1 /* -----------------------------------------------------------------------------
2 * $Id: HeapStackCheck.hc,v 1.9 1999/08/25 16:11:48 simonmar Exp $
4 * (c) The GHC Team, 1998-1999
6 * Canned Heap-Check and Stack-Check sequences.
8 * ---------------------------------------------------------------------------*/
11 #include "Storage.h" /* for CurrentTSO */
12 #include "StgRun.h" /* for StgReturn and register saving */
13 #include "Schedule.h" /* for context_switch */
14 #include "HeapStackCheck.h"
16 /* Stack/Heap Check Failure
17 * ------------------------
19 * On discovering that a stack or heap check has failed, we do the following:
21 * - If the context_switch flag is set, indicating that there are more
22 * threads waiting to run, we yield to the scheduler
23 * (return ThreadYeilding).
25 * - If Hp > HpLim, we've had a heap check failure. This means we've
26 * come to the end of the current heap block, so we try to chain
27 * another block on with ExtendNursery().
29 * - If this succeeds, we carry on without returning to the
32 * - If it fails, we return to the scheduler claiming HeapOverflow
33 * so that a garbage collection can be performed.
35 * - If Hp <= HpLim, it must have been a stack check that failed. In
36 * which case, we return to the scheduler claiming StackOverflow, the
37 * scheduler will either increase the size of our stack, or flag
38 * an error if the stack is already too big.
40 * The effect of checking for context switch only in the heap/stack check
41 * failure code is that we'll switch threads after the current thread has
42 * reached the end of its heap block. If a thread isn't allocating
43 * at all, it won't yield. Hopefully this won't be a problem in practice.
46 /* Remember that the return address is *removed* when returning to a
47 * ThreadRunGHC thread.
53 if (ExtendNursery(Hp,HpLim)) { \
54 if (context_switch) { \
55 R1.i = ThreadYielding; \
58 JMP_(ENTRY_CODE(Sp[-1])); \
61 R1.i = HeapOverflow; \
64 R1.i = StackOverflow; \
67 CurrentTSO->whatNext = ThreadRunGHC; \
72 if (ExtendNursery(Hp,HpLim)) { \
73 if (context_switch) { \
74 R1.i = ThreadYielding; \
78 JMP_(ENTRY_CODE(*R1.p)); \
81 R1.i = HeapOverflow; \
84 R1.i = StackOverflow; \
87 CurrentTSO->whatNext = ThreadEnterGHC; \
92 CurrentTSO->whatNext = ThreadRunGHC; \
93 R1.i = HeapOverflow; \
98 CurrentTSO->whatNext = ThreadRunGHC; \
99 R1.i = StackOverflow; \
102 #define YIELD_GENERIC \
104 CurrentTSO->whatNext = ThreadRunGHC; \
105 R1.i = ThreadYielding; \
108 #define YIELD_TO_HUGS \
110 CurrentTSO->whatNext = ThreadEnterHugs; \
111 R1.i = ThreadYielding; \
114 #define BLOCK_GENERIC \
116 CurrentTSO->whatNext = ThreadRunGHC; \
117 R1.i = ThreadBlocked; \
120 #define BLOCK_ENTER \
122 CurrentTSO->whatNext = ThreadEnterGHC;\
123 R1.i = ThreadBlocked; \
126 /* -----------------------------------------------------------------------------
128 -------------------------------------------------------------------------- */
131 * This one is used when we want to *enter* the top thing on the stack
132 * when we return, instead of the just returning to an address. See
133 * UpdatePAP for an example.
136 EXTFUN(stg_gc_entertop)
143 /* -----------------------------------------------------------------------------
144 Heap checks in non-top-level thunks/functions.
146 In these cases, node always points to the function closure. This gives
147 us an easy way to return to the function: just leave R1 on the top of
148 the stack, and have the scheduler enter it to return.
150 There are canned sequences for 'n' pointer values in registers.
151 -------------------------------------------------------------------------- */
153 EXTFUN(stg_gc_enter_1)
162 /*- 2 Regs--------------------------------------------------------------------*/
164 EXTFUN(stg_gc_enter_2)
174 /*- 3 Regs -------------------------------------------------------------------*/
176 EXTFUN(stg_gc_enter_3)
187 /*- 4 Regs -------------------------------------------------------------------*/
189 EXTFUN(stg_gc_enter_4)
201 /*- 5 Regs -------------------------------------------------------------------*/
203 EXTFUN(stg_gc_enter_5)
216 /*- 6 Regs -------------------------------------------------------------------*/
218 EXTFUN(stg_gc_enter_6)
232 /*- 7 Regs -------------------------------------------------------------------*/
234 EXTFUN(stg_gc_enter_7)
249 /*- 8 Regs -------------------------------------------------------------------*/
251 EXTFUN(stg_gc_enter_8)
267 /* -----------------------------------------------------------------------------
268 For a case expression on a polymorphic or function-typed object, if
269 the default branch (there can only be one branch) of the case fails
270 a heap-check, instead of using stg_gc_enter_1 as normal, we must
271 push a new SEQ frame on the stack, followed by the object returned.
273 Otherwise, if the object is a function, it won't return to the
274 correct activation record on returning from garbage collection. It will
275 assume it has some arguments and apply itself.
276 -------------------------------------------------------------------------- */
281 Sp -= 1 + sizeofW(StgSeqFrame);
282 PUSH_SEQ_FRAME(Sp+1);
288 /* -----------------------------------------------------------------------------
289 Heap checks in Primitive case alternatives
291 A primitive case alternative is entered with a value either in
292 R1, FloatReg1 or D1 depending on the return convention. All the
293 cases are covered below.
294 -------------------------------------------------------------------------- */
296 /*-- No regsiters live (probably a void return) ----------------------------- */
298 /* If we change the policy for thread startup to *not* remove the
299 * return address from the stack, we can get rid of this little
300 * function/info table...
302 INFO_TABLE_SRT_BITMAP(stg_gc_noregs_ret_info, stg_gc_noregs_ret, 0/*BITMAP*/,
303 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
304 RET_SMALL,, EF_, 0, 0);
306 EXTFUN(stg_gc_noregs_ret)
309 JMP_(ENTRY_CODE(Sp[0]));
313 EXTFUN(stg_gc_noregs)
317 Sp[0] = (W_)&stg_gc_noregs_ret_info;
322 /*-- R1 is boxed/unpointed -------------------------------------------------- */
324 INFO_TABLE_SRT_BITMAP(stg_gc_unpt_r1_info, stg_gc_unpt_r1_entry, 0/*BITMAP*/,
325 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
326 RET_SMALL,, EF_, 0, 0);
328 EXTFUN(stg_gc_unpt_r1_entry)
333 JMP_(ENTRY_CODE(Sp[0]));
337 EXTFUN(stg_gc_unpt_r1)
342 Sp[0] = (W_)&stg_gc_unpt_r1_info;
347 /*-- R1 is unboxed -------------------------------------------------- */
349 INFO_TABLE_SRT_BITMAP(stg_gc_unbx_r1_info, stg_gc_unbx_r1_entry, 1/*BITMAP*/,
350 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
351 RET_SMALL,, EF_, 0, 0);
352 /* the 1 is a bitmap - i.e. 1 non-pointer word on the stack. */
354 EXTFUN(stg_gc_unbx_r1_entry)
359 JMP_(ENTRY_CODE(Sp[0]));
363 EXTFUN(stg_gc_unbx_r1)
368 Sp[0] = (W_)&stg_gc_unbx_r1_info;
373 /*-- F1 contains a float ------------------------------------------------- */
375 INFO_TABLE_SRT_BITMAP(stg_gc_f1_info, stg_gc_f1_entry, 1/*BITMAP*/,
376 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
377 RET_SMALL,, EF_, 0, 0);
379 EXTFUN(stg_gc_f1_entry)
384 JMP_(ENTRY_CODE(Sp[0]));
392 ASSIGN_FLT(Sp+1, F1);
393 Sp[0] = (W_)&stg_gc_f1_info;
398 /*-- D1 contains a double ------------------------------------------------- */
400 /* we support doubles of either 1 or 2 words in size */
402 #if SIZEOF_DOUBLE == SIZEOF_VOID_P
403 # define DBL_BITMAP 1
405 # define DBL_BITMAP 3
408 INFO_TABLE_SRT_BITMAP(stg_gc_d1_info, stg_gc_d1_entry, DBL_BITMAP,
409 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
410 RET_SMALL,, EF_, 0, 0);
412 EXTFUN(stg_gc_d1_entry)
416 Sp += sizeofW(StgDouble);
417 JMP_(ENTRY_CODE(Sp[0]));
424 Sp -= 1 + sizeofW(StgDouble);
426 Sp[0] = (W_)&stg_gc_d1_info;
431 /* -----------------------------------------------------------------------------
432 Heap checks for unboxed tuple case alternatives
436 - for an unboxed tuple with n components, we rearrange the components
437 with pointers first followed by non-pointers. (NB: not done yet)
439 - The first k components are allocated registers, where k is the
440 number of components that will fit in real registers.
442 - The rest are placed on the stack, with space left for tagging
443 of the non-pointer block if necessary.
445 - On failure of a heap check:
446 - the tag is filled in if necessary,
447 - we load Ri with the address of the continuation,
448 where i is the lowest unused vanilla register.
449 - jump to 'stg_gc_ut_x_y' where x is the number of pointer
450 registers and y the number of non-pointers.
451 - if the required canned sequence isn't available, it will
452 have to be generated at compile-time by the code
453 generator (this will probably happen if there are
454 floating-point values, for instance).
456 For now, just deal with R1, hence R2 contains the sequel address.
457 -------------------------------------------------------------------------- */
459 /*---- R1 contains a pointer: ------ */
461 INFO_TABLE_SRT_BITMAP(stg_gc_ut_1_0_info, stg_gc_ut_1_0_entry, 1/*BITMAP*/,
462 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
463 RET_SMALL,, EF_, 0, 0);
465 EXTFUN(stg_gc_ut_1_0_entry)
470 JMP_(ENTRY_CODE(Sp[-2]));
474 EXTFUN(stg_gc_ut_1_0)
480 Sp[0] = (W_)&stg_gc_ut_1_0_info;
485 /*---- R1 contains a non-pointer: ------ */
487 INFO_TABLE_SRT_BITMAP(stg_gc_ut_0_1_info, stg_gc_ut_0_1_entry, 3/*BITMAP*/,
488 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
489 RET_SMALL,, EF_, 0, 0);
491 EXTFUN(stg_gc_ut_0_1_entry)
496 JMP_(ENTRY_CODE(Sp[-2]));
500 EXTFUN(stg_gc_ut_0_1)
504 Sp[0] = (W_)&stg_gc_ut_0_1_info;
511 /* -----------------------------------------------------------------------------
512 Standard top-level fast-entry heap checks.
514 - we want to make the stack look like it should at the slow entry
515 point for the function. That way we can just push the slow
516 entry point on the stack and return using ThreadRunGHC.
518 - The compiler will generate code to fill in any tags on the stack,
519 in case we arrived directly at the fast entry point and these tags
522 - The rest is hopefully handled by jumping to a canned sequence.
523 We currently have canned sequences for 0-8 pointer registers. If
524 any registers contain non-pointers, we must reduce to an all-pointers
525 situation by pushing as many registers on the stack as necessary.
527 eg. if R1, R2 contain pointers and R3 contains a word, the heap check
528 failure sequence looks like this:
535 after pushing R3, we have pointers in R1 and R2 which corresponds
536 to the 2-pointer canned sequence.
538 -------------------------------------------------------------------------- */
540 /*- 0 Regs -------------------------------------------------------------------*/
551 /*- 1 Reg --------------------------------------------------------------------*/
563 /*- 1 Reg (non-ptr) ----------------------------------------------------------*/
570 Sp[1] = WORD_TAG; /* ToDo: or maybe its an int? */
576 /*- 2 Regs--------------------------------------------------------------------*/
589 /*- 3 Regs -------------------------------------------------------------------*/
603 /*- 4 Regs -------------------------------------------------------------------*/
618 /*- 5 Regs -------------------------------------------------------------------*/
634 /*- 6 Regs -------------------------------------------------------------------*/
651 /*- 7 Regs -------------------------------------------------------------------*/
669 /*- 8 Regs -------------------------------------------------------------------*/
688 /* -----------------------------------------------------------------------------
689 Generic Heap Check Code.
691 Called with Liveness mask in R9, Return address in R10.
692 Stack must be consistent (tagged, and containing all necessary info pointers
695 We also define an stg_gen_yield here, because it's very similar.
696 -------------------------------------------------------------------------- */
698 #if SIZEOF_DOUBLE > SIZEOF_VOID_P
700 #define RESTORE_EVERYTHING \
701 D2 = PK_DBL(Sp+16); \
702 D1 = PK_DBL(Sp+14); \
703 F4 = PK_FLT(Sp+13); \
704 F3 = PK_FLT(Sp+12); \
705 F2 = PK_FLT(Sp+11); \
706 F1 = PK_FLT(Sp+10); \
717 #define RET_OFFSET (-17)
719 #define SAVE_EVERYTHING \
720 ASSIGN_DBL(Sp-2,D2); \
721 ASSIGN_DBL(Sp-4,D1); \
722 ASSIGN_FLT(Sp-5,F4); \
723 ASSIGN_FLT(Sp-6,F3); \
724 ASSIGN_FLT(Sp-7,F2); \
725 ASSIGN_FLT(Sp-8,F1); \
734 Sp[-17] = R10.w; /* return address */ \
735 Sp[-18] = R9.w; /* liveness mask */ \
736 Sp[-19] = (W_)&stg_gen_chk_info; \
741 #define RESTORE_EVERYTHING \
742 D2 = PK_DBL(Sp+15); \
743 D1 = PK_DBL(Sp+14); \
744 F4 = PK_FLT(Sp+13); \
745 F3 = PK_FLT(Sp+12); \
746 F2 = PK_FLT(Sp+11); \
747 F1 = PK_FLT(Sp+10); \
758 #define RET_OFFSET (-15)
760 #define SAVE_EVERYTHING \
761 ASSIGN_DBL(Sp-1,D2); \
762 ASSIGN_DBL(Sp-2,D1); \
763 ASSIGN_FLT(Sp-3,F4); \
764 ASSIGN_FLT(Sp-4,F3); \
765 ASSIGN_FLT(Sp-5,F2); \
766 ASSIGN_FLT(Sp-6,F1); \
775 Sp[-15] = R10.w; /* return address */ \
776 Sp[-16] = R9.w; /* liveness mask */ \
777 Sp[-17] = (W_)&stg_gen_chk_info; \
782 INFO_TABLE_SRT_BITMAP(stg_gen_chk_info, stg_gen_chk_ret, 0,
783 0/*SRT*/, 0/*SRT_OFF*/, 0/*SRT_LEN*/,
784 RET_DYN,, EF_, 0, 0);
786 /* bitmap in the above info table is unused, the real one is on the stack.
793 JMP_(Sp[RET_OFFSET]); /* NO ENTRY_CODE() - this is a direct ret address */
806 * stg_gen_hp is used by MAYBE_GC, where we can't use GC_GENERIC
807 * because we've just failed doYouWantToGC(), not a standard heap
808 * check. GC_GENERIC would end up returning StackOverflow.
818 /* -----------------------------------------------------------------------------
820 -------------------------------------------------------------------------- */
830 FN_(stg_yield_noregs)
834 Sp[0] = (W_)&stg_gc_noregs_ret_info;
839 FN_(stg_yield_to_Hugs)
842 /* No need to save everything - no live registers */
847 /* -----------------------------------------------------------------------------
849 -------------------------------------------------------------------------- */
859 FN_(stg_block_noregs)
863 Sp[0] = (W_)&stg_gc_noregs_ret_info;