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.
168 We don't support this optimisation when compiling into Windows DLLs yet
169 because they don't support cross package data references well.
174 buildDynCon binder _ con [arg_amode]
175 | maybeIntLikeCon con
176 #if defined(mingw32_TARGET_OS)
179 , (_, CmmLit (CmmInt val _)) <- arg_amode
180 , let val_int = (fromIntegral val) :: Int
181 , val_int <= mAX_INTLIKE && val_int >= mIN_INTLIKE
182 = do { let intlike_lbl = mkCmmGcPtrLabel rtsPackageId (fsLit "stg_INTLIKE_closure")
183 offsetW = (val_int - mIN_INTLIKE) * (fixedHdrSize + 1)
184 -- INTLIKE closures consist of a header and one word payload
185 intlike_amode = CmmLit (cmmLabelOffW intlike_lbl offsetW)
186 ; returnFC (taggedStableIdInfo binder intlike_amode (mkConLFInfo con) con) }
188 buildDynCon binder _ con [arg_amode]
189 | maybeCharLikeCon con
190 #if defined(mingw32_TARGET_OS)
193 , (_, CmmLit (CmmInt val _)) <- arg_amode
194 , let val_int = (fromIntegral val) :: Int
195 , val_int <= mAX_CHARLIKE && val_int >= mIN_CHARLIKE
196 = do { let charlike_lbl = mkCmmGcPtrLabel rtsPackageId (fsLit "stg_CHARLIKE_closure")
197 offsetW = (val_int - mIN_CHARLIKE) * (fixedHdrSize + 1)
198 -- CHARLIKE closures consist of a header and one word payload
199 charlike_amode = CmmLit (cmmLabelOffW charlike_lbl offsetW)
200 ; returnFC (taggedStableIdInfo binder charlike_amode (mkConLFInfo con) con) }
204 Now the general case.
207 buildDynCon binder ccs con args
210 (closure_info, amodes_w_offsets) = layOutDynConstr con args
212 ; hp_off <- allocDynClosure closure_info use_cc blame_cc amodes_w_offsets
213 ; returnFC (taggedHeapIdInfo binder hp_off lf_info con) }
215 lf_info = mkConLFInfo con
217 use_cc -- cost-centre to stick in the object
218 | currentOrSubsumedCCS ccs = curCCS
219 | otherwise = CmmLit (mkCCostCentreStack ccs)
221 blame_cc = use_cc -- cost-centre on which to blame the alloc (same)
225 %************************************************************************
227 %* constructor-related utility function: *
228 %* bindConArgs is called from cgAlt of a case *
230 %************************************************************************
231 \subsection[constructor-utilities]{@bindConArgs@: constructor-related utility}
233 @bindConArgs@ $con args$ augments the environment with bindings for the
234 binders $args$, assuming that we have just returned from a @case@ which
238 bindConArgs :: DataCon -> [Id] -> Code
242 -- The binding below forces the masking out of the tag bits
243 -- when accessing the constructor field.
244 bind_arg (arg, offset) = bindNewToUntagNode arg offset (mkLFArgument arg) (tagForCon con)
245 (_, args_w_offsets) = layOutDynConstr con (addIdReps args)
247 ASSERT(not (isUnboxedTupleCon con)) return ()
248 mapCs bind_arg args_w_offsets
251 Unboxed tuples are handled slightly differently - the object is
252 returned in registers and on the stack instead of the heap.
255 bindUnboxedTupleComponents
257 -> FCode ([(Id,GlobalReg)], -- Regs assigned
258 WordOff, -- Number of pointer stack slots
259 WordOff, -- Number of non-pointer stack slots
260 VirtualSpOffset) -- Offset of return address slot
261 -- (= realSP on entry)
263 bindUnboxedTupleComponents args
268 -- Assign as many components as possible to registers
269 ; let (reg_args, stk_args) = assignReturnRegs (addIdReps args)
271 -- Separate the rest of the args into pointers and non-pointers
272 (ptr_args, nptr_args) = separateByPtrFollowness stk_args
274 -- Allocate the rest on the stack
275 -- The real SP points to the return address, above which any
276 -- leftover unboxed-tuple components will be allocated
277 (ptr_sp, ptr_offsets) = mkVirtStkOffsets rsp ptr_args
278 (nptr_sp, nptr_offsets) = mkVirtStkOffsets ptr_sp nptr_args
280 nptrs = nptr_sp - ptr_sp
282 -- The stack pointer points to the last stack-allocated component
283 ; setRealAndVirtualSp nptr_sp
285 -- We have just allocated slots starting at real SP + 1, and set the new
286 -- virtual SP to the topmost allocated slot.
287 -- If the virtual SP started *below* the real SP, we've just jumped over
288 -- some slots that won't be in the free-list, so put them there
289 -- This commonly happens because we've freed the return-address slot
290 -- (trimming back the virtual SP), but the real SP still points to that slot
291 ; freeStackSlots [vsp+1,vsp+2 .. rsp]
293 ; bindArgsToRegs reg_args
294 ; bindArgsToStack ptr_offsets
295 ; bindArgsToStack nptr_offsets
297 ; returnFC (reg_args, ptrs, nptrs, rsp) }
300 %************************************************************************
302 Actually generate code for a constructor return
304 %************************************************************************
307 Note: it's the responsibility of the @cgReturnDataCon@ caller to be
308 sure the @amodes@ passed don't conflict with each other.
310 cgReturnDataCon :: DataCon -> [(CgRep, CmmExpr)] -> Code
312 cgReturnDataCon con amodes
313 | isUnboxedTupleCon con = returnUnboxedTuple amodes
314 -- when profiling we can't shortcut here, we have to enter the closure
315 -- for it to be marked as "used" for LDV profiling.
316 | opt_SccProfilingOn = build_it_then enter_it
318 = ASSERT( amodes `lengthIs` dataConRepArity con )
319 do { EndOfBlockInfo _ sequel <- getEndOfBlockInfo
321 CaseAlts _ (Just (alts, deflt_lbl)) bndr
322 -> -- Ho! We know the constructor so we can
323 -- go straight to the right alternative
324 case assocMaybe alts (dataConTagZ con) of {
325 Just join_lbl -> build_it_then (jump_to join_lbl);
327 -- Special case! We're returning a constructor to the default case
328 -- of an enclosing case. For example:
330 -- case (case e of (a,b) -> C a b) of
332 -- y -> ...<returning here!>...
335 -- if the default is a non-bind-default (ie does not use y),
336 -- then we should simply jump to the default join point;
338 | isDeadBinder bndr -> performReturn (jump_to deflt_lbl)
339 | otherwise -> build_it_then (jump_to deflt_lbl) }
341 _otherwise -- The usual case
342 -> build_it_then emitReturnInstr
345 enter_it = stmtsC [ CmmAssign nodeReg (cmmUntag (CmmReg nodeReg)),
346 CmmJump (entryCode (closureInfoPtr (CmmReg nodeReg))) [] ]
347 jump_to lbl = stmtC (CmmJump (CmmLit lbl) [])
348 build_it_then return_code
349 = do { -- BUILD THE OBJECT IN THE HEAP
350 -- The first "con" says that the name bound to this
351 -- closure is "con", which is a bit of a fudge, but it only
354 -- This Id is also used to get a unique for a
355 -- temporary variable, if the closure is a CHARLIKE.
356 -- funnily enough, this makes the unique always come
358 tickyReturnNewCon (length amodes)
359 ; idinfo <- buildDynCon (dataConWorkId con) currentCCS con amodes
360 ; amode <- idInfoToAmode idinfo
361 ; checkedAbsC (CmmAssign nodeReg amode)
362 ; performReturn return_code }
366 %************************************************************************
368 Generating static stuff for algebraic data types
370 %************************************************************************
372 [These comments are rather out of date]
375 Info tbls & Macro & Kind of constructor \\
377 info & @CONST_INFO_TABLE@& Zero arity (no info -- compiler uses static closure)\\
378 info & @CHARLIKE_INFO_TABLE@& Charlike (no info -- compiler indexes fixed array)\\
379 info & @INTLIKE_INFO_TABLE@& Intlike; the one macro generates both info tbls\\
380 info & @SPEC_INFO_TABLE@& SPECish, and bigger than or equal to @MIN_UPD_SIZE@\\
381 info & @GEN_INFO_TABLE@& GENish (hence bigger than or equal to @MIN_UPD_SIZE@)\\
384 Possible info tables for constructor con:
388 Used for dynamically let(rec)-bound occurrences of
389 the constructor, and for updates. For constructors
390 which are int-like, char-like or nullary, when GC occurs,
391 the closure tries to get rid of itself.
393 \item[@_static_info@:]
394 Static occurrences of the constructor
395 macro: @STATIC_INFO_TABLE@.
398 For zero-arity constructors, \tr{con}, we NO LONGER generate a static closure;
399 it's place is taken by the top level defn of the constructor.
401 For charlike and intlike closures there is a fixed array of static
402 closures predeclared.
405 cgTyCon :: TyCon -> FCode [Cmm] -- each constructor gets a separate Cmm
407 = do { constrs <- mapM (getCmm . cgDataCon) (tyConDataCons tycon)
409 -- Generate a table of static closures for an enumeration type
410 -- Put the table after the data constructor decls, because the
411 -- datatype closure table (for enumeration types)
412 -- to (say) PrelBase_$wTrue_closure, which is defined in code_stuff
413 -- Note that the closure pointers are tagged.
415 -- XXX comment says to put table after constructor decls, but
416 -- code appears to put it before --- NR 16 Aug 2007
418 if isEnumerationTyCon tycon then do
419 tbl <- getCmm (emitRODataLits "cgTyCon" (mkLocalClosureTableLabel (tyConName tycon) NoCafRefs)
420 [ CmmLabelOff (mkLocalClosureLabel (dataConName con) NoCafRefs) (tagForCon con)
421 | con <- tyConDataCons tycon])
426 ; return (extra ++ constrs)
430 Generate the entry code, info tables, and (for niladic constructor) the
431 static closure, for a constructor.
434 cgDataCon :: DataCon -> Code
436 = do { -- Don't need any dynamic closure code for zero-arity constructors
439 -- To allow the debuggers, interpreters, etc to cope with
440 -- static data structures (ie those built at compile
441 -- time), we take care that info-table contains the
442 -- information we need.
443 (static_cl_info, _) =
444 layOutStaticConstr data_con arg_reps
446 (dyn_cl_info, arg_things) =
447 layOutDynConstr data_con arg_reps
449 emit_info cl_info ticky_code
450 = do { code_blks <- getCgStmts the_code
451 ; emitClosureCodeAndInfoTable cl_info [] code_blks }
453 the_code = do { _ <- ticky_code
454 ; ldvEnter (CmmReg nodeReg)
457 arg_reps :: [(CgRep, Type)]
458 arg_reps = [(typeCgRep ty, ty) | ty <- dataConRepArgTys data_con]
461 -- NB: We don't set CC when entering data (WDP 94/06)
462 tickyReturnOldCon (length arg_things)
463 -- The case continuation code is expecting a tagged pointer
464 ; stmtC (CmmAssign nodeReg
465 (tagCons data_con (CmmReg nodeReg)))
466 ; performReturn emitReturnInstr }
467 -- noStmts: Ptr to thing already in Node
469 ; whenC (not (isNullaryRepDataCon data_con))
470 (emit_info dyn_cl_info tickyEnterDynCon)
472 -- Dynamic-Closure first, to reduce forward references
473 ; emit_info static_cl_info tickyEnterStaticCon }