1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2004
5 * Out-of-line primitive operations
7 * This file contains the implementations of all the primitive
8 * operations ("primops") which are not expanded inline. See
9 * ghc/compiler/prelude/primops.txt.pp for a list of all the primops;
10 * this file contains code for most of those with the attribute
13 * Entry convention: the entry convention for a primop is that all the
14 * args are in Stg registers (R1, R2, etc.). This is to make writing
15 * the primops easier. (see compiler/codeGen/CgCallConv.hs).
17 * Return convention: results from a primop are generally returned
18 * using the ordinary unboxed tuple return convention. The C-- parser
19 * implements the RET_xxxx() macros to perform unboxed-tuple returns
20 * based on the prevailing return convention.
22 * This file is written in a subset of C--, extended with various
23 * features specific to GHC. It is compiled by GHC directly. For the
24 * syntax of .cmm files, see the parser in ghc/compiler/cmm/CmmParse.y.
26 * ---------------------------------------------------------------------------*/
31 import pthread_mutex_lock;
32 import pthread_mutex_unlock;
34 import base_ControlziExceptionziBase_nestedAtomically_closure;
35 import EnterCriticalSection;
36 import LeaveCriticalSection;
37 import ghczmprim_GHCziBool_False_closure;
38 #if !defined(mingw32_HOST_OS)
42 /*-----------------------------------------------------------------------------
45 Basically just new*Array - the others are all inline macros.
47 The size arg is always passed in R1, and the result returned in R1.
49 The slow entry point is for returning from a heap check, the saved
50 size argument must be re-loaded from the stack.
51 -------------------------------------------------------------------------- */
53 /* for objects that are *less* than the size of a word, make sure we
54 * round up to the nearest word for the size of the array.
59 W_ words, payload_words, n, p;
60 MAYBE_GC(NO_PTRS,stg_newByteArrayzh);
62 payload_words = ROUNDUP_BYTES_TO_WDS(n);
63 words = BYTES_TO_WDS(SIZEOF_StgArrWords) + payload_words;
64 ("ptr" p) = foreign "C" allocate(MyCapability() "ptr",words) [];
65 TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);
66 SET_HDR(p, stg_ARR_WORDS_info, W_[CCCS]);
67 StgArrWords_words(p) = payload_words;
72 #define BA_MASK (BA_ALIGN-1)
74 stg_newPinnedByteArrayzh
76 W_ words, bytes, payload_words, p;
78 MAYBE_GC(NO_PTRS,stg_newPinnedByteArrayzh);
80 /* payload_words is what we will tell the profiler we had to allocate */
81 payload_words = ROUNDUP_BYTES_TO_WDS(bytes);
82 /* When we actually allocate memory, we need to allow space for the
84 bytes = bytes + SIZEOF_StgArrWords;
85 /* And we want to align to BA_ALIGN bytes, so we need to allow space
86 to shift up to BA_ALIGN - 1 bytes: */
87 bytes = bytes + BA_ALIGN - 1;
88 /* Now we convert to a number of words: */
89 words = ROUNDUP_BYTES_TO_WDS(bytes);
91 ("ptr" p) = foreign "C" allocatePinned(MyCapability() "ptr", words) [];
92 TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);
94 /* Now we need to move p forward so that the payload is aligned
96 p = p + ((-p - SIZEOF_StgArrWords) & BA_MASK);
98 SET_HDR(p, stg_ARR_WORDS_info, W_[CCCS]);
99 StgArrWords_words(p) = payload_words;
103 stg_newAlignedPinnedByteArrayzh
105 W_ words, bytes, payload_words, p, alignment;
107 MAYBE_GC(NO_PTRS,stg_newAlignedPinnedByteArrayzh);
111 /* payload_words is what we will tell the profiler we had to allocate */
112 payload_words = ROUNDUP_BYTES_TO_WDS(bytes);
114 /* When we actually allocate memory, we need to allow space for the
116 bytes = bytes + SIZEOF_StgArrWords;
117 /* And we want to align to <alignment> bytes, so we need to allow space
118 to shift up to <alignment - 1> bytes: */
119 bytes = bytes + alignment - 1;
120 /* Now we convert to a number of words: */
121 words = ROUNDUP_BYTES_TO_WDS(bytes);
123 ("ptr" p) = foreign "C" allocatePinned(MyCapability() "ptr", words) [];
124 TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);
126 /* Now we need to move p forward so that the payload is aligned
127 to <alignment> bytes. Note that we are assuming that
128 <alignment> is a power of 2, which is technically not guaranteed */
129 p = p + ((-p - SIZEOF_StgArrWords) & (alignment - 1));
131 SET_HDR(p, stg_ARR_WORDS_info, W_[CCCS]);
132 StgArrWords_words(p) = payload_words;
138 W_ words, n, init, arr, p, size;
139 /* Args: R1 = words, R2 = initialisation value */
142 MAYBE_GC(R2_PTR,stg_newArrayzh);
144 // the mark area contains one byte for each 2^MUT_ARR_PTRS_CARD_BITS words
145 // in the array, making sure we round up, and then rounding up to a whole
147 size = n + mutArrPtrsCardWords(n);
148 words = BYTES_TO_WDS(SIZEOF_StgMutArrPtrs) + size;
149 ("ptr" arr) = foreign "C" allocate(MyCapability() "ptr",words) [R2];
150 TICK_ALLOC_PRIM(SIZEOF_StgMutArrPtrs, WDS(n), 0);
152 SET_HDR(arr, stg_MUT_ARR_PTRS_DIRTY_info, W_[CCCS]);
153 StgMutArrPtrs_ptrs(arr) = n;
154 StgMutArrPtrs_size(arr) = size;
156 // Initialise all elements of the the array with the value in R2
158 p = arr + SIZEOF_StgMutArrPtrs;
160 if (p < arr + WDS(words)) {
165 // Initialise the mark bits with 0
167 if (p < arr + WDS(size)) {
176 stg_unsafeThawArrayzh
178 // SUBTLETY TO DO WITH THE OLD GEN MUTABLE LIST
180 // A MUT_ARR_PTRS lives on the mutable list, but a MUT_ARR_PTRS_FROZEN
181 // normally doesn't. However, when we freeze a MUT_ARR_PTRS, we leave
182 // it on the mutable list for the GC to remove (removing something from
183 // the mutable list is not easy).
185 // So that we can tell whether a MUT_ARR_PTRS_FROZEN is on the mutable list,
186 // when we freeze it we set the info ptr to be MUT_ARR_PTRS_FROZEN0
187 // to indicate that it is still on the mutable list.
189 // So, when we thaw a MUT_ARR_PTRS_FROZEN, we must cope with two cases:
190 // either it is on a mut_list, or it isn't. We adopt the convention that
191 // the closure type is MUT_ARR_PTRS_FROZEN0 if it is on the mutable list,
192 // and MUT_ARR_PTRS_FROZEN otherwise. In fact it wouldn't matter if
193 // we put it on the mutable list more than once, but it would get scavenged
194 // multiple times during GC, which would be unnecessarily slow.
196 if (StgHeader_info(R1) != stg_MUT_ARR_PTRS_FROZEN0_info) {
197 SET_INFO(R1,stg_MUT_ARR_PTRS_DIRTY_info);
198 recordMutable(R1, R1);
199 // must be done after SET_INFO, because it ASSERTs closure_MUTABLE()
202 SET_INFO(R1,stg_MUT_ARR_PTRS_DIRTY_info);
207 /* -----------------------------------------------------------------------------
209 -------------------------------------------------------------------------- */
214 /* Args: R1 = initialisation value */
216 ALLOC_PRIM( SIZEOF_StgMutVar, R1_PTR, stg_newMutVarzh);
218 mv = Hp - SIZEOF_StgMutVar + WDS(1);
219 SET_HDR(mv,stg_MUT_VAR_DIRTY_info,W_[CCCS]);
220 StgMutVar_var(mv) = R1;
225 stg_atomicModifyMutVarzh
227 W_ mv, f, z, x, y, r, h;
228 /* Args: R1 :: MutVar#, R2 :: a -> (a,b) */
230 /* If x is the current contents of the MutVar#, then
231 We want to make the new contents point to
235 and the return value is
239 obviously we can share (f x).
241 z = [stg_ap_2 f x] (max (HS + 2) MIN_UPD_SIZE)
242 y = [stg_sel_0 z] (max (HS + 1) MIN_UPD_SIZE)
243 r = [stg_sel_1 z] (max (HS + 1) MIN_UPD_SIZE)
247 #define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
248 #define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),WDS(MIN_UPD_SIZE-1))
250 #define THUNK_1_SIZE (SIZEOF_StgThunkHeader + WDS(1))
251 #define TICK_ALLOC_THUNK_1() TICK_ALLOC_UP_THK(WDS(1),0)
255 #define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(MIN_UPD_SIZE))
256 #define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),WDS(MIN_UPD_SIZE-2))
258 #define THUNK_2_SIZE (SIZEOF_StgThunkHeader + WDS(2))
259 #define TICK_ALLOC_THUNK_2() TICK_ALLOC_UP_THK(WDS(2),0)
262 #define SIZE (THUNK_2_SIZE + THUNK_1_SIZE + THUNK_1_SIZE)
264 HP_CHK_GEN_TICKY(SIZE, R1_PTR & R2_PTR, stg_atomicModifyMutVarzh);
269 TICK_ALLOC_THUNK_2();
270 CCCS_ALLOC(THUNK_2_SIZE);
271 z = Hp - THUNK_2_SIZE + WDS(1);
272 SET_HDR(z, stg_ap_2_upd_info, W_[CCCS]);
273 LDV_RECORD_CREATE(z);
274 StgThunk_payload(z,0) = f;
276 TICK_ALLOC_THUNK_1();
277 CCCS_ALLOC(THUNK_1_SIZE);
278 y = z - THUNK_1_SIZE;
279 SET_HDR(y, stg_sel_0_upd_info, W_[CCCS]);
280 LDV_RECORD_CREATE(y);
281 StgThunk_payload(y,0) = z;
283 TICK_ALLOC_THUNK_1();
284 CCCS_ALLOC(THUNK_1_SIZE);
285 r = y - THUNK_1_SIZE;
286 SET_HDR(r, stg_sel_1_upd_info, W_[CCCS]);
287 LDV_RECORD_CREATE(r);
288 StgThunk_payload(r,0) = z;
291 x = StgMutVar_var(mv);
292 StgThunk_payload(z,1) = x;
294 (h) = foreign "C" cas(mv + SIZEOF_StgHeader + OFFSET_StgMutVar_var, x, y) [];
295 if (h != x) { goto retry; }
297 StgMutVar_var(mv) = y;
300 if (GET_INFO(mv) == stg_MUT_VAR_CLEAN_info) {
301 foreign "C" dirty_MUT_VAR(BaseReg "ptr", mv "ptr") [];
307 /* -----------------------------------------------------------------------------
308 Weak Pointer Primitives
309 -------------------------------------------------------------------------- */
311 STRING(stg_weak_msg,"New weak pointer at %p\n")
317 R3 = finalizer (or NULL)
322 R3 = stg_NO_FINALIZER_closure;
325 ALLOC_PRIM( SIZEOF_StgWeak, R1_PTR & R2_PTR & R3_PTR, stg_mkWeakzh );
327 w = Hp - SIZEOF_StgWeak + WDS(1);
328 SET_HDR(w, stg_WEAK_info, W_[CCCS]);
330 // We don't care about cfinalizer here.
331 // Should StgWeak_cfinalizer(w) be stg_NO_FINALIZER_closure or
335 StgWeak_value(w) = R2;
336 StgWeak_finalizer(w) = R3;
337 StgWeak_cfinalizer(w) = stg_NO_FINALIZER_closure;
339 ACQUIRE_LOCK(sm_mutex);
340 StgWeak_link(w) = W_[weak_ptr_list];
341 W_[weak_ptr_list] = w;
342 RELEASE_LOCK(sm_mutex);
344 IF_DEBUG(weak, foreign "C" debugBelch(stg_weak_msg,w) []);
349 stg_mkWeakForeignEnvzh
355 R5 = has environment (0 or 1)
358 W_ w, payload_words, words, p;
360 W_ key, val, fptr, ptr, flag, eptr;
369 ALLOC_PRIM( SIZEOF_StgWeak, R1_PTR & R2_PTR, stg_mkWeakForeignEnvzh );
371 w = Hp - SIZEOF_StgWeak + WDS(1);
372 SET_HDR(w, stg_WEAK_info, W_[CCCS]);
375 words = BYTES_TO_WDS(SIZEOF_StgArrWords) + payload_words;
376 ("ptr" p) = foreign "C" allocate(MyCapability() "ptr", words) [];
378 TICK_ALLOC_PRIM(SIZEOF_StgArrWords,WDS(payload_words),0);
379 SET_HDR(p, stg_ARR_WORDS_info, W_[CCCS]);
381 StgArrWords_words(p) = payload_words;
382 StgArrWords_payload(p,0) = fptr;
383 StgArrWords_payload(p,1) = ptr;
384 StgArrWords_payload(p,2) = eptr;
385 StgArrWords_payload(p,3) = flag;
387 // We don't care about the value here.
388 // Should StgWeak_value(w) be stg_NO_FINALIZER_closure or something else?
390 StgWeak_key(w) = key;
391 StgWeak_value(w) = val;
392 StgWeak_finalizer(w) = stg_NO_FINALIZER_closure;
393 StgWeak_cfinalizer(w) = p;
395 ACQUIRE_LOCK(sm_mutex);
396 StgWeak_link(w) = W_[weak_ptr_list];
397 W_[weak_ptr_list] = w;
398 RELEASE_LOCK(sm_mutex);
400 IF_DEBUG(weak, foreign "C" debugBelch(stg_weak_msg,w) []);
414 if (GET_INFO(w) == stg_DEAD_WEAK_info) {
415 RET_NP(0,stg_NO_FINALIZER_closure);
421 // A weak pointer is inherently used, so we do not need to call
422 // LDV_recordDead_FILL_SLOP_DYNAMIC():
423 // LDV_recordDead_FILL_SLOP_DYNAMIC((StgClosure *)w);
424 // or, LDV_recordDead():
425 // LDV_recordDead((StgClosure *)w, sizeofW(StgWeak) - sizeofW(StgProfHeader));
426 // Furthermore, when PROFILING is turned on, dead weak pointers are exactly as
427 // large as weak pointers, so there is no need to fill the slop, either.
428 // See stg_DEAD_WEAK_info in StgMiscClosures.hc.
432 // Todo: maybe use SET_HDR() and remove LDV_recordCreate()?
434 SET_INFO(w,stg_DEAD_WEAK_info);
435 LDV_RECORD_CREATE(w);
437 f = StgWeak_finalizer(w);
438 arr = StgWeak_cfinalizer(w);
440 StgDeadWeak_link(w) = StgWeak_link(w);
442 if (arr != stg_NO_FINALIZER_closure) {
443 foreign "C" runCFinalizer(StgArrWords_payload(arr,0),
444 StgArrWords_payload(arr,1),
445 StgArrWords_payload(arr,2),
446 StgArrWords_payload(arr,3)) [];
449 /* return the finalizer */
450 if (f == stg_NO_FINALIZER_closure) {
451 RET_NP(0,stg_NO_FINALIZER_closure);
463 if (GET_INFO(w) == stg_WEAK_info) {
465 val = StgWeak_value(w);
473 /* -----------------------------------------------------------------------------
474 Floating point operations.
475 -------------------------------------------------------------------------- */
477 stg_decodeFloatzuIntzh
484 STK_CHK_GEN( WDS(2), NO_PTRS, stg_decodeFloatzuIntzh );
486 mp_tmp1 = Sp - WDS(1);
487 mp_tmp_w = Sp - WDS(2);
489 /* arguments: F1 = Float# */
492 /* Perform the operation */
493 foreign "C" __decodeFloat_Int(mp_tmp1 "ptr", mp_tmp_w "ptr", arg) [];
495 /* returns: (Int# (mantissa), Int# (exponent)) */
496 RET_NN(W_[mp_tmp1], W_[mp_tmp_w]);
499 stg_decodeDoublezu2Intzh
508 STK_CHK_GEN( WDS(4), NO_PTRS, stg_decodeDoublezu2Intzh );
510 mp_tmp1 = Sp - WDS(1);
511 mp_tmp2 = Sp - WDS(2);
512 mp_result1 = Sp - WDS(3);
513 mp_result2 = Sp - WDS(4);
515 /* arguments: D1 = Double# */
518 /* Perform the operation */
519 foreign "C" __decodeDouble_2Int(mp_tmp1 "ptr", mp_tmp2 "ptr",
520 mp_result1 "ptr", mp_result2 "ptr",
524 (Int# (mant sign), Word# (mant high), Word# (mant low), Int# (expn)) */
525 RET_NNNN(W_[mp_tmp1], W_[mp_tmp2], W_[mp_result1], W_[mp_result2]);
528 /* -----------------------------------------------------------------------------
529 * Concurrency primitives
530 * -------------------------------------------------------------------------- */
534 /* args: R1 = closure to spark */
536 MAYBE_GC(R1_PTR, stg_forkzh);
542 ("ptr" threadid) = foreign "C" createIOThread( MyCapability() "ptr",
543 RtsFlags_GcFlags_initialStkSize(RtsFlags),
546 /* start blocked if the current thread is blocked */
547 StgTSO_flags(threadid) = %lobits16(
548 TO_W_(StgTSO_flags(threadid)) |
549 TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
551 foreign "C" scheduleThread(MyCapability() "ptr", threadid "ptr") [];
553 // context switch soon, but not immediately: we don't want every
554 // forkIO to force a context-switch.
555 Capability_context_switch(MyCapability()) = 1 :: CInt;
562 /* args: R1 = cpu, R2 = closure to spark */
564 MAYBE_GC(R2_PTR, stg_forkOnzh);
572 ("ptr" threadid) = foreign "C" createIOThread( MyCapability() "ptr",
573 RtsFlags_GcFlags_initialStkSize(RtsFlags),
576 /* start blocked if the current thread is blocked */
577 StgTSO_flags(threadid) = %lobits16(
578 TO_W_(StgTSO_flags(threadid)) |
579 TO_W_(StgTSO_flags(CurrentTSO)) & (TSO_BLOCKEX | TSO_INTERRUPTIBLE));
581 foreign "C" scheduleThreadOn(MyCapability() "ptr", cpu, threadid "ptr") [];
583 // context switch soon, but not immediately: we don't want every
584 // forkIO to force a context-switch.
585 Capability_context_switch(MyCapability()) = 1 :: CInt;
592 jump stg_yield_noregs;
607 foreign "C" labelThread(R1 "ptr", R2 "ptr") [];
609 jump %ENTRY_CODE(Sp(0));
612 stg_isCurrentThreadBoundzh
616 (r) = foreign "C" isThreadBound(CurrentTSO) [];
622 /* args: R1 :: ThreadId# */
630 if (TO_W_(StgTSO_what_next(tso)) == ThreadRelocated) {
631 tso = StgTSO__link(tso);
635 what_next = TO_W_(StgTSO_what_next(tso));
636 why_blocked = TO_W_(StgTSO_why_blocked(tso));
637 // Note: these two reads are not atomic, so they might end up
638 // being inconsistent. It doesn't matter, since we
639 // only return one or the other. If we wanted to return the
640 // contents of block_info too, then we'd have to do some synchronisation.
642 if (what_next == ThreadComplete) {
643 ret = 16; // NB. magic, matches up with GHC.Conc.threadStatus
645 if (what_next == ThreadKilled) {
654 /* -----------------------------------------------------------------------------
656 * -------------------------------------------------------------------------- */
660 // Catch retry frame ------------------------------------------------------------
662 INFO_TABLE_RET(stg_catch_retry_frame, CATCH_RETRY_FRAME,
663 #if defined(PROFILING)
664 W_ unused1, W_ unused2,
666 W_ unused3, P_ unused4, P_ unused5)
668 W_ r, frame, trec, outer;
671 trec = StgTSO_trec(CurrentTSO);
672 outer = StgTRecHeader_enclosing_trec(trec);
673 (r) = foreign "C" stmCommitNestedTransaction(MyCapability() "ptr", trec "ptr") [];
675 /* Succeeded (either first branch or second branch) */
676 StgTSO_trec(CurrentTSO) = outer;
677 Sp = Sp + SIZEOF_StgCatchRetryFrame;
678 jump %ENTRY_CODE(Sp(SP_OFF));
680 /* Did not commit: re-execute */
682 ("ptr" new_trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", outer "ptr") [];
683 StgTSO_trec(CurrentTSO) = new_trec;
684 if (StgCatchRetryFrame_running_alt_code(frame) != 0::I32) {
685 R1 = StgCatchRetryFrame_alt_code(frame);
687 R1 = StgCatchRetryFrame_first_code(frame);
694 // Atomically frame ------------------------------------------------------------
696 INFO_TABLE_RET(stg_atomically_frame, ATOMICALLY_FRAME,
697 #if defined(PROFILING)
698 W_ unused1, W_ unused2,
700 P_ code, P_ next_invariant_to_check, P_ result)
702 W_ frame, trec, valid, next_invariant, q, outer;
705 trec = StgTSO_trec(CurrentTSO);
707 outer = StgTRecHeader_enclosing_trec(trec);
709 if (outer == NO_TREC) {
710 /* First time back at the atomically frame -- pick up invariants */
711 ("ptr" q) = foreign "C" stmGetInvariantsToCheck(MyCapability() "ptr", trec "ptr") [];
712 StgAtomicallyFrame_next_invariant_to_check(frame) = q;
713 StgAtomicallyFrame_result(frame) = result;
716 /* Second/subsequent time back at the atomically frame -- abort the
717 * tx that's checking the invariant and move on to the next one */
718 StgTSO_trec(CurrentTSO) = outer;
719 q = StgAtomicallyFrame_next_invariant_to_check(frame);
720 StgInvariantCheckQueue_my_execution(q) = trec;
721 foreign "C" stmAbortTransaction(MyCapability() "ptr", trec "ptr") [];
722 /* Don't free trec -- it's linked from q and will be stashed in the
723 * invariant if we eventually commit. */
724 q = StgInvariantCheckQueue_next_queue_entry(q);
725 StgAtomicallyFrame_next_invariant_to_check(frame) = q;
729 q = StgAtomicallyFrame_next_invariant_to_check(frame);
731 if (q != END_INVARIANT_CHECK_QUEUE) {
732 /* We can't commit yet: another invariant to check */
733 ("ptr" trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", trec "ptr") [];
734 StgTSO_trec(CurrentTSO) = trec;
736 next_invariant = StgInvariantCheckQueue_invariant(q);
737 R1 = StgAtomicInvariant_code(next_invariant);
742 /* We've got no more invariants to check, try to commit */
743 (valid) = foreign "C" stmCommitTransaction(MyCapability() "ptr", trec "ptr") [];
745 /* Transaction was valid: commit succeeded */
746 StgTSO_trec(CurrentTSO) = NO_TREC;
747 R1 = StgAtomicallyFrame_result(frame);
748 Sp = Sp + SIZEOF_StgAtomicallyFrame;
749 jump %ENTRY_CODE(Sp(SP_OFF));
751 /* Transaction was not valid: try again */
752 ("ptr" trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", NO_TREC "ptr") [];
753 StgTSO_trec(CurrentTSO) = trec;
754 StgAtomicallyFrame_next_invariant_to_check(frame) = END_INVARIANT_CHECK_QUEUE;
755 R1 = StgAtomicallyFrame_code(frame);
761 INFO_TABLE_RET(stg_atomically_waiting_frame, ATOMICALLY_FRAME,
762 #if defined(PROFILING)
763 W_ unused1, W_ unused2,
765 P_ code, P_ next_invariant_to_check, P_ result)
767 W_ frame, trec, valid;
771 /* The TSO is currently waiting: should we stop waiting? */
772 (valid) = foreign "C" stmReWait(MyCapability() "ptr", CurrentTSO "ptr") [];
774 /* Previous attempt is still valid: no point trying again yet */
775 jump stg_block_noregs;
777 /* Previous attempt is no longer valid: try again */
778 ("ptr" trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", NO_TREC "ptr") [];
779 StgTSO_trec(CurrentTSO) = trec;
780 StgHeader_info(frame) = stg_atomically_frame_info;
781 R1 = StgAtomicallyFrame_code(frame);
786 // STM catch frame --------------------------------------------------------------
790 /* Catch frames are very similar to update frames, but when entering
791 * one we just pop the frame off the stack and perform the correct
792 * kind of return to the activation record underneath us on the stack.
795 INFO_TABLE_RET(stg_catch_stm_frame, CATCH_STM_FRAME,
796 #if defined(PROFILING)
797 W_ unused1, W_ unused2,
799 P_ unused3, P_ unused4)
801 W_ r, frame, trec, outer;
803 trec = StgTSO_trec(CurrentTSO);
804 outer = StgTRecHeader_enclosing_trec(trec);
805 (r) = foreign "C" stmCommitNestedTransaction(MyCapability() "ptr", trec "ptr") [];
807 /* Commit succeeded */
808 StgTSO_trec(CurrentTSO) = outer;
809 Sp = Sp + SIZEOF_StgCatchSTMFrame;
814 ("ptr" new_trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", outer "ptr") [];
815 StgTSO_trec(CurrentTSO) = new_trec;
816 R1 = StgCatchSTMFrame_code(frame);
822 // Primop definition ------------------------------------------------------------
830 // stmStartTransaction may allocate
831 MAYBE_GC (R1_PTR, stg_atomicallyzh);
833 /* Args: R1 = m :: STM a */
834 STK_CHK_GEN(SIZEOF_StgAtomicallyFrame + WDS(1), R1_PTR, stg_atomicallyzh);
836 old_trec = StgTSO_trec(CurrentTSO);
838 /* Nested transactions are not allowed; raise an exception */
839 if (old_trec != NO_TREC) {
840 R1 = base_ControlziExceptionziBase_nestedAtomically_closure;
844 /* Set up the atomically frame */
845 Sp = Sp - SIZEOF_StgAtomicallyFrame;
848 SET_HDR(frame,stg_atomically_frame_info, W_[CCCS]);
849 StgAtomicallyFrame_code(frame) = R1;
850 StgAtomicallyFrame_result(frame) = NO_TREC;
851 StgAtomicallyFrame_next_invariant_to_check(frame) = END_INVARIANT_CHECK_QUEUE;
853 /* Start the memory transcation */
854 ("ptr" new_trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", old_trec "ptr") [R1];
855 StgTSO_trec(CurrentTSO) = new_trec;
857 /* Apply R1 to the realworld token */
866 /* Args: R1 :: STM a */
867 /* Args: R2 :: Exception -> STM a */
868 STK_CHK_GEN(SIZEOF_StgCatchSTMFrame + WDS(1), R1_PTR & R2_PTR, stg_catchSTMzh);
870 /* Set up the catch frame */
871 Sp = Sp - SIZEOF_StgCatchSTMFrame;
874 SET_HDR(frame, stg_catch_stm_frame_info, W_[CCCS]);
875 StgCatchSTMFrame_handler(frame) = R2;
876 StgCatchSTMFrame_code(frame) = R1;
878 /* Start a nested transaction to run the body of the try block in */
881 cur_trec = StgTSO_trec(CurrentTSO);
882 ("ptr" new_trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", cur_trec "ptr");
883 StgTSO_trec(CurrentTSO) = new_trec;
885 /* Apply R1 to the realworld token */
896 // stmStartTransaction may allocate
897 MAYBE_GC (R1_PTR & R2_PTR, stg_catchRetryzh);
899 /* Args: R1 :: STM a */
900 /* Args: R2 :: STM a */
901 STK_CHK_GEN(SIZEOF_StgCatchRetryFrame + WDS(1), R1_PTR & R2_PTR, stg_catchRetryzh);
903 /* Start a nested transaction within which to run the first code */
904 trec = StgTSO_trec(CurrentTSO);
905 ("ptr" new_trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", trec "ptr") [R1,R2];
906 StgTSO_trec(CurrentTSO) = new_trec;
908 /* Set up the catch-retry frame */
909 Sp = Sp - SIZEOF_StgCatchRetryFrame;
912 SET_HDR(frame, stg_catch_retry_frame_info, W_[CCCS]);
913 StgCatchRetryFrame_running_alt_code(frame) = 0 :: CInt; // false;
914 StgCatchRetryFrame_first_code(frame) = R1;
915 StgCatchRetryFrame_alt_code(frame) = R2;
917 /* Apply R1 to the realworld token */
930 MAYBE_GC (NO_PTRS, stg_retryzh); // STM operations may allocate
932 // Find the enclosing ATOMICALLY_FRAME or CATCH_RETRY_FRAME
934 StgTSO_sp(CurrentTSO) = Sp;
935 (frame_type) = foreign "C" findRetryFrameHelper(CurrentTSO "ptr") [];
936 Sp = StgTSO_sp(CurrentTSO);
938 trec = StgTSO_trec(CurrentTSO);
939 outer = StgTRecHeader_enclosing_trec(trec);
941 if (frame_type == CATCH_RETRY_FRAME) {
942 // The retry reaches a CATCH_RETRY_FRAME before the atomic frame
943 ASSERT(outer != NO_TREC);
944 // Abort the transaction attempting the current branch
945 foreign "C" stmAbortTransaction(MyCapability() "ptr", trec "ptr") [];
946 foreign "C" stmFreeAbortedTRec(MyCapability() "ptr", trec "ptr") [];
947 if (!StgCatchRetryFrame_running_alt_code(frame) != 0::I32) {
948 // Retry in the first branch: try the alternative
949 ("ptr" trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", outer "ptr") [];
950 StgTSO_trec(CurrentTSO) = trec;
951 StgCatchRetryFrame_running_alt_code(frame) = 1 :: CInt; // true;
952 R1 = StgCatchRetryFrame_alt_code(frame);
955 // Retry in the alternative code: propagate the retry
956 StgTSO_trec(CurrentTSO) = outer;
957 Sp = Sp + SIZEOF_StgCatchRetryFrame;
958 goto retry_pop_stack;
962 // We've reached the ATOMICALLY_FRAME: attempt to wait
963 ASSERT(frame_type == ATOMICALLY_FRAME);
964 if (outer != NO_TREC) {
965 // We called retry while checking invariants, so abort the current
966 // invariant check (merging its TVar accesses into the parents read
967 // set so we'll wait on them)
968 foreign "C" stmAbortTransaction(MyCapability() "ptr", trec "ptr") [];
969 foreign "C" stmFreeAbortedTRec(MyCapability() "ptr", trec "ptr") [];
971 StgTSO_trec(CurrentTSO) = trec;
972 outer = StgTRecHeader_enclosing_trec(trec);
974 ASSERT(outer == NO_TREC);
976 (r) = foreign "C" stmWait(MyCapability() "ptr", CurrentTSO "ptr", trec "ptr") [];
978 // Transaction was valid: stmWait put us on the TVars' queues, we now block
979 StgHeader_info(frame) = stg_atomically_waiting_frame_info;
981 // Fix up the stack in the unregisterised case: the return convention is different.
982 R3 = trec; // passing to stmWaitUnblock()
983 jump stg_block_stmwait;
985 // Transaction was not valid: retry immediately
986 ("ptr" trec) = foreign "C" stmStartTransaction(MyCapability() "ptr", outer "ptr") [];
987 StgTSO_trec(CurrentTSO) = trec;
988 R1 = StgAtomicallyFrame_code(frame);
999 /* Args: R1 = invariant closure */
1000 MAYBE_GC (R1_PTR, stg_checkzh);
1002 trec = StgTSO_trec(CurrentTSO);
1004 foreign "C" stmAddInvariantToCheck(MyCapability() "ptr",
1008 jump %ENTRY_CODE(Sp(0));
1017 /* Args: R1 = initialisation value */
1019 MAYBE_GC (R1_PTR, stg_newTVarzh);
1021 ("ptr" tv) = foreign "C" stmNewTVar(MyCapability() "ptr", new_value "ptr") [];
1032 /* Args: R1 = TVar closure */
1034 MAYBE_GC (R1_PTR, stg_readTVarzh); // Call to stmReadTVar may allocate
1035 trec = StgTSO_trec(CurrentTSO);
1037 ("ptr" result) = foreign "C" stmReadTVar(MyCapability() "ptr", trec "ptr", tvar "ptr") [];
1047 result = StgTVar_current_value(R1);
1048 if (%INFO_PTR(result) == stg_TREC_HEADER_info) {
1060 /* Args: R1 = TVar closure */
1061 /* R2 = New value */
1063 MAYBE_GC (R1_PTR & R2_PTR, stg_writeTVarzh); // Call to stmWriteTVar may allocate
1064 trec = StgTSO_trec(CurrentTSO);
1067 foreign "C" stmWriteTVar(MyCapability() "ptr", trec "ptr", tvar "ptr", new_value "ptr") [];
1069 jump %ENTRY_CODE(Sp(0));
1073 /* -----------------------------------------------------------------------------
1076 * take & putMVar work as follows. Firstly, an important invariant:
1078 * If the MVar is full, then the blocking queue contains only
1079 * threads blocked on putMVar, and if the MVar is empty then the
1080 * blocking queue contains only threads blocked on takeMVar.
1083 * MVar empty : then add ourselves to the blocking queue
1084 * MVar full : remove the value from the MVar, and
1085 * blocking queue empty : return
1086 * blocking queue non-empty : perform the first blocked putMVar
1087 * from the queue, and wake up the
1088 * thread (MVar is now full again)
1090 * putMVar is just the dual of the above algorithm.
1092 * How do we "perform a putMVar"? Well, we have to fiddle around with
1093 * the stack of the thread waiting to do the putMVar. See
1094 * stg_block_putmvar and stg_block_takemvar in HeapStackCheck.c for
1095 * the stack layout, and the PerformPut and PerformTake macros below.
1097 * It is important that a blocked take or put is woken up with the
1098 * take/put already performed, because otherwise there would be a
1099 * small window of vulnerability where the thread could receive an
1100 * exception and never perform its take or put, and we'd end up with a
1103 * -------------------------------------------------------------------------- */
1107 /* args: R1 = MVar closure */
1109 if (StgMVar_value(R1) == stg_END_TSO_QUEUE_closure) {
1121 ALLOC_PRIM ( SIZEOF_StgMVar, NO_PTRS, stg_newMVarzh );
1123 mvar = Hp - SIZEOF_StgMVar + WDS(1);
1124 SET_HDR(mvar,stg_MVAR_DIRTY_info,W_[CCCS]);
1125 // MVARs start dirty: generation 0 has no mutable list
1126 StgMVar_head(mvar) = stg_END_TSO_QUEUE_closure;
1127 StgMVar_tail(mvar) = stg_END_TSO_QUEUE_closure;
1128 StgMVar_value(mvar) = stg_END_TSO_QUEUE_closure;
1133 #define PerformTake(tso, value) \
1134 W_[StgTSO_sp(tso) + WDS(1)] = value; \
1135 W_[StgTSO_sp(tso) + WDS(0)] = stg_gc_unpt_r1_info;
1137 #define PerformPut(tso,lval) \
1138 StgTSO_sp(tso) = StgTSO_sp(tso) + WDS(3); \
1139 lval = W_[StgTSO_sp(tso) - WDS(1)];
1143 W_ mvar, val, info, tso, q;
1145 /* args: R1 = MVar closure */
1148 #if defined(THREADED_RTS)
1149 ("ptr" info) = foreign "C" lockClosure(mvar "ptr") [];
1151 info = GET_INFO(mvar);
1154 if (info == stg_MVAR_CLEAN_info) {
1155 foreign "C" dirty_MVAR(BaseReg "ptr", mvar "ptr") [];
1158 /* If the MVar is empty, put ourselves on its blocking queue,
1159 * and wait until we're woken up.
1161 if (StgMVar_value(mvar) == stg_END_TSO_QUEUE_closure) {
1163 // Note [mvar-heap-check] We want to do the heap check in the
1164 // branch here, to avoid the conditional in the common case.
1165 // However, we've already locked the MVar above, so we better
1166 // be careful to unlock it again if the the heap check fails.
1167 // Unfortunately we don't have an easy way to inject any code
1168 // into the heap check generated by the code generator, so we
1169 // have to do it in stg_gc_gen (see HeapStackCheck.cmm).
1170 HP_CHK_GEN_TICKY(SIZEOF_StgMVarTSOQueue, R1_PTR, stg_takeMVarzh);
1172 q = Hp - SIZEOF_StgMVarTSOQueue + WDS(1);
1174 SET_HDR(q, stg_MVAR_TSO_QUEUE_info, CCS_SYSTEM);
1175 StgMVarTSOQueue_link(q) = END_TSO_QUEUE;
1176 StgMVarTSOQueue_tso(q) = CurrentTSO;
1178 if (StgMVar_head(mvar) == stg_END_TSO_QUEUE_closure) {
1179 StgMVar_head(mvar) = q;
1181 StgMVarTSOQueue_link(StgMVar_tail(mvar)) = q;
1182 foreign "C" recordClosureMutated(MyCapability() "ptr",
1183 StgMVar_tail(mvar)) [];
1185 StgTSO__link(CurrentTSO) = q;
1186 StgTSO_block_info(CurrentTSO) = mvar;
1187 StgTSO_why_blocked(CurrentTSO) = BlockedOnMVar::I16;
1188 StgMVar_tail(mvar) = q;
1191 jump stg_block_takemvar;
1194 /* we got the value... */
1195 val = StgMVar_value(mvar);
1197 q = StgMVar_head(mvar);
1199 if (q == stg_END_TSO_QUEUE_closure) {
1200 /* No further putMVars, MVar is now empty */
1201 StgMVar_value(mvar) = stg_END_TSO_QUEUE_closure;
1202 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1205 if (StgHeader_info(q) == stg_IND_info ||
1206 StgHeader_info(q) == stg_MSG_NULL_info) {
1207 q = StgInd_indirectee(q);
1211 // There are putMVar(s) waiting... wake up the first thread on the queue
1213 tso = StgMVarTSOQueue_tso(q);
1214 StgMVar_head(mvar) = StgMVarTSOQueue_link(q);
1215 if (StgMVar_head(mvar) == stg_END_TSO_QUEUE_closure) {
1216 StgMVar_tail(mvar) = stg_END_TSO_QUEUE_closure;
1220 if (TO_W_(StgTSO_what_next(tso)) == ThreadRelocated) {
1221 tso = StgTSO__link(tso);
1225 ASSERT(StgTSO_why_blocked(tso) == BlockedOnMVar::I16);
1226 ASSERT(StgTSO_block_info(tso) == mvar);
1228 // actually perform the putMVar for the thread that we just woke up
1229 PerformPut(tso,StgMVar_value(mvar));
1231 // indicate that the MVar operation has now completed.
1232 StgTSO__link(tso) = stg_END_TSO_QUEUE_closure;
1234 // no need to mark the TSO dirty, we have only written END_TSO_QUEUE.
1236 foreign "C" tryWakeupThread_(MyCapability() "ptr", tso) [];
1238 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1245 W_ mvar, val, info, tso, q;
1247 /* args: R1 = MVar closure */
1250 #if defined(THREADED_RTS)
1251 ("ptr" info) = foreign "C" lockClosure(mvar "ptr") [];
1253 info = GET_INFO(mvar);
1256 /* If the MVar is empty, put ourselves on its blocking queue,
1257 * and wait until we're woken up.
1259 if (StgMVar_value(mvar) == stg_END_TSO_QUEUE_closure) {
1260 #if defined(THREADED_RTS)
1261 unlockClosure(mvar, info);
1263 /* HACK: we need a pointer to pass back,
1264 * so we abuse NO_FINALIZER_closure
1266 RET_NP(0, stg_NO_FINALIZER_closure);
1269 if (info == stg_MVAR_CLEAN_info) {
1270 foreign "C" dirty_MVAR(BaseReg "ptr", mvar "ptr") [];
1273 /* we got the value... */
1274 val = StgMVar_value(mvar);
1276 q = StgMVar_head(mvar);
1278 if (q == stg_END_TSO_QUEUE_closure) {
1279 /* No further putMVars, MVar is now empty */
1280 StgMVar_value(mvar) = stg_END_TSO_QUEUE_closure;
1281 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1284 if (StgHeader_info(q) == stg_IND_info ||
1285 StgHeader_info(q) == stg_MSG_NULL_info) {
1286 q = StgInd_indirectee(q);
1290 // There are putMVar(s) waiting... wake up the first thread on the queue
1292 tso = StgMVarTSOQueue_tso(q);
1293 StgMVar_head(mvar) = StgMVarTSOQueue_link(q);
1294 if (StgMVar_head(mvar) == stg_END_TSO_QUEUE_closure) {
1295 StgMVar_tail(mvar) = stg_END_TSO_QUEUE_closure;
1299 if (TO_W_(StgTSO_what_next(tso)) == ThreadRelocated) {
1300 tso = StgTSO__link(tso);
1304 ASSERT(StgTSO_why_blocked(tso) == BlockedOnMVar::I16);
1305 ASSERT(StgTSO_block_info(tso) == mvar);
1307 // actually perform the putMVar for the thread that we just woke up
1308 PerformPut(tso,StgMVar_value(mvar));
1310 // indicate that the MVar operation has now completed.
1311 StgTSO__link(tso) = stg_END_TSO_QUEUE_closure;
1313 // no need to mark the TSO dirty, we have only written END_TSO_QUEUE.
1315 foreign "C" tryWakeupThread_(MyCapability() "ptr", tso) [];
1317 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1324 W_ mvar, val, info, tso, q;
1326 /* args: R1 = MVar, R2 = value */
1330 #if defined(THREADED_RTS)
1331 ("ptr" info) = foreign "C" lockClosure(mvar "ptr") [];
1333 info = GET_INFO(mvar);
1336 if (info == stg_MVAR_CLEAN_info) {
1337 foreign "C" dirty_MVAR(BaseReg "ptr", mvar "ptr");
1340 if (StgMVar_value(mvar) != stg_END_TSO_QUEUE_closure) {
1342 // see Note [mvar-heap-check] above
1343 HP_CHK_GEN_TICKY(SIZEOF_StgMVarTSOQueue, R1_PTR & R2_PTR, stg_putMVarzh);
1345 q = Hp - SIZEOF_StgMVarTSOQueue + WDS(1);
1347 SET_HDR(q, stg_MVAR_TSO_QUEUE_info, CCS_SYSTEM);
1348 StgMVarTSOQueue_link(q) = END_TSO_QUEUE;
1349 StgMVarTSOQueue_tso(q) = CurrentTSO;
1351 if (StgMVar_head(mvar) == stg_END_TSO_QUEUE_closure) {
1352 StgMVar_head(mvar) = q;
1354 StgMVarTSOQueue_link(StgMVar_tail(mvar)) = q;
1355 foreign "C" recordClosureMutated(MyCapability() "ptr",
1356 StgMVar_tail(mvar)) [];
1358 StgTSO__link(CurrentTSO) = q;
1359 StgTSO_block_info(CurrentTSO) = mvar;
1360 StgTSO_why_blocked(CurrentTSO) = BlockedOnMVar::I16;
1361 StgMVar_tail(mvar) = q;
1365 jump stg_block_putmvar;
1368 q = StgMVar_head(mvar);
1370 if (q == stg_END_TSO_QUEUE_closure) {
1371 /* No further takes, the MVar is now full. */
1372 StgMVar_value(mvar) = val;
1373 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1374 jump %ENTRY_CODE(Sp(0));
1376 if (StgHeader_info(q) == stg_IND_info ||
1377 StgHeader_info(q) == stg_MSG_NULL_info) {
1378 q = StgInd_indirectee(q);
1382 // There are takeMVar(s) waiting: wake up the first one
1384 tso = StgMVarTSOQueue_tso(q);
1385 StgMVar_head(mvar) = StgMVarTSOQueue_link(q);
1386 if (StgMVar_head(mvar) == stg_END_TSO_QUEUE_closure) {
1387 StgMVar_tail(mvar) = stg_END_TSO_QUEUE_closure;
1391 if (TO_W_(StgTSO_what_next(tso)) == ThreadRelocated) {
1392 tso = StgTSO__link(tso);
1396 ASSERT(StgTSO_why_blocked(tso) == BlockedOnMVar::I16);
1397 ASSERT(StgTSO_block_info(tso) == mvar);
1399 // actually perform the takeMVar
1400 PerformTake(tso, val);
1402 // indicate that the MVar operation has now completed.
1403 StgTSO__link(tso) = stg_END_TSO_QUEUE_closure;
1405 if (TO_W_(StgTSO_dirty(tso)) == 0) {
1406 foreign "C" dirty_TSO(MyCapability() "ptr", tso "ptr") [];
1409 foreign "C" tryWakeupThread_(MyCapability() "ptr", tso) [];
1411 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1412 jump %ENTRY_CODE(Sp(0));
1418 W_ mvar, val, info, tso, q;
1420 /* args: R1 = MVar, R2 = value */
1424 #if defined(THREADED_RTS)
1425 ("ptr" info) = foreign "C" lockClosure(mvar "ptr") [];
1427 info = GET_INFO(mvar);
1430 if (info == stg_MVAR_CLEAN_info) {
1431 foreign "C" dirty_MVAR(BaseReg "ptr", mvar "ptr");
1434 if (StgMVar_value(mvar) != stg_END_TSO_QUEUE_closure) {
1435 #if defined(THREADED_RTS)
1436 unlockClosure(mvar, info);
1441 q = StgMVar_head(mvar);
1443 if (q == stg_END_TSO_QUEUE_closure) {
1444 /* No further takes, the MVar is now full. */
1445 StgMVar_value(mvar) = val;
1446 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1447 jump %ENTRY_CODE(Sp(0));
1449 if (StgHeader_info(q) == stg_IND_info ||
1450 StgHeader_info(q) == stg_MSG_NULL_info) {
1451 q = StgInd_indirectee(q);
1455 // There are takeMVar(s) waiting: wake up the first one
1457 tso = StgMVarTSOQueue_tso(q);
1458 StgMVar_head(mvar) = StgMVarTSOQueue_link(q);
1459 if (StgMVar_head(mvar) == stg_END_TSO_QUEUE_closure) {
1460 StgMVar_tail(mvar) = stg_END_TSO_QUEUE_closure;
1464 if (TO_W_(StgTSO_what_next(tso)) == ThreadRelocated) {
1465 tso = StgTSO__link(tso);
1469 ASSERT(StgTSO_why_blocked(tso) == BlockedOnMVar::I16);
1470 ASSERT(StgTSO_block_info(tso) == mvar);
1472 // actually perform the takeMVar
1473 PerformTake(tso, val);
1475 // indicate that the MVar operation has now completed.
1476 StgTSO__link(tso) = stg_END_TSO_QUEUE_closure;
1478 if (TO_W_(StgTSO_dirty(tso)) == 0) {
1479 foreign "C" dirty_TSO(MyCapability() "ptr", tso "ptr") [];
1482 foreign "C" tryWakeupThread_(MyCapability() "ptr", tso) [];
1484 unlockClosure(mvar, stg_MVAR_DIRTY_info);
1485 jump %ENTRY_CODE(Sp(0));
1489 /* -----------------------------------------------------------------------------
1490 Stable pointer primitives
1491 ------------------------------------------------------------------------- */
1493 stg_makeStableNamezh
1497 ALLOC_PRIM( SIZEOF_StgStableName, R1_PTR, stg_makeStableNamezh );
1499 (index) = foreign "C" lookupStableName(R1 "ptr") [];
1501 /* Is there already a StableName for this heap object?
1502 * stable_ptr_table is a pointer to an array of snEntry structs.
1504 if ( snEntry_sn_obj(W_[stable_ptr_table] + index*SIZEOF_snEntry) == NULL ) {
1505 sn_obj = Hp - SIZEOF_StgStableName + WDS(1);
1506 SET_HDR(sn_obj, stg_STABLE_NAME_info, W_[CCCS]);
1507 StgStableName_sn(sn_obj) = index;
1508 snEntry_sn_obj(W_[stable_ptr_table] + index*SIZEOF_snEntry) = sn_obj;
1510 sn_obj = snEntry_sn_obj(W_[stable_ptr_table] + index*SIZEOF_snEntry);
1521 MAYBE_GC(R1_PTR, stg_makeStablePtrzh);
1522 ("ptr" sp) = foreign "C" getStablePtr(R1 "ptr") [];
1526 stg_deRefStablePtrzh
1528 /* Args: R1 = the stable ptr */
1531 r = snEntry_addr(W_[stable_ptr_table] + sp*SIZEOF_snEntry);
1535 /* -----------------------------------------------------------------------------
1536 Bytecode object primitives
1537 ------------------------------------------------------------------------- */
1547 W_ bco, bitmap_arr, bytes, words;
1551 words = BYTES_TO_WDS(SIZEOF_StgBCO) + StgArrWords_words(bitmap_arr);
1554 ALLOC_PRIM( bytes, R1_PTR&R2_PTR&R3_PTR&R5_PTR, stg_newBCOzh );
1556 bco = Hp - bytes + WDS(1);
1557 SET_HDR(bco, stg_BCO_info, W_[CCCS]);
1559 StgBCO_instrs(bco) = R1;
1560 StgBCO_literals(bco) = R2;
1561 StgBCO_ptrs(bco) = R3;
1562 StgBCO_arity(bco) = HALF_W_(R4);
1563 StgBCO_size(bco) = HALF_W_(words);
1565 // Copy the arity/bitmap info into the BCO
1569 if (i < StgArrWords_words(bitmap_arr)) {
1570 StgBCO_bitmap(bco,i) = StgArrWords_payload(bitmap_arr,i);
1581 // R1 = the BCO# for the AP
1585 // This function is *only* used to wrap zero-arity BCOs in an
1586 // updatable wrapper (see ByteCodeLink.lhs). An AP thunk is always
1587 // saturated and always points directly to a FUN or BCO.
1588 ASSERT(%INFO_TYPE(%GET_STD_INFO(R1)) == HALF_W_(BCO) &&
1589 StgBCO_arity(R1) == HALF_W_(0));
1591 HP_CHK_GEN_TICKY(SIZEOF_StgAP, R1_PTR, stg_mkApUpd0zh);
1592 TICK_ALLOC_UP_THK(0, 0);
1593 CCCS_ALLOC(SIZEOF_StgAP);
1595 ap = Hp - SIZEOF_StgAP + WDS(1);
1596 SET_HDR(ap, stg_AP_info, W_[CCCS]);
1598 StgAP_n_args(ap) = HALF_W_(0);
1606 /* args: R1 = closure to analyze */
1607 // TODO: Consider the absence of ptrs or nonptrs as a special case ?
1609 W_ info, ptrs, nptrs, p, ptrs_arr, nptrs_arr;
1610 info = %GET_STD_INFO(UNTAG(R1));
1612 // Some closures have non-standard layout, so we omit those here.
1614 type = TO_W_(%INFO_TYPE(info));
1615 switch [0 .. N_CLOSURE_TYPES] type {
1616 case THUNK_SELECTOR : {
1621 case THUNK, THUNK_1_0, THUNK_0_1, THUNK_2_0, THUNK_1_1,
1622 THUNK_0_2, THUNK_STATIC, AP, PAP, AP_STACK, BCO : {
1628 ptrs = TO_W_(%INFO_PTRS(info));
1629 nptrs = TO_W_(%INFO_NPTRS(info));
1634 W_ ptrs_arr_sz, ptrs_arr_cards, nptrs_arr_sz;
1635 nptrs_arr_sz = SIZEOF_StgArrWords + WDS(nptrs);
1636 ptrs_arr_cards = mutArrPtrsCardWords(ptrs);
1637 ptrs_arr_sz = SIZEOF_StgMutArrPtrs + WDS(ptrs) + WDS(ptrs_arr_cards);
1639 ALLOC_PRIM (ptrs_arr_sz + nptrs_arr_sz, R1_PTR, stg_unpackClosurezh);
1644 ptrs_arr = Hp - nptrs_arr_sz - ptrs_arr_sz + WDS(1);
1645 nptrs_arr = Hp - nptrs_arr_sz + WDS(1);
1647 SET_HDR(ptrs_arr, stg_MUT_ARR_PTRS_FROZEN_info, W_[CCCS]);
1648 StgMutArrPtrs_ptrs(ptrs_arr) = ptrs;
1649 StgMutArrPtrs_size(ptrs_arr) = ptrs + ptrs_arr_cards;
1654 W_[ptrs_arr + SIZEOF_StgMutArrPtrs + WDS(p)] = StgClosure_payload(clos,p);
1658 /* We can leave the card table uninitialised, since the array is
1659 allocated in the nursery. The GC will fill it in if/when the array
1662 SET_HDR(nptrs_arr, stg_ARR_WORDS_info, W_[CCCS]);
1663 StgArrWords_words(nptrs_arr) = nptrs;
1667 W_[BYTE_ARR_CTS(nptrs_arr) + WDS(p)] = StgClosure_payload(clos, p+ptrs);
1671 RET_NPP(info, ptrs_arr, nptrs_arr);
1674 /* -----------------------------------------------------------------------------
1675 Thread I/O blocking primitives
1676 -------------------------------------------------------------------------- */
1678 /* Add a thread to the end of the blocked queue. (C-- version of the C
1679 * macro in Schedule.h).
1681 #define APPEND_TO_BLOCKED_QUEUE(tso) \
1682 ASSERT(StgTSO__link(tso) == END_TSO_QUEUE); \
1683 if (W_[blocked_queue_hd] == END_TSO_QUEUE) { \
1684 W_[blocked_queue_hd] = tso; \
1686 foreign "C" setTSOLink(MyCapability() "ptr", W_[blocked_queue_tl] "ptr", tso) []; \
1688 W_[blocked_queue_tl] = tso;
1694 foreign "C" barf("waitRead# on threaded RTS") never returns;
1697 ASSERT(StgTSO_why_blocked(CurrentTSO) == NotBlocked::I16);
1698 StgTSO_why_blocked(CurrentTSO) = BlockedOnRead::I16;
1699 StgTSO_block_info(CurrentTSO) = R1;
1700 // No locking - we're not going to use this interface in the
1701 // threaded RTS anyway.
1702 APPEND_TO_BLOCKED_QUEUE(CurrentTSO);
1703 jump stg_block_noregs;
1711 foreign "C" barf("waitWrite# on threaded RTS") never returns;
1714 ASSERT(StgTSO_why_blocked(CurrentTSO) == NotBlocked::I16);
1715 StgTSO_why_blocked(CurrentTSO) = BlockedOnWrite::I16;
1716 StgTSO_block_info(CurrentTSO) = R1;
1717 // No locking - we're not going to use this interface in the
1718 // threaded RTS anyway.
1719 APPEND_TO_BLOCKED_QUEUE(CurrentTSO);
1720 jump stg_block_noregs;
1725 STRING(stg_delayzh_malloc_str, "stg_delayzh")
1728 #ifdef mingw32_HOST_OS
1736 foreign "C" barf("delay# on threaded RTS") never returns;
1739 /* args: R1 (microsecond delay amount) */
1740 ASSERT(StgTSO_why_blocked(CurrentTSO) == NotBlocked::I16);
1741 StgTSO_why_blocked(CurrentTSO) = BlockedOnDelay::I16;
1743 #ifdef mingw32_HOST_OS
1745 /* could probably allocate this on the heap instead */
1746 ("ptr" ares) = foreign "C" stgMallocBytes(SIZEOF_StgAsyncIOResult,
1747 stg_delayzh_malloc_str);
1748 (reqID) = foreign "C" addDelayRequest(R1);
1749 StgAsyncIOResult_reqID(ares) = reqID;
1750 StgAsyncIOResult_len(ares) = 0;
1751 StgAsyncIOResult_errCode(ares) = 0;
1752 StgTSO_block_info(CurrentTSO) = ares;
1754 /* Having all async-blocked threads reside on the blocked_queue
1755 * simplifies matters, so change the status to OnDoProc put the
1756 * delayed thread on the blocked_queue.
1758 StgTSO_why_blocked(CurrentTSO) = BlockedOnDoProc::I16;
1759 APPEND_TO_BLOCKED_QUEUE(CurrentTSO);
1760 jump stg_block_async_void;
1766 (time) = foreign "C" getourtimeofday() [R1];
1767 divisor = TO_W_(RtsFlags_MiscFlags_tickInterval(RtsFlags));
1771 divisor = divisor * 1000;
1772 target = ((R1 + divisor - 1) / divisor) /* divide rounding up */
1773 + time + 1; /* Add 1 as getourtimeofday rounds down */
1774 StgTSO_block_info(CurrentTSO) = target;
1776 /* Insert the new thread in the sleeping queue. */
1778 t = W_[sleeping_queue];
1780 if (t != END_TSO_QUEUE && StgTSO_block_info(t) < target) {
1782 t = StgTSO__link(t);
1786 StgTSO__link(CurrentTSO) = t;
1788 W_[sleeping_queue] = CurrentTSO;
1790 foreign "C" setTSOLink(MyCapability() "ptr", prev "ptr", CurrentTSO) [];
1792 jump stg_block_noregs;
1794 #endif /* !THREADED_RTS */
1798 #ifdef mingw32_HOST_OS
1799 STRING(stg_asyncReadzh_malloc_str, "stg_asyncReadzh")
1806 foreign "C" barf("asyncRead# on threaded RTS") never returns;
1809 /* args: R1 = fd, R2 = isSock, R3 = len, R4 = buf */
1810 ASSERT(StgTSO_why_blocked(CurrentTSO) == NotBlocked::I16);
1811 StgTSO_why_blocked(CurrentTSO) = BlockedOnRead::I16;
1813 /* could probably allocate this on the heap instead */
1814 ("ptr" ares) = foreign "C" stgMallocBytes(SIZEOF_StgAsyncIOResult,
1815 stg_asyncReadzh_malloc_str)
1817 (reqID) = foreign "C" addIORequest(R1, 0/*FALSE*/,R2,R3,R4 "ptr") [];
1818 StgAsyncIOResult_reqID(ares) = reqID;
1819 StgAsyncIOResult_len(ares) = 0;
1820 StgAsyncIOResult_errCode(ares) = 0;
1821 StgTSO_block_info(CurrentTSO) = ares;
1822 APPEND_TO_BLOCKED_QUEUE(CurrentTSO);
1823 jump stg_block_async;
1827 STRING(stg_asyncWritezh_malloc_str, "stg_asyncWritezh")
1834 foreign "C" barf("asyncWrite# on threaded RTS") never returns;
1837 /* args: R1 = fd, R2 = isSock, R3 = len, R4 = buf */
1838 ASSERT(StgTSO_why_blocked(CurrentTSO) == NotBlocked::I16);
1839 StgTSO_why_blocked(CurrentTSO) = BlockedOnWrite::I16;
1841 ("ptr" ares) = foreign "C" stgMallocBytes(SIZEOF_StgAsyncIOResult,
1842 stg_asyncWritezh_malloc_str)
1844 (reqID) = foreign "C" addIORequest(R1, 1/*TRUE*/,R2,R3,R4 "ptr") [];
1846 StgAsyncIOResult_reqID(ares) = reqID;
1847 StgAsyncIOResult_len(ares) = 0;
1848 StgAsyncIOResult_errCode(ares) = 0;
1849 StgTSO_block_info(CurrentTSO) = ares;
1850 APPEND_TO_BLOCKED_QUEUE(CurrentTSO);
1851 jump stg_block_async;
1855 STRING(stg_asyncDoProczh_malloc_str, "stg_asyncDoProczh")
1862 foreign "C" barf("asyncDoProc# on threaded RTS") never returns;
1865 /* args: R1 = proc, R2 = param */
1866 ASSERT(StgTSO_why_blocked(CurrentTSO) == NotBlocked::I16);
1867 StgTSO_why_blocked(CurrentTSO) = BlockedOnDoProc::I16;
1869 /* could probably allocate this on the heap instead */
1870 ("ptr" ares) = foreign "C" stgMallocBytes(SIZEOF_StgAsyncIOResult,
1871 stg_asyncDoProczh_malloc_str)
1873 (reqID) = foreign "C" addDoProcRequest(R1 "ptr",R2 "ptr") [];
1874 StgAsyncIOResult_reqID(ares) = reqID;
1875 StgAsyncIOResult_len(ares) = 0;
1876 StgAsyncIOResult_errCode(ares) = 0;
1877 StgTSO_block_info(CurrentTSO) = ares;
1878 APPEND_TO_BLOCKED_QUEUE(CurrentTSO);
1879 jump stg_block_async;
1884 /* -----------------------------------------------------------------------------
1887 * noDuplicate# tries to ensure that none of the thunks under
1888 * evaluation by the current thread are also under evaluation by
1889 * another thread. It relies on *both* threads doing noDuplicate#;
1890 * the second one will get blocked if they are duplicating some work.
1892 * The idea is that noDuplicate# is used within unsafePerformIO to
1893 * ensure that the IO operation is performed at most once.
1894 * noDuplicate# calls threadPaused which acquires an exclusive lock on
1895 * all the thunks currently under evaluation by the current thread.
1897 * Consider the following scenario. There is a thunk A, whose
1898 * evaluation requires evaluating thunk B, where thunk B is an
1899 * unsafePerformIO. Two threads, 1 and 2, bother enter A. Thread 2
1900 * is pre-empted before it enters B, and claims A by blackholing it
1901 * (in threadPaused). Thread 1 now enters B, and calls noDuplicate#.
1904 * +-----------+ +---------------+
1905 * | -------+-----> A <-------+------- |
1906 * | update | BLACKHOLE | marked_update |
1907 * +-----------+ +---------------+
1910 * | | +---------------+
1913 * | update | BLACKHOLE
1916 * At this point: A is a blackhole, owned by thread 2. noDuplicate#
1917 * calls threadPaused, which walks up the stack and
1918 * - claims B on behalf of thread 1
1919 * - then it reaches the update frame for A, which it sees is already
1920 * a BLACKHOLE and is therefore owned by another thread. Since
1921 * thread 1 is duplicating work, the computation up to the update
1922 * frame for A is suspended, including thunk B.
1923 * - thunk B, which is an unsafePerformIO, has now been reverted to
1924 * an AP_STACK which could be duplicated - BAD!
1925 * - The solution is as follows: before calling threadPaused, we
1926 * leave a frame on the stack (stg_noDuplicate_info) that will call
1927 * noDuplicate# again if the current computation is suspended and
1930 * See the test program in concurrent/prog003 for a way to demonstrate
1931 * this. It needs to be run with +RTS -N3 or greater, and the bug
1932 * only manifests occasionally (once very 10 runs or so).
1933 * -------------------------------------------------------------------------- */
1935 INFO_TABLE_RET(stg_noDuplicate, RET_SMALL)
1938 jump stg_noDuplicatezh;
1943 STK_CHK_GEN( WDS(1), NO_PTRS, stg_noDuplicatezh );
1944 // leave noDuplicate frame in case the current
1945 // computation is suspended and restarted (see above).
1947 Sp(0) = stg_noDuplicate_info;
1949 SAVE_THREAD_STATE();
1950 ASSERT(StgTSO_what_next(CurrentTSO) == ThreadRunGHC::I16);
1951 foreign "C" threadPaused (MyCapability() "ptr", CurrentTSO "ptr") [];
1953 if (StgTSO_what_next(CurrentTSO) == ThreadKilled::I16) {
1954 jump stg_threadFinished;
1956 LOAD_THREAD_STATE();
1957 ASSERT(StgTSO_what_next(CurrentTSO) == ThreadRunGHC::I16);
1958 // remove the stg_noDuplicate frame if it is still there.
1959 if (Sp(0) == stg_noDuplicate_info) {
1962 jump %ENTRY_CODE(Sp(0));
1966 /* -----------------------------------------------------------------------------
1968 -------------------------------------------------------------------------- */
1972 W_ ap_stack, offset, val, ok;
1974 /* args: R1 = AP_STACK, R2 = offset */
1978 if (%INFO_PTR(ap_stack) == stg_AP_STACK_info) {
1980 val = StgAP_STACK_payload(ap_stack,offset);
1988 // Write the cost center stack of the first argument on stderr; return
1989 // the second. Possibly only makes sense for already evaluated
1996 ccs = StgHeader_ccs(UNTAG(R1));
1997 foreign "C" fprintCCS_stderr(ccs "ptr") [R2];
2008 #ifndef THREADED_RTS
2009 RET_NP(0,ghczmprim_GHCziBool_False_closure);
2011 (spark) = foreign "C" findSpark(MyCapability());
2015 RET_NP(0,ghczmprim_GHCziBool_False_closure);
2025 #if defined(TRACING) || defined(DEBUG)
2027 foreign "C" traceUserMsg(MyCapability() "ptr", msg "ptr") [];
2029 #elif defined(DTRACE)
2033 // We should go through the macro HASKELLEVENT_USER_MSG_ENABLED from
2034 // RtsProbes.h, but that header file includes unistd.h, which doesn't
2036 (enabled) = foreign "C" __dtrace_isenabled$HaskellEvent$user__msg$v1() [];
2038 foreign "C" dtraceUserMsgWrapper(MyCapability() "ptr", msg "ptr") [];
2042 jump %ENTRY_CODE(Sp(0));