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 import PositionIndependentCode
25 import CmmOpt ( cmmMiniInline, cmmMachOpFold )
26 import PprCmm ( pprStmt, pprCmms )
31 import Unique ( Unique, getUnique )
34 import List ( groupBy, sortBy )
35 import ErrUtils ( dumpIfSet_dyn )
37 import StaticFlags ( opt_Static, opt_PIC )
39 import Config ( cProjectVersion )
42 import qualified Pretty
50 import List ( intersperse )
59 The native-code generator has machine-independent and
60 machine-dependent modules.
62 This module ("AsmCodeGen") is the top-level machine-independent
63 module. Before entering machine-dependent land, we do some
64 machine-independent optimisations (defined below) on the
67 We convert to the machine-specific 'Instr' datatype with
68 'cmmCodeGen', assuming an infinite supply of registers. We then use
69 a machine-independent register allocator ('regAlloc') to rejoin
70 reality. Obviously, 'regAlloc' has machine-specific helper
71 functions (see about "RegAllocInfo" below).
73 Finally, we order the basic blocks of the function so as to minimise
74 the number of jumps between blocks, by utilising fallthrough wherever
77 The machine-dependent bits break down as follows:
79 * ["MachRegs"] Everything about the target platform's machine
80 registers (and immediate operands, and addresses, which tend to
81 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', ...)
87 * ["MachCodeGen"] is where 'Cmm' stuff turns into
90 * ["PprMach"] 'pprInstr' turns an 'Instr' into text (well, really
93 * ["RegAllocInfo"] In the register allocator, we manipulate
94 'MRegsState's, which are 'BitSet's, one bit per machine register.
95 When we want to say something about a specific machine register
96 (e.g., ``it gets clobbered by this instruction''), we set/unset
97 its bit. Obviously, we do this 'BitSet' thing for efficiency
100 The 'RegAllocInfo' module collects together the machine-specific
101 info needed to do register allocation.
103 * ["RegisterAlloc"] The (machine-independent) register allocator.
106 -- -----------------------------------------------------------------------------
107 -- Top-level of the native codegen
109 -- NB. We *lazilly* compile each block of code for space reasons.
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))
116 cgCmm :: [RawCmmTop] -> UniqSM (RawCmm, Pretty.Doc, [CLabel])
118 lazyMapUs (cmmNativeGen dflags) tops `thenUs` \ results ->
119 case unzip3 results of { (cmms,docs,imps) ->
120 returnUs (Cmm cmms, my_vcat docs, concat imps)
123 case res of { (ppr_cmms, insn_sdoc, imports) -> do
124 dumpIfSet_dyn dflags Opt_D_dump_opt_cmm "Optimised Cmm" (pprCmms [ppr_cmms])
125 return (insn_sdoc Pretty.$$ dyld_stubs imports
126 #if HAVE_SUBSECTIONS_VIA_SYMBOLS
127 -- On recent versions of Darwin, the linker supports
128 -- dead-stripping of code and data on a per-symbol basis.
129 -- There's a hack to make this work in PprMach.pprNatCmmTop.
130 Pretty.$$ Pretty.text ".subsections_via_symbols"
132 #if HAVE_GNU_NONEXEC_STACK
133 -- On recent GNU ELF systems one can mark an object file
134 -- as not requiring an executable stack. If all objects
135 -- linked into a program have this note then the program
136 -- will not use an executable stack, which is good for
137 -- security. GHC generated code does not need an executable
138 -- stack so add the note in:
139 Pretty.$$ Pretty.text ".section .note.GNU-stack,\"\",@progbits"
141 #if !defined(darwin_TARGET_OS)
142 -- And just because every other compiler does, lets stick in
143 -- an identifier directive: .ident "GHC x.y.z"
144 Pretty.$$ let compilerIdent = Pretty.text "GHC" Pretty.<+>
145 Pretty.text cProjectVersion
146 in Pretty.text ".ident" Pretty.<+>
147 Pretty.doubleQuotes compilerIdent
155 | dopt Opt_SplitObjs dflags = split_marker : tops
158 split_marker = CmmProc [] mkSplitMarkerLabel [] []
160 -- Generate "symbol stubs" for all external symbols that might
161 -- come from a dynamic library.
162 {- dyld_stubs imps = Pretty.vcat $ map pprDyldSymbolStub $
163 map head $ group $ sort imps-}
165 -- (Hack) sometimes two Labels pretty-print the same, but have
166 -- different uniques; so we compare their text versions...
168 | needImportedSymbols
170 (pprGotDeclaration :) $
171 map (pprImportedSymbol . fst . head) $
172 groupBy (\(_,a) (_,b) -> a == b) $
173 sortBy (\(_,a) (_,b) -> compare a b) $
179 where doPpr lbl = (lbl, Pretty.render $ pprCLabel lbl astyle)
180 astyle = mkCodeStyle AsmStyle
183 my_vcat sds = Pretty.vcat sds
185 my_vcat sds = Pretty.vcat (
188 Pretty.$$ Pretty.ptext SLIT("# ___ncg_debug_marker")
189 Pretty.$$ Pretty.char ' '
196 -- Complete native code generation phase for a single top-level chunk
199 cmmNativeGen :: DynFlags -> RawCmmTop -> UniqSM (RawCmmTop, Pretty.Doc, [CLabel])
200 cmmNativeGen dflags cmm
201 = {-# SCC "fixAssigns" #-}
202 fixAssignsTop cmm `thenUs` \ fixed_cmm ->
203 {-# SCC "genericOpt" #-}
204 cmmToCmm fixed_cmm `bind` \ (cmm, imports) ->
205 (if dopt Opt_D_dump_opt_cmm dflags -- space leak avoidance
207 else CmmData Text []) `bind` \ ppr_cmm ->
208 {-# SCC "genMachCode" #-}
209 genMachCode cmm `thenUs` \ (pre_regalloc, lastMinuteImports) ->
210 {-# SCC "regAlloc" #-}
211 mapUs regAlloc pre_regalloc `thenUs` \ with_regs ->
212 {-# SCC "shortcutBranches" #-}
213 shortcutBranches dflags with_regs `bind` \ shorted ->
214 {-# SCC "sequenceBlocks" #-}
215 map sequenceTop shorted `bind` \ sequenced ->
216 {-# SCC "x86fp_kludge" #-}
217 map x86fp_kludge sequenced `bind` \ final_mach_code ->
219 Pretty.vcat (map pprNatCmmTop final_mach_code) `bind` \ final_sdoc ->
221 returnUs (ppr_cmm, final_sdoc Pretty.$$ Pretty.text "", lastMinuteImports ++ imports)
223 x86fp_kludge :: NatCmmTop -> NatCmmTop
224 x86fp_kludge top@(CmmData _ _) = top
226 x86fp_kludge top@(CmmProc info lbl params code) =
227 CmmProc info lbl params (map bb_i386_insert_ffrees code)
229 bb_i386_insert_ffrees (BasicBlock id instrs) =
230 BasicBlock id (i386_insert_ffrees instrs)
232 x86fp_kludge top = top
235 -- -----------------------------------------------------------------------------
236 -- Sequencing the basic blocks
238 -- Cmm BasicBlocks are self-contained entities: they always end in a
239 -- jump, either non-local or to another basic block in the same proc.
240 -- In this phase, we attempt to place the basic blocks in a sequence
241 -- such that as many of the local jumps as possible turn into
244 sequenceTop :: NatCmmTop -> NatCmmTop
245 sequenceTop top@(CmmData _ _) = top
246 sequenceTop (CmmProc info lbl params blocks) =
247 CmmProc info lbl params (makeFarBranches $ sequenceBlocks blocks)
249 -- The algorithm is very simple (and stupid): we make a graph out of
250 -- the blocks where there is an edge from one block to another iff the
251 -- first block ends by jumping to the second. Then we topologically
252 -- sort this graph. Then traverse the list: for each block, we first
253 -- output the block, then if it has an out edge, we move the
254 -- destination of the out edge to the front of the list, and continue.
256 sequenceBlocks :: [NatBasicBlock] -> [NatBasicBlock]
257 sequenceBlocks [] = []
258 sequenceBlocks (entry:blocks) =
259 seqBlocks (mkNode entry : reverse (flattenSCCs (sccBlocks blocks)))
260 -- the first block is the entry point ==> it must remain at the start.
262 sccBlocks :: [NatBasicBlock] -> [SCC (NatBasicBlock,Unique,[Unique])]
263 sccBlocks blocks = stronglyConnCompR (map mkNode blocks)
265 getOutEdges :: [Instr] -> [Unique]
266 getOutEdges instrs = case jumpDests (last instrs) [] of
267 [one] -> [getUnique one]
269 -- we're only interested in the last instruction of
270 -- the block, and only if it has a single destination.
272 mkNode block@(BasicBlock id instrs) = (block, getUnique id, getOutEdges instrs)
275 seqBlocks ((block,_,[]) : rest)
276 = block : seqBlocks rest
277 seqBlocks ((block@(BasicBlock id instrs),_,[next]) : rest)
278 | can_fallthrough = BasicBlock id (init instrs) : seqBlocks rest'
279 | otherwise = block : seqBlocks rest'
281 (can_fallthrough, rest') = reorder next [] rest
282 -- TODO: we should do a better job for cycles; try to maximise the
283 -- fallthroughs within a loop.
284 seqBlocks _ = panic "AsmCodegen:seqBlocks"
286 reorder id accum [] = (False, reverse accum)
287 reorder id accum (b@(block,id',out) : rest)
288 | id == id' = (True, (block,id,out) : reverse accum ++ rest)
289 | otherwise = reorder id (b:accum) rest
292 -- -----------------------------------------------------------------------------
293 -- Making far branches
295 -- Conditional branches on PowerPC are limited to +-32KB; if our Procs get too
296 -- big, we have to work around this limitation.
298 makeFarBranches :: [NatBasicBlock] -> [NatBasicBlock]
300 #if powerpc_TARGET_ARCH
301 makeFarBranches blocks
302 | last blockAddresses < nearLimit = blocks
303 | otherwise = zipWith handleBlock blockAddresses blocks
305 blockAddresses = scanl (+) 0 $ map blockLen blocks
306 blockLen (BasicBlock _ instrs) = length instrs
308 handleBlock addr (BasicBlock id instrs)
309 = BasicBlock id (zipWith makeFar [addr..] instrs)
311 makeFar addr (BCC ALWAYS tgt) = BCC ALWAYS tgt
312 makeFar addr (BCC cond tgt)
313 | abs (addr - targetAddr) >= nearLimit
317 where Just targetAddr = lookupUFM blockAddressMap tgt
318 makeFar addr other = other
320 nearLimit = 7000 -- 8192 instructions are allowed; let's keep some
321 -- distance, as we have a few pseudo-insns that are
322 -- pretty-printed as multiple instructions,
323 -- and it's just not worth the effort to calculate
326 blockAddressMap = listToUFM $ zip (map blockId blocks) blockAddresses
331 -- -----------------------------------------------------------------------------
334 shortcutBranches :: DynFlags -> [NatCmmTop] -> [NatCmmTop]
335 shortcutBranches dflags tops
336 | optLevel dflags < 1 = tops -- only with -O or higher
337 | otherwise = map (apply_mapping mapping) tops'
339 (tops', mappings) = mapAndUnzip build_mapping tops
340 mapping = foldr plusUFM emptyUFM mappings
342 build_mapping top@(CmmData _ _) = (top, emptyUFM)
343 build_mapping (CmmProc info lbl params [])
344 = (CmmProc info lbl params [], emptyUFM)
345 build_mapping (CmmProc info lbl params (head:blocks))
346 = (CmmProc info lbl params (head:others), mapping)
347 -- drop the shorted blocks, but don't ever drop the first one,
348 -- because it is pointed to by a global label.
350 -- find all the blocks that just consist of a jump that can be
352 (shortcut_blocks, others) = partitionWith split blocks
353 split (BasicBlock id [insn]) | Just dest <- canShortcut insn
355 split other = Right other
357 -- build a mapping from BlockId to JumpDest for shorting branches
358 mapping = foldl add emptyUFM shortcut_blocks
359 add ufm (id,dest) = addToUFM ufm id dest
361 apply_mapping ufm (CmmData sec statics)
362 = CmmData sec (map (shortcutStatic (lookupUFM ufm)) statics)
363 -- we need to get the jump tables, so apply the mapping to the entries
365 apply_mapping ufm (CmmProc info lbl params blocks)
366 = CmmProc info lbl params (map short_bb blocks)
368 short_bb (BasicBlock id insns) = BasicBlock id $! map short_insn insns
369 short_insn i = shortcutJump (lookupUFM ufm) i
370 -- shortcutJump should apply the mapping repeatedly,
371 -- just in case we can short multiple branches.
373 -- -----------------------------------------------------------------------------
374 -- Instruction selection
376 -- Native code instruction selection for a chunk of stix code. For
377 -- this part of the computation, we switch from the UniqSM monad to
378 -- the NatM monad. The latter carries not only a Unique, but also an
379 -- Int denoting the current C stack pointer offset in the generated
380 -- code; this is needed for creating correct spill offsets on
381 -- architectures which don't offer, or for which it would be
382 -- prohibitively expensive to employ, a frame pointer register. Viz,
385 -- The offset is measured in bytes, and indicates the difference
386 -- between the current (simulated) C stack-ptr and the value it was at
387 -- the beginning of the block. For stacks which grow down, this value
388 -- should be either zero or negative.
390 -- Switching between the two monads whilst carrying along the same
391 -- Unique supply breaks abstraction. Is that bad?
393 genMachCode :: RawCmmTop -> UniqSM ([NatCmmTop], [CLabel])
396 = do { initial_us <- getUs
397 ; let initial_st = mkNatM_State initial_us 0
398 (new_tops, final_st) = initNat initial_st (cmmTopCodeGen cmm_top)
399 final_us = natm_us final_st
400 final_delta = natm_delta final_st
401 final_imports = natm_imports final_st
402 ; if final_delta == 0
403 then return (new_tops, final_imports)
404 else pprPanic "genMachCode: nonzero final delta" (int final_delta)
407 -- -----------------------------------------------------------------------------
408 -- Fixup assignments to global registers so that they assign to
409 -- locations within the RegTable, if appropriate.
411 -- Note that we currently don't fixup reads here: they're done by
412 -- the generic optimiser below, to avoid having two separate passes
415 fixAssignsTop :: RawCmmTop -> UniqSM RawCmmTop
416 fixAssignsTop top@(CmmData _ _) = returnUs top
417 fixAssignsTop (CmmProc info lbl params blocks) =
418 mapUs fixAssignsBlock blocks `thenUs` \ blocks' ->
419 returnUs (CmmProc info lbl params blocks')
421 fixAssignsBlock :: CmmBasicBlock -> UniqSM CmmBasicBlock
422 fixAssignsBlock (BasicBlock id stmts) =
423 fixAssigns stmts `thenUs` \ stmts' ->
424 returnUs (BasicBlock id stmts')
426 fixAssigns :: [CmmStmt] -> UniqSM [CmmStmt]
428 mapUs fixAssign stmts `thenUs` \ stmtss ->
429 returnUs (concat stmtss)
431 fixAssign :: CmmStmt -> UniqSM [CmmStmt]
432 fixAssign (CmmAssign (CmmGlobal reg) src)
433 | Left realreg <- reg_or_addr
434 = returnUs [CmmAssign (CmmGlobal reg) src]
435 | Right baseRegAddr <- reg_or_addr
436 = returnUs [CmmStore baseRegAddr src]
437 -- Replace register leaves with appropriate StixTrees for
438 -- the given target. GlobalRegs which map to a reg on this
439 -- arch are left unchanged. Assigning to BaseReg is always
440 -- illegal, so we check for that.
442 reg_or_addr = get_GlobalReg_reg_or_addr reg
444 fixAssign other_stmt = returnUs [other_stmt]
446 -- -----------------------------------------------------------------------------
447 -- Generic Cmm optimiser
453 (b) Simple inlining: a temporary which is assigned to and then
454 used, once, can be shorted.
455 (c) Replacement of references to GlobalRegs which do not have
456 machine registers by the appropriate memory load (eg.
457 Hp ==> *(BaseReg + 34) ).
458 (d) Position independent code and dynamic linking
459 (i) introduce the appropriate indirections
460 and position independent refs
461 (ii) compile a list of imported symbols
463 Ideas for other things we could do (ToDo):
465 - shortcut jumps-to-jumps
466 - eliminate dead code blocks
467 - simple CSE: if an expr is assigned to a temp, then replace later occs of
468 that expr with the temp, until the expr is no longer valid (can push through
469 temp assignments, and certain assigns to mem...)
472 cmmToCmm :: RawCmmTop -> (RawCmmTop, [CLabel])
473 cmmToCmm top@(CmmData _ _) = (top, [])
474 cmmToCmm (CmmProc info lbl params blocks) = runCmmOpt $ do
475 blocks' <- mapM cmmBlockConFold (cmmMiniInline blocks)
476 return $ CmmProc info lbl params blocks'
478 newtype CmmOptM a = CmmOptM ([CLabel] -> (# a, [CLabel] #))
480 instance Monad CmmOptM where
481 return x = CmmOptM $ \imports -> (# x,imports #)
483 CmmOptM $ \imports ->
487 CmmOptM g' -> g' imports'
489 addImportCmmOpt :: CLabel -> CmmOptM ()
490 addImportCmmOpt lbl = CmmOptM $ \imports -> (# (), lbl:imports #)
492 runCmmOpt :: CmmOptM a -> (a, [CLabel])
493 runCmmOpt (CmmOptM f) = case f [] of
494 (# result, imports #) -> (result, imports)
496 cmmBlockConFold :: CmmBasicBlock -> CmmOptM CmmBasicBlock
497 cmmBlockConFold (BasicBlock id stmts) = do
498 stmts' <- mapM cmmStmtConFold stmts
499 return $ BasicBlock id stmts'
504 -> do src' <- cmmExprConFold DataReference src
505 return $ case src' of
506 CmmReg reg' | reg == reg' -> CmmNop
507 new_src -> CmmAssign reg new_src
510 -> do addr' <- cmmExprConFold DataReference addr
511 src' <- cmmExprConFold DataReference src
512 return $ CmmStore addr' src'
515 -> do addr' <- cmmExprConFold JumpReference addr
516 return $ CmmJump addr' regs
518 CmmCall target regs args srt
519 -> do target' <- case target of
520 CmmForeignCall e conv -> do
521 e' <- cmmExprConFold CallReference e
522 return $ CmmForeignCall e' conv
523 other -> return other
524 args' <- mapM (\(arg, hint) -> do
525 arg' <- cmmExprConFold DataReference arg
526 return (arg', hint)) args
527 return $ CmmCall target' regs args' srt
529 CmmCondBranch test dest
530 -> do test' <- cmmExprConFold DataReference test
531 return $ case test' of
532 CmmLit (CmmInt 0 _) ->
533 CmmComment (mkFastString ("deleted: " ++
534 showSDoc (pprStmt stmt)))
536 CmmLit (CmmInt n _) -> CmmBranch dest
537 other -> CmmCondBranch test' dest
540 -> do expr' <- cmmExprConFold DataReference expr
541 return $ CmmSwitch expr' ids
547 cmmExprConFold referenceKind expr
550 -> do addr' <- cmmExprConFold DataReference addr
551 return $ CmmLoad addr' rep
554 -- For MachOps, we first optimize the children, and then we try
555 -- our hand at some constant-folding.
556 -> do args' <- mapM (cmmExprConFold DataReference) args
557 return $ cmmMachOpFold mop args'
559 CmmLit (CmmLabel lbl)
560 -> cmmMakeDynamicReference addImportCmmOpt referenceKind lbl
561 CmmLit (CmmLabelOff lbl off)
562 -> do dynRef <- cmmMakeDynamicReference addImportCmmOpt referenceKind lbl
563 return $ cmmMachOpFold (MO_Add wordRep) [
565 (CmmLit $ CmmInt (fromIntegral off) wordRep)
568 #if powerpc_TARGET_ARCH
569 -- On powerpc (non-PIC), it's easier to jump directly to a label than
570 -- to use the register table, so we replace these registers
571 -- with the corresponding labels:
572 CmmReg (CmmGlobal GCEnter1)
574 -> cmmExprConFold referenceKind $
575 CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_enter_1")))
576 CmmReg (CmmGlobal GCFun)
578 -> cmmExprConFold referenceKind $
579 CmmLit (CmmLabel (mkRtsCodeLabel SLIT( "__stg_gc_fun")))
582 CmmReg (CmmGlobal mid)
583 -- Replace register leaves with appropriate StixTrees for
584 -- the given target. MagicIds which map to a reg on this
585 -- arch are left unchanged. For the rest, BaseReg is taken
586 -- to mean the address of the reg table in MainCapability,
587 -- and for all others we generate an indirection to its
588 -- location in the register table.
589 -> case get_GlobalReg_reg_or_addr mid of
590 Left realreg -> return expr
593 BaseReg -> cmmExprConFold DataReference baseRegAddr
594 other -> cmmExprConFold DataReference
595 (CmmLoad baseRegAddr (globalRegRep mid))
596 -- eliminate zero offsets
598 -> cmmExprConFold referenceKind (CmmReg reg)
600 CmmRegOff (CmmGlobal mid) offset
601 -- RegOf leaves are just a shorthand form. If the reg maps
602 -- to a real reg, we keep the shorthand, otherwise, we just
603 -- expand it and defer to the above code.
604 -> case get_GlobalReg_reg_or_addr mid of
605 Left realreg -> return expr
607 -> cmmExprConFold DataReference (CmmMachOp (MO_Add wordRep) [
608 CmmReg (CmmGlobal mid),
609 CmmLit (CmmInt (fromIntegral offset)
614 -- -----------------------------------------------------------------------------