2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 % $Id: CgHeapery.lhs,v 1.19 1999/10/13 16:39:15 simonmar Exp $
6 \section[CgHeapery]{Heap management functions}
10 fastEntryChecks, altHeapCheck, thunkChecks,
11 allocDynClosure, inPlaceAllocDynClosure
13 -- new functions, basically inserting macro calls into Code -- HWL
14 ,fetchAndReschedule, yield
17 #include "HsVersions.h"
23 import CgStackery ( getFinalStackHW, mkTaggedStkAmodes, mkTagAssts )
24 import SMRep ( fixedHdrSize )
25 import AbsCUtils ( mkAbstractCs, getAmodeRep )
26 import CgUsages ( getVirtAndRealHp, getRealSp, setVirtHp, setRealHp,
29 import ClosureInfo ( closureSize, closureGoodStuffSize,
30 slopSize, allocProfilingMsg, ClosureInfo,
33 import PrimRep ( PrimRep(..), isFollowableRep )
34 import Unique ( Unique )
35 import CmdLineOpts ( opt_SccProfilingOn )
40 import PprAbsC ( pprMagicId ) -- tmp
44 %************************************************************************
46 \subsection[CgHeapery-heap-overflow]{Heap overflow checking}
48 %************************************************************************
50 The new code for heapChecks. For GrAnSim the code for doing a heap check
51 and doing a context switch has been separated. Especially, the HEAP_CHK
52 macro only performs a heap check. THREAD_CONTEXT_SWITCH should be used for
53 doing a context switch. GRAN_FETCH_AND_RESCHEDULE must be put at the
54 beginning of every slow entry code in order to simulate the fetching of
55 closures. If fetching is necessary (i.e. current closure is not local) then
56 an automatic context switch is done.
58 -----------------------------------------------------------------------------
59 A heap/stack check at a fast entry point.
64 :: [MagicId] -- Live registers
65 -> [(VirtualSpOffset,Int)] -- stack slots to tag
66 -> CLabel -- return point
67 -> Bool -- node points to closure
71 fastEntryChecks regs tags ret node_points code
72 = mkTagAssts tags `thenFC` \tag_assts ->
73 getFinalStackHW (\ spHw ->
74 getRealSp `thenFC` \ sp ->
75 let stk_words = spHw - sp in
76 initHeapUsage (\ hp_words ->
78 getTickyCtrLabel `thenFC` \ ticky_ctr ->
80 ( if all_pointers then -- heap checks are quite easy
81 absC (checking_code stk_words hp_words tag_assts
82 free_reg (length regs) ticky_ctr)
84 else -- they are complicated
86 -- save all registers on the stack and adjust the stack pointer.
87 -- ToDo: find the initial all-pointer segment and don't save them.
89 mkTaggedStkAmodes sp addrmode_regs
90 `thenFC` \(new_sp, stk_assts, more_tag_assts) ->
92 -- only let the extra stack assignments affect the stack
93 -- high water mark if we were doing a stack check anyway;
94 -- otherwise we end up generating unnecessary stack checks.
95 -- Careful about knot-tying loops!
96 let real_stk_words = if new_sp - sp > stk_words && stk_words /= 0
101 let adjust_sp = CAssign (CReg Sp) (CAddr (spRel sp new_sp)) in
103 absC (checking_code real_stk_words hp_words
104 (mkAbstractCs [tag_assts, stk_assts, more_tag_assts,
106 (CReg node) 0 ticky_ctr)
110 setRealHp hp_words `thenC`
115 checking_code stk hp assts ret regs ctr
118 if hp == 0 then AbsCNop
119 else profCtrAbsC SLIT("TICK_ALLOC_HEAP")
120 [ mkIntCLit hp, CLbl ctr DataPtrRep ]
124 | node_points = do_checks_np stk hp assts (regs+1)
125 | otherwise = do_checks stk hp assts ret regs
127 -- When node points to the closure for the function:
130 :: Int -- stack headroom
131 -> Int -- heap headroom
132 -> AbstractC -- assignments to perform on failure
133 -> Int -- number of pointer registers live
135 do_checks_np 0 0 _ _ = AbsCNop
136 do_checks_np 0 hp_words tag_assts ptrs =
142 do_checks_np stk_words 0 tag_assts ptrs =
148 do_checks_np stk_words hp_words tag_assts ptrs =
149 CCheck HP_STK_CHK_NP [
156 -- When node doesn't point to the closure (we need an explicit retn addr)
159 :: Int -- stack headroom
160 -> Int -- heap headroom
161 -> AbstractC -- assignments to perform on failure
162 -> CAddrMode -- a register to hold the retn addr.
163 -> Int -- number of pointer registers live
166 do_checks 0 0 _ _ _ = AbsCNop
167 do_checks 0 hp_words tag_assts ret_reg ptrs =
175 do_checks stk_words 0 tag_assts ret_reg ptrs =
183 do_checks stk_words hp_words tag_assts ret_reg ptrs =
193 free_reg = case length regs + 1 of
194 IBOX(x) -> CReg (VanillaReg PtrRep x)
196 all_pointers = all pointer regs
197 pointer (VanillaReg rep _) = isFollowableRep rep
200 addrmode_regs = map CReg regs
202 -- Checking code for thunks is just a special case of fast entry points:
204 thunkChecks :: CLabel -> Bool -> Code -> Code
205 thunkChecks ret node_points code = fastEntryChecks [] [] ret node_points code
208 Heap checks in a case alternative are nice and easy, provided this is
209 a bog-standard algebraic case. We have in our hand:
211 * one return address, on the stack,
212 * one return value, in Node.
214 the canned code for this heap check failure just pushes Node on the
215 stack, saying 'EnterGHC' to return. The scheduler will return by
216 entering the top value on the stack, which in turn will return through
217 the return address, getting us back to where we were. This is
218 therefore only valid if the return value is *lifted* (just being
219 boxed isn't good enough). Only a PtrRep will do.
221 For primitive returns, we have an unlifted value in some register
222 (either R1 or FloatReg1 or DblReg1). This means using specialised
223 heap-check code for these cases.
225 For unboxed tuple returns, there are an arbitrary number of possibly
226 unboxed return values, some of which will be in registers, and the
227 others will be on the stack, with gaps left for tagging the unboxed
228 objects. If a heap check is required, we need to fill in these tags.
230 The code below will cover all cases for the x86 architecture (where R1
231 is the only VanillaReg ever used). For other architectures, we'll
232 have to do something about saving and restoring the other registers.
236 :: Bool -- is an algebraic alternative
237 -> [MagicId] -- live registers
238 -> [(VirtualSpOffset,Int)] -- stack slots to tag
240 -> Maybe Unique -- uniq of ret address (possibly)
244 -- unboxed tuple alternatives and let-no-escapes (the two most annoying
245 -- constructs to generate code for!):
247 altHeapCheck is_fun regs tags fail_code (Just ret_addr) code
248 = mkTagAssts tags `thenFC` \tag_assts1 ->
249 let tag_assts = mkAbstractCs [fail_code, tag_assts1]
251 initHeapUsage (\ hHw -> do_heap_chk hHw tag_assts `thenC` code)
253 do_heap_chk words_required tag_assts
254 = getTickyCtrLabel `thenFC` \ ctr ->
255 absC ( if words_required == 0
258 [ checking_code tag_assts,
259 profCtrAbsC SLIT("TICK_ALLOC_HEAP")
260 [ mkIntCLit words_required, CLbl ctr DataPtrRep ]
263 setRealHp words_required
266 non_void_regs = filter (/= VoidReg) regs
268 checking_code tag_assts =
269 case non_void_regs of
271 {- no: there might be stuff on top of the retn. addr. on the stack.
274 [mkIntCLit words_required]
277 -- this will cover all cases for x86
278 [VanillaReg rep ILIT(1)]
280 | isFollowableRep rep ->
282 [mkIntCLit words_required, mkIntCLit 1, mkIntCLit 0,
283 CReg (VanillaReg RetRep ILIT(2)),
284 CLbl (mkReturnInfoLabel ret_addr) RetRep]
289 [mkIntCLit words_required, mkIntCLit 0, mkIntCLit 1,
290 CReg (VanillaReg RetRep ILIT(2)),
291 CLbl (mkReturnInfoLabel ret_addr) RetRep]
295 let liveness = mkRegLiveness several_regs
298 [mkIntCLit words_required,
299 mkIntCLit (IBOX(word2Int# liveness)),
300 -- HP_CHK_GEN needs a direct return address,
301 -- not an info table (might be different if
302 -- we're not assembly-mangling/tail-jumping etc.)
303 CLbl (mkReturnPtLabel ret_addr) RetRep]
306 -- normal algebraic and primitive case alternatives:
308 altHeapCheck is_fun regs [] AbsCNop Nothing code
309 = initHeapUsage (\ hHw -> do_heap_chk hHw `thenC` code)
312 do_heap_chk :: HeapOffset -> Code
313 do_heap_chk words_required
314 = getTickyCtrLabel `thenFC` \ ctr ->
315 absC ( if words_required == 0
319 profCtrAbsC SLIT("TICK_ALLOC_HEAP")
320 [ mkIntCLit words_required, CLbl ctr DataPtrRep ]
323 setRealHp words_required
326 non_void_regs = filter (/= VoidReg) regs
329 case non_void_regs of
331 -- No regs live: probably a Void return
333 CCheck HP_CHK_NOREGS [mkIntCLit words_required] AbsCNop
335 -- The SEQ case (polymophic/function typed case branch)
336 -- We need this case because the closure in Node won't return
337 -- directly when we enter it (it could be a function), so the
338 -- heap check code needs to push a seq frame on top of the stack.
339 [VanillaReg rep ILIT(1)]
343 [mkIntCLit words_required, mkIntCLit 1{-regs live-}]
346 -- R1 is lifted (the common case)
347 [VanillaReg rep ILIT(1)]
350 [mkIntCLit words_required, mkIntCLit 1{-regs live-}]
353 -- R1 is boxed, but unlifted
354 | isFollowableRep rep ->
355 CCheck HP_CHK_UNPT_R1 [mkIntCLit words_required] AbsCNop
359 CCheck HP_CHK_UNBX_R1 [mkIntCLit words_required] AbsCNop
362 [FloatReg ILIT(1)] ->
363 CCheck HP_CHK_F1 [mkIntCLit words_required] AbsCNop
366 [DoubleReg ILIT(1)] ->
367 CCheck HP_CHK_D1 [mkIntCLit words_required] AbsCNop
370 [LongReg _ ILIT(1)] ->
371 CCheck HP_CHK_L1 [mkIntCLit words_required] AbsCNop
374 _ -> panic ("CgHeapery.altHeapCheck: unimplemented heap-check, live regs = " ++ showSDoc (sep (map pprMagicId non_void_regs)))
377 -- build up a bitmap of the live pointer registers
379 mkRegLiveness :: [MagicId] -> Word#
380 mkRegLiveness [] = int2Word# 0#
381 mkRegLiveness (VanillaReg rep i : regs) | isFollowableRep rep
382 = ((int2Word# 1#) `shiftL#` (i -# 1#)) `or#` mkRegLiveness regs
383 mkRegLiveness (_ : regs) = mkRegLiveness regs
385 -- Emit macro for simulating a fetch and then reschedule
387 fetchAndReschedule :: [MagicId] -- Live registers
388 -> Bool -- Node reqd?
391 fetchAndReschedule regs node_reqd =
392 if (node `elem` regs || node_reqd)
393 then fetch_code `thenC` reschedule_code
396 all_regs = if node_reqd then node:regs else regs
397 liveness_mask = 0 {-XXX: mkLiveRegsMask all_regs-}
399 reschedule_code = absC (CMacroStmt GRAN_RESCHEDULE [
400 mkIntCLit liveness_mask,
401 mkIntCLit (if node_reqd then 1 else 0)])
403 --HWL: generate GRAN_FETCH macro for GrAnSim
404 -- currently GRAN_FETCH and GRAN_FETCH_AND_RESCHEDULE are miai
405 fetch_code = absC (CMacroStmt GRAN_FETCH [])
408 The @GRAN_YIELD@ macro is taken from JSM's code for Concurrent Haskell. It
409 allows to context-switch at places where @node@ is not alive (it uses the
410 @Continue@ rather than the @EnterNodeCode@ function in the RTS). We emit
411 this kind of macro at the beginning of the following kinds of basic bocks:
413 \item Slow entry code where node is not alive (see @CgClosure.lhs@). Normally
414 we use @fetchAndReschedule@ at a slow entry code.
415 \item Fast entry code (see @CgClosure.lhs@).
416 \item Alternatives in case expressions (@CLabelledCode@ structures), provided
417 that they are not inlined (see @CgCases.lhs@). These alternatives will
418 be turned into separate functions.
422 yield :: [MagicId] -- Live registers
423 -> Bool -- Node reqd?
426 yield regs node_reqd =
427 -- NB: node is not alive; that's why we use DO_YIELD rather than
431 all_regs = if node_reqd then node:regs else regs
432 liveness_mask = 0 {-XXX: mkLiveRegsMask all_regs-}
434 yield_code = absC (CMacroStmt GRAN_YIELD [mkIntCLit liveness_mask])
437 %************************************************************************
439 \subsection[initClosure]{Initialise a dynamic closure}
441 %************************************************************************
443 @allocDynClosure@ puts the thing in the heap, and modifies the virtual Hp
449 -> CAddrMode -- Cost Centre to stick in the object
450 -> CAddrMode -- Cost Centre to blame for this alloc
451 -- (usually the same; sometimes "OVERHEAD")
453 -> [(CAddrMode, VirtualHeapOffset)] -- Offsets from start of the object
454 -- ie Info ptr has offset zero.
455 -> FCode VirtualHeapOffset -- Returns virt offset of object
457 allocDynClosure closure_info use_cc blame_cc amodes_with_offsets
458 = getVirtAndRealHp `thenFC` \ (virtHp, realHp) ->
460 -- FIND THE OFFSET OF THE INFO-PTR WORD
461 -- virtHp points to last allocated word, ie 1 *before* the
462 -- info-ptr word of new object.
463 let info_offset = virtHp + 1
465 -- do_move IS THE ASSIGNMENT FUNCTION
466 do_move (amode, offset_from_start)
467 = CAssign (CVal (hpRel realHp
468 (info_offset + offset_from_start))
472 -- SAY WHAT WE ARE ABOUT TO DO
473 profCtrC (allocProfilingMsg closure_info)
474 [mkIntCLit (closureGoodStuffSize closure_info),
475 mkIntCLit slop_size] `thenC`
478 absC ( mkAbstractCs (
479 [ cInitHdr closure_info (hpRel realHp info_offset) use_cc ]
480 ++ (map do_move amodes_with_offsets))) `thenC`
482 -- GENERATE CC PROFILING MESSAGES
483 costCentresC SLIT("CCS_ALLOC") [blame_cc, mkIntCLit closure_size]
486 -- BUMP THE VIRTUAL HEAP POINTER
487 setVirtHp (virtHp + closure_size) `thenC`
489 -- RETURN PTR TO START OF OBJECT
492 closure_size = closureSize closure_info
493 slop_size = slopSize closure_info
496 Occasionally we can update a closure in place instead of allocating
497 new space for it. This is the function that does the business, assuming:
499 - node points to the closure to be overwritten
501 - the new closure doesn't contain any pointers if we're
502 using a generational collector.
505 inPlaceAllocDynClosure
507 -> CAddrMode -- Pointer to beginning of closure
508 -> CAddrMode -- Cost Centre to stick in the object
510 -> [(CAddrMode, VirtualHeapOffset)] -- Offsets from start of the object
511 -- ie Info ptr has offset zero.
514 inPlaceAllocDynClosure closure_info head use_cc amodes_with_offsets
515 = let -- do_move IS THE ASSIGNMENT FUNCTION
516 do_move (amode, offset_from_start)
517 = CAssign (CVal (CIndex head (mkIntCLit offset_from_start) WordRep)
522 absC ( mkAbstractCs (
523 [ CInitHdr closure_info head use_cc ]
524 ++ (map do_move amodes_with_offsets)))
526 -- Avoid hanging on to anything in the CC field when we're not profiling.
528 cInitHdr closure_info amode cc
529 | opt_SccProfilingOn = CInitHdr closure_info (CAddr amode) cc
530 | otherwise = CInitHdr closure_info (CAddr amode) (panic "absent cc")