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"
22 #elif i386_TARGET_ARCH || x86_64_TARGET_ARCH
28 #elif sparc_TARGET_ARCH
30 import SPARC.CodeGen.Expand
34 import SPARC.ShortcutJump
36 #elif powerpc_TARGET_ARCH
45 #error "AsmCodeGen: unknown architecture"
49 import RegAlloc.Liveness
50 import qualified RegAlloc.Linear.Main as Linear
52 import qualified GraphColor as Color
53 import qualified RegAlloc.Graph.Main as Color
54 import qualified RegAlloc.Graph.Stats as Color
55 import qualified RegAlloc.Graph.TrivColorable as Color
65 import CgUtils ( fixStgRegisters )
67 import CmmOpt ( cmmMiniInline, cmmMachOpFold )
72 import Unique ( Unique, getUnique )
80 import qualified Pretty
95 import Distribution.System
98 The native-code generator has machine-independent and
99 machine-dependent modules.
101 This module ("AsmCodeGen") is the top-level machine-independent
102 module. Before entering machine-dependent land, we do some
103 machine-independent optimisations (defined below) on the
106 We convert to the machine-specific 'Instr' datatype with
107 'cmmCodeGen', assuming an infinite supply of registers. We then use
108 a machine-independent register allocator ('regAlloc') to rejoin
109 reality. Obviously, 'regAlloc' has machine-specific helper
110 functions (see about "RegAllocInfo" below).
112 Finally, we order the basic blocks of the function so as to minimise
113 the number of jumps between blocks, by utilising fallthrough wherever
116 The machine-dependent bits break down as follows:
118 * ["MachRegs"] Everything about the target platform's machine
119 registers (and immediate operands, and addresses, which tend to
120 intermingle/interact with registers).
122 * ["MachInstrs"] Includes the 'Instr' datatype (possibly should
123 have a module of its own), plus a miscellany of other things
124 (e.g., 'targetDoubleSize', 'smStablePtrTable', ...)
126 * ["MachCodeGen"] is where 'Cmm' stuff turns into
127 machine instructions.
129 * ["PprMach"] 'pprInstr' turns an 'Instr' into text (well, really
132 * ["RegAllocInfo"] In the register allocator, we manipulate
133 'MRegsState's, which are 'BitSet's, one bit per machine register.
134 When we want to say something about a specific machine register
135 (e.g., ``it gets clobbered by this instruction''), we set/unset
136 its bit. Obviously, we do this 'BitSet' thing for efficiency
139 The 'RegAllocInfo' module collects together the machine-specific
140 info needed to do register allocation.
142 * ["RegisterAlloc"] The (machine-independent) register allocator.
145 -- -----------------------------------------------------------------------------
146 -- Top-level of the native codegen
149 nativeCodeGen :: DynFlags -> Handle -> UniqSupply -> [RawCmm] -> IO ()
150 nativeCodeGen dflags h us cmms
152 let split_cmms = concat $ map add_split cmms
154 -- BufHandle is a performance hack. We could hide it inside
155 -- Pretty if it weren't for the fact that we do lots of little
156 -- printDocs here (in order to do codegen in constant space).
157 bufh <- newBufHandle h
158 (imports, prof) <- cmmNativeGens dflags bufh us split_cmms [] [] 0
161 let (native, colorStats, linearStats)
166 Opt_D_dump_asm "Asm code"
167 (vcat $ map (docToSDoc . pprNatCmmTop) $ concat native)
169 -- dump global NCG stats for graph coloring allocator
170 (case concat $ catMaybes colorStats of
173 -- build the global register conflict graph
175 = foldl Color.union Color.initGraph
176 $ [ Color.raGraph stat
177 | stat@Color.RegAllocStatsStart{} <- stats]
179 dumpSDoc dflags Opt_D_dump_asm_stats "NCG stats"
180 $ Color.pprStats stats graphGlobal
183 Opt_D_dump_asm_conflicts "Register conflict graph"
187 targetVirtualRegSqueeze
188 targetRealRegSqueeze)
192 -- dump global NCG stats for linear allocator
193 (case concat $ catMaybes linearStats of
195 stats -> dumpSDoc dflags Opt_D_dump_asm_stats "NCG stats"
196 $ Linear.pprStats (concat native) stats)
198 -- write out the imports
199 Pretty.printDoc Pretty.LeftMode h
200 $ makeImportsDoc dflags (concat imports)
204 where add_split (Cmm tops)
205 | dopt Opt_SplitObjs dflags = split_marker : tops
208 split_marker = CmmProc [] mkSplitMarkerLabel (ListGraph [])
211 -- | Do native code generation on all these cmms.
213 cmmNativeGens :: DynFlags
218 -> [ ([NatCmmTop Instr],
219 Maybe [Color.RegAllocStats Instr],
220 Maybe [Linear.RegAllocStats]) ]
224 Maybe [Color.RegAllocStats Instr],
225 Maybe [Linear.RegAllocStats])] )
227 cmmNativeGens _ _ _ [] impAcc profAcc _
228 = return (reverse impAcc, reverse profAcc)
230 cmmNativeGens dflags h us (cmm : cmms) impAcc profAcc count
232 (us', native, imports, colorStats, linearStats)
233 <- cmmNativeGen dflags us cmm count
235 Pretty.bufLeftRender h
236 $ {-# SCC "pprNativeCode" #-} Pretty.vcat $ map pprNatCmmTop native
238 -- carefully evaluate this strictly. Binding it with 'let'
239 -- and then using 'seq' doesn't work, because the let
240 -- apparently gets inlined first.
241 lsPprNative <- return $!
242 if dopt Opt_D_dump_asm dflags
243 || dopt Opt_D_dump_asm_stats dflags
247 count' <- return $! count + 1;
249 -- force evaulation all this stuff to avoid space leaks
250 seqString (showSDoc $ vcat $ map ppr imports) `seq` return ()
252 cmmNativeGens dflags h us' cmms
254 ((lsPprNative, colorStats, linearStats) : profAcc)
257 where seqString [] = ()
258 seqString (x:xs) = x `seq` seqString xs `seq` ()
261 -- | Complete native code generation phase for a single top-level chunk of Cmm.
262 -- Dumping the output of each stage along the way.
263 -- Global conflict graph and NGC stats
267 -> RawCmmTop -- ^ the cmm to generate code for
268 -> Int -- ^ sequence number of this top thing
270 , [NatCmmTop Instr] -- native code
271 , [CLabel] -- things imported by this cmm
272 , Maybe [Color.RegAllocStats Instr] -- stats for the coloring register allocator
273 , Maybe [Linear.RegAllocStats]) -- stats for the linear register allocators
275 cmmNativeGen dflags us cmm count
278 -- rewrite assignments to global regs
280 {-# SCC "fixStgRegisters" #-}
283 -- cmm to cmm optimisations
284 let (opt_cmm, imports) =
285 {-# SCC "cmmToCmm" #-}
286 cmmToCmm dflags fixed_cmm
289 Opt_D_dump_opt_cmm "Optimised Cmm"
290 (pprCmm $ Cmm [opt_cmm])
292 -- generate native code from cmm
293 let ((native, lastMinuteImports), usGen) =
294 {-# SCC "genMachCode" #-}
295 initUs us $ genMachCode dflags opt_cmm
298 Opt_D_dump_asm_native "Native code"
299 (vcat $ map (docToSDoc . pprNatCmmTop) native)
301 -- tag instructions with register liveness information
302 let (withLiveness, usLive) =
303 {-# SCC "regLiveness" #-}
306 $ map natCmmTopToLive native
309 Opt_D_dump_asm_liveness "Liveness annotations added"
310 (vcat $ map ppr withLiveness)
312 -- allocate registers
313 (alloced, usAlloc, ppr_raStatsColor, ppr_raStatsLinear) <-
314 if ( dopt Opt_RegsGraph dflags
315 || dopt Opt_RegsIterative dflags)
317 -- the regs usable for allocation
318 let (alloc_regs :: UniqFM (UniqSet RealReg))
319 = foldr (\r -> plusUFM_C unionUniqSets
320 $ unitUFM (targetClassOfRealReg r) (unitUniqSet r))
324 -- do the graph coloring register allocation
325 let ((alloced, regAllocStats), usAlloc)
326 = {-# SCC "RegAlloc" #-}
331 (mkUniqSet [0..maxSpillSlots])
334 -- dump out what happened during register allocation
336 Opt_D_dump_asm_regalloc "Registers allocated"
337 (vcat $ map (docToSDoc . pprNatCmmTop) alloced)
340 Opt_D_dump_asm_regalloc_stages "Build/spill stages"
341 (vcat $ map (\(stage, stats)
342 -> text "# --------------------------"
343 $$ text "# cmm " <> int count <> text " Stage " <> int stage
345 $ zip [0..] regAllocStats)
348 if dopt Opt_D_dump_asm_stats dflags
349 then Just regAllocStats else Nothing
351 -- force evaluation of the Maybe to avoid space leak
352 mPprStats `seq` return ()
354 return ( alloced, usAlloc
359 -- do linear register allocation
360 let ((alloced, regAllocStats), usAlloc)
361 = {-# SCC "RegAlloc" #-}
364 $ mapUs Linear.regAlloc withLiveness
367 Opt_D_dump_asm_regalloc "Registers allocated"
368 (vcat $ map (docToSDoc . pprNatCmmTop) alloced)
371 if dopt Opt_D_dump_asm_stats dflags
372 then Just (catMaybes regAllocStats) else Nothing
374 -- force evaluation of the Maybe to avoid space leak
375 mPprStats `seq` return ()
377 return ( alloced, usAlloc
381 ---- shortcut branches
383 {-# SCC "shortcutBranches" #-}
384 shortcutBranches dflags alloced
388 {-# SCC "sequenceBlocks" #-}
389 map sequenceTop shorted
394 {-# SCC "x86fp_kludge" #-}
395 map x86fp_kludge sequenced
400 ---- expansion of SPARC synthetic instrs
401 #if sparc_TARGET_ARCH
403 {-# SCC "sparc_expand" #-}
404 map expandTop kludged
407 Opt_D_dump_asm_expanded "Synthetic instructions expanded"
408 (vcat $ map (docToSDoc . pprNatCmmTop) expanded)
416 , lastMinuteImports ++ imports
422 x86fp_kludge :: NatCmmTop Instr -> NatCmmTop Instr
423 x86fp_kludge top@(CmmData _ _) = top
424 x86fp_kludge (CmmProc info lbl (ListGraph code)) =
425 CmmProc info lbl (ListGraph $ i386_insert_ffrees code)
429 -- | Build a doc for all the imports.
431 makeImportsDoc :: DynFlags -> [CLabel] -> Pretty.Doc
432 makeImportsDoc dflags imports
435 #if HAVE_SUBSECTIONS_VIA_SYMBOLS
436 -- On recent versions of Darwin, the linker supports
437 -- dead-stripping of code and data on a per-symbol basis.
438 -- There's a hack to make this work in PprMach.pprNatCmmTop.
439 Pretty.$$ Pretty.text ".subsections_via_symbols"
441 #if HAVE_GNU_NONEXEC_STACK
442 -- On recent GNU ELF systems one can mark an object file
443 -- as not requiring an executable stack. If all objects
444 -- linked into a program have this note then the program
445 -- will not use an executable stack, which is good for
446 -- security. GHC generated code does not need an executable
447 -- stack so add the note in:
448 Pretty.$$ Pretty.text ".section .note.GNU-stack,\"\",@progbits"
450 #if !defined(darwin_TARGET_OS)
451 -- And just because every other compiler does, lets stick in
452 -- an identifier directive: .ident "GHC x.y.z"
453 Pretty.$$ let compilerIdent = Pretty.text "GHC" Pretty.<+>
454 Pretty.text cProjectVersion
455 in Pretty.text ".ident" Pretty.<+>
456 Pretty.doubleQuotes compilerIdent
460 -- Generate "symbol stubs" for all external symbols that might
461 -- come from a dynamic library.
462 dyld_stubs :: [CLabel] -> Pretty.Doc
463 {- dyld_stubs imps = Pretty.vcat $ map pprDyldSymbolStub $
464 map head $ group $ sort imps-}
466 arch = platformArch $ targetPlatform dflags
467 os = platformOS $ targetPlatform dflags
469 -- (Hack) sometimes two Labels pretty-print the same, but have
470 -- different uniques; so we compare their text versions...
472 | needImportedSymbols arch os
474 (pprGotDeclaration arch os :) $
475 map ( pprImportedSymbol arch os . fst . head) $
476 groupBy (\(_,a) (_,b) -> a == b) $
477 sortBy (\(_,a) (_,b) -> compare a b) $
483 doPpr lbl = (lbl, Pretty.render $ pprCLabel lbl astyle)
484 astyle = mkCodeStyle AsmStyle
487 -- -----------------------------------------------------------------------------
488 -- Sequencing the basic blocks
490 -- Cmm BasicBlocks are self-contained entities: they always end in a
491 -- jump, either non-local or to another basic block in the same proc.
492 -- In this phase, we attempt to place the basic blocks in a sequence
493 -- such that as many of the local jumps as possible turn into
500 sequenceTop top@(CmmData _ _) = top
501 sequenceTop (CmmProc info lbl (ListGraph blocks)) =
502 CmmProc info lbl (ListGraph $ makeFarBranches $ sequenceBlocks blocks)
504 -- The algorithm is very simple (and stupid): we make a graph out of
505 -- the blocks where there is an edge from one block to another iff the
506 -- first block ends by jumping to the second. Then we topologically
507 -- sort this graph. Then traverse the list: for each block, we first
508 -- output the block, then if it has an out edge, we move the
509 -- destination of the out edge to the front of the list, and continue.
511 -- FYI, the classic layout for basic blocks uses postorder DFS; this
512 -- algorithm is implemented in Hoopl.
516 => [NatBasicBlock instr]
517 -> [NatBasicBlock instr]
519 sequenceBlocks [] = []
520 sequenceBlocks (entry:blocks) =
521 seqBlocks (mkNode entry : reverse (flattenSCCs (sccBlocks blocks)))
522 -- the first block is the entry point ==> it must remain at the start.
527 => [NatBasicBlock instr]
528 -> [SCC ( NatBasicBlock instr
532 sccBlocks blocks = stronglyConnCompFromEdgedVerticesR (map mkNode blocks)
534 -- we're only interested in the last instruction of
535 -- the block, and only if it has a single destination.
538 => [instr] -> [Unique]
541 = case jumpDestsOfInstr (last instrs) of
542 [one] -> [getUnique one]
545 mkNode :: (Instruction t)
547 -> (GenBasicBlock t, Unique, [Unique])
548 mkNode block@(BasicBlock id instrs) = (block, getUnique id, getOutEdges instrs)
550 seqBlocks :: (Eq t) => [(GenBasicBlock t1, t, [t])] -> [GenBasicBlock t1]
552 seqBlocks ((block,_,[]) : rest)
553 = block : seqBlocks rest
554 seqBlocks ((block@(BasicBlock id instrs),_,[next]) : rest)
555 | can_fallthrough = BasicBlock id (init instrs) : seqBlocks rest'
556 | otherwise = block : seqBlocks rest'
558 (can_fallthrough, rest') = reorder next [] rest
559 -- TODO: we should do a better job for cycles; try to maximise the
560 -- fallthroughs within a loop.
561 seqBlocks _ = panic "AsmCodegen:seqBlocks"
563 reorder :: (Eq a) => a -> [(t, a, t1)] -> [(t, a, t1)] -> (Bool, [(t, a, t1)])
564 reorder _ accum [] = (False, reverse accum)
565 reorder id accum (b@(block,id',out) : rest)
566 | id == id' = (True, (block,id,out) : reverse accum ++ rest)
567 | otherwise = reorder id (b:accum) rest
570 -- -----------------------------------------------------------------------------
571 -- Making far branches
573 -- Conditional branches on PowerPC are limited to +-32KB; if our Procs get too
574 -- big, we have to work around this limitation.
577 :: [NatBasicBlock Instr]
578 -> [NatBasicBlock Instr]
580 #if powerpc_TARGET_ARCH
581 makeFarBranches blocks
582 | last blockAddresses < nearLimit = blocks
583 | otherwise = zipWith handleBlock blockAddresses blocks
585 blockAddresses = scanl (+) 0 $ map blockLen blocks
586 blockLen (BasicBlock _ instrs) = length instrs
588 handleBlock addr (BasicBlock id instrs)
589 = BasicBlock id (zipWith makeFar [addr..] instrs)
591 makeFar _ (BCC ALWAYS tgt) = BCC ALWAYS tgt
592 makeFar addr (BCC cond tgt)
593 | abs (addr - targetAddr) >= nearLimit
597 where Just targetAddr = lookupUFM blockAddressMap tgt
598 makeFar _ other = other
600 nearLimit = 7000 -- 8192 instructions are allowed; let's keep some
601 -- distance, as we have a few pseudo-insns that are
602 -- pretty-printed as multiple instructions,
603 -- and it's just not worth the effort to calculate
606 blockAddressMap = listToUFM $ zip (map blockId blocks) blockAddresses
611 -- -----------------------------------------------------------------------------
619 shortcutBranches dflags tops
620 | optLevel dflags < 1 = tops -- only with -O or higher
621 | otherwise = map (apply_mapping mapping) tops'
623 (tops', mappings) = mapAndUnzip build_mapping tops
624 mapping = foldr plusUFM emptyUFM mappings
626 build_mapping :: GenCmmTop d t (ListGraph Instr)
627 -> (GenCmmTop d t (ListGraph Instr), UniqFM JumpDest)
628 build_mapping top@(CmmData _ _) = (top, emptyUFM)
629 build_mapping (CmmProc info lbl (ListGraph []))
630 = (CmmProc info lbl (ListGraph []), emptyUFM)
631 build_mapping (CmmProc info lbl (ListGraph (head:blocks)))
632 = (CmmProc info lbl (ListGraph (head:others)), mapping)
633 -- drop the shorted blocks, but don't ever drop the first one,
634 -- because it is pointed to by a global label.
636 -- find all the blocks that just consist of a jump that can be
638 -- Don't completely eliminate loops here -- that can leave a dangling jump!
639 (_, shortcut_blocks, others) = foldl split (emptyBlockSet, [], []) blocks
640 split (s, shortcut_blocks, others) b@(BasicBlock id [insn])
641 | Just (DestBlockId dest) <- canShortcut insn,
642 (setMember dest s) || dest == id -- loop checks
643 = (s, shortcut_blocks, b : others)
644 split (s, shortcut_blocks, others) (BasicBlock id [insn])
645 | Just dest <- canShortcut insn
646 = (setInsert id s, (id,dest) : shortcut_blocks, others)
647 split (s, shortcut_blocks, others) other = (s, shortcut_blocks, other : others)
650 -- build a mapping from BlockId to JumpDest for shorting branches
651 mapping = foldl add emptyUFM shortcut_blocks
652 add ufm (id,dest) = addToUFM ufm id dest
654 apply_mapping :: UniqFM JumpDest
655 -> GenCmmTop CmmStatic h (ListGraph Instr)
656 -> GenCmmTop CmmStatic h (ListGraph Instr)
657 apply_mapping ufm (CmmData sec statics)
658 = CmmData sec (map (shortcutStatic (lookupUFM ufm)) statics)
659 -- we need to get the jump tables, so apply the mapping to the entries
661 apply_mapping ufm (CmmProc info lbl (ListGraph blocks))
662 = CmmProc info lbl (ListGraph $ map short_bb blocks)
664 short_bb (BasicBlock id insns) = BasicBlock id $! map short_insn insns
665 short_insn i = shortcutJump (lookupUFM ufm) i
666 -- shortcutJump should apply the mapping repeatedly,
667 -- just in case we can short multiple branches.
669 -- -----------------------------------------------------------------------------
670 -- Instruction selection
672 -- Native code instruction selection for a chunk of stix code. For
673 -- this part of the computation, we switch from the UniqSM monad to
674 -- the NatM monad. The latter carries not only a Unique, but also an
675 -- Int denoting the current C stack pointer offset in the generated
676 -- code; this is needed for creating correct spill offsets on
677 -- architectures which don't offer, or for which it would be
678 -- prohibitively expensive to employ, a frame pointer register. Viz,
681 -- The offset is measured in bytes, and indicates the difference
682 -- between the current (simulated) C stack-ptr and the value it was at
683 -- the beginning of the block. For stacks which grow down, this value
684 -- should be either zero or negative.
686 -- Switching between the two monads whilst carrying along the same
687 -- Unique supply breaks abstraction. Is that bad?
696 genMachCode dflags cmm_top
697 = do { initial_us <- getUs
698 ; let initial_st = mkNatM_State initial_us 0 dflags
699 (new_tops, final_st) = initNat initial_st (cmmTopCodeGen dflags cmm_top)
700 final_delta = natm_delta final_st
701 final_imports = natm_imports final_st
702 ; if final_delta == 0
703 then return (new_tops, final_imports)
704 else pprPanic "genMachCode: nonzero final delta" (int final_delta)
707 -- -----------------------------------------------------------------------------
708 -- Generic Cmm optimiser
714 (b) Simple inlining: a temporary which is assigned to and then
715 used, once, can be shorted.
716 (c) Position independent code and dynamic linking
717 (i) introduce the appropriate indirections
718 and position independent refs
719 (ii) compile a list of imported symbols
721 Ideas for other things we could do (ToDo):
723 - shortcut jumps-to-jumps
724 - eliminate dead code blocks
725 - simple CSE: if an expr is assigned to a temp, then replace later occs of
726 that expr with the temp, until the expr is no longer valid (can push through
727 temp assignments, and certain assigns to mem...)
730 cmmToCmm :: DynFlags -> RawCmmTop -> (RawCmmTop, [CLabel])
731 cmmToCmm _ top@(CmmData _ _) = (top, [])
732 cmmToCmm dflags (CmmProc info lbl (ListGraph blocks)) = runCmmOpt dflags $ do
733 blocks' <- mapM cmmBlockConFold (cmmMiniInline blocks)
734 return $ CmmProc info lbl (ListGraph blocks')
736 newtype CmmOptM a = CmmOptM (([CLabel], DynFlags) -> (# a, [CLabel] #))
738 instance Monad CmmOptM where
739 return x = CmmOptM $ \(imports, _) -> (# x,imports #)
741 CmmOptM $ \(imports, dflags) ->
742 case f (imports, dflags) of
745 CmmOptM g' -> g' (imports', dflags)
747 addImportCmmOpt :: CLabel -> CmmOptM ()
748 addImportCmmOpt lbl = CmmOptM $ \(imports, _dflags) -> (# (), lbl:imports #)
750 getDynFlagsCmmOpt :: CmmOptM DynFlags
751 getDynFlagsCmmOpt = CmmOptM $ \(imports, dflags) -> (# dflags, imports #)
753 runCmmOpt :: DynFlags -> CmmOptM a -> (a, [CLabel])
754 runCmmOpt dflags (CmmOptM f) = case f ([], dflags) of
755 (# result, imports #) -> (result, imports)
757 cmmBlockConFold :: CmmBasicBlock -> CmmOptM CmmBasicBlock
758 cmmBlockConFold (BasicBlock id stmts) = do
759 stmts' <- mapM cmmStmtConFold stmts
760 return $ BasicBlock id stmts'
762 cmmStmtConFold :: CmmStmt -> CmmOptM CmmStmt
766 -> do src' <- cmmExprConFold DataReference src
767 return $ case src' of
768 CmmReg reg' | reg == reg' -> CmmNop
769 new_src -> CmmAssign reg new_src
772 -> do addr' <- cmmExprConFold DataReference addr
773 src' <- cmmExprConFold DataReference src
774 return $ CmmStore addr' src'
777 -> do addr' <- cmmExprConFold JumpReference addr
778 return $ CmmJump addr' regs
780 CmmCall target regs args srt returns
781 -> do target' <- case target of
782 CmmCallee e conv -> do
783 e' <- cmmExprConFold CallReference e
784 return $ CmmCallee e' conv
785 other -> return other
786 args' <- mapM (\(CmmHinted arg hint) -> do
787 arg' <- cmmExprConFold DataReference arg
788 return (CmmHinted arg' hint)) args
789 return $ CmmCall target' regs args' srt returns
791 CmmCondBranch test dest
792 -> do test' <- cmmExprConFold DataReference test
793 return $ case test' of
794 CmmLit (CmmInt 0 _) ->
795 CmmComment (mkFastString ("deleted: " ++
796 showSDoc (pprStmt stmt)))
798 CmmLit (CmmInt _ _) -> CmmBranch dest
799 _other -> CmmCondBranch test' dest
802 -> do expr' <- cmmExprConFold DataReference expr
803 return $ CmmSwitch expr' ids
809 cmmExprConFold :: ReferenceKind -> CmmExpr -> CmmOptM CmmExpr
810 cmmExprConFold referenceKind expr
813 -> do addr' <- cmmExprConFold DataReference addr
814 return $ CmmLoad addr' rep
817 -- For MachOps, we first optimize the children, and then we try
818 -- our hand at some constant-folding.
819 -> do args' <- mapM (cmmExprConFold DataReference) args
820 return $ cmmMachOpFold mop args'
822 CmmLit (CmmLabel lbl)
824 dflags <- getDynFlagsCmmOpt
825 cmmMakeDynamicReference dflags addImportCmmOpt referenceKind lbl
826 CmmLit (CmmLabelOff lbl off)
828 dflags <- getDynFlagsCmmOpt
829 dynRef <- cmmMakeDynamicReference dflags addImportCmmOpt referenceKind lbl
830 return $ cmmMachOpFold (MO_Add wordWidth) [
832 (CmmLit $ CmmInt (fromIntegral off) wordWidth)
835 -- On powerpc (non-PIC), it's easier to jump directly to a label than
836 -- to use the register table, so we replace these registers
837 -- with the corresponding labels:
838 CmmReg (CmmGlobal EagerBlackholeInfo)
839 | cTargetArch == PPC && not opt_PIC
840 -> cmmExprConFold referenceKind $
841 CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit "__stg_EAGER_BLACKHOLE_info")))
842 CmmReg (CmmGlobal GCEnter1)
843 | cTargetArch == PPC && not opt_PIC
844 -> cmmExprConFold referenceKind $
845 CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit "__stg_gc_enter_1")))
846 CmmReg (CmmGlobal GCFun)
847 | cTargetArch == PPC && not opt_PIC
848 -> cmmExprConFold referenceKind $
849 CmmLit (CmmLabel (mkCmmCodeLabel rtsPackageId (fsLit "__stg_gc_fun")))