Fix unregisterised builds, and building on non-x86/amd64/powerpc
[ghc.git] / includes / SMP.h
1 /* ----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 2005
4 *
5 * Macros for THREADED_RTS support
6 *
7 * -------------------------------------------------------------------------- */
8
9 #ifndef SMP_H
10 #define SMP_H
11
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 #if defined(TICKY_TICKY)
22 #error Build options incompatible with THREADED_RTS.
23 #endif
24
25 /*
26 * XCHG - the atomic exchange instruction. Used for locking closures
27 * during updates (see lockClosure() below) and the MVar primops.
28 *
29 * NB: the xchg instruction is implicitly locked, so we do not need
30 * a lock prefix here.
31 */
32 INLINE_HEADER StgWord
33 xchg(StgPtr p, StgWord w)
34 {
35 StgWord result;
36 #if i386_HOST_ARCH || x86_64_HOST_ARCH
37 result = w;
38 __asm__ __volatile__ (
39 "xchg %1,%0"
40 :"+r" (result), "+m" (*p)
41 : /* no input-only operands */
42 );
43 #elif powerpc_HOST_ARCH
44 __asm__ __volatile__ (
45 "1: lwarx %0, 0, %2\n"
46 " stwcx. %1, 0, %2\n"
47 " bne- 1b"
48 :"=r" (result)
49 :"r" (w), "r" (p)
50 );
51 #elif !defined(WITHSMP)
52 result = *p;
53 *p = w;
54 #else
55 #error xchg() unimplemented on this architecture
56 #endif
57 return result;
58 }
59
60 /*
61 * CMPXCHG - the single-word atomic compare-and-exchange instruction. Used
62 * in the STM implementation.
63 */
64 INLINE_HEADER StgWord
65 cas(StgVolatilePtr p, StgWord o, StgWord n)
66 {
67 #if i386_HOST_ARCH || x86_64_HOST_ARCH
68 __asm__ __volatile__ (
69 "lock/cmpxchg %3,%1"
70 :"=a"(o), "=m" (*(volatile unsigned int *)p)
71 :"0" (o), "r" (n));
72 return o;
73 #elif powerpc_HOST_ARCH
74 StgWord result;
75 __asm__ __volatile__ (
76 "1: lwarx %0, 0, %3\n"
77 " cmpw %0, %1\n"
78 " bne 2f\n"
79 " stwcx. %2, 0, %3\n"
80 " bne- 1b\n"
81 "2:"
82 :"=&r" (result)
83 :"r" (o), "r" (n), "r" (p)
84 :"cc", "memory"
85 );
86 return result;
87 #elif !defined(WITHSMP)
88 StgWord result;
89 result = *p;
90 if (result == o) {
91 *p = n;
92 }
93 return result;
94 #else
95 #error cas() unimplemented on this architecture
96 #endif
97 }
98
99 /*
100 * Write barrier - ensure that all preceding writes have happened
101 * before all following writes.
102 *
103 * We need to tell both the compiler AND the CPU about the barrier.
104 * This is a brute force solution; better results might be obtained by
105 * using volatile type declarations to get fine-grained ordering
106 * control in C, and optionally a memory barrier instruction on CPUs
107 * that require it (not x86 or x86_64).
108 */
109 INLINE_HEADER void
110 write_barrier(void) {
111 #if i386_HOST_ARCH || x86_64_HOST_ARCH
112 __asm__ __volatile__ ("" : : : "memory");
113 #elif powerpc_HOST_ARCH
114 __asm__ __volatile__ ("lwsync" : : : "memory");
115 #elif !defined(WITHSMP)
116 return;
117 #else
118 #error memory barriers unimplemented on this architecture
119 #endif
120 }
121
122 /*
123 * Locking/unlocking closures
124 *
125 * This is used primarily in the implementation of MVars.
126 */
127 #define SPIN_COUNT 4000
128
129 INLINE_HEADER StgInfoTable *
130 lockClosure(StgClosure *p)
131 {
132 StgWord info;
133 do {
134 nat i = 0;
135 do {
136 info = xchg((P_)(void *)&p->header.info, (W_)&stg_WHITEHOLE_info);
137 if (info != (W_)&stg_WHITEHOLE_info) return (StgInfoTable *)info;
138 } while (++i < SPIN_COUNT);
139 yieldThread();
140 } while (1);
141 }
142
143 INLINE_HEADER void
144 unlockClosure(StgClosure *p, StgInfoTable *info)
145 {
146 // This is a strictly ordered write, so we need a wb():
147 write_barrier();
148 p->header.info = info;
149 }
150
151 #else /* !THREADED_RTS */
152
153 #define write_barrier() /* nothing */
154
155 INLINE_HEADER StgWord
156 xchg(StgPtr p, StgWord w)
157 {
158 StgWord old = *p;
159 *p = w;
160 return old;
161 }
162
163 INLINE_HEADER StgInfoTable *
164 lockClosure(StgClosure *p)
165 { return (StgInfoTable *)p->header.info; }
166
167 INLINE_HEADER void
168 unlockClosure(StgClosure *p STG_UNUSED, StgInfoTable *info STG_UNUSED)
169 { /* nothing */ }
170
171 #endif /* !THREADED_RTS */
172
173 // Handy specialised versions of lockClosure()/unlockClosure()
174 INLINE_HEADER void lockTSO(StgTSO *tso)
175 { lockClosure((StgClosure *)tso); }
176
177 INLINE_HEADER void unlockTSO(StgTSO *tso)
178 { unlockClosure((StgClosure*)tso, (StgInfoTable*)&stg_TSO_info); }
179
180 #endif /* SMP_H */