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"
55 %************************************************************************
57 \subsection[toplevel-constructors]{Top-level constructors}
59 %************************************************************************
62 cgTopRhsCon :: Id -- Name of thing bound to this RHS
65 -> FCode (Id, CgIdInfo)
66 cgTopRhsCon id con args
69 -- Windows DLLs have a problem with static cross-DLL refs.
70 ; this_pkg <- getThisPackage
71 ; ASSERT( not (isDllConApp this_pkg con args) ) return ()
73 ; ASSERT( args `lengthIs` dataConRepArity con ) return ()
76 ; amodes <- getArgAmodes args
80 lf_info = mkConLFInfo con
81 closure_label = mkClosureLabel name $ idCafInfo id
82 caffy = any stgArgHasCafRefs args
83 (closure_info, amodes_w_offsets) = layOutStaticConstr con amodes
84 closure_rep = mkStaticClosureFields
86 dontCareCCS -- Because it's static data
90 payload = map get_lit amodes_w_offsets
91 get_lit (CmmLit lit, _offset) = lit
92 get_lit other = pprPanic "CgCon.get_lit" (ppr other)
93 -- NB1: amodes_w_offsets is sorted into ptrs first, then non-ptrs
94 -- NB2: all the amodes should be Lits!
97 ; emitDataLits closure_label closure_rep
100 ; returnFC (id, taggedStableIdInfo id (mkLblExpr closure_label) lf_info con) }
103 %************************************************************************
105 %* non-top-level constructors *
107 %************************************************************************
108 \subsection[code-for-constructors]{The code for constructors}
111 buildDynCon :: Id -- Name of the thing to which this constr will
113 -> CostCentreStack -- Where to grab cost centre from;
114 -- current CCS if currentOrSubsumedCCS
115 -> DataCon -- The data constructor
116 -> [(CgRep,CmmExpr)] -- Its args
117 -> FCode CgIdInfo -- Return details about how to find it
119 -- We used to pass a boolean indicating whether all the
120 -- args were of size zero, so we could use a static
121 -- construtor; but I concluded that it just isn't worth it.
122 -- Now I/O uses unboxed tuples there just aren't any constructors
123 -- with all size-zero args.
125 -- The reason for having a separate argument, rather than looking at
126 -- the addr modes of the args is that we may be in a "knot", and
127 -- premature looking at the args will cause the compiler to black-hole!
130 First we deal with the case of zero-arity constructors. Now, they
131 will probably be unfolded, so we don't expect to see this case much,
132 if at all, but it does no harm, and sets the scene for characters.
134 In the case of zero-arity constructors, or, more accurately, those
135 which have exclusively size-zero (VoidRep) args, we generate no code
139 buildDynCon binder _ con []
140 = returnFC (taggedStableIdInfo binder
141 (mkLblExpr (mkClosureLabel (dataConName con)
147 The following three paragraphs about @Char@-like and @Int@-like
148 closures are obsolete, but I don't understand the details well enough
149 to properly word them, sorry. I've changed the treatment of @Char@s to
150 be analogous to @Int@s: only a subset is preallocated, because @Char@
151 has now 31 bits. Only literals are handled here. -- Qrczak
153 Now for @Char@-like closures. We generate an assignment of the
154 address of the closure to a temporary. It would be possible simply to
155 generate no code, and record the addressing mode in the environment,
156 but we'd have to be careful if the argument wasn't a constant --- so
157 for simplicity we just always asssign to a temporary.
159 Last special case: @Int@-like closures. We only special-case the
160 situation in which the argument is a literal in the range
161 @mIN_INTLIKE@..@mAX_INTLILKE@. NB: for @Char@-like closures we can
162 work with any old argument, but for @Int@-like ones the argument has
163 to be a literal. Reason: @Char@ like closures have an argument type
164 which is guaranteed in range.
166 Because of this, we use can safely return an addressing mode.
169 buildDynCon binder _ con [arg_amode]
170 | maybeIntLikeCon con
171 , (_, CmmLit (CmmInt val _)) <- arg_amode
172 , let val_int = (fromIntegral val) :: Int
173 , val_int <= mAX_INTLIKE && val_int >= mIN_INTLIKE
174 = do { let intlike_lbl = mkCmmGcPtrLabel rtsPackageId (fsLit "stg_INTLIKE_closure")
175 offsetW = (val_int - mIN_INTLIKE) * (fixedHdrSize + 1)
176 -- INTLIKE closures consist of a header and one word payload
177 intlike_amode = CmmLit (cmmLabelOffW intlike_lbl offsetW)
178 ; returnFC (taggedStableIdInfo binder intlike_amode (mkConLFInfo con) con) }
180 buildDynCon binder _ con [arg_amode]
181 | maybeCharLikeCon con
182 , (_, CmmLit (CmmInt val _)) <- arg_amode
183 , let val_int = (fromIntegral val) :: Int
184 , val_int <= mAX_CHARLIKE && val_int >= mIN_CHARLIKE
185 = do { let charlike_lbl = mkCmmGcPtrLabel rtsPackageId (fsLit "stg_CHARLIKE_closure")
186 offsetW = (val_int - mIN_CHARLIKE) * (fixedHdrSize + 1)
187 -- CHARLIKE closures consist of a header and one word payload
188 charlike_amode = CmmLit (cmmLabelOffW charlike_lbl offsetW)
189 ; returnFC (taggedStableIdInfo binder charlike_amode (mkConLFInfo con) con) }
192 Now the general case.
195 buildDynCon binder ccs con args
198 (closure_info, amodes_w_offsets) = layOutDynConstr con args
200 ; hp_off <- allocDynClosure closure_info use_cc blame_cc amodes_w_offsets
201 ; returnFC (taggedHeapIdInfo binder hp_off lf_info con) }
203 lf_info = mkConLFInfo con
205 use_cc -- cost-centre to stick in the object
206 | currentOrSubsumedCCS ccs = curCCS
207 | otherwise = CmmLit (mkCCostCentreStack ccs)
209 blame_cc = use_cc -- cost-centre on which to blame the alloc (same)
213 %************************************************************************
215 %* constructor-related utility function: *
216 %* bindConArgs is called from cgAlt of a case *
218 %************************************************************************
219 \subsection[constructor-utilities]{@bindConArgs@: constructor-related utility}
221 @bindConArgs@ $con args$ augments the environment with bindings for the
222 binders $args$, assuming that we have just returned from a @case@ which
226 bindConArgs :: DataCon -> [Id] -> Code
230 -- The binding below forces the masking out of the tag bits
231 -- when accessing the constructor field.
232 bind_arg (arg, offset) = bindNewToUntagNode arg offset (mkLFArgument arg) (tagForCon con)
233 (_, args_w_offsets) = layOutDynConstr con (addIdReps args)
235 ASSERT(not (isUnboxedTupleCon con)) return ()
236 mapCs bind_arg args_w_offsets
239 Unboxed tuples are handled slightly differently - the object is
240 returned in registers and on the stack instead of the heap.
243 bindUnboxedTupleComponents
245 -> FCode ([(Id,GlobalReg)], -- Regs assigned
246 WordOff, -- Number of pointer stack slots
247 WordOff, -- Number of non-pointer stack slots
248 VirtualSpOffset) -- Offset of return address slot
249 -- (= realSP on entry)
251 bindUnboxedTupleComponents args
256 -- Assign as many components as possible to registers
257 ; let (reg_args, stk_args) = assignReturnRegs (addIdReps args)
259 -- Separate the rest of the args into pointers and non-pointers
260 (ptr_args, nptr_args) = separateByPtrFollowness stk_args
262 -- Allocate the rest on the stack
263 -- The real SP points to the return address, above which any
264 -- leftover unboxed-tuple components will be allocated
265 (ptr_sp, ptr_offsets) = mkVirtStkOffsets rsp ptr_args
266 (nptr_sp, nptr_offsets) = mkVirtStkOffsets ptr_sp nptr_args
268 nptrs = nptr_sp - ptr_sp
270 -- The stack pointer points to the last stack-allocated component
271 ; setRealAndVirtualSp nptr_sp
273 -- We have just allocated slots starting at real SP + 1, and set the new
274 -- virtual SP to the topmost allocated slot.
275 -- If the virtual SP started *below* the real SP, we've just jumped over
276 -- some slots that won't be in the free-list, so put them there
277 -- This commonly happens because we've freed the return-address slot
278 -- (trimming back the virtual SP), but the real SP still points to that slot
279 ; freeStackSlots [vsp+1,vsp+2 .. rsp]
281 ; bindArgsToRegs reg_args
282 ; bindArgsToStack ptr_offsets
283 ; bindArgsToStack nptr_offsets
285 ; returnFC (reg_args, ptrs, nptrs, rsp) }
288 %************************************************************************
290 Actually generate code for a constructor return
292 %************************************************************************
295 Note: it's the responsibility of the @cgReturnDataCon@ caller to be
296 sure the @amodes@ passed don't conflict with each other.
298 cgReturnDataCon :: DataCon -> [(CgRep, CmmExpr)] -> Code
300 cgReturnDataCon con amodes
301 | isUnboxedTupleCon con = returnUnboxedTuple amodes
302 -- when profiling we can't shortcut here, we have to enter the closure
303 -- for it to be marked as "used" for LDV profiling.
304 | opt_SccProfilingOn = build_it_then enter_it
306 = ASSERT( amodes `lengthIs` dataConRepArity con )
307 do { EndOfBlockInfo _ sequel <- getEndOfBlockInfo
309 CaseAlts _ (Just (alts, deflt_lbl)) bndr
310 -> -- Ho! We know the constructor so we can
311 -- go straight to the right alternative
312 case assocMaybe alts (dataConTagZ con) of {
313 Just join_lbl -> build_it_then (jump_to join_lbl);
315 -- Special case! We're returning a constructor to the default case
316 -- of an enclosing case. For example:
318 -- case (case e of (a,b) -> C a b) of
320 -- y -> ...<returning here!>...
323 -- if the default is a non-bind-default (ie does not use y),
324 -- then we should simply jump to the default join point;
326 | isDeadBinder bndr -> performReturn (jump_to deflt_lbl)
327 | otherwise -> build_it_then (jump_to deflt_lbl) }
329 _otherwise -- The usual case
330 -> build_it_then emitReturnInstr
333 enter_it = stmtsC [ CmmAssign nodeReg (cmmUntag (CmmReg nodeReg)),
334 CmmJump (entryCode (closureInfoPtr (CmmReg nodeReg))) [] ]
335 jump_to lbl = stmtC (CmmJump (CmmLit lbl) [])
336 build_it_then return_code
337 = do { -- BUILD THE OBJECT IN THE HEAP
338 -- The first "con" says that the name bound to this
339 -- closure is "con", which is a bit of a fudge, but it only
342 -- This Id is also used to get a unique for a
343 -- temporary variable, if the closure is a CHARLIKE.
344 -- funnily enough, this makes the unique always come
346 tickyReturnNewCon (length amodes)
347 ; idinfo <- buildDynCon (dataConWorkId con) currentCCS con amodes
348 ; amode <- idInfoToAmode idinfo
349 ; checkedAbsC (CmmAssign nodeReg amode)
350 ; performReturn return_code }
354 %************************************************************************
356 Generating static stuff for algebraic data types
358 %************************************************************************
360 [These comments are rather out of date]
363 Info tbls & Macro & Kind of constructor \\
365 info & @CONST_INFO_TABLE@& Zero arity (no info -- compiler uses static closure)\\
366 info & @CHARLIKE_INFO_TABLE@& Charlike (no info -- compiler indexes fixed array)\\
367 info & @INTLIKE_INFO_TABLE@& Intlike; the one macro generates both info tbls\\
368 info & @SPEC_INFO_TABLE@& SPECish, and bigger than or equal to @MIN_UPD_SIZE@\\
369 info & @GEN_INFO_TABLE@& GENish (hence bigger than or equal to @MIN_UPD_SIZE@)\\
372 Possible info tables for constructor con:
376 Used for dynamically let(rec)-bound occurrences of
377 the constructor, and for updates. For constructors
378 which are int-like, char-like or nullary, when GC occurs,
379 the closure tries to get rid of itself.
381 \item[@_static_info@:]
382 Static occurrences of the constructor
383 macro: @STATIC_INFO_TABLE@.
386 For zero-arity constructors, \tr{con}, we NO LONGER generate a static closure;
387 it's place is taken by the top level defn of the constructor.
389 For charlike and intlike closures there is a fixed array of static
390 closures predeclared.
393 cgTyCon :: TyCon -> FCode [Cmm] -- each constructor gets a separate Cmm
395 = do { constrs <- mapM (getCmm . cgDataCon) (tyConDataCons tycon)
397 -- Generate a table of static closures for an enumeration type
398 -- Put the table after the data constructor decls, because the
399 -- datatype closure table (for enumeration types)
400 -- to (say) PrelBase_$wTrue_closure, which is defined in code_stuff
401 -- Note that the closure pointers are tagged.
403 -- XXX comment says to put table after constructor decls, but
404 -- code appears to put it before --- NR 16 Aug 2007
406 if isEnumerationTyCon tycon then do
407 tbl <- getCmm (emitRODataLits "cgTyCon" (mkLocalClosureTableLabel (tyConName tycon) NoCafRefs)
408 [ CmmLabelOff (mkLocalClosureLabel (dataConName con) NoCafRefs) (tagForCon con)
409 | con <- tyConDataCons tycon])
414 ; return (extra ++ constrs)
418 Generate the entry code, info tables, and (for niladic constructor) the
419 static closure, for a constructor.
422 cgDataCon :: DataCon -> Code
424 = do { -- Don't need any dynamic closure code for zero-arity constructors
427 -- To allow the debuggers, interpreters, etc to cope with
428 -- static data structures (ie those built at compile
429 -- time), we take care that info-table contains the
430 -- information we need.
431 (static_cl_info, _) =
432 layOutStaticConstr data_con arg_reps
434 (dyn_cl_info, arg_things) =
435 layOutDynConstr data_con arg_reps
437 emit_info cl_info ticky_code
438 = do { code_blks <- getCgStmts the_code
439 ; emitClosureCodeAndInfoTable cl_info [] code_blks }
441 the_code = do { _ <- ticky_code
442 ; ldvEnter (CmmReg nodeReg)
445 arg_reps :: [(CgRep, Type)]
446 arg_reps = [(typeCgRep ty, ty) | ty <- dataConRepArgTys data_con]
449 -- NB: We don't set CC when entering data (WDP 94/06)
450 tickyReturnOldCon (length arg_things)
451 -- The case continuation code is expecting a tagged pointer
452 ; stmtC (CmmAssign nodeReg
453 (tagCons data_con (CmmReg nodeReg)))
454 ; performReturn emitReturnInstr }
455 -- noStmts: Ptr to thing already in Node
457 ; whenC (not (isNullaryRepDataCon data_con))
458 (emit_info dyn_cl_info tickyEnterDynCon)
460 -- Dynamic-Closure first, to reduce forward references
461 ; emit_info static_cl_info tickyEnterStaticCon }