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"
56 %************************************************************************
58 \subsection[toplevel-constructors]{Top-level constructors}
60 %************************************************************************
63 cgTopRhsCon :: Id -- Name of thing bound to this RHS
66 -> FCode (Id, CgIdInfo)
67 cgTopRhsCon id con args
70 -- Windows DLLs have a problem with static cross-DLL refs.
71 ; this_pkg <- getThisPackage
72 ; ASSERT( not (isDllConApp this_pkg con args) ) return ()
74 ; ASSERT( args `lengthIs` dataConRepArity con ) return ()
77 ; amodes <- getArgAmodes args
81 lf_info = mkConLFInfo con
82 closure_label = mkClosureLabel name $ idCafInfo id
83 caffy = any stgArgHasCafRefs args
84 (closure_info, amodes_w_offsets) = layOutStaticConstr con amodes
85 closure_rep = mkStaticClosureFields
87 dontCareCCS -- Because it's static data
91 payload = map get_lit amodes_w_offsets
92 get_lit (CmmLit lit, _offset) = lit
93 get_lit other = pprPanic "CgCon.get_lit" (ppr other)
94 -- NB1: amodes_w_offsets is sorted into ptrs first, then non-ptrs
95 -- NB2: all the amodes should be Lits!
98 ; emitDataLits closure_label closure_rep
101 ; returnFC (id, taggedStableIdInfo id (mkLblExpr closure_label) lf_info con) }
104 %************************************************************************
106 %* non-top-level constructors *
108 %************************************************************************
109 \subsection[code-for-constructors]{The code for constructors}
112 buildDynCon :: Id -- Name of the thing to which this constr will
114 -> CostCentreStack -- Where to grab cost centre from;
115 -- current CCS if currentOrSubsumedCCS
116 -> DataCon -- The data constructor
117 -> [(CgRep,CmmExpr)] -- Its args
118 -> FCode CgIdInfo -- Return details about how to find it
120 -- We used to pass a boolean indicating whether all the
121 -- args were of size zero, so we could use a static
122 -- construtor; but I concluded that it just isn't worth it.
123 -- Now I/O uses unboxed tuples there just aren't any constructors
124 -- with all size-zero args.
126 -- The reason for having a separate argument, rather than looking at
127 -- the addr modes of the args is that we may be in a "knot", and
128 -- premature looking at the args will cause the compiler to black-hole!
131 First we deal with the case of zero-arity constructors. Now, they
132 will probably be unfolded, so we don't expect to see this case much,
133 if at all, but it does no harm, and sets the scene for characters.
135 In the case of zero-arity constructors, or, more accurately, those
136 which have exclusively size-zero (VoidRep) args, we generate no code
140 buildDynCon binder _ con []
141 = returnFC (taggedStableIdInfo binder
142 (mkLblExpr (mkClosureLabel (dataConName con)
148 The following three paragraphs about @Char@-like and @Int@-like
149 closures are obsolete, but I don't understand the details well enough
150 to properly word them, sorry. I've changed the treatment of @Char@s to
151 be analogous to @Int@s: only a subset is preallocated, because @Char@
152 has now 31 bits. Only literals are handled here. -- Qrczak
154 Now for @Char@-like closures. We generate an assignment of the
155 address of the closure to a temporary. It would be possible simply to
156 generate no code, and record the addressing mode in the environment,
157 but we'd have to be careful if the argument wasn't a constant --- so
158 for simplicity we just always asssign to a temporary.
160 Last special case: @Int@-like closures. We only special-case the
161 situation in which the argument is a literal in the range
162 @mIN_INTLIKE@..@mAX_INTLILKE@. NB: for @Char@-like closures we can
163 work with any old argument, but for @Int@-like ones the argument has
164 to be a literal. Reason: @Char@ like closures have an argument type
165 which is guaranteed in range.
167 Because of this, we use can safely return an addressing mode.
170 buildDynCon binder _ con [arg_amode]
171 | maybeIntLikeCon con
172 , (_, CmmLit (CmmInt val _)) <- arg_amode
173 , let val_int = (fromIntegral val) :: Int
174 , val_int <= mAX_INTLIKE && val_int >= mIN_INTLIKE
175 = do { let intlike_lbl = mkRtsGcPtrLabel (sLit "stg_INTLIKE_closure")
176 offsetW = (val_int - mIN_INTLIKE) * (fixedHdrSize + 1)
177 -- INTLIKE closures consist of a header and one word payload
178 intlike_amode = CmmLit (cmmLabelOffW intlike_lbl offsetW)
179 ; returnFC (taggedStableIdInfo binder intlike_amode (mkConLFInfo con) con) }
181 buildDynCon binder _ con [arg_amode]
182 | maybeCharLikeCon con
183 , (_, CmmLit (CmmInt val _)) <- arg_amode
184 , let val_int = (fromIntegral val) :: Int
185 , val_int <= mAX_CHARLIKE && val_int >= mIN_CHARLIKE
186 = do { let charlike_lbl = mkRtsGcPtrLabel (sLit "stg_CHARLIKE_closure")
187 offsetW = (val_int - mIN_CHARLIKE) * (fixedHdrSize + 1)
188 -- CHARLIKE closures consist of a header and one word payload
189 charlike_amode = CmmLit (cmmLabelOffW charlike_lbl offsetW)
190 ; returnFC (taggedStableIdInfo binder charlike_amode (mkConLFInfo con) con) }
193 Now the general case.
196 buildDynCon binder ccs con args
199 (closure_info, amodes_w_offsets) = layOutDynConstr con args
201 ; hp_off <- allocDynClosure closure_info use_cc blame_cc amodes_w_offsets
202 ; returnFC (taggedHeapIdInfo binder hp_off lf_info con) }
204 lf_info = mkConLFInfo con
206 use_cc -- cost-centre to stick in the object
207 | currentOrSubsumedCCS ccs = curCCS
208 | otherwise = CmmLit (mkCCostCentreStack ccs)
210 blame_cc = use_cc -- cost-centre on which to blame the alloc (same)
214 %************************************************************************
216 %* constructor-related utility function: *
217 %* bindConArgs is called from cgAlt of a case *
219 %************************************************************************
220 \subsection[constructor-utilities]{@bindConArgs@: constructor-related utility}
222 @bindConArgs@ $con args$ augments the environment with bindings for the
223 binders $args$, assuming that we have just returned from a @case@ which
227 bindConArgs :: DataCon -> [Id] -> Code
231 -- The binding below forces the masking out of the tag bits
232 -- when accessing the constructor field.
233 bind_arg (arg, offset) = bindNewToUntagNode arg offset (mkLFArgument arg) (tagForCon con)
234 (_, args_w_offsets) = layOutDynConstr con (addIdReps args)
236 ASSERT(not (isUnboxedTupleCon con)) return ()
237 mapCs bind_arg args_w_offsets
240 Unboxed tuples are handled slightly differently - the object is
241 returned in registers and on the stack instead of the heap.
244 bindUnboxedTupleComponents
246 -> FCode ([(Id,GlobalReg)], -- Regs assigned
247 WordOff, -- Number of pointer stack slots
248 WordOff, -- Number of non-pointer stack slots
249 VirtualSpOffset) -- Offset of return address slot
250 -- (= realSP on entry)
252 bindUnboxedTupleComponents args
257 -- Assign as many components as possible to registers
258 ; let (reg_args, stk_args) = assignReturnRegs (addIdReps args)
260 -- Separate the rest of the args into pointers and non-pointers
261 (ptr_args, nptr_args) = separateByPtrFollowness stk_args
263 -- Allocate the rest on the stack
264 -- The real SP points to the return address, above which any
265 -- leftover unboxed-tuple components will be allocated
266 (ptr_sp, ptr_offsets) = mkVirtStkOffsets rsp ptr_args
267 (nptr_sp, nptr_offsets) = mkVirtStkOffsets ptr_sp nptr_args
269 nptrs = nptr_sp - ptr_sp
271 -- The stack pointer points to the last stack-allocated component
272 ; setRealAndVirtualSp nptr_sp
274 -- We have just allocated slots starting at real SP + 1, and set the new
275 -- virtual SP to the topmost allocated slot.
276 -- If the virtual SP started *below* the real SP, we've just jumped over
277 -- some slots that won't be in the free-list, so put them there
278 -- This commonly happens because we've freed the return-address slot
279 -- (trimming back the virtual SP), but the real SP still points to that slot
280 ; freeStackSlots [vsp+1,vsp+2 .. rsp]
282 ; bindArgsToRegs reg_args
283 ; bindArgsToStack ptr_offsets
284 ; bindArgsToStack nptr_offsets
286 ; returnFC (reg_args, ptrs, nptrs, rsp) }
289 %************************************************************************
291 Actually generate code for a constructor return
293 %************************************************************************
296 Note: it's the responsibility of the @cgReturnDataCon@ caller to be
297 sure the @amodes@ passed don't conflict with each other.
299 cgReturnDataCon :: DataCon -> [(CgRep, CmmExpr)] -> Code
301 cgReturnDataCon con amodes
302 | isUnboxedTupleCon con = returnUnboxedTuple amodes
303 -- when profiling we can't shortcut here, we have to enter the closure
304 -- for it to be marked as "used" for LDV profiling.
305 | opt_SccProfilingOn = build_it_then enter_it
307 = ASSERT( amodes `lengthIs` dataConRepArity con )
308 do { EndOfBlockInfo _ sequel <- getEndOfBlockInfo
310 CaseAlts _ (Just (alts, deflt_lbl)) bndr
311 -> -- Ho! We know the constructor so we can
312 -- go straight to the right alternative
313 case assocMaybe alts (dataConTagZ con) of {
314 Just join_lbl -> build_it_then (jump_to join_lbl);
316 -- Special case! We're returning a constructor to the default case
317 -- of an enclosing case. For example:
319 -- case (case e of (a,b) -> C a b) of
321 -- y -> ...<returning here!>...
324 -- if the default is a non-bind-default (ie does not use y),
325 -- then we should simply jump to the default join point;
327 | isDeadBinder bndr -> performReturn (jump_to deflt_lbl)
328 | otherwise -> build_it_then (jump_to deflt_lbl) }
330 _otherwise -- The usual case
331 -> build_it_then emitReturnInstr
334 enter_it = stmtsC [ CmmAssign nodeReg (cmmUntag (CmmReg nodeReg)),
335 CmmJump (entryCode (closureInfoPtr (CmmReg nodeReg))) [] ]
336 jump_to lbl = stmtC (CmmJump (CmmLit lbl) [])
337 build_it_then return_code
338 = do { -- BUILD THE OBJECT IN THE HEAP
339 -- The first "con" says that the name bound to this
340 -- closure is "con", which is a bit of a fudge, but it only
343 -- This Id is also used to get a unique for a
344 -- temporary variable, if the closure is a CHARLIKE.
345 -- funnily enough, this makes the unique always come
347 tickyReturnNewCon (length amodes)
348 ; idinfo <- buildDynCon (dataConWorkId con) currentCCS con amodes
349 ; amode <- idInfoToAmode idinfo
350 ; checkedAbsC (CmmAssign nodeReg amode)
351 ; performReturn return_code }
355 %************************************************************************
357 Generating static stuff for algebraic data types
359 %************************************************************************
361 [These comments are rather out of date]
364 Info tbls & Macro & Kind of constructor \\
366 info & @CONST_INFO_TABLE@& Zero arity (no info -- compiler uses static closure)\\
367 info & @CHARLIKE_INFO_TABLE@& Charlike (no info -- compiler indexes fixed array)\\
368 info & @INTLIKE_INFO_TABLE@& Intlike; the one macro generates both info tbls\\
369 info & @SPEC_INFO_TABLE@& SPECish, and bigger than or equal to @MIN_UPD_SIZE@\\
370 info & @GEN_INFO_TABLE@& GENish (hence bigger than or equal to @MIN_UPD_SIZE@)\\
373 Possible info tables for constructor con:
377 Used for dynamically let(rec)-bound occurrences of
378 the constructor, and for updates. For constructors
379 which are int-like, char-like or nullary, when GC occurs,
380 the closure tries to get rid of itself.
382 \item[@_static_info@:]
383 Static occurrences of the constructor
384 macro: @STATIC_INFO_TABLE@.
387 For zero-arity constructors, \tr{con}, we NO LONGER generate a static closure;
388 it's place is taken by the top level defn of the constructor.
390 For charlike and intlike closures there is a fixed array of static
391 closures predeclared.
394 cgTyCon :: TyCon -> FCode [Cmm] -- each constructor gets a separate Cmm
396 = do { constrs <- mapM (getCmm . cgDataCon) (tyConDataCons tycon)
398 -- Generate a table of static closures for an enumeration type
399 -- Put the table after the data constructor decls, because the
400 -- datatype closure table (for enumeration types)
401 -- to (say) PrelBase_$wTrue_closure, which is defined in code_stuff
402 -- Note that the closure pointers are tagged.
404 -- XXX comment says to put table after constructor decls, but
405 -- code appears to put it before --- NR 16 Aug 2007
407 if isEnumerationTyCon tycon then do
408 tbl <- getCmm (emitRODataLits "cgTyCon" (mkLocalClosureTableLabel (tyConName tycon) NoCafRefs)
409 [ CmmLabelOff (mkLocalClosureLabel (dataConName con) NoCafRefs) (tagForCon con)
410 | con <- tyConDataCons tycon])
415 ; return (extra ++ constrs)
419 Generate the entry code, info tables, and (for niladic constructor) the
420 static closure, for a constructor.
423 cgDataCon :: DataCon -> Code
425 = do { -- Don't need any dynamic closure code for zero-arity constructors
428 -- To allow the debuggers, interpreters, etc to cope with
429 -- static data structures (ie those built at compile
430 -- time), we take care that info-table contains the
431 -- information we need.
432 (static_cl_info, _) =
433 layOutStaticConstr data_con arg_reps
435 (dyn_cl_info, arg_things) =
436 layOutDynConstr data_con arg_reps
438 emit_info cl_info ticky_code
439 = do { code_blks <- getCgStmts the_code
440 ; emitClosureCodeAndInfoTable cl_info [] code_blks }
442 the_code = do { ticky_code
443 ; ldvEnter (CmmReg nodeReg)
446 arg_reps :: [(CgRep, Type)]
447 arg_reps = [(typeCgRep ty, ty) | ty <- dataConRepArgTys data_con]
450 -- NB: We don't set CC when entering data (WDP 94/06)
451 tickyReturnOldCon (length arg_things)
452 -- The case continuation code is expecting a tagged pointer
453 ; stmtC (CmmAssign nodeReg
454 (tagCons data_con (CmmReg nodeReg)))
455 ; performReturn emitReturnInstr }
456 -- noStmts: Ptr to thing already in Node
458 ; whenC (not (isNullaryRepDataCon data_con))
459 (emit_info dyn_cl_info tickyEnterDynCon)
461 -- Dynamic-Closure first, to reduce forward references
462 ; emit_info static_cl_info tickyEnterStaticCon }