SMP.h 2.88 KB
Newer Older
1
2
/* ----------------------------------------------------------------------------
 *
3
 * (c) The GHC Team, 2005
4
 *
5
 * Macros for THREADED_RTS support
6
7
8
9
10
11
 *
 * -------------------------------------------------------------------------- */

#ifndef SMP_H
#define SMP_H

12
/* THREADED_RTS is currently not compatible with the following options:
13
 *
14
 *      PROFILING (but only 1 CPU supported)
15
 *      TICKY_TICKY
16
 *      Unregisterised builds are ok, but only 1 CPU supported.
17
18
 */

19
#if defined(THREADED_RTS)
20

21
22
#if  defined(TICKY_TICKY)
#error Build options incompatible with THREADED_RTS.
23
24
25
26
#endif

/* 
 * XCHG - the atomic exchange instruction.  Used for locking closures
27
 * during updates (see lockClosure() below) and the MVar primops.
28
29
30
 *
 * NB: the xchg instruction is implicitly locked, so we do not need
 * a lock prefix here. 
31
 */
32
33
34
35
36
37
INLINE_HEADER StgWord
xchg(StgPtr p, StgWord w)
{
    StgWord result;
    result = w;
    __asm__ __volatile__ (
38
 	  "xchg %1,%0"
39
40
41
42
43
44
          :"+r" (result), "+m" (*p)
          : /* no input-only operands */
	);
    return result;
}

45
46
47
48
49
50
51
52
53
54
55
56
57
58
/* 
 * CMPXCHG - the single-word atomic compare-and-exchange instruction.  Used 
 * in the STM implementation.
 */
INLINE_HEADER StgWord
cas(StgVolatilePtr p, StgWord o, StgWord n)
{
    __asm__ __volatile__ (
 	  "lock cmpxchg %3,%1"
          :"=a"(o), "=m" (*(volatile unsigned int *)p) 
          :"0" (o), "r" (n));
    return o;
}

59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
/*
 * Write barrier - ensure that all preceding writes have happened
 * before all following writes.  
 *
 * We need to tell both the compiler AND the CPU about the barrier.
 * This is a brute force solution; better results might be obtained by
 * using volatile type declarations to get fine-grained ordering
 * control in C, and optionally a memory barrier instruction on CPUs
 * that require it (not x86 or x86_64).
 */
INLINE_HEADER void
wb(void) {
#if i386_HOST_ARCH || x86_64_HOST_ARCH
    __asm__ __volatile__ ("" : : : "memory");
#else
#error memory barriers unimplemented on this architecture
#endif
}

/*
 * Locking/unlocking closures
 *
 * This is used primarily in the implementation of MVars.
 */
#define SPIN_COUNT 4000

85
86
87
INLINE_HEADER StgInfoTable *
lockClosure(StgClosure *p)
{
88
#if i386_HOST_ARCH || x86_64_HOST_ARCH
89
90
    StgWord info;
    do {
91
92
93
94
95
	nat i = 0;
	do {
	    info = xchg((P_)&p->header.info, (W_)&stg_WHITEHOLE_info);
	    if (info != (W_)&stg_WHITEHOLE_info) return (StgInfoTable *)info;
	} while (++i < SPIN_COUNT);
96
97
	yieldThread();
    } while (1);
98
#else
99
100
101
102
103
104
105
106
   ACQUIRE_SM_LOCK
#endif
}

INLINE_HEADER void
unlockClosure(StgClosure *p, StgInfoTable *info)
{
#if i386_HOST_ARCH || x86_64_HOST_ARCH
107
108
    // This is a strictly ordered write, so we need a wb():
    wb();
109
110
111
    p->header.info = info;
#else
    RELEASE_SM_LOCK;
112
#endif
113
}
114

115
#else /* !THREADED_RTS */
116
117
118
119
120
121
122
123
124
125
126

#define wb() /* nothing */

INLINE_HEADER StgWord
xchg(StgPtr p, StgWord w)
{
    StgWord old = *p;
    *p = w;
    return old;
}

127
#endif /* !THREADED_RTS */
128
129

#endif /* SMP_H */