1 /* -----------------------------------------------------------------------------
2 * $Id: StgMacros.h,v 1.25 2000/05/08 15:05:36 simonmar Exp $
4 * (c) The GHC Team, 1998-1999
6 * Macros used for writing STG-ish C code.
8 * ---------------------------------------------------------------------------*/
13 /* -----------------------------------------------------------------------------
14 The following macros create function headers.
16 Each basic block is represented by a C function with no arguments.
17 We therefore always begin with either
25 The macros can be used either to define the function itself, or to provide
26 prototypes (by following with a ';').
28 Note: the various I*_ shorthands in the second block below are used to
29 declare forward references to local symbols. These shorthands *have* to
30 use the 'extern' type specifier and not 'static'. The reason for this is
31 that 'static' declares a reference as being a static/local variable,
32 and *not* as a forward reference to a static variable.
34 This might seem obvious, but it had me stumped as to why my info tables
35 were suddenly all filled with 0s.
39 --------------------------------------------------------------------------- */
41 #define STGFUN(f) StgFunPtr f(void)
42 #define EXTFUN(f) extern StgFunPtr f(void)
43 #define EXTFUN_RTS(f) extern DLL_IMPORT_RTS StgFunPtr f(void)
44 #define FN_(f) F_ f(void)
45 #define IFN_(f) static F_ f(void)
46 #define IF_(f) static F_ f(void)
47 #define EF_(f) extern F_ f(void)
48 #define EDF_(f) extern DLLIMPORT F_ f(void)
51 #define EDD_ extern DLLIMPORT
52 #define ED_RO_ extern const
54 #define ID_RO_ static const
55 #define EI_ extern INFO_TBL_CONST StgInfoTable
56 #define EDI_ extern DLLIMPORT INFO_TBL_CONST StgInfoTable
57 #define II_ static INFO_TBL_CONST StgInfoTable
58 #define EC_ extern StgClosure
59 #define EDC_ extern DLLIMPORT StgClosure
60 #define IC_ static StgClosure
61 #define ECP_(x) extern const StgClosure *(x)[]
62 #define EDCP_(x) extern DLLIMPORT StgClosure *(x)[]
63 #define ICP_(x) static const StgClosure *(x)[]
65 /* -----------------------------------------------------------------------------
68 For a block of non-pointer words on the stack, we precede the
69 block with a small-integer tag giving the number of non-pointer
71 -------------------------------------------------------------------------- */
73 #define ARGTAG_MAX 16 /* probably arbitrary */
74 #define ARG_TAG(n) (n)
75 #define ARG_SIZE(n) (StgWord)n
79 INT_TAG = sizeofW(StgInt),
80 INT64_TAG = sizeofW(StgInt64),
81 WORD_TAG = sizeofW(StgWord),
82 ADDR_TAG = sizeofW(StgAddr),
83 CHAR_TAG = sizeofW(StgChar),
84 FLOAT_TAG = sizeofW(StgFloat),
85 DOUBLE_TAG = sizeofW(StgDouble),
86 STABLE_TAG = sizeofW(StgWord),
89 static inline int IS_ARG_TAG( StgWord p );
90 static inline int IS_ARG_TAG( StgWord p ) { return p <= ARGTAG_MAX; }
92 /* -----------------------------------------------------------------------------
95 If (Sp + <n_args>) > Su { JMP_(stg_update_PAP); }
97 Sp points to the topmost used word on the stack, and Su points to
98 the most recently pushed update frame.
100 Remember that <n_args> must include any tagging of unboxed values.
102 ARGS_CHK_LOAD_NODE is for top-level functions, whose entry
103 convention doesn't require that Node is loaded with a pointer to
104 the closure. Thus we must load node before calling stg_updatePAP if
105 the argument check fails.
106 -------------------------------------------------------------------------- */
108 #define ARGS_CHK(n) \
109 if ((P_)(Sp + (n)) > (P_)Su) { \
110 JMP_(stg_update_PAP); \
113 #define ARGS_CHK_LOAD_NODE(n,closure) \
114 if ((P_)(Sp + (n)) > (P_)Su) { \
115 R1.p = (P_)closure; \
116 JMP_(stg_update_PAP); \
119 /* -----------------------------------------------------------------------------
122 When failing a check, we save a return address on the stack and
123 jump to a pre-compiled code fragment that saves the live registers
124 and returns to the scheduler.
126 The return address in most cases will be the beginning of the basic
127 block in which the check resides, since we need to perform the check
128 again on re-entry because someone else might have stolen the resource
130 ------------------------------------------------------------------------- */
132 #define STK_CHK(headroom,ret,r,layout,tag_assts) \
133 if (Sp - headroom < SpLim) { \
134 EXTFUN_RTS(stg_chk_##layout); \
137 JMP_(stg_chk_##layout); \
140 #define HP_CHK(headroom,ret,r,layout,tag_assts) \
141 DO_GRAN_ALLOCATE(headroom) \
142 if ((Hp += headroom) > HpLim) { \
143 EXTFUN_RTS(stg_chk_##layout); \
146 JMP_(stg_chk_##layout); \
149 #define HP_STK_CHK(stk_headroom,hp_headroom,ret,r,layout,tag_assts) \
150 DO_GRAN_ALLOCATE(hp_headroom) \
151 if (Sp - stk_headroom < SpLim || (Hp += hp_headroom) > HpLim) { \
152 EXTFUN_RTS(stg_chk_##layout); \
155 JMP_(stg_chk_##layout); \
158 /* -----------------------------------------------------------------------------
159 A Heap Check in a case alternative are much simpler: everything is
160 on the stack and covered by a liveness mask already, and there is
161 even a return address with an SRT info table there as well.
163 Just push R1 and return to the scheduler saying 'EnterGHC'
165 {STK,HP,HP_STK}_CHK_NP are the various checking macros for
166 bog-standard case alternatives, thunks, and non-top-level
167 functions. In all these cases, node points to a closure that we
168 can just enter to restart the heap check (the NP stands for 'node points').
170 In the NP case GranSim absolutely has to check whether the current node
171 resides on the current processor. Otherwise a FETCH event has to be
172 scheduled. All that is done in GranSimFetch. -- HWL
174 HpLim points to the LAST WORD of valid allocation space.
175 -------------------------------------------------------------------------- */
177 #define STK_CHK_NP(headroom,ptrs,tag_assts) \
178 if ((Sp - (headroom)) < SpLim) { \
179 EXTFUN_RTS(stg_gc_enter_##ptrs); \
181 JMP_(stg_gc_enter_##ptrs); \
184 #define HP_CHK_NP(headroom,ptrs,tag_assts) \
185 DO_GRAN_ALLOCATE(headroom) \
186 if ((Hp += (headroom)) > HpLim) { \
187 EXTFUN_RTS(stg_gc_enter_##ptrs); \
189 JMP_(stg_gc_enter_##ptrs); \
192 #define HP_CHK_SEQ_NP(headroom,ptrs,tag_assts) \
193 DO_GRAN_ALLOCATE(headroom) \
194 if ((Hp += (headroom)) > HpLim) { \
195 EXTFUN_RTS(stg_gc_seq_##ptrs); \
197 JMP_(stg_gc_seq_##ptrs); \
200 #define HP_STK_CHK_NP(stk_headroom, hp_headroom, ptrs, tag_assts) \
201 DO_GRAN_ALLOCATE(hp_headroom) \
202 if ((Sp - (stk_headroom)) < SpLim || (Hp += (hp_headroom)) > HpLim) { \
203 EXTFUN_RTS(stg_gc_enter_##ptrs); \
205 JMP_(stg_gc_enter_##ptrs); \
209 /* Heap checks for branches of a primitive case / unboxed tuple return */
211 #define GEN_HP_CHK_ALT(headroom,lbl,tag_assts) \
212 DO_GRAN_ALLOCATE(headroom) \
213 if ((Hp += (headroom)) > HpLim) { \
219 #define HP_CHK_NOREGS(headroom,tag_assts) \
220 GEN_HP_CHK_ALT(headroom,stg_gc_noregs,tag_assts);
221 #define HP_CHK_UNPT_R1(headroom,tag_assts) \
222 GEN_HP_CHK_ALT(headroom,stg_gc_unpt_r1,tag_assts);
223 #define HP_CHK_UNBX_R1(headroom,tag_assts) \
224 GEN_HP_CHK_ALT(headroom,stg_gc_unbx_r1,tag_assts);
225 #define HP_CHK_F1(headroom,tag_assts) \
226 GEN_HP_CHK_ALT(headroom,stg_gc_f1,tag_assts);
227 #define HP_CHK_D1(headroom,tag_assts) \
228 GEN_HP_CHK_ALT(headroom,stg_gc_d1,tag_assts);
230 #define HP_CHK_L1(headroom,tag_assts) \
231 GEN_HP_CHK_ALT(headroom,stg_gc_d1,tag_assts);
233 #define HP_CHK_UT_ALT(headroom, ptrs, nptrs, r, ret, tag_assts) \
234 GEN_HP_CHK_ALT(headroom, stg_gc_ut_##ptrs##_##nptrs, \
235 tag_assts r = (P_)ret;)
237 /* -----------------------------------------------------------------------------
240 These are slow, but have the advantage of being usable in a variety
243 The one restriction is that any relevant SRTs must already be pointed
244 to from the stack. The return address doesn't need to have an info
245 table attached: hence it can be any old code pointer.
247 The liveness mask is a logical 'XOR' of NO_PTRS and zero or more
248 Rn_PTR constants defined below. All registers will be saved, but
249 the garbage collector needs to know which ones contain pointers.
251 Good places to use a generic heap check:
253 - case alternatives (the return address with an SRT is already
256 - primitives (no SRT required).
258 The stack layout is like this:
267 so the liveness mask depends on the size of an StgDouble (FltRegs
268 and R<n> are guaranteed to be 1 word in size).
270 -------------------------------------------------------------------------- */
272 /* VERY MAGIC CONSTANTS!
273 * must agree with code in HeapStackCheck.c, stg_gen_chk
276 #if SIZEOF_DOUBLE > SIZEOF_VOID_P
277 #define ALL_NON_PTRS 0xffff
278 #else /* SIZEOF_DOUBLE == SIZEOF_VOID_P */
279 #define ALL_NON_PTRS 0x3fff
282 #define LIVENESS_MASK(ptr_regs) (ALL_NON_PTRS ^ (ptr_regs))
294 #define HP_CHK_GEN(headroom,liveness,reentry,tag_assts) \
295 if ((Hp += (headroom)) > HpLim ) { \
298 R9.w = (W_)LIVENESS_MASK(liveness); \
299 R10.w = (W_)reentry; \
303 #define HP_CHK_GEN_TICKY(headroom,liveness,reentry,tag_assts) \
304 HP_CHK_GEN(headroom,liveness,reentry,tag_assts); \
305 TICK_ALLOC_HEAP_NOCTR(headroom)
307 #define STK_CHK_GEN(headroom,liveness,reentry,tag_assts) \
308 if ((Sp - (headroom)) < SpLim) { \
311 R9.w = (W_)LIVENESS_MASK(liveness); \
312 R10.w = (W_)reentry; \
316 #define MAYBE_GC(liveness,reentry) \
317 if (doYouWantToGC()) { \
319 R9.w = (W_)LIVENESS_MASK(liveness); \
320 R10.w = (W_)reentry; \
324 /* -----------------------------------------------------------------------------
325 Voluntary Yields/Blocks
327 We only have a generic version of this at the moment - if it turns
328 out to be slowing us down we can make specialised ones.
329 -------------------------------------------------------------------------- */
334 #define YIELD(liveness,reentry) \
336 R9.w = (W_)LIVENESS_MASK(liveness); \
337 R10.w = (W_)reentry; \
338 JMP_(stg_gen_yield); \
341 #define BLOCK(liveness,reentry) \
343 R9.w = (W_)LIVENESS_MASK(liveness); \
344 R10.w = (W_)reentry; \
345 JMP_(stg_gen_block); \
348 #define BLOCK_NP(ptrs) \
350 EF_(stg_block_##ptrs); \
351 JMP_(stg_block_##ptrs); \
356 Similar to BLOCK_NP but separates the saving of the thread state from the
357 actual jump via an StgReturn
360 #define SAVE_THREAD_STATE(ptrs) \
366 #define THREAD_RETURN(ptrs) \
368 CurrentTSO->what_next = ThreadEnterGHC; \
369 R1.i = ThreadBlocked; \
373 /* -----------------------------------------------------------------------------
374 CCall_GC needs to push a dummy stack frame containing the contents
375 of volatile registers and variables.
377 We use a RET_DYN frame the same as for a dynamic heap check.
378 ------------------------------------------------------------------------- */
381 EI_(stg_gen_chk_info);
383 EDI_(stg_gen_chk_info);
385 /* -----------------------------------------------------------------------------
388 RETVEC(p,t) where 'p' is a pointer to the info table for a
389 vectored return address, returns the address of the return code for
392 Return vectors are placed in *reverse order* immediately before the info
393 table for the return address. Hence the formula for computing the
394 actual return address is (addr - sizeof(InfoTable) - tag - 1).
395 The extra subtraction of one word is because tags start at zero.
396 -------------------------------------------------------------------------- */
398 #ifdef TABLES_NEXT_TO_CODE
399 #define RET_VEC(p,t) (*((P_)(p) - sizeofW(StgInfoTable) - t - 1))
401 #define RET_VEC(p,t) (((StgInfoTable *)p)->vector[t])
404 /* -----------------------------------------------------------------------------
406 -------------------------------------------------------------------------- */
409 /* set the tag register (if we have one) */
410 #define SET_TAG(t) /* nothing */
412 #ifdef EAGER_BLACKHOLING
414 # define UPD_BH_UPDATABLE(info) \
415 TICK_UPD_BH_UPDATABLE(); \
417 bdescr *bd = Bdescr(R1.p); \
418 if (bd->back != (bdescr *)BaseReg) { \
419 if (bd->gen->no >= 1 || bd->step->no >= 1) { \
422 EXTFUN_RTS(stg_gc_enter_1_hponly); \
423 JMP_(stg_gc_enter_1_hponly); \
427 SET_INFO(R1.cl,&BLACKHOLE_info)
428 # define UPD_BH_SINGLE_ENTRY(info) \
429 TICK_UPD_BH_SINGLE_ENTRY(); \
431 bdescr *bd = Bdescr(R1.p); \
432 if (bd->back != (bdescr *)BaseReg) { \
433 if (bd->gen->no >= 1 || bd->step->no >= 1) { \
436 EXTFUN_RTS(stg_gc_enter_1_hponly); \
437 JMP_(stg_gc_enter_1_hponly); \
441 SET_INFO(R1.cl,&BLACKHOLE_info)
443 # define UPD_BH_UPDATABLE(info) \
444 TICK_UPD_BH_UPDATABLE(); \
445 SET_INFO(R1.cl,&BLACKHOLE_info)
446 # define UPD_BH_SINGLE_ENTRY(info) \
447 TICK_UPD_BH_SINGLE_ENTRY(); \
448 SET_INFO(R1.cl,&SE_BLACKHOLE_info)
450 #else /* !EAGER_BLACKHOLING */
451 # define UPD_BH_UPDATABLE(thunk) /* nothing */
452 # define UPD_BH_SINGLE_ENTRY(thunk) /* nothing */
453 #endif /* EAGER_BLACKHOLING */
455 #define UPD_FRAME_UPDATEE(p) (((StgUpdateFrame *)(p))->updatee)
456 #define UPDATE_SU_FROM_UPD_FRAME(p) (Su=((StgUpdateFrame *)(p))->link)
458 /* -----------------------------------------------------------------------------
459 Moving Floats and Doubles
461 ASSIGN_FLT is for assigning a float to memory (usually the
462 stack/heap). The memory address is guaranteed to be
463 StgWord aligned (currently == sizeof(long)).
465 PK_FLT is for pulling a float out of memory. The memory is
466 guaranteed to be StgWord aligned.
467 -------------------------------------------------------------------------- */
469 static inline void ASSIGN_FLT (W_ [], StgFloat);
470 static inline StgFloat PK_FLT (W_ []);
472 #if ALIGNMENT_FLOAT <= ALIGNMENT_LONG
474 static inline void ASSIGN_FLT(W_ p_dest[], StgFloat src) { *(StgFloat *)p_dest = src; }
475 static inline StgFloat PK_FLT (W_ p_src[]) { return *(StgFloat *)p_src; }
477 #else /* ALIGNMENT_FLOAT > ALIGNMENT_UNSIGNED_INT */
479 static inline void ASSIGN_FLT(W_ p_dest[], StgFloat src)
486 static inline StgFloat PK_FLT(W_ p_src[])
493 #endif /* ALIGNMENT_FLOAT > ALIGNMENT_LONG */
495 #if ALIGNMENT_DOUBLE <= ALIGNMENT_LONG
497 static inline void ASSIGN_DBL (W_ [], StgDouble);
498 static inline StgDouble PK_DBL (W_ []);
500 static inline void ASSIGN_DBL(W_ p_dest[], StgDouble src) { *(StgDouble *)p_dest = src; }
501 static inline StgDouble PK_DBL (W_ p_src[]) { return *(StgDouble *)p_src; }
503 #else /* ALIGNMENT_DOUBLE > ALIGNMENT_LONG */
505 /* Sparc uses two floating point registers to hold a double. We can
506 * write ASSIGN_DBL and PK_DBL by directly accessing the registers
507 * independently - unfortunately this code isn't writable in C, we
508 * have to use inline assembler.
510 #if sparc_TARGET_ARCH
512 #define ASSIGN_DBL(dst0,src) \
513 { StgPtr dst = (StgPtr)(dst0); \
514 __asm__("st %2,%0\n\tst %R2,%1" : "=m" (((P_)(dst))[0]), \
515 "=m" (((P_)(dst))[1]) : "f" (src)); \
518 #define PK_DBL(src0) \
519 ( { StgPtr src = (StgPtr)(src0); \
521 __asm__("ld %1,%0\n\tld %2,%R0" : "=f" (d) : \
522 "m" (((P_)(src))[0]), "m" (((P_)(src))[1])); d; \
525 #else /* ! sparc_TARGET_ARCH */
527 static inline void ASSIGN_DBL (W_ [], StgDouble);
528 static inline StgDouble PK_DBL (W_ []);
540 static inline void ASSIGN_DBL(W_ p_dest[], StgDouble src)
544 p_dest[0] = y.du.dhi;
545 p_dest[1] = y.du.dlo;
548 /* GCC also works with this version, but it generates
549 the same code as the previous one, and is not ANSI
551 #define ASSIGN_DBL( p_dest, src ) \
552 *p_dest = ((double_thing) src).du.dhi; \
553 *(p_dest+1) = ((double_thing) src).du.dlo \
556 static inline StgDouble PK_DBL(W_ p_src[])
564 #endif /* ! sparc_TARGET_ARCH */
566 #endif /* ALIGNMENT_DOUBLE > ALIGNMENT_UNSIGNED_INT */
568 #ifdef SUPPORT_LONG_LONGS
573 } unpacked_double_word;
577 unpacked_double_word iu;
582 unpacked_double_word wu;
585 static inline void ASSIGN_Word64(W_ p_dest[], StgWord64 src)
589 p_dest[0] = y.wu.dhi;
590 p_dest[1] = y.wu.dlo;
593 static inline StgWord64 PK_Word64(W_ p_src[])
601 static inline void ASSIGN_Int64(W_ p_dest[], StgInt64 src)
605 p_dest[0] = y.iu.dhi;
606 p_dest[1] = y.iu.dlo;
609 static inline StgInt64 PK_Int64(W_ p_src[])
618 /* -----------------------------------------------------------------------------
620 -------------------------------------------------------------------------- */
622 extern DLL_IMPORT_DATA const StgPolyInfoTable catch_frame_info;
624 /* -----------------------------------------------------------------------------
627 A seq frame is very like an update frame, except that it doesn't do
629 -------------------------------------------------------------------------- */
631 extern DLL_IMPORT_DATA const StgPolyInfoTable seq_frame_info;
633 #define PUSH_SEQ_FRAME(sp) \
635 StgSeqFrame *__frame; \
636 TICK_SEQF_PUSHED(); \
637 __frame = (StgSeqFrame *)(sp); \
638 SET_HDR_(__frame,&seq_frame_info,CCCS); \
639 __frame->link = Su; \
640 Su = (StgUpdateFrame *)__frame; \
643 /* -----------------------------------------------------------------------------
645 -------------------------------------------------------------------------- */
647 #if defined(USE_SPLIT_MARKERS)
648 #define __STG_SPLIT_MARKER(n) FN_(__stg_split_marker##n) { }
650 #define __STG_SPLIT_MARKER(n) /* nothing */
653 /* -----------------------------------------------------------------------------
654 Closure and Info Macros with casting.
656 We don't want to mess around with casts in the generated C code, so
657 we use these casting versions of the closure/info tables macros.
658 -------------------------------------------------------------------------- */
660 #define SET_HDR_(c,info,ccs) \
661 SET_HDR((StgClosure *)(c),(StgInfoTable *)(info),ccs)
663 /* -----------------------------------------------------------------------------
664 Saving context for exit from the STG world, and loading up context
665 on entry to STG code.
667 We save all the STG registers (that is, the ones that are mapped to
668 machine registers) in their places in the TSO.
670 The stack registers go into the current stack object, and the
671 current nursery is updated from the heap pointer.
673 These functions assume that BaseReg is loaded appropriately (if
675 -------------------------------------------------------------------------- */
679 static __inline__ void
680 SaveThreadState(void)
684 /* Don't need to save REG_Base, it won't have changed. */
692 #ifdef REG_CurrentTSO
693 SAVE_CurrentTSO = tso;
695 #ifdef REG_CurrentNursery
696 SAVE_CurrentNursery = CurrentNursery;
698 #if defined(PROFILING)
699 CurrentTSO->prof.CCCS = CCCS;
703 static __inline__ void
704 LoadThreadState (void)
708 #ifdef REG_CurrentTSO
709 CurrentTSO = SAVE_CurrentTSO;
716 OpenNursery(Hp,HpLim);
718 #ifdef REG_CurrentNursery
719 CurrentNursery = SAVE_CurrentNursery;
721 # if defined(PROFILING)
722 CCCS = CurrentTSO->prof.CCCS;
728 /* -----------------------------------------------------------------------------
729 Module initialisation
730 -------------------------------------------------------------------------- */
732 #define PUSH_INIT_STACK(reg_function) \
733 *(Sp++) = (W_)reg_function
735 #define POP_INIT_STACK() \
738 #define START_MOD_INIT(reg_mod_name) \
739 static int _module_registered = 0; \
740 FN_(reg_mod_name) { \
742 if (! _module_registered) { \
743 _module_registered = 1; \
745 /* extern decls go here, followed by init code */
747 #define REGISTER_FOREIGN_EXPORT(reg_fe_binder) \
748 STGCALL1(getStablePtr,reg_fe_binder)
750 #define REGISTER_IMPORT(reg_mod_name) \
751 do { EF_(reg_mod_name); \
752 PUSH_INIT_STACK(reg_mod_name) ; \
755 #define END_MOD_INIT() \
757 JMP_(POP_INIT_STACK()); \
760 /* -----------------------------------------------------------------------------
761 Support for _ccall_GC_ and _casm_GC.
762 -------------------------------------------------------------------------- */
765 * Suspending/resuming threads for doing external C-calls (_ccall_GC).
766 * These functions are defined in rts/Schedule.c.
768 StgInt suspendThread ( StgRegTable *cap );
769 StgRegTable * resumeThread ( StgInt );
771 #define SUSPEND_THREAD(token) \
773 token = suspendThread(BaseReg);
776 #define RESUME_THREAD(token) \
777 BaseReg = resumeThread(token); \
780 #define RESUME_THREAD(token) \
781 (void)resumeThread(token); \
785 #endif /* STGMACROS_H */