From 51d46f8e4690456942e014c8313f89de2fb1260f Mon Sep 17 00:00:00 2001
From: Kees van Reeuwijk <reeuwijk@few.vu.nl>
Date: Tue, 4 May 2010 21:02:44 +0000
Subject: [PATCH] Let memory allocation be aligned on 8-byte boundaries.

---
 lib/libc/ansi/calloc.c           | 11 ++++++++++-
 lib/libc/ansi/malloc.c           | 19 ++++++++++++++-----
 lib/libc/arch/i386/misc/alloca.S |  4 ++--
 3 files changed, 26 insertions(+), 8 deletions(-)

diff --git a/lib/libc/ansi/calloc.c b/lib/libc/ansi/calloc.c
index cccd79ab7..0c4323902 100644
--- a/lib/libc/ansi/calloc.c
+++ b/lib/libc/ansi/calloc.c
@@ -1,7 +1,16 @@
 /* $Header$ */
 #include	<stdlib.h>
 
-#define ALIGN(x)	(((x) + (sizeof(size_t) - 1)) & ~(sizeof(size_t) - 1))
+/* replace undef by define */
+#define  ALIGN_EIGHT_BYTES /* Use 8-byte alignment. */
+
+#ifdef  ALIGN_EIGHT_BYTES
+#define ALIGN_SIZE 8
+#else
+#define ALIGN_SIZE sizeof(size_t)
+#endif
+
+#define ALIGN(x)	(((x) + (ALIGN_SIZE - 1)) & ~(ALIGN_SIZE - 1))
 
 void *
 calloc(size_t nelem, size_t elsize)
diff --git a/lib/libc/ansi/malloc.c b/lib/libc/ansi/malloc.c
index d6990fef8..f4644f1c2 100644
--- a/lib/libc/ansi/malloc.c
+++ b/lib/libc/ansi/malloc.c
@@ -1,8 +1,9 @@
 /* $Header$ */
 
 /* replace undef by define */
-#define	 DEBUG		/* check assertions */
-#undef	 SLOWDEBUG	/* some extra test loops (requires DEBUG) */
+#define  ALIGN_EIGHT_BYTES /* Use 8-byte alignment. */
+#define	 DEBUG		   /* check assertions */
+#undef	 SLOWDEBUG	   /* some extra test loops (requires DEBUG) */
 
 #ifndef DEBUG
 #define NDEBUG
@@ -24,7 +25,11 @@
 #else
 #define BRKSIZE		4096
 #endif
+#ifdef ALIGN_EIGHT_BYTES
+#define PTRSIZE		8
+#else
 #define	PTRSIZE		((int) sizeof(void *))
+#endif
 #define Align(x,a)	(((x) + (a - 1)) & ~(a - 1))
 #define NextSlot(p)	(* (void **) ((p) - PTRSIZE))
 #define NextFree(p)	(* (void **) (p))
@@ -41,6 +46,9 @@
  * linked together by a pointer at the start of the
  * user visable part, so just after the next-slot pointer.
  * Free slots are merged together by free().
+ *
+ * Since modern processors prefer 8-byte alignment, we now pretend
+ * our pointers are 8 bytes wide.
  */
 
 extern void *_sbrk(int);
@@ -67,16 +75,17 @@ static int grow(size_t len)
 }
 
 void *
-malloc(size_t size)
+malloc(const size_t size)
 {
   register char *prev, *p, *next, *new;
-  register unsigned len, ntries;
+  unsigned ntries;
 
   if (size == 0)
 	return NULL;
 
   for (ntries = 0; ntries < 2; ntries++) {
-	if ((len = Align(size, PTRSIZE) + PTRSIZE) < 2 * PTRSIZE) {
+	unsigned len = Align(size, PTRSIZE) + PTRSIZE;
+	if (len < 2 * PTRSIZE) {
 		errno = ENOMEM;
 		return NULL;
 	}
diff --git a/lib/libc/arch/i386/misc/alloca.S b/lib/libc/arch/i386/misc/alloca.S
index 24de081fb..0c9b598fe 100644
--- a/lib/libc/arch/i386/misc/alloca.S
+++ b/lib/libc/arch/i386/misc/alloca.S
@@ -10,7 +10,7 @@ _alloca:
 	pop	%ecx	/* Return address */
 	pop	%eax	/* Bytes to allocate */
 	addl	$2*4+3, %eax	/* Add space for two saved register variables */
-	andb	$0xFC, %al	/* Align */
+	andb	$0xF8, %al	/* Align */
 	movl	%esp, %ebx	/* Keep current esp */
 	subl	%eax, %esp	/* Lower stack */
 	movl	%esp, %eax	/* Return value */
@@ -24,7 +24,7 @@ _alloca:
 	pop	%ecx	/* Return address */
 	pop	%eax	/* Bytes to allocate */
 	addl	$3, %eax
-	andb	$0xFC, %al	/* Align */
+	andb	$0xF8, %al	/* Align */
 	subl	%eax, %esp	/* Lower stack */
 	movl	%esp, %eax	/* Return value */
 	push	%eax	/* Dummy argument */