1 -- -----------------------------------------------------------------------------
3 -- (c) The University of Glasgow 1993-2004
5 -- This is the top-level module in the native code generator.
7 -- -----------------------------------------------------------------------------
10 module AsmCodeGen ( nativeCodeGen ) where
12 #include "HsVersions.h"
13 #include "nativeGen/NCG.h"
21 import PositionIndependentCode
24 import qualified RegAllocLinear as Linear
25 import qualified RegAllocColor as Color
26 import qualified RegAllocStats as Color
27 import qualified GraphColor as Color
30 import CmmOpt ( cmmMiniInline, cmmMachOpFold )
31 import PprCmm ( pprStmt, pprCmms, pprCmm )
36 import Unique ( Unique, getUnique )
39 import List ( groupBy, sortBy )
40 import ErrUtils ( dumpIfSet_dyn )
42 import StaticFlags ( opt_Static, opt_PIC )
44 import Config ( cProjectVersion )
48 import qualified Pretty
65 The native-code generator has machine-independent and
66 machine-dependent modules.
68 This module ("AsmCodeGen") is the top-level machine-independent
69 module. Before entering machine-dependent land, we do some
70 machine-independent optimisations (defined below) on the
73 We convert to the machine-specific 'Instr' datatype with
74 'cmmCodeGen', assuming an infinite supply of registers. We then use
75 a machine-independent register allocator ('regAlloc') to rejoin
76 reality. Obviously, 'regAlloc' has machine-specific helper
77 functions (see about "RegAllocInfo" below).
79 Finally, we order the basic blocks of the function so as to minimise
80 the number of jumps between blocks, by utilising fallthrough wherever
83 The machine-dependent bits break down as follows:
85 * ["MachRegs"] Everything about the target platform's machine
86 registers (and immediate operands, and addresses, which tend to
87 intermingle/interact with registers).
89 * ["MachInstrs"] Includes the 'Instr' datatype (possibly should
90 have a module of its own), plus a miscellany of other things
91 (e.g., 'targetDoubleSize', 'smStablePtrTable', ...)
93 * ["MachCodeGen"] is where 'Cmm' stuff turns into
96 * ["PprMach"] 'pprInstr' turns an 'Instr' into text (well, really
99 * ["RegAllocInfo"] In the register allocator, we manipulate
100 'MRegsState's, which are 'BitSet's, one bit per machine register.
101 When we want to say something about a specific machine register
102 (e.g., ``it gets clobbered by this instruction''), we set/unset
103 its bit. Obviously, we do this 'BitSet' thing for efficiency
106 The 'RegAllocInfo' module collects together the machine-specific
107 info needed to do register allocation.
109 * ["RegisterAlloc"] The (machine-independent) register allocator.
112 -- -----------------------------------------------------------------------------
113 -- Top-level of the native codegen
115 -- NB. We *lazilly* compile each block of code for space reasons.
118 nativeCodeGen :: DynFlags -> Module -> ModLocation -> [RawCmm] -> UniqSupply -> IO Pretty.Doc
119 nativeCodeGen dflags mod modLocation cmms us
120 = let (res, _) = initUs us $
121 cgCmm (concat (map add_split cmms))
123 cgCmm :: [RawCmmTop] -> UniqSM ( [CmmNativeGenDump], Pretty.Doc, [CLabel])
125 lazyMapUs (cmmNativeGen dflags) tops `thenUs` \ results ->
126 case unzip3 results of { (dump,docs,imps) ->
127 returnUs (dump, my_vcat docs, concat imps)
130 case res of { (dump, insn_sdoc, imports) -> do
132 cmmNativeGenDump dflags mod modLocation dump
134 return (insn_sdoc Pretty.$$ dyld_stubs imports
136 #if HAVE_SUBSECTIONS_VIA_SYMBOLS
137 -- On recent versions of Darwin, the linker supports
138 -- dead-stripping of code and data on a per-symbol basis.
139 -- There's a hack to make this work in PprMach.pprNatCmmTop.
140 Pretty.$$ Pretty.text ".subsections_via_symbols"
142 #if HAVE_GNU_NONEXEC_STACK
143 -- On recent GNU ELF systems one can mark an object file
144 -- as not requiring an executable stack. If all objects
145 -- linked into a program have this note then the program
146 -- will not use an executable stack, which is good for
147 -- security. GHC generated code does not need an executable
148 -- stack so add the note in:
149 Pretty.$$ Pretty.text ".section .note.GNU-stack,\"\",@progbits"
151 #if !defined(darwin_TARGET_OS)
152 -- And just because every other compiler does, lets stick in
153 -- an identifier directive: .ident "GHC x.y.z"
154 Pretty.$$ let compilerIdent = Pretty.text "GHC" Pretty.<+>
155 Pretty.text cProjectVersion
156 in Pretty.text ".ident" Pretty.<+>
157 Pretty.doubleQuotes compilerIdent
165 | dopt Opt_SplitObjs dflags = split_marker : tops
168 split_marker = CmmProc [] mkSplitMarkerLabel [] []
170 -- Generate "symbol stubs" for all external symbols that might
171 -- come from a dynamic library.
172 {- dyld_stubs imps = Pretty.vcat $ map pprDyldSymbolStub $
173 map head $ group $ sort imps-}
175 -- (Hack) sometimes two Labels pretty-print the same, but have
176 -- different uniques; so we compare their text versions...
178 | needImportedSymbols
180 (pprGotDeclaration :) $
181 map (pprImportedSymbol . fst . head) $
182 groupBy (\(_,a) (_,b) -> a == b) $
183 sortBy (\(_,a) (_,b) -> compare a b) $
189 where doPpr lbl = (lbl, Pretty.render $ pprCLabel lbl astyle)
190 astyle = mkCodeStyle AsmStyle
193 my_vcat sds = Pretty.vcat sds
195 my_vcat sds = Pretty.vcat (
198 Pretty.$$ Pretty.ptext SLIT("# ___ncg_debug_marker")
199 Pretty.$$ Pretty.char ' '
206 -- Carries output of the code generator passes, for dumping.
207 -- Make sure to only fill the one's we're interested in to avoid
208 -- creating space leaks.
210 data CmmNativeGenDump
212 { cdCmmOpt :: RawCmmTop
213 , cdNative :: [NatCmmTop]
214 , cdLiveness :: [LiveCmmTop]
215 , cdCoalesce :: Maybe [LiveCmmTop]
216 , cdRegAllocStats :: Maybe [Color.RegAllocStats]
217 , cdRegAllocStatsLinear :: [Linear.RegAllocStats]
218 , cdColoredGraph :: Maybe (Color.Graph Reg RegClass Reg)
219 , cdAlloced :: [NatCmmTop] }
221 dchoose dflags opt a b
222 | dopt opt dflags = a
225 dchooses dflags opts a b
226 | or $ map ( (flip dopt) dflags) opts = a
229 -- | Complete native code generation phase for a single top-level chunk of Cmm.
230 -- Unless they're being dumped, intermediate data structures are squashed after
231 -- every stage to avoid creating space leaks.
233 -- TODO: passing data via CmmNativeDump/squashing structs has become a horrible mess.
234 -- it might be better to forgo trying to keep all the outputs for each
235 -- stage together and just thread IO() through cmmNativeGen so we can dump
236 -- what we want to after each stage.
238 cmmNativeGen :: DynFlags -> RawCmmTop -> UniqSM (CmmNativeGenDump, Pretty.Doc, [CLabel])
239 cmmNativeGen dflags cmm
243 <- {-# SCC "fixAssigns" #-}
246 ---- cmm to cmm optimisations
247 (cmm, imports, ppr_cmm)
249 -> {-# SCC "genericOpt" #-}
250 do let (cmm, imports) = cmmToCmm dflags fixed_cmm
254 , dchoose dflags Opt_D_dump_cmm cmm (CmmData Text []))
258 ---- generate native code from cmm
259 (native, lastMinuteImports, ppr_native)
261 -> {-# SCC "genMachCode" #-}
262 do (machCode, lastMinuteImports)
263 <- genMachCode dflags cmm
267 , dchoose dflags Opt_D_dump_asm_native machCode [])
271 ---- tag instructions with register liveness information
272 (withLiveness, ppr_withLiveness)
274 -> {-# SCC "regLiveness" #-}
276 withLiveness <- mapUs regLiveness native
278 return ( withLiveness
279 , dchoose dflags Opt_D_dump_asm_liveness withLiveness []))
282 ---- allocate registers
283 ( alloced, ppr_alloced, ppr_coalesce
284 , ppr_regAllocStats, ppr_regAllocStatsLinear, ppr_coloredGraph)
286 -> {-# SCC "regAlloc" #-}
288 if dopt Opt_RegsGraph dflags
290 -- the regs usable for allocation
292 = foldr (\r -> plusUFM_C unionUniqSets
293 $ unitUFM (regClass r) (unitUniqSet r))
295 $ map RealReg allocatableRegs
297 -- aggressively coalesce moves between virtual regs
298 coalesced <- regCoalesce withLiveness
300 -- graph coloring register allocation
301 (alloced, regAllocStats)
304 (mkUniqSet [0..maxSpillSlots])
308 , dchoose dflags Opt_D_dump_asm_regalloc
310 , dchoose dflags Opt_D_dump_asm_coalesce
311 (Just coalesced) Nothing
313 [ Opt_D_dump_asm_regalloc_stages
314 , Opt_D_drop_asm_stats]
315 (Just regAllocStats) Nothing
317 , dchoose dflags Opt_D_dump_asm_conflicts
321 -- do linear register allocation
324 $ mapUs Linear.regAlloc withLiveness
327 , dchoose dflags Opt_D_dump_asm_regalloc
331 , dchoose dflags Opt_D_drop_asm_stats
337 ---- shortcut branches
339 {-# SCC "shortcutBranches" #-}
340 shortcutBranches dflags alloced
344 {-# SCC "sequenceBlocks" #-}
345 map sequenceTop shorted
348 let final_mach_code =
350 {-# SCC "x86fp_kludge" #-}
351 map x86fp_kludge sequenced
359 Pretty.vcat (map pprNatCmmTop final_mach_code)
364 , cdNative = ppr_native
365 , cdLiveness = ppr_withLiveness
366 , cdCoalesce = ppr_coalesce
367 , cdRegAllocStats = ppr_regAllocStats
368 , cdRegAllocStatsLinear = ppr_regAllocStatsLinear
369 , cdColoredGraph = ppr_coloredGraph
370 , cdAlloced = ppr_alloced }
372 returnUs (dump, final_sdoc Pretty.$$ Pretty.text "", lastMinuteImports ++ imports)
375 x86fp_kludge :: NatCmmTop -> NatCmmTop
376 x86fp_kludge top@(CmmData _ _) = top
377 x86fp_kludge top@(CmmProc info lbl params code) =
378 CmmProc info lbl params (map bb_i386_insert_ffrees code)
380 bb_i386_insert_ffrees (BasicBlock id instrs) =
381 BasicBlock id (i386_insert_ffrees instrs)
385 -- Dump output of native code generator passes
386 -- stripe across the outputs for each block so all the information for a
387 -- certain stage is concurrent in the dumps.
389 cmmNativeGenDump :: DynFlags -> Module -> ModLocation -> [CmmNativeGenDump] -> IO ()
390 cmmNativeGenDump dflags mod modLocation dump
393 Opt_D_dump_opt_cmm "Optimised Cmm"
394 (pprCmm $ Cmm $ map cdCmmOpt dump)
397 Opt_D_dump_asm_native "Native code"
398 (vcat $ map (docToSDoc . pprNatCmmTop) $ concatMap cdNative dump)
401 Opt_D_dump_asm_liveness "Liveness annotations added"
402 (vcat $ map (ppr . cdLiveness) dump)
405 Opt_D_dump_asm_coalesce "Reg-Reg moves coalesced"
406 (vcat $ map (fromMaybe empty . liftM ppr . cdCoalesce) dump)
409 Opt_D_dump_asm_regalloc "Registers allocated"
410 (vcat $ map (docToSDoc . pprNatCmmTop) $ concatMap cdAlloced dump)
412 -- with the graph coloring allocator, show the result of each build/spill stage
413 -- for each block in turn.
414 when (dopt Opt_D_dump_asm_regalloc_stages dflags)
417 $ vcat $ map (\(stage, stats) ->
418 text "-- Stage " <> int stage
421 $ map (fromMaybe [] . cdRegAllocStats) dump
423 -- Build a global register conflict graph.
424 -- If you want to see the graph for just one basic block then use asm-regalloc-stages instead.
426 Opt_D_dump_asm_conflicts "Register conflict graph"
427 $ Color.dotGraph Color.regDotColor trivColorable
428 $ foldl Color.union Color.initGraph
429 $ catMaybes $ map cdColoredGraph dump
431 -- Drop native code generator statistics.
432 -- This is potentially a large amount of information, and we want to be able
433 -- to collect it while running nofib. Drop a new file instead of emitting
434 -- it to stdout/stderr.
436 when (dopt Opt_D_drop_asm_stats dflags)
437 $ do -- make the drop file name based on the object file name
438 let dropFile = (init $ ml_obj_file modLocation) ++ "drop-asm-stats"
440 -- slurp out all the regalloc stats
441 let stats = concat $ catMaybes $ map cdRegAllocStats dump
443 -- build a global conflict graph
444 let graph = foldl Color.union Color.initGraph $ map Color.raGraph stats
446 -- pretty print the various sections and write out the file.
447 let outSpills = Color.pprStatsSpills stats
448 let outLife = Color.pprStatsLifetimes stats
449 let outConflict = Color.pprStatsConflict stats
450 let outScatter = Color.pprStatsLifeConflict stats graph
453 (showSDoc $ vcat [outSpills, outLife, outConflict, outScatter])
457 -- -----------------------------------------------------------------------------
458 -- Sequencing the basic blocks
460 -- Cmm BasicBlocks are self-contained entities: they always end in a
461 -- jump, either non-local or to another basic block in the same proc.
462 -- In this phase, we attempt to place the basic blocks in a sequence
463 -- such that as many of the local jumps as possible turn into
466 sequenceTop :: NatCmmTop -> NatCmmTop
467 sequenceTop top@(CmmData _ _) = top
468 sequenceTop (CmmProc info lbl params blocks) =
469 CmmProc info lbl params (makeFarBranches $ sequenceBlocks blocks)
471 -- The algorithm is very simple (and stupid): we make a graph out of
472 -- the blocks where there is an edge from one block to another iff the
473 -- first block ends by jumping to the second. Then we topologically
474 -- sort this graph. Then traverse the list: for each block, we first
475 -- output the block, then if it has an out edge, we move the
476 -- destination of the out edge to the front of the list, and continue.
478 sequenceBlocks :: [NatBasicBlock] -> [NatBasicBlock]
479 sequenceBlocks [] = []
480 sequenceBlocks (entry:blocks) =
481 seqBlocks (mkNode entry : reverse (flattenSCCs (sccBlocks blocks)))
482 -- the first block is the entry point ==> it must remain at the start.
484 sccBlocks :: [NatBasicBlock] -> [SCC (NatBasicBlock,Unique,[Unique])]
485 sccBlocks blocks = stronglyConnCompR (map mkNode blocks)
487 getOutEdges :: [Instr] -> [Unique]
488 getOutEdges instrs = case jumpDests (last instrs) [] of
489 [one] -> [getUnique one]
491 -- we're only interested in the last instruction of
492 -- the block, and only if it has a single destination.
494 mkNode block@(BasicBlock id instrs) = (block, getUnique id, getOutEdges instrs)
497 seqBlocks ((block,_,[]) : rest)
498 = block : seqBlocks rest
499 seqBlocks ((block@(BasicBlock id instrs),_,[next]) : rest)
500 | can_fallthrough = BasicBlock id (init instrs) : seqBlocks rest'
501 | otherwise = block : seqBlocks rest'
503 (can_fallthrough, rest') = reorder next [] rest
504 -- TODO: we should do a better job for cycles; try to maximise the
505 -- fallthroughs within a loop.
506 seqBlocks _ = panic "AsmCodegen:seqBlocks"
508 reorder id accum [] = (False, reverse accum)
509 reorder id accum (b@(block,id',out) : rest)
510 | id == id' = (True, (block,id,out) : reverse accum ++ rest)
511 | otherwise = reorder id (b:accum) rest
514 -- -----------------------------------------------------------------------------
515 -- Making far branches
517 -- Conditional branches on PowerPC are limited to +-32KB; if our Procs get too
518 -- big, we have to work around this limitation.
520 makeFarBranches :: [NatBasicBlock] -> [NatBasicBlock]
522 #if powerpc_TARGET_ARCH
523 makeFarBranches blocks
524 | last blockAddresses < nearLimit = blocks
525 | otherwise = zipWith handleBlock blockAddresses blocks
527 blockAddresses = scanl (+) 0 $ map blockLen blocks
528 blockLen (BasicBlock _ instrs) = length instrs
530 handleBlock addr (BasicBlock id instrs)
531 = BasicBlock id (zipWith makeFar [addr..] instrs)
533 makeFar addr (BCC ALWAYS tgt) = BCC ALWAYS tgt
534 makeFar addr (BCC cond tgt)
535 | abs (addr - targetAddr) >= nearLimit
539 where Just targetAddr = lookupUFM blockAddressMap tgt
540 makeFar addr other = other
542 nearLimit = 7000 -- 8192 instructions are allowed; let's keep some
543 -- distance, as we have a few pseudo-insns that are
544 -- pretty-printed as multiple instructions,
545 -- and it's just not worth the effort to calculate
548 blockAddressMap = listToUFM $ zip (map blockId blocks) blockAddresses
553 -- -----------------------------------------------------------------------------
556 shortcutBranches :: DynFlags -> [NatCmmTop] -> [NatCmmTop]
557 shortcutBranches dflags tops
558 | optLevel dflags < 1 = tops -- only with -O or higher
559 | otherwise = map (apply_mapping mapping) tops'
561 (tops', mappings) = mapAndUnzip build_mapping tops
562 mapping = foldr plusUFM emptyUFM mappings
564 build_mapping top@(CmmData _ _) = (top, emptyUFM)
565 build_mapping (CmmProc info lbl params [])
566 = (CmmProc info lbl params [], emptyUFM)
567 build_mapping (CmmProc info lbl params (head:blocks))
568 = (CmmProc info lbl params (head:others), mapping)
569 -- drop the shorted blocks, but don't ever drop the first one,
570 -- because it is pointed to by a global label.
572 -- find all the blocks that just consist of a jump that can be
574 (shortcut_blocks, others) = partitionWith split blocks
575 split (BasicBlock id [insn]) | Just dest <- canShortcut insn
577 split other = Right other
579 -- build a mapping from BlockId to JumpDest for shorting branches
580 mapping = foldl add emptyUFM shortcut_blocks
581 add ufm (id,dest) = addToUFM ufm id dest
583 apply_mapping ufm (CmmData sec statics)
584 = CmmData sec (map (shortcutStatic (lookupUFM ufm)) statics)
585 -- we need to get the jump tables, so apply the mapping to the entries
587 apply_mapping ufm (CmmProc info lbl params blocks)
588 = CmmProc info lbl params (map short_bb blocks)
590 short_bb (BasicBlock id insns) = BasicBlock id $! map short_insn insns
591 short_insn i = shortcutJump (lookupUFM ufm) i
592 -- shortcutJump should apply the mapping repeatedly,
593 -- just in case we can short multiple branches.
595 -- -----------------------------------------------------------------------------
596 -- Instruction selection
598 -- Native code instruction selection for a chunk of stix code. For
599 -- this part of the computation, we switch from the UniqSM monad to
600 -- the NatM monad. The latter carries not only a Unique, but also an
601 -- Int denoting the current C stack pointer offset in the generated
602 -- code; this is needed for creating correct spill offsets on
603 -- architectures which don't offer, or for which it would be
604 -- prohibitively expensive to employ, a frame pointer register. Viz,
607 -- The offset is measured in bytes, and indicates the difference
608 -- between the current (simulated) C stack-ptr and the value it was at
609 -- the beginning of the block. For stacks which grow down, this value
610 -- should be either zero or negative.
612 -- Switching between the two monads whilst carrying along the same
613 -- Unique supply breaks abstraction. Is that bad?
615 genMachCode :: DynFlags -> RawCmmTop -> UniqSM ([NatCmmTop], [CLabel])
617 genMachCode dflags cmm_top
618 = do { initial_us <- getUs
619 ; let initial_st = mkNatM_State initial_us 0 dflags
620 (new_tops, final_st) = initNat initial_st (cmmTopCodeGen cmm_top)
621 final_delta = natm_delta final_st
622 final_imports = natm_imports final_st
623 ; if final_delta == 0
624 then return (new_tops, final_imports)
625 else pprPanic "genMachCode: nonzero final delta" (int final_delta)
628 -- -----------------------------------------------------------------------------
629 -- Fixup assignments to global registers so that they assign to
630 -- locations within the RegTable, if appropriate.
632 -- Note that we currently don't fixup reads here: they're done by
633 -- the generic optimiser below, to avoid having two separate passes
636 fixAssignsTop :: RawCmmTop -> UniqSM RawCmmTop
637 fixAssignsTop top@(CmmData _ _) = returnUs top
638 fixAssignsTop (CmmProc info lbl params blocks) =
639 mapUs fixAssignsBlock blocks `thenUs` \ blocks' ->
640 returnUs (CmmProc info lbl params blocks')
642 fixAssignsBlock :: CmmBasicBlock -> UniqSM CmmBasicBlock
643 fixAssignsBlock (BasicBlock id stmts) =
644 fixAssigns stmts `thenUs` \ stmts' ->
645 returnUs (BasicBlock id stmts')
647 fixAssigns :: [CmmStmt] -> UniqSM [CmmStmt]
649 mapUs fixAssign stmts `thenUs` \ stmtss ->
650 returnUs (concat stmtss)
652 fixAssign :: CmmStmt -> UniqSM [CmmStmt]
653 fixAssign (CmmAssign (CmmGlobal reg) src)
654 | Left realreg <- reg_or_addr
655 = returnUs [CmmAssign (CmmGlobal reg) src]
656 | Right baseRegAddr <- reg_or_addr
657 = returnUs [CmmStore baseRegAddr src]
658 -- Replace register leaves with appropriate StixTrees for
659 -- the given target. GlobalRegs which map to a reg on this
660 -- arch are left unchanged. Assigning to BaseReg is always
661 -- illegal, so we check for that.
663 reg_or_addr = get_GlobalReg_reg_or_addr reg
665 fixAssign other_stmt = returnUs [other_stmt]
667 -- -----------------------------------------------------------------------------
668 -- Generic Cmm optimiser
674 (b) Simple inlining: a temporary which is assigned to and then
675 used, once, can be shorted.
676 (c) Replacement of references to GlobalRegs which do not have
677 machine registers by the appropriate memory load (eg.
678 Hp ==> *(BaseReg + 34) ).
679 (d) Position independent code and dynamic linking
680 (i) introduce the appropriate indirections
681 and position independent refs
682 (ii) compile a list of imported symbols
684 Ideas for other things we could do (ToDo):
686 - shortcut jumps-to-jumps
687 - eliminate dead code blocks
688 - simple CSE: if an expr is assigned to a temp, then replace later occs of
689 that expr with the temp, until the expr is no longer valid (can push through
690 temp assignments, and certain assigns to mem...)
693 cmmToCmm :: DynFlags -> RawCmmTop -> (RawCmmTop, [CLabel])
694 cmmToCmm _ top@(CmmData _ _) = (top, [])
695 cmmToCmm dflags (CmmProc info lbl params blocks) = runCmmOpt dflags $ do
696 blocks' <- mapM cmmBlockConFold (cmmMiniInline blocks)
697 return $ CmmProc info lbl params blocks'
699 newtype CmmOptM a = CmmOptM (([CLabel], DynFlags) -> (# a, [CLabel] #))
701 instance Monad CmmOptM where
702 return x = CmmOptM $ \(imports, _) -> (# x,imports #)
704 CmmOptM $ \(imports, dflags) ->
705 case f (imports, dflags) of
708 CmmOptM g' -> g' (imports', dflags)
710 addImportCmmOpt :: CLabel -> CmmOptM ()
711 addImportCmmOpt lbl = CmmOptM $ \(imports, dflags) -> (# (), lbl:imports #)
713 getDynFlagsCmmOpt :: CmmOptM DynFlags
714 getDynFlagsCmmOpt = CmmOptM $ \(imports, dflags) -> (# dflags, imports #)
716 runCmmOpt :: DynFlags -> CmmOptM a -> (a, [CLabel])
717 runCmmOpt dflags (CmmOptM f) = case f ([], dflags) of
718 (# result, imports #) -> (result, imports)
720 cmmBlockConFold :: CmmBasicBlock -> CmmOptM CmmBasicBlock
721 cmmBlockConFold (BasicBlock id stmts) = do
722 stmts' <- mapM cmmStmtConFold stmts
723 return $ BasicBlock id stmts'
728 -> do src' <- cmmExprConFold DataReference src
729 return $ case src' of
730 CmmReg reg' | reg == reg' -> CmmNop
731 new_src -> CmmAssign reg new_src
734 -> do addr' <- cmmExprConFold DataReference addr
735 src' <- cmmExprConFold DataReference src
736 return $ CmmStore addr' src'
739 -> do addr' <- cmmExprConFold JumpReference addr
740 return $ CmmJump addr' regs
742 CmmCall target regs args srt returns
743 -> do target' <- case target of
744 CmmCallee e conv -> do
745 e' <- cmmExprConFold CallReference e
746 return $ CmmCallee e' conv
747 other -> return other
748 args' <- mapM (\(arg, hint) -> do
749 arg' <- cmmExprConFold DataReference arg
750 return (arg', hint)) args
751 return $ CmmCall target' regs args' srt returns
753 CmmCondBranch test dest
754 -> do test' <- cmmExprConFold DataReference test
755 return $ case test' of
756 CmmLit (CmmInt 0 _) ->
757 CmmComment (mkFastString ("deleted: " ++
758 showSDoc (pprStmt stmt)))
760 CmmLit (CmmInt n _) -> CmmBranch dest
761 other -> CmmCondBranch test' dest
764 -> do expr' <- cmmExprConFold DataReference expr
765 return $ CmmSwitch expr' ids
771 cmmExprConFold referenceKind expr
774 -> do addr' <- cmmExprConFold DataReference addr
775 return $ CmmLoad addr' rep
778 -- For MachOps, we first optimize the children, and then we try
779 -- our hand at some constant-folding.
780 -> do args' <- mapM (cmmExprConFold DataReference) args
781 return $ cmmMachOpFold mop args'
783 CmmLit (CmmLabel lbl)
785 dflags <- getDynFlagsCmmOpt
786 cmmMakeDynamicReference dflags addImportCmmOpt referenceKind lbl
787 CmmLit (CmmLabelOff lbl off)
789 dflags <- getDynFlagsCmmOpt
790 dynRef <- cmmMakeDynamicReference dflags addImportCmmOpt referenceKind lbl
791 return $ cmmMachOpFold (MO_Add wordRep) [
793 (CmmLit $ CmmInt (fromIntegral off) wordRep)
796 #if powerpc_TARGET_ARCH
797 -- On powerpc (non-PIC), it's easier to jump directly to a label than
798 -- to use the register table, so we replace these registers
799 -- with the corresponding labels:
800 CmmReg (CmmGlobal GCEnter1)
802 -> cmmExprConFold referenceKind $
803 CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_enter_1")))
804 CmmReg (CmmGlobal GCFun)
806 -> cmmExprConFold referenceKind $
807 CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_fun")))
810 CmmReg (CmmGlobal mid)
811 -- Replace register leaves with appropriate StixTrees for
812 -- the given target. MagicIds which map to a reg on this
813 -- arch are left unchanged. For the rest, BaseReg is taken
814 -- to mean the address of the reg table in MainCapability,
815 -- and for all others we generate an indirection to its
816 -- location in the register table.
817 -> case get_GlobalReg_reg_or_addr mid of
818 Left realreg -> return expr
821 BaseReg -> cmmExprConFold DataReference baseRegAddr
822 other -> cmmExprConFold DataReference
823 (CmmLoad baseRegAddr (globalRegRep mid))
824 -- eliminate zero offsets
826 -> cmmExprConFold referenceKind (CmmReg reg)
828 CmmRegOff (CmmGlobal mid) offset
829 -- RegOf leaves are just a shorthand form. If the reg maps
830 -- to a real reg, we keep the shorthand, otherwise, we just
831 -- expand it and defer to the above code.
832 -> case get_GlobalReg_reg_or_addr mid of
833 Left realreg -> return expr
835 -> cmmExprConFold DataReference (CmmMachOp (MO_Add wordRep) [
836 CmmReg (CmmGlobal mid),
837 CmmLit (CmmInt (fromIntegral offset)
842 -- -----------------------------------------------------------------------------