From 06ddffe09a37acd77de2aa93cdd99f4c6f27923b Mon Sep 17 00:00:00 2001
From: Stefan Lankes <lankes@lfbs.rwth-aachen.de>
Date: Wed, 5 Oct 2011 22:42:09 -0700
Subject: [PATCH] use optimized memcpy functions for non-cacheable memory

---
 drivers/net/mmnif.c | 82 +++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 79 insertions(+), 3 deletions(-)

diff --git a/drivers/net/mmnif.c b/drivers/net/mmnif.c
index 754f7f34..6b4f1c25 100644
--- a/drivers/net/mmnif.c
+++ b/drivers/net/mmnif.c
@@ -266,6 +266,76 @@ typedef struct mmnif {
 	sem_t com_poll;
 } mmnif_t;
 
+inline static void* memcpy_from_nc(void *dest, const void *src, size_t count)
+{
+#if 0
+	size_t i;
+
+	for(i=0; i<count; i++)
+		((uint8_t*) dest)[i] = ((uint8_t*) src)[i];
+
+	return dest;
+#else
+	int32_t h, i, j, k, l, m;
+
+	asm volatile    ("cld;\n\t"
+			"1: cmpl $0, %%eax ; je 3f\n\t"
+			"movl (%%edi), %%edx\n\t"
+			"cmpl $1, %%eax ; je 2f\n\t"
+			"movl 32(%%edi), %%edx\n\t"
+			"2: movl 0(%%esi), %%ecx\n\t"
+			"movl 4(%%esi), %%edx\n\t"
+			"movl %%ecx, 0(%%edi)\n\t"
+			"movl %%edx, 4(%%edi)\n\t"
+			"movl 8(%%esi), %%ecx\n\t"
+			"movl 12(%%esi), %%edx\n\t"
+			"movl %%ecx, 8(%%edi)\n\t"
+			"movl %%edx, 12(%%edi)\n\t"
+			"movl 16(%%esi), %%ecx\n\t"
+			"movl 20(%%esi), %%edx\n\t"
+			"movl %%ecx, 16(%%edi)\n\t"
+			"movl %%edx, 20(%%edi)\n\t"
+			"movl 24(%%esi), %%ecx\n\t"
+			"movl 28(%%esi), %%edx\n\t"
+			"movl %%ecx, 24(%%edi)\n\t"
+			"movl %%edx, 28(%%edi)\n\t"
+			"addl $32, %%esi\n\t"
+			"addl $32, %%edi\n\t"
+			"dec %%eax ; jmp 1b\n\t"
+			"3: movl %%ebx, %%ecx\n\t"
+			"movl (%%edi), %%edx\n\t"
+			"andl $31, %%ecx\n\t"
+			"rep ; movsb\n\t" : "=&a"(h), "=&D"(i), "=&S"(j), "=&b"(k), "=&c"(l), "=&d"(m)
+			: "0"(count / 32), "1"(dest), "2"(src), "3"(count) : "memory","cc");
+
+	return dest;
+#endif
+}
+
+inline static void* memcpy_to_nc(void* dest, const void *src, size_t count)
+{
+#if 0
+	size_t i;
+        
+	for(i=0; i<count; i++)
+		((uint8_t*) dest)[i] = ((uint8_t*) src)[i];
+
+	return dest;
+#else
+	int32_t i, j, k;
+
+	asm volatile (
+		"cld; rep movsl\n\t"
+		"movl %4, %%ecx\n\t"
+		"andl $3, %%ecx\n\t"
+		"rep movsb\n\t"
+		: "=&c"(i), "=&D"(j), "=&S"(k)
+		: "0"(count/4), "g"(count), "1"(dest), "2"(src) : "memory","cc");
+
+	return dest;
+#endif
+}
+
 /*
  *	memory maped interface helper functions
  */
@@ -630,7 +700,9 @@ realloc:
 
 	for (q = p, i = 0; q != 0; q = q->next)
 	{
-#if MMNIF_USE_MPB
+#if !MMNIF_USE_MBO && !USE_CACHE
+		memcpy_to_nc((char*) write_address+ i, q->payload, q->len);
+#elif MMNIF_USE_MPB
 		memcpy_put((char*) write_address + i, q->payload, q->len);
 #else
 		memcpy((char*) write_address+ i, q->payload, q->len);
@@ -794,7 +866,9 @@ realloc:
 	asm volatile (".byte 0x0f; .byte 0x0a;\n");
 #endif
 
-#if MMNIF_USE_MPB
+#if !MMNIF_USE_MBO && !USE_CACHE
+	 memcpy_to_nc((void*) write_address, pbuff, CLINE_ALIGN(size));
+#elif MMNIF_USE_MPB
 	memcpy_put(write_address, pbuff, CLINE_ALIGN(size));
 #else
 	memcpy((void*) write_address, pbuff, CLINE_ALIGN(size));
@@ -1174,7 +1248,9 @@ anotherpacket:
 	/* copy packet to pbuf structure going through linked list */
 	for (q = p, i = 0; q != NULL; q = q->next)
 	{
-#if MMNIF_USE_MPB
+#if !USE_CACHE || !MMNIF_USE_MBP
+		memcpy_from_nc((uint8_t *) q->payload, &packet[i], q->len);
+#elif MMNIF_USE_MPB
 		memcpy_get((uint8_t *) q->payload, &packet[i], q->len);
 #else
 		memcpy((uint8_t *) q->payload, &packet[i], q->len);