AsmCodeGen.lhs 20.2 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

import Cmm
25
import CmmOpt		( cmmMiniInline, cmmMachOpFold )
26
27
import PprCmm		( pprStmt, pprCmms )
import MachOp
28
import CLabel           ( CLabel, mkSplitMarkerLabel, mkAsmTempLabel )
29
30
31
#if powerpc_TARGET_ARCH
import CLabel           ( mkRtsCodeLabel )
#endif
32

33
34
35
36
import UniqFM
import Unique		( Unique, getUnique )
import UniqSupply
import FastTypes
37
38
import List		( groupBy, sortBy )
import CLabel           ( pprCLabel )
39
import ErrUtils		( dumpIfSet_dyn )
40
import DynFlags		( DynFlags, DynFlag(..), dopt )
41
import StaticFlags	( opt_Static, opt_PIC )
42
import Config           ( cProjectVersion )
43

44
import Digraph
45
import qualified Pretty
46
import Outputable
47
import FastString
48

49
50
-- DEBUGGING ONLY
--import OrdList
51

52
#ifdef NCG_DEBUG
53
import List		( intersperse )
54
#endif
55

56
57
58
59
import DATA_INT
import DATA_WORD
import DATA_BITS
import GLAEXTS
60

61
62
63
{-
The native-code generator has machine-independent and
machine-dependent modules.
64

65
66
67
68
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.
69

70
71
72
73
74
75
76
77
78
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.
79
80

The machine-dependent bits break down as follows:
81
82

  * ["MachRegs"]  Everything about the target platform's machine
83
84
85
    registers (and immediate operands, and addresses, which tend to
    intermingle/interact with registers).

86
  * ["MachInstrs"]  Includes the 'Instr' datatype (possibly should
87
    have a module of its own), plus a miscellany of other things
88
    (e.g., 'targetDoubleSize', 'smStablePtrTable', ...)
89

90
  * ["MachCodeGen"]  is where 'Cmm' stuff turns into
91
    machine instructions.
92

93
94
  * ["PprMach"] 'pprInstr' turns an 'Instr' into text (well, really
    a 'Doc').
95

96
97
  * ["RegAllocInfo"] In the register allocator, we manipulate
    'MRegsState's, which are 'BitSet's, one bit per machine register.
98
99
    When we want to say something about a specific machine register
    (e.g., ``it gets clobbered by this instruction''), we set/unset
100
    its bit.  Obviously, we do this 'BitSet' thing for efficiency
101
102
    reasons.

103
    The 'RegAllocInfo' module collects together the machine-specific
104
105
    info needed to do register allocation.

106
107
   * ["RegisterAlloc"] The (machine-independent) register allocator.
-}
108

109
110
-- -----------------------------------------------------------------------------
-- Top-level of the native codegen
111

112
-- NB. We *lazilly* compile each block of code for space reasons.
113

114
115
nativeCodeGen :: DynFlags -> [Cmm] -> UniqSupply -> IO Pretty.Doc
nativeCodeGen dflags cmms us
116
  = let (res, _) = initUs us $
117
	   cgCmm (concat (map add_split cmms))
118

119
	cgCmm :: [CmmTop] -> UniqSM (Cmm, Pretty.Doc, [CLabel])
120
121
	cgCmm tops = 
	   lazyMapUs (cmmNativeGen dflags) tops  `thenUs` \ results -> 
122
	   case unzip3 results of { (cmms,docs,imps) ->
123
	   returnUs (Cmm cmms, my_vcat docs, concat imps)
124
125
126
	   }
    in 
    case res of { (ppr_cmms, insn_sdoc, imports) -> do
127
    dumpIfSet_dyn dflags Opt_D_dump_opt_cmm "Optimised Cmm" (pprCmms [ppr_cmms])
128
129
130
131
132
133
    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"
134
135
136
137
138
139
140
141
142
#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.$$ Pretty.text ".section .note.GNU-stack,\"\",@progbits"
143
#endif
144
145
146
147
148
149
                -- And just because every other compiler does, lets stick in
		-- an identifier directive: .ident "GHC x.y.z"
	    Pretty.$$ let compilerIdent = Pretty.text "GHC" Pretty.<+>
	                                  Pretty.text cProjectVersion
                       in Pretty.text ".ident" Pretty.<+>
                          Pretty.doubleQuotes compilerIdent
150
            )
151
   }
152

153
  where
154

155
    add_split (Cmm tops)
156
157
	| dopt Opt_SplitObjs dflags = split_marker : tops
	| otherwise		    = tops
158

159
    split_marker = CmmProc [] mkSplitMarkerLabel [] []
160

161
162
     	 -- Generate "symbol stubs" for all external symbols that might
	 -- come from a dynamic library.
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
{-    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
182

183
184
185
186
187
188
189
190
191
192
193
194
#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
195
196


197
198
199
-- Complete native code generation phase for a single top-level chunk
-- of Cmm.

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

220
        returnUs (ppr_cmm, final_sdoc Pretty.$$ Pretty.text "", lastMinuteImports ++ imports)
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
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
     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?

310
genMachCode :: CmmTop -> UniqSM ([NatCmmTop], [CLabel])
311
312

genMachCode cmm_top initial_us
313
  = let initial_st             = mkNatM_State initial_us 0
314
315
316
317
        (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
318
319
    in
        if   final_delta == 0
320
        then ((new_tops, final_imports), final_us)
321
322
323
        else pprPanic "genMachCode: nonzero final delta"
                      (int final_delta)

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
-- -----------------------------------------------------------------------------
-- 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
354
  = returnUs [CmmAssign (CmmGlobal reg) src]
355
356
357
358
359
360
361
362
363
364
365
  | 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) ->
366
367
368
369
    returnUs (caller_save ++
	      CmmCall target results' args vols :
	      caller_restore ++
	      concat stores)
370
  where
371
372
373
	-- we also save/restore any caller-saves STG registers here
	(caller_save, caller_restore) = callerSaveVolatileRegs vols

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
	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) ).
399
400
401
402
  (d) Position independent code and dynamic linking
        (i)  introduce the appropriate indirections
             and position independent refs
        (ii) compile a list of imported symbols
403
404
405
406
407

Ideas for other things we could do (ToDo):

  - shortcut jumps-to-jumps
  - eliminate dead code blocks
408
409
410
  - 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...)
411
412
-}

413
414
415
cmmToCmm :: CmmTop -> (CmmTop, [CLabel])
cmmToCmm top@(CmmData _ _) = (top, [])
cmmToCmm (CmmProc info lbl params blocks) = runCmmOpt $ do
416
  blocks' <- mapM cmmBlockConFold (cmmMiniInline blocks)
417
  return $ CmmProc info lbl params blocks'
418

419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
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'
441
442
443
444

cmmStmtConFold stmt
   = case stmt of
        CmmAssign reg src
445
446
447
448
           -> do src' <- cmmExprConFold False src
                 return $ case src' of
		   CmmReg reg' | reg == reg' -> CmmNop
		   new_src -> CmmAssign reg new_src
449
450

        CmmStore addr src
451
452
453
           -> do addr' <- cmmExprConFold False addr
                 src'  <- cmmExprConFold False src
                 return $ CmmStore addr' src'
454
455

        CmmJump addr regs
456
457
           -> do addr' <- cmmExprConFold True addr
                 return $ CmmJump addr' regs
458
459

	CmmCall target regs args vols
460
461
462
463
464
465
466
467
468
	   -> 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
469
470

        CmmCondBranch test dest
471
472
473
474
           -> do test' <- cmmExprConFold False test
	         return $ case test' of
		   CmmLit (CmmInt 0 _) -> 
		     CmmComment (mkFastString ("deleted: " ++ 
475
					showSDoc (pprStmt stmt)))
476

477
478
		   CmmLit (CmmInt n _) -> CmmBranch dest
		   other -> CmmCondBranch test' dest
479

480
	CmmSwitch expr ids
481
482
	   -> do expr' <- cmmExprConFold False expr
	         return $ CmmSwitch expr' ids
483
484

        other
485
           -> return other
486

487

488
cmmExprConFold isJumpTarget expr
489
   = case expr of
490
        CmmLoad addr rep
491
492
           -> do addr' <- cmmExprConFold False addr
                 return $ CmmLoad addr' rep
493
494
495

        CmmMachOp mop args
           -- For MachOps, we first optimize the children, and then we try 
496
           -- our hand at some constant-folding.
497
498
499
500
501
502
503
504
505
506
507
           -> 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)
                   ]
508
509

#if powerpc_TARGET_ARCH
510
           -- On powerpc (non-PIC), it's easier to jump directly to a label than
511
512
513
           -- to use the register table, so we replace these registers
           -- with the corresponding labels:
        CmmReg (CmmGlobal GCEnter1)
514
515
516
          | not opt_PIC
          -> cmmExprConFold isJumpTarget $
             CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_enter_1"))) 
517
        CmmReg (CmmGlobal GCFun)
518
519
520
          | not opt_PIC
          -> cmmExprConFold isJumpTarget $
             CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_fun")))
521
522
523
524
525
526
527
528
529
530
#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
531
                 Left  realreg -> return expr
532
                 Right baseRegAddr 
533
                    -> case mid of 
534
535
                          BaseReg -> cmmExprConFold False baseRegAddr
                          other   -> cmmExprConFold False (CmmLoad baseRegAddr 
536
537
538
							(globalRegRep mid))
	   -- eliminate zero offsets
	CmmRegOff reg 0
539
	   -> cmmExprConFold False (CmmReg reg)
540
541
542
543
544
545

        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
546
                Left  realreg -> return expr
547
                Right baseRegAddr
548
                   -> cmmExprConFold False (CmmMachOp (MO_Add wordRep) [
549
550
551
                                        CmmReg (CmmGlobal mid),
                                        CmmLit (CmmInt (fromIntegral offset)
                                                       wordRep)])
552
        other
553
           -> return other
554

555
556
557
558
559
-- -----------------------------------------------------------------------------
-- Utils

bind f x = x $! f

560
561
\end{code}