diff -urN linux-2.2.14/arch/alpha/kernel/alpha_ksyms.c linux-2.2.14.SuSE/arch/alpha/kernel/alpha_ksyms.c
--- linux-2.2.14/arch/alpha/kernel/alpha_ksyms.c	Tue Jan  4 23:15:00 2000
+++ linux-2.2.14.SuSE/arch/alpha/kernel/alpha_ksyms.c	Tue Jan  4 23:42:15 2000
@@ -30,6 +30,7 @@
 #include <asm/machvec.h>
 #include <asm/pgtable.h>
 #include <asm/semaphore.h>
+#include <asm/delay.h>
 
 #define __KERNEL_SYSCALLS__
 #include <asm/unistd.h>
@@ -96,6 +97,7 @@
 EXPORT_SYMBOL(__memset);
 EXPORT_SYMBOL(__memsetw);
 EXPORT_SYMBOL(__constant_c_memset);
+EXPORT_SYMBOL(___delay);
 
 EXPORT_SYMBOL(dump_thread);
 EXPORT_SYMBOL(dump_fpu);
diff -urN linux-2.2.14/arch/alpha/kernel/head.S linux-2.2.14.SuSE/arch/alpha/kernel/head.S
--- linux-2.2.14/arch/alpha/kernel/head.S	Sat Jun 12 20:52:52 1999
+++ linux-2.2.14.SuSE/arch/alpha/kernel/head.S	Tue Jan  4 23:42:15 2000
@@ -95,3 +95,21 @@
 	.prologue 0
 	call_pal PAL_halt
 	.end	halt
+
+	#
+	# Having the delay loop out of line guarantees that we wont
+	# run into weird alignment conditions (on new processors)
+	# that vary the speed of the loop.
+	#
+	.align 5
+	.globl	___delay
+	.ent	___delay
+___delay:
+	.set noat
+	.frame	$30,0,$28,0
+	.prologue 0
+1:	subq	$0,1,$0
+	bge	$0,1b
+	ret	$31,($28),0
+	.set at
+	.end	___delay
diff -urN linux-2.2.14/include/asm-alpha/delay.h linux-2.2.14.SuSE/include/asm-alpha/delay.h
--- linux-2.2.14/include/asm-alpha/delay.h	Tue Jan  4 23:15:04 2000
+++ linux-2.2.14.SuSE/include/asm-alpha/delay.h	Tue Jan  4 23:44:49 2000
@@ -9,18 +9,22 @@
  * Delay routines, using a pre-computed "loops_per_second" value.
  */
 
-/* We can make the delay loop inline, but we have to be very careful
- * WRT scheduling for EV6 machines, to keep it consistent for all locations
- * of invocations. This is a reasonable compromise.
- */
+/* We must make the innermost loop external, ie called via "jsr/bsr", so
+   that timings will be consistent accross ALL invocations, which is not
+   the case for a completely inline implementation. */
+
+extern void ___delay(void);
 
 extern __inline__ void
 __delay(unsigned long loops)
 {
-	__asm__ __volatile__(".align 4\n"
-		     "1:\tsubq %0,1,%0\n\t"
-		     "bge %0,1b\n\t"
-		     "nop": "=r" (loops) : "0" (loops));
+	register unsigned long r0 __asm__("$0") = loops;
+#ifdef MODULE
+	__asm__ __volatile__("lda $28,___delay; jsr $28,($28),0"
+			     : "=r"(r0) : "r"(r0) : "$28");
+#else
+	__asm__ __volatile__("bsr $28,___delay" : "=r"(r0) : "r"(r0) : "$28");
+#endif
 }
 
 /*