2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 \section[StgSyn]{Shared term graph (STG) syntax for spineless-tagless code generation}
6 This data type represents programs just before code generation
7 (conversion to @AbstractC@): basically, what we have is a stylised
8 form of @CoreSyntax@, the style being one that happens to be ideally
9 suited to spineless tagless code generation.
16 GenStgBinding(..), GenStgExpr(..), GenStgRhs(..),
17 GenStgAlt, AltType(..),
19 UpdateFlag(..), isUpdatable,
22 noBinderInfo, stgSatOcc, stgUnsatOcc, satCallsOnly,
25 -- a set of synonyms for the most common (only :-) parameterisation
27 StgBinding, StgExpr, StgRhs, StgAlt,
36 stgBindHasCafRefs, stgArgHasCafRefs, stgRhsArity,
37 isDllConApp, isStgTypeArg,
40 pprStgBinding, pprStgBindings, pprStgBindingsWithSRTs
47 #include "HsVersions.h"
49 import CostCentre ( CostCentreStack, CostCentre )
50 import VarSet ( IdSet, isEmptyVarSet )
52 import Id ( Id, idName, idType, idCafInfo )
53 import IdInfo ( mayHaveCafRefs )
54 import Packages ( isDllName )
55 import PackageConfig ( PackageId )
56 import Literal ( Literal, literalType )
57 import ForeignCall ( ForeignCall )
58 import DataCon ( DataCon, dataConName )
59 import CoreSyn ( AltCon )
60 import PprCore ( {- instances -} )
61 import PrimOp ( PrimOp )
65 import TyCon ( TyCon )
66 import UniqSet ( isEmptyUniqSet, uniqSetToList, UniqSet )
67 import Unique ( Unique )
69 import StaticFlags ( opt_SccProfilingOn )
70 import Module ( Module, pprModule )
73 %************************************************************************
75 \subsection{@GenStgBinding@}
77 %************************************************************************
79 As usual, expressions are interesting; other things are boring. Here
80 are the boring things [except note the @GenStgRhs@], parameterised
81 with respect to binder and occurrence information (just as in
84 There is one SRT for each group of bindings.
87 data GenStgBinding bndr occ
88 = StgNonRec bndr (GenStgRhs bndr occ)
89 | StgRec [(bndr, GenStgRhs bndr occ)]
92 %************************************************************************
94 \subsection{@GenStgArg@}
96 %************************************************************************
102 | StgTypeArg Type -- For when we want to preserve all type info
106 isStgTypeArg (StgTypeArg _) = True
107 isStgTypeArg other = False
109 isDllArg :: PackageId -> StgArg -> Bool
110 -- Does this argument refer to something in a different DLL?
111 isDllArg this_pkg (StgTypeArg v) = False
112 isDllArg this_pkg (StgVarArg v) = isDllName this_pkg (idName v)
113 isDllArg this_pkg (StgLitArg lit) = False
115 isDllConApp :: PackageId -> DataCon -> [StgArg] -> Bool
116 -- Does this constructor application refer to
117 -- anything in a different DLL?
118 -- If so, we can't allocate it statically
119 isDllConApp this_pkg con args
120 = isDllName this_pkg (dataConName con) || any (isDllArg this_pkg) args
122 stgArgType :: StgArg -> Type
123 -- Very half baked becase we have lost the type arguments
124 stgArgType (StgVarArg v) = idType v
125 stgArgType (StgLitArg lit) = literalType lit
126 stgArgType (StgTypeArg lit) = panic "stgArgType called on stgTypeArg"
129 %************************************************************************
131 \subsection{STG expressions}
133 %************************************************************************
135 The @GenStgExpr@ data type is parameterised on binder and occurrence
138 %************************************************************************
140 \subsubsection{@GenStgExpr@ application}
142 %************************************************************************
144 An application is of a function to a list of atoms [not expressions].
145 Operationally, we want to push the arguments on the stack and call the
146 function. (If the arguments were expressions, we would have to build
147 their closures first.)
149 There is no constructor for a lone variable; it would appear as
152 type GenStgLiveVars occ = UniqSet occ
154 data GenStgExpr bndr occ
157 [GenStgArg occ] -- arguments; may be empty
160 %************************************************************************
162 \subsubsection{@StgConApp@ and @StgPrimApp@---saturated applications}
164 %************************************************************************
166 There are a specialised forms of application, for
167 constructors, primitives, and literals.
172 [GenStgArg occ] -- Saturated
174 | StgOpApp StgOp -- Primitive op or foreign call
175 [GenStgArg occ] -- Saturated
176 Type -- Result type; we need to know the result type
177 -- so that we can assign result registers.
180 %************************************************************************
182 \subsubsection{@StgLam@}
184 %************************************************************************
186 StgLam is used *only* during CoreToStg's work. Before CoreToStg has finished
187 it encodes (\x -> e) as (let f = \x -> e in f)
191 Type -- Type of whole lambda (useful when making a binder for it)
193 StgExpr -- Body of lambda
197 %************************************************************************
199 \subsubsection{@GenStgExpr@: case-expressions}
201 %************************************************************************
203 This has the same boxed/unboxed business as Core case expressions.
206 (GenStgExpr bndr occ)
207 -- the thing to examine
209 (GenStgLiveVars occ) -- Live vars of whole case expression,
210 -- plus everything that happens after the case
211 -- i.e., those which mustn't be overwritten
213 (GenStgLiveVars occ) -- Live vars of RHSs (plus what happens afterwards)
214 -- i.e., those which must be saved before eval.
216 -- note that an alt's constructor's
217 -- binder-variables are NOT counted in the
218 -- free vars for the alt's RHS
220 bndr -- binds the result of evaluating the scrutinee
222 SRT -- The SRT for the continuation
226 [GenStgAlt bndr occ] -- The DEFAULT case is always *first*
227 -- if it is there at all
230 %************************************************************************
232 \subsubsection{@GenStgExpr@: @let(rec)@-expressions}
234 %************************************************************************
236 The various forms of let(rec)-expression encode most of the
237 interesting things we want to do.
241 let-closure x = [free-vars] expr [args]
246 let x = (\free-vars -> \args -> expr) free-vars
248 \tr{args} may be empty (and is for most closures). It isn't under
249 circumstances like this:
255 let-closure x = [z] [y] (y+z)
257 The idea is that we compile code for @(y+z)@ in an environment in which
258 @z@ is bound to an offset from \tr{Node}, and @y@ is bound to an
259 offset from the stack pointer.
261 (A let-closure is an @StgLet@ with a @StgRhsClosure@ RHS.)
265 let-constructor x = Constructor [args]
269 (A let-constructor is an @StgLet@ with a @StgRhsCon@ RHS.)
272 Letrec-expressions are essentially the same deal as
273 let-closure/let-constructor, so we use a common structure and
274 distinguish between them with an @is_recursive@ boolean flag.
278 let-unboxed u = an arbitrary arithmetic expression in unboxed values
281 All the stuff on the RHS must be fully evaluated. No function calls either!
283 (We've backed away from this toward case-expressions with
284 suitably-magical alts ...)
287 ~[Advanced stuff here! Not to start with, but makes pattern matching
288 generate more efficient code.]
291 let-escapes-not fail = expr
294 Here the idea is that @e'@ guarantees not to put @fail@ in a data structure,
295 or pass it to another function. All @e'@ will ever do is tail-call @fail@.
296 Rather than build a closure for @fail@, all we need do is to record the stack
297 level at the moment of the @let-escapes-not@; then entering @fail@ is just
298 a matter of adjusting the stack pointer back down to that point and entering
303 f x y = let z = huge-expression in
309 (A let-escapes-not is an @StgLetNoEscape@.)
312 We may eventually want:
314 let-literal x = Literal
318 (ToDo: is this obsolete?)
321 And so the code for let(rec)-things:
324 (GenStgBinding bndr occ) -- right hand sides (see below)
325 (GenStgExpr bndr occ) -- body
327 | StgLetNoEscape -- remember: ``advanced stuff''
328 (GenStgLiveVars occ) -- Live in the whole let-expression
329 -- Mustn't overwrite these stack slots
330 -- *Doesn't* include binders of the let(rec).
332 (GenStgLiveVars occ) -- Live in the right hand sides (only)
333 -- These are the ones which must be saved on
334 -- the stack if they aren't there already
335 -- *Does* include binders of the let(rec) if recursive.
337 (GenStgBinding bndr occ) -- right hand sides (see below)
338 (GenStgExpr bndr occ) -- body
341 %************************************************************************
343 \subsubsection{@GenStgExpr@: @scc@ expressions}
345 %************************************************************************
347 Finally for @scc@ expressions we introduce a new STG construct.
351 CostCentre -- label of SCC expression
352 (GenStgExpr bndr occ) -- scc expression
355 %************************************************************************
357 \subsubsection{@GenStgExpr@: @hpc@ expressions}
359 %************************************************************************
361 Finally for @scc@ expressions we introduce a new STG construct.
365 Module -- the module of the source of this tick
367 (GenStgExpr bndr occ) -- sub expression
371 %************************************************************************
373 \subsection{STG right-hand sides}
375 %************************************************************************
377 Here's the rest of the interesting stuff for @StgLet@s; the first
378 flavour is for closures:
380 data GenStgRhs bndr occ
382 CostCentreStack -- CCS to be attached (default is CurrentCCS)
383 StgBinderInfo -- Info about how this binder is used (see below)
384 [occ] -- non-global free vars; a list, rather than
385 -- a set, because order is important
386 !UpdateFlag -- ReEntrant | Updatable | SingleEntry
387 SRT -- The SRT reference
388 [bndr] -- arguments; if empty, then not a function;
389 -- as above, order is important.
390 (GenStgExpr bndr occ) -- body
392 An example may be in order. Consider:
394 let t = \x -> \y -> ... x ... y ... p ... q in e
396 Pulling out the free vars and stylising somewhat, we get the equivalent:
398 let t = (\[p,q] -> \[x,y] -> ... x ... y ... p ...q) p q
400 Stg-operationally, the @[x,y]@ are on the stack, the @[p,q]@ are
401 offsets from @Node@ into the closure, and the code ptr for the closure
402 will be exactly that in parentheses above.
404 The second flavour of right-hand-side is for constructors (simple but important):
407 CostCentreStack -- CCS to be attached (default is CurrentCCS).
408 -- Top-level (static) ones will end up with
409 -- DontCareCCS, because we don't count static
410 -- data in heap profiles, and we don't set CCCS
411 -- from static closure.
412 DataCon -- constructor
413 [GenStgArg occ] -- args
417 stgRhsArity :: StgRhs -> Int
418 stgRhsArity (StgRhsClosure _ _ _ _ _ bndrs _) = count isId bndrs
419 -- The arity never includes type parameters, so
420 -- when keeping type arguments and binders in the Stg syntax
421 -- (opt_RuntimeTypes) we have to fliter out the type binders.
422 stgRhsArity (StgRhsCon _ _ _) = 0
426 stgBindHasCafRefs :: GenStgBinding bndr Id -> Bool
427 stgBindHasCafRefs (StgNonRec _ rhs) = rhsHasCafRefs rhs
428 stgBindHasCafRefs (StgRec binds) = any rhsHasCafRefs (map snd binds)
430 rhsHasCafRefs (StgRhsClosure _ _ _ upd srt _ _)
431 = isUpdatable upd || nonEmptySRT srt
432 rhsHasCafRefs (StgRhsCon _ _ args)
433 = any stgArgHasCafRefs args
435 stgArgHasCafRefs (StgVarArg id) = mayHaveCafRefs (idCafInfo id)
436 stgArgHasCafRefs _ = False
439 Here's the @StgBinderInfo@ type, and its combining op:
443 | SatCallsOnly -- All occurrences are *saturated* *function* calls
444 -- This means we don't need to build an info table and
445 -- slow entry code for the thing
446 -- Thunks never get this value
448 noBinderInfo = NoStgBinderInfo
449 stgUnsatOcc = NoStgBinderInfo
450 stgSatOcc = SatCallsOnly
452 satCallsOnly :: StgBinderInfo -> Bool
453 satCallsOnly SatCallsOnly = True
454 satCallsOnly NoStgBinderInfo = False
456 combineStgBinderInfo :: StgBinderInfo -> StgBinderInfo -> StgBinderInfo
457 combineStgBinderInfo SatCallsOnly SatCallsOnly = SatCallsOnly
458 combineStgBinderInfo info1 info2 = NoStgBinderInfo
461 pp_binder_info NoStgBinderInfo = empty
462 pp_binder_info SatCallsOnly = ptext SLIT("sat-only")
465 %************************************************************************
467 \subsection[Stg-case-alternatives]{STG case alternatives}
469 %************************************************************************
471 Very like in @CoreSyntax@ (except no type-world stuff).
473 The type constructor is guaranteed not to be abstract; that is, we can
474 see its representation. This is important because the code generator
475 uses it to determine return conventions etc. But it's not trivial
476 where there's a moduule loop involved, because some versions of a type
477 constructor might not have all the constructors visible. So
478 mkStgAlgAlts (in CoreToStg) ensures that it gets the TyCon from the
479 constructors or literals (which are guaranteed to have the Real McCoy)
480 rather than from the scrutinee type.
483 type GenStgAlt bndr occ
484 = (AltCon, -- alts: data constructor,
485 [bndr], -- constructor's parameters,
486 [Bool], -- "use mask", same length as
487 -- parameters; a True in a
488 -- param's position if it is
490 GenStgExpr bndr occ) -- ...right-hand side.
493 = PolyAlt -- Polymorphic (a type variable)
494 | UbxTupAlt TyCon -- Unboxed tuple
495 | AlgAlt TyCon -- Algebraic data type; the AltCons will be DataAlts
496 | PrimAlt TyCon -- Primitive data type; the AltCons will be LitAlts
499 %************************************************************************
501 \subsection[Stg]{The Plain STG parameterisation}
503 %************************************************************************
505 This happens to be the only one we use at the moment.
508 type StgBinding = GenStgBinding Id Id
509 type StgArg = GenStgArg Id
510 type StgLiveVars = GenStgLiveVars Id
511 type StgExpr = GenStgExpr Id Id
512 type StgRhs = GenStgRhs Id Id
513 type StgAlt = GenStgAlt Id Id
516 %************************************************************************
518 \subsubsection[UpdateFlag-datatype]{@UpdateFlag@}
520 %************************************************************************
522 This is also used in @LambdaFormInfo@ in the @ClosureInfo@ module.
524 A @ReEntrant@ closure may be entered multiple times, but should not be
525 updated or blackholed. An @Updatable@ closure should be updated after
526 evaluation (and may be blackholed during evaluation). A @SingleEntry@
527 closure will only be entered once, and so need not be updated but may
528 safely be blackholed.
531 data UpdateFlag = ReEntrant | Updatable | SingleEntry
533 instance Outputable UpdateFlag where
535 = char (case u of { ReEntrant -> 'r'; Updatable -> 'u'; SingleEntry -> 's' })
537 isUpdatable ReEntrant = False
538 isUpdatable SingleEntry = False
539 isUpdatable Updatable = True
542 %************************************************************************
544 \subsubsection{StgOp}
546 %************************************************************************
548 An StgOp allows us to group together PrimOps and ForeignCalls.
549 It's quite useful to move these around together, notably
550 in StgOpApp and COpStmt.
553 data StgOp = StgPrimOp PrimOp
555 | StgFCallOp ForeignCall Unique
556 -- The Unique is occasionally needed by the C pretty-printer
557 -- (which lacks a unique supply), notably when generating a
558 -- typedef for foreign-export-dynamic
562 %************************************************************************
564 \subsubsection[Static Reference Tables]{@SRT@}
566 %************************************************************************
568 There is one SRT per top-level function group. Each local binding and
569 case expression within this binding group has a subrange of the whole
570 SRT, expressed as an offset and length.
572 In CoreToStg we collect the list of CafRefs at each SRT site, which is later
573 converted into the length and offset form by the SRT pass.
578 -- generated by CoreToStg
579 | SRT !Int{-offset-} !Int{-length-} !Bitmap{-bitmap-}
580 -- generated by computeSRTs
585 nonEmptySRT NoSRT = False
586 nonEmptySRT (SRTEntries vs) = not (isEmptyVarSet vs)
589 pprSRT (NoSRT) = ptext SLIT("_no_srt_")
590 pprSRT (SRTEntries ids) = text "SRT:" <> ppr ids
591 pprSRT (SRT off length bitmap) = parens (ppr off <> comma <> text "*bitmap*")
594 %************************************************************************
596 \subsection[Stg-pretty-printing]{Pretty-printing}
598 %************************************************************************
600 Robin Popplestone asked for semi-colon separators on STG binds; here's
601 hoping he likes terminators instead... Ditto for case alternatives.
604 pprGenStgBinding :: (Outputable bndr, Outputable bdee, Ord bdee)
605 => GenStgBinding bndr bdee -> SDoc
607 pprGenStgBinding (StgNonRec bndr rhs)
608 = hang (hsep [ppr bndr, equals])
609 4 ((<>) (ppr rhs) semi)
611 pprGenStgBinding (StgRec pairs)
612 = vcat ((ifPprDebug (ptext SLIT("{- StgRec (begin) -}"))) :
613 (map (ppr_bind) pairs) ++ [(ifPprDebug (ptext SLIT("{- StgRec (end) -}")))])
615 ppr_bind (bndr, expr)
616 = hang (hsep [ppr bndr, equals])
617 4 ((<>) (ppr expr) semi)
619 pprStgBinding :: StgBinding -> SDoc
620 pprStgBinding bind = pprGenStgBinding bind
622 pprStgBindings :: [StgBinding] -> SDoc
623 pprStgBindings binds = vcat (map pprGenStgBinding binds)
625 pprGenStgBindingWithSRT
626 :: (Outputable bndr, Outputable bdee, Ord bdee)
627 => (GenStgBinding bndr bdee,[(Id,[Id])]) -> SDoc
629 pprGenStgBindingWithSRT (bind,srts)
630 = vcat (pprGenStgBinding bind : map pprSRT srts)
631 where pprSRT (id,srt) =
632 ptext SLIT("SRT") <> parens (ppr id) <> ptext SLIT(": ") <> ppr srt
634 pprStgBindingsWithSRTs :: [(StgBinding,[(Id,[Id])])] -> SDoc
635 pprStgBindingsWithSRTs binds = vcat (map pprGenStgBindingWithSRT binds)
639 instance (Outputable bdee) => Outputable (GenStgArg bdee) where
642 instance (Outputable bndr, Outputable bdee, Ord bdee)
643 => Outputable (GenStgBinding bndr bdee) where
644 ppr = pprGenStgBinding
646 instance (Outputable bndr, Outputable bdee, Ord bdee)
647 => Outputable (GenStgExpr bndr bdee) where
650 instance (Outputable bndr, Outputable bdee, Ord bdee)
651 => Outputable (GenStgRhs bndr bdee) where
652 ppr rhs = pprStgRhs rhs
656 pprStgArg :: (Outputable bdee) => GenStgArg bdee -> SDoc
658 pprStgArg (StgVarArg var) = ppr var
659 pprStgArg (StgLitArg con) = ppr con
660 pprStgArg (StgTypeArg ty) = char '@' <+> ppr ty
664 pprStgExpr :: (Outputable bndr, Outputable bdee, Ord bdee)
665 => GenStgExpr bndr bdee -> SDoc
667 pprStgExpr (StgLit lit) = ppr lit
670 pprStgExpr (StgApp func args)
672 4 (sep (map (ppr) args))
676 pprStgExpr (StgConApp con args)
677 = hsep [ ppr con, brackets (interppSP args)]
679 pprStgExpr (StgOpApp op args _)
680 = hsep [ pprStgOp op, brackets (interppSP args)]
682 pprStgExpr (StgLam _ bndrs body)
683 =sep [ char '\\' <+> ppr bndrs <+> ptext SLIT("->"),
688 -- special case: let v = <very specific thing>
694 -- Very special! Suspicious! (SLPJ)
697 pprStgExpr (StgLet srt (StgNonRec bndr (StgRhsClosure cc bi free_vars upd_flag args rhs))
700 (hang (hcat [ptext SLIT("let { "), ppr bndr, ptext SLIT(" = "),
703 ptext SLIT(" ["), ifPprDebug (interppSP free_vars), ptext SLIT("] \\"),
704 ppr upd_flag, ptext SLIT(" ["),
705 interppSP args, char ']'])
706 8 (sep [hsep [ppr rhs, ptext SLIT("} in")]]))
710 -- special case: let ... in let ...
712 pprStgExpr (StgLet bind expr@(StgLet _ _))
714 (sep [hang (ptext SLIT("let {"))
715 2 (hsep [pprGenStgBinding bind, ptext SLIT("} in")])])
719 pprStgExpr (StgLet bind expr)
720 = sep [hang (ptext SLIT("let {")) 2 (pprGenStgBinding bind),
721 hang (ptext SLIT("} in ")) 2 (ppr expr)]
723 pprStgExpr (StgLetNoEscape lvs_whole lvs_rhss bind expr)
724 = sep [hang (ptext SLIT("let-no-escape {"))
725 2 (pprGenStgBinding bind),
726 hang ((<>) (ptext SLIT("} in "))
729 hcat [ptext SLIT("-- lvs: ["), interppSP (uniqSetToList lvs_whole),
730 ptext SLIT("]; rhs lvs: ["), interppSP (uniqSetToList lvs_rhss),
734 pprStgExpr (StgSCC cc expr)
735 = sep [ hsep [ptext SLIT("_scc_"), ppr cc],
738 pprStgExpr (StgTick m n expr)
739 = sep [ hsep [ptext SLIT("_tick_"), pprModule m,text (show n)],
742 pprStgExpr (StgCase expr lvs_whole lvs_rhss bndr srt alt_type alts)
743 = sep [sep [ptext SLIT("case"),
744 nest 4 (hsep [pprStgExpr expr,
745 ifPprDebug (dcolon <+> ppr alt_type)]),
746 ptext SLIT("of"), ppr bndr, char '{'],
749 hcat [ptext SLIT("-- lvs: ["), interppSP (uniqSetToList lvs_whole),
750 ptext SLIT("]; rhs lvs: ["), interppSP (uniqSetToList lvs_rhss),
753 nest 2 (vcat (map pprStgAlt alts)),
756 pprStgAlt (con, params, use_mask, expr)
757 = hang (hsep [ppr con, interppSP params, ptext SLIT("->")])
760 pprStgOp (StgPrimOp op) = ppr op
761 pprStgOp (StgFCallOp op _) = ppr op
763 instance Outputable AltType where
764 ppr PolyAlt = ptext SLIT("Polymorphic")
765 ppr (UbxTupAlt tc) = ptext SLIT("UbxTup") <+> ppr tc
766 ppr (AlgAlt tc) = ptext SLIT("Alg") <+> ppr tc
767 ppr (PrimAlt tc) = ptext SLIT("Prim") <+> ppr tc
771 pprStgLVs :: Outputable occ => GenStgLiveVars occ -> SDoc
773 = getPprStyle $ \ sty ->
774 if userStyle sty || isEmptyUniqSet lvs then
777 hcat [text "{-lvs:", interpp'SP (uniqSetToList lvs), text "-}"]
781 pprStgRhs :: (Outputable bndr, Outputable bdee, Ord bdee)
782 => GenStgRhs bndr bdee -> SDoc
785 pprStgRhs (StgRhsClosure cc bi [free_var] upd_flag srt [{-no args-}] (StgApp func []))
788 brackets (ifPprDebug (ppr free_var)),
789 ptext SLIT(" \\"), ppr upd_flag, pprMaybeSRT srt, ptext SLIT(" [] "), ppr func ]
792 pprStgRhs (StgRhsClosure cc bi free_vars upd_flag srt args body)
793 = hang (hsep [if opt_SccProfilingOn then ppr cc else empty,
795 ifPprDebug (brackets (interppSP free_vars)),
796 char '\\' <> ppr upd_flag, pprMaybeSRT srt, brackets (interppSP args)])
799 pprStgRhs (StgRhsCon cc con args)
801 space, ppr con, ptext SLIT("! "), brackets (interppSP args)]
803 pprMaybeSRT (NoSRT) = empty
804 pprMaybeSRT srt = ptext SLIT("srt:") <> pprSRT srt