provide a more accurate assembly delay() routine

ok miod@

>From: NetBSD

2 files changed, 57 insertions, 51 deletions

diff --git a/sys/arch/mac68k/mac68k/clock.c b/sys/arch/mac68k/mac68k/clock.c
index b8e7eeefeb8..9971328f047 100644
--- a/sys/arch/mac68k/mac68k/clock.c
+++ b/sys/arch/mac68k/mac68k/clock.c
@@ -1,5 +1,5 @@
-/*	$OpenBSD: clock.c,v 1.17 2004/11/26 21:21:28 miod Exp $	*/
-/*	$NetBSD: clock.c,v 1.36 1997/10/07 03:04:55 scottr Exp $	*/
+/*	$OpenBSD: clock.c,v 1.18 2004/11/30 01:44:22 martin Exp $	*/
+/*	$NetBSD: clock.c,v 1.39 1999/11/05 19:14:56 scottr Exp $	*/
 
 /*
  * Copyright (c) 1988 University of Utah.
@@ -408,6 +408,8 @@ resettodr()
 		    "to the PRAM on this system.\n");
 #endif
 }
+
+
 /*
  * The Macintosh timers decrement once every 1.2766 microseconds.
  * MGFH2, p. 180
@@ -415,55 +417,11 @@ resettodr()
 #define	CLK_RATE	12766
 
 #define	DELAY_CALIBRATE	(0xffffff << 7)	/* Large value for calibration */
-#define	LARGE_DELAY	0x40000		/* About 335 msec */
 
 u_int		delay_factor = DELAY_CALIBRATE;
 volatile int	delay_flag = 1;
 
-/*
- * delay(usec)
- *	Delay usec microseconds.
- *
- * The delay_factor is scaled up by a factor of 128 to avoid loss
- * of precision for small delays.  As a result of this, note that
- * delays larger that LARGE_DELAY will be up to 128 usec too short,
- * due to adjustments for calculations involving 32 bit values.
- */
-void
-delay(usec)
-	unsigned usec;
-{
-	volatile unsigned int cycles;
-
-	if (usec > LARGE_DELAY)
-		cycles = (usec >> 7) * delay_factor;
-	else
-		cycles = ((usec > 0 ? usec : 1) * delay_factor) >> 7;
-
-	while ((cycles-- > 0) && delay_flag);
-}
-
-static unsigned	dummy_delay(unsigned);
-/*
- * Dummy delay calibration.  Functionally identical to delay(), but
- * returns the number of times through the loop.
- */
-static unsigned
-dummy_delay(usec)
-	unsigned usec;
-{
-	volatile unsigned int cycles;
-
-	if (usec > LARGE_DELAY)
-		cycles = (usec >> 7) * delay_factor;
-	else
-		cycles = ((usec > 0 ? usec : 1) * delay_factor) >> 7;
-
-	while ((cycles-- > 0) && delay_flag);
-
-	return ((delay_factor >> 7) - cycles);
-}
-
+int		_delay(u_int);
 static int	delay_timer1_irq(void *);
 
 static int
@@ -480,8 +438,9 @@ delay_timer1_irq(dummy)
 void
 mac68k_calibrate_delay()
 {
-	int n;
-	unsigned sum;
+	u_int sum, n;
+
+	(void)spl0();
 
 	/* Disable VIA1 timer 1 interrupts and set up service routine */
 	via_reg(VIA1, vIER) = V1IF_T1;
@@ -501,7 +460,7 @@ mac68k_calibrate_delay()
 		delay_flag = 1;
 		via_reg(VIA1, vT1C) = 0;	/* 1024 clock ticks */
 		via_reg(VIA1, vT1CH) = 4;	/* (approx 1.3 msec) */
-		sum += dummy_delay(1);
+		sum += ((delay_factor >> 7) - _delay(1));
 	}
 
 	/* Disable timer interrupts and reset service routine */
@@ -533,4 +492,6 @@ mac68k_calibrate_delay()
 	if (clock_debug)
 		printf("mac68k_calibrate_delay(): delay_factor calibrated\n");
 #endif
+
+	(void)splhigh();
 }
diff --git a/sys/arch/mac68k/mac68k/locore.s b/sys/arch/mac68k/mac68k/locore.s
index b49530ff836..762c40d819c 100644
--- a/sys/arch/mac68k/mac68k/locore.s
+++ b/sys/arch/mac68k/mac68k/locore.s
@@ -1,4 +1,4 @@
-/*	$OpenBSD: locore.s,v 1.38 2004/11/27 14:26:32 miod Exp $	*/
+/*	$OpenBSD: locore.s,v 1.39 2004/11/30 01:44:22 martin Exp $	*/
 /*	$NetBSD: locore.s,v 1.103 1998/07/09 06:02:50 scottr Exp $	*/
 
 /*
@@ -1506,6 +1506,51 @@ Lm68881rdone:
 	rts
 
 /*
+ * delay() - delay for a specified number of microseconds
+ * _delay() - calibrator helper for delay()
+ *
+ * Notice that delay_factor is scaled up by a factor of 128 to avoid loss
+ * of precision for small delays.  As a result of this we need to avoid
+ * overflow.
+ *
+ * The branch target for the loops must be aligned on a half-line (8-byte)
+ * boundary to minimize cache effects.  This guarantees both that there
+ * will be no prefetch stalls due to cache line burst operations and that
+ * the loops will run from a single cache half-line.
+ */
+	.align	8			| align to half-line boundary
+
+ALTENTRY(_delay, _delay)
+ENTRY(delay)
+	movl	sp@(4),d0		| get microseconds to delay
+	cmpl	#0x40000,d0		| is it a "large" delay?
+	bls	Ldelayshort		| no, normal calculation
+	movql	#0x7f,d1		| adjust for scaled multipler (to
+	addl	d1,d0			|   avoid overflow)
+	lsrl	#7,d0
+	mulul	_C_LABEL(delay_factor),d0 | calculate number of loop iterations
+	bra	Ldelaysetup		| go do it!
+Ldelayshort:
+	mulul	_C_LABEL(delay_factor),d0 | calculate number of loop iterations
+	lsrl	#7,d0			| adjust for scaled multiplier
+Ldelaysetup:
+	jeq	Ldelayexit		| bail out if nothing to do
+	movql	#0,d1			| put bits 15-0 in d1 for the
+	movw	d0,d1			|   inner loop, and move bits
+	movw	#0,d0			|   31-16 to the low-order word
+	subql	#1,d1			|   of d0 for the outer loop
+	swap	d0
+Ldelay:
+	tstl	_C_LABEL(delay_flag)	| this never changes for delay()!
+	dbeq	d1,Ldelay		|   (used only for timing purposes)
+	dbeq	d0,Ldelay
+	addqw	#1,d1			| adjust end count and
+	swap	d0			|    return the longword result
+	orl	d1,d0
+Ldelayexit:
+	rts
+
+/*
  * Handle the nitty-gritty of rebooting the machine.
  * Basically we just turn off the MMU and jump to the appropriate ROM routine.
  * Note that we must be running in an address range that is mapped one-to-one