We now following the ABI and always clear cld on function entry, so remove

the extra CLD instructions from when that wasn't true

testing miod@ krw@

8 files changed, 65 insertions, 87 deletions

diff --git a/sys/arch/amd64/amd64/bus_space.c b/sys/arch/amd64/amd64/bus_space.c
index 5cf9a745314..87813aab39d 100644
--- a/sys/arch/amd64/amd64/bus_space.c
+++ b/sys/arch/amd64/amd64/bus_space.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: bus_space.c,v 1.25 2014/11/16 12:30:56 deraadt Exp $	*/
+/*	$OpenBSD: bus_space.c,v 1.26 2015/04/25 21:31:24 guenther Exp $	*/
 /*	$NetBSD: bus_space.c,v 1.2 2003/03/14 18:47:53 christos Exp $	*/
 
 /*-
@@ -694,7 +694,7 @@ x86_bus_space_io_read_region_1(bus_space_handle_t h,
 	int dummy3;
 	int __x;
 	u_int32_t port = h + o;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"1:	inb %w1,%%al				;"
 	"	stosb					;"
 	"	incl %1					;"
@@ -714,7 +714,7 @@ x86_bus_space_io_read_region_2(bus_space_handle_t h,
 	int dummy3;
 	int __x;
 	u_int32_t port = h + o;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"1:	inw %w1,%%ax				;"
 	"	stosw					;"
 	"	addl $2,%1				;"
@@ -734,7 +734,7 @@ x86_bus_space_io_read_region_4(bus_space_handle_t h,
 	int dummy3;
 	int __x;
 	u_int32_t port = h + o;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	inl %w1,%%eax				;"
 	"	stosl					;"
 	"	addl $4,%1				;"
@@ -813,7 +813,7 @@ x86_bus_space_io_write_region_1(bus_space_handle_t h,
 	int dummy3;
 	int __x;
 	u_int32_t port = h + o;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsb					;"
 	"	outb %%al,%w1				;"
 	"	incl %1					;"
@@ -833,7 +833,7 @@ x86_bus_space_io_write_region_2(bus_space_handle_t h,
 	int dummy3;
 	int __x;
 	u_int32_t port = h + o;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsw					;"
 	"	outw %%ax,%w1				;"
 	"	addl $2,%1				;"
@@ -853,7 +853,7 @@ x86_bus_space_io_write_region_4(bus_space_handle_t h,
 	int dummy3;
 	int __x;
 	u_int32_t port = h + o;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"1:	lodsl					;"
 	"	outl %%eax,%w1				;"
 	"	addl $4,%1				;"
@@ -1066,7 +1066,7 @@ x86_bus_space_mem_read_multi_1(bus_space_handle_t h, bus_size_t o,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"1:	movb (%2),%%al				;"
 	"	stosb					;"
 	"	loop 1b"				:
@@ -1083,7 +1083,7 @@ x86_bus_space_mem_read_multi_2(bus_space_handle_t h, bus_size_t o,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"1:	movw (%2),%%ax				;"
 	"	stosw					;"
 	"	loop 1b"				:
@@ -1100,7 +1100,7 @@ x86_bus_space_mem_read_multi_4(bus_space_handle_t h, bus_size_t o,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"1:	movl (%2),%%eax				;"
 	"	stosl					;"
 	"	loop 1b"				:
@@ -1117,7 +1117,7 @@ x86_bus_space_mem_read_multi_8(bus_space_handle_t h, bus_size_t o,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"1:	movq (%2),%%rax				;"
 	"	stosq					;"
 	"	loop 1b"				:
@@ -1133,7 +1133,7 @@ x86_bus_space_mem_read_region_1(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsb"					:
 	    "=S" (dummy1), "=D" (dummy2), "=c" (dummy3)	:
@@ -1148,7 +1148,7 @@ x86_bus_space_mem_read_region_2(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile(" cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsw"					:
 	    "=S" (dummy1), "=D" (dummy2), "=c" (dummy3)	:
@@ -1163,7 +1163,7 @@ x86_bus_space_mem_read_region_4(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsl"					:
 	    "=S" (dummy1), "=D" (dummy2), "=c" (dummy3)	:
@@ -1178,7 +1178,7 @@ x86_bus_space_mem_read_region_8(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsq"					:
 	    "=S" (dummy1), "=D" (dummy2), "=c" (dummy3)	:
@@ -1206,7 +1206,7 @@ x86_bus_space_mem_write_multi_1(bus_space_handle_t h,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsb					;"
 	"	movb %%al,(%2)				;"
 	"	loop 1b"				:
@@ -1222,7 +1222,7 @@ x86_bus_space_mem_write_multi_2(bus_space_handle_t h,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsw					;"
 	"	movw %%ax,(%2)				;"
 	"	loop 1b"				:
@@ -1238,7 +1238,7 @@ x86_bus_space_mem_write_multi_4(bus_space_handle_t h,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsl					;"
 	"	movl %%eax,(%2)				;"
 	"	loop 1b"				:
@@ -1254,7 +1254,7 @@ x86_bus_space_mem_write_multi_8(bus_space_handle_t h,
 	int dummy2;
 	void *dummy3;
 	int __x;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsq					;"
 	"	movq %%rax,(%2)				;"
 	"	loop 1b"				:
@@ -1269,7 +1269,7 @@ x86_bus_space_mem_write_region_1(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsb"					:
 	    "=D" (dummy1), "=S" (dummy2), "=c" (dummy3)	:
@@ -1284,7 +1284,7 @@ x86_bus_space_mem_write_region_2(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsw"					:
 	    "=D" (dummy1), "=S" (dummy2), "=c" (dummy3)	:
@@ -1299,7 +1299,7 @@ x86_bus_space_mem_write_region_4(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsl"					:
 	    "=D" (dummy1), "=S" (dummy2), "=c" (dummy3)	:
@@ -1314,7 +1314,7 @@ x86_bus_space_mem_write_region_8(bus_space_handle_t h,
 	int dummy1;
 	void *dummy2;
 	int dummy3;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	movsq"					:
 	    "=D" (dummy1), "=S" (dummy2), "=c" (dummy3)	:
diff --git a/sys/arch/amd64/amd64/copy.S b/sys/arch/amd64/amd64/copy.S
index c638e2d1683..ba71b7fa2be 100644
--- a/sys/arch/amd64/amd64/copy.S
+++ b/sys/arch/amd64/amd64/copy.S
@@ -1,4 +1,4 @@
-/*	$OpenBSD: copy.S,v 1.6 2015/01/16 10:17:51 sf Exp $	*/
+/*	$OpenBSD: copy.S,v 1.7 2015/04/25 21:31:24 guenther Exp $	*/
 /*	$NetBSD: copy.S,v 1.1 2003/04/26 18:39:26 fvdl Exp $	*/
 
 /*
@@ -81,8 +81,7 @@ ENTRY(kcopy)
 	subq	%rsi,%rax
 	cmpq	%rcx,%rax		# overlapping?
 	jb	1f
-	cld				# nope, copy forward
-	shrq	$3,%rcx			# copy by 64-bit words
+	shrq	$3,%rcx			# nope, copy forward by 64-bit words
 	rep
 	movsq
 
@@ -134,7 +133,6 @@ ENTRY(copyout)
 	leaq	_C_LABEL(copy_fault)(%rip),%r11
 	movq	%r11,PCB_ONFAULT(%rdx)
 	SMAP_STAC
-	cld
 	movq	%rax,%rcx
 	shrq	$3,%rcx
 	rep
@@ -165,7 +163,6 @@ ENTRY(copyin)
 	ja	_C_LABEL(copy_efault)
 
 3:	/* bcopy(%rsi, %rdi, %rax); */
-	cld
 	movq	%rax,%rcx
 	shrq	$3,%rcx
 	rep
@@ -211,7 +208,6 @@ ENTRY(copyoutstr)
 	movq	%rax,%r8
 
 1:	incq	%rdx
-	cld
 
 1:	decq	%rdx
 	jz	2f
@@ -254,7 +250,6 @@ ENTRY(copyinstr)
 	movq	%rax,%r8
 
 1:	incq	%rdx
-	cld
 
 1:	decq	%rdx
 	jz	2f
@@ -296,7 +291,6 @@ ENTRY(copystr)
 	movq	%rdx,%r8
 
 	incq	%rdx
-	cld
 
 1:	decq	%rdx
 	jz	4f
diff --git a/sys/arch/amd64/include/pio.h b/sys/arch/amd64/include/pio.h
index 928564c73d8..8715a186b19 100644
--- a/sys/arch/amd64/include/pio.h
+++ b/sys/arch/amd64/include/pio.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pio.h,v 1.4 2014/03/29 18:09:28 guenther Exp $	*/
+/*	$OpenBSD: pio.h,v 1.5 2015/04/25 21:31:24 guenther Exp $	*/
 /*	$NetBSD: pio.h,v 1.2 2003/02/27 11:22:46 fvdl Exp $	*/
 
 /*-
@@ -78,7 +78,7 @@ insb(unsigned port, void *addr, int cnt)
 {
 	void *dummy1;
 	int dummy2;
-	__asm volatile("cld\n\trepne\n\tinsb"			:
+	__asm volatile("repne\n\tinsb"				:
 			 "=D" (dummy1), "=c" (dummy2) 		:
 			 "d" (port), "0" (addr), "1" (cnt)	:
 			 "memory");
@@ -108,7 +108,7 @@ insw(unsigned port, void *addr, int cnt)
 {
 	void *dummy1;
 	int dummy2;
-	__asm volatile("cld\n\trepne\n\tinsw"			:
+	__asm volatile("repne\n\tinsw"				:
 			 "=D" (dummy1), "=c" (dummy2)		:
 			 "d" (port), "0" (addr), "1" (cnt)	:
 			 "memory");
@@ -138,7 +138,7 @@ insl(unsigned port, void *addr, int cnt)
 {
 	void *dummy1;
 	int dummy2;
-	__asm volatile("cld\n\trepne\n\tinsl"			:
+	__asm volatile("repne\n\tinsl"				:
 			 "=D" (dummy1), "=c" (dummy2)		:
 			 "d" (port), "0" (addr), "1" (cnt)	:
 			 "memory");
@@ -165,7 +165,7 @@ outsb(unsigned port, const void *addr, int cnt)
 {
 	void *dummy1;
 	int dummy2;
-	__asm volatile("cld\n\trepne\n\toutsb"			:
+	__asm volatile("repne\n\toutsb"				:
 			 "=S" (dummy1), "=c" (dummy2)		:
 			 "d" (port), "0" (addr), "1" (cnt));
 }
@@ -191,7 +191,7 @@ outsw(unsigned port, const void *addr, int cnt)
 {
 	void *dummy1;
 	int dummy2;
-	__asm volatile("cld\n\trepne\n\toutsw"			:
+	__asm volatile("repne\n\toutsw"				:
 			 "=S" (dummy1), "=c" (dummy2)		:
 			 "d" (port), "0" (addr), "1" (cnt));
 }
@@ -217,7 +217,7 @@ outsl(unsigned port, const void *addr, int cnt)
 {
 	void *dummy1;
 	int dummy2;
-	__asm volatile("cld\n\trepne\n\toutsl"			:
+	__asm volatile("repne\n\toutsl"				:
 			 "=S" (dummy1), "=c" (dummy2)		:
 			 "d" (port), "0" (addr), "1" (cnt));
 }
diff --git a/sys/arch/i386/i386/bus_space.c b/sys/arch/i386/i386/bus_space.c
index 4b1c9246962..10dee17eb3f 100644
--- a/sys/arch/i386/i386/bus_space.c
+++ b/sys/arch/i386/i386/bus_space.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: bus_space.c,v 1.8 2014/10/17 20:37:57 sthen Exp $ */
+/*	$OpenBSD: bus_space.c,v 1.9 2015/04/25 21:31:24 guenther Exp $ */
 /*-
  * Copyright (c) 1996, 1997 The NetBSD Foundation, Inc.
  * All rights reserved.
@@ -303,7 +303,7 @@ i386_bus_space_io_read_region_1(bus_space_handle_t h,
 	void *_addr = a;
 	int _port = h + o;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	inb %w2,%%al				;"
 	"	stosb					;"
 	"	incl %2					;"
@@ -320,7 +320,7 @@ i386_bus_space_io_read_region_2(bus_space_handle_t h,
 	void *_addr = a;
 	int _port = h + o;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	inw %w2,%%ax				;"
 	"	stosw					;"
 	"	addl $2,%2				;"
@@ -337,7 +337,7 @@ i386_bus_space_io_read_region_4(bus_space_handle_t h,
 	void *_addr = a;
 	int _port = h + o;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	inl %w2,%%eax				;"
 	"	stosl					;"
 	"	addl $4,%2				;"
@@ -393,7 +393,7 @@ i386_bus_space_io_write_region_1(bus_space_handle_t h,
 	const void *_addr = a;
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsb					;"
 	"	outb %%al,%w0				;"
 	"	incl %0					;"
@@ -410,7 +410,7 @@ i386_bus_space_io_write_region_2(bus_space_handle_t h,
 	const void *_addr = a;
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsw					;"
 	"	outw %%ax,%w0				;"
 	"	addl $2,%0				;"
@@ -427,7 +427,7 @@ i386_bus_space_io_write_region_4(bus_space_handle_t h,
 	const void *_addr = a;
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsl					;"
 	"	outl %%eax,%w0				;"
 	"	addl $4,%0				;"
@@ -442,7 +442,7 @@ i386_bus_space_io_set_multi_1(bus_space_handle_t h,
 {
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	outb %b2, %w1				;"
 	"	loop 1b"				:
 	    "+c" (_cnt) : "d" (h + o), "a" (v)		:
@@ -455,7 +455,7 @@ i386_bus_space_io_set_multi_2(bus_space_handle_t h,
 {
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	outw %w2, %w1				;"
 	"	loop 1b"				:
 	    "+c" (_cnt) : "d" (h + o), "a" (v)	:
@@ -468,7 +468,7 @@ i386_bus_space_io_set_multi_4(bus_space_handle_t h,
 {
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	outl %2,%w1				;"
 	"	loop 1b"				:
 	    "+c" (_cnt) : "d" (h + o), "a" (v)	:
@@ -617,7 +617,7 @@ i386_bus_space_mem_read_multi_1(bus_space_handle_t h, bus_size_t o,
 {
 	void *_addr=a;
 	int _cnt=cnt;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	movb (%2),%%al				;"
 	"	stosb					;"
 	"	loop 1b"				:
@@ -631,7 +631,7 @@ i386_bus_space_mem_read_multi_2(bus_space_handle_t h, bus_size_t o,
 {
 	void *_addr=a;
 	int _cnt=cnt;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	movw (%2),%%ax				;"
 	"	stosw					;"
 	"	loop 1b"				:
@@ -645,7 +645,7 @@ i386_bus_space_mem_read_multi_4(bus_space_handle_t h, bus_size_t o,
 {
 	void *_addr=a;
 	int _cnt=cnt;
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	movl (%2),%%eax				;"
 	"	stosl					;"
 	"	loop 1b"				:
@@ -711,7 +711,7 @@ i386_bus_space_mem_write_multi_1(bus_space_handle_t h,
 	const void *_addr=a;
 	int _cnt=cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsb					;"
 	"	movb %%al,(%2)				;"
 	"	loop 1b"				:
@@ -726,7 +726,7 @@ i386_bus_space_mem_write_multi_2(bus_space_handle_t h,
 	const void *_addr = a;
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsw					;"
 	"	movw %%ax,(%2)				;"
 	"	loop 1b"				:
@@ -741,7 +741,7 @@ i386_bus_space_mem_write_multi_4(bus_space_handle_t h,
 	const void *_addr=a;
 	int _cnt=cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	lodsl					;"
 	"	movl %%eax,(%2)				;"
 	"	loop 1b"				:
@@ -788,7 +788,7 @@ i386_bus_space_mem_set_multi_1(bus_space_handle_t h,
 {
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	movb %b2, (%1)				;"
 	"	loop 1b"				:
 	    "+c" (_cnt) : "D" (h + o), "a" (v)		:
@@ -801,7 +801,7 @@ i386_bus_space_mem_set_multi_2(bus_space_handle_t h,
 {
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	movw %w2, (%1)				;"
 	"	loop 1b"				:
 	    "+c" (_cnt) : "D" (h + o), "a" (v)		:
@@ -814,7 +814,7 @@ i386_bus_space_mem_set_multi_4(bus_space_handle_t h,
 {
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"1:	movl %2,(%1)				;"
 	"	loop 1b"				:
 	    "+c" (_cnt) : "D" (h + o), "a" (v)	:
@@ -828,7 +828,7 @@ i386_bus_space_mem_set_region_1(bus_space_handle_t h,
 	int _port = h + o;
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	stosb"					:
 	    "+D" (_port), "+c" (_cnt) : "a" (v)	:
@@ -842,7 +842,7 @@ i386_bus_space_mem_set_region_2(bus_space_handle_t h,
 	int _port = h + o;
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	stosw"					:
 	    "+D" (_port), "+c" (_cnt) : "a" (v)	:
@@ -856,7 +856,7 @@ i386_bus_space_mem_set_region_4(bus_space_handle_t h,
 	int _port = h + o;
 	int _cnt = cnt;
 
-	__asm volatile("cld				;"
+	__asm volatile(
 	"	repne					;"
 	"	stosl"					:
 	    "+D" (_port), "+c" (_cnt) : "a" (v)	:
diff --git a/sys/arch/i386/i386/kvm86call.S b/sys/arch/i386/i386/kvm86call.S
index 444a72fdfb2..f4c20ae423b 100644
--- a/sys/arch/i386/i386/kvm86call.S
+++ b/sys/arch/i386/i386/kvm86call.S
@@ -1,4 +1,4 @@
-/* $OpenBSD: kvm86call.S,v 1.6 2010/04/05 22:42:43 kettenis Exp $ */
+/* $OpenBSD: kvm86call.S,v 1.7 2015/04/25 21:31:24 guenther Exp $ */
 /* $NetBSD: kvm86call.S,v 1.7 2006/04/11 17:14:07 drochner Exp $ */
 
 /*-
@@ -74,7 +74,6 @@ ENTRY(kvm86_call)
 	movl	vm86frame,%edi		/* target frame location */
 	movl    SCRARGFRAME,%esi	/* source (set on entry) */
 	movl	$FRAMESIZE/4,%ecx	/* sizeof(struct trapframe)/4 */
-	cld
 	rep
 	movsl				/* copy frame to new stack */
 
@@ -163,7 +162,6 @@ ENTRY(kvm86_ret)
 	movl	8(%ebp),%esi		/* source */
 	movl	SCRARGFRAME,%edi	/* destination */
 	movl	$FRAMESIZE/4,%ecx	/* size */
-	cld
 	rep
 	movsl				/* copy frame to original frame */
 
diff --git a/sys/arch/i386/i386/locore.s b/sys/arch/i386/i386/locore.s
index bb8a1d3857d..1e8b4d159a6 100644
--- a/sys/arch/i386/i386/locore.s
+++ b/sys/arch/i386/i386/locore.s
@@ -1,4 +1,4 @@
-/*	$OpenBSD: locore.s,v 1.155 2015/04/24 12:52:38 kettenis Exp $	*/
+/*	$OpenBSD: locore.s,v 1.156 2015/04/25 21:31:24 guenther Exp $	*/
 /*	$NetBSD: locore.s,v 1.145 1996/05/03 19:41:19 christos Exp $	*/
 
 /*-
@@ -544,7 +544,6 @@ try586:	/* Use the `cpuid' instruction. */
 	subl	%edi,%ecx			# size of tables
 	shrl	$2,%ecx
 	xorl	%eax, %eax
-	cld
 	rep
 	stosl
 
@@ -764,8 +763,7 @@ ENTRY(kcopy)
 	subl	%esi,%eax
 	cmpl	%ecx,%eax		# overlapping?
 	jb	1f
-	cld				# nope, copy forward
-	shrl	$2,%ecx			# copy by 32-bit words
+	shrl	$2,%ecx			# nope, copy forward by 32-bit words
 	rep
 	movsl
 	movl	24+FPADD(%esp),%ecx
@@ -852,7 +850,6 @@ ENTRY(copyout)
 	SMAP_STAC
 
 	/* bcopy(%esi, %edi, %eax); */
-	cld
 	movl	%eax,%ecx
 	shrl	$2,%ecx
 	rep
@@ -904,7 +901,6 @@ ENTRY(copyin)
 	ja	_C_LABEL(copy_fault)
 
 	/* bcopy(%esi, %edi, %eax); */
-	cld
 	movl	%eax,%ecx
 	shrl	$2,%ecx
 	rep
@@ -974,7 +970,6 @@ ENTRY(copyoutstr)
 	movl	%eax,20+FPADD(%esp)
 
 1:	incl	%edx
-	cld
 
 1:	decl	%edx
 	jz	2f
@@ -1030,7 +1025,6 @@ ENTRY(copyinstr)
 	movl	%eax,20+FPADD(%esp)
 
 1:	incl	%edx
-	cld
 
 1:	decl	%edx
 	jz	2f
@@ -1090,7 +1084,6 @@ ENTRY(copystr)
 	movl	16+FPADD(%esp),%edi		# edi = to
 	movl	20+FPADD(%esp),%edx		# edx = maxlen
 	incl	%edx
-	cld
 
 1:	decl	%edx
 	jz	4f
@@ -1538,7 +1531,6 @@ ENTRY(bzero)
 	movl	8(%esp),%edi
 	movl	12(%esp),%edx
 
-	cld				/* set fill direction forward */
 	xorl	%eax,%eax		/* set fill data to 0 */
 
 	/*
@@ -1618,7 +1610,6 @@ ENTRY(i686_pagezero)
 
 	movl	12(%esp), %edi
 	movl	$1024, %ecx
-	cld
 
 	ALIGN_TEXT
 1:
@@ -1679,7 +1670,6 @@ ENTRY(cpu_paenable)
 	movl	%edi, %cr3
 	addl	$KERNBASE, %edi /* and make it back virtual again */
 	movl	$8, %ecx
-	cld
 	rep
 	movsl
 
diff --git a/sys/arch/i386/i386/vector.s b/sys/arch/i386/i386/vector.s
index f865b217fa0..8f3aadff113 100644
--- a/sys/arch/i386/i386/vector.s
+++ b/sys/arch/i386/i386/vector.s
@@ -1,4 +1,4 @@
-/*	$OpenBSD: vector.s,v 1.18 2013/11/28 19:30:46 brad Exp $	*/
+/*	$OpenBSD: vector.s,v 1.19 2015/04/25 21:31:24 guenther Exp $	*/
 /*	$NetBSD: vector.s,v 1.32 1996/01/07 21:29:47 mycroft Exp $	*/
 
 /*
@@ -49,10 +49,6 @@
  * If the interrupt frame is made more flexible,  INTR can push %eax first and
  * decide the ipending case with less overhead, e.g., by avoiding loading the
  * segment registers.
- *
- * XXX
- * Should we do a cld on every system entry to avoid the requirement for
- * scattered cld's?
  */
 
 	.globl	_C_LABEL(isa_strayintr)
diff --git a/sys/arch/i386/include/pio.h b/sys/arch/i386/include/pio.h
index 284c7d702f9..f70d503f396 100644
--- a/sys/arch/i386/include/pio.h
+++ b/sys/arch/i386/include/pio.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pio.h,v 1.11 2014/03/29 18:09:29 guenther Exp $	*/
+/*	$OpenBSD: pio.h,v 1.12 2015/04/25 21:31:24 guenther Exp $	*/
 /*	$NetBSD: pio.h,v 1.13 1996/03/08 20:15:23 cgd Exp $	*/
 
 /*
@@ -76,7 +76,7 @@ __inb(int port)
 static __inline void
 insb(int port, void *addr, int cnt)
 {
-	__asm volatile("cld\n\trepne\n\tinsb"
+	__asm volatile("repne\n\tinsb"
 	    : "+D" (addr), "+c" (cnt) : "d" (port) : "memory", "cc");
 }
 
@@ -102,7 +102,7 @@ __inw(int port)
 static __inline void
 insw(int port, void *addr, int cnt)
 {
-	__asm volatile("cld\n\trepne\n\tinsw"
+	__asm volatile("repne\n\tinsw"
 	    : "+D" (addr), "+c" (cnt) : "d" (port) : "memory", "cc");
 }
 
@@ -128,7 +128,7 @@ __inl(int port)
 static __inline void
 insl(int port, void *addr, int cnt)
 {
-	__asm volatile("cld\n\trepne\n\tinsl"
+	__asm volatile("repne\n\tinsl"
 	    : "+D" (addr), "+c" (cnt) : "d" (port) : "memory", "cc");
 }
 
@@ -150,7 +150,7 @@ __outb(int port, u_int8_t data)
 static __inline void
 outsb(int port, const void *addr, int cnt)
 {
-	__asm volatile("cld\n\trepne\n\toutsb"
+	__asm volatile("repne\n\toutsb"
 	    : "+S" (addr), "+c" (cnt) : "d" (port) : "cc");
 }
 
@@ -172,7 +172,7 @@ __outw(int port, u_int16_t data)
 static __inline void
 outsw(int port, const void *addr, int cnt)
 {
-	__asm volatile("cld\n\trepne\n\toutsw"
+	__asm volatile("repne\n\toutsw"
 	    : "+S" (addr), "+c" (cnt) : "d" (port) : "cc");
 }
 
@@ -194,7 +194,7 @@ __outl(int port, u_int32_t data)
 static __inline void
 outsl(int port, const void *addr, int cnt)
 {
-	__asm volatile("cld\n\trepne\n\toutsl"
+	__asm volatile("repne\n\toutsl"
 	    : "+S" (addr), "+c" (cnt) : "d" (port) : "cc");
 }