CgUtils.hs 31.4 KB
Newer Older
1 2 3 4
-----------------------------------------------------------------------------
--
-- Code generator utilities; mostly monadic
--
Simon Marlow's avatar
Simon Marlow committed
5
-- (c) The University of Glasgow 2004-2006
6 7 8 9 10 11
--
-----------------------------------------------------------------------------

module CgUtils (
	addIdReps,
	cgLit,
12 13 14
	emitDataLits, mkDataLits,
        emitRODataLits, mkRODataLits,
        emitIf, emitIfThenElse,
15
	emitRtsCall, emitRtsCallWithVols, emitRtsCallWithResult,
16 17
	assignNonPtrTemp, newNonPtrTemp,
	assignPtrTemp, newPtrTemp,
18 19 20 21
	emitSimultaneously,
	emitSwitch, emitLitSwitch,
	tagToClosure,

22 23
        callerSaveVolatileRegs, get_GlobalReg_addr,

24 25 26 27 28 29 30 31 32 33
	cmmAndWord, cmmOrWord, cmmNegate, cmmEqWord, cmmNeWord,
	cmmOffsetExprW, cmmOffsetExprB,
	cmmRegOffW, cmmRegOffB,
	cmmLabelOffW, cmmLabelOffB,
	cmmOffsetW, cmmOffsetB,
	cmmOffsetLitW, cmmOffsetLitB,
	cmmLoadIndexW,

	addToMem, addToMemE,
	mkWordCLit,
34
	mkStringCLit, mkByteStringCLit,
35
	packHalfWordsCLit,
36 37 38
	blankWord,

	getSRTInfo
39 40 41
  ) where

#include "HsVersions.h"
42
#include "MachRegs.h"
43 44

import CgMonad
Simon Marlow's avatar
Simon Marlow committed
45 46 47 48
import TyCon
import Id
import Constants
import SMRep
49 50 51 52
import PprCmm		( {- instances -} )
import Cmm
import CLabel
import CmmUtils
Simon Marlow's avatar
Simon Marlow committed
53 54
import MachOp
import ForeignCall
55 56
import ClosureInfo
import StgSyn (SRT(..))
Simon Marlow's avatar
Simon Marlow committed
57 58 59 60 61 62 63
import Literal
import Digraph
import ListSetOps
import Util
import DynFlags
import FastString
import PackageConfig
64 65
import Outputable

Simon Marlow's avatar
Simon Marlow committed
66 67 68 69
import Data.Char
import Data.Bits
import Data.Word
import Data.Maybe
70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86

-------------------------------------------------------------------------
--
--	Random small functions
--
-------------------------------------------------------------------------

addIdReps :: [Id] -> [(CgRep, Id)]
addIdReps ids = [(idCgRep id, id) | id <- ids]

-------------------------------------------------------------------------
--
--	Literals
--
-------------------------------------------------------------------------

cgLit :: Literal -> FCode CmmLit
87 88
cgLit (MachStr s) = mkByteStringCLit (bytesFS s)
 -- not unpackFS; we want the UTF-8 byte stream.
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
cgLit other_lit   = return (mkSimpleLit other_lit)

mkSimpleLit :: Literal -> CmmLit
mkSimpleLit (MachChar	c)    = CmmInt (fromIntegral (ord c)) wordRep
mkSimpleLit MachNullAddr      = zeroCLit
mkSimpleLit (MachInt i)       = CmmInt i wordRep
mkSimpleLit (MachInt64 i)     = CmmInt i I64
mkSimpleLit (MachWord i)      = CmmInt i wordRep
mkSimpleLit (MachWord64 i)    = CmmInt i I64
mkSimpleLit (MachFloat r)     = CmmFloat r F32
mkSimpleLit (MachDouble r)    = CmmFloat r F64
mkSimpleLit (MachLabel fs ms) = CmmLabel (mkForeignLabel fs ms is_dyn)
			      where
				is_dyn = False	-- ToDo: fix me
	
mkLtOp :: Literal -> MachOp
-- On signed literals we must do a signed comparison
mkLtOp (MachInt _)    = MO_S_Lt wordRep
mkLtOp (MachFloat _)  = MO_S_Lt F32
mkLtOp (MachDouble _) = MO_S_Lt F64
mkLtOp lit	      = MO_U_Lt (cmmLitRep (mkSimpleLit lit))


---------------------------------------------------
--
--	Cmm data type functions
--
---------------------------------------------------

-----------------------
-- The "B" variants take byte offsets
cmmRegOffB :: CmmReg -> ByteOff -> CmmExpr
cmmRegOffB = cmmRegOff

cmmOffsetB :: CmmExpr -> ByteOff -> CmmExpr
cmmOffsetB = cmmOffset

cmmOffsetExprB :: CmmExpr -> CmmExpr -> CmmExpr
cmmOffsetExprB = cmmOffsetExpr

cmmLabelOffB :: CLabel -> ByteOff -> CmmLit
cmmLabelOffB = cmmLabelOff

cmmOffsetLitB :: CmmLit -> ByteOff -> CmmLit
cmmOffsetLitB = cmmOffsetLit

-----------------------
-- The "W" variants take word offsets
cmmOffsetExprW :: CmmExpr -> CmmExpr -> CmmExpr
-- The second arg is a *word* offset; need to change it to bytes
cmmOffsetExprW e (CmmLit (CmmInt n _)) = cmmOffsetW e (fromInteger n)
cmmOffsetExprW e wd_off = cmmIndexExpr wordRep e wd_off

cmmOffsetW :: CmmExpr -> WordOff -> CmmExpr
cmmOffsetW e n = cmmOffsetB e (wORD_SIZE * n)

cmmRegOffW :: CmmReg -> WordOff -> CmmExpr
cmmRegOffW reg wd_off = cmmRegOffB reg (wd_off * wORD_SIZE)

cmmOffsetLitW :: CmmLit -> WordOff -> CmmLit
cmmOffsetLitW lit wd_off = cmmOffsetLitB lit (wORD_SIZE * wd_off)

cmmLabelOffW :: CLabel -> WordOff -> CmmLit
cmmLabelOffW lbl wd_off = cmmLabelOffB lbl (wORD_SIZE * wd_off)

cmmLoadIndexW :: CmmExpr -> Int -> CmmExpr
cmmLoadIndexW base off
  = CmmLoad (cmmOffsetW base off) wordRep

-----------------------
cmmNeWord, cmmEqWord, cmmOrWord, cmmAndWord :: CmmExpr -> CmmExpr -> CmmExpr
cmmOrWord  e1 e2 = CmmMachOp mo_wordOr  [e1, e2]
cmmAndWord e1 e2 = CmmMachOp mo_wordAnd [e1, e2]
cmmNeWord  e1 e2 = CmmMachOp mo_wordNe  [e1, e2]
cmmEqWord  e1 e2 = CmmMachOp mo_wordEq  [e1, e2]
164
cmmULtWord e1 e2 = CmmMachOp mo_wordULt [e1, e2]
165
cmmUGeWord e1 e2 = CmmMachOp mo_wordUGe [e1, e2]
166
cmmUGtWord e1 e2 = CmmMachOp mo_wordUGt [e1, e2]
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221

cmmNegate :: CmmExpr -> CmmExpr
cmmNegate (CmmLit (CmmInt n rep)) = CmmLit (CmmInt (-n) rep)
cmmNegate e			  = CmmMachOp (MO_S_Neg (cmmExprRep e)) [e]

blankWord :: CmmStatic
blankWord = CmmUninitialised wORD_SIZE

-----------------------
--	Making literals

mkWordCLit :: StgWord -> CmmLit
mkWordCLit wd = CmmInt (fromIntegral wd) wordRep

packHalfWordsCLit :: (Integral a, Integral b) => a -> b -> CmmLit
-- Make a single word literal in which the lower_half_word is
-- at the lower address, and the upper_half_word is at the 
-- higher address
-- ToDo: consider using half-word lits instead
-- 	 but be careful: that's vulnerable when reversed
packHalfWordsCLit lower_half_word upper_half_word
#ifdef WORDS_BIGENDIAN
   = mkWordCLit ((fromIntegral lower_half_word `shiftL` hALF_WORD_SIZE_IN_BITS)
		 .|. fromIntegral upper_half_word)
#else 
   = mkWordCLit ((fromIntegral lower_half_word) 
		 .|. (fromIntegral upper_half_word `shiftL` hALF_WORD_SIZE_IN_BITS))
#endif

--------------------------------------------------------------------------
--
-- Incrementing a memory location
--
--------------------------------------------------------------------------

addToMem :: MachRep 	-- rep of the counter
	 -> CmmExpr	-- Address
	 -> Int		-- What to add (a word)
	 -> CmmStmt
addToMem rep ptr n = addToMemE rep ptr (CmmLit (CmmInt (toInteger n) rep))

addToMemE :: MachRep 	-- rep of the counter
	  -> CmmExpr	-- Address
	  -> CmmExpr	-- What to add (a word-typed expression)
	  -> CmmStmt
addToMemE rep ptr n
  = CmmStore ptr (CmmMachOp (MO_Add rep) [CmmLoad ptr rep, n])

-------------------------------------------------------------------------
--
--	Converting a closure tag to a closure for enumeration types
--      (this is the implementation of tagToEnum#).
--
-------------------------------------------------------------------------

Simon Marlow's avatar
Simon Marlow committed
222 223
tagToClosure :: PackageId -> TyCon -> CmmExpr -> CmmExpr
tagToClosure this_pkg tycon tag
224
  = CmmLoad (cmmOffsetExprW closure_tbl tag) wordRep
225
  where closure_tbl = CmmLit (CmmLabel lbl)
Simon Marlow's avatar
Simon Marlow committed
226
	lbl = mkClosureTableLabel this_pkg (tyConName tycon)
227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268

-------------------------------------------------------------------------
--
--	Conditionals and rts calls
--
-------------------------------------------------------------------------

emitIf :: CmmExpr 	-- Boolean
       -> Code		-- Then part
       -> Code		
-- Emit (if e then x)
-- ToDo: reverse the condition to avoid the extra branch instruction if possible
-- (some conditionals aren't reversible. eg. floating point comparisons cannot
-- be inverted because there exist some values for which both comparisons
-- return False, such as NaN.)
emitIf cond then_part
  = do { then_id <- newLabelC
       ; join_id <- newLabelC
       ; stmtC (CmmCondBranch cond then_id)
       ; stmtC (CmmBranch join_id)
       ; labelC then_id
       ; then_part
       ; labelC join_id
       }

emitIfThenElse :: CmmExpr 	-- Boolean
       		-> Code		-- Then part
       		-> Code		-- Else part
       		-> Code		
-- Emit (if e then x else y)
emitIfThenElse cond then_part else_part
  = do { then_id <- newLabelC
       ; else_id <- newLabelC
       ; join_id <- newLabelC
       ; stmtC (CmmCondBranch cond then_id)
       ; else_part
       ; stmtC (CmmBranch join_id)
       ; labelC then_id
       ; then_part
       ; labelC join_id
       }

269 270
emitRtsCall :: LitString -> [(CmmExpr,MachHint)] -> Bool -> Code
emitRtsCall fun args safe = emitRtsCall' [] fun args Nothing safe
271 272
   -- The 'Nothing' says "save all global registers"

273 274 275
emitRtsCallWithVols :: LitString -> [(CmmExpr,MachHint)] -> [GlobalReg] -> Bool -> Code
emitRtsCallWithVols fun args vols safe
   = emitRtsCall' [] fun args (Just vols) safe
276

277
emitRtsCallWithResult :: LocalReg -> MachHint -> LitString
278 279 280
	-> [(CmmExpr,MachHint)] -> Bool -> Code
emitRtsCallWithResult res hint fun args safe
   = emitRtsCall' [(res,hint)] fun args Nothing safe
281 282 283

-- Make a call to an RTS C procedure
emitRtsCall'
284
   :: CmmHintFormals
285 286 287
   -> LitString
   -> [(CmmExpr,MachHint)]
   -> Maybe [GlobalReg]
288
   -> Bool -- True <=> CmmSafe call
289
   -> Code
290 291 292 293 294 295 296
emitRtsCall' res fun args vols safe = do
  safety <- if safe
            then getSRTInfo >>= (return . CmmSafe)
            else return CmmUnsafe
  stmtsC caller_save
  stmtC (CmmCall target res args safety)
  stmtsC caller_load
297
  where
298
    (caller_save, caller_load) = callerSaveVolatileRegs vols
299 300 301
    target   = CmmForeignCall fun_expr CCallConv
    fun_expr = mkLblExpr (mkRtsCodeLabel fun)

302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483
-----------------------------------------------------------------------------
--
--	Caller-Save Registers
--
-----------------------------------------------------------------------------

-- Here we generate the sequence of saves/restores required around a
-- foreign call instruction.

-- TODO: reconcile with includes/Regs.h
--  * Regs.h claims that BaseReg should be saved last and loaded first
--    * This might not have been tickled before since BaseReg is callee save
--  * Regs.h saves SparkHd, ParkT1, SparkBase and SparkLim
callerSaveVolatileRegs :: Maybe [GlobalReg] -> ([CmmStmt], [CmmStmt])
callerSaveVolatileRegs vols = (caller_save, caller_load)
  where
    caller_save = foldr ($!) [] (map callerSaveGlobalReg    regs_to_save)
    caller_load = foldr ($!) [] (map callerRestoreGlobalReg regs_to_save)

    system_regs = [Sp,SpLim,Hp,HpLim,CurrentTSO,CurrentNursery,
		   {-SparkHd,SparkTl,SparkBase,SparkLim,-}BaseReg ]

    regs_to_save = system_regs ++ vol_list

    vol_list = case vols of Nothing -> all_of_em; Just regs -> regs

    all_of_em = [ VanillaReg n | n <- [0..mAX_Vanilla_REG] ]
	     ++ [ FloatReg   n | n <- [0..mAX_Float_REG] ]
	     ++ [ DoubleReg  n | n <- [0..mAX_Double_REG] ]
	     ++ [ LongReg    n | n <- [0..mAX_Long_REG] ]

    callerSaveGlobalReg reg next
	| callerSaves reg = 
		CmmStore (get_GlobalReg_addr reg) 
			 (CmmReg (CmmGlobal reg)) : next
	| otherwise = next

    callerRestoreGlobalReg reg next
	| callerSaves reg = 
		CmmAssign (CmmGlobal reg)
			  (CmmLoad (get_GlobalReg_addr reg) (globalRegRep reg))
			: next
	| otherwise = next

-- -----------------------------------------------------------------------------
-- Global registers

-- We map STG registers onto appropriate CmmExprs.  Either they map
-- to real machine registers or stored as offsets from BaseReg.  Given
-- a GlobalReg, get_GlobalReg_addr always produces the 
-- register table address for it.
-- (See also get_GlobalReg_reg_or_addr in MachRegs)

get_GlobalReg_addr              :: GlobalReg -> CmmExpr
get_GlobalReg_addr BaseReg = regTableOffset 0
get_GlobalReg_addr mid     = get_Regtable_addr_from_offset 
				(globalRegRep mid) (baseRegOffset mid)

-- Calculate a literal representing an offset into the register table.
-- Used when we don't have an actual BaseReg to offset from.
regTableOffset n = 
  CmmLit (CmmLabelOff mkMainCapabilityLabel (oFFSET_Capability_r + n))

get_Regtable_addr_from_offset   :: MachRep -> Int -> CmmExpr
get_Regtable_addr_from_offset rep offset =
#ifdef REG_Base
  CmmRegOff (CmmGlobal BaseReg) offset
#else
  regTableOffset offset
#endif


-- | Returns 'True' if this global register is stored in a caller-saves
-- machine register.

callerSaves :: GlobalReg -> Bool

#ifdef CALLER_SAVES_Base
callerSaves BaseReg		= True
#endif
#ifdef CALLER_SAVES_R1
callerSaves (VanillaReg 1)	= True
#endif
#ifdef CALLER_SAVES_R2
callerSaves (VanillaReg 2)	= True
#endif
#ifdef CALLER_SAVES_R3
callerSaves (VanillaReg 3)	= True
#endif
#ifdef CALLER_SAVES_R4
callerSaves (VanillaReg 4)	= True
#endif
#ifdef CALLER_SAVES_R5
callerSaves (VanillaReg 5)	= True
#endif
#ifdef CALLER_SAVES_R6
callerSaves (VanillaReg 6)	= True
#endif
#ifdef CALLER_SAVES_R7
callerSaves (VanillaReg 7)	= True
#endif
#ifdef CALLER_SAVES_R8
callerSaves (VanillaReg 8)	= True
#endif
#ifdef CALLER_SAVES_F1
callerSaves (FloatReg 1)	= True
#endif
#ifdef CALLER_SAVES_F2
callerSaves (FloatReg 2)	= True
#endif
#ifdef CALLER_SAVES_F3
callerSaves (FloatReg 3)	= True
#endif
#ifdef CALLER_SAVES_F4
callerSaves (FloatReg 4)	= True
#endif
#ifdef CALLER_SAVES_D1
callerSaves (DoubleReg 1)	= True
#endif
#ifdef CALLER_SAVES_D2
callerSaves (DoubleReg 2)	= True
#endif
#ifdef CALLER_SAVES_L1
callerSaves (LongReg 1)		= True
#endif
#ifdef CALLER_SAVES_Sp
callerSaves Sp			= True
#endif
#ifdef CALLER_SAVES_SpLim
callerSaves SpLim		= True
#endif
#ifdef CALLER_SAVES_Hp
callerSaves Hp			= True
#endif
#ifdef CALLER_SAVES_HpLim
callerSaves HpLim		= True
#endif
#ifdef CALLER_SAVES_CurrentTSO
callerSaves CurrentTSO		= True
#endif
#ifdef CALLER_SAVES_CurrentNursery
callerSaves CurrentNursery	= True
#endif
callerSaves _			= False


-- -----------------------------------------------------------------------------
-- Information about global registers

baseRegOffset :: GlobalReg -> Int

baseRegOffset (VanillaReg 1)      = oFFSET_StgRegTable_rR1
baseRegOffset (VanillaReg 2)      = oFFSET_StgRegTable_rR2
baseRegOffset (VanillaReg 3)      = oFFSET_StgRegTable_rR3
baseRegOffset (VanillaReg 4)      = oFFSET_StgRegTable_rR4
baseRegOffset (VanillaReg 5)      = oFFSET_StgRegTable_rR5
baseRegOffset (VanillaReg 6)      = oFFSET_StgRegTable_rR6
baseRegOffset (VanillaReg 7)      = oFFSET_StgRegTable_rR7
baseRegOffset (VanillaReg 8)      = oFFSET_StgRegTable_rR8
baseRegOffset (VanillaReg 9)      = oFFSET_StgRegTable_rR9
baseRegOffset (VanillaReg 10)     = oFFSET_StgRegTable_rR10
baseRegOffset (FloatReg  1)       = oFFSET_StgRegTable_rF1
baseRegOffset (FloatReg  2)       = oFFSET_StgRegTable_rF2
baseRegOffset (FloatReg  3)       = oFFSET_StgRegTable_rF3
baseRegOffset (FloatReg  4)       = oFFSET_StgRegTable_rF4
baseRegOffset (DoubleReg 1)       = oFFSET_StgRegTable_rD1
baseRegOffset (DoubleReg 2)       = oFFSET_StgRegTable_rD2
baseRegOffset Sp		  = oFFSET_StgRegTable_rSp
baseRegOffset SpLim		  = oFFSET_StgRegTable_rSpLim
baseRegOffset (LongReg 1)         = oFFSET_StgRegTable_rL1
baseRegOffset Hp		  = oFFSET_StgRegTable_rHp
baseRegOffset HpLim		  = oFFSET_StgRegTable_rHpLim
baseRegOffset CurrentTSO	  = oFFSET_StgRegTable_rCurrentTSO
baseRegOffset CurrentNursery	  = oFFSET_StgRegTable_rCurrentNursery
baseRegOffset HpAlloc		  = oFFSET_StgRegTable_rHpAlloc
baseRegOffset GCEnter1		  = oFFSET_stgGCEnter1
baseRegOffset GCFun		  = oFFSET_stgGCFun
#ifdef DEBUG
baseRegOffset BaseReg		  = panic "baseRegOffset:BaseReg"
baseRegOffset _			  = panic "baseRegOffset:other"
#endif

484 485 486 487 488 489 490 491 492 493 494 495

-------------------------------------------------------------------------
--
--	Strings gnerate a top-level data block
--
-------------------------------------------------------------------------

emitDataLits :: CLabel -> [CmmLit] -> Code
-- Emit a data-segment data block
emitDataLits lbl lits
  = emitData Data (CmmDataLabel lbl : map CmmStaticLit lits)

496 497 498 499 500
mkDataLits :: CLabel -> [CmmLit] -> GenCmmTop CmmStatic info stmt
-- Emit a data-segment data block
mkDataLits lbl lits
  = CmmData Data (CmmDataLabel lbl : map CmmStaticLit lits)

501 502 503
emitRODataLits :: CLabel -> [CmmLit] -> Code
-- Emit a read-only data block
emitRODataLits lbl lits
504 505 506 507 508 509
  = emitData section (CmmDataLabel lbl : map CmmStaticLit lits)
  where section | any needsRelocation lits = RelocatableReadOnlyData
                | otherwise                = ReadOnlyData
        needsRelocation (CmmLabel _)      = True
        needsRelocation (CmmLabelOff _ _) = True
        needsRelocation _                 = False
510

511 512 513 514 515 516 517 518 519
mkRODataLits :: CLabel -> [CmmLit] -> GenCmmTop CmmStatic info stmt
mkRODataLits lbl lits
  = CmmData section (CmmDataLabel lbl : map CmmStaticLit lits)
  where section | any needsRelocation lits = RelocatableReadOnlyData
                | otherwise                = ReadOnlyData
        needsRelocation (CmmLabel _)      = True
        needsRelocation (CmmLabelOff _ _) = True
        needsRelocation _                 = False

520 521 522
mkStringCLit :: String -> FCode CmmLit
-- Make a global definition for the string,
-- and return its label
523 524 525 526
mkStringCLit str = mkByteStringCLit (map (fromIntegral.ord) str)

mkByteStringCLit :: [Word8] -> FCode CmmLit
mkByteStringCLit bytes
527
  = do 	{ uniq <- newUnique
528
	; let lbl = mkStringLitLabel uniq
529
	; emitData ReadOnlyData [CmmDataLabel lbl, CmmString bytes]
530 531 532 533 534 535 536 537
	; return (CmmLabel lbl) }

-------------------------------------------------------------------------
--
--	Assigning expressions to temporaries
--
-------------------------------------------------------------------------

538
assignNonPtrTemp :: CmmExpr -> FCode CmmExpr
539 540
-- For a non-trivial expression, e, create a local
-- variable and assign the expression to it
541
assignNonPtrTemp e 
542
  | isTrivialCmmExpr e = return e
543 544 545
  | otherwise 	       = do { reg <- newNonPtrTemp (cmmExprRep e) 
			    ; stmtC (CmmAssign (CmmLocal reg) e)
			    ; return (CmmReg (CmmLocal reg)) }
546

547 548 549 550 551 552 553 554
assignPtrTemp :: CmmExpr -> FCode CmmExpr
-- For a non-trivial expression, e, create a local
-- variable and assign the expression to it
assignPtrTemp e 
  | isTrivialCmmExpr e = return e
  | otherwise 	       = do { reg <- newPtrTemp (cmmExprRep e) 
			    ; stmtC (CmmAssign (CmmLocal reg) e)
			    ; return (CmmReg (CmmLocal reg)) }
555

556 557 558 559 560
newNonPtrTemp :: MachRep -> FCode LocalReg
newNonPtrTemp rep = do { uniq <- newUnique; return (LocalReg uniq rep KindNonPtr) }

newPtrTemp :: MachRep -> FCode LocalReg
newPtrTemp rep = do { uniq <- newUnique; return (LocalReg uniq rep KindPtr) }
561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588


-------------------------------------------------------------------------
--
--	Building case analysis
--
-------------------------------------------------------------------------

emitSwitch
	:: CmmExpr  		  -- Tag to switch on
	-> [(ConTagZ, CgStmts)]	  -- Tagged branches
	-> Maybe CgStmts	  -- Default branch (if any)
	-> ConTagZ -> ConTagZ	  -- Min and Max possible values; behaviour
				  -- 	outside this range is undefined
	-> Code

-- ONLY A DEFAULT BRANCH: no case analysis to do
emitSwitch tag_expr [] (Just stmts) _ _
  = emitCgStmts stmts

-- Right, off we go
emitSwitch tag_expr branches mb_deflt lo_tag hi_tag
  = 	-- Just sort the branches before calling mk_sritch
    do	{ mb_deflt_id <-
		case mb_deflt of
		  Nothing    -> return Nothing
		  Just stmts -> do id <- forkCgStmts stmts; return (Just id)

589 590 591 592
	; dflags <- getDynFlags
	; let via_C | HscC <- hscTarget dflags = True
		    | otherwise                = False

593
	; stmts <- mk_switch tag_expr (sortLe le branches) 
594
			mb_deflt_id lo_tag hi_tag via_C
595 596 597
	; emitCgStmts stmts
	}
  where
598
    (t1,_) `le` (t2,_) = t1 <= t2
599 600 601


mk_switch :: CmmExpr -> [(ConTagZ, CgStmts)]
602
	  -> Maybe BlockId -> ConTagZ -> ConTagZ -> Bool
603 604 605
	  -> FCode CgStmts

-- SINGLETON TAG RANGE: no case analysis to do
606
mk_switch tag_expr [(tag,stmts)] _ lo_tag hi_tag via_C
607 608 609 610 611
  | lo_tag == hi_tag
  = ASSERT( tag == lo_tag )
    return stmts

-- SINGLETON BRANCH, NO DEFUALT: no case analysis to do
612
mk_switch tag_expr [(tag,stmts)] Nothing lo_tag hi_tag via_C
613 614 615 616 617 618 619 620
  = return stmts
	-- The simplifier might have eliminated a case
	-- 	 so we may have e.g. case xs of 
	--				 [] -> e
	-- In that situation we can be sure the (:) case 
	-- can't happen, so no need to test

-- SINGLETON BRANCH: one equality check to do
621
mk_switch tag_expr [(tag,stmts)] (Just deflt) lo_tag hi_tag via_C
622 623 624 625 626 627 628 629 630 631
  = return (CmmCondBranch cond deflt `consCgStmt` stmts)
  where
    cond  =  cmmNeWord tag_expr (CmmLit (mkIntCLit tag))
	-- We have lo_tag < hi_tag, but there's only one branch, 
	-- so there must be a default

-- ToDo: we might want to check for the two branch case, where one of
-- the branches is the tag 0, because comparing '== 0' is likely to be
-- more efficient than other kinds of comparison.

632 633 634 635 636 637 638 639 640 641
-- DENSE TAG RANGE: use a switch statment.
--
-- We also use a switch uncoditionally when compiling via C, because
-- this will get emitted as a C switch statement and the C compiler
-- should do a good job of optimising it.  Also, older GCC versions
-- (2.95 in particular) have problems compiling the complicated
-- if-trees generated by this code, so compiling to a switch every
-- time works around that problem.
--
mk_switch tag_expr branches mb_deflt lo_tag hi_tag via_C
642
  | use_switch 	-- Use a switch
643
  = do	{ branch_ids <- mapM forkCgStmts (map snd branches)
644
	; let 
645
		tagged_blk_ids = zip (map fst branches) (map Just branch_ids)
646

647 648
		find_branch :: ConTagZ -> Maybe BlockId
		find_branch i = assocDefault mb_deflt tagged_blk_ids i
649

650 651 652 653
		-- NB. we have eliminated impossible branches at
		-- either end of the range (see below), so the first
		-- tag of a real branch is real_lo_tag (not lo_tag).
		arms = [ find_branch i | i <- [real_lo_tag..real_hi_tag]]
654

655
	        switch_stmt = CmmSwitch (cmmOffset tag_expr (- real_lo_tag)) arms
656

657 658
	; ASSERT(not (all isNothing arms)) 
	  return (oneCgStmt switch_stmt)
659 660
	}

661 662
  -- if we can knock off a bunch of default cases with one if, then do so
  | Just deflt <- mb_deflt, (lowest_branch - lo_tag) >= n_branches
663
  = do { (assign_tag, tag_expr') <- assignNonPtrTemp' tag_expr
664
       ; let cond = cmmULtWord tag_expr' (CmmLit (mkIntCLit lowest_branch))
665
	     branch = CmmCondBranch cond deflt
666 667
       ; stmts <- mk_switch tag_expr' branches mb_deflt 
			lowest_branch hi_tag via_C
668
       ; return (assign_tag `consCgStmt` (branch `consCgStmt` stmts))
669 670 671
       }

  | Just deflt <- mb_deflt, (hi_tag - highest_branch) >= n_branches
672
  = do { (assign_tag, tag_expr') <- assignNonPtrTemp' tag_expr
673
       ; let cond = cmmUGtWord tag_expr' (CmmLit (mkIntCLit highest_branch))
674
	     branch = CmmCondBranch cond deflt
675 676
       ; stmts <- mk_switch tag_expr' branches mb_deflt 
			lo_tag highest_branch via_C
677
       ; return (assign_tag `consCgStmt` (branch `consCgStmt` stmts))
678 679
       }

680
  | otherwise	-- Use an if-tree
681
  = do	{ (assign_tag, tag_expr') <- assignNonPtrTemp' tag_expr
682
		-- To avoid duplication
683 684 685 686
	; lo_stmts <- mk_switch tag_expr' lo_branches mb_deflt 
				lo_tag (mid_tag-1) via_C
	; hi_stmts <- mk_switch tag_expr' hi_branches mb_deflt 
				mid_tag hi_tag via_C
687 688 689 690
	; hi_id <- forkCgStmts hi_stmts
	; let cond = cmmUGeWord tag_expr' (CmmLit (mkIntCLit mid_tag))
	      branch_stmt = CmmCondBranch cond hi_id
	; return (assign_tag `consCgStmt` (branch_stmt `consCgStmt` lo_stmts)) 
691
	}
692 693 694 695 696 697 698
	-- we test (e >= mid_tag) rather than (e < mid_tag), because
	-- the former works better when e is a comparison, and there
	-- are two tags 0 & 1 (mid_tag == 1).  In this case, the code
	-- generator can reduce the condition to e itself without
	-- having to reverse the sense of the comparison: comparisons
	-- can't always be easily reversed (eg. floating
	-- pt. comparisons).
699
  where
700 701
    use_switch 	 = {- pprTrace "mk_switch" (
			ppr tag_expr <+> text "n_tags:" <+> int n_tags <+>
702
                        text "branches:" <+> ppr (map fst branches) <+>
703
			text "n_branches:" <+> int n_branches <+>
704 705 706 707
			text "lo_tag:" <+> int lo_tag <+>
			text "hi_tag:" <+> int hi_tag <+>
			text "real_lo_tag:" <+> int real_lo_tag <+>
			text "real_hi_tag:" <+> int real_hi_tag) $ -}
708
		   ASSERT( n_branches > 1 && n_tags > 1 ) 
709 710 711 712 713
		   n_tags > 2 && (via_C || (dense && big_enough))
		 -- up to 4 branches we use a decision tree, otherwise
                 -- a switch (== jump table in the NCG).  This seems to be
                 -- optimal, and corresponds with what gcc does.
    big_enough 	 = n_branches > 4
714 715 716
    dense      	 = n_branches > (n_tags `div` 2)
    n_branches   = length branches
    
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731
    -- ignore default slots at each end of the range if there's 
    -- no default branch defined.
    lowest_branch  = fst (head branches)
    highest_branch = fst (last branches)

    real_lo_tag
	| isNothing mb_deflt = lowest_branch
	| otherwise          = lo_tag

    real_hi_tag
	| isNothing mb_deflt = highest_branch
	| otherwise          = hi_tag

    n_tags = real_hi_tag - real_lo_tag + 1

732 733 734 735 736 737 738 739 740 741 742 743 744 745
	-- INVARIANT: Provided hi_tag > lo_tag (which is true)
	--	lo_tag <= mid_tag < hi_tag
	--	lo_branches have tags <  mid_tag
	--	hi_branches have tags >= mid_tag

    (mid_tag,_) = branches !! (n_branches `div` 2)
	-- 2 branches => n_branches `div` 2 = 1
	--	      => branches !! 1 give the *second* tag
	-- There are always at least 2 branches here

    (lo_branches, hi_branches) = span is_lo branches
    is_lo (t,_) = t < mid_tag


746
assignNonPtrTemp' e
747
  | isTrivialCmmExpr e = return (CmmNop, e)
748 749
  | otherwise          = do { reg <- newNonPtrTemp (cmmExprRep e)
                            ; return (CmmAssign (CmmLocal reg) e, CmmReg (CmmLocal reg)) }
750 751 752 753 754 755 756 757

emitLitSwitch :: CmmExpr			-- Tag to switch on
	      -> [(Literal, CgStmts)]		-- Tagged branches
	      -> CgStmts			-- Default branch (always)
	      -> Code				-- Emit the code
-- Used for general literals, whose size might not be a word, 
-- where there is always a default case, and where we don't know
-- the range of values for certain.  For simplicity we always generate a tree.
758 759 760
--
-- ToDo: for integers we could do better here, perhaps by generalising
-- mk_switch and using that.  --SDM 15/09/2004
761 762 763
emitLitSwitch scrut [] deflt 
  = emitCgStmts deflt
emitLitSwitch scrut branches deflt_blk
764
  = do	{ scrut' <- assignNonPtrTemp scrut
765
	; deflt_blk_id <- forkCgStmts deflt_blk
766
	; blk <- mk_lit_switch scrut' deflt_blk_id (sortLe le branches)
767 768
	; emitCgStmts blk }
  where
769
    le (t1,_) (t2,_) = t1 <= t2
770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855

mk_lit_switch :: CmmExpr -> BlockId 
 	      -> [(Literal,CgStmts)]
	      -> FCode CgStmts
mk_lit_switch scrut deflt_blk_id [(lit,blk)] 
  = return (consCgStmt if_stmt blk)
  where
    cmm_lit = mkSimpleLit lit
    rep     = cmmLitRep cmm_lit
    cond    = CmmMachOp (MO_Ne rep) [scrut, CmmLit cmm_lit]
    if_stmt = CmmCondBranch cond deflt_blk_id

mk_lit_switch scrut deflt_blk_id branches
  = do	{ hi_blk <- mk_lit_switch scrut deflt_blk_id hi_branches
 	; lo_blk <- mk_lit_switch scrut deflt_blk_id lo_branches
	; lo_blk_id <- forkCgStmts lo_blk
	; let if_stmt = CmmCondBranch cond lo_blk_id
	; return (if_stmt `consCgStmt` hi_blk) }
  where
    n_branches = length branches
    (mid_lit,_) = branches !! (n_branches `div` 2)
	-- See notes above re mid_tag

    (lo_branches, hi_branches) = span is_lo branches
    is_lo (t,_) = t < mid_lit

    cond    = CmmMachOp (mkLtOp mid_lit) 
			[scrut, CmmLit (mkSimpleLit mid_lit)]

-------------------------------------------------------------------------
--
--	Simultaneous assignment
--
-------------------------------------------------------------------------


emitSimultaneously :: CmmStmts -> Code
-- Emit code to perform the assignments in the
-- input simultaneously, using temporary variables when necessary.
--
-- The Stmts must be:
--	CmmNop, CmmComment, CmmAssign, CmmStore
-- and nothing else


-- We use the strongly-connected component algorithm, in which
--	* the vertices are the statements
--	* an edge goes from s1 to s2 iff
--		s1 assigns to something s2 uses
--	  that is, if s1 should *follow* s2 in the final order

type CVertex = (Int, CmmStmt)	-- Give each vertex a unique number,
				-- for fast comparison

emitSimultaneously stmts
  = codeOnly $
    case filterOut isNopStmt (stmtList stmts) of 
	-- Remove no-ops
      []     	-> nopC
      [stmt] 	-> stmtC stmt	-- It's often just one stmt
      stmt_list -> doSimultaneously1 (zip [(1::Int)..] stmt_list)

doSimultaneously1 :: [CVertex] -> Code
doSimultaneously1 vertices
  = let
	edges = [ (vertex, key1, edges_from stmt1)
		| vertex@(key1, stmt1) <- vertices
		]
	edges_from stmt1 = [ key2 | (key2, stmt2) <- vertices, 
				    stmt1 `mustFollow` stmt2
			   ]
	components = stronglyConnComp edges

	-- do_components deal with one strongly-connected component
	-- Not cyclic, or singleton?  Just do it
	do_component (AcyclicSCC (n,stmt))  = stmtC stmt
	do_component (CyclicSCC [(n,stmt)]) = stmtC stmt

		-- Cyclic?  Then go via temporaries.  Pick one to
		-- break the loop and try again with the rest.
	do_component (CyclicSCC ((n,first_stmt) : rest))
	  = do	{ from_temp <- go_via_temp first_stmt
		; doSimultaneously1 rest
		; stmtC from_temp }

	go_via_temp (CmmAssign dest src)
856 857 858
	  = do	{ tmp <- newNonPtrTemp (cmmRegRep dest) -- TODO FIXME NOW if the pair of assignments move across a call this will be wrong
		; stmtC (CmmAssign (CmmLocal tmp) src)
		; return (CmmAssign dest (CmmReg (CmmLocal tmp))) }
859
	go_via_temp (CmmStore dest src)
860 861 862
	  = do	{ tmp <- newNonPtrTemp (cmmExprRep src) -- TODO FIXME NOW if the pair of assignemnts move across a call this will be wrong
		; stmtC (CmmAssign (CmmLocal tmp) src)
		; return (CmmStore dest (CmmReg (CmmLocal tmp))) }
863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910
    in
    mapCs do_component components

mustFollow :: CmmStmt -> CmmStmt -> Bool
CmmAssign reg _  `mustFollow` stmt = anySrc (reg `regUsedIn`) stmt
CmmStore loc e   `mustFollow` stmt = anySrc (locUsedIn loc (cmmExprRep e)) stmt
CmmNop           `mustFollow` stmt = False
CmmComment _     `mustFollow` stmt = False


anySrc :: (CmmExpr -> Bool) -> CmmStmt -> Bool
-- True if the fn is true of any input of the stmt
anySrc p (CmmAssign _ e)    = p e
anySrc p (CmmStore e1 e2)   = p e1 || p e2	-- Might be used in either side
anySrc p (CmmComment _)	    = False
anySrc p CmmNop		    = False
anySrc p other		    = True		-- Conservative

regUsedIn :: CmmReg -> CmmExpr -> Bool
reg `regUsedIn` CmmLit _ 	 = False
reg `regUsedIn` CmmLoad e  _ 	 = reg `regUsedIn` e
reg `regUsedIn` CmmReg reg' 	 = reg == reg'
reg `regUsedIn` CmmRegOff reg' _ = reg == reg'
reg `regUsedIn` CmmMachOp _ es   = any (reg `regUsedIn`) es

locUsedIn :: CmmExpr -> MachRep -> CmmExpr -> Bool
-- (locUsedIn a r e) checks whether writing to r[a] could affect the value of
-- 'e'.  Returns True if it's not sure.
locUsedIn loc rep (CmmLit _) 	     = False
locUsedIn loc rep (CmmLoad e ld_rep) = possiblySameLoc loc rep e ld_rep
locUsedIn loc rep (CmmReg reg')      = False
locUsedIn loc rep (CmmRegOff reg' _) = False
locUsedIn loc rep (CmmMachOp _ es)   = any (locUsedIn loc rep) es

possiblySameLoc :: CmmExpr -> MachRep -> CmmExpr -> MachRep -> Bool
-- Assumes that distinct registers (eg Hp, Sp) do not 
-- point to the same location, nor any offset thereof.
possiblySameLoc (CmmReg r1)       rep1 (CmmReg r2)      rep2  = r1==r2
possiblySameLoc (CmmReg r1)       rep1 (CmmRegOff r2 0) rep2  = r1==r2
possiblySameLoc (CmmRegOff r1 0)  rep1 (CmmReg r2)      rep2  = r1==r2
possiblySameLoc (CmmRegOff r1 start1) rep1 (CmmRegOff r2 start2) rep2 
  = r1==r2 && end1 > start2 && end2 > start1
  where
    end1 = start1 + machRepByteWidth rep1
    end2 = start2 + machRepByteWidth rep2

possiblySameLoc l1 rep1 (CmmLit _) rep2 = False
possiblySameLoc l1 rep1 l2	   rep2 = True	-- Conservative
911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945

-------------------------------------------------------------------------
--
--	Static Reference Tables
--
-------------------------------------------------------------------------

-- There is just one SRT for each top level binding; all the nested
-- bindings use sub-sections of this SRT.  The label is passed down to
-- the nested bindings via the monad.

getSRTInfo :: FCode C_SRT
getSRTInfo = do
  srt_lbl <- getSRTLabel
  srt <- getSRT
  case srt of
    -- TODO: Should we panic in this case?
    -- Someone obviously thinks there should be an SRT
    NoSRT -> return NoC_SRT
    SRT off len bmp
      | len > hALF_WORD_SIZE_IN_BITS || bmp == [fromIntegral srt_escape]
      -> do id <- newUnique
            let srt_desc_lbl = mkLargeSRTLabel id
	    emitRODataLits srt_desc_lbl
             ( cmmLabelOffW srt_lbl off
	       : mkWordCLit (fromIntegral len)
	       : map mkWordCLit bmp)
	    return (C_SRT srt_desc_lbl 0 srt_escape)

    SRT off len bmp
      | otherwise 
      -> return (C_SRT srt_lbl off (fromIntegral (head bmp)))
		-- The fromIntegral converts to StgHalfWord

srt_escape = (-1) :: StgHalfWord