From d3359db4d5b4cb6f1c2758dc03e74edf4427a769 Mon Sep 17 00:00:00 2001
From: John Hodge <tpg@mutabah.net>
Date: Fri, 11 Nov 2011 19:22:34 +0800
Subject: [PATCH] Kernel - Fixing bugs exposed by enabling -O3

---
 Kernel/Makefile       |  1 +
 Kernel/arch/x86/lib.c | 40 +++++++++++++++++++++++++++++++++++++---
 Kernel/threads.c      |  2 +-
 Kernel/vfs/select.c   |  2 +-
 4 files changed, 40 insertions(+), 5 deletions(-)

diff --git a/Kernel/Makefile b/Kernel/Makefile
index 287d01a9..900a46e6 100644
--- a/Kernel/Makefile
+++ b/Kernel/Makefile
@@ -23,6 +23,7 @@ CPPFLAGS	+= -DARCH=$(ARCH) -DARCHDIR=$(ARCHDIR) -DARCHDIR_IS_$(ARCHDIR)=1
 CPPFLAGS	+= -DKERNEL_VERSION=$(KERNEL_VERSION)
 CFLAGS  	+= -Wall -Werror -fno-stack-protector -Wstrict-prototypes -g
 CFLAGS  	+= -Wshadow -Wpointer-arith -Wcast-align -Wwrite-strings -Wmissing-prototypes -Wmissing-declarations -Wredundant-decls -Wnested-externs -Winline -Wuninitialized
+CFLAGS          += -O3
 LDFLAGS		+= -T arch/$(ARCHDIR)/link.ld -g
 
 OBJDIR := obj-$(ARCH)/
diff --git a/Kernel/arch/x86/lib.c b/Kernel/arch/x86/lib.c
index 0b802d1e..db05d936 100644
--- a/Kernel/arch/x86/lib.c
+++ b/Kernel/arch/x86/lib.c
@@ -324,11 +324,45 @@ int memcmp(const void *m1, const void *m2, size_t Num)
  */
 void *memcpy(void *Dest, const void *Src, size_t Num)
 {
-	if( ((Uint)Dest & 3) || ((Uint)Src & 3) ) {
-		__asm__ __volatile__ ("rep movsb" :: "D" (Dest), "S" (Src), "c" (Num));
+	tVAddr	dst = (tVAddr)Dest;
+	tVAddr	src = (tVAddr)Src;
+	if( (dst & 3) != (src & 3) )
+	{
+		__asm__ __volatile__ ("rep movsb" :: "D" (dst), "S" (src), "c" (Num));
 //		Debug("\nmemcpy:Num=0x%x by %p (UA)", Num, __builtin_return_address(0));
 	}
-	else {
+	#if 1
+	else if( Num > 128 && (dst & 15) == (src & 15) )
+	{
+		 int	count = 16 - (dst & 15);
+//		Debug("\nmemcpy:Num=0x%x by %p (SSE)", Num, __builtin_return_address(0));
+		if( count < 16 )
+		{
+			Num -= count;
+//			Debug("dst = %p, src = %p, count = %i (Num=0x%x) (head)", dst, src, count, Num);
+			__asm__ __volatile__ ("rep movsb" : "=D"(dst),"=S"(src): "0"(dst), "1"(src), "c"(count));
+		}
+		
+		count = Num / 16;
+//		Debug("dst = %p, src = %p, count = %i (bulk)", dst, src, count);
+		__asm__ __volatile__ (
+			"1:\n\t"
+			"movdqa 0(%1), %%xmm0;\n\t"
+			"movdqa %%xmm0, 0(%0);\n\t"
+			"add $16,%0;\n\t"
+			"add $16,%1;\n\t"
+			"loop 1b;\n\t"
+			: "=r"(dst),"=r"(src) : "0"(dst), "1"(src), "c"(count)
+			);
+
+		count = Num & 15;
+//		Debug("dst = %p, src = %p, count = %i (tail)", dst, src, count);
+		if(count)
+			__asm__ __volatile__ ("rep movsb" :: "D"(dst), "S"(src), "c"(count));
+	}
+	#endif
+	else
+	{
 //		Debug("\nmemcpy:Num=0x%x by %p", Num, __builtin_return_address(0));
 		__asm__ __volatile__ (
 			"rep movsl;\n\t"
diff --git a/Kernel/threads.c b/Kernel/threads.c
index 6933dd60..879fbdda 100644
--- a/Kernel/threads.c
+++ b/Kernel/threads.c
@@ -26,7 +26,7 @@
 #define	DEFAULT_QUANTUM	5
 #define	DEFAULT_PRIORITY	5
 #define MIN_PRIORITY		10
-const enum eConfigTypes	cCONFIG_TYPES[] = {
+const enum eConfigTypes	cCONFIG_TYPES[NUM_CFG_ENTRIES] = {
 	CFGT_HEAPSTR,	// e.g. CFG_VFS_CWD
 	CFGT_INT,	// e.g. CFG_VFS_MAXFILES
 	CFGT_NULL
diff --git a/Kernel/vfs/select.c b/Kernel/vfs/select.c
index 18fdab76..bbca5471 100644
--- a/Kernel/vfs/select.c
+++ b/Kernel/vfs/select.c
@@ -430,7 +430,7 @@ int VFS_int_Select_AddThread(tVFS_SelectList *List, tVFS_SelectThread *Thread, i
 void VFS_int_Select_RemThread(tVFS_SelectList *List, tVFS_SelectThread *Thread)
 {
 	 int	i;
-	tVFS_SelectListEnt	*block, *prev;
+	tVFS_SelectListEnt	*block, *prev = NULL;
 	
 	ENTER("pList pThread", List, Thread);
 	
-- 
2.20.1