AsmCodeGen.lhs 31.9 KB
Newer Older
1 2 3 4 5 6 7
-- -----------------------------------------------------------------------------
--
-- (c) The University of Glasgow 1993-2004
-- 
-- This is the top-level module in the native code generator.
--
-- -----------------------------------------------------------------------------
8 9

\begin{code}
10
module AsmCodeGen ( nativeCodeGen ) where
11

12
#include "HsVersions.h"
rrt's avatar
rrt committed
13
#include "NCG.h"
14

15
import MachInstrs
16
import MachRegs
17
import MachCodeGen
18
import PprMach
19 20 21
import RegisterAlloc
import RegAllocInfo	( jumpDests )
import NCGMonad
22
import PositionIndependentCode
23 24 25 26

import Cmm
import PprCmm		( pprStmt, pprCmms )
import MachOp
27
import CLabel           ( CLabel, mkSplitMarkerLabel, mkAsmTempLabel )
28 29 30
#if powerpc_TARGET_ARCH
import CLabel           ( mkRtsCodeLabel )
#endif
31

32 33 34 35
import UniqFM
import Unique		( Unique, getUnique )
import UniqSupply
import FastTypes
36 37
import List		( groupBy, sortBy )
import CLabel           ( pprCLabel )
38
import ErrUtils		( dumpIfSet_dyn )
39
import DynFlags		( DynFlags, DynFlag(..), dopt )
40
import StaticFlags	( opt_Static, opt_PIC )
41

42
import Digraph
43
import qualified Pretty
44
import Outputable
45
import FastString
46

47 48
-- DEBUGGING ONLY
--import OrdList
49

50
#ifdef NCG_DEBUG
51
import List		( intersperse )
52
#endif
53

54 55 56 57
import DATA_INT
import DATA_WORD
import DATA_BITS
import GLAEXTS
58

59 60 61
{-
The native-code generator has machine-independent and
machine-dependent modules.
62

63 64 65 66
This module ("AsmCodeGen") is the top-level machine-independent
module.  Before entering machine-dependent land, we do some
machine-independent optimisations (defined below) on the
'CmmStmts's.
67

68 69 70 71 72 73 74 75 76
We convert to the machine-specific 'Instr' datatype with
'cmmCodeGen', assuming an infinite supply of registers.  We then use
a machine-independent register allocator ('regAlloc') to rejoin
reality.  Obviously, 'regAlloc' has machine-specific helper
functions (see about "RegAllocInfo" below).

Finally, we order the basic blocks of the function so as to minimise
the number of jumps between blocks, by utilising fallthrough wherever
possible.
77 78

The machine-dependent bits break down as follows:
79 80

  * ["MachRegs"]  Everything about the target platform's machine
81 82 83
    registers (and immediate operands, and addresses, which tend to
    intermingle/interact with registers).

84
  * ["MachInstrs"]  Includes the 'Instr' datatype (possibly should
85
    have a module of its own), plus a miscellany of other things
86
    (e.g., 'targetDoubleSize', 'smStablePtrTable', ...)
87

88
  * ["MachCodeGen"]  is where 'Cmm' stuff turns into
89
    machine instructions.
90

91 92
  * ["PprMach"] 'pprInstr' turns an 'Instr' into text (well, really
    a 'Doc').
93

94 95
  * ["RegAllocInfo"] In the register allocator, we manipulate
    'MRegsState's, which are 'BitSet's, one bit per machine register.
96 97
    When we want to say something about a specific machine register
    (e.g., ``it gets clobbered by this instruction''), we set/unset
98
    its bit.  Obviously, we do this 'BitSet' thing for efficiency
99 100
    reasons.

101
    The 'RegAllocInfo' module collects together the machine-specific
102 103
    info needed to do register allocation.

104 105
   * ["RegisterAlloc"] The (machine-independent) register allocator.
-}
106

107 108
-- -----------------------------------------------------------------------------
-- Top-level of the native codegen
109

110
-- NB. We *lazilly* compile each block of code for space reasons.
111

112 113 114 115
nativeCodeGen :: DynFlags -> [Cmm] -> UniqSupply -> IO Pretty.Doc
nativeCodeGen dflags cmms us
  = let ((ppr_cmms, insn_sdoc, imports), _) = initUs us $
	   cgCmm (concat (map add_split cmms))
116

117
	cgCmm :: [CmmTop] -> UniqSM (Cmm, Pretty.Doc, [CLabel])
118 119 120 121 122 123
	cgCmm tops = 
	   lazyMapUs (cmmNativeGen dflags) tops  `thenUs` \ results -> 
	   let (cmms,docs,imps) = unzip3 results in
	   returnUs (Cmm cmms, my_vcat docs, concat imps)
    in do
    dumpIfSet_dyn dflags Opt_D_dump_opt_cmm "Optimised Cmm" (pprCmms [ppr_cmms])
124 125 126 127 128 129 130 131
    return (insn_sdoc Pretty.$$ dyld_stubs imports
#if HAVE_SUBSECTIONS_VIA_SYMBOLS
                -- On recent versions of Darwin, the linker supports
                -- dead-stripping of code and data on a per-symbol basis.
                -- There's a hack to make this work in PprMach.pprNatCmmTop.
            Pretty.$$ Pretty.text ".subsections_via_symbols"
#endif
            )
132

133
  where
134

135
    add_split (Cmm tops)
136 137
	| dopt Opt_SplitObjs dflags = split_marker : tops
	| otherwise		    = tops
138

139
    split_marker = CmmProc [] mkSplitMarkerLabel [] []
140

141 142
     	 -- Generate "symbol stubs" for all external symbols that might
	 -- come from a dynamic library.
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
{-    dyld_stubs imps = Pretty.vcat $ map pprDyldSymbolStub $
				    map head $ group $ sort imps-}
				    
	-- (Hack) sometimes two Labels pretty-print the same, but have
	-- different uniques; so we compare their text versions...
    dyld_stubs imps 
        | needImportedSymbols
          = Pretty.vcat $
            (pprGotDeclaration :) $
            map (pprImportedSymbol . fst . head) $
            groupBy (\(_,a) (_,b) -> a == b) $
            sortBy (\(_,a) (_,b) -> compare a b) $
            map doPpr $
            imps
        | otherwise
          = Pretty.empty
        
        where doPpr lbl = (lbl, Pretty.render $ pprCLabel lbl astyle)
              astyle = mkCodeStyle AsmStyle
162

163 164 165 166 167 168 169 170 171 172 173 174
#ifndef NCG_DEBUG
    my_vcat sds = Pretty.vcat sds
#else
    my_vcat sds = Pretty.vcat (
                      intersperse (
                         Pretty.char ' ' 
                            Pretty.$$ Pretty.ptext SLIT("# ___ncg_debug_marker")
                            Pretty.$$ Pretty.char ' '
                      ) 
                      sds
                   )
#endif
175 176


177 178 179
-- Complete native code generation phase for a single top-level chunk
-- of Cmm.

180
cmmNativeGen :: DynFlags -> CmmTop -> UniqSM (CmmTop, Pretty.Doc, [CLabel])
181 182 183 184
cmmNativeGen dflags cmm
   = {-# SCC "fixAssigns"       #-} 
 	fixAssignsTop cmm	     `thenUs` \ fixed_cmm ->
     {-# SCC "genericOpt"       #-} 
185
	cmmToCmm fixed_cmm           `bind`   \ (cmm, imports) ->
186 187 188 189
        (if dopt Opt_D_dump_opt_cmm dflags  -- space leak avoidance
	   then cmm 
	   else CmmData Text [])     `bind`   \ ppr_cmm ->
     {-# SCC "genMachCode"      #-}
190
	genMachCode cmm              `thenUs` \ (pre_regalloc, lastMinuteImports) ->
191 192 193 194 195 196 197 198 199
     {-# SCC "regAlloc"         #-}
	map regAlloc pre_regalloc    `bind`   \ with_regs ->
     {-# SCC "sequenceBlocks"   #-}
	map sequenceTop with_regs    `bind`   \ sequenced ->
     {-# SCC "x86fp_kludge"     #-}
	map x86fp_kludge sequenced   `bind`   \ final_mach_code ->
     {-# SCC "vcat"             #-}
	Pretty.vcat (map pprNatCmmTop final_mach_code)  `bind`   \ final_sdoc ->

200
        returnUs (ppr_cmm, final_sdoc Pretty.$$ Pretty.text "", lastMinuteImports ++ imports)
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
     where
        x86fp_kludge :: NatCmmTop -> NatCmmTop
        x86fp_kludge top@(CmmData _ _) = top
#if i386_TARGET_ARCH
        x86fp_kludge top@(CmmProc info lbl params code) = 
		CmmProc info lbl params (map bb_i386_insert_ffrees code)
		where
		  bb_i386_insert_ffrees (BasicBlock id instrs) =
			BasicBlock id (i386_insert_ffrees instrs)
#else
        x86fp_kludge top =  top
#endif

-- -----------------------------------------------------------------------------
-- Sequencing the basic blocks

-- Cmm BasicBlocks are self-contained entities: they always end in a
-- jump, either non-local or to another basic block in the same proc.
-- In this phase, we attempt to place the basic blocks in a sequence
-- such that as many of the local jumps as possible turn into
-- fallthroughs.

sequenceTop :: NatCmmTop -> NatCmmTop
sequenceTop top@(CmmData _ _) = top
sequenceTop (CmmProc info lbl params blocks) = 
  CmmProc info lbl params (sequenceBlocks blocks)

-- The algorithm is very simple (and stupid): we make a graph out of
-- the blocks where there is an edge from one block to another iff the
-- first block ends by jumping to the second.  Then we topologically
-- sort this graph.  Then traverse the list: for each block, we first
-- output the block, then if it has an out edge, we move the
-- destination of the out edge to the front of the list, and continue.

sequenceBlocks :: [NatBasicBlock] -> [NatBasicBlock]
sequenceBlocks [] = []
sequenceBlocks (entry:blocks) = 
  seqBlocks (mkNode entry : reverse (flattenSCCs (sccBlocks blocks)))
  -- the first block is the entry point ==> it must remain at the start.

sccBlocks :: [NatBasicBlock] -> [SCC (NatBasicBlock,Unique,[Unique])]
sccBlocks blocks = stronglyConnCompR (map mkNode blocks)

getOutEdges :: [Instr] -> [Unique]
getOutEdges instrs = case jumpDests (last instrs) [] of
			[one] -> [getUnique one]
			_many -> []
		-- we're only interested in the last instruction of
		-- the block, and only if it has a single destination.

mkNode block@(BasicBlock id instrs) = (block, getUnique id, getOutEdges instrs)

seqBlocks [] = []
seqBlocks ((block,_,[]) : rest)
  = block : seqBlocks rest
seqBlocks ((block@(BasicBlock id instrs),_,[next]) : rest)
  | can_fallthrough = BasicBlock id (init instrs) : seqBlocks rest'
  | otherwise       = block : seqBlocks rest'
  where
	(can_fallthrough, rest') = reorder next [] rest
	  -- TODO: we should do a better job for cycles; try to maximise the
	  -- fallthroughs within a loop.
seqBlocks _ = panic "AsmCodegen:seqBlocks"

reorder id accum [] = (False, reverse accum)
reorder id accum (b@(block,id',out) : rest)
  | id == id'  = (True, (block,id,out) : reverse accum ++ rest)
  | otherwise  = reorder id (b:accum) rest

-- -----------------------------------------------------------------------------
-- Instruction selection

-- Native code instruction selection for a chunk of stix code.  For
-- this part of the computation, we switch from the UniqSM monad to
-- the NatM monad.  The latter carries not only a Unique, but also an
-- Int denoting the current C stack pointer offset in the generated
-- code; this is needed for creating correct spill offsets on
-- architectures which don't offer, or for which it would be
-- prohibitively expensive to employ, a frame pointer register.  Viz,
-- x86.

-- The offset is measured in bytes, and indicates the difference
-- between the current (simulated) C stack-ptr and the value it was at
-- the beginning of the block.  For stacks which grow down, this value
-- should be either zero or negative.

-- Switching between the two monads whilst carrying along the same
-- Unique supply breaks abstraction.  Is that bad?

290
genMachCode :: CmmTop -> UniqSM ([NatCmmTop], [CLabel])
291 292

genMachCode cmm_top initial_us
293
  = let initial_st             = mkNatM_State initial_us 0
294 295 296 297
        (new_tops, final_st)   = initNat initial_st (cmmTopCodeGen cmm_top)
        final_us               = natm_us final_st
        final_delta            = natm_delta final_st
	final_imports          = natm_imports final_st
298 299
    in
        if   final_delta == 0
300
        then ((new_tops, final_imports), final_us)
301 302 303
        else pprPanic "genMachCode: nonzero final delta"
                      (int final_delta)

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
-- -----------------------------------------------------------------------------
-- Fixup assignments to global registers so that they assign to 
-- locations within the RegTable, if appropriate.

-- Note that we currently don't fixup reads here: they're done by
-- the generic optimiser below, to avoid having two separate passes
-- over the Cmm.

fixAssignsTop :: CmmTop -> UniqSM CmmTop
fixAssignsTop top@(CmmData _ _) = returnUs top
fixAssignsTop (CmmProc info lbl params blocks) =
  mapUs fixAssignsBlock blocks `thenUs` \ blocks' ->
  returnUs (CmmProc info lbl params blocks')

fixAssignsBlock :: CmmBasicBlock -> UniqSM CmmBasicBlock
fixAssignsBlock (BasicBlock id stmts) =
  fixAssigns stmts `thenUs` \ stmts' ->
  returnUs (BasicBlock id stmts')

fixAssigns :: [CmmStmt] -> UniqSM [CmmStmt]
fixAssigns stmts =
  mapUs fixAssign stmts `thenUs` \ stmtss ->
  returnUs (concat stmtss)

fixAssign :: CmmStmt -> UniqSM [CmmStmt]
fixAssign (CmmAssign (CmmGlobal BaseReg) src)
   = panic "cmmStmtConFold: assignment to BaseReg";

fixAssign (CmmAssign (CmmGlobal reg) src)
  | Left  realreg <- reg_or_addr
334
  = returnUs [CmmAssign (CmmGlobal reg) src]
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
  | Right baseRegAddr <- reg_or_addr
  = returnUs [CmmStore baseRegAddr src]
           -- Replace register leaves with appropriate StixTrees for
           -- the given target. GlobalRegs which map to a reg on this
           -- arch are left unchanged.  Assigning to BaseReg is always
           -- illegal, so we check for that.
  where
	reg_or_addr = get_GlobalReg_reg_or_addr reg

fixAssign (CmmCall target results args vols)
  = mapAndUnzipUs fixResult results `thenUs` \ (results',stores) ->
    returnUs (CmmCall target results' args vols : concat stores)
  where
	fixResult g@(CmmGlobal reg,hint) = 
	  case get_GlobalReg_reg_or_addr reg of
		Left realreg -> returnUs (g, [])
		Right baseRegAddr ->
		    getUniqueUs `thenUs` \ uq ->
		    let local = CmmLocal (LocalReg uq (globalRegRep reg)) in
		    returnUs ((local,hint), 
			      [CmmStore baseRegAddr (CmmReg local)])
	fixResult other =
	  returnUs (other,[])

fixAssign other_stmt = returnUs [other_stmt]

-- -----------------------------------------------------------------------------
-- Generic Cmm optimiser

{-
Here we do:

  (a) Constant folding
  (b) Simple inlining: a temporary which is assigned to and then
      used, once, can be shorted.
  (c) Replacement of references to GlobalRegs which do not have
      machine registers by the appropriate memory load (eg.
      Hp ==>  *(BaseReg + 34) ).
373 374 375 376
  (d) Position independent code and dynamic linking
        (i)  introduce the appropriate indirections
             and position independent refs
        (ii) compile a list of imported symbols
377 378 379 380 381 382 383

Ideas for other things we could do (ToDo):

  - shortcut jumps-to-jumps
  - eliminate dead code blocks
-}

384 385 386 387 388
cmmToCmm :: CmmTop -> (CmmTop, [CLabel])
cmmToCmm top@(CmmData _ _) = (top, [])
cmmToCmm (CmmProc info lbl params blocks) = runCmmOpt $ do
  blocks' <- mapM cmmBlockConFold (cmmPeep blocks)
  return $ CmmProc info lbl params blocks'
389

390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411
newtype CmmOptM a = CmmOptM ([CLabel] -> (# a, [CLabel] #))

instance Monad CmmOptM where
  return x = CmmOptM $ \imports -> (# x,imports #)
  (CmmOptM f) >>= g =
    CmmOptM $ \imports ->
                case f imports of
                  (# x, imports' #) ->
                    case g x of
                      CmmOptM g' -> g' imports'

addImportCmmOpt :: CLabel -> CmmOptM ()
addImportCmmOpt lbl = CmmOptM $ \imports -> (# (), lbl:imports #)

runCmmOpt :: CmmOptM a -> (a, [CLabel])
runCmmOpt (CmmOptM f) = case f [] of
                        (# result, imports #) -> (result, imports)

cmmBlockConFold :: CmmBasicBlock -> CmmOptM CmmBasicBlock
cmmBlockConFold (BasicBlock id stmts) = do
  stmts' <- mapM cmmStmtConFold stmts
  return $ BasicBlock id stmts'
412 413 414 415

cmmStmtConFold stmt
   = case stmt of
        CmmAssign reg src
416 417 418 419
           -> do src' <- cmmExprConFold False src
                 return $ case src' of
		   CmmReg reg' | reg == reg' -> CmmNop
		   new_src -> CmmAssign reg new_src
420 421

        CmmStore addr src
422 423 424
           -> do addr' <- cmmExprConFold False addr
                 src'  <- cmmExprConFold False src
                 return $ CmmStore addr' src'
425 426

        CmmJump addr regs
427 428
           -> do addr' <- cmmExprConFold True addr
                 return $ CmmJump addr' regs
429 430

	CmmCall target regs args vols
431 432 433 434 435 436 437 438 439
	   -> do target' <- case target of
			      CmmForeignCall e conv -> do
			        e' <- cmmExprConFold True e
			        return $ CmmForeignCall e' conv
			      other -> return other
                 args' <- mapM (\(arg, hint) -> do
                                  arg' <- cmmExprConFold False arg
                                  return (arg', hint)) args
	         return $ CmmCall target' regs args' vols
440 441

        CmmCondBranch test dest
442 443 444 445
           -> do test' <- cmmExprConFold False test
	         return $ case test' of
		   CmmLit (CmmInt 0 _) -> 
		     CmmComment (mkFastString ("deleted: " ++ 
446
					showSDoc (pprStmt stmt)))
447

448 449
		   CmmLit (CmmInt n _) -> CmmBranch dest
		   other -> CmmCondBranch test' dest
450

451
	CmmSwitch expr ids
452 453
	   -> do expr' <- cmmExprConFold False expr
	         return $ CmmSwitch expr' ids
454 455

        other
456
           -> return other
457

458

459
cmmExprConFold isJumpTarget expr
460
   = case expr of
461
        CmmLoad addr rep
462 463
           -> do addr' <- cmmExprConFold False addr
                 return $ CmmLoad addr' rep
464 465 466

        CmmMachOp mop args
           -- For MachOps, we first optimize the children, and then we try 
467
           -- our hand at some constant-folding.
468 469 470 471 472 473 474 475 476 477 478
           -> do args' <- mapM (cmmExprConFold False) args
                 return $ cmmMachOpFold mop args'

        CmmLit (CmmLabel lbl)
           -> cmmMakeDynamicReference addImportCmmOpt isJumpTarget lbl
        CmmLit (CmmLabelOff lbl off)
           -> do dynRef <- cmmMakeDynamicReference addImportCmmOpt isJumpTarget lbl
                 return $ cmmMachOpFold (MO_Add wordRep) [
                     dynRef,
                     (CmmLit $ CmmInt (fromIntegral off) wordRep)
                   ]
479 480

#if powerpc_TARGET_ARCH
481
           -- On powerpc (non-PIC), it's easier to jump directly to a label than
482 483 484
           -- to use the register table, so we replace these registers
           -- with the corresponding labels:
        CmmReg (CmmGlobal GCEnter1)
485 486 487
          | not opt_PIC
          -> cmmExprConFold isJumpTarget $
             CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_enter_1"))) 
488
        CmmReg (CmmGlobal GCFun)
489 490 491
          | not opt_PIC
          -> cmmExprConFold isJumpTarget $
             CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_fun")))
492 493 494 495 496 497 498 499 500 501
#endif

        CmmReg (CmmGlobal mid)
           -- Replace register leaves with appropriate StixTrees for
           -- the given target.  MagicIds which map to a reg on this
           -- arch are left unchanged.  For the rest, BaseReg is taken
           -- to mean the address of the reg table in MainCapability,
           -- and for all others we generate an indirection to its
           -- location in the register table.
           -> case get_GlobalReg_reg_or_addr mid of
502
                 Left  realreg -> return expr
503
                 Right baseRegAddr 
504
                    -> case mid of 
505 506
                          BaseReg -> cmmExprConFold False baseRegAddr
                          other   -> cmmExprConFold False (CmmLoad baseRegAddr 
507 508 509
							(globalRegRep mid))
	   -- eliminate zero offsets
	CmmRegOff reg 0
510
	   -> cmmExprConFold False (CmmReg reg)
511 512 513 514 515 516

        CmmRegOff (CmmGlobal mid) offset
           -- RegOf leaves are just a shorthand form. If the reg maps
           -- to a real reg, we keep the shorthand, otherwise, we just
           -- expand it and defer to the above code. 
           -> case get_GlobalReg_reg_or_addr mid of
517
                Left  realreg -> return expr
518
                Right baseRegAddr
519
                   -> cmmExprConFold False (CmmMachOp (MO_Add wordRep) [
520 521 522
                                        CmmReg (CmmGlobal mid),
                                        CmmLit (CmmInt (fromIntegral offset)
                                                       wordRep)])
523
        other
524
           -> return other
525 526


527 528
-- -----------------------------------------------------------------------------
-- MachOp constant folder
529

530 531
-- Now, try to constant-fold the MachOps.  The arguments have already
-- been optimized and folded.
532

533 534 535 536
cmmMachOpFold
    :: MachOp	    	-- The operation from an CmmMachOp
    -> [CmmExpr]   	-- The optimized arguments
    -> CmmExpr
537

538 539 540 541
cmmMachOpFold op arg@[CmmLit (CmmInt x rep)]
  = case op of
      MO_S_Neg r -> CmmLit (CmmInt (-x) rep)
      MO_Not r   -> CmmLit (CmmInt (complement x) rep)
542

543 544 545 546 547 548 549
	-- these are interesting: we must first narrow to the 
	-- "from" type, in order to truncate to the correct size.
	-- The final narrow/widen to the destination type
	-- is implicit in the CmmLit.
      MO_S_Conv from to -> CmmLit (CmmInt (narrowS from x) to)
      MO_U_Conv from to -> CmmLit (CmmInt (narrowU from x) to)
      _  -> panic "cmmMachOpFold: unknown unary op"
550

551 552 553
-- Eliminate conversion NOPs
cmmMachOpFold (MO_S_Conv rep1 rep2) [x] | rep1 == rep2 = x
cmmMachOpFold (MO_U_Conv rep1 rep2) [x] | rep1 == rep2 = x
554

555 556
-- ToDo: eliminate multiple conversions.  Be careful though: can't remove
-- a narrowing, and can't remove conversions to/from floating point types.
557

558 559 560
-- ToDo: eliminate nested comparisons:
--    CmmMachOp MO_Lt [CmmMachOp MO_Eq [x,y], CmmLit (CmmInt 0 _)]
-- turns into a simple equality test.
561

562
cmmMachOpFold mop args@[CmmLit (CmmInt x xrep), CmmLit (CmmInt y _)]
563
  = case mop of
564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
	-- for comparisons: don't forget to narrow the arguments before
	-- comparing, since they might be out of range.
    	MO_Eq r   -> CmmLit (CmmInt (if x_u == y_u then 1 else 0) wordRep)
    	MO_Ne r   -> CmmLit (CmmInt (if x_u /= y_u then 1 else 0) wordRep)

    	MO_U_Gt r -> CmmLit (CmmInt (if x_u >  y_u then 1 else 0) wordRep)
    	MO_U_Ge r -> CmmLit (CmmInt (if x_u >= y_u then 1 else 0) wordRep)
    	MO_U_Lt r -> CmmLit (CmmInt (if x_u <  y_u then 1 else 0) wordRep)
    	MO_U_Le r -> CmmLit (CmmInt (if x_u <= y_u then 1 else 0) wordRep)

    	MO_S_Gt r -> CmmLit (CmmInt (if x_s >  y_s then 1 else 0) wordRep) 
    	MO_S_Ge r -> CmmLit (CmmInt (if x_s >= y_s then 1 else 0) wordRep)
    	MO_S_Lt r -> CmmLit (CmmInt (if x_s <  y_s then 1 else 0) wordRep)
    	MO_S_Le r -> CmmLit (CmmInt (if x_s <= y_s then 1 else 0) wordRep)

    	MO_Add r -> CmmLit (CmmInt (x + y) r)
    	MO_Sub r -> CmmLit (CmmInt (x - y) r)
    	MO_Mul r -> CmmLit (CmmInt (x * y) r)
    	MO_S_Quot r | y /= 0 -> CmmLit (CmmInt (x `quot` y) r)
    	MO_S_Rem  r | y /= 0 -> CmmLit (CmmInt (x `rem` y) r)

	MO_And   r -> CmmLit (CmmInt (x .&. y) r)
	MO_Or    r -> CmmLit (CmmInt (x .|. y) r)
	MO_Xor   r -> CmmLit (CmmInt (x `xor` y) r)

        MO_Shl   r -> CmmLit (CmmInt (x `shiftL` fromIntegral y) r)
        MO_U_Shr r -> CmmLit (CmmInt (x_u `shiftR` fromIntegral y) r)
        MO_S_Shr r -> CmmLit (CmmInt (x `shiftR` fromIntegral y) r)

	other      -> CmmMachOp mop args

   where
	x_u = narrowU xrep x
	y_u = narrowU xrep y
	x_s = narrowS xrep x
	y_s = narrowS xrep y
	

-- When possible, shift the constants to the right-hand side, so that we
-- can match for strength reductions.  Note that the code generator will
-- also assume that constants have been shifted to the right when
-- possible.

cmmMachOpFold op [x@(CmmLit _), y]
   | not (isLit y) && isCommutableMachOp op 
   = cmmMachOpFold op [y, x]

-- Turn (a+b)+c into a+(b+c) where possible.  Because literals are
-- moved to the right, it is more likely that we will find
-- opportunities for constant folding when the expression is
-- right-associated.
--
-- ToDo: this appears to introduce a quadratic behaviour due to the
-- nested cmmMachOpFold.  Can we fix this?
618 619 620 621 622 623
--
-- Why do we check isLit arg1?  If arg1 is a lit, it means that arg2
-- is also a lit (otherwise arg1 would be on the right).  If we
-- put arg1 on the left of the rearranged expression, we'll get into a
-- loop:  (x1+x2)+x3 => x1+(x2+x3)  => (x2+x3)+x1 => x2+(x3+x1) ...
--
624
cmmMachOpFold mop1 [CmmMachOp mop2 [arg1,arg2], arg3]
625
   | mop1 == mop2 && isAssociativeMachOp mop1 && not (isLit arg1)
626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695
   = cmmMachOpFold mop1 [arg1, cmmMachOpFold mop2 [arg2,arg3]]

-- Make a RegOff if we can
cmmMachOpFold (MO_Add _) [CmmReg reg, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (fromIntegral (narrowS rep n))
cmmMachOpFold (MO_Add _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (off + fromIntegral (narrowS rep n))
cmmMachOpFold (MO_Sub _) [CmmReg reg, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (- fromIntegral (narrowS rep n))
cmmMachOpFold (MO_Sub _) [CmmRegOff reg off, CmmLit (CmmInt n rep)]
  = CmmRegOff reg (off - fromIntegral (narrowS rep n))

-- Fold label(+/-)offset into a CmmLit where possible

cmmMachOpFold (MO_Add _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
  = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
cmmMachOpFold (MO_Add _) [CmmLit (CmmInt i rep), CmmLit (CmmLabel lbl)]
  = CmmLit (CmmLabelOff lbl (fromIntegral (narrowU rep i)))
cmmMachOpFold (MO_Sub _) [CmmLit (CmmLabel lbl), CmmLit (CmmInt i rep)]
  = CmmLit (CmmLabelOff lbl (fromIntegral (negate (narrowU rep i))))

-- We can often do something with constants of 0 and 1 ...

cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 0 _))]
  = case mop of
    	MO_Add   r -> x
    	MO_Sub   r -> x
    	MO_Mul   r -> y
    	MO_And   r -> y
    	MO_Or    r -> x
    	MO_Xor   r -> x
    	MO_Shl   r -> x
    	MO_S_Shr r -> x
    	MO_U_Shr r -> x
        MO_Ne    r | isComparisonExpr x -> x
	MO_Eq    r | Just x' <- maybeInvertConditionalExpr x -> x'
	MO_U_Gt  r | isComparisonExpr x -> x
	MO_S_Gt  r | isComparisonExpr x -> x
	MO_U_Lt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_S_Lt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_U_Ge  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
	MO_S_Ge  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
	MO_U_Le  r | Just x' <- maybeInvertConditionalExpr x -> x'
	MO_S_Le  r | Just x' <- maybeInvertConditionalExpr x -> x'
    	other    -> CmmMachOp mop args

cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt 1 rep))]
  = case mop of
    	MO_Mul    r -> x
    	MO_S_Quot r -> x
    	MO_U_Quot r -> x
    	MO_S_Rem  r -> CmmLit (CmmInt 0 rep)
    	MO_U_Rem  r -> CmmLit (CmmInt 0 rep)
        MO_Ne    r | Just x' <- maybeInvertConditionalExpr x -> x'
	MO_Eq    r | isComparisonExpr x -> x
	MO_U_Lt  r | Just x' <- maybeInvertConditionalExpr x -> x'
	MO_S_Lt  r | Just x' <- maybeInvertConditionalExpr x -> x'
	MO_U_Gt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_S_Gt  r | isComparisonExpr x -> CmmLit (CmmInt 0 wordRep)
	MO_U_Le  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
	MO_S_Le  r | isComparisonExpr x -> CmmLit (CmmInt 1 wordRep)
	MO_U_Ge  r | isComparisonExpr x -> x
	MO_S_Ge  r | isComparisonExpr x -> x
    	other       -> CmmMachOp mop args

-- Now look for multiplication/division by powers of 2 (integers).

cmmMachOpFold mop args@[x, y@(CmmLit (CmmInt n _))]
  = case mop of
    	MO_Mul rep
696 697
           -> case exactLog2 n of
                 Nothing -> unchanged
698 699
                 Just p  -> CmmMachOp (MO_Shl rep) [x, CmmLit (CmmInt p rep)]
    	MO_S_Quot rep
700 701
           -> case exactLog2 n of
                 Nothing -> unchanged
702
                 Just p  -> CmmMachOp (MO_S_Shr rep) [x, CmmLit (CmmInt p rep)]
703 704 705
    	other 
           -> unchanged
    where
706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806
       unchanged = CmmMachOp mop args

-- Anything else is just too hard.

cmmMachOpFold mop args = CmmMachOp mop args

-- -----------------------------------------------------------------------------
-- exactLog2

-- This algorithm for determining the $\log_2$ of exact powers of 2 comes
-- from GCC.  It requires bit manipulation primitives, and we use GHC
-- extensions.  Tough.
-- 
-- Used to be in MachInstrs --SDM.
-- ToDo: remove use of unboxery --SDM.

w2i x = word2Int# x
i2w x = int2Word# x

exactLog2 :: Integer -> Maybe Integer
exactLog2 x
  = if (x <= 0 || x >= 2147483648) then
       Nothing
    else
       case iUnbox (fromInteger x) of { x# ->
       if (w2i ((i2w x#) `and#` (i2w (0# -# x#))) /=# x#) then
	  Nothing
       else
	  Just (toInteger (iBox (pow2 x#)))
       }
  where
    pow2 x# | x# ==# 1# = 0#
            | otherwise = 1# +# pow2 (w2i (i2w x# `shiftRL#` 1#))


-- -----------------------------------------------------------------------------
-- widening / narrowing

narrowU :: MachRep -> Integer -> Integer
narrowU I8  x = fromIntegral (fromIntegral x :: Word8)
narrowU I16 x = fromIntegral (fromIntegral x :: Word16)
narrowU I32 x = fromIntegral (fromIntegral x :: Word32)
narrowU I64 x = fromIntegral (fromIntegral x :: Word64)
narrowU _ _ = panic "narrowTo"

narrowS :: MachRep -> Integer -> Integer
narrowS I8  x = fromIntegral (fromIntegral x :: Int8)
narrowS I16 x = fromIntegral (fromIntegral x :: Int16)
narrowS I32 x = fromIntegral (fromIntegral x :: Int32)
narrowS I64 x = fromIntegral (fromIntegral x :: Int64)
narrowS _ _ = panic "narrowTo"

-- -----------------------------------------------------------------------------
-- The mini-inliner

-- This pass inlines assignments to temporaries that are used just
-- once in the very next statement only.  Generalising this would be
-- quite difficult (have to take into account aliasing of memory
-- writes, and so on), but at the moment it catches a number of useful
-- cases and lets the code generator generate much better code.

-- NB. This assumes that temporaries are single-assignment.

cmmPeep :: [CmmBasicBlock] -> [CmmBasicBlock]
cmmPeep blocks = map do_inline blocks 
  where 
	blockUses (BasicBlock _ stmts)
	 = foldr (plusUFM_C (+)) emptyUFM (map getStmtUses stmts)

	uses = foldr (plusUFM_C (+)) emptyUFM (map blockUses blocks)

	do_inline (BasicBlock id stmts)
	 = BasicBlock id (cmmMiniInline uses stmts)


cmmMiniInline :: UniqFM Int -> [CmmStmt] -> [CmmStmt]
cmmMiniInline uses [] = []
cmmMiniInline uses (stmt@(CmmAssign (CmmLocal (LocalReg u _)) expr) : stmts)
  | Just 1 <- lookupUFM uses u,
    Just stmts' <- lookForInline u expr stmts
  = 
#ifdef NCG_DEBUG
     trace ("nativeGen: inlining " ++ showSDoc (pprStmt stmt)) $
#endif
     cmmMiniInline uses stmts'

cmmMiniInline uses (stmt:stmts)
  = stmt : cmmMiniInline uses stmts


-- Try to inline a temporary assignment.  We can skip over assignments to
-- other tempoararies, because we know that expressions aren't side-effecting
-- and temporaries are single-assignment.
lookForInline u expr (stmt@(CmmAssign (CmmLocal (LocalReg u' _)) rhs) : rest)
  | u /= u' 
  = case lookupUFM (getExprUses rhs) u of
	Just 1 -> Just (inlineStmt u expr stmt : rest)
	_other -> case lookForInline u expr rest of
		     Nothing    -> Nothing
		     Just stmts -> Just (stmt:stmts)

807 808 809
lookForInline u expr (CmmNop : rest)
  = lookForInline u expr rest

810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867
lookForInline u expr (stmt:stmts)
  = case lookupUFM (getStmtUses stmt) u of
	Just 1 -> Just (inlineStmt u expr stmt : stmts)
	_other -> Nothing

-- -----------------------------------------------------------------------------
-- Boring Cmm traversals for collecting usage info and substitutions.

getStmtUses :: CmmStmt -> UniqFM Int
getStmtUses (CmmAssign _ e) = getExprUses e
getStmtUses (CmmStore e1 e2) = plusUFM_C (+) (getExprUses e1) (getExprUses e2)
getStmtUses (CmmCall target _ es _)
   = plusUFM_C (+) (uses target) (getExprsUses (map fst es))
   where uses (CmmForeignCall e _) = getExprUses e
	 uses _ = emptyUFM
getStmtUses (CmmCondBranch e _) = getExprUses e
getStmtUses (CmmSwitch e _) = getExprUses e
getStmtUses (CmmJump e _) = getExprUses e
getStmtUses _ = emptyUFM

getExprUses :: CmmExpr -> UniqFM Int
getExprUses (CmmReg (CmmLocal (LocalReg u _))) = unitUFM u 1
getExprUses (CmmRegOff (CmmLocal (LocalReg u _)) _) = unitUFM u 1
getExprUses (CmmLoad e _) = getExprUses e
getExprUses (CmmMachOp _ es) = getExprsUses es
getExprUses _other = emptyUFM

getExprsUses es = foldr (plusUFM_C (+)) emptyUFM (map getExprUses es)

inlineStmt :: Unique -> CmmExpr -> CmmStmt -> CmmStmt
inlineStmt u a (CmmAssign r e) = CmmAssign r (inlineExpr u a e)
inlineStmt u a (CmmStore e1 e2) = CmmStore (inlineExpr u a e1) (inlineExpr u a e2)
inlineStmt u a (CmmCall target regs es vols)
   = CmmCall (infn target) regs es' vols
   where infn (CmmForeignCall fn cconv) = CmmForeignCall fn cconv
	 infn (CmmPrim p) = CmmPrim p
	 es' = [ (inlineExpr u a e, hint) | (e,hint) <- es ]
inlineStmt u a (CmmCondBranch e d) = CmmCondBranch (inlineExpr u a e) d
inlineStmt u a (CmmSwitch e d) = CmmSwitch (inlineExpr u a e) d
inlineStmt u a (CmmJump e d) = CmmJump (inlineExpr u a e) d
inlineStmt u a other_stmt = other_stmt

inlineExpr :: Unique -> CmmExpr -> CmmExpr -> CmmExpr
inlineExpr u a e@(CmmReg (CmmLocal (LocalReg u' _)))
  | u == u' = a
  | otherwise = e
inlineExpr u a e@(CmmRegOff (CmmLocal (LocalReg u' rep)) off)
  | u == u' = CmmMachOp (MO_Add rep) [a, CmmLit (CmmInt (fromIntegral off) rep)]
  | otherwise = e
inlineExpr u a (CmmLoad e rep) = CmmLoad (inlineExpr u a e) rep
inlineExpr u a (CmmMachOp op es) = CmmMachOp op (map (inlineExpr u a) es)
inlineExpr u a other_expr = other_expr

-- -----------------------------------------------------------------------------
-- Utils

bind f x = x $! f

868 869 870
isLit (CmmLit _) = True
isLit _          = False

871 872 873 874 875 876 877 878
isComparisonExpr :: CmmExpr -> Bool
isComparisonExpr (CmmMachOp op _) = isComparisonMachOp op
isComparisonExpr _other 	    = False

maybeInvertConditionalExpr :: CmmExpr -> Maybe CmmExpr
maybeInvertConditionalExpr (CmmMachOp op args) 
  | Just op' <- maybeInvertComparison op = Just (CmmMachOp op' args)
maybeInvertConditionalExpr _ = Nothing
879 880
\end{code}