AsmCodeGen.lhs 23.1 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"
13
#include "nativeGen/NCG.h"
14

15
import MachInstrs
16
import MachRegs
17
import MachCodeGen
18
import PprMach
19
import RegisterAlloc
20
import RegAllocInfo
21
import NCGMonad
22
import PositionIndependentCode
23 24

import Cmm
25
import CmmOpt		( cmmMiniInline, cmmMachOpFold )
26 27
import PprCmm		( pprStmt, pprCmms )
import MachOp
28
import CLabel
29

30 31 32 33
import UniqFM
import Unique		( Unique, getUnique )
import UniqSupply
import FastTypes
34
import List		( groupBy, sortBy )
35
import ErrUtils		( dumpIfSet_dyn )
36
import DynFlags
37
import StaticFlags	( opt_Static, opt_PIC )
38
import Util
39
import Config           ( cProjectVersion )
40

41
import Digraph
42
import qualified Pretty
43
import Outputable
44
import FastString
45

46 47
-- DEBUGGING ONLY
--import OrdList
48

49
#ifdef NCG_DEBUG
50
import List		( intersperse )
51
#endif
52

Simon Marlow's avatar
Simon Marlow committed
53 54 55 56
import Data.Int
import Data.Word
import Data.Bits
import GHC.Exts
57

58 59 60
{-
The native-code generator has machine-independent and
machine-dependent modules.
61

62 63 64 65
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.
66

67 68 69 70 71 72 73 74 75
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.
76 77

The machine-dependent bits break down as follows:
78 79

  * ["MachRegs"]  Everything about the target platform's machine
80 81 82
    registers (and immediate operands, and addresses, which tend to
    intermingle/interact with registers).

83
  * ["MachInstrs"]  Includes the 'Instr' datatype (possibly should
84
    have a module of its own), plus a miscellany of other things
85
    (e.g., 'targetDoubleSize', 'smStablePtrTable', ...)
86

87
  * ["MachCodeGen"]  is where 'Cmm' stuff turns into
88
    machine instructions.
89

90 91
  * ["PprMach"] 'pprInstr' turns an 'Instr' into text (well, really
    a 'Doc').
92

93 94
  * ["RegAllocInfo"] In the register allocator, we manipulate
    'MRegsState's, which are 'BitSet's, one bit per machine register.
95 96
    When we want to say something about a specific machine register
    (e.g., ``it gets clobbered by this instruction''), we set/unset
97
    its bit.  Obviously, we do this 'BitSet' thing for efficiency
98 99
    reasons.

100
    The 'RegAllocInfo' module collects together the machine-specific
101 102
    info needed to do register allocation.

103 104
   * ["RegisterAlloc"] The (machine-independent) register allocator.
-}
105

106 107
-- -----------------------------------------------------------------------------
-- Top-level of the native codegen
108

109
-- NB. We *lazilly* compile each block of code for space reasons.
110

111
nativeCodeGen :: DynFlags -> [RawCmm] -> UniqSupply -> IO Pretty.Doc
112
nativeCodeGen dflags cmms us
113
  = let (res, _) = initUs us $
114
	   cgCmm (concat (map add_split cmms))
115

116
	cgCmm :: [RawCmmTop] -> UniqSM (RawCmm, Pretty.Doc, [CLabel])
117 118
	cgCmm tops = 
	   lazyMapUs (cmmNativeGen dflags) tops  `thenUs` \ results -> 
119
	   case unzip3 results of { (cmms,docs,imps) ->
120
	   returnUs (Cmm cmms, my_vcat docs, concat imps)
121 122 123
	   }
    in 
    case res of { (ppr_cmms, insn_sdoc, imports) -> do
124
    dumpIfSet_dyn dflags Opt_D_dump_opt_cmm "Optimised Cmm" (pprCmms [ppr_cmms])
125 126 127 128 129 130
    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"
131 132 133 134 135 136 137 138 139
#endif
#if HAVE_GNU_NONEXEC_STACK
                -- On recent GNU ELF systems one can mark an object file
                -- as not requiring an executable stack. If all objects
                -- linked into a program have this note then the program
                -- will not use an executable stack, which is good for
                -- security. GHC generated code does not need an executable
                -- stack so add the note in:
            Pretty.$$ Pretty.text ".section .note.GNU-stack,\"\",@progbits"
140
#endif
141
#if !defined(darwin_TARGET_OS)
142 143 144 145 146 147
                -- And just because every other compiler does, lets stick in
		-- an identifier directive: .ident "GHC x.y.z"
	    Pretty.$$ let compilerIdent = Pretty.text "GHC" Pretty.<+>
	                                  Pretty.text cProjectVersion
                       in Pretty.text ".ident" Pretty.<+>
                          Pretty.doubleQuotes compilerIdent
148
#endif
149
            )
150
   }
151

152
  where
153

154
    add_split (Cmm tops)
155 156
	| dopt Opt_SplitObjs dflags = split_marker : tops
	| otherwise		    = tops
157

158
    split_marker = CmmProc [] mkSplitMarkerLabel [] []
159

160 161
     	 -- Generate "symbol stubs" for all external symbols that might
	 -- come from a dynamic library.
162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180
{-    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
181

182 183 184 185 186 187 188 189 190 191 192 193
#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
194 195


196 197 198
-- Complete native code generation phase for a single top-level chunk
-- of Cmm.

199
cmmNativeGen :: DynFlags -> RawCmmTop -> UniqSM (RawCmmTop, Pretty.Doc, [CLabel])
200 201 202 203
cmmNativeGen dflags cmm
   = {-# SCC "fixAssigns"       #-} 
 	fixAssignsTop cmm	     `thenUs` \ fixed_cmm ->
     {-# SCC "genericOpt"       #-} 
204
	cmmToCmm dflags fixed_cmm           `bind`   \ (cmm, imports) ->
205 206 207 208
        (if dopt Opt_D_dump_opt_cmm dflags  -- space leak avoidance
	   then cmm 
	   else CmmData Text [])     `bind`   \ ppr_cmm ->
     {-# SCC "genMachCode"      #-}
209
	genMachCode dflags cmm       `thenUs` \ (pre_regalloc, lastMinuteImports) ->
210
     {-# SCC "regAlloc"         #-}
211
	mapUs regAlloc pre_regalloc `thenUs`   \ with_regs ->
212 213
     {-# SCC "shortcutBranches"   #-}
        shortcutBranches dflags with_regs `bind` \ shorted -> 
214
     {-# SCC "sequenceBlocks"   #-}
215
	map sequenceTop shorted        `bind`   \ sequenced ->
216 217 218 219 220
     {-# SCC "x86fp_kludge"     #-}
	map x86fp_kludge sequenced   `bind`   \ final_mach_code ->
     {-# SCC "vcat"             #-}
	Pretty.vcat (map pprNatCmmTop final_mach_code)  `bind`   \ final_sdoc ->

221
        returnUs (ppr_cmm, final_sdoc Pretty.$$ Pretty.text "", lastMinuteImports ++ imports)
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
     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) = 
247
  CmmProc info lbl params (makeFarBranches $ sequenceBlocks blocks)
248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290

-- 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

291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330

-- -----------------------------------------------------------------------------
-- Making far branches

-- Conditional branches on PowerPC are limited to +-32KB; if our Procs get too
-- big, we have to work around this limitation.

makeFarBranches :: [NatBasicBlock] -> [NatBasicBlock]

#if powerpc_TARGET_ARCH
makeFarBranches blocks
    | last blockAddresses < nearLimit = blocks
    | otherwise = zipWith handleBlock blockAddresses blocks
    where
        blockAddresses = scanl (+) 0 $ map blockLen blocks
        blockLen (BasicBlock _ instrs) = length instrs
        
        handleBlock addr (BasicBlock id instrs)
                = BasicBlock id (zipWith makeFar [addr..] instrs)
        
        makeFar addr (BCC ALWAYS tgt) = BCC ALWAYS tgt
        makeFar addr (BCC cond tgt)
            | abs (addr - targetAddr) >= nearLimit
            = BCCFAR cond tgt
            | otherwise
            = BCC cond tgt
            where Just targetAddr = lookupUFM blockAddressMap tgt
        makeFar addr other            = other
        
        nearLimit = 7000 -- 8192 instructions are allowed; let's keep some
                         -- distance, as we have a few pseudo-insns that are
                         -- pretty-printed as multiple instructions,
                         -- and it's just not worth the effort to calculate
                         -- things exactly
        
        blockAddressMap = listToUFM $ zip (map blockId blocks) blockAddresses
#else
makeFarBranches = id
#endif

331 332 333 334 335 336 337 338 339 340 341 342
-- -----------------------------------------------------------------------------
-- Shortcut branches

shortcutBranches :: DynFlags -> [NatCmmTop] -> [NatCmmTop]
shortcutBranches dflags tops
  | optLevel dflags < 1 = tops    -- only with -O or higher
  | otherwise           = map (apply_mapping mapping) tops'
  where
    (tops', mappings) = mapAndUnzip build_mapping tops
    mapping = foldr plusUFM emptyUFM mappings

build_mapping top@(CmmData _ _) = (top, emptyUFM)
Ian Lynagh's avatar
Ian Lynagh committed
343
build_mapping (CmmProc info lbl params [])
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
  = (CmmProc info lbl params [], emptyUFM)
build_mapping (CmmProc info lbl params (head:blocks))
  = (CmmProc info lbl params (head:others), mapping)
        -- drop the shorted blocks, but don't ever drop the first one,
        -- because it is pointed to by a global label.
  where
    -- find all the blocks that just consist of a jump that can be
    -- shorted.
    (shortcut_blocks, others) = partitionWith split blocks
    split (BasicBlock id [insn]) | Just dest <- canShortcut insn 
                                 = Left (id,dest)
    split other = Right other

    -- build a mapping from BlockId to JumpDest for shorting branches
    mapping = foldl add emptyUFM shortcut_blocks
    add ufm (id,dest) = addToUFM ufm id dest
    
apply_mapping ufm (CmmData sec statics) 
  = CmmData sec (map (shortcutStatic (lookupUFM ufm)) statics)
  -- we need to get the jump tables, so apply the mapping to the entries
  -- of a CmmData too.
apply_mapping ufm (CmmProc info lbl params blocks)
  = CmmProc info lbl params (map short_bb blocks)
  where
    short_bb (BasicBlock id insns) = BasicBlock id $! map short_insn insns
    short_insn i = shortcutJump (lookupUFM ufm) i
                 -- shortcutJump should apply the mapping repeatedly,
                 -- just in case we can short multiple branches.

373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392
-- -----------------------------------------------------------------------------
-- 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?

393
genMachCode :: DynFlags -> RawCmmTop -> UniqSM ([NatCmmTop], [CLabel])
394

395
genMachCode dflags cmm_top
396
  = do	{ initial_us <- getUs
397
	; let initial_st           = mkNatM_State initial_us 0 dflags
398 399 400 401 402 403 404
	      (new_tops, final_st) = initNat initial_st (cmmTopCodeGen cmm_top)
	      final_delta          = natm_delta final_st
	      final_imports        = natm_imports final_st
	; if   final_delta == 0
          then return (new_tops, final_imports)
          else pprPanic "genMachCode: nonzero final delta" (int final_delta)
    }
405

406 407 408 409 410 411 412 413
-- -----------------------------------------------------------------------------
-- 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.

414
fixAssignsTop :: RawCmmTop -> UniqSM RawCmmTop
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
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 reg) src)
  | Left  realreg <- reg_or_addr
433
  = returnUs [CmmAssign (CmmGlobal reg) src]
434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456
  | 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 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) ).
457 458 459 460
  (d) Position independent code and dynamic linking
        (i)  introduce the appropriate indirections
             and position independent refs
        (ii) compile a list of imported symbols
461 462 463 464 465

Ideas for other things we could do (ToDo):

  - shortcut jumps-to-jumps
  - eliminate dead code blocks
466 467 468
  - simple CSE: if an expr is assigned to a temp, then replace later occs of
    that expr with the temp, until the expr is no longer valid (can push through
    temp assignments, and certain assigns to mem...)
469 470
-}

471 472 473
cmmToCmm :: DynFlags -> RawCmmTop -> (RawCmmTop, [CLabel])
cmmToCmm _ top@(CmmData _ _) = (top, [])
cmmToCmm dflags (CmmProc info lbl params blocks) = runCmmOpt dflags $ do
474
  blocks' <- mapM cmmBlockConFold (cmmMiniInline blocks)
475
  return $ CmmProc info lbl params blocks'
476

477
newtype CmmOptM a = CmmOptM (([CLabel], DynFlags) -> (# a, [CLabel] #))
478 479

instance Monad CmmOptM where
480
  return x = CmmOptM $ \(imports, _) -> (# x,imports #)
481
  (CmmOptM f) >>= g =
482 483
    CmmOptM $ \(imports, dflags) ->
                case f (imports, dflags) of
484 485
                  (# x, imports' #) ->
                    case g x of
486
                      CmmOptM g' -> g' (imports', dflags)
487 488

addImportCmmOpt :: CLabel -> CmmOptM ()
489
addImportCmmOpt lbl = CmmOptM $ \(imports, dflags) -> (# (), lbl:imports #)
490

491 492 493 494 495
getDynFlagsCmmOpt :: CmmOptM DynFlags
getDynFlagsCmmOpt = CmmOptM $ \(imports, dflags) -> (# dflags, imports #)

runCmmOpt :: DynFlags -> CmmOptM a -> (a, [CLabel])
runCmmOpt dflags (CmmOptM f) = case f ([], dflags) of
496 497 498 499 500 501
                        (# result, imports #) -> (result, imports)

cmmBlockConFold :: CmmBasicBlock -> CmmOptM CmmBasicBlock
cmmBlockConFold (BasicBlock id stmts) = do
  stmts' <- mapM cmmStmtConFold stmts
  return $ BasicBlock id stmts'
502 503 504 505

cmmStmtConFold stmt
   = case stmt of
        CmmAssign reg src
506
           -> do src' <- cmmExprConFold DataReference src
507 508 509
                 return $ case src' of
		   CmmReg reg' | reg == reg' -> CmmNop
		   new_src -> CmmAssign reg new_src
510 511

        CmmStore addr src
512 513
           -> do addr' <- cmmExprConFold DataReference addr
                 src'  <- cmmExprConFold DataReference src
514
                 return $ CmmStore addr' src'
515 516

        CmmJump addr regs
517
           -> do addr' <- cmmExprConFold JumpReference addr
518
                 return $ CmmJump addr' regs
519

520
	CmmCall target regs args srt returns
521
	   -> do target' <- case target of
522
			      CmmCallee e conv -> do
523
			        e' <- cmmExprConFold CallReference e
524
			        return $ CmmCallee e' conv
525 526
			      other -> return other
                 args' <- mapM (\(arg, hint) -> do
527
                                  arg' <- cmmExprConFold DataReference arg
528
                                  return (arg', hint)) args
529
	         return $ CmmCall target' regs args' srt returns
530 531

        CmmCondBranch test dest
532
           -> do test' <- cmmExprConFold DataReference test
533 534 535
	         return $ case test' of
		   CmmLit (CmmInt 0 _) -> 
		     CmmComment (mkFastString ("deleted: " ++ 
536
					showSDoc (pprStmt stmt)))
537

538 539
		   CmmLit (CmmInt n _) -> CmmBranch dest
		   other -> CmmCondBranch test' dest
540

541
	CmmSwitch expr ids
542
	   -> do expr' <- cmmExprConFold DataReference expr
543
	         return $ CmmSwitch expr' ids
544 545

        other
546
           -> return other
547

548

549
cmmExprConFold referenceKind expr
550
   = case expr of
551
        CmmLoad addr rep
552
           -> do addr' <- cmmExprConFold DataReference addr
553
                 return $ CmmLoad addr' rep
554 555 556

        CmmMachOp mop args
           -- For MachOps, we first optimize the children, and then we try 
557
           -- our hand at some constant-folding.
558
           -> do args' <- mapM (cmmExprConFold DataReference) args
559 560 561
                 return $ cmmMachOpFold mop args'

        CmmLit (CmmLabel lbl)
562 563 564
           -> do
		dflags <- getDynFlagsCmmOpt
		cmmMakeDynamicReference dflags addImportCmmOpt referenceKind lbl
565
        CmmLit (CmmLabelOff lbl off)
566 567 568
           -> do
		 dflags <- getDynFlagsCmmOpt
		 dynRef <- cmmMakeDynamicReference dflags addImportCmmOpt referenceKind lbl
569 570 571 572
                 return $ cmmMachOpFold (MO_Add wordRep) [
                     dynRef,
                     (CmmLit $ CmmInt (fromIntegral off) wordRep)
                   ]
573 574

#if powerpc_TARGET_ARCH
575
           -- On powerpc (non-PIC), it's easier to jump directly to a label than
576 577 578
           -- to use the register table, so we replace these registers
           -- with the corresponding labels:
        CmmReg (CmmGlobal GCEnter1)
579
          | not opt_PIC
580
          -> cmmExprConFold referenceKind $
581
             CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_enter_1"))) 
582
        CmmReg (CmmGlobal GCFun)
583
          | not opt_PIC
584
          -> cmmExprConFold referenceKind $
585
             CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_fun")))
586 587 588 589 590 591 592 593 594 595
#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
596
                 Left  realreg -> return expr
597
                 Right baseRegAddr 
598
                    -> case mid of 
599 600 601
                          BaseReg -> cmmExprConFold DataReference baseRegAddr
                          other   -> cmmExprConFold DataReference
                                        (CmmLoad baseRegAddr (globalRegRep mid))
602 603
	   -- eliminate zero offsets
	CmmRegOff reg 0
604
	   -> cmmExprConFold referenceKind (CmmReg reg)
605 606 607 608 609 610

        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
611
                Left  realreg -> return expr
612
                Right baseRegAddr
613
                   -> cmmExprConFold DataReference (CmmMachOp (MO_Add wordRep) [
614 615 616
                                        CmmReg (CmmGlobal mid),
                                        CmmLit (CmmInt (fromIntegral offset)
                                                       wordRep)])
617
        other
618
           -> return other
619

620 621 622 623 624
-- -----------------------------------------------------------------------------
-- Utils

bind f x = x $! f

625 626
\end{code}