2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 % $Id: CgHeapery.lhs,v 1.32 2002/08/29 15:44:13 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 AbsCUtils ( mkAbstractCs, getAmodeRep )
25 import CgUsages ( getVirtAndRealHp, getRealSp, setVirtHp, setRealHp,
28 import ClosureInfo ( closureSize, closureGoodStuffSize,
29 slopSize, allocProfilingMsg, ClosureInfo
31 import PrimRep ( PrimRep(..), isFollowableRep )
32 import Unique ( Unique )
33 import CmdLineOpts ( opt_GranMacros )
37 import PprAbsC ( pprMagicId ) -- tmp
43 %************************************************************************
45 \subsection[CgHeapery-heap-overflow]{Heap overflow checking}
47 %************************************************************************
49 The new code for heapChecks. For GrAnSim the code for doing a heap check
50 and doing a context switch has been separated. Especially, the HEAP_CHK
51 macro only performs a heap check. THREAD_CONTEXT_SWITCH should be used for
52 doing a context switch. GRAN_FETCH_AND_RESCHEDULE must be put at the
53 beginning of every slow entry code in order to simulate the fetching of
54 closures. If fetching is necessary (i.e. current closure is not local) then
55 an automatic context switch is done.
57 -----------------------------------------------------------------------------
58 A heap/stack check at a fast entry point.
63 :: [MagicId] -- Live registers
64 -> [(VirtualSpOffset,Int)] -- stack slots to tag
65 -> CLabel -- return point
66 -> Bool -- node points to closure
70 fastEntryChecks regs tags ret node_points code
71 = mkTagAssts tags `thenFC` \tag_assts ->
72 getFinalStackHW (\ spHw ->
73 getRealSp `thenFC` \ sp ->
74 let stk_words = spHw - sp in
75 initHeapUsage (\ hHw ->
77 getTickyCtrLabel `thenFC` \ ticky_ctr ->
79 ( if all_pointers then -- heap checks are quite easy
80 -- HWL: gran-yield immediately before heap check proper
81 --(if node `elem` regs
82 -- then yield regs True
83 -- else absC AbsCNop ) `thenC`
84 absC (checking_code stk_words hHw tag_assts
85 free_reg (length regs) ticky_ctr)
87 else -- they are complicated
89 -- save all registers on the stack and adjust the stack pointer.
90 -- ToDo: find the initial all-pointer segment and don't save them.
92 mkTaggedStkAmodes sp addrmode_regs
93 `thenFC` \(new_sp, stk_assts, more_tag_assts) ->
95 -- only let the extra stack assignments affect the stack
96 -- high water mark if we were doing a stack check anyway;
97 -- otherwise we end up generating unnecessary stack checks.
98 -- Careful about knot-tying loops!
99 let real_stk_words = if new_sp - sp > stk_words && stk_words /= 0
104 let adjust_sp = CAssign (CReg Sp) (CAddr (spRel sp new_sp)) in
106 absC (checking_code real_stk_words hHw
107 (mkAbstractCs [tag_assts, stk_assts, more_tag_assts,
109 (CReg node) 0 ticky_ctr)
113 setRealHp hHw `thenC`
118 checking_code stk hp assts ret regs ctr
121 if hp == 0 then AbsCNop
122 else profCtrAbsC FSLIT("TICK_ALLOC_HEAP")
123 [ mkIntCLit hp, CLbl ctr DataPtrRep ]
127 | node_points = do_checks_np stk hp assts (regs+1)
128 | otherwise = do_checks stk hp assts ret regs
130 -- When node points to the closure for the function:
133 :: Int -- stack headroom
134 -> Int -- heap headroom
135 -> AbstractC -- assignments to perform on failure
136 -> Int -- number of pointer registers live
138 do_checks_np 0 0 _ _ = AbsCNop
139 do_checks_np 0 hp_words tag_assts ptrs =
145 do_checks_np stk_words 0 tag_assts ptrs =
151 do_checks_np stk_words hp_words tag_assts ptrs =
152 CCheck HP_STK_CHK_NP [
159 -- When node doesn't point to the closure (we need an explicit retn addr)
162 :: Int -- stack headroom
163 -> Int -- heap headroom
164 -> AbstractC -- assignments to perform on failure
165 -> CAddrMode -- a register to hold the retn addr.
166 -> Int -- number of pointer registers live
169 do_checks 0 0 _ _ _ = AbsCNop
170 do_checks 0 hp_words tag_assts ret_reg ptrs =
178 do_checks stk_words 0 tag_assts ret_reg ptrs =
186 do_checks stk_words hp_words tag_assts ret_reg ptrs =
196 free_reg = case length regs + 1 of
197 I# x -> CReg (VanillaReg PtrRep x)
199 all_pointers = all pointer regs
200 pointer (VanillaReg rep _) = isFollowableRep rep
203 addrmode_regs = map CReg regs
205 -- Checking code for thunks is just a special case of fast entry points:
207 thunkChecks :: CLabel -> Bool -> Code -> Code
208 thunkChecks ret node_points code = fastEntryChecks [] [] ret node_points code
211 Heap checks in a case alternative are nice and easy, provided this is
212 a bog-standard algebraic case. We have in our hand:
214 * one return address, on the stack,
215 * one return value, in Node.
217 the canned code for this heap check failure just pushes Node on the
218 stack, saying 'EnterGHC' to return. The scheduler will return by
219 entering the top value on the stack, which in turn will return through
220 the return address, getting us back to where we were. This is
221 therefore only valid if the return value is *lifted* (just being
222 boxed isn't good enough). Only a PtrRep will do.
224 For primitive returns, we have an unlifted value in some register
225 (either R1 or FloatReg1 or DblReg1). This means using specialised
226 heap-check code for these cases.
228 For unboxed tuple returns, there are an arbitrary number of possibly
229 unboxed return values, some of which will be in registers, and the
230 others will be on the stack, with gaps left for tagging the unboxed
231 objects. If a heap check is required, we need to fill in these tags.
233 The code below will cover all cases for the x86 architecture (where R1
234 is the only VanillaReg ever used). For other architectures, we'll
235 have to do something about saving and restoring the other registers.
239 :: Bool -- is an algebraic alternative
240 -> [MagicId] -- live registers
241 -> [(VirtualSpOffset,Int)] -- stack slots to tag
243 -> Maybe Unique -- uniq of ret address (possibly)
247 -- unboxed tuple alternatives and let-no-escapes (the two most annoying
248 -- constructs to generate code for!):
250 altHeapCheck is_fun regs tags fail_code (Just ret_addr) code
251 = mkTagAssts tags `thenFC` \tag_assts1 ->
252 let tag_assts = mkAbstractCs [fail_code, tag_assts1]
254 initHeapUsage (\ hHw -> do_heap_chk hHw tag_assts `thenC` code)
256 do_heap_chk words_required tag_assts
257 = getTickyCtrLabel `thenFC` \ ctr ->
258 absC ( if words_required == 0
261 [ checking_code tag_assts,
262 profCtrAbsC FSLIT("TICK_ALLOC_HEAP")
263 [ mkIntCLit words_required, CLbl ctr DataPtrRep ]
266 setRealHp words_required
269 non_void_regs = filter (/= VoidReg) regs
271 checking_code tag_assts =
272 case non_void_regs of
274 {- no: there might be stuff on top of the retn. addr. on the stack.
277 [mkIntCLit words_required]
280 -- this will cover all cases for x86
283 | isFollowableRep rep ->
285 [mkIntCLit words_required, mkIntCLit 1, mkIntCLit 0,
286 CReg (VanillaReg RetRep 2#),
287 CLbl (mkReturnInfoLabel ret_addr) RetRep]
292 [mkIntCLit words_required, mkIntCLit 0, mkIntCLit 1,
293 CReg (VanillaReg RetRep 2#),
294 CLbl (mkReturnInfoLabel ret_addr) RetRep]
298 let liveness = mkRegLiveness several_regs
301 [mkIntCLit words_required,
302 mkIntCLit (I# (word2Int# liveness)),
303 -- HP_CHK_GEN needs a direct return address,
304 -- not an info table (might be different if
305 -- we're not assembly-mangling/tail-jumping etc.)
306 CLbl (mkReturnPtLabel ret_addr) RetRep]
309 -- normal algebraic and primitive case alternatives:
311 altHeapCheck is_fun regs [] AbsCNop Nothing code
312 = initHeapUsage (\ hHw -> do_heap_chk hHw `thenC` code)
314 do_heap_chk :: HeapOffset -> Code
315 do_heap_chk words_required
316 = getTickyCtrLabel `thenFC` \ ctr ->
317 absC ( if words_required == 0
321 profCtrAbsC FSLIT("TICK_ALLOC_HEAP")
322 [ mkIntCLit words_required, CLbl ctr DataPtrRep ]
325 setRealHp words_required
328 non_void_regs = filter (/= VoidReg) regs
331 case non_void_regs of
333 -- No regs live: probably a Void return
335 CCheck HP_CHK_NOREGS [mkIntCLit words_required] AbsCNop
337 -- The SEQ case (polymophic/function typed case branch)
338 -- We need this case because the closure in Node won't return
339 -- directly when we enter it (it could be a function), so the
340 -- heap check code needs to push a seq frame on top of the stack.
345 [mkIntCLit words_required, mkIntCLit 1{-regs live-}]
348 -- R1 is lifted (the common case)
352 [mkIntCLit words_required, mkIntCLit 1{-regs live-}]
355 -- R1 is boxed, but unlifted
356 | isFollowableRep rep ->
357 CCheck HP_CHK_UNPT_R1 [mkIntCLit words_required] AbsCNop
361 CCheck HP_CHK_UNBX_R1 [mkIntCLit words_required] AbsCNop
365 CCheck HP_CHK_F1 [mkIntCLit words_required] AbsCNop
369 CCheck HP_CHK_D1 [mkIntCLit words_required] AbsCNop
373 CCheck HP_CHK_L1 [mkIntCLit words_required] AbsCNop
376 _ -> panic ("CgHeapery.altHeapCheck: unimplemented heap-check, live regs = " ++ showSDoc (sep (map pprMagicId non_void_regs)))
379 -- build up a bitmap of the live pointer registers
381 #if __GLASGOW_HASKELL__ >= 503
382 shiftL = uncheckedShiftL#
387 mkRegLiveness :: [MagicId] -> Word#
388 mkRegLiveness [] = int2Word# 0#
389 mkRegLiveness (VanillaReg rep i : regs) | isFollowableRep rep
390 = ((int2Word# 1#) `shiftL` (i -# 1#)) `or#` mkRegLiveness regs
391 mkRegLiveness (_ : regs) = mkRegLiveness regs
393 -- The two functions below are only used in a GranSim setup
394 -- Emit macro for simulating a fetch and then reschedule
396 fetchAndReschedule :: [MagicId] -- Live registers
397 -> Bool -- Node reqd?
400 fetchAndReschedule regs node_reqd =
401 if (node `elem` regs || node_reqd)
402 then fetch_code `thenC` reschedule_code
405 liveness_mask = mkRegLiveness regs
406 reschedule_code = absC (CMacroStmt GRAN_RESCHEDULE [
407 mkIntCLit (I# (word2Int# liveness_mask)),
408 mkIntCLit (if node_reqd then 1 else 0)])
410 --HWL: generate GRAN_FETCH macro for GrAnSim
411 -- currently GRAN_FETCH and GRAN_FETCH_AND_RESCHEDULE are miai
412 fetch_code = absC (CMacroStmt GRAN_FETCH [])
415 The @GRAN_YIELD@ macro is taken from JSM's code for Concurrent Haskell. It
416 allows to context-switch at places where @node@ is not alive (it uses the
417 @Continue@ rather than the @EnterNodeCode@ function in the RTS). We emit
418 this kind of macro at the beginning of the following kinds of basic bocks:
420 \item Slow entry code where node is not alive (see @CgClosure.lhs@). Normally
421 we use @fetchAndReschedule@ at a slow entry code.
422 \item Fast entry code (see @CgClosure.lhs@).
423 \item Alternatives in case expressions (@CLabelledCode@ structures), provided
424 that they are not inlined (see @CgCases.lhs@). These alternatives will
425 be turned into separate functions.
429 yield :: [MagicId] -- Live registers
430 -> Bool -- Node reqd?
433 yield regs node_reqd =
434 if opt_GranMacros && node_reqd
438 liveness_mask = mkRegLiveness regs
440 absC (CMacroStmt GRAN_YIELD
441 [mkIntCLit (I# (word2Int# liveness_mask))])
444 %************************************************************************
446 \subsection[initClosure]{Initialise a dynamic closure}
448 %************************************************************************
450 @allocDynClosure@ puts the thing in the heap, and modifies the virtual Hp
456 -> CAddrMode -- Cost Centre to stick in the object
457 -> CAddrMode -- Cost Centre to blame for this alloc
458 -- (usually the same; sometimes "OVERHEAD")
460 -> [(CAddrMode, VirtualHeapOffset)] -- Offsets from start of the object
461 -- ie Info ptr has offset zero.
462 -> FCode VirtualHeapOffset -- Returns virt offset of object
464 allocDynClosure closure_info use_cc blame_cc amodes_with_offsets
465 = getVirtAndRealHp `thenFC` \ (virtHp, realHp) ->
467 -- FIND THE OFFSET OF THE INFO-PTR WORD
468 -- virtHp points to last allocated word, ie 1 *before* the
469 -- info-ptr word of new object.
470 let info_offset = virtHp + 1
472 -- do_move IS THE ASSIGNMENT FUNCTION
473 do_move (amode, offset_from_start)
474 = CAssign (CVal (hpRel realHp
475 (info_offset + offset_from_start))
479 -- SAY WHAT WE ARE ABOUT TO DO
480 profCtrC (allocProfilingMsg closure_info)
481 [mkIntCLit (closureGoodStuffSize closure_info),
482 mkIntCLit slop_size] `thenC`
485 absC ( mkAbstractCs (
486 [ CInitHdr closure_info
487 (CAddr (hpRel realHp info_offset))
488 use_cc closure_size ]
489 ++ (map do_move amodes_with_offsets))) `thenC`
491 -- BUMP THE VIRTUAL HEAP POINTER
492 setVirtHp (virtHp + closure_size) `thenC`
494 -- RETURN PTR TO START OF OBJECT
497 closure_size = closureSize closure_info
498 slop_size = slopSize closure_info
501 Occasionally we can update a closure in place instead of allocating
502 new space for it. This is the function that does the business, assuming:
504 - node points to the closure to be overwritten
506 - the new closure doesn't contain any pointers if we're
507 using a generational collector.
510 inPlaceAllocDynClosure
512 -> CAddrMode -- Pointer to beginning of closure
513 -> CAddrMode -- Cost Centre to stick in the object
515 -> [(CAddrMode, VirtualHeapOffset)] -- Offsets from start of the object
516 -- ie Info ptr has offset zero.
519 inPlaceAllocDynClosure closure_info head use_cc amodes_with_offsets
520 = let -- do_move IS THE ASSIGNMENT FUNCTION
521 do_move (amode, offset_from_start)
522 = CAssign (CVal (CIndex head (mkIntCLit offset_from_start) WordRep)
527 absC ( mkAbstractCs (
528 [ CInitHdr closure_info head use_cc 0{-no alloc-} ]
529 ++ (map do_move amodes_with_offsets)))