1 /* -----------------------------------------------------------------------------
2 * $Id: StgMacros.h,v 1.50 2002/12/11 15:36:39 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 IF_(f) static F_ f(void)
46 #define EF_(f) extern F_ f(void)
47 #define EDF_(f) extern DLLIMPORT F_ f(void)
49 #define EXTINFO_RTS extern DLL_IMPORT_RTS const StgInfoTable
50 #define ETI_RTS extern DLL_IMPORT_RTS const StgThunkInfoTable
52 // Info tables as generated by the compiler are simply arrays of words.
53 typedef StgWord StgWordArray[];
56 #define EDD_ extern DLLIMPORT
57 #define ED_RO_ extern const
59 #define ID_RO_ static const
60 #define EI_ extern StgWordArray
61 #define ERI_ extern const StgRetInfoTable
62 #define II_ static StgWordArray
63 #define IRI_ static const StgRetInfoTable
64 #define EC_ extern StgClosure
65 #define EDC_ extern DLLIMPORT StgClosure
66 #define IC_ static StgClosure
67 #define ECP_(x) extern const StgClosure *(x)[]
68 #define EDCP_(x) extern DLLIMPORT StgClosure *(x)[]
69 #define ICP_(x) static const StgClosure *(x)[]
71 /* -----------------------------------------------------------------------------
74 It isn't safe to "enter" every closure. Functions in particular
75 have no entry code as such; their entry point contains the code to
77 -------------------------------------------------------------------------- */
82 switch (get_itbl(R1.cl)->type) { \
86 case IND_OLDGEN_PERM: \
88 R1.cl = ((StgInd *)R1.cl)->indirectee; \
99 JMP_(ENTRY_CODE(Sp[0])); \
101 JMP_(GET_ENTRY(R1.cl)); \
105 /* -----------------------------------------------------------------------------
108 When failing a check, we save a return address on the stack and
109 jump to a pre-compiled code fragment that saves the live registers
110 and returns to the scheduler.
112 The return address in most cases will be the beginning of the basic
113 block in which the check resides, since we need to perform the check
114 again on re-entry because someone else might have stolen the resource
116 ------------------------------------------------------------------------- */
118 #define STK_CHK_FUN(headroom,assts) \
119 if (Sp - headroom < SpLim) { \
124 #define HP_CHK_FUN(headroom,assts) \
125 DO_GRAN_ALLOCATE(headroom) \
126 if ((Hp += headroom) > HpLim) { \
127 HpAlloc = (headroom); \
132 // When doing both a heap and a stack check, don't move the heap
133 // pointer unless the stack check succeeds. Otherwise we might end up
134 // with slop at the end of the current block, which can confuse the
136 #define HP_STK_CHK_FUN(stk_headroom,hp_headroom,assts) \
137 DO_GRAN_ALLOCATE(hp_headroom) \
138 if (Sp - stk_headroom < SpLim || (Hp += hp_headroom) > HpLim) { \
139 HpAlloc = (hp_headroom); \
144 /* -----------------------------------------------------------------------------
145 A Heap Check in a case alternative are much simpler: everything is
146 on the stack and covered by a liveness mask already, and there is
147 even a return address with an SRT info table there as well.
149 Just push R1 and return to the scheduler saying 'EnterGHC'
151 {STK,HP,HP_STK}_CHK_NP are the various checking macros for
152 bog-standard case alternatives, thunks, and non-top-level
153 functions. In all these cases, node points to a closure that we
154 can just enter to restart the heap check (the NP stands for 'node points').
156 In the NP case GranSim absolutely has to check whether the current node
157 resides on the current processor. Otherwise a FETCH event has to be
158 scheduled. All that is done in GranSimFetch. -- HWL
160 HpLim points to the LAST WORD of valid allocation space.
161 -------------------------------------------------------------------------- */
163 #define STK_CHK_NP(headroom,tag_assts) \
164 if ((Sp - (headroom)) < SpLim) { \
166 JMP_(stg_gc_enter_1); \
169 #define HP_CHK_NP(headroom,tag_assts) \
170 DO_GRAN_ALLOCATE(headroom) \
171 if ((Hp += (headroom)) > HpLim) { \
172 HpAlloc = (headroom); \
174 JMP_(stg_gc_enter_1); \
177 // See comment on HP_STK_CHK_FUN above.
178 #define HP_STK_CHK_NP(stk_headroom, hp_headroom, tag_assts) \
179 DO_GRAN_ALLOCATE(hp_headroom) \
180 if ((Sp - (stk_headroom)) < SpLim || (Hp += (hp_headroom)) > HpLim) { \
181 HpAlloc = (hp_headroom); \
183 JMP_(stg_gc_enter_1); \
187 /* Heap checks for branches of a primitive case / unboxed tuple return */
189 #define GEN_HP_CHK_ALT(headroom,lbl,tag_assts) \
190 DO_GRAN_ALLOCATE(headroom) \
191 if ((Hp += (headroom)) > HpLim) { \
192 HpAlloc = (headroom); \
197 #define HP_CHK_NOREGS(headroom,tag_assts) \
198 GEN_HP_CHK_ALT(headroom,stg_gc_noregs,tag_assts);
199 #define HP_CHK_UNPT_R1(headroom,tag_assts) \
200 GEN_HP_CHK_ALT(headroom,stg_gc_unpt_r1,tag_assts);
201 #define HP_CHK_UNBX_R1(headroom,tag_assts) \
202 GEN_HP_CHK_ALT(headroom,stg_gc_unbx_r1,tag_assts);
203 #define HP_CHK_F1(headroom,tag_assts) \
204 GEN_HP_CHK_ALT(headroom,stg_gc_f1,tag_assts);
205 #define HP_CHK_D1(headroom,tag_assts) \
206 GEN_HP_CHK_ALT(headroom,stg_gc_d1,tag_assts);
207 #define HP_CHK_L1(headroom,tag_assts) \
208 GEN_HP_CHK_ALT(headroom,stg_gc_l1,tag_assts);
210 /* -----------------------------------------------------------------------------
213 These are slow, but have the advantage of being usable in a variety
216 The one restriction is that any relevant SRTs must already be pointed
217 to from the stack. The return address doesn't need to have an info
218 table attached: hence it can be any old code pointer.
220 The liveness mask is a logical 'XOR' of NO_PTRS and zero or more
221 Rn_PTR constants defined below. All registers will be saved, but
222 the garbage collector needs to know which ones contain pointers.
224 Good places to use a generic heap check:
226 - case alternatives (the return address with an SRT is already
229 - primitives (no SRT required).
231 The stack frame layout for a RET_DYN is like this:
242 we assume that the size of a double is always 2 pointers (wasting a
243 word when it is only one pointer, but avoiding lots of #ifdefs).
245 -------------------------------------------------------------------------- */
247 // VERY MAGIC CONSTANTS!
248 // must agree with code in HeapStackCheck.c, stg_gen_chk
250 #define ALL_NON_PTRS 0xffff
251 #define RET_DYN_SIZE 16
253 #define LIVENESS_MASK(ptr_regs) (ALL_NON_PTRS ^ (ptr_regs))
255 // We can have up to 255 pointers and 255 nonpointers in the stack
257 #define N_NONPTRS(n) ((n)<<16)
258 #define N_PTRS(n) ((n)<<24)
260 #define GET_NONPTRS(l) ((l)>>16 & 0xff)
261 #define GET_PTRS(l) ((l)>>24 & 0xff)
262 #define GET_LIVENESS(l) ((l) & 0xffff)
274 #define HP_CHK_UNBX_TUPLE(headroom,liveness,code) \
275 if ((Hp += (headroom)) > HpLim ) { \
276 HpAlloc = (headroom); \
278 R9.w = (W_)LIVENESS_MASK(liveness); \
282 #define HP_CHK_GEN(headroom,liveness,reentry) \
283 if ((Hp += (headroom)) > HpLim ) { \
284 HpAlloc = (headroom); \
285 R9.w = (W_)LIVENESS_MASK(liveness); \
286 R10.w = (W_)reentry; \
290 #define HP_CHK_GEN_TICKY(headroom,liveness,reentry) \
291 HP_CHK_GEN(headroom,liveness,reentry); \
292 TICK_ALLOC_HEAP_NOCTR(headroom)
294 #define STK_CHK_GEN(headroom,liveness,reentry) \
295 if ((Sp - (headroom)) < SpLim) { \
296 R9.w = (W_)LIVENESS_MASK(liveness); \
297 R10.w = (W_)reentry; \
301 #define MAYBE_GC(liveness,reentry) \
302 if (doYouWantToGC()) { \
303 R9.w = (W_)LIVENESS_MASK(liveness); \
304 R10.w = (W_)reentry; \
305 JMP_(stg_gc_gen_hp); \
308 /* -----------------------------------------------------------------------------
309 Voluntary Yields/Blocks
311 We only have a generic version of this at the moment - if it turns
312 out to be slowing us down we can make specialised ones.
313 -------------------------------------------------------------------------- */
315 EXTFUN_RTS(stg_gen_yield);
316 EXTFUN_RTS(stg_gen_block);
318 #define YIELD(liveness,reentry) \
320 R9.w = (W_)LIVENESS_MASK(liveness); \
321 R10.w = (W_)reentry; \
322 JMP_(stg_gen_yield); \
325 #define BLOCK(liveness,reentry) \
327 R9.w = (W_)LIVENESS_MASK(liveness); \
328 R10.w = (W_)reentry; \
329 JMP_(stg_gen_block); \
332 #define BLOCK_NP(ptrs) \
334 EXTFUN_RTS(stg_block_##ptrs); \
335 JMP_(stg_block_##ptrs); \
340 Similar to BLOCK_NP but separates the saving of the thread state from the
341 actual jump via an StgReturn
344 #define SAVE_THREAD_STATE(ptrs) \
350 #define THREAD_RETURN(ptrs) \
352 CurrentTSO->what_next = ThreadEnterGHC; \
353 R1.i = ThreadBlocked; \
357 /* -----------------------------------------------------------------------------
358 CCall_GC needs to push a dummy stack frame containing the contents
359 of volatile registers and variables.
361 We use a RET_DYN frame the same as for a dynamic heap check.
362 ------------------------------------------------------------------------- */
364 /* -----------------------------------------------------------------------------
367 RETVEC(p,t) where 'p' is a pointer to the info table for a
368 vectored return address, returns the address of the return code for
371 Return vectors are placed in *reverse order* immediately before the info
372 table for the return address. Hence the formula for computing the
373 actual return address is (addr - sizeof(RetInfoTable) - tag - 1).
374 The extra subtraction of one word is because tags start at zero.
375 -------------------------------------------------------------------------- */
377 #ifdef TABLES_NEXT_TO_CODE
378 #define RET_VEC(p,t) (*((P_)(p) - sizeofW(StgRetInfoTable) - t - 1))
380 #define RET_VEC(p,t) (((StgRetInfoTable *)p)->vector[t])
383 /* -----------------------------------------------------------------------------
385 -------------------------------------------------------------------------- */
388 /* set the tag register (if we have one) */
389 #define SET_TAG(t) /* nothing */
391 #ifdef EAGER_BLACKHOLING
393 # define UPD_BH_UPDATABLE(info) \
394 TICK_UPD_BH_UPDATABLE(); \
396 bdescr *bd = Bdescr(R1.p); \
397 if (bd->u.back != (bdescr *)BaseReg) { \
398 if (bd->gen_no >= 1 || bd->step->no >= 1) { \
401 EXTFUN_RTS(stg_gc_enter_1_hponly); \
402 JMP_(stg_gc_enter_1_hponly); \
406 SET_INFO(R1.cl,&stg_BLACKHOLE_info)
407 # define UPD_BH_SINGLE_ENTRY(info) \
408 TICK_UPD_BH_SINGLE_ENTRY(); \
410 bdescr *bd = Bdescr(R1.p); \
411 if (bd->u.back != (bdescr *)BaseReg) { \
412 if (bd->gen_no >= 1 || bd->step->no >= 1) { \
415 EXTFUN_RTS(stg_gc_enter_1_hponly); \
416 JMP_(stg_gc_enter_1_hponly); \
420 SET_INFO(R1.cl,&stg_BLACKHOLE_info)
423 # define UPD_BH_UPDATABLE(info) \
424 TICK_UPD_BH_UPDATABLE(); \
425 SET_INFO(R1.cl,&stg_BLACKHOLE_info)
426 # define UPD_BH_SINGLE_ENTRY(info) \
427 TICK_UPD_BH_SINGLE_ENTRY(); \
428 SET_INFO(R1.cl,&stg_SE_BLACKHOLE_info)
430 // An object is replaced by a blackhole, so we fill the slop with zeros.
432 // Todo: maybe use SET_HDR() and remove LDV_recordCreate()?
434 # define UPD_BH_UPDATABLE(info) \
435 TICK_UPD_BH_UPDATABLE(); \
436 LDV_recordDead_FILL_SLOP_DYNAMIC(R1.cl); \
437 SET_INFO(R1.cl,&stg_BLACKHOLE_info); \
438 LDV_recordCreate(R1.cl)
439 # define UPD_BH_SINGLE_ENTRY(info) \
440 TICK_UPD_BH_SINGLE_ENTRY(); \
441 LDV_recordDead_FILL_SLOP_DYNAMIC(R1.cl); \
442 SET_INFO(R1.cl,&stg_SE_BLACKHOLE_info) \
443 LDV_recordCreate(R1.cl)
444 # endif /* PROFILING */
446 #else /* !EAGER_BLACKHOLING */
447 # define UPD_BH_UPDATABLE(thunk) /* nothing */
448 # define UPD_BH_SINGLE_ENTRY(thunk) /* nothing */
449 #endif /* EAGER_BLACKHOLING */
451 #define UPD_FRAME_UPDATEE(p) ((P_)(((StgUpdateFrame *)(p))->updatee))
453 /* -----------------------------------------------------------------------------
454 Moving Floats and Doubles
456 ASSIGN_FLT is for assigning a float to memory (usually the
457 stack/heap). The memory address is guaranteed to be
458 StgWord aligned (currently == sizeof(void *)).
460 PK_FLT is for pulling a float out of memory. The memory is
461 guaranteed to be StgWord aligned.
462 -------------------------------------------------------------------------- */
464 static inline void ASSIGN_FLT (W_ [], StgFloat);
465 static inline StgFloat PK_FLT (W_ []);
467 #if ALIGNMENT_FLOAT <= ALIGNMENT_LONG
469 static inline void ASSIGN_FLT(W_ p_dest[], StgFloat src) { *(StgFloat *)p_dest = src; }
470 static inline StgFloat PK_FLT (W_ p_src[]) { return *(StgFloat *)p_src; }
472 #else /* ALIGNMENT_FLOAT > ALIGNMENT_UNSIGNED_INT */
474 static inline void ASSIGN_FLT(W_ p_dest[], StgFloat src)
481 static inline StgFloat PK_FLT(W_ p_src[])
488 #endif /* ALIGNMENT_FLOAT > ALIGNMENT_LONG */
490 #if ALIGNMENT_DOUBLE <= ALIGNMENT_LONG
492 static inline void ASSIGN_DBL (W_ [], StgDouble);
493 static inline StgDouble PK_DBL (W_ []);
495 static inline void ASSIGN_DBL(W_ p_dest[], StgDouble src) { *(StgDouble *)p_dest = src; }
496 static inline StgDouble PK_DBL (W_ p_src[]) { return *(StgDouble *)p_src; }
498 #else /* ALIGNMENT_DOUBLE > ALIGNMENT_LONG */
500 /* Sparc uses two floating point registers to hold a double. We can
501 * write ASSIGN_DBL and PK_DBL by directly accessing the registers
502 * independently - unfortunately this code isn't writable in C, we
503 * have to use inline assembler.
505 #if sparc_TARGET_ARCH
507 #define ASSIGN_DBL(dst0,src) \
508 { StgPtr dst = (StgPtr)(dst0); \
509 __asm__("st %2,%0\n\tst %R2,%1" : "=m" (((P_)(dst))[0]), \
510 "=m" (((P_)(dst))[1]) : "f" (src)); \
513 #define PK_DBL(src0) \
514 ( { StgPtr src = (StgPtr)(src0); \
516 __asm__("ld %1,%0\n\tld %2,%R0" : "=f" (d) : \
517 "m" (((P_)(src))[0]), "m" (((P_)(src))[1])); d; \
520 #else /* ! sparc_TARGET_ARCH */
522 static inline void ASSIGN_DBL (W_ [], StgDouble);
523 static inline StgDouble PK_DBL (W_ []);
535 static inline void ASSIGN_DBL(W_ p_dest[], StgDouble src)
539 p_dest[0] = y.du.dhi;
540 p_dest[1] = y.du.dlo;
543 /* GCC also works with this version, but it generates
544 the same code as the previous one, and is not ANSI
546 #define ASSIGN_DBL( p_dest, src ) \
547 *p_dest = ((double_thing) src).du.dhi; \
548 *(p_dest+1) = ((double_thing) src).du.dlo \
551 static inline StgDouble PK_DBL(W_ p_src[])
559 #endif /* ! sparc_TARGET_ARCH */
561 #endif /* ALIGNMENT_DOUBLE > ALIGNMENT_UNSIGNED_INT */
563 #ifdef SUPPORT_LONG_LONGS
568 } unpacked_double_word;
572 unpacked_double_word iu;
577 unpacked_double_word wu;
580 static inline void ASSIGN_Word64(W_ p_dest[], StgWord64 src)
584 p_dest[0] = y.wu.dhi;
585 p_dest[1] = y.wu.dlo;
588 static inline StgWord64 PK_Word64(W_ p_src[])
596 static inline void ASSIGN_Int64(W_ p_dest[], StgInt64 src)
600 p_dest[0] = y.iu.dhi;
601 p_dest[1] = y.iu.dlo;
604 static inline StgInt64 PK_Int64(W_ p_src[])
612 #elif SIZEOF_VOID_P == 8
614 static inline void ASSIGN_Word64(W_ p_dest[], StgWord64 src)
619 static inline StgWord64 PK_Word64(W_ p_src[])
624 static inline void ASSIGN_Int64(W_ p_dest[], StgInt64 src)
629 static inline StgInt64 PK_Int64(W_ p_src[])
636 /* -----------------------------------------------------------------------------
638 -------------------------------------------------------------------------- */
640 extern DLL_IMPORT_RTS const StgPolyInfoTable stg_catch_frame_info;
642 /* -----------------------------------------------------------------------------
644 -------------------------------------------------------------------------- */
646 #if defined(USE_SPLIT_MARKERS)
647 #if defined(LEADING_UNDERSCORE)
648 #define __STG_SPLIT_MARKER __asm__("\n___stg_split_marker:");
650 #define __STG_SPLIT_MARKER __asm__("\n__stg_split_marker:");
653 #define __STG_SPLIT_MARKER /* nothing */
656 /* -----------------------------------------------------------------------------
657 Closure and Info Macros with casting.
659 We don't want to mess around with casts in the generated C code, so
660 we use this casting versions of the closure macro.
662 This version of SET_HDR also includes CCS_ALLOC for profiling - the
663 reason we don't use two separate macros is that the cost centre
664 field is sometimes a non-simple expression and we want to share its
665 value between SET_HDR and CCS_ALLOC.
666 -------------------------------------------------------------------------- */
668 #define SET_HDR_(c,info,ccs,size) \
670 CostCentreStack *tmp = (ccs); \
671 SET_HDR((StgClosure *)(c),(StgInfoTable *)(info),tmp); \
672 CCS_ALLOC(tmp,size); \
675 /* -----------------------------------------------------------------------------
676 Saving context for exit from the STG world, and loading up context
677 on entry to STG code.
679 We save all the STG registers (that is, the ones that are mapped to
680 machine registers) in their places in the TSO.
682 The stack registers go into the current stack object, and the
683 current nursery is updated from the heap pointer.
685 These functions assume that BaseReg is loaded appropriately (if
687 -------------------------------------------------------------------------- */
691 static __inline__ void
692 SaveThreadState(void)
696 /* Don't need to save REG_Base, it won't have changed. */
702 #ifdef REG_CurrentTSO
703 SAVE_CurrentTSO = tso;
705 #ifdef REG_CurrentNursery
706 SAVE_CurrentNursery = CurrentNursery;
708 #if defined(PROFILING)
709 CurrentTSO->prof.CCCS = CCCS;
713 static __inline__ void
714 LoadThreadState (void)
718 #ifdef REG_CurrentTSO
719 CurrentTSO = SAVE_CurrentTSO;
724 SpLim = (P_)&(tso->stack) + RESERVED_STACK_WORDS;
725 OpenNursery(Hp,HpLim);
727 #ifdef REG_CurrentNursery
728 CurrentNursery = SAVE_CurrentNursery;
730 # if defined(PROFILING)
731 CCCS = CurrentTSO->prof.CCCS;
737 /* -----------------------------------------------------------------------------
738 Module initialisation
740 The module initialisation code looks like this, roughly:
743 JMP_(__stginit_Foo_1_p)
746 FN(__stginit_Foo_1_p) {
750 We have one version of the init code with a module version and the
751 'way' attached to it. The version number helps to catch cases
752 where modules are not compiled in dependency order before being
753 linked: if a module has been compiled since any modules which depend on
754 it, then the latter modules will refer to a different version in their
755 init blocks and a link error will ensue.
757 The 'way' suffix helps to catch cases where modules compiled in different
758 ways are linked together (eg. profiled and non-profiled).
760 We provide a plain, unadorned, version of the module init code
761 which just jumps to the version with the label and way attached. The
762 reason for this is that when using foreign exports, the caller of
763 startupHaskell() must supply the name of the init function for the "top"
764 module in the program, and we don't want to require that this name
765 has the version and way info appended to it.
766 -------------------------------------------------------------------------- */
768 #define PUSH_INIT_STACK(reg_function) \
769 *(Sp++) = (W_)reg_function
771 #define POP_INIT_STACK() \
774 #define MOD_INIT_WRAPPER(label,real_init) \
777 #define START_MOD_INIT(plain_lbl, real_lbl) \
778 static int _module_registered = 0; \
787 if (! _module_registered) { \
788 _module_registered = 1; \
790 /* extern decls go here, followed by init code */
792 #define REGISTER_FOREIGN_EXPORT(reg_fe_binder) \
793 STGCALL1(getStablePtr,reg_fe_binder)
795 #define REGISTER_IMPORT(reg_mod_name) \
796 PUSH_INIT_STACK(reg_mod_name)
798 #define END_MOD_INIT() \
800 JMP_(POP_INIT_STACK()); \
803 /* -----------------------------------------------------------------------------
804 Support for _ccall_GC_ and _casm_GC.
805 -------------------------------------------------------------------------- */
808 * Suspending/resuming threads for doing external C-calls (_ccall_GC).
809 * These functions are defined in rts/Schedule.c.
811 StgInt suspendThread ( StgRegTable *, rtsBool);
812 StgRegTable * resumeThread ( StgInt, rtsBool );
814 #define SUSPEND_THREAD(token,threaded) \
816 token = suspendThread(BaseReg,threaded);
819 #define RESUME_THREAD(token,threaded) \
820 BaseReg = resumeThread(token,threaded); \
823 #define RESUME_THREAD(token,threaded) \
824 (void)resumeThread(token,threaded); \
828 #endif /* STGMACROS_H */