Override cpu_switch() with a faster version if we can use BWX instructions.

From RusticBSD, ok deraadt@

2 files changed, 43 insertions, 7 deletions

diff --git a/sys/arch/alpha/alpha/locore.s b/sys/arch/alpha/alpha/locore.s
index 8ed599af4ff..f517807db3e 100644
--- a/sys/arch/alpha/alpha/locore.s
+++ b/sys/arch/alpha/alpha/locore.s
@@ -1,4 +1,4 @@
-/* $OpenBSD: locore.s,v 1.27 2004/12/24 22:50:27 miod Exp $ */
+/* $OpenBSD: locore.s,v 1.28 2005/06/17 21:54:14 miod Exp $ */
 /* $NetBSD: locore.s,v 1.94 2001/04/26 03:10:44 ross Exp $ */
 
 /*-
@@ -891,10 +891,7 @@ cpu_switch_queuescan:
 	 *
 	 * Note: GET_CPUINFO clobbers v0, t0, t8...t11.
 	 */
-#ifdef __alpha_bwx__
-	ldiq	t0, SONPROC			/* p->p_stat = SONPROC */
-	stb	t0, P_STAT(s2)
-#else
+EXPORT(__bwx_switch0)
 	addq	s2, P_STAT, t3			/* p->p_stat = SONPROC */
 	ldq_u	t1, 0(t3)
 	ldiq	t0, SONPROC
@@ -902,7 +899,7 @@ cpu_switch_queuescan:
 	mskbl	t1, t3, t1
 	or	t0, t1, t0
 	stq_u	t0, 0(t3)
-#endif /* __alpha_bwx__ */
+EXPORT(__bwx_switch1)
 
 	GET_CPUINFO
 	/* p->p_cpu initialized in fork1() for single-processor */
@@ -935,6 +932,19 @@ cpu_switch_queuescan:
 	RET
 	END(cpu_switch)
 
+#ifndef SMALL_KERNEL
+	/*
+	 * BWX-enhanced version of the p->p_stat assignment, to be copied
+	 * over the __bwx_switch0 area.
+
+	 * Do not put anything between the end of cpu_switch and this!
+	 */
+EXPORT(__bwx_switch2)
+	ldiq	t0, SONPROC			/* p->p_stat = SONPROC */
+	stb	t0, P_STAT(s2)
+EXPORT(__bwx_switch3)
+#endif
+
 /*
  * switch_trampoline()
  *
diff --git a/sys/arch/alpha/alpha/machdep.c b/sys/arch/alpha/alpha/machdep.c
index 268ccba2a2c..ed271257d0b 100644
--- a/sys/arch/alpha/alpha/machdep.c
+++ b/sys/arch/alpha/alpha/machdep.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: machdep.c,v 1.91 2005/04/28 17:19:27 deraadt Exp $ */
+/* $OpenBSD: machdep.c,v 1.92 2005/06/17 21:54:14 miod Exp $ */
 /* $NetBSD: machdep.c,v 1.210 2000/06/01 17:12:38 thorpej Exp $ */
 
 /*-
@@ -412,6 +412,32 @@ nobootinfo:
 	 */
 #endif
 
+#ifndef SMALL_KERNEL
+	/*
+	 * If we run on a BWX-capable processor, override cpu_switch
+	 * with a faster version.
+	 * We do this now because the kernel text might be mapped
+	 * read-only eventually (although this is not the case at the moment).
+	 */
+	if (alpha_implver() >= ALPHA_IMPLVER_EV5) {
+		if (~alpha_amask(ALPHA_AMASK_BWX) != 0) {
+			extern vaddr_t __bwx_switch0, __bwx_switch1,
+			    __bwx_switch2, __bwx_switch3;
+			u_int32_t *dst, *src, *end;
+
+			src = (u_int32_t *)&__bwx_switch2;
+			end = (u_int32_t *)&__bwx_switch3;
+			dst = (u_int32_t *)&__bwx_switch0;
+			while (src != end)
+				*dst++ = *src++;
+			src = (u_int32_t *)&__bwx_switch1;
+			end = (u_int32_t *)&__bwx_switch2;
+			while (src != end)
+				*dst++ = *src++;
+		}
+	}
+#endif
+
 	/*
 	 * find out this system's page size
 	 */