SMP support (xchg(), cas() and mb()) for PowerPC

[ghc-hetmet.git] / ghc / includes / SMP.h

diff --git a/ghc/includes/SMP.h b/ghc/includes/SMP.h
index e3916bc..f4c8605 100644 (file)
--- a/ghc/includes/SMP.h
+++ b/ghc/includes/SMP.h
@@ -2,52 +2,128 @@
  *
  * (c) The GHC Team, 2005
  *
- * Macros for SMP support
+ * Macros for THREADED_RTS support
  *
  * -------------------------------------------------------------------------- */
 
 #ifndef SMP_H
 #define SMP_H
 
-/* SMP is currently not compatible with the following options:
+/* THREADED_RTS is currently not compatible with the following options:
  *
- *      INTERPRETER
- *      PROFILING
+ *      PROFILING (but only 1 CPU supported)
  *      TICKY_TICKY
- *      and unregisterised builds.
+ *      Unregisterised builds are ok, but only 1 CPU supported.
  */
 
-#if defined(SMP)
+#if defined(THREADED_RTS)
 
-#if  defined(PROFILING)  || defined(TICKY_TICKY)
-#error Build options incompatible with SMP.
+#if  defined(TICKY_TICKY)
+#error Build options incompatible with THREADED_RTS.
 #endif
 
 /* 
  * XCHG - the atomic exchange instruction.  Used for locking closures
- * during updates (see LOCK_CLOSURE below) and the MVar primops.
+ * during updates (see lockClosure() below) and the MVar primops.
+ *
+ * NB: the xchg instruction is implicitly locked, so we do not need
+ * a lock prefix here. 
  */
 INLINE_HEADER StgWord
 xchg(StgPtr p, StgWord w)
 {
     StgWord result;
+#if i386_HOST_ARCH || x86_64_HOST_ARCH
     result = w;
     __asm__ __volatile__ (
          "xchg %1,%0"
           :"+r" (result), "+m" (*p)
           : /* no input-only operands */
        );
+#elif powerpc_HOST_ARCH
+    __asm__ __volatile__ (
+        "1:     lwarx     %0, 0, %2\n"
+        "       stwcx.    %1, 0, %2\n"
+        "       bne-      1b"
+        :"=r" (result)
+        :"r" (w), "r" (p)
+    );
+#else
+#error xchg() unimplemented on this architecture
+#endif
     return result;
 }
 
+/* 
+ * CMPXCHG - the single-word atomic compare-and-exchange instruction.  Used 
+ * in the STM implementation.
+ */
+INLINE_HEADER StgWord
+cas(StgVolatilePtr p, StgWord o, StgWord n)
+{
+#if i386_HOST_ARCH || x86_64_HOST_ARCH
+    __asm__ __volatile__ (
+         "lock cmpxchg %3,%1"
+          :"=a"(o), "=m" (*(volatile unsigned int *)p) 
+          :"0" (o), "r" (n));
+    return o;
+#elif powerpc_HOST_ARCH
+    StgWord result;
+    __asm__ __volatile__ (
+        "1:     lwarx     %0, 0, %3\n"
+        "       cmpw      %0, %1\n"
+        "       bne       2f\n"
+        "       stwcx.    %2, 0, %3\n"
+        "       bne-      1b\n"
+        "2:"
+        :"=r" (result)
+        :"r" (o), "r" (n), "r" (p)
+    );
+    return result;
+#else
+#error cas() unimplemented on this architecture
+#endif
+}
+
+/*
+ * 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");
+#elif powerpc_HOST_ARCH
+    __asm__ __volatile__ ("lwsync" : : : "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
+
 INLINE_HEADER StgInfoTable *
 lockClosure(StgClosure *p)
 {
-#if i386_HOST_ARCH || x86_64_HOST_ARCH
+#if i386_HOST_ARCH || x86_64_HOST_ARCH || powerpc_HOST_ARCH
     StgWord info;
     do {
-       info = xchg((P_)&p->header.info, (W_)&stg_WHITEHOLE_info);
-       if (info != (W_)&stg_WHITEHOLE_info) return (StgInfoTable *)info;
+       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);
        yieldThread();
     } while (1);
 #else
@@ -58,14 +134,27 @@ lockClosure(StgClosure *p)
 INLINE_HEADER void
 unlockClosure(StgClosure *p, StgInfoTable *info)
 {
-#if i386_HOST_ARCH || x86_64_HOST_ARCH
-    // This is safe enough, because lockClosure() does the memory barrier:
+#if i386_HOST_ARCH || x86_64_HOST_ARCH || powerpc_HOST_ARCH
+    // This is a strictly ordered write, so we need a wb():
+    wb();
     p->header.info = info;
 #else
     RELEASE_SM_LOCK;
 #endif
 }
 
-#endif /* SMP */
+#else /* !THREADED_RTS */
+
+#define wb() /* nothing */
+
+INLINE_HEADER StgWord
+xchg(StgPtr p, StgWord w)
+{
+    StgWord old = *p;
+    *p = w;
+    return old;
+}
+
+#endif /* !THREADED_RTS */
 
 #endif /* SMP_H */