2 % (c) The University of Glasgow 2006
3 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
5 \section[CgClosure]{Code generation for closures}
7 This module provides the support code for @StgToAbstractC@ to deal
8 with {\em closures} on the RHSs of let(rec)s. See also
9 @CgCon@, which deals with constructors.
12 module CgClosure ( cgTopRhsClosure,
18 #include "HsVersions.h"
20 import {-# SOURCE #-} CgExpr ( cgExpr )
53 %********************************************************
55 \subsection[closures-no-free-vars]{Top-level closures}
57 %********************************************************
59 For closures bound at top level, allocate in static space.
60 They should have no free variables.
64 -> CostCentreStack -- Optional cost centre annotation
69 -> FCode (Id, CgIdInfo)
71 cgTopRhsClosure id ccs binder_info upd_flag args body = do
72 { -- LAY OUT THE OBJECT
74 ; lf_info <- mkClosureLFInfo id TopLevel [] upd_flag args
75 ; srt_info <- getSRTInfo
76 ; mod_name <- getModuleName
77 ; let descr = closureDescription mod_name name
78 closure_info = mkClosureInfo True id lf_info 0 0 srt_info descr
79 closure_label = mkLocalClosureLabel name
80 cg_id_info = stableIdInfo id (mkLblExpr closure_label) lf_info
81 closure_rep = mkStaticClosureFields closure_info ccs True []
83 -- BUILD THE OBJECT, AND GENERATE INFO TABLE (IF NECESSARY)
84 ; emitDataLits closure_label closure_rep
85 ; forkClosureBody (closureCodeBody binder_info closure_info
88 ; returnFC (id, cg_id_info) }
91 %********************************************************
93 \subsection[non-top-level-closures]{Non top-level closures}
95 %********************************************************
97 For closures with free vars, allocate in heap.
102 -> CostCentreStack -- Optional cost centre annotation
108 -> [StgArg] -- payload
109 -> FCode (Id, CgIdInfo)
111 cgStdRhsClosure bndr cc bndr_info fvs args body lf_info payload
112 = do -- AHA! A STANDARD-FORM THUNK
113 { -- LAY OUT THE OBJECT
114 amodes <- getArgAmodes payload
115 ; mod_name <- getModuleName
116 ; let (tot_wds, ptr_wds, amodes_w_offsets)
117 = mkVirtHeapOffsets (isLFThunk lf_info) amodes
119 descr = closureDescription mod_name (idName bndr)
120 closure_info = mkClosureInfo False -- Not static
121 bndr lf_info tot_wds ptr_wds
122 NoC_SRT -- No SRT for a std-form closure
125 ; (use_cc, blame_cc) <- chooseDynCostCentres cc args body
128 ; heap_offset <- allocDynClosure closure_info use_cc blame_cc amodes_w_offsets
131 ; returnFC (bndr, heapIdInfo bndr heap_offset lf_info) }
134 Here's the general case.
138 -> CostCentreStack -- Optional cost centre annotation
144 -> FCode (Id, CgIdInfo)
146 cgRhsClosure bndr cc bndr_info fvs upd_flag args body = do
147 { -- LAY OUT THE OBJECT
148 -- If the binder is itself a free variable, then don't store
149 -- it in the closure. Instead, just bind it to Node on entry.
150 -- NB we can be sure that Node will point to it, because we
151 -- havn't told mkClosureLFInfo about this; so if the binder
152 -- _was_ a free var of its RHS, mkClosureLFInfo thinks it *is*
153 -- stored in the closure itself, so it will make sure that
154 -- Node points to it...
157 is_elem = isIn "cgRhsClosure"
158 bndr_is_a_fv = bndr `is_elem` fvs
159 reduced_fvs | bndr_is_a_fv = fvs `minusList` [bndr]
162 ; lf_info <- mkClosureLFInfo bndr NotTopLevel fvs upd_flag args
163 ; fv_infos <- mapFCs getCgIdInfo reduced_fvs
164 ; srt_info <- getSRTInfo
165 ; mod_name <- getModuleName
166 ; let bind_details :: [(CgIdInfo, VirtualHpOffset)]
167 (tot_wds, ptr_wds, bind_details)
168 = mkVirtHeapOffsets (isLFThunk lf_info) (map add_rep fv_infos)
170 add_rep info = (cgIdInfoArgRep info, info)
172 descr = closureDescription mod_name name
173 closure_info = mkClosureInfo False -- Not static
174 bndr lf_info tot_wds ptr_wds
177 -- BUILD ITS INFO TABLE AND CODE
178 ; forkClosureBody (do
181 -- A function closure pointer may be tagged, so we
182 -- must take it into account when accessing the free variables.
183 mbtag = tagForArity (length args)
184 bind_fv (info, offset)
186 = bindNewToUntagNode (cgIdInfoId info) offset (cgIdInfoLF info) tag
188 = bindNewToNode (cgIdInfoId info) offset (cgIdInfoLF info)
189 ; mapCs bind_fv bind_details
191 -- Bind the binder itself, if it is a free var
192 ; whenC bndr_is_a_fv (bindNewToReg bndr nodeReg lf_info)
195 ; closureCodeBody bndr_info closure_info cc args body })
199 to_amode (info, offset) = do { amode <- idInfoToAmode info
200 ; return (amode, offset) }
201 ; (use_cc, blame_cc) <- chooseDynCostCentres cc args body
202 ; amodes_w_offsets <- mapFCs to_amode bind_details
203 ; heap_offset <- allocDynClosure closure_info use_cc blame_cc amodes_w_offsets
206 ; returnFC (bndr, heapIdInfo bndr heap_offset lf_info) }
209 mkClosureLFInfo :: Id -- The binder
210 -> TopLevelFlag -- True of top level
212 -> UpdateFlag -- Update flag
214 -> FCode LambdaFormInfo
215 mkClosureLFInfo bndr top fvs upd_flag args
216 | null args = return (mkLFThunk (idType bndr) top fvs upd_flag)
217 | otherwise = do { arg_descr <- mkArgDescr (idName bndr) args
218 ; return (mkLFReEntrant top fvs args arg_descr) }
222 %************************************************************************
224 \subsection[code-for-closures]{The code for closures}
226 %************************************************************************
229 closureCodeBody :: StgBinderInfo
230 -> ClosureInfo -- Lots of information about this closure
231 -> CostCentreStack -- Optional cost centre attached to closure
237 There are two main cases for the code for closures. If there are {\em
238 no arguments}, then the closure is a thunk, and not in normal form.
239 So it should set up an update frame (if it is shared).
240 NB: Thunks cannot have a primitive type!
243 closureCodeBody binder_info cl_info cc [{- No args i.e. thunk -}] body = do
244 { body_absC <- getCgStmts $ do
245 { tickyEnterThunk cl_info
246 ; ldvEnterClosure cl_info -- NB: Node always points when profiling
247 ; thunkWrapper cl_info $ do
248 -- We only enter cc after setting up update so
249 -- that cc of enclosing scope will be recorded
250 -- in update frame CAF/DICT functions will be
251 -- subsumed by this enclosing cc
252 { enterCostCentre cl_info cc body
256 ; emitClosureCodeAndInfoTable cl_info [] body_absC }
259 If there is /at least one argument/, then this closure is in
260 normal form, so there is no need to set up an update frame.
262 The Macros for GrAnSim are produced at the beginning of the
263 argSatisfactionCheck (by calling fetchAndReschedule). There info if
264 Node points to closure is available. -- HWL
267 closureCodeBody binder_info cl_info cc args body
268 = ASSERT( length args > 0 )
269 do { -- Get the current virtual Sp (it might not be zero,
270 -- eg. if we're compiling a let-no-escape).
272 ; let (reg_args, other_args) = assignCallRegs (addIdReps args)
273 (sp_top, stk_args) = mkVirtStkOffsets vSp other_args
275 -- Allocate the global ticky counter
276 ; let ticky_ctr_lbl = mkRednCountsLabel (closureName cl_info)
277 ; emitTickyCounter cl_info args sp_top
279 -- ...and establish the ticky-counter
280 -- label for this block
281 ; setTickyCtrLabel ticky_ctr_lbl $ do
283 -- Emit the slow-entry code
284 { reg_save_code <- mkSlowEntryCode cl_info reg_args
286 -- Emit the main entry code
288 mkFunEntryCode cl_info cc reg_args stk_args
289 sp_top reg_save_code body
290 ; emitClosureCodeAndInfoTable cl_info [] blks
295 mkFunEntryCode :: ClosureInfo
297 -> [(Id,GlobalReg)] -- Args in regs
298 -> [(Id,VirtualSpOffset)] -- Args on stack
299 -> VirtualSpOffset -- Last allocated word on stack
300 -> CmmStmts -- Register-save code in case of GC
303 -- The main entry code for the closure
304 mkFunEntryCode cl_info cc reg_args stk_args sp_top reg_save_code body = do
305 { -- Bind args to regs/stack as appropriate,
306 -- and record expected position of sps
307 ; bindArgsToRegs reg_args
308 ; bindArgsToStack stk_args
309 ; setRealAndVirtualSp sp_top
311 -- Enter the cost-centre, if required
312 -- ToDo: It's not clear why this is outside the funWrapper,
313 -- but the tickyEnterFun is inside. Perhaps we can put
315 ; enterCostCentre cl_info cc body
318 ; funWrapper cl_info reg_args reg_save_code $ do
319 { tickyEnterFun cl_info
324 The "slow entry" code for a function. This entry point takes its
325 arguments on the stack. It loads the arguments into registers
326 according to the calling convention, and jumps to the function's
327 normal entry point. The function's closure is assumed to be in
330 The slow entry point is used in two places:
332 (a) unknown calls: eg. stg_PAP_entry
333 (b) returning from a heap-check failure
336 mkSlowEntryCode :: ClosureInfo -> [(Id,GlobalReg)] -> FCode CmmStmts
337 -- If this function doesn't have a specialised ArgDescr, we need
338 -- to generate the function's arg bitmap, slow-entry code, and
339 -- register-save code for the heap-check failure
340 -- Here, we emit the slow-entry code, and
341 -- return the register-save assignments
342 mkSlowEntryCode cl_info reg_args
343 | Just (_, ArgGen _) <- closureFunInfo cl_info
344 = do { emitSimpleProc slow_lbl (emitStmts load_stmts)
345 ; return save_stmts }
346 | otherwise = return noStmts
348 name = closureName cl_info
349 slow_lbl = mkSlowEntryLabel name
351 load_stmts = mkStmts load_assts `plusStmts` mkStmts [stk_adj_pop, jump_to_entry]
352 save_stmts = oneStmt stk_adj_push `plusStmts` mkStmts save_assts
354 reps_w_regs :: [(CgRep,GlobalReg)]
355 reps_w_regs = [(idCgRep id, reg) | (id,reg) <- reverse reg_args]
356 (final_stk_offset, stk_offsets)
357 = mapAccumL (\off (rep,_) -> (off + cgRepSizeW rep, off))
360 load_assts = zipWithEqual "mk_load" mk_load reps_w_regs stk_offsets
361 mk_load (rep,reg) offset = CmmAssign (CmmGlobal reg)
362 (CmmLoad (cmmRegOffW spReg offset)
365 save_assts = zipWithEqual "mk_save" mk_save reps_w_regs stk_offsets
366 mk_save (rep,reg) offset = ASSERT( argMachRep rep == globalRegRep reg )
367 CmmStore (cmmRegOffW spReg offset)
368 (CmmReg (CmmGlobal reg))
370 stk_adj_pop = CmmAssign spReg (cmmRegOffW spReg final_stk_offset)
371 stk_adj_push = CmmAssign spReg (cmmRegOffW spReg (- final_stk_offset))
372 jump_to_entry = CmmJump (mkLblExpr (enterLocalIdLabel name)) []
376 %************************************************************************
378 \subsubsection[closure-code-wrappers]{Wrappers around closure code}
380 %************************************************************************
383 thunkWrapper:: ClosureInfo -> Code -> Code
384 thunkWrapper closure_info thunk_code = do
385 { let node_points = nodeMustPointToIt (closureLFInfo closure_info)
387 -- HWL: insert macros for GrAnSim; 2 versions depending on liveness of node
388 -- (we prefer fetchAndReschedule-style context switches to yield ones)
390 then granFetchAndReschedule [] node_points
391 else granYield [] node_points
393 -- Stack and/or heap checks
394 ; thunkEntryChecks closure_info $ do
395 { -- Overwrite with black hole if necessary
396 whenC (blackHoleOnEntry closure_info && node_points)
397 (blackHoleIt closure_info)
398 ; setupUpdate closure_info thunk_code }
399 -- setupUpdate *encloses* the thunk_code
402 funWrapper :: ClosureInfo -- Closure whose code body this is
403 -> [(Id,GlobalReg)] -- List of argument registers (if any)
404 -> CmmStmts -- reg saves for the heap check failure
405 -> Code -- Body of function being compiled
407 funWrapper closure_info arg_regs reg_save_code fun_body = do
408 { let node_points = nodeMustPointToIt (closureLFInfo closure_info)
411 -- Debugging: check that R1 has the correct tag
412 ; let tag = funTag closure_info
413 ; whenC (tag /= 0 && node_points) $ do
415 stmtC (CmmCondBranch (CmmMachOp mo_wordEq [cmmGetTag (CmmReg nodeReg),
416 CmmLit (mkIntCLit tag)]) l)
417 stmtC (CmmStore (CmmLit (mkWordCLit 0)) (CmmLit (mkWordCLit 0)))
421 -- Enter for Ldv profiling
422 ; whenC node_points (ldvEnterClosure closure_info)
424 -- GranSim yeild poin
425 ; granYield arg_regs node_points
427 -- Heap and/or stack checks wrap the function body
428 ; funEntryChecks closure_info reg_save_code
434 %************************************************************************
436 \subsubsubsection[update-and-BHs]{Update and black-hole wrappers}
438 %************************************************************************
442 blackHoleIt :: ClosureInfo -> Code
443 -- Only called for closures with no args
444 -- Node points to the closure
445 blackHoleIt closure_info = emitBlackHoleCode (closureSingleEntry closure_info)
447 emitBlackHoleCode :: Bool -> Code
448 emitBlackHoleCode is_single_entry
449 | eager_blackholing = do
450 tickyBlackHole (not is_single_entry)
451 stmtC (CmmStore (CmmReg nodeReg) (CmmLit (CmmLabel bh_lbl)))
455 bh_lbl | is_single_entry = mkRtsDataLabel SLIT("stg_SE_BLACKHOLE_info")
456 | otherwise = mkRtsDataLabel SLIT("stg_BLACKHOLE_info")
458 -- If we wanted to do eager blackholing with slop filling,
459 -- we'd need to do it at the *end* of a basic block, otherwise
460 -- we overwrite the free variables in the thunk that we still
461 -- need. We have a patch for this from Andy Cheadle, but not
462 -- incorporated yet. --SDM [6/2004]
464 -- Profiling needs slop filling (to support LDV profiling), so
465 -- currently eager blackholing doesn't work with profiling.
467 -- Previously, eager blackholing was enabled when ticky-ticky
468 -- was on. But it didn't work, and it wasn't strictly necessary
469 -- to bring back minimal ticky-ticky, so now EAGER_BLACKHOLING
470 -- is unconditionally disabled. -- krc 1/2007
472 eager_blackholing = False
476 setupUpdate :: ClosureInfo -> Code -> Code -- Only called for closures with no args
477 -- Nota Bene: this function does not change Node (even if it's a CAF),
478 -- so that the cost centre in the original closure can still be
479 -- extracted by a subsequent enterCostCentre
480 setupUpdate closure_info code
481 | closureReEntrant closure_info
484 | not (isStaticClosure closure_info)
485 = if closureUpdReqd closure_info
486 then do { tickyPushUpdateFrame; pushUpdateFrame (CmmReg nodeReg) code }
487 else do { tickyUpdateFrameOmitted; code }
489 | otherwise -- A static closure
490 = do { tickyUpdateBhCaf closure_info
492 ; if closureUpdReqd closure_info
493 then do -- Blackhole the (updatable) CAF:
494 { upd_closure <- link_caf closure_info True
495 ; pushUpdateFrame upd_closure code }
497 { -- krc: removed some ticky-related code here.
498 ; tickyUpdateFrameOmitted
503 -----------------------------------------------------------------------------
506 -- When a CAF is first entered, it creates a black hole in the heap,
507 -- and updates itself with an indirection to this new black hole.
509 -- We update the CAF with an indirection to a newly-allocated black
510 -- hole in the heap. We also set the blocking queue on the newly
511 -- allocated black hole to be empty.
513 -- Why do we make a black hole in the heap when we enter a CAF?
515 -- - for a generational garbage collector, which needs a fast
516 -- test for whether an updatee is in an old generation or not
518 -- - for the parallel system, which can implement updates more
519 -- easily if the updatee is always in the heap. (allegedly).
521 -- When debugging, we maintain a separate CAF list so we can tell when
522 -- a CAF has been garbage collected.
524 -- newCAF must be called before the itbl ptr is overwritten, since
525 -- newCAF records the old itbl ptr in order to do CAF reverting
526 -- (which Hugs needs to do in order that combined mode works right.)
529 -- ToDo [Feb 04] This entire link_caf nonsense could all be moved
530 -- into the "newCAF" RTS procedure, which we call anyway, including
531 -- the allocation of the black-hole indirection closure.
532 -- That way, code size would fall, the CAF-handling code would
533 -- be closer together, and the compiler wouldn't need to know
534 -- about off_indirectee etc.
536 link_caf :: ClosureInfo
537 -> Bool -- True <=> updatable, False <=> single-entry
538 -> FCode CmmExpr -- Returns amode for closure to be updated
539 -- To update a CAF we must allocate a black hole, link the CAF onto the
540 -- CAF list, then update the CAF to point to the fresh black hole.
541 -- This function returns the address of the black hole, so it can be
542 -- updated with the new value when available. The reason for all of this
543 -- is that we only want to update dynamic heap objects, not static ones,
544 -- so that generational GC is easier.
545 link_caf cl_info is_upd = do
546 { -- Alloc black hole specifying CC_HDR(Node) as the cost centre
547 ; let use_cc = costCentreFrom (CmmReg nodeReg)
549 ; hp_offset <- allocDynClosure bh_cl_info use_cc blame_cc []
550 ; hp_rel <- getHpRelOffset hp_offset
552 -- Call the RTS function newCAF to add the CAF to the CafList
553 -- so that the garbage collector can find them
554 -- This must be done *before* the info table pointer is overwritten,
555 -- because the old info table ptr is needed for reversion
556 ; emitRtsCallWithVols SLIT("newCAF") [(CmmReg nodeReg,PtrHint)] [node] False
557 -- node is live, so save it.
559 -- Overwrite the closure with a (static) indirection
560 -- to the newly-allocated black hole
561 ; stmtsC [ CmmStore (cmmRegOffW nodeReg off_indirectee) hp_rel
562 , CmmStore (CmmReg nodeReg) ind_static_info ]
566 bh_cl_info :: ClosureInfo
567 bh_cl_info | is_upd = cafBlackHoleClosureInfo cl_info
568 | otherwise = seCafBlackHoleClosureInfo cl_info
570 ind_static_info :: CmmExpr
571 ind_static_info = mkLblExpr mkIndStaticInfoLabel
573 off_indirectee :: WordOff
574 off_indirectee = fixedHdrSize + oFFSET_StgInd_indirectee*wORD_SIZE
578 %************************************************************************
580 \subsection[CgClosure-Description]{Profiling Closure Description.}
582 %************************************************************************
584 For "global" data constructors the description is simply occurrence
585 name of the data constructor itself. Otherwise it is determined by
586 @closureDescription@ from the let binding information.
589 closureDescription :: Module -- Module
590 -> Name -- Id of closure binding
592 -- Not called for StgRhsCon which have global info tables built in
593 -- CgConTbls.lhs with a description generated from the data constructor
594 closureDescription mod_name name
595 = showSDocDump (char '<' <>
596 (if isExternalName name
597 then ppr name -- ppr will include the module name prefix
598 else pprModule mod_name <> char '.' <> ppr name) <>
600 -- showSDocDump, because we want to see the unique on the Name.