2 % (c) The University of Glasgow 2006
3 % (c) The GRASP Project, Glasgow University, 1992-1998
5 \section[CgCon]{Code generation for constructors}
7 This module provides the support code for @StgToAbstractC@ to deal
8 with {\em constructors} on the RHSs of let(rec)s. See also
9 @CgClosure@, which deals with closures.
13 cgTopRhsCon, buildDynCon,
14 bindConArgs, bindUnboxedTupleComponents,
19 #include "HsVersions.h"
48 import Util ( lengthIs )
53 %************************************************************************
55 \subsection[toplevel-constructors]{Top-level constructors}
57 %************************************************************************
60 cgTopRhsCon :: Id -- Name of thing bound to this RHS
63 -> FCode (Id, CgIdInfo)
64 cgTopRhsCon id con args
67 -- Windows DLLs have a problem with static cross-DLL refs.
68 ; this_pkg <- getThisPackage
69 ; ASSERT( not (isDllConApp this_pkg con args) ) return ()
71 ; ASSERT( args `lengthIs` dataConRepArity con ) return ()
74 ; amodes <- getArgAmodes args
78 lf_info = mkConLFInfo con
79 closure_label = mkClosureLabel name
80 caffy = any stgArgHasCafRefs args
81 (closure_info, amodes_w_offsets) = layOutStaticConstr con amodes
82 closure_rep = mkStaticClosureFields
84 dontCareCCS -- Because it's static data
88 payload = map get_lit amodes_w_offsets
89 get_lit (CmmLit lit, _offset) = lit
90 get_lit other = pprPanic "CgCon.get_lit" (ppr other)
91 -- NB1: amodes_w_offsets is sorted into ptrs first, then non-ptrs
92 -- NB2: all the amodes should be Lits!
95 ; emitDataLits closure_label closure_rep
98 ; returnFC (id, taggedStableIdInfo id (mkLblExpr closure_label) lf_info con) }
101 %************************************************************************
103 %* non-top-level constructors *
105 %************************************************************************
106 \subsection[code-for-constructors]{The code for constructors}
109 buildDynCon :: Id -- Name of the thing to which this constr will
111 -> CostCentreStack -- Where to grab cost centre from;
112 -- current CCS if currentOrSubsumedCCS
113 -> DataCon -- The data constructor
114 -> [(CgRep,CmmExpr)] -- Its args
115 -> FCode CgIdInfo -- Return details about how to find it
117 -- We used to pass a boolean indicating whether all the
118 -- args were of size zero, so we could use a static
119 -- construtor; but I concluded that it just isn't worth it.
120 -- Now I/O uses unboxed tuples there just aren't any constructors
121 -- with all size-zero args.
123 -- The reason for having a separate argument, rather than looking at
124 -- the addr modes of the args is that we may be in a "knot", and
125 -- premature looking at the args will cause the compiler to black-hole!
128 First we deal with the case of zero-arity constructors. Now, they
129 will probably be unfolded, so we don't expect to see this case much,
130 if at all, but it does no harm, and sets the scene for characters.
132 In the case of zero-arity constructors, or, more accurately, those
133 which have exclusively size-zero (VoidRep) args, we generate no code
137 buildDynCon binder cc con []
138 = returnFC (taggedStableIdInfo binder
139 (mkLblExpr (mkClosureLabel (dataConName con)))
144 The following three paragraphs about @Char@-like and @Int@-like
145 closures are obsolete, but I don't understand the details well enough
146 to properly word them, sorry. I've changed the treatment of @Char@s to
147 be analogous to @Int@s: only a subset is preallocated, because @Char@
148 has now 31 bits. Only literals are handled here. -- Qrczak
150 Now for @Char@-like closures. We generate an assignment of the
151 address of the closure to a temporary. It would be possible simply to
152 generate no code, and record the addressing mode in the environment,
153 but we'd have to be careful if the argument wasn't a constant --- so
154 for simplicity we just always asssign to a temporary.
156 Last special case: @Int@-like closures. We only special-case the
157 situation in which the argument is a literal in the range
158 @mIN_INTLIKE@..@mAX_INTLILKE@. NB: for @Char@-like closures we can
159 work with any old argument, but for @Int@-like ones the argument has
160 to be a literal. Reason: @Char@ like closures have an argument type
161 which is guaranteed in range.
163 Because of this, we use can safely return an addressing mode.
166 buildDynCon binder cc con [arg_amode]
167 | maybeIntLikeCon con
168 , (_, CmmLit (CmmInt val _)) <- arg_amode
169 , let val_int = (fromIntegral val) :: Int
170 , val_int <= mAX_INTLIKE && val_int >= mIN_INTLIKE
171 = do { let intlike_lbl = mkRtsDataLabel SLIT("stg_INTLIKE_closure")
172 offsetW = (val_int - mIN_INTLIKE) * (fixedHdrSize + 1)
173 -- INTLIKE closures consist of a header and one word payload
174 intlike_amode = CmmLit (cmmLabelOffW intlike_lbl offsetW)
175 ; returnFC (taggedStableIdInfo binder intlike_amode (mkConLFInfo con) con) }
177 buildDynCon binder cc con [arg_amode]
178 | maybeCharLikeCon con
179 , (_, CmmLit (CmmInt val _)) <- arg_amode
180 , let val_int = (fromIntegral val) :: Int
181 , val_int <= mAX_CHARLIKE && val_int >= mIN_CHARLIKE
182 = do { let charlike_lbl = mkRtsDataLabel SLIT("stg_CHARLIKE_closure")
183 offsetW = (val_int - mIN_CHARLIKE) * (fixedHdrSize + 1)
184 -- CHARLIKE closures consist of a header and one word payload
185 charlike_amode = CmmLit (cmmLabelOffW charlike_lbl offsetW)
186 ; returnFC (taggedStableIdInfo binder charlike_amode (mkConLFInfo con) con) }
189 Now the general case.
192 buildDynCon binder ccs con args
195 (closure_info, amodes_w_offsets) = layOutDynConstr con args
197 ; hp_off <- allocDynClosure closure_info use_cc blame_cc amodes_w_offsets
198 ; returnFC (taggedHeapIdInfo binder hp_off lf_info con) }
200 lf_info = mkConLFInfo con
202 use_cc -- cost-centre to stick in the object
203 | currentOrSubsumedCCS ccs = curCCS
204 | otherwise = CmmLit (mkCCostCentreStack ccs)
206 blame_cc = use_cc -- cost-centre on which to blame the alloc (same)
210 %************************************************************************
212 %* constructor-related utility function: *
213 %* bindConArgs is called from cgAlt of a case *
215 %************************************************************************
216 \subsection[constructor-utilities]{@bindConArgs@: constructor-related utility}
218 @bindConArgs@ $con args$ augments the environment with bindings for the
219 binders $args$, assuming that we have just returned from a @case@ which
223 bindConArgs :: DataCon -> [Id] -> Code
227 -- The binding below forces the masking out of the tag bits
228 -- when accessing the constructor field.
229 bind_arg (arg, offset) = bindNewToUntagNode arg offset (mkLFArgument arg) (tagForCon con)
230 (_, args_w_offsets) = layOutDynConstr con (addIdReps args)
232 ASSERT(not (isUnboxedTupleCon con)) return ()
233 mapCs bind_arg args_w_offsets
236 Unboxed tuples are handled slightly differently - the object is
237 returned in registers and on the stack instead of the heap.
240 bindUnboxedTupleComponents
242 -> FCode ([(Id,GlobalReg)], -- Regs assigned
243 WordOff, -- Number of pointer stack slots
244 WordOff, -- Number of non-pointer stack slots
245 VirtualSpOffset) -- Offset of return address slot
246 -- (= realSP on entry)
248 bindUnboxedTupleComponents args
253 -- Assign as many components as possible to registers
254 ; let (reg_args, stk_args) = assignReturnRegs (addIdReps args)
256 -- Separate the rest of the args into pointers and non-pointers
257 (ptr_args, nptr_args) = separateByPtrFollowness stk_args
259 -- Allocate the rest on the stack
260 -- The real SP points to the return address, above which any
261 -- leftover unboxed-tuple components will be allocated
262 (ptr_sp, ptr_offsets) = mkVirtStkOffsets rsp ptr_args
263 (nptr_sp, nptr_offsets) = mkVirtStkOffsets ptr_sp nptr_args
265 nptrs = nptr_sp - ptr_sp
267 -- The stack pointer points to the last stack-allocated component
268 ; setRealAndVirtualSp nptr_sp
270 -- We have just allocated slots starting at real SP + 1, and set the new
271 -- virtual SP to the topmost allocated slot.
272 -- If the virtual SP started *below* the real SP, we've just jumped over
273 -- some slots that won't be in the free-list, so put them there
274 -- This commonly happens because we've freed the return-address slot
275 -- (trimming back the virtual SP), but the real SP still points to that slot
276 ; freeStackSlots [vsp+1,vsp+2 .. rsp]
278 ; bindArgsToRegs reg_args
279 ; bindArgsToStack ptr_offsets
280 ; bindArgsToStack nptr_offsets
282 ; returnFC (reg_args, ptrs, nptrs, rsp) }
285 %************************************************************************
287 Actually generate code for a constructor return
289 %************************************************************************
292 Note: it's the responsibility of the @cgReturnDataCon@ caller to be
293 sure the @amodes@ passed don't conflict with each other.
295 cgReturnDataCon :: DataCon -> [(CgRep, CmmExpr)] -> Code
297 cgReturnDataCon con amodes
298 = ASSERT( amodes `lengthIs` dataConRepArity con )
299 do { EndOfBlockInfo _ sequel <- getEndOfBlockInfo
301 CaseAlts _ (Just (alts, deflt_lbl)) bndr
302 -> -- Ho! We know the constructor so we can
303 -- go straight to the right alternative
304 case assocMaybe alts (dataConTagZ con) of {
305 Just join_lbl -> build_it_then (jump_to join_lbl);
307 -- Special case! We're returning a constructor to the default case
308 -- of an enclosing case. For example:
310 -- case (case e of (a,b) -> C a b) of
312 -- y -> ...<returning here!>...
315 -- if the default is a non-bind-default (ie does not use y),
316 -- then we should simply jump to the default join point;
318 | isDeadBinder bndr -> performReturn (jump_to deflt_lbl)
319 | otherwise -> build_it_then (jump_to deflt_lbl) }
321 other_sequel -- The usual case
322 | isUnboxedTupleCon con -> returnUnboxedTuple amodes
323 | otherwise -> build_it_then emitReturnInstr
326 jump_to lbl = stmtC (CmmJump (CmmLit lbl) [])
327 build_it_then return_code
328 = do { -- BUILD THE OBJECT IN THE HEAP
329 -- The first "con" says that the name bound to this
330 -- closure is "con", which is a bit of a fudge, but it only
333 -- This Id is also used to get a unique for a
334 -- temporary variable, if the closure is a CHARLIKE.
335 -- funnily enough, this makes the unique always come
337 tickyReturnNewCon (length amodes)
338 ; idinfo <- buildDynCon (dataConWorkId con) currentCCS con amodes
339 ; amode <- idInfoToAmode idinfo
340 ; checkedAbsC (CmmAssign nodeReg amode)
341 ; performReturn return_code }
345 %************************************************************************
347 Generating static stuff for algebraic data types
349 %************************************************************************
351 [These comments are rather out of date]
354 Info tbls & Macro & Kind of constructor \\
356 info & @CONST_INFO_TABLE@& Zero arity (no info -- compiler uses static closure)\\
357 info & @CHARLIKE_INFO_TABLE@& Charlike (no info -- compiler indexes fixed array)\\
358 info & @INTLIKE_INFO_TABLE@& Intlike; the one macro generates both info tbls\\
359 info & @SPEC_INFO_TABLE@& SPECish, and bigger than or equal to @MIN_UPD_SIZE@\\
360 info & @GEN_INFO_TABLE@& GENish (hence bigger than or equal to @MIN_UPD_SIZE@)\\
363 Possible info tables for constructor con:
367 Used for dynamically let(rec)-bound occurrences of
368 the constructor, and for updates. For constructors
369 which are int-like, char-like or nullary, when GC occurs,
370 the closure tries to get rid of itself.
372 \item[@_static_info@:]
373 Static occurrences of the constructor
374 macro: @STATIC_INFO_TABLE@.
377 For zero-arity constructors, \tr{con}, we NO LONGER generate a static closure;
378 it's place is taken by the top level defn of the constructor.
380 For charlike and intlike closures there is a fixed array of static
381 closures predeclared.
384 cgTyCon :: TyCon -> FCode [Cmm] -- each constructor gets a separate Cmm
386 = do { constrs <- mapM (getCmm . cgDataCon) (tyConDataCons tycon)
388 -- Generate a table of static closures for an enumeration type
389 -- Put the table after the data constructor decls, because the
390 -- datatype closure table (for enumeration types)
391 -- to (say) PrelBase_$wTrue_closure, which is defined in code_stuff
392 -- Note that the closure pointers are tagged.
394 if isEnumerationTyCon tycon then do
395 tbl <- getCmm (emitRODataLits (mkLocalClosureTableLabel
397 [ CmmLabelOff (mkLocalClosureLabel (dataConName con)) (tagForCon con)
398 | con <- tyConDataCons tycon])
403 ; return (extra ++ constrs)
407 Generate the entry code, info tables, and (for niladic constructor) the
408 static closure, for a constructor.
411 cgDataCon :: DataCon -> Code
413 = do { -- Don't need any dynamic closure code for zero-arity constructors
416 -- To allow the debuggers, interpreters, etc to cope with
417 -- static data structures (ie those built at compile
418 -- time), we take care that info-table contains the
419 -- information we need.
420 (static_cl_info, _) =
421 layOutStaticConstr data_con arg_reps
423 (dyn_cl_info, arg_things) =
424 layOutDynConstr data_con arg_reps
426 emit_info cl_info ticky_code
427 = do { code_blks <- getCgStmts the_code
428 ; emitClosureCodeAndInfoTable cl_info [] code_blks }
430 the_code = do { ticky_code
431 ; ldvEnter (CmmReg nodeReg)
434 arg_reps :: [(CgRep, Type)]
435 arg_reps = [(typeCgRep ty, ty) | ty <- dataConRepArgTys data_con]
438 -- NB: We don't set CC when entering data (WDP 94/06)
439 tickyReturnOldCon (length arg_things)
440 -- The case continuation code is expecting a tagged pointer
441 ; stmtC (CmmAssign nodeReg
442 (tagCons data_con (CmmReg nodeReg)))
443 ; performReturn emitReturnInstr }
444 -- noStmts: Ptr to thing already in Node
446 ; whenC (not (isNullaryRepDataCon data_con))
447 (emit_info dyn_cl_info tickyEnterDynCon)
449 -- Dynamic-Closure first, to reduce forward references
450 ; emit_info static_cl_info tickyEnterStaticCon }