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.
171 -- StgConApp is vital for returning unboxed tuples
172 -- which can't be let-bound first
174 [GenStgArg occ] -- Saturated
176 | StgOpApp StgOp -- Primitive op or foreign call
177 [GenStgArg occ] -- Saturated
179 -- We need to know this so that we can
180 -- assign result registers
183 %************************************************************************
185 \subsubsection{@StgLam@}
187 %************************************************************************
189 StgLam is used *only* during CoreToStg's work. Before CoreToStg has finished
190 it encodes (\x -> e) as (let f = \x -> e in f)
194 Type -- Type of whole lambda (useful when making a binder for it)
196 StgExpr -- Body of lambda
200 %************************************************************************
202 \subsubsection{@GenStgExpr@: case-expressions}
204 %************************************************************************
206 This has the same boxed/unboxed business as Core case expressions.
209 (GenStgExpr bndr occ)
210 -- the thing to examine
212 (GenStgLiveVars occ) -- Live vars of whole case expression,
213 -- plus everything that happens after the case
214 -- i.e., those which mustn't be overwritten
216 (GenStgLiveVars occ) -- Live vars of RHSs (plus what happens afterwards)
217 -- i.e., those which must be saved before eval.
219 -- note that an alt's constructor's
220 -- binder-variables are NOT counted in the
221 -- free vars for the alt's RHS
223 bndr -- binds the result of evaluating the scrutinee
225 SRT -- The SRT for the continuation
229 [GenStgAlt bndr occ] -- The DEFAULT case is always *first*
230 -- if it is there at all
233 %************************************************************************
235 \subsubsection{@GenStgExpr@: @let(rec)@-expressions}
237 %************************************************************************
239 The various forms of let(rec)-expression encode most of the
240 interesting things we want to do.
244 let-closure x = [free-vars] expr [args]
249 let x = (\free-vars -> \args -> expr) free-vars
251 \tr{args} may be empty (and is for most closures). It isn't under
252 circumstances like this:
258 let-closure x = [z] [y] (y+z)
260 The idea is that we compile code for @(y+z)@ in an environment in which
261 @z@ is bound to an offset from \tr{Node}, and @y@ is bound to an
262 offset from the stack pointer.
264 (A let-closure is an @StgLet@ with a @StgRhsClosure@ RHS.)
268 let-constructor x = Constructor [args]
272 (A let-constructor is an @StgLet@ with a @StgRhsCon@ RHS.)
275 Letrec-expressions are essentially the same deal as
276 let-closure/let-constructor, so we use a common structure and
277 distinguish between them with an @is_recursive@ boolean flag.
281 let-unboxed u = an arbitrary arithmetic expression in unboxed values
284 All the stuff on the RHS must be fully evaluated. No function calls either!
286 (We've backed away from this toward case-expressions with
287 suitably-magical alts ...)
290 ~[Advanced stuff here! Not to start with, but makes pattern matching
291 generate more efficient code.]
294 let-escapes-not fail = expr
297 Here the idea is that @e'@ guarantees not to put @fail@ in a data structure,
298 or pass it to another function. All @e'@ will ever do is tail-call @fail@.
299 Rather than build a closure for @fail@, all we need do is to record the stack
300 level at the moment of the @let-escapes-not@; then entering @fail@ is just
301 a matter of adjusting the stack pointer back down to that point and entering
306 f x y = let z = huge-expression in
312 (A let-escapes-not is an @StgLetNoEscape@.)
315 We may eventually want:
317 let-literal x = Literal
321 (ToDo: is this obsolete?)
324 And so the code for let(rec)-things:
327 (GenStgBinding bndr occ) -- right hand sides (see below)
328 (GenStgExpr bndr occ) -- body
330 | StgLetNoEscape -- remember: ``advanced stuff''
331 (GenStgLiveVars occ) -- Live in the whole let-expression
332 -- Mustn't overwrite these stack slots
333 -- *Doesn't* include binders of the let(rec).
335 (GenStgLiveVars occ) -- Live in the right hand sides (only)
336 -- These are the ones which must be saved on
337 -- the stack if they aren't there already
338 -- *Does* include binders of the let(rec) if recursive.
340 (GenStgBinding bndr occ) -- right hand sides (see below)
341 (GenStgExpr bndr occ) -- body
344 %************************************************************************
346 \subsubsection{@GenStgExpr@: @scc@ expressions}
348 %************************************************************************
350 Finally for @scc@ expressions we introduce a new STG construct.
354 CostCentre -- label of SCC expression
355 (GenStgExpr bndr occ) -- scc expression
358 %************************************************************************
360 \subsubsection{@GenStgExpr@: @hpc@ expressions}
362 %************************************************************************
364 Finally for @scc@ expressions we introduce a new STG construct.
368 Module -- the module of the source of this tick
370 (GenStgExpr bndr occ) -- sub expression
374 %************************************************************************
376 \subsection{STG right-hand sides}
378 %************************************************************************
380 Here's the rest of the interesting stuff for @StgLet@s; the first
381 flavour is for closures:
383 data GenStgRhs bndr occ
385 CostCentreStack -- CCS to be attached (default is CurrentCCS)
386 StgBinderInfo -- Info about how this binder is used (see below)
387 [occ] -- non-global free vars; a list, rather than
388 -- a set, because order is important
389 !UpdateFlag -- ReEntrant | Updatable | SingleEntry
390 SRT -- The SRT reference
391 [bndr] -- arguments; if empty, then not a function;
392 -- as above, order is important.
393 (GenStgExpr bndr occ) -- body
395 An example may be in order. Consider:
397 let t = \x -> \y -> ... x ... y ... p ... q in e
399 Pulling out the free vars and stylising somewhat, we get the equivalent:
401 let t = (\[p,q] -> \[x,y] -> ... x ... y ... p ...q) p q
403 Stg-operationally, the @[x,y]@ are on the stack, the @[p,q]@ are
404 offsets from @Node@ into the closure, and the code ptr for the closure
405 will be exactly that in parentheses above.
407 The second flavour of right-hand-side is for constructors (simple but important):
410 CostCentreStack -- CCS to be attached (default is CurrentCCS).
411 -- Top-level (static) ones will end up with
412 -- DontCareCCS, because we don't count static
413 -- data in heap profiles, and we don't set CCCS
414 -- from static closure.
415 DataCon -- constructor
416 [GenStgArg occ] -- args
420 stgRhsArity :: StgRhs -> Int
421 stgRhsArity (StgRhsClosure _ _ _ _ _ bndrs _) = count isId bndrs
422 -- The arity never includes type parameters, so
423 -- when keeping type arguments and binders in the Stg syntax
424 -- (opt_RuntimeTypes) we have to fliter out the type binders.
425 stgRhsArity (StgRhsCon _ _ _) = 0
429 stgBindHasCafRefs :: GenStgBinding bndr Id -> Bool
430 stgBindHasCafRefs (StgNonRec _ rhs) = rhsHasCafRefs rhs
431 stgBindHasCafRefs (StgRec binds) = any rhsHasCafRefs (map snd binds)
433 rhsHasCafRefs (StgRhsClosure _ _ _ upd srt _ _)
434 = isUpdatable upd || nonEmptySRT srt
435 rhsHasCafRefs (StgRhsCon _ _ args)
436 = any stgArgHasCafRefs args
438 stgArgHasCafRefs (StgVarArg id) = mayHaveCafRefs (idCafInfo id)
439 stgArgHasCafRefs _ = False
442 Here's the @StgBinderInfo@ type, and its combining op:
446 | SatCallsOnly -- All occurrences are *saturated* *function* calls
447 -- This means we don't need to build an info table and
448 -- slow entry code for the thing
449 -- Thunks never get this value
451 noBinderInfo = NoStgBinderInfo
452 stgUnsatOcc = NoStgBinderInfo
453 stgSatOcc = SatCallsOnly
455 satCallsOnly :: StgBinderInfo -> Bool
456 satCallsOnly SatCallsOnly = True
457 satCallsOnly NoStgBinderInfo = False
459 combineStgBinderInfo :: StgBinderInfo -> StgBinderInfo -> StgBinderInfo
460 combineStgBinderInfo SatCallsOnly SatCallsOnly = SatCallsOnly
461 combineStgBinderInfo info1 info2 = NoStgBinderInfo
464 pp_binder_info NoStgBinderInfo = empty
465 pp_binder_info SatCallsOnly = ptext SLIT("sat-only")
468 %************************************************************************
470 \subsection[Stg-case-alternatives]{STG case alternatives}
472 %************************************************************************
474 Very like in @CoreSyntax@ (except no type-world stuff).
476 The type constructor is guaranteed not to be abstract; that is, we can
477 see its representation. This is important because the code generator
478 uses it to determine return conventions etc. But it's not trivial
479 where there's a moduule loop involved, because some versions of a type
480 constructor might not have all the constructors visible. So
481 mkStgAlgAlts (in CoreToStg) ensures that it gets the TyCon from the
482 constructors or literals (which are guaranteed to have the Real McCoy)
483 rather than from the scrutinee type.
486 type GenStgAlt bndr occ
487 = (AltCon, -- alts: data constructor,
488 [bndr], -- constructor's parameters,
489 [Bool], -- "use mask", same length as
490 -- parameters; a True in a
491 -- param's position if it is
493 GenStgExpr bndr occ) -- ...right-hand side.
496 = PolyAlt -- Polymorphic (a type variable)
497 | UbxTupAlt TyCon -- Unboxed tuple
498 | AlgAlt TyCon -- Algebraic data type; the AltCons will be DataAlts
499 | PrimAlt TyCon -- Primitive data type; the AltCons will be LitAlts
502 %************************************************************************
504 \subsection[Stg]{The Plain STG parameterisation}
506 %************************************************************************
508 This happens to be the only one we use at the moment.
511 type StgBinding = GenStgBinding Id Id
512 type StgArg = GenStgArg Id
513 type StgLiveVars = GenStgLiveVars Id
514 type StgExpr = GenStgExpr Id Id
515 type StgRhs = GenStgRhs Id Id
516 type StgAlt = GenStgAlt Id Id
519 %************************************************************************
521 \subsubsection[UpdateFlag-datatype]{@UpdateFlag@}
523 %************************************************************************
525 This is also used in @LambdaFormInfo@ in the @ClosureInfo@ module.
527 A @ReEntrant@ closure may be entered multiple times, but should not be
528 updated or blackholed. An @Updatable@ closure should be updated after
529 evaluation (and may be blackholed during evaluation). A @SingleEntry@
530 closure will only be entered once, and so need not be updated but may
531 safely be blackholed.
534 data UpdateFlag = ReEntrant | Updatable | SingleEntry
536 instance Outputable UpdateFlag where
538 = char (case u of { ReEntrant -> 'r'; Updatable -> 'u'; SingleEntry -> 's' })
540 isUpdatable ReEntrant = False
541 isUpdatable SingleEntry = False
542 isUpdatable Updatable = True
545 %************************************************************************
547 \subsubsection{StgOp}
549 %************************************************************************
551 An StgOp allows us to group together PrimOps and ForeignCalls.
552 It's quite useful to move these around together, notably
553 in StgOpApp and COpStmt.
556 data StgOp = StgPrimOp PrimOp
558 | StgFCallOp ForeignCall Unique
559 -- The Unique is occasionally needed by the C pretty-printer
560 -- (which lacks a unique supply), notably when generating a
561 -- typedef for foreign-export-dynamic
565 %************************************************************************
567 \subsubsection[Static Reference Tables]{@SRT@}
569 %************************************************************************
571 There is one SRT per top-level function group. Each local binding and
572 case expression within this binding group has a subrange of the whole
573 SRT, expressed as an offset and length.
575 In CoreToStg we collect the list of CafRefs at each SRT site, which is later
576 converted into the length and offset form by the SRT pass.
581 -- generated by CoreToStg
582 | SRT !Int{-offset-} !Int{-length-} !Bitmap{-bitmap-}
583 -- generated by computeSRTs
588 nonEmptySRT NoSRT = False
589 nonEmptySRT (SRTEntries vs) = not (isEmptyVarSet vs)
592 pprSRT (NoSRT) = ptext SLIT("_no_srt_")
593 pprSRT (SRTEntries ids) = text "SRT:" <> ppr ids
594 pprSRT (SRT off length bitmap) = parens (ppr off <> comma <> text "*bitmap*")
597 %************************************************************************
599 \subsection[Stg-pretty-printing]{Pretty-printing}
601 %************************************************************************
603 Robin Popplestone asked for semi-colon separators on STG binds; here's
604 hoping he likes terminators instead... Ditto for case alternatives.
607 pprGenStgBinding :: (Outputable bndr, Outputable bdee, Ord bdee)
608 => GenStgBinding bndr bdee -> SDoc
610 pprGenStgBinding (StgNonRec bndr rhs)
611 = hang (hsep [ppr bndr, equals])
612 4 ((<>) (ppr rhs) semi)
614 pprGenStgBinding (StgRec pairs)
615 = vcat ((ifPprDebug (ptext SLIT("{- StgRec (begin) -}"))) :
616 (map (ppr_bind) pairs) ++ [(ifPprDebug (ptext SLIT("{- StgRec (end) -}")))])
618 ppr_bind (bndr, expr)
619 = hang (hsep [ppr bndr, equals])
620 4 ((<>) (ppr expr) semi)
622 pprStgBinding :: StgBinding -> SDoc
623 pprStgBinding bind = pprGenStgBinding bind
625 pprStgBindings :: [StgBinding] -> SDoc
626 pprStgBindings binds = vcat (map pprGenStgBinding binds)
628 pprGenStgBindingWithSRT
629 :: (Outputable bndr, Outputable bdee, Ord bdee)
630 => (GenStgBinding bndr bdee,[(Id,[Id])]) -> SDoc
632 pprGenStgBindingWithSRT (bind,srts)
633 = vcat (pprGenStgBinding bind : map pprSRT srts)
634 where pprSRT (id,srt) =
635 ptext SLIT("SRT") <> parens (ppr id) <> ptext SLIT(": ") <> ppr srt
637 pprStgBindingsWithSRTs :: [(StgBinding,[(Id,[Id])])] -> SDoc
638 pprStgBindingsWithSRTs binds = vcat (map pprGenStgBindingWithSRT binds)
642 instance (Outputable bdee) => Outputable (GenStgArg bdee) where
645 instance (Outputable bndr, Outputable bdee, Ord bdee)
646 => Outputable (GenStgBinding bndr bdee) where
647 ppr = pprGenStgBinding
649 instance (Outputable bndr, Outputable bdee, Ord bdee)
650 => Outputable (GenStgExpr bndr bdee) where
653 instance (Outputable bndr, Outputable bdee, Ord bdee)
654 => Outputable (GenStgRhs bndr bdee) where
655 ppr rhs = pprStgRhs rhs
659 pprStgArg :: (Outputable bdee) => GenStgArg bdee -> SDoc
661 pprStgArg (StgVarArg var) = ppr var
662 pprStgArg (StgLitArg con) = ppr con
663 pprStgArg (StgTypeArg ty) = char '@' <+> ppr ty
667 pprStgExpr :: (Outputable bndr, Outputable bdee, Ord bdee)
668 => GenStgExpr bndr bdee -> SDoc
670 pprStgExpr (StgLit lit) = ppr lit
673 pprStgExpr (StgApp func args)
675 4 (sep (map (ppr) args))
679 pprStgExpr (StgConApp con args)
680 = hsep [ ppr con, brackets (interppSP args)]
682 pprStgExpr (StgOpApp op args _)
683 = hsep [ pprStgOp op, brackets (interppSP args)]
685 pprStgExpr (StgLam _ bndrs body)
686 =sep [ char '\\' <+> ppr bndrs <+> ptext SLIT("->"),
691 -- special case: let v = <very specific thing>
697 -- Very special! Suspicious! (SLPJ)
700 pprStgExpr (StgLet srt (StgNonRec bndr (StgRhsClosure cc bi free_vars upd_flag args rhs))
703 (hang (hcat [ptext SLIT("let { "), ppr bndr, ptext SLIT(" = "),
706 ptext SLIT(" ["), ifPprDebug (interppSP free_vars), ptext SLIT("] \\"),
707 ppr upd_flag, ptext SLIT(" ["),
708 interppSP args, char ']'])
709 8 (sep [hsep [ppr rhs, ptext SLIT("} in")]]))
713 -- special case: let ... in let ...
715 pprStgExpr (StgLet bind expr@(StgLet _ _))
717 (sep [hang (ptext SLIT("let {"))
718 2 (hsep [pprGenStgBinding bind, ptext SLIT("} in")])])
722 pprStgExpr (StgLet bind expr)
723 = sep [hang (ptext SLIT("let {")) 2 (pprGenStgBinding bind),
724 hang (ptext SLIT("} in ")) 2 (ppr expr)]
726 pprStgExpr (StgLetNoEscape lvs_whole lvs_rhss bind expr)
727 = sep [hang (ptext SLIT("let-no-escape {"))
728 2 (pprGenStgBinding bind),
729 hang ((<>) (ptext SLIT("} in "))
732 hcat [ptext SLIT("-- lvs: ["), interppSP (uniqSetToList lvs_whole),
733 ptext SLIT("]; rhs lvs: ["), interppSP (uniqSetToList lvs_rhss),
737 pprStgExpr (StgSCC cc expr)
738 = sep [ hsep [ptext SLIT("_scc_"), ppr cc],
741 pprStgExpr (StgTick m n expr)
742 = sep [ hsep [ptext SLIT("_tick_"), pprModule m,text (show n)],
745 pprStgExpr (StgCase expr lvs_whole lvs_rhss bndr srt alt_type alts)
746 = sep [sep [ptext SLIT("case"),
747 nest 4 (hsep [pprStgExpr expr,
748 ifPprDebug (dcolon <+> ppr alt_type)]),
749 ptext SLIT("of"), ppr bndr, char '{'],
752 hcat [ptext SLIT("-- lvs: ["), interppSP (uniqSetToList lvs_whole),
753 ptext SLIT("]; rhs lvs: ["), interppSP (uniqSetToList lvs_rhss),
756 nest 2 (vcat (map pprStgAlt alts)),
759 pprStgAlt (con, params, use_mask, expr)
760 = hang (hsep [ppr con, interppSP params, ptext SLIT("->")])
763 pprStgOp (StgPrimOp op) = ppr op
764 pprStgOp (StgFCallOp op _) = ppr op
766 instance Outputable AltType where
767 ppr PolyAlt = ptext SLIT("Polymorphic")
768 ppr (UbxTupAlt tc) = ptext SLIT("UbxTup") <+> ppr tc
769 ppr (AlgAlt tc) = ptext SLIT("Alg") <+> ppr tc
770 ppr (PrimAlt tc) = ptext SLIT("Prim") <+> ppr tc
774 pprStgLVs :: Outputable occ => GenStgLiveVars occ -> SDoc
776 = getPprStyle $ \ sty ->
777 if userStyle sty || isEmptyUniqSet lvs then
780 hcat [text "{-lvs:", interpp'SP (uniqSetToList lvs), text "-}"]
784 pprStgRhs :: (Outputable bndr, Outputable bdee, Ord bdee)
785 => GenStgRhs bndr bdee -> SDoc
788 pprStgRhs (StgRhsClosure cc bi [free_var] upd_flag srt [{-no args-}] (StgApp func []))
791 brackets (ifPprDebug (ppr free_var)),
792 ptext SLIT(" \\"), ppr upd_flag, pprMaybeSRT srt, ptext SLIT(" [] "), ppr func ]
795 pprStgRhs (StgRhsClosure cc bi free_vars upd_flag srt args body)
796 = hang (hsep [if opt_SccProfilingOn then ppr cc else empty,
798 ifPprDebug (brackets (interppSP free_vars)),
799 char '\\' <> ppr upd_flag, pprMaybeSRT srt, brackets (interppSP args)])
802 pprStgRhs (StgRhsCon cc con args)
804 space, ppr con, ptext SLIT("! "), brackets (interppSP args)]
806 pprMaybeSRT (NoSRT) = empty
807 pprMaybeSRT srt = ptext SLIT("srt:") <> pprSRT srt