCG). This seems to be
779 -- optimal, and corresponds with what gcc does.
780 big_enough = n_branches > 4
781 dense = n_branches > (n_tags `div` 2)
782 n_branches = length branches
784 -- ignore default slots at each end of the range if there's
785 -- no default branch defined.
786 lowest_branch = fst (head branches)
787 highest_branch = fst (last branches)
790 | isNothing mb_deflt = lowest_branch
794 | isNothing mb_deflt = highest_branch
797 n_tags = real_hi_tag - real_lo_tag + 1
799 -- INVARIANT: Provided hi_tag > lo_tag (which is true)
800 -- lo_tag <= mid_tag < hi_tag
801 -- lo_branches have tags < mid_tag
802 -- hi_branches have tags >= mid_tag
804 (mid_tag,_) = branches !! (n_branches `div` 2)
805 -- 2 branches => n_branches `div` 2 = 1
806 -- => branches !! 1 give the *second* tag
807 -- There are always at least 2 branches here
809 (lo_branches, hi_branches) = span is_lo branches
810 is_lo (t,_) = t < mid_tag
814 | isTrivialCmmExpr e = return (CmmNop, e)
815 | otherwise = do { reg <- newNonPtrTemp (cmmExprRep e)
816 ; return (CmmAssign (CmmLocal reg) e, CmmReg (CmmLocal reg)) }
818 emitLitSwitch :: CmmExpr -- Tag to switch on
819 -> [(Literal, CgStmts)] -- Tagged branches
820 -> CgStmts -- Default branch (always)
821 -> Code -- Emit the code
822 -- Used for general literals, whose size might not be a word,
823 -- where there is always a default case, and where we don't know
824 -- the range of values for certain. For simplicity we always generate a tree.
826 -- ToDo: for integers we could do better here, perhaps by generalising
827 -- mk_switch and using that. --SDM 15/09/2004
828 emitLitSwitch scrut [] deflt
830 emitLitSwitch scrut branches deflt_blk
831 = do { scrut' <- assignNonPtrTemp scrut
832 ; deflt_blk_id <- forkCgStmts deflt_blk
833 ; blk <- mk_lit_switch scrut' deflt_blk_id (sortLe le branches)
836 le (t1,_) (t2,_) = t1 <= t2
838 mk_lit_switch :: CmmExpr -> BlockId
839 -> [(Literal,CgStmts)]
841 mk_lit_switch scrut deflt_blk_id [(lit,blk)]
842 = return (consCgStmt if_stmt blk)
844 cmm_lit = mkSimpleLit lit
845 rep = cmmLitRep cmm_lit
846 cond = CmmMachOp (MO_Ne rep) [scrut, CmmLit cmm_lit]
847 if_stmt = CmmCondBranch cond deflt_blk_id
849 mk_lit_switch scrut deflt_blk_id branches
850 = do { hi_blk <- mk_lit_switch scrut deflt_blk_id hi_branches
851 ; lo_blk <- mk_lit_switch scrut deflt_blk_id lo_branches
852 ; lo_blk_id <- forkCgStmts lo_blk
853 ; let if_stmt = CmmCondBranch cond lo_blk_id
854 ; return (if_stmt `consCgStmt` hi_blk) }
856 n_branches = length branches
857 (mid_lit,_) = branches !! (n_branches `div` 2)
858 -- See notes above re mid_tag
860 (lo_branches, hi_branches) = span is_lo branches
861 is_lo (t,_) = t < mid_lit
863 cond = CmmMachOp (mkLtOp mid_lit)
864 [scrut, CmmLit (mkSimpleLit mid_lit)]
866 -------------------------------------------------------------------------
868 -- Simultaneous assignment
870 -------------------------------------------------------------------------
873 emitSimultaneously :: CmmStmts -> Code
874 -- Emit code to perform the assignments in the
875 -- input simultaneously, using temporary variables when necessary.
877 -- The Stmts must be:
878 -- CmmNop, CmmComment, CmmAssign, CmmStore
882 -- We use the strongly-connected component algorithm, in which
883 -- * the vertices are the statements
884 -- * an edge goes from s1 to s2 iff
885 -- s1 assigns to something s2 uses
886 -- that is, if s1 should *follow* s2 in the final order
888 type CVertex = (Int, CmmStmt) -- Give each vertex a unique number,
889 -- for fast comparison
891 emitSimultaneously stmts
893 case filterOut isNopStmt (stmtList stmts) of
896 [stmt] -> stmtC stmt -- It's often just one stmt
897 stmt_list -> doSimultaneously1 (zip [(1::Int)..] stmt_list)
899 doSimultaneously1 :: [CVertex] -> Code
900 doSimultaneously1 vertices
902 edges = [ (vertex, key1, edges_from stmt1)
903 | vertex@(key1, stmt1) <- vertices
905 edges_from stmt1 = [ key2 | (key2, stmt2) <- vertices,
906 stmt1 `mustFollow` stmt2
908 components = stronglyConnCompFromEdgedVertices edges
910 -- do_components deal with one strongly-connected component
911 -- Not cyclic, or singleton? Just do it
912 do_component (AcyclicSCC (n,stmt)) = stmtC stmt
913 do_component (CyclicSCC [(n,stmt)]) = stmtC stmt
915 -- Cyclic? Then go via temporaries. Pick one to
916 -- break the loop and try again with the rest.
917 do_component (CyclicSCC ((n,first_stmt) : rest))
918 = do { from_temp <- go_via_temp first_stmt
919 ; doSimultaneously1 rest
922 go_via_temp (CmmAssign dest src)
923 = do { tmp <- newNonPtrTemp (cmmRegRep dest) -- TODO FIXME NOW if the pair of assignments move across a call this will be wrong
924 ; stmtC (CmmAssign (CmmLocal tmp) src)
925 ; return (CmmAssign dest (CmmReg (CmmLocal tmp))) }
926 go_via_temp (CmmStore dest src)
927 = do { tmp <- newNonPtrTemp (cmmExprRep src) -- TODO FIXME NOW if the pair of assignemnts move across a call this will be wrong
928 ; stmtC (CmmAssign (CmmLocal tmp) src)
929 ; return (CmmStore dest (CmmReg (CmmLocal tmp))) }
931 mapCs do_component components
933 mustFollow :: CmmStmt -> CmmStmt -> Bool
934 CmmAssign reg _ `mustFollow` stmt = anySrc (reg `regUsedIn`) stmt
935 CmmStore loc e `mustFollow` stmt = anySrc (locUsedIn loc (cmmExprRep e)) stmt
936 CmmNop `mustFollow` stmt = False
937 CmmComment _ `mustFollow` stmt = False
940 anySrc :: (CmmExpr -> Bool) -> CmmStmt -> Bool
941 -- True if the fn is true of any input of the stmt
942 anySrc p (CmmAssign _ e) = p e
943 anySrc p (CmmStore e1 e2) = p e1 || p e2 -- Might be used in either side
944 anySrc p (CmmComment _) = False
945 anySrc p CmmNop = False
946 anySrc p other = True -- Conservative
948 regUsedIn :: CmmReg -> CmmExpr -> Bool
949 reg `regUsedIn` CmmLit _ = False
950 reg `regUsedIn` CmmLoad e _ = reg `regUsedIn` e
951 reg `regUsedIn` CmmReg reg' = reg == reg'
952 reg `regUsedIn` CmmRegOff reg' _ = reg == reg'
953 reg `regUsedIn` CmmMachOp _ es = any (reg `regUsedIn`) es
955 locUsedIn :: CmmExpr -> MachRep -> CmmExpr -> Bool
956 -- (locUsedIn a r e) checks whether writing to r[a] could affect the value of
957 -- 'e'. Returns True if it's not sure.
958 locUsedIn loc rep (CmmLit _) = False
959 locUsedIn loc rep (CmmLoad e ld_rep) = possiblySameLoc loc rep e ld_rep
960 locUsedIn loc rep (CmmReg reg') = False
961 locUsedIn loc rep (CmmRegOff reg' _) = False
962 locUsedIn loc rep (CmmMachOp _ es) = any (locUsedIn loc rep) es
964 possiblySameLoc :: CmmExpr -> MachRep -> CmmExpr -> MachRep -> Bool
965 -- Assumes that distinct registers (eg Hp, Sp) do not
966 -- point to the same location, nor any offset thereof.
967 possiblySameLoc (CmmReg r1) rep1 (CmmReg r2) rep2 = r1==r2
968 possiblySameLoc (CmmReg r1) rep1 (CmmRegOff r2 0) rep2 = r1==r2
969 possiblySameLoc (CmmRegOff r1 0) rep1 (CmmReg r2) rep2 = r1==r2
970 possiblySameLoc (CmmRegOff r1 start1) rep1 (CmmRegOff r2 start2) rep2
971 = r1==r2 && end1 > start2 && end2 > start1
973 end1 = start1 + machRepByteWidth rep1
974 end2 = start2 + machRepByteWidth rep2
976 possiblySameLoc l1 rep1 (CmmLit _) rep2 = False
977 possiblySameLoc l1 rep1 l2 rep2 = True -- Conservative
979 -------------------------------------------------------------------------
981 -- Static Reference Tables
983 -------------------------------------------------------------------------
985 -- There is just one SRT for each top level binding; all the nested
986 -- bindings use sub-sections of this SRT. The label is passed down to
987 -- the nested bindings via the monad.
989 getSRTInfo :: FCode C_SRT
991 srt_lbl <- getSRTLabel
994 -- TODO: Should we panic in this case?
995 -- Someone obviously thinks there should be an SRT
996 NoSRT -> return NoC_SRT
997 SRTEntries {} -> panic "getSRTInfo: SRTEntries. Perhaps you forgot to run SimplStg?"
999 | len > hALF_WORD_SIZE_IN_BITS || bmp == [fromIntegral srt_escape]
1000 -> do id <- newUnique
1001 let srt_desc_lbl = mkLargeSRTLabel id
1002 emitRODataLits srt_desc_lbl
1003 ( cmmLabelOffW srt_lbl off
1004 : mkWordCLit (fromIntegral len)
1005 : map mkWordCLit bmp)
1006 return (C_SRT srt_desc_lbl 0 srt_escape)
1010 -> return (C_SRT srt_lbl off (fromIntegral (head bmp)))
1011 -- The fromIntegral converts to StgHalfWord
1013 srt_escape = (-1) :: StgHalfWord