From 6a3dfed789e30191406785ad013619984fba6dda Mon Sep 17 00:00:00 2001
From: "H. Peter Anvin" <hpa@zytor.com>
Date: Mon, 3 Mar 2008 16:55:15 -0800
Subject: [PATCH] Drop unnecessary CLD instructions.

The EFI ABI spec states that DF=0 on ABI boundaries, and gcc expects
this.  Assume it is safe to rely upon everywhere; drop unnecessary cld
instructions, except on entry and (obviously) after std.
---
 com32/lib/memcpy.S           |  1 -
 com32/lib/mempcpy.S          |  1 -
 com32/lib/memset.c           | 16 ++--------------
 com32/lib/sys/vesa/drawtxt.c |  2 +-
 com32/lib/sys/vesa/fill.h    |  6 +++---
 com32/modules/mboot.c        |  6 +++---
 6 files changed, 9 insertions(+), 23 deletions(-)

diff --git a/com32/lib/memcpy.S b/com32/lib/memcpy.S
index ed8e20e..a893d24 100644
--- a/com32/lib/memcpy.S
+++ b/com32/lib/memcpy.S
@@ -19,7 +19,6 @@ memcpy:
 	movl	%ecx, %edx
 
 	shrl	$2, %ecx
-	cld
 	rep ; movsl
 
 	jnc	1f		# The shrl had carry out if odd word count
diff --git a/com32/lib/mempcpy.S b/com32/lib/mempcpy.S
index aa17f47..eeab918 100644
--- a/com32/lib/mempcpy.S
+++ b/com32/lib/mempcpy.S
@@ -19,7 +19,6 @@ mempcpy:
 	movl	%ecx, %edx
 
 	shrl	$2, %ecx
-	cld
 	rep ; movsl
 
 	jnc	1f		# The shrl had carry out if odd word count
diff --git a/com32/lib/memset.c b/com32/lib/memset.c
index 522cc59..516ad47 100644
--- a/com32/lib/memset.c
+++ b/com32/lib/memset.c
@@ -8,23 +8,11 @@
 void *memset(void *dst, int c, size_t n)
 {
   char *q = dst;
-
-#if defined(__i386__)
   size_t nl = n >> 2;
-  asm volatile("cld ; rep ; stosl ; movl %3,%0 ; rep ; stosb"
+
+  asm volatile("rep ; stosl ; movl %3,%0 ; rep ; stosb"
 	       : "+c" (nl), "+D" (q)
 	       : "a" ((unsigned char)c * 0x01010101U), "r" (n & 3));
-#elif defined(__x86_64__)
-  size_t nq = n >> 3;
-  asm volatile("cld ; rep ; stosq ; movl %3,%%ecx ; rep ; stosb"
-	       : "+c" (nq), "+D" (q)
-	       : "a" ((unsigned char)c * 0x0101010101010101U),
-	       "r" ((uint32_t)n & 7));
-#else
-  while ( n-- ) {
-    *q++ = c;
-  }
-#endif
 
   return dst;
 }
diff --git a/com32/lib/sys/vesa/drawtxt.c b/com32/lib/sys/vesa/drawtxt.c
index 6e21a34..662eebe 100644
--- a/com32/lib/sys/vesa/drawtxt.c
+++ b/com32/lib/sys/vesa/drawtxt.c
@@ -40,7 +40,7 @@ static int cursor_x, cursor_y;
 
 static inline void *copy_dword(void *dst, void *src, size_t dword_count)
 {
-  asm volatile("cld; rep; movsl"
+  asm volatile("rep; movsl"
 	       : "+D" (dst), "+S" (src), "+c" (dword_count));
   return dst;			/* Updated destination pointer */
 }
diff --git a/com32/lib/sys/vesa/fill.h b/com32/lib/sys/vesa/fill.h
index 4d4a9f9..63ca577 100644
--- a/com32/lib/sys/vesa/fill.h
+++ b/com32/lib/sys/vesa/fill.h
@@ -37,19 +37,19 @@ static inline struct vesa_char *vesacon_fill(struct vesa_char *ptr,
 {
   switch (sizeof(struct vesa_char)) {
   case 1:
-    asm volatile("cld; rep; stosb"
+    asm volatile("rep; stosb"
 		 : "+D" (ptr), "+c" (count)
 		 : "a" (fill)
 		 : "memory");
     break;
   case 2:
-    asm volatile("cld; rep; stosw"
+    asm volatile("rep; stosw"
 		 : "+D" (ptr), "+c" (count)
 		 : "a" (fill)
 		 : "memory");
     break;
   case 4:
-    asm volatile("cld; rep; stosl"
+    asm volatile("rep; stosl"
 		 : "+D" (ptr), "+c" (count)
 		 : "a" (fill)
 		 : "memory");
diff --git a/com32/modules/mboot.c b/com32/modules/mboot.c
index 3d74425..6c3ce34 100644
--- a/com32/modules/mboot.c
+++ b/com32/modules/mboot.c
@@ -835,7 +835,7 @@ static void trampoline_start(section_t *secs, int sec_count,
             /* asm bzero() code from com32/lib/memset.c */
             char *q = (char *) secs[i].dest;
             size_t nl = secs[i].size >> 2;
-            asm volatile("cld ; rep ; stosl ; movl %3,%0 ; rep ; stosb"
+            asm volatile("rep ; stosl ; movl %3,%0 ; rep ; stosb"
                          : "+c" (nl), "+D" (q)
                          : "a" (0x0U), "r" (secs[i].size & 3));
         } else {
@@ -844,12 +844,12 @@ static void trampoline_start(section_t *secs, int sec_count,
             char *q = (char *) secs[i].dest;
             size_t n = secs[i].size;
             if ( q < p ) {
-                asm volatile("cld ; rep ; movsb"
+                asm volatile("rep ; movsb"
                              : "+c" (n), "+S" (p), "+D" (q));
             } else {
                 p += (n-1);
                 q += (n-1);
-                asm volatile("std ; rep ; movsb"
+                asm volatile("std ; rep ; movsb ; cld"
                              : "+c" (n), "+S" (p), "+D" (q));
             }
         }
-- 
2.7.4