2 % (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
4 % $Id: CgHeapery.lhs,v 1.17 1999/05/26 14:12:13 simonmar Exp $
6 \section[CgHeapery]{Heap management functions}
10 fastEntryChecks, altHeapCheck, thunkChecks,
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 ( if all_pointers then -- heap checks are quite easy
79 absC (checking_code stk_words hp_words tag_assts
80 free_reg (length regs))
82 else -- they are complicated
84 -- save all registers on the stack and adjust the stack pointer.
85 -- ToDo: find the initial all-pointer segment and don't save them.
87 mkTaggedStkAmodes sp addrmode_regs
88 `thenFC` \(new_sp, stk_assts, more_tag_assts) ->
90 -- only let the extra stack assignments affect the stack
91 -- high water mark if we were doing a stack check anyway;
92 -- otherwise we end up generating unnecessary stack checks.
93 -- Careful about knot-tying loops!
94 let real_stk_words = if new_sp - sp > stk_words && stk_words /= 0
99 let adjust_sp = CAssign (CReg Sp) (CAddr (spRel sp new_sp)) in
101 absC (checking_code real_stk_words hp_words
102 (mkAbstractCs [tag_assts, stk_assts, more_tag_assts,
108 setRealHp hp_words `thenC`
113 checking_code stk hp assts ret regs
114 | node_points = do_checks_np stk hp assts (regs+1) -- ret not required
115 | otherwise = do_checks stk hp assts ret regs
117 -- When node points to the closure for the function:
120 :: Int -- stack headroom
121 -> Int -- heap headroom
122 -> AbstractC -- assignments to perform on failure
123 -> Int -- number of pointer registers live
125 do_checks_np 0 0 _ _ = AbsCNop
126 do_checks_np 0 hp_words tag_assts ptrs =
132 do_checks_np stk_words 0 tag_assts ptrs =
138 do_checks_np stk_words hp_words tag_assts ptrs =
139 CCheck HP_STK_CHK_NP [
146 -- When node doesn't point to the closure (we need an explicit retn addr)
149 :: Int -- stack headroom
150 -> Int -- heap headroom
151 -> AbstractC -- assignments to perform on failure
152 -> CAddrMode -- a register to hold the retn addr.
153 -> Int -- number of pointer registers live
156 do_checks 0 0 _ _ _ = AbsCNop
157 do_checks 0 hp_words tag_assts ret_reg ptrs =
165 do_checks stk_words 0 tag_assts ret_reg ptrs =
173 do_checks stk_words hp_words tag_assts ret_reg ptrs =
183 free_reg = case length regs + 1 of
184 IBOX(x) -> CReg (VanillaReg PtrRep x)
186 all_pointers = all pointer regs
187 pointer (VanillaReg rep _) = isFollowableRep rep
190 addrmode_regs = map CReg regs
192 -- Checking code for thunks is just a special case of fast entry points:
194 thunkChecks :: CLabel -> Bool -> Code -> Code
195 thunkChecks ret node_points code = fastEntryChecks [] [] ret node_points code
198 Heap checks in a case alternative are nice and easy, provided this is
199 a bog-standard algebraic case. We have in our hand:
201 * one return address, on the stack,
202 * one return value, in Node.
204 the canned code for this heap check failure just pushes Node on the
205 stack, saying 'EnterGHC' to return. The scheduler will return by
206 entering the top value on the stack, which in turn will return through
207 the return address, getting us back to where we were. This is
208 therefore only valid if the return value is *lifted* (just being
209 boxed isn't good enough). Only a PtrRep will do.
211 For primitive returns, we have an unlifted value in some register
212 (either R1 or FloatReg1 or DblReg1). This means using specialised
213 heap-check code for these cases.
215 For unboxed tuple returns, there are an arbitrary number of possibly
216 unboxed return values, some of which will be in registers, and the
217 others will be on the stack, with gaps left for tagging the unboxed
218 objects. If a heap check is required, we need to fill in these tags.
220 The code below will cover all cases for the x86 architecture (where R1
221 is the only VanillaReg ever used). For other architectures, we'll
222 have to do something about saving and restoring the other registers.
226 :: Bool -- is an algebraic alternative
227 -> [MagicId] -- live registers
228 -> [(VirtualSpOffset,Int)] -- stack slots to tag
230 -> Maybe Unique -- uniq of ret address (possibly)
234 -- unboxed tuple alternatives and let-no-escapes (the two most annoying
235 -- constructs to generate code for!):
237 altHeapCheck is_fun regs tags fail_code (Just ret_addr) code
238 = mkTagAssts tags `thenFC` \tag_assts1 ->
239 let tag_assts = mkAbstractCs [fail_code, tag_assts1]
241 initHeapUsage (\ hHw -> do_heap_chk hHw tag_assts `thenC` code)
243 do_heap_chk words_required tag_assts
244 = absC (if words_required == 0
246 else checking_code tag_assts) `thenC`
247 setRealHp words_required
250 non_void_regs = filter (/= VoidReg) regs
252 checking_code tag_assts =
253 case non_void_regs of
255 {- no: there might be stuff on top of the retn. addr. on the stack.
258 [mkIntCLit words_required]
261 -- this will cover all cases for x86
262 [VanillaReg rep ILIT(1)]
264 | isFollowableRep rep ->
266 [mkIntCLit words_required, mkIntCLit 1, mkIntCLit 0,
267 CReg (VanillaReg RetRep ILIT(2)),
268 CLbl (mkReturnInfoLabel ret_addr) RetRep]
273 [mkIntCLit words_required, mkIntCLit 0, mkIntCLit 1,
274 CReg (VanillaReg RetRep ILIT(2)),
275 CLbl (mkReturnInfoLabel ret_addr) RetRep]
279 let liveness = mkRegLiveness several_regs
282 [mkIntCLit words_required,
283 mkIntCLit (IBOX(word2Int# liveness)),
284 -- HP_CHK_GEN needs a direct return address,
285 -- not an info table (might be different if
286 -- we're not assembly-mangling/tail-jumping etc.)
287 CLbl (mkReturnPtLabel ret_addr) RetRep]
290 -- normal algebraic and primitive case alternatives:
292 altHeapCheck is_fun regs [] AbsCNop Nothing code
293 = initHeapUsage (\ hHw -> do_heap_chk hHw `thenC` code)
295 do_heap_chk :: HeapOffset -> Code
296 do_heap_chk words_required
297 = absC (if words_required == 0
299 else checking_code) `thenC`
300 setRealHp words_required
303 non_void_regs = filter (/= VoidReg) regs
306 case non_void_regs of
308 -- No regs live: probably a Void return
310 CCheck HP_CHK_NOREGS [mkIntCLit words_required] AbsCNop
312 -- The SEQ case (polymophic/function typed case branch)
313 -- We need this case because the closure in Node won't return
314 -- directly when we enter it (it could be a function), so the
315 -- heap check code needs to push a seq frame on top of the stack.
316 [VanillaReg rep ILIT(1)]
320 [mkIntCLit words_required, mkIntCLit 1{-regs live-}]
323 -- R1 is lifted (the common case)
324 [VanillaReg rep ILIT(1)]
327 [mkIntCLit words_required, mkIntCLit 1{-regs live-}]
330 -- R1 is boxed, but unlifted
331 | isFollowableRep rep ->
332 CCheck HP_CHK_UNPT_R1 [mkIntCLit words_required] AbsCNop
336 CCheck HP_CHK_UNBX_R1 [mkIntCLit words_required] AbsCNop
339 [FloatReg ILIT(1)] ->
340 CCheck HP_CHK_F1 [mkIntCLit words_required] AbsCNop
343 [DoubleReg ILIT(1)] ->
344 CCheck HP_CHK_D1 [mkIntCLit words_required] AbsCNop
347 [LongReg _ ILIT(1)] ->
348 CCheck HP_CHK_L1 [mkIntCLit words_required] AbsCNop
351 _ -> panic ("CgHeapery.altHeapCheck: unimplemented heap-check, live regs = " ++ showSDoc (sep (map pprMagicId non_void_regs)))
354 -- build up a bitmap of the live pointer registers
356 mkRegLiveness :: [MagicId] -> Word#
357 mkRegLiveness [] = int2Word# 0#
358 mkRegLiveness (VanillaReg rep i : regs) | isFollowableRep rep
359 = ((int2Word# 1#) `shiftL#` (i -# 1#)) `or#` mkRegLiveness regs
360 mkRegLiveness (_ : regs) = mkRegLiveness regs
362 -- Emit macro for simulating a fetch and then reschedule
364 fetchAndReschedule :: [MagicId] -- Live registers
365 -> Bool -- Node reqd?
368 fetchAndReschedule regs node_reqd =
369 if (node `elem` regs || node_reqd)
370 then fetch_code `thenC` reschedule_code
373 all_regs = if node_reqd then node:regs else regs
374 liveness_mask = 0 {-XXX: mkLiveRegsMask all_regs-}
376 reschedule_code = absC (CMacroStmt GRAN_RESCHEDULE [
377 mkIntCLit liveness_mask,
378 mkIntCLit (if node_reqd then 1 else 0)])
380 --HWL: generate GRAN_FETCH macro for GrAnSim
381 -- currently GRAN_FETCH and GRAN_FETCH_AND_RESCHEDULE are miai
382 fetch_code = absC (CMacroStmt GRAN_FETCH [])
385 The @GRAN_YIELD@ macro is taken from JSM's code for Concurrent Haskell. It
386 allows to context-switch at places where @node@ is not alive (it uses the
387 @Continue@ rather than the @EnterNodeCode@ function in the RTS). We emit
388 this kind of macro at the beginning of the following kinds of basic bocks:
390 \item Slow entry code where node is not alive (see @CgClosure.lhs@). Normally
391 we use @fetchAndReschedule@ at a slow entry code.
392 \item Fast entry code (see @CgClosure.lhs@).
393 \item Alternatives in case expressions (@CLabelledCode@ structures), provided
394 that they are not inlined (see @CgCases.lhs@). These alternatives will
395 be turned into separate functions.
399 yield :: [MagicId] -- Live registers
400 -> Bool -- Node reqd?
403 yield regs node_reqd =
404 -- NB: node is not alive; that's why we use DO_YIELD rather than
408 all_regs = if node_reqd then node:regs else regs
409 liveness_mask = 0 {-XXX: mkLiveRegsMask all_regs-}
411 yield_code = absC (CMacroStmt GRAN_YIELD [mkIntCLit liveness_mask])
414 %************************************************************************
416 \subsection[initClosure]{Initialise a dynamic closure}
418 %************************************************************************
420 @allocDynClosure@ puts the thing in the heap, and modifies the virtual Hp
426 -> CAddrMode -- Cost Centre to stick in the object
427 -> CAddrMode -- Cost Centre to blame for this alloc
428 -- (usually the same; sometimes "OVERHEAD")
430 -> [(CAddrMode, VirtualHeapOffset)] -- Offsets from start of the object
431 -- ie Info ptr has offset zero.
432 -> FCode VirtualHeapOffset -- Returns virt offset of object
434 allocDynClosure closure_info use_cc blame_cc amodes_with_offsets
435 = getVirtAndRealHp `thenFC` \ (virtHp, realHp) ->
437 -- FIND THE OFFSET OF THE INFO-PTR WORD
438 -- virtHp points to last allocated word, ie 1 *before* the
439 -- info-ptr word of new object.
440 let info_offset = virtHp + 1
442 -- do_move IS THE ASSIGNMENT FUNCTION
443 do_move (amode, offset_from_start)
444 = CAssign (CVal (hpRel realHp
445 (info_offset + offset_from_start))
449 -- SAY WHAT WE ARE ABOUT TO DO
450 profCtrC (allocProfilingMsg closure_info)
451 [mkIntCLit (closureGoodStuffSize closure_info),
452 mkIntCLit slop_size] `thenC`
455 absC ( mkAbstractCs (
456 [ cInitHdr closure_info (hpRel realHp info_offset) use_cc ]
457 ++ (map do_move amodes_with_offsets))) `thenC`
459 -- GENERATE CC PROFILING MESSAGES
460 costCentresC SLIT("CCS_ALLOC") [blame_cc, mkIntCLit closure_size]
463 -- BUMP THE VIRTUAL HEAP POINTER
464 setVirtHp (virtHp + closure_size) `thenC`
466 -- RETURN PTR TO START OF OBJECT
469 closure_size = closureSize closure_info
470 slop_size = slopSize closure_info
472 -- Avoid hanging on to anything in the CC field when we're not profiling.
474 cInitHdr closure_info amode cc
475 | opt_SccProfilingOn = CInitHdr closure_info amode cc
476 | otherwise = CInitHdr closure_info amode (panic "absent cc")