1 -----------------------------------------------------------------------------
3 -- Monad for Stg to C-- code generation
5 -- (c) The University of Glasgow 2004-2006
7 -----------------------------------------------------------------------------
12 initC, thenC, thenFC, listCs, listFCs, mapCs, mapFCs,
13 returnFC, fixC, fixC_, nopC, whenC,
14 newUnique, newUniqSupply,
16 emit, emitData, emitProc, emitProcWithConvention, emitSimpleProc,
18 getCmm, cgStmtsToBlocks,
19 getCodeR, getCode, getHeapUsage,
21 forkClosureBody, forkStatics, forkAlts, forkProc, codeOnly,
26 withSequel, getSequel,
28 setSRTLabel, getSRTLabel,
29 setTickyCtrLabel, getTickyCtrLabel,
31 withUpdFrameOff, getUpdFrameOff, initUpdFrameOff,
33 HeapUsage(..), VirtualHpOffset, initHpUsage,
34 getHpUsage, setHpUsage, heapHWM,
35 setVirtHp, getVirtHp, setRealHp,
39 -- ideally we wouldn't export these, but some other modules access internal state
40 getState, setState, getInfoDown, getDynFlags, getThisPackage,
42 -- more localised access to monad state
43 CgIdInfo(..), CgLoc(..),
44 getBinds, setBinds, getStaticBinds,
46 -- out of general friendliness, we also export ...
47 CgInfoDownwards(..), CgState(..) -- non-abstract
50 #include "HsVersions.h"
55 import ZipCfgCmmRep (UpdFrameOffset)
59 import TyCon ( PrimRep )
68 import FastString(sLit)
73 import Prelude hiding( sequence )
74 import qualified Prelude( sequence )
76 infixr 9 `thenC` -- Right-associative!
80 --------------------------------------------------------
81 -- The FCode monad and its types
82 --------------------------------------------------------
84 newtype FCode a = FCode (CgInfoDownwards -> CgState -> (a, CgState))
86 instance Monad FCode where
92 {-# INLINE returnFC #-}
94 initC :: DynFlags -> Module -> FCode a -> IO a
95 initC dflags mod (FCode code)
96 = do { uniqs <- mkSplitUniqSupply 'c'
97 ; case code (initCgInfoDown dflags mod) (initCgState uniqs) of
98 (res, _) -> return res
101 returnFC :: a -> FCode a
102 returnFC val = FCode (\_info_down state -> (val, state))
104 thenC :: FCode () -> FCode a -> FCode a
105 thenC (FCode m) (FCode k) =
106 FCode (\info_down state -> let (_,new_state) = m info_down state in
107 k info_down new_state)
112 whenC :: Bool -> FCode () -> FCode ()
113 whenC True code = code
114 whenC False _code = nopC
116 listCs :: [FCode ()] -> FCode ()
117 listCs [] = return ()
122 mapCs :: (a -> FCode ()) -> [a] -> FCode ()
125 thenFC :: FCode a -> (a -> FCode c) -> FCode c
126 thenFC (FCode m) k = FCode (
129 (m_result, new_state) = m info_down state
130 (FCode kcode) = k m_result
132 kcode info_down new_state
135 listFCs :: [FCode a] -> FCode [a]
136 listFCs = Prelude.sequence
138 mapFCs :: (a -> FCode b) -> [a] -> FCode [b]
141 fixC :: (a -> FCode a) -> FCode a
146 result@(v,_) = fc info_down state
152 fixC_ :: (a -> FCode a) -> FCode ()
153 fixC_ fcode = fixC fcode >> return ()
155 --------------------------------------------------------
156 -- The code generator environment
157 --------------------------------------------------------
159 -- This monadery has some information that it only passes
160 -- *downwards*, as well as some ``state'' which is modified
163 data CgInfoDownwards -- information only passed *downwards* by the monad
165 cgd_dflags :: DynFlags,
166 cgd_mod :: Module, -- Module being compiled
167 cgd_statics :: CgBindings, -- [Id -> info] : static environment
168 cgd_srt_lbl :: CLabel, -- Label of the current top-level SRT
169 cgd_updfr_off :: UpdFrameOffset, -- Size of current update frame
170 cgd_ticky :: CLabel, -- Current destination for ticky counts
171 cgd_sequel :: Sequel -- What to do at end of basic block
174 type CgBindings = IdEnv CgIdInfo
178 { cg_id :: Id -- Id that this is the info for
179 -- Can differ from the Id at occurrence sites by
180 -- virtue of being externalised, for splittable C
181 , cg_lf :: LambdaFormInfo
182 , cg_loc :: CgLoc -- CmmExpr for the *tagged* value
183 , cg_rep :: PrimRep -- Cache for (idPrimRep id)
184 , cg_tag :: {-# UNPACK #-} !DynTag -- Cache for (lfDynTag cg_lf)
188 = CmmLoc CmmExpr -- A stable CmmExpr; that is, one not mentioning
189 -- Hp, so that it remains valid across calls
191 | LneLoc BlockId [LocalReg] -- A join point
192 -- A join point (= let-no-escape) should only
193 -- be tail-called, and in a saturated way.
194 -- To tail-call it, assign to these locals,
195 -- and branch to the block id
197 instance Outputable CgIdInfo where
198 ppr (CgIdInfo { cg_id = id, cg_loc = loc })
199 = ppr id <+> ptext (sLit "-->") <+> ppr loc
201 instance Outputable CgLoc where
202 ppr (CmmLoc e) = ptext (sLit "cmm") <+> ppr e
203 ppr (LneLoc b rs) = ptext (sLit "lne") <+> ppr b <+> ppr rs
206 -- Sequel tells what to do with the result of this expression
208 = Return Bool -- Return result(s) to continuation found on the stack
209 -- True <=> the continuation is update code (???)
212 [LocalReg] -- Put result(s) in these regs and fall through
213 -- NB: no void arguments here
214 Bool -- Should we adjust the heap pointer back to recover
215 -- space that's unused on this path?
216 -- We need to do this only if the expression may
217 -- allocate (e.g. it's a foreign call or allocating primOp)
218 instance Show Sequel where
219 show (Return _) = "Sequel: Return"
220 show (AssignTo _ _) = "Sequel: Assign"
222 initCgInfoDown :: DynFlags -> Module -> CgInfoDownwards
223 initCgInfoDown dflags mod
224 = MkCgInfoDown { cgd_dflags = dflags,
226 cgd_statics = emptyVarEnv,
227 cgd_srt_lbl = error "initC: srt_lbl",
228 cgd_updfr_off = initUpdFrameOff,
229 cgd_ticky = mkTopTickyCtrLabel,
230 cgd_sequel = initSequel }
233 initSequel = Return False
235 initUpdFrameOff :: UpdFrameOffset
236 initUpdFrameOff = widthInBytes wordWidth -- space for the RA
239 --------------------------------------------------------
240 -- The code generator state
241 --------------------------------------------------------
245 cgs_stmts :: CmmAGraph, -- Current procedure
247 cgs_tops :: OrdList CmmTopZ,
248 -- Other procedures and data blocks in this compilation unit
249 -- Both are ordered only so that we can
250 -- reduce forward references, when it's easy to do so
252 cgs_binds :: CgBindings, -- [Id -> info] : *local* bindings environment
253 -- Bindings for top-level things are given in
254 -- the info-down part
256 cgs_hp_usg :: HeapUsage,
258 cgs_uniqs :: UniqSupply }
262 virtHp :: VirtualHpOffset, -- Virtual offset of highest-allocated word
263 realHp :: VirtualHpOffset -- realHp: Virtual offset of real heap ptr
266 type VirtualHpOffset = WordOff
268 initCgState :: UniqSupply -> CgState
270 = MkCgState { cgs_stmts = mkNop, cgs_tops = nilOL,
271 cgs_binds = emptyVarEnv,
272 cgs_hp_usg = initHpUsage,
275 stateIncUsage :: CgState -> CgState -> CgState
276 -- stateIncUsage@ e1 e2 incorporates in e1
277 -- the heap high water mark found in e2.
278 stateIncUsage s1 s2@(MkCgState { cgs_hp_usg = hp_usg })
279 = s1 { cgs_hp_usg = cgs_hp_usg s1 `maxHpHw` virtHp hp_usg }
280 `addCodeBlocksFrom` s2
282 addCodeBlocksFrom :: CgState -> CgState -> CgState
283 -- Add code blocks from the latter to the former
284 -- (The cgs_stmts will often be empty, but not always; see codeOnly)
285 s1 `addCodeBlocksFrom` s2
286 = s1 { cgs_stmts = cgs_stmts s1 <*> cgs_stmts s2,
287 cgs_tops = cgs_tops s1 `appOL` cgs_tops s2 }
290 -- The heap high water mark is the larger of virtHp and hwHp. The latter is
291 -- only records the high water marks of forked-off branches, so to find the
292 -- heap high water mark you have to take the max of virtHp and hwHp. Remember,
293 -- virtHp never retreats!
295 -- Note Jan 04: ok, so why do we only look at the virtual Hp??
297 heapHWM :: HeapUsage -> VirtualHpOffset
300 initHpUsage :: HeapUsage
301 initHpUsage = HeapUsage { virtHp = 0, realHp = 0 }
303 maxHpHw :: HeapUsage -> VirtualHpOffset -> HeapUsage
304 hp_usg `maxHpHw` hw = hp_usg { virtHp = virtHp hp_usg `max` hw }
307 --------------------------------------------------------
308 -- Operators for getting and setting the state and "info_down".
309 --------------------------------------------------------
311 getState :: FCode CgState
312 getState = FCode $ \_info_down state -> (state,state)
314 setState :: CgState -> FCode ()
315 setState state = FCode $ \_info_down _ -> ((),state)
317 getHpUsage :: FCode HeapUsage
320 return $ cgs_hp_usg state
322 setHpUsage :: HeapUsage -> FCode ()
323 setHpUsage new_hp_usg = do
325 setState $ state {cgs_hp_usg = new_hp_usg}
327 setVirtHp :: VirtualHpOffset -> FCode ()
329 = do { hp_usage <- getHpUsage
330 ; setHpUsage (hp_usage {virtHp = new_virtHp}) }
332 getVirtHp :: FCode VirtualHpOffset
334 = do { hp_usage <- getHpUsage
335 ; return (virtHp hp_usage) }
337 setRealHp :: VirtualHpOffset -> FCode ()
339 = do { hp_usage <- getHpUsage
340 ; setHpUsage (hp_usage {realHp = new_realHp}) }
342 getBinds :: FCode CgBindings
345 return $ cgs_binds state
347 setBinds :: CgBindings -> FCode ()
348 setBinds new_binds = do
350 setState $ state {cgs_binds = new_binds}
352 getStaticBinds :: FCode CgBindings
355 return (cgd_statics info)
357 withState :: FCode a -> CgState -> FCode (a,CgState)
358 withState (FCode fcode) newstate = FCode $ \info_down state ->
359 let (retval, state2) = fcode info_down newstate in ((retval,state2), state)
361 newUniqSupply :: FCode UniqSupply
364 let (us1, us2) = splitUniqSupply (cgs_uniqs state)
365 setState $ state { cgs_uniqs = us1 }
368 newUnique :: FCode Unique
371 return (uniqFromSupply us)
374 getInfoDown :: FCode CgInfoDownwards
375 getInfoDown = FCode $ \info_down state -> (info_down,state)
377 getDynFlags :: FCode DynFlags
378 getDynFlags = liftM cgd_dflags getInfoDown
380 getThisPackage :: FCode PackageId
381 getThisPackage = liftM thisPackage getDynFlags
383 withInfoDown :: FCode a -> CgInfoDownwards -> FCode a
384 withInfoDown (FCode fcode) info_down = FCode $ \_ state -> fcode info_down state
386 doFCode :: FCode a -> CgInfoDownwards -> CgState -> (a,CgState)
387 doFCode (FCode fcode) info_down state = fcode info_down state
390 -- ----------------------------------------------------------------------------
391 -- Get the current module name
393 getModuleName :: FCode Module
394 getModuleName = do { info <- getInfoDown; return (cgd_mod info) }
396 -- ----------------------------------------------------------------------------
397 -- Get/set the end-of-block info
399 withSequel :: Sequel -> FCode () -> FCode ()
400 withSequel sequel code
401 = do { info <- getInfoDown
402 ; withInfoDown code (info {cgd_sequel = sequel }) }
404 getSequel :: FCode Sequel
405 getSequel = do { info <- getInfoDown
406 ; return (cgd_sequel info) }
408 -- ----------------------------------------------------------------------------
409 -- Get/set the current SRT label
411 -- There is just one SRT for each top level binding; all the nested
412 -- bindings use sub-sections of this SRT. The label is passed down to
413 -- the nested bindings via the monad.
415 getSRTLabel :: FCode CLabel -- Used only by cgPanic
416 getSRTLabel = do info <- getInfoDown
417 return (cgd_srt_lbl info)
419 setSRTLabel :: CLabel -> FCode a -> FCode a
420 setSRTLabel srt_lbl code
421 = do info <- getInfoDown
422 withInfoDown code (info { cgd_srt_lbl = srt_lbl})
424 -- ----------------------------------------------------------------------------
425 -- Get/set the size of the update frame
427 -- We keep track of the size of the update frame so that we
428 -- can set the stack pointer to the proper address on return
429 -- (or tail call) from the closure.
430 -- There should be at most one update frame for each closure.
431 -- Note: I'm including the size of the original return address
432 -- in the size of the update frame -- hence the default case on `get'.
434 withUpdFrameOff :: UpdFrameOffset -> FCode () -> FCode ()
435 withUpdFrameOff size code
436 = do { info <- getInfoDown
437 ; withInfoDown code (info {cgd_updfr_off = size }) }
439 getUpdFrameOff :: FCode UpdFrameOffset
441 = do { info <- getInfoDown
442 ; return $ cgd_updfr_off info }
444 -- ----------------------------------------------------------------------------
445 -- Get/set the current ticky counter label
447 getTickyCtrLabel :: FCode CLabel
448 getTickyCtrLabel = do
450 return (cgd_ticky info)
452 setTickyCtrLabel :: CLabel -> FCode () -> FCode ()
453 setTickyCtrLabel ticky code = do
455 withInfoDown code (info {cgd_ticky = ticky})
458 --------------------------------------------------------
460 --------------------------------------------------------
462 forkClosureBody :: FCode () -> FCode ()
463 -- forkClosureBody takes a code, $c$, and compiles it in a
464 -- fresh environment, except that:
465 -- - compilation info and statics are passed in unchanged.
466 -- - local bindings are passed in unchanged
467 -- (it's up to the enclosed code to re-bind the
468 -- free variables to a field of the closure)
470 -- The current state is passed on completely unaltered, except that
471 -- C-- from the fork is incorporated.
473 forkClosureBody body_code
474 = do { info <- getInfoDown
475 ; us <- newUniqSupply
477 ; let body_info_down = info { cgd_sequel = initSequel
478 , cgd_updfr_off = initUpdFrameOff }
479 fork_state_in = (initCgState us) { cgs_binds = cgs_binds state }
481 = doFCode body_code body_info_down fork_state_in
482 ; setState $ state `addCodeBlocksFrom` fork_state_out }
484 forkStatics :: FCode a -> FCode a
485 -- @forkStatics@ $fc$ compiles $fc$ in an environment whose *statics* come
486 -- from the current *local bindings*, but which is otherwise freshly initialised.
487 -- The Abstract~C returned is attached to the current state, but the
488 -- bindings and usage information is otherwise unchanged.
489 forkStatics body_code
490 = do { info <- getInfoDown
491 ; us <- newUniqSupply
493 ; let rhs_info_down = info { cgd_statics = cgs_binds state
494 , cgd_sequel = initSequel
495 , cgd_updfr_off = initUpdFrameOff }
496 (result, fork_state_out) = doFCode body_code rhs_info_down
498 ; setState (state `addCodeBlocksFrom` fork_state_out)
501 forkProc :: FCode a -> FCode a
502 -- 'forkProc' takes a code and compiles it in the *current* environment,
503 -- returning the graph thus constructed.
505 -- The current environment is passed on completely unchanged to
506 -- the successor. In particular, any heap usage from the enclosed
507 -- code is discarded; it should deal with its own heap consumption
509 = do { info_down <- getInfoDown
510 ; us <- newUniqSupply
512 ; let info_down' = info_down -- { cgd_sequel = initSequel }
513 fork_state_in = (initCgState us) { cgs_binds = cgs_binds state }
514 (result, fork_state_out) = doFCode body_code info_down' fork_state_in
515 ; setState $ state `addCodeBlocksFrom` fork_state_out
518 codeOnly :: FCode () -> FCode ()
519 -- Emit any code from the inner thing into the outer thing
520 -- Do not affect anything else in the outer state
521 -- Used in almost-circular code to prevent false loop dependencies
523 = do { info_down <- getInfoDown
524 ; us <- newUniqSupply
526 ; let fork_state_in = (initCgState us) { cgs_binds = cgs_binds state,
527 cgs_hp_usg = cgs_hp_usg state }
528 ((), fork_state_out) = doFCode body_code info_down fork_state_in
529 ; setState $ state `addCodeBlocksFrom` fork_state_out }
531 forkAlts :: [FCode a] -> FCode [a]
532 -- (forkAlts' bs d) takes fcodes 'bs' for the branches of a 'case', and
533 -- an fcode for the default case 'd', and compiles each in the current
534 -- environment. The current environment is passed on unmodified, except
535 -- that the virtual Hp is moved on to the worst virtual Hp for the branches
537 forkAlts branch_fcodes
538 = do { info_down <- getInfoDown
539 ; us <- newUniqSupply
541 ; let compile us branch
542 = (us2, doFCode branch info_down branch_state)
544 (us1,us2) = splitUniqSupply us
545 branch_state = (initCgState us1) {
546 cgs_binds = cgs_binds state,
547 cgs_hp_usg = cgs_hp_usg state }
549 (_us, results) = mapAccumL compile us branch_fcodes
550 (branch_results, branch_out_states) = unzip results
551 ; setState $ foldl stateIncUsage state branch_out_states
552 -- NB foldl. state is the *left* argument to stateIncUsage
553 ; return branch_results }
555 -- collect the code emitted by an FCode computation
556 getCodeR :: FCode a -> FCode (a, CmmAGraph)
558 = do { state1 <- getState
559 ; (a, state2) <- withState fcode (state1 { cgs_stmts = mkNop })
560 ; setState $ state2 { cgs_stmts = cgs_stmts state1 }
561 ; return (a, cgs_stmts state2) }
563 getCode :: FCode a -> FCode CmmAGraph
564 getCode fcode = do { (_,stmts) <- getCodeR fcode; return stmts }
566 -- 'getHeapUsage' applies a function to the amount of heap that it uses.
567 -- It initialises the heap usage to zeros, and passes on an unchanged
570 -- It is usually a prelude to performing a GC check, so everything must
571 -- be in a tidy and consistent state.
573 -- Note the slightly subtle fixed point behaviour needed here
575 getHeapUsage :: (VirtualHpOffset -> FCode a) -> FCode a
577 = do { info_down <- getInfoDown
579 ; let fstate_in = state { cgs_hp_usg = initHpUsage }
580 (r, fstate_out) = doFCode (fcode hp_hw) info_down fstate_in
581 hp_hw = heapHWM (cgs_hp_usg fstate_out) -- Loop here!
583 ; setState $ fstate_out { cgs_hp_usg = cgs_hp_usg state }
586 -- ----------------------------------------------------------------------------
587 -- Combinators for emitting code
589 emit :: CmmAGraph -> FCode ()
591 = do { state <- getState
592 ; setState $ state { cgs_stmts = cgs_stmts state <*> ag } }
594 emitData :: Section -> [CmmStatic] -> FCode ()
596 = do { state <- getState
597 ; setState $ state { cgs_tops = cgs_tops state `snocOL` data_block } }
599 data_block = CmmData sect lits
601 emitProcWithConvention :: Convention -> CmmInfo -> CLabel -> CmmFormals ->
602 CmmAGraph -> FCode ()
603 emitProcWithConvention conv info lbl args blocks
604 = do { us <- newUniqSupply
605 ; let (offset, entry) = mkEntry (mkBlockId $ uniqFromSupply us) conv args
606 blks = initUs_ us $ lgraphOfAGraph $ entry <*> blocks
607 ; let proc_block = CmmProc info lbl args ((offset, Just initUpdFrameOff), blks)
609 ; setState $ state { cgs_tops = cgs_tops state `snocOL` proc_block } }
611 emitProc :: CmmInfo -> CLabel -> CmmFormals -> CmmAGraph -> FCode ()
612 emitProc = emitProcWithConvention NativeNodeCall
614 emitSimpleProc :: CLabel -> CmmAGraph -> FCode ()
615 emitSimpleProc lbl code =
616 emitProc (CmmInfo Nothing Nothing CmmNonInfoTable) lbl [] code
618 getCmm :: FCode () -> FCode CmmZ
619 -- Get all the CmmTops (there should be no stmts)
620 -- Return a single Cmm which may be split from other Cmms by
621 -- object splitting (at a later stage)
623 = do { state1 <- getState
624 ; ((), state2) <- withState code (state1 { cgs_tops = nilOL })
625 ; setState $ state2 { cgs_tops = cgs_tops state1 }
626 ; return (Cmm (fromOL (cgs_tops state2))) }
628 -- ----------------------------------------------------------------------------
631 -- These functions deal in terms of CgStmts, which is an abstract type
632 -- representing the code in the current proc.
634 -- turn CgStmts into [CmmBasicBlock], for making a new proc.
635 cgStmtsToBlocks :: CmmAGraph -> FCode CmmGraph
636 cgStmtsToBlocks stmts
637 = do { us <- newUniqSupply
638 ; return (initUs_ us (lgraphOfAGraph stmts)) }