2 % (c) The University of Glasgow 2006
3 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
8 -- The above warning supression flag is a temporary kludge.
9 -- While working on this module you are encouraged to remove it and fix
10 -- any warnings in the module. See
11 -- http://hackage.haskell.org/trac/ghc/wiki/Commentary/CodingStyle#Warnings
14 module CgCase ( cgCase, saveVolatileVarsAndRegs,
15 restoreCurrentCostCentre
18 #include "HsVersions.h"
20 import {-# SOURCE #-} CgExpr ( cgExpr )
54 = GCMayHappen -- The scrutinee may involve GC, so everything must be
55 -- tidy before the code for the scrutinee.
57 | NoGC -- The scrutinee is a primitive value, or a call to a
58 -- primitive op which does no GC. Hence the case can
59 -- be done inline, without tidying up first.
62 It is quite interesting to decide whether to put a heap-check
63 at the start of each alternative. Of course we certainly have
64 to do so if the case forces an evaluation, or if there is a primitive
65 op which can trigger GC.
67 A more interesting situation is this:
74 default -> !C!; ...C...
77 where \tr{!x!} indicates a possible heap-check point. The heap checks
78 in the alternatives {\em can} be omitted, in which case the topmost
79 heapcheck will take their worst case into account.
81 In favour of omitting \tr{!B!}, \tr{!C!}:
83 - {\em May} save a heap overflow test,
84 if ...A... allocates anything. The other advantage
85 of this is that we can use relative addressing
86 from a single Hp to get at all the closures so allocated.
88 - No need to save volatile vars etc across the case
92 - May do more allocation than reqd. This sometimes bites us
93 badly. For example, nfib (ha!) allocates about 30\% more space if the
94 worst-casing is done, because many many calls to nfib are leaf calls
95 which don't need to allocate anything.
97 This never hurts us if there is only one alternative.
109 Special case #1: case of literal.
112 cgCase (StgLit lit) live_in_whole_case live_in_alts bndr
113 alt_type@(PrimAlt tycon) alts
114 = do { tmp_reg <- bindNewToTemp bndr
115 ; cm_lit <- cgLit lit
116 ; stmtC (CmmAssign (CmmLocal tmp_reg) (CmmLit cm_lit))
117 ; cgPrimAlts NoGC alt_type (CmmLocal tmp_reg) alts }
120 Special case #2: scrutinising a primitive-typed variable. No
121 evaluation required. We don't save volatile variables, nor do we do a
122 heap-check in the alternatives. Instead, the heap usage of the
123 alternatives is worst-cased and passed upstream. This can result in
124 allocating more heap than strictly necessary, but it will sometimes
125 eliminate a heap check altogether.
128 cgCase (StgApp v []) live_in_whole_case live_in_alts bndr
129 alt_type@(PrimAlt tycon) alts
130 = do { -- Careful! we can't just bind the default binder to the same thing
131 -- as the scrutinee, since it might be a stack location, and having
132 -- two bindings pointing at the same stack locn doesn't work (it
133 -- confuses nukeDeadBindings). Hence, use a new temp.
134 v_info <- getCgIdInfo v
135 ; amode <- idInfoToAmode v_info
136 ; tmp_reg <- bindNewToTemp bndr
137 ; stmtC (CmmAssign (CmmLocal tmp_reg) amode)
138 ; cgPrimAlts NoGC alt_type (CmmLocal tmp_reg) alts }
141 Special case #3: inline PrimOps and foreign calls.
144 cgCase (StgOpApp op@(StgPrimOp primop) args _)
145 live_in_whole_case live_in_alts bndr alt_type alts
146 | not (primOpOutOfLine primop)
147 = cgInlinePrimOp primop args bndr alt_type live_in_alts alts
150 TODO: Case-of-case of primop can probably be done inline too (but
151 maybe better to translate it out beforehand). See
152 ghc/lib/misc/PackedString.lhs for examples where this crops up (with
155 Special case #4: inline foreign calls: an unsafe foreign call can be done
156 right here, just like an inline primop.
159 cgCase (StgOpApp op@(StgFCallOp fcall _) args _)
160 live_in_whole_case live_in_alts bndr alt_type alts
161 | unsafe_foreign_call
162 = ASSERT( isSingleton alts )
163 do -- *must* be an unboxed tuple alt.
164 -- exactly like the cgInlinePrimOp case for unboxed tuple alts..
165 { res_tmps <- mapFCs bindNewToTemp non_void_res_ids
166 ; let res_hints = map (typeForeignHint.idType) non_void_res_ids
167 ; cgForeignCall (zipWith CmmHinted res_tmps res_hints) fcall args live_in_alts
170 (_, res_ids, _, rhs) = head alts
171 non_void_res_ids = filter (nonVoidArg . idCgRep) res_ids
175 CCall (CCallSpec _ _ s) -> not (playSafe s)
179 Special case: scrutinising a non-primitive variable.
180 This can be done a little better than the general case, because
181 we can reuse/trim the stack slot holding the variable (if it is in one).
184 cgCase (StgApp fun args)
185 live_in_whole_case live_in_alts bndr alt_type alts
186 = do { fun_info <- getCgIdInfo fun
187 ; arg_amodes <- getArgAmodes args
189 -- Nuking dead bindings *before* calculating the saves is the
190 -- value-add here. We might end up freeing up some slots currently
191 -- occupied by variables only required for the call.
192 -- NOTE: we need to look up the variables used in the call before
193 -- doing this, because some of them may not be in the environment
195 ; nukeDeadBindings live_in_alts
196 ; (save_assts, alts_eob_info, maybe_cc_slot)
197 <- saveVolatileVarsAndRegs live_in_alts
200 <- forkEval alts_eob_info
201 (allocStackTop retAddrSizeW >> nopC)
202 (do { deAllocStackTop retAddrSizeW
203 ; cgEvalAlts maybe_cc_slot bndr alt_type alts })
205 ; setEndOfBlockInfo scrut_eob_info
206 (performTailCall fun_info arg_amodes save_assts) }
209 Note about return addresses: we *always* push a return address, even
210 if because of an optimisation we end up jumping direct to the return
211 code (not through the address itself). The alternatives always assume
212 that the return address is on the stack. The return address is
213 required in case the alternative performs a heap check, since it
214 encodes the liveness of the slots in the activation record.
216 On entry to the case alternative, we can re-use the slot containing
217 the return address immediately after the heap check. That's what the
218 deAllocStackTop call is doing above.
220 Finally, here is the general case.
223 cgCase expr live_in_whole_case live_in_alts bndr alt_type alts
224 = do { -- Figure out what volatile variables to save
225 nukeDeadBindings live_in_whole_case
227 ; (save_assts, alts_eob_info, maybe_cc_slot)
228 <- saveVolatileVarsAndRegs live_in_alts
230 -- Save those variables right now!
231 ; emitStmts save_assts
233 -- generate code for the alts
235 <- forkEval alts_eob_info
236 (do { nukeDeadBindings live_in_alts
237 ; allocStackTop retAddrSizeW -- space for retn address
239 (do { deAllocStackTop retAddrSizeW
240 ; cgEvalAlts maybe_cc_slot bndr alt_type alts })
242 ; setEndOfBlockInfo scrut_eob_info (cgExpr expr)
246 There's a lot of machinery going on behind the scenes to manage the
247 stack pointer here. forkEval takes the virtual Sp and free list from
248 the first argument, and turns that into the *real* Sp for the second
249 argument. It also uses this virtual Sp as the args-Sp in the EOB info
250 returned, so that the scrutinee will trim the real Sp back to the
251 right place before doing whatever it does.
252 --SDM (who just spent an hour figuring this out, and didn't want to
255 Why don't we push the return address just before evaluating the
256 scrutinee? Because the slot reserved for the return address might
257 contain something useful, so we wait until performing a tail call or
258 return before pushing the return address (see
259 CgTailCall.pushReturnAddress).
261 This also means that the environment doesn't need to know about the
262 free stack slot for the return address (for generating bitmaps),
263 because we don't reserve it until just before the eval.
265 TODO!! Problem: however, we have to save the current cost centre
266 stack somewhere, because at the eval point the current CCS might be
267 different. So we pick a free stack slot and save CCCS in it. One
268 consequence of this is that activation records on the stack don't
269 follow the layout of closures when we're profiling. The CCS could be
270 anywhere within the record).
272 %************************************************************************
276 %************************************************************************
279 cgInlinePrimOp primop args bndr (PrimAlt tycon) live_in_alts alts
280 | isVoidArg (idCgRep bndr)
281 = ASSERT( con == DEFAULT && isSingleton alts && null bs )
282 do { -- VOID RESULT; just sequencing,
283 -- so get in there and do it
284 cgPrimOp [] primop args live_in_alts
287 (con,bs,_,rhs) = head alts
289 cgInlinePrimOp primop args bndr (PrimAlt tycon) live_in_alts alts
290 = do { -- PRIMITIVE ALTS, with non-void result
291 tmp_reg <- bindNewToTemp bndr
292 ; cgPrimOp [tmp_reg] primop args live_in_alts
293 ; cgPrimAlts NoGC (PrimAlt tycon) (CmmLocal tmp_reg) alts }
295 cgInlinePrimOp primop args bndr (UbxTupAlt tycon) live_in_alts alts
296 = ASSERT( isSingleton alts )
297 do { -- UNBOXED TUPLE ALTS
298 -- No heap check, no yield, just get in there and do it.
299 -- NB: the case binder isn't bound to anything;
300 -- it has a unboxed tuple type
302 res_tmps <- mapFCs bindNewToTemp non_void_res_ids
303 ; cgPrimOp res_tmps primop args live_in_alts
306 (_, res_ids, _, rhs) = head alts
307 non_void_res_ids = filter (nonVoidArg . idCgRep) res_ids
309 cgInlinePrimOp primop args bndr (AlgAlt tycon) live_in_alts alts
310 = do { -- ENUMERATION TYPE RETURN
311 -- Typical: case a ># b of { True -> ..; False -> .. }
312 -- The primop itself returns an index into the table of
313 -- closures for the enumeration type.
314 tag_amode <- ASSERT( isEnumerationTyCon tycon )
315 do_enum_primop primop
317 -- Bind the default binder if necessary
318 -- (avoiding it avoids the assignment)
319 -- The deadness info is set by StgVarInfo
320 ; whenC (not (isDeadBinder bndr))
321 (do { tmp_reg <- bindNewToTemp bndr
324 (tagToClosure tycon tag_amode)) })
327 ; (branches, mb_deflt) <- cgAlgAlts NoGC Nothing{-cc_slot-}
331 ; emitSwitch tag_amode branches mb_deflt 0 (tyConFamilySize tycon - 1)
335 do_enum_primop :: PrimOp -> FCode CmmExpr -- Returns amode for result
336 do_enum_primop TagToEnumOp -- No code!
338 (_,e) <- getArgAmode arg
340 do_enum_primop primop
341 = do tmp <- newTemp bWord
342 cgPrimOp [tmp] primop args live_in_alts
343 returnFC (CmmReg (CmmLocal tmp))
345 cgInlinePrimOp primop arg_amodes bndr PolyAlt live_in_alts alts
346 = pprPanic "cgCase: case of primop has polymorphic type" (ppr bndr)
349 %************************************************************************
351 \subsection[CgCase-alts]{Alternatives}
353 %************************************************************************
355 @cgEvalAlts@ returns an addressing mode for a continuation for the
356 alternatives of a @case@, used in a context when there
357 is some evaluation to be done.
360 cgEvalAlts :: Maybe VirtualSpOffset -- Offset of cost-centre to be restored, if any
364 -> FCode Sequel -- Any addr modes inside are guaranteed
365 -- to be a label so that we can duplicate it
366 -- without risk of duplicating code
368 cgEvalAlts cc_slot bndr alt_type@(PrimAlt tycon) alts
369 = do { let rep = tyConCgRep tycon
370 reg = dataReturnConvPrim rep -- Bottom for voidRep
372 ; abs_c <- forkProc $ do
373 { -- Bind the case binder, except if it's void
374 -- (reg is bottom in that case)
375 whenC (nonVoidArg rep) $
376 bindNewToReg bndr reg (mkLFArgument bndr)
377 ; restoreCurrentCostCentre cc_slot True
378 ; cgPrimAlts GCMayHappen alt_type reg alts }
380 ; lbl <- emitReturnTarget (idName bndr) abs_c
381 ; returnFC (CaseAlts lbl Nothing bndr) }
383 cgEvalAlts cc_slot bndr (UbxTupAlt _) [(con,args,_,rhs)]
384 = -- Unboxed tuple case
385 -- By now, the simplifier should have have turned it
386 -- into case e of (# a,b #) -> e
387 -- There shouldn't be a
388 -- case e of DEFAULT -> e
389 ASSERT2( case con of { DataAlt _ -> True; other -> False },
390 text "cgEvalAlts: dodgy case of unboxed tuple type" )
391 do { -- forkAbsC for the RHS, so that the envt is
392 -- not changed for the emitReturn call
393 abs_c <- forkProc $ do
394 { (live_regs, ptrs, nptrs, _) <- bindUnboxedTupleComponents args
395 -- Restore the CC *after* binding the tuple components,
396 -- so that we get the stack offset of the saved CC right.
397 ; restoreCurrentCostCentre cc_slot True
398 -- Generate a heap check if necessary
399 -- and finally the code for the alternative
400 ; unbxTupleHeapCheck live_regs ptrs nptrs noStmts
402 ; lbl <- emitReturnTarget (idName bndr) abs_c
403 ; returnFC (CaseAlts lbl Nothing bndr) }
405 cgEvalAlts cc_slot bndr alt_type alts
406 = -- Algebraic and polymorphic case
407 do { -- Bind the default binder
408 bindNewToReg bndr nodeReg (mkLFArgument bndr)
410 -- Generate sequel info for use downstream
411 -- At the moment, we only do it if the type is vector-returnable.
412 -- Reason: if not, then it costs extra to label the
413 -- alternatives, because we'd get return code like:
415 -- switch TagReg { 0 : JMP(alt_1); 1 : JMP(alt_2) ..etc }
417 -- which is worse than having the alt code in the switch statement
419 ; (alts, mb_deflt) <- cgAlgAlts GCMayHappen cc_slot alt_type alts
421 ; (lbl, branches) <- emitAlgReturnTarget (idName bndr)
424 ; returnFC (CaseAlts lbl branches bndr) }
426 fam_sz = case alt_type of
427 AlgAlt tc -> tyConFamilySize tc
432 HWL comment on {\em GrAnSim\/} (adding GRAN_YIELDs for context switch): If
433 we do an inlining of the case no separate functions for returning are
434 created, so we don't have to generate a GRAN_YIELD in that case. This info
435 must be propagated to cgAlgAltRhs (where the GRAN_YIELD macro might be
436 emitted). Hence, the new Bool arg to cgAlgAltRhs.
438 %************************************************************************
440 \subsection[CgCase-alg-alts]{Algebraic alternatives}
442 %************************************************************************
444 In @cgAlgAlts@, none of the binders in the alternatives are
445 assumed to be yet bound.
447 HWL comment on {\em GrAnSim\/} (adding GRAN_YIELDs for context switch): The
448 last arg of cgAlgAlts indicates if we want a context switch at the
449 beginning of each alternative. Normally we want that. The only exception
450 are inlined alternatives.
454 -> Maybe VirtualSpOffset
455 -> AltType -- ** AlgAlt or PolyAlt only **
456 -> [StgAlt] -- The alternatives
457 -> FCode ( [(ConTagZ, CgStmts)], -- The branches
458 Maybe CgStmts ) -- The default case
460 cgAlgAlts gc_flag cc_slot alt_type alts
461 = do alts <- forkAlts [ cgAlgAlt gc_flag cc_slot alt_type alt | alt <- alts]
463 mb_deflt = case alts of -- DEFAULT is always first, if present
464 ((DEFAULT,blks) : _) -> Just blks
467 branches = [(dataConTagZ con, blks)
468 | (DataAlt con, blks) <- alts]
470 return (branches, mb_deflt)
474 -> Maybe VirtualSpOffset -- Turgid state
475 -> AltType -- ** AlgAlt or PolyAlt only **
477 -> FCode (AltCon, CgStmts)
479 cgAlgAlt gc_flag cc_slot alt_type (con, args, use_mask, rhs)
480 = do { abs_c <- getCgStmts $ do
481 { bind_con_args con args
482 ; restoreCurrentCostCentre cc_slot True
483 ; maybeAltHeapCheck gc_flag alt_type (cgExpr rhs) }
484 ; return (con, abs_c) }
486 bind_con_args DEFAULT args = nopC
487 bind_con_args (DataAlt dc) args = bindConArgs dc args
491 %************************************************************************
493 \subsection[CgCase-prim-alts]{Primitive alternatives}
495 %************************************************************************
497 @cgPrimAlts@ generates suitable a @CSwitch@
498 for dealing with the alternatives of a primitive @case@, given an
499 addressing mode for the thing to scrutinise. It also keeps track of
500 the maximum stack depth encountered down any branch.
502 As usual, no binders in the alternatives are yet bound.
506 -> AltType -- Always PrimAlt, but passed to maybeAltHeapCheck
507 -> CmmReg -- Scrutinee
508 -> [StgAlt] -- Alternatives
510 -- NB: cgPrimAlts emits code that does the case analysis.
511 -- It's often used in inline situations, rather than to genearte
512 -- a labelled return point. That's why its interface is a little
513 -- different to cgAlgAlts
515 -- INVARIANT: the default binder is already bound
516 cgPrimAlts gc_flag alt_type scrutinee alts
517 = do { tagged_absCs <- forkAlts (map (cgPrimAlt gc_flag alt_type) alts)
518 ; let ((DEFAULT, deflt_absC) : others) = tagged_absCs -- There is always a default
519 alt_absCs = [(lit,rhs) | (LitAlt lit, rhs) <- others]
520 ; emitLitSwitch (CmmReg scrutinee) alt_absCs deflt_absC }
524 -> StgAlt -- The alternative
525 -> FCode (AltCon, CgStmts) -- Its compiled form
527 cgPrimAlt gc_flag alt_type (con, [], [], rhs)
528 = ASSERT( case con of { DEFAULT -> True; LitAlt _ -> True; other -> False } )
529 do { abs_c <- getCgStmts (maybeAltHeapCheck gc_flag alt_type (cgExpr rhs))
530 ; returnFC (con, abs_c) }
534 %************************************************************************
536 \subsection[CgCase-tidy]{Code for tidying up prior to an eval}
538 %************************************************************************
543 -> AltType -- PolyAlt, PrimAlt, AlgAlt, but *not* UbxTupAlt
544 -> Code -- Continuation
546 maybeAltHeapCheck NoGC _ code = code
547 maybeAltHeapCheck GCMayHappen alt_type code = altHeapCheck alt_type code
549 saveVolatileVarsAndRegs
550 :: StgLiveVars -- Vars which should be made safe
551 -> FCode (CmmStmts, -- Assignments to do the saves
552 EndOfBlockInfo, -- sequel for the alts
553 Maybe VirtualSpOffset) -- Slot for current cost centre
555 saveVolatileVarsAndRegs vars
556 = do { var_saves <- saveVolatileVars vars
557 ; (maybe_cc_slot, cc_save) <- saveCurrentCostCentre
558 ; eob_info <- getEndOfBlockInfo
559 ; returnFC (var_saves `plusStmts` cc_save,
564 saveVolatileVars :: StgLiveVars -- Vars which should be made safe
565 -> FCode CmmStmts -- Assignments to to the saves
567 saveVolatileVars vars
568 = do { stmts_s <- mapFCs save_it (varSetElems vars)
569 ; return (foldr plusStmts noStmts stmts_s) }
572 = do { v <- getCAddrModeIfVolatile var
574 Nothing -> return noStmts -- Non-volatile
575 Just vol_amode -> save_var var vol_amode -- Aha! It's volatile
578 save_var var vol_amode
579 = do { slot <- allocPrimStack (idCgRep var)
580 ; rebindToStack var slot
581 ; sp_rel <- getSpRelOffset slot
582 ; returnFC (oneStmt (CmmStore sp_rel vol_amode)) }
585 ---------------------------------------------------------------------------
587 When we save the current cost centre (which is done for lexical
588 scoping), we allocate a free stack location, and return (a)~the
589 virtual offset of the location, to pass on to the alternatives, and
590 (b)~the assignment to do the save (just as for @saveVolatileVars@).
593 saveCurrentCostCentre ::
594 FCode (Maybe VirtualSpOffset, -- Where we decide to store it
595 CmmStmts) -- Assignment to save it
597 saveCurrentCostCentre
598 | not opt_SccProfilingOn
599 = returnFC (Nothing, noStmts)
601 = do { slot <- allocPrimStack PtrArg
602 ; sp_rel <- getSpRelOffset slot
603 ; returnFC (Just slot,
604 oneStmt (CmmStore sp_rel curCCS)) }
606 -- Sometimes we don't free the slot containing the cost centre after restoring it
607 -- (see CgLetNoEscape.cgLetNoEscapeBody).
608 restoreCurrentCostCentre :: Maybe VirtualSpOffset -> Bool -> Code
609 restoreCurrentCostCentre Nothing _freeit = nopC
610 restoreCurrentCostCentre (Just slot) freeit
611 = do { sp_rel <- getSpRelOffset slot
612 ; whenC freeit (freeStackSlots [slot])
613 ; stmtC (CmmStore curCCSAddr (CmmLoad sp_rel bWord)) }