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 -- The above warning supression flag is a temporary kludge.
14 -- While working on this module you are encouraged to remove it and fix
15 -- any warnings in the module. See
16 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
20 cgTopRhsCon, buildDynCon,
21 bindConArgs, bindUnboxedTupleComponents,
26 #include "HsVersions.h"
58 %************************************************************************
60 \subsection[toplevel-constructors]{Top-level constructors}
62 %************************************************************************
65 cgTopRhsCon :: Id -- Name of thing bound to this RHS
68 -> FCode (Id, CgIdInfo)
69 cgTopRhsCon id con args
72 -- Windows DLLs have a problem with static cross-DLL refs.
73 ; this_pkg <- getThisPackage
74 ; ASSERT( not (isDllConApp this_pkg con args) ) return ()
76 ; ASSERT( args `lengthIs` dataConRepArity con ) return ()
79 ; amodes <- getArgAmodes args
83 lf_info = mkConLFInfo con
84 closure_label = mkClosureLabel name
85 caffy = any stgArgHasCafRefs args
86 (closure_info, amodes_w_offsets) = layOutStaticConstr con amodes
87 closure_rep = mkStaticClosureFields
89 dontCareCCS -- Because it's static data
93 payload = map get_lit amodes_w_offsets
94 get_lit (CmmLit lit, _offset) = lit
95 get_lit other = pprPanic "CgCon.get_lit" (ppr other)
96 -- NB1: amodes_w_offsets is sorted into ptrs first, then non-ptrs
97 -- NB2: all the amodes should be Lits!
100 ; emitDataLits closure_label closure_rep
103 ; returnFC (id, taggedStableIdInfo id (mkLblExpr closure_label) lf_info con) }
106 %************************************************************************
108 %* non-top-level constructors *
110 %************************************************************************
111 \subsection[code-for-constructors]{The code for constructors}
114 buildDynCon :: Id -- Name of the thing to which this constr will
116 -> CostCentreStack -- Where to grab cost centre from;
117 -- current CCS if currentOrSubsumedCCS
118 -> DataCon -- The data constructor
119 -> [(CgRep,CmmExpr)] -- Its args
120 -> FCode CgIdInfo -- Return details about how to find it
122 -- We used to pass a boolean indicating whether all the
123 -- args were of size zero, so we could use a static
124 -- construtor; but I concluded that it just isn't worth it.
125 -- Now I/O uses unboxed tuples there just aren't any constructors
126 -- with all size-zero args.
128 -- The reason for having a separate argument, rather than looking at
129 -- the addr modes of the args is that we may be in a "knot", and
130 -- premature looking at the args will cause the compiler to black-hole!
133 First we deal with the case of zero-arity constructors. Now, they
134 will probably be unfolded, so we don't expect to see this case much,
135 if at all, but it does no harm, and sets the scene for characters.
137 In the case of zero-arity constructors, or, more accurately, those
138 which have exclusively size-zero (VoidRep) args, we generate no code
142 buildDynCon binder cc con []
143 = returnFC (taggedStableIdInfo binder
144 (mkLblExpr (mkClosureLabel (dataConName con)))
149 The following three paragraphs about @Char@-like and @Int@-like
150 closures are obsolete, but I don't understand the details well enough
151 to properly word them, sorry. I've changed the treatment of @Char@s to
152 be analogous to @Int@s: only a subset is preallocated, because @Char@
153 has now 31 bits. Only literals are handled here. -- Qrczak
155 Now for @Char@-like closures. We generate an assignment of the
156 address of the closure to a temporary. It would be possible simply to
157 generate no code, and record the addressing mode in the environment,
158 but we'd have to be careful if the argument wasn't a constant --- so
159 for simplicity we just always asssign to a temporary.
161 Last special case: @Int@-like closures. We only special-case the
162 situation in which the argument is a literal in the range
163 @mIN_INTLIKE@..@mAX_INTLILKE@. NB: for @Char@-like closures we can
164 work with any old argument, but for @Int@-like ones the argument has
165 to be a literal. Reason: @Char@ like closures have an argument type
166 which is guaranteed in range.
168 Because of this, we use can safely return an addressing mode.
171 buildDynCon binder cc con [arg_amode]
172 | maybeIntLikeCon con
173 , (_, CmmLit (CmmInt val _)) <- arg_amode
174 , let val_int = (fromIntegral val) :: Int
175 , val_int <= mAX_INTLIKE && val_int >= mIN_INTLIKE
176 = do { let intlike_lbl = mkRtsDataLabel SLIT("stg_INTLIKE_closure")
177 offsetW = (val_int - mIN_INTLIKE) * (fixedHdrSize + 1)
178 -- INTLIKE closures consist of a header and one word payload
179 intlike_amode = CmmLit (cmmLabelOffW intlike_lbl offsetW)
180 ; returnFC (taggedStableIdInfo binder intlike_amode (mkConLFInfo con) con) }
182 buildDynCon binder cc con [arg_amode]
183 | maybeCharLikeCon con
184 , (_, CmmLit (CmmInt val _)) <- arg_amode
185 , let val_int = (fromIntegral val) :: Int
186 , val_int <= mAX_CHARLIKE && val_int >= mIN_CHARLIKE
187 = do { let charlike_lbl = mkRtsDataLabel SLIT("stg_CHARLIKE_closure")
188 offsetW = (val_int - mIN_CHARLIKE) * (fixedHdrSize + 1)
189 -- CHARLIKE closures consist of a header and one word payload
190 charlike_amode = CmmLit (cmmLabelOffW charlike_lbl offsetW)
191 ; returnFC (taggedStableIdInfo binder charlike_amode (mkConLFInfo con) con) }
194 Now the general case.
197 buildDynCon binder ccs con args
200 (closure_info, amodes_w_offsets) = layOutDynConstr con args
202 ; hp_off <- allocDynClosure closure_info use_cc blame_cc amodes_w_offsets
203 ; returnFC (taggedHeapIdInfo binder hp_off lf_info con) }
205 lf_info = mkConLFInfo con
207 use_cc -- cost-centre to stick in the object
208 | currentOrSubsumedCCS ccs = curCCS
209 | otherwise = CmmLit (mkCCostCentreStack ccs)
211 blame_cc = use_cc -- cost-centre on which to blame the alloc (same)
215 %************************************************************************
217 %* constructor-related utility function: *
218 %* bindConArgs is called from cgAlt of a case *
220 %************************************************************************
221 \subsection[constructor-utilities]{@bindConArgs@: constructor-related utility}
223 @bindConArgs@ $con args$ augments the environment with bindings for the
224 binders $args$, assuming that we have just returned from a @case@ which
228 bindConArgs :: DataCon -> [Id] -> Code
232 -- The binding below forces the masking out of the tag bits
233 -- when accessing the constructor field.
234 bind_arg (arg, offset) = bindNewToUntagNode arg offset (mkLFArgument arg) (tagForCon con)
235 (_, args_w_offsets) = layOutDynConstr con (addIdReps args)
237 ASSERT(not (isUnboxedTupleCon con)) return ()
238 mapCs bind_arg args_w_offsets
241 Unboxed tuples are handled slightly differently - the object is
242 returned in registers and on the stack instead of the heap.
245 bindUnboxedTupleComponents
247 -> FCode ([(Id,GlobalReg)], -- Regs assigned
248 WordOff, -- Number of pointer stack slots
249 WordOff, -- Number of non-pointer stack slots
250 VirtualSpOffset) -- Offset of return address slot
251 -- (= realSP on entry)
253 bindUnboxedTupleComponents args
258 -- Assign as many components as possible to registers
259 ; let (reg_args, stk_args) = assignReturnRegs (addIdReps args)
261 -- Separate the rest of the args into pointers and non-pointers
262 (ptr_args, nptr_args) = separateByPtrFollowness stk_args
264 -- Allocate the rest on the stack
265 -- The real SP points to the return address, above which any
266 -- leftover unboxed-tuple components will be allocated
267 (ptr_sp, ptr_offsets) = mkVirtStkOffsets rsp ptr_args
268 (nptr_sp, nptr_offsets) = mkVirtStkOffsets ptr_sp nptr_args
270 nptrs = nptr_sp - ptr_sp
272 -- The stack pointer points to the last stack-allocated component
273 ; setRealAndVirtualSp nptr_sp
275 -- We have just allocated slots starting at real SP + 1, and set the new
276 -- virtual SP to the topmost allocated slot.
277 -- If the virtual SP started *below* the real SP, we've just jumped over
278 -- some slots that won't be in the free-list, so put them there
279 -- This commonly happens because we've freed the return-address slot
280 -- (trimming back the virtual SP), but the real SP still points to that slot
281 ; freeStackSlots [vsp+1,vsp+2 .. rsp]
283 ; bindArgsToRegs reg_args
284 ; bindArgsToStack ptr_offsets
285 ; bindArgsToStack nptr_offsets
287 ; returnFC (reg_args, ptrs, nptrs, rsp) }
290 %************************************************************************
292 Actually generate code for a constructor return
294 %************************************************************************
297 Note: it's the responsibility of the @cgReturnDataCon@ caller to be
298 sure the @amodes@ passed don't conflict with each other.
300 cgReturnDataCon :: DataCon -> [(CgRep, CmmExpr)] -> Code
302 cgReturnDataCon con amodes
303 = ASSERT( amodes `lengthIs` dataConRepArity con )
304 do { EndOfBlockInfo _ sequel <- getEndOfBlockInfo
306 CaseAlts _ (Just (alts, deflt_lbl)) bndr
307 -> -- Ho! We know the constructor so we can
308 -- go straight to the right alternative
309 case assocMaybe alts (dataConTagZ con) of {
310 Just join_lbl -> build_it_then (jump_to join_lbl);
312 -- Special case! We're returning a constructor to the default case
313 -- of an enclosing case. For example:
315 -- case (case e of (a,b) -> C a b) of
317 -- y -> ...<returning here!>...
320 -- if the default is a non-bind-default (ie does not use y),
321 -- then we should simply jump to the default join point;
323 | isDeadBinder bndr -> performReturn (jump_to deflt_lbl)
324 | otherwise -> build_it_then (jump_to deflt_lbl) }
326 other_sequel -- The usual case
327 | isUnboxedTupleCon con -> returnUnboxedTuple amodes
328 | otherwise -> build_it_then emitReturnInstr
331 jump_to lbl = stmtC (CmmJump (CmmLit lbl) [])
332 build_it_then return_code
333 = do { -- BUILD THE OBJECT IN THE HEAP
334 -- The first "con" says that the name bound to this
335 -- closure is "con", which is a bit of a fudge, but it only
338 -- This Id is also used to get a unique for a
339 -- temporary variable, if the closure is a CHARLIKE.
340 -- funnily enough, this makes the unique always come
342 tickyReturnNewCon (length amodes)
343 ; idinfo <- buildDynCon (dataConWorkId con) currentCCS con amodes
344 ; amode <- idInfoToAmode idinfo
345 ; checkedAbsC (CmmAssign nodeReg amode)
346 ; performReturn return_code }
350 %************************************************************************
352 Generating static stuff for algebraic data types
354 %************************************************************************
356 [These comments are rather out of date]
359 Info tbls & Macro & Kind of constructor \\
361 info & @CONST_INFO_TABLE@& Zero arity (no info -- compiler uses static closure)\\
362 info & @CHARLIKE_INFO_TABLE@& Charlike (no info -- compiler indexes fixed array)\\
363 info & @INTLIKE_INFO_TABLE@& Intlike; the one macro generates both info tbls\\
364 info & @SPEC_INFO_TABLE@& SPECish, and bigger than or equal to @MIN_UPD_SIZE@\\
365 info & @GEN_INFO_TABLE@& GENish (hence bigger than or equal to @MIN_UPD_SIZE@)\\
368 Possible info tables for constructor con:
372 Used for dynamically let(rec)-bound occurrences of
373 the constructor, and for updates. For constructors
374 which are int-like, char-like or nullary, when GC occurs,
375 the closure tries to get rid of itself.
377 \item[@_static_info@:]
378 Static occurrences of the constructor
379 macro: @STATIC_INFO_TABLE@.
382 For zero-arity constructors, \tr{con}, we NO LONGER generate a static closure;
383 it's place is taken by the top level defn of the constructor.
385 For charlike and intlike closures there is a fixed array of static
386 closures predeclared.
389 cgTyCon :: TyCon -> FCode [Cmm] -- each constructor gets a separate Cmm
391 = do { constrs <- mapM (getCmm . cgDataCon) (tyConDataCons tycon)
393 -- Generate a table of static closures for an enumeration type
394 -- Put the table after the data constructor decls, because the
395 -- datatype closure table (for enumeration types)
396 -- to (say) PrelBase_$wTrue_closure, which is defined in code_stuff
397 -- Note that the closure pointers are tagged.
399 -- XXX comment says to put table after constructor decls, but
400 -- code appears to put it before --- NR 16 Aug 2007
402 if isEnumerationTyCon tycon then do
403 tbl <- getCmm (emitRODataLits (mkLocalClosureTableLabel
405 [ CmmLabelOff (mkLocalClosureLabel (dataConName con)) (tagForCon con)
406 | con <- tyConDataCons tycon])
411 ; return (extra ++ constrs)
415 Generate the entry code, info tables, and (for niladic constructor) the
416 static closure, for a constructor.
419 cgDataCon :: DataCon -> Code
421 = do { -- Don't need any dynamic closure code for zero-arity constructors
424 -- To allow the debuggers, interpreters, etc to cope with
425 -- static data structures (ie those built at compile
426 -- time), we take care that info-table contains the
427 -- information we need.
428 (static_cl_info, _) =
429 layOutStaticConstr data_con arg_reps
431 (dyn_cl_info, arg_things) =
432 layOutDynConstr data_con arg_reps
434 emit_info cl_info ticky_code
435 = do { code_blks <- getCgStmts the_code
436 ; emitClosureCodeAndInfoTable cl_info [] code_blks }
438 the_code = do { ticky_code
439 ; ldvEnter (CmmReg nodeReg)
442 arg_reps :: [(CgRep, Type)]
443 arg_reps = [(typeCgRep ty, ty) | ty <- dataConRepArgTys data_con]
446 -- NB: We don't set CC when entering data (WDP 94/06)
447 tickyReturnOldCon (length arg_things)
448 -- The case continuation code is expecting a tagged pointer
449 ; stmtC (CmmAssign nodeReg
450 (tagCons data_con (CmmReg nodeReg)))
451 ; performReturn emitReturnInstr }
452 -- noStmts: Ptr to thing already in Node
454 ; whenC (not (isNullaryRepDataCon data_con))
455 (emit_info dyn_cl_info tickyEnterDynCon)
457 -- Dynamic-Closure first, to reduce forward references
458 ; emit_info static_cl_info tickyEnterStaticCon }