Better implementation of buf_space_*_raw_multi_* that deals well

with unaligned buffers

4 files changed, 291 insertions, 31 deletions

diff --git a/sys/arch/alpha/include/bus.h b/sys/arch/alpha/include/bus.h
index 01a2cb5a259..1a5189833ab 100644
--- a/sys/arch/alpha/include/bus.h
+++ b/sys/arch/alpha/include/bus.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: bus.h,v 1.6 1997/03/21 00:16:32 niklas Exp $	*/
+/*	$OpenBSD: bus.h,v 1.7 1997/04/02 22:08:07 niklas Exp $	*/
 /*	$NetBSD: bus.h,v 1.10 1996/12/02 22:19:32 cgd Exp $	*/
 
 /*
@@ -155,6 +155,24 @@ struct alpha_bus_space {
 			    bus_space_handle_t, bus_size_t, bus_size_t));
 	void		(*abs_c_8) __P((void *, bus_space_handle_t, bus_size_t,
 			    bus_space_handle_t, bus_size_t, bus_size_t));
+
+	/* OpenBSD extensions follows */
+
+	/* read multiple raw */
+	void		(*abs_rrm_2) __P((void *, bus_space_handle_t,
+			    bus_size_t, u_int8_t *, bus_size_t));
+	void		(*abs_rrm_4) __P((void *, bus_space_handle_t,
+			    bus_size_t, u_int8_t *, bus_size_t));
+	void		(*abs_rrm_8) __P((void *, bus_space_handle_t,
+			    bus_size_t, u_int8_t *, bus_size_t));
+
+	/* write multiple raw */
+	void		(*abs_wrm_2) __P((void *, bus_space_handle_t,
+			    bus_size_t, const u_int8_t *, bus_size_t));
+	void		(*abs_wrm_4) __P((void *, bus_space_handle_t,
+			    bus_size_t, const u_int8_t *, bus_size_t));
+	void		(*abs_wrm_8) __P((void *, bus_space_handle_t,
+			    bus_size_t, const u_int8_t *, bus_size_t));
 };
 
 
@@ -168,11 +186,14 @@ struct alpha_bus_space {
 	(*(t)->__abs_opname(r,sz))((t)->abs_cookie, h, o)
 #define	__abs_ws(sz, t, h, o, v)					\
 	(*(t)->__abs_opname(w,sz))((t)->abs_cookie, h, o, v)
-#ifndef DEBUG
 #define	__abs_nonsingle(type, sz, t, h, o, a, c)			\
 	(*(t)->__abs_opname(type,sz))((t)->abs_cookie, h, o, a, c)
+#ifndef DEBUG
+#define	__abs_aligned_nonsingle(type, sz, t, h, o, a, c)		\
+	__abs_nonsingle((type), (sz), (t), (h), (o), (a), (c))
+
 #else
-#define	__abs_nonsingle(type, sz, t, h, o, a, c)			\
+#define	__abs_aligend_nonsingle(type, sz, t, h, o, a, c)		\
     do {								\
 	if (((unsigned long)a & (sz - 1)) != 0)				\
 		panic("bus non-single %d-byte unaligned (to %p) at %s:%d", \
@@ -185,7 +206,6 @@ struct alpha_bus_space {
 #define	__abs_copy(sz, t, h1, o1, h2, o2, cnt)			\
 	(*(t)->__abs_opname(c,sz))((t)->abs_cookie, h1, o1, h2, o2, cnt)
 
-
 /*
  * Mapping and unmapping operations.
  */
@@ -232,11 +252,11 @@ struct alpha_bus_space {
 #define	bus_space_read_multi_1(t, h, o, a, c)				\
 	__abs_nonsingle(rm,1,(t),(h),(o),(a),(c))
 #define	bus_space_read_multi_2(t, h, o, a, c)				\
-	__abs_nonsingle(rm,2,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(rm,2,(t),(h),(o),(a),(c))
 #define	bus_space_read_multi_4(t, h, o, a, c)				\
-	__abs_nonsingle(rm,4,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(rm,4,(t),(h),(o),(a),(c))
 #define	bus_space_read_multi_8(t, h, o, a, c)				\
-	__abs_nonsingle(rm,8,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(rm,8,(t),(h),(o),(a),(c))
 
 
 /*
@@ -250,12 +270,12 @@ struct alpha_bus_space {
  * possible byte-swapping should be done by these functions.
  */
 
-#define	bus_space_read_raw_multi_2(t, h, o, a, c) \
-    bus_space_read_multi_2((t), (h), (o), (u_int16_t *)(a), (c) >> 1)
-#define	bus_space_read_raw_multi_4(t, h, o, a, c) \
-    bus_space_read_multi_4((t), (h), (o), (u_int32_t *)(a), (c) >> 2)
-#define	bus_space_read_raw_multi_8(t, h, o, a, c) \
-    bus_space_read_multi_8((t), (h), (o), (u_int64_t *)(a), (c) >> 4)
+#define	bus_space_read_raw_multi_2(t, h, o, a, c)			\
+	__abs_nonsingle(rrm,2,(t),(h),(o),(a),(c))
+#define	bus_space_read_raw_multi_4(t, h, o, a, c)			\
+	__abs_nonsingle(rrm,4,(t),(h),(o),(a),(c))
+#define	bus_space_read_raw_multi_8(t, h, o, a, c)			\
+	__abs_nonsingle(rrm,8,(t),(h),(o),(a),(c))
 
 /*
  * Bus read region operations.
@@ -263,11 +283,11 @@ struct alpha_bus_space {
 #define	bus_space_read_region_1(t, h, o, a, c)				\
 	__abs_nonsingle(rr,1,(t),(h),(o),(a),(c))
 #define	bus_space_read_region_2(t, h, o, a, c)				\
-	__abs_nonsingle(rr,2,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(rr,2,(t),(h),(o),(a),(c))
 #define	bus_space_read_region_4(t, h, o, a, c)				\
-	__abs_nonsingle(rr,4,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(rr,4,(t),(h),(o),(a),(c))
 #define	bus_space_read_region_8(t, h, o, a, c)				\
-	__abs_nonsingle(rr,8,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(rr,8,(t),(h),(o),(a),(c))
 
 
 /*
@@ -285,11 +305,11 @@ struct alpha_bus_space {
 #define	bus_space_write_multi_1(t, h, o, a, c)				\
 	__abs_nonsingle(wm,1,(t),(h),(o),(a),(c))
 #define	bus_space_write_multi_2(t, h, o, a, c)				\
-	__abs_nonsingle(wm,2,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(wm,2,(t),(h),(o),(a),(c))
 #define	bus_space_write_multi_4(t, h, o, a, c)				\
-	__abs_nonsingle(wm,4,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(wm,4,(t),(h),(o),(a),(c))
 #define	bus_space_write_multi_8(t, h, o, a, c)				\
-	__abs_nonsingle(wm,8,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(wm,8,(t),(h),(o),(a),(c))
 
 /*
  *	void bus_space_write_raw_multi_N __P((bus_space_tag_t tag,
@@ -302,12 +322,12 @@ struct alpha_bus_space {
  * possible byte-swapping should be done by these functions.
  */
 
-#define	bus_space_write_raw_multi_2(t, h, o, a, c) \
-    bus_space_write_multi_2((t), (h), (o), (u_int16_t *)(a), (c) >> 1)
-#define	bus_space_write_raw_multi_4(t, h, o, a, c) \
-    bus_space_write_multi_4((t), (h), (o), (u_int32_t *)(a), (c) >> 2)
-#define	bus_space_write_raw_multi_8(t, h, o, a, c) \
-    bus_space_write_multi_8((t), (h), (o), (u_int64_t *)(a), (c) >> 4)
+#define	bus_space_write_raw_multi_2(t, h, o, a, c)			\
+	__abs_nonsingle(wrm,2,(t),(h),(o),(a),(c))
+#define	bus_space_write_raw_multi_4(t, h, o, a, c)			\
+	__abs_nonsingle(wrm,4,(t),(h),(o),(a),(c))
+#define	bus_space_write_raw_multi_8(t, h, o, a, c)			\
+	__abs_nonsingle(wrm,8,(t),(h),(o),(a),(c))
 
 /*
  * Bus write region operations.
@@ -315,11 +335,11 @@ struct alpha_bus_space {
 #define	bus_space_write_region_1(t, h, o, a, c)				\
 	__abs_nonsingle(wr,1,(t),(h),(o),(a),(c))
 #define	bus_space_write_region_2(t, h, o, a, c)				\
-	__abs_nonsingle(wr,2,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(wr,2,(t),(h),(o),(a),(c))
 #define	bus_space_write_region_4(t, h, o, a, c)				\
-	__abs_nonsingle(wr,4,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(wr,4,(t),(h),(o),(a),(c))
 #define	bus_space_write_region_8(t, h, o, a, c)				\
-	__abs_nonsingle(wr,8,(t),(h),(o),(a),(c))
+	__abs_aligned_nonsingle(wr,8,(t),(h),(o),(a),(c))
 
 
 /*
diff --git a/sys/arch/alpha/pci/pcs_bus_io_common.c b/sys/arch/alpha/pci/pcs_bus_io_common.c
index 89cc1418cd1..f393cd2e7ef 100644
--- a/sys/arch/alpha/pci/pcs_bus_io_common.c
+++ b/sys/arch/alpha/pci/pcs_bus_io_common.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pcs_bus_io_common.c,v 1.5 1997/01/24 19:57:55 niklas Exp $	*/
+/*	$OpenBSD: pcs_bus_io_common.c,v 1.6 1997/04/02 22:08:09 niklas Exp $	*/
 /*	$NetBSD: pcs_bus_io_common.c,v 1.14 1996/12/02 22:19:35 cgd Exp $	*/
 
 /*
@@ -151,6 +151,25 @@ void		__C(CHIP,_io_copy_4) __P((void *, bus_space_handle_t,
 void		__C(CHIP,_io_copy_8) __P((void *, bus_space_handle_t,
 		    bus_size_t, bus_space_handle_t, bus_size_t, bus_size_t));
 
+/* read multiple raw */
+void		__C(CHIP,_io_read_raw_multi_2) __P((void *, bus_space_handle_t,
+		    bus_size_t, u_int8_t *, bus_size_t));
+void		__C(CHIP,_io_read_raw_multi_4) __P((void *, bus_space_handle_t,
+		    bus_size_t, u_int8_t *, bus_size_t));
+void		__C(CHIP,_io_read_raw_multi_8) __P((void *, bus_space_handle_t,
+		    bus_size_t, u_int8_t *, bus_size_t));
+
+/* write multiple raw */
+void		__C(CHIP,_io_write_raw_multi_2) __P((void *,
+		    bus_space_handle_t, bus_size_t, const u_int8_t *,
+		    bus_size_t));
+void		__C(CHIP,_io_write_raw_multi_4) __P((void *,
+		    bus_space_handle_t, bus_size_t, const u_int8_t *,
+		    bus_size_t));
+void		__C(CHIP,_io_write_raw_multi_8) __P((void *,
+		    bus_space_handle_t, bus_size_t, const u_int8_t *,
+		    bus_size_t));
+
 static long
     __C(CHIP,_io_ex_storage)[EXTENT_FIXED_STORAGE_SIZE(8) / sizeof(long)];
 
@@ -223,6 +242,16 @@ static struct alpha_bus_space __C(CHIP,_io_space) = {
 	__C(CHIP,_io_copy_2),
 	__C(CHIP,_io_copy_4),
 	__C(CHIP,_io_copy_8),
+
+	/* read multiple raw */
+	__C(CHIP,_io_read_raw_multi_2),
+	__C(CHIP,_io_read_raw_multi_4),
+	__C(CHIP,_io_read_raw_multi_8),
+	
+	/* write multiple raw*/
+	__C(CHIP,_io_write_raw_multi_2),
+	__C(CHIP,_io_write_raw_multi_4),
+	__C(CHIP,_io_write_raw_multi_8),
 };
 
 bus_space_tag_t
@@ -716,3 +745,55 @@ CHIP_io_copy_N(1)
 CHIP_io_copy_N(2)
 CHIP_io_copy_N(4)
 CHIP_io_copy_N(8)
+
+#define CHIP_io_read_raw_multi_N(BYTES,TYPE)				\
+void									\
+__C(__C(CHIP,_io_read_raw_multi_),BYTES)(v, h, o, a, c)			\
+	void *v;							\
+	bus_space_handle_t h;						\
+	bus_size_t o, c;						\
+	u_int8_t *a;							\
+{									\
+	TYPE temp;							\
+	int i;								\
+									\
+	while (c > 0) {							\
+		__C(CHIP,_io_barrier)(v, h, o, BYTES, BUS_BARRIER_READ); \
+		temp = __C(__C(CHIP,_io_read_),BYTES)(v, h, o);		\
+		for (i = 0; i < BYTES; i++) {				\
+			*a++ = temp & 0xff;				\
+			temp >>= 8;					\
+		}							\
+		c -= BYTES;						\
+	}								\
+}
+CHIP_io_read_raw_multi_N(2,u_int16_t)
+CHIP_io_read_raw_multi_N(4,u_int32_t)
+CHIP_io_read_raw_multi_N(8,u_int64_t)
+
+#define CHIP_io_write_raw_multi_N(BYTES,TYPE)				\
+void									\
+__C(__C(CHIP,_io_write_raw_multi_),BYTES)(v, h, o, a, c)		\
+	void *v;							\
+	bus_space_handle_t h;						\
+	bus_size_t o, c;						\
+	const u_int8_t *a;						\
+{									\
+	TYPE temp;							\
+	int i;								\
+									\
+	while (c > 0) {							\
+		temp = 0;						\
+		for (i = BYTES - 1; i >= 0; i--) {			\
+			temp <<= 8;					\
+			temp |= *(a + i);				\
+		}							\
+		__C(__C(CHIP,_io_write_),BYTES)(v, h, o, temp);		\
+		__C(CHIP,_io_barrier)(v, h, o, BYTES, BUS_BARRIER_WRITE); \
+		c -= BYTES;						\
+		a += BYTES;						\
+	}								\
+}
+CHIP_io_write_raw_multi_N(2,u_int16_t)
+CHIP_io_write_raw_multi_N(4,u_int32_t)
+CHIP_io_write_raw_multi_N(8,u_int64_t)
diff --git a/sys/arch/alpha/pci/pcs_bus_mem_common.c b/sys/arch/alpha/pci/pcs_bus_mem_common.c
index 44d0c90cf2f..0cd2f784206 100644
--- a/sys/arch/alpha/pci/pcs_bus_mem_common.c
+++ b/sys/arch/alpha/pci/pcs_bus_mem_common.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: pcs_bus_mem_common.c,v 1.5 1997/01/24 19:57:56 niklas Exp $	*/
+/*	$OpenBSD: pcs_bus_mem_common.c,v 1.6 1997/04/02 22:08:09 niklas Exp $	*/
 /*	$NetBSD: pcs_bus_mem_common.c,v 1.15 1996/12/02 22:19:36 cgd Exp $	*/
 
 /*
@@ -152,6 +152,25 @@ void		__C(CHIP,_mem_copy_4) __P((void *, bus_space_handle_t,
 void		__C(CHIP,_mem_copy_8) __P((void *, bus_space_handle_t,
 		    bus_size_t, bus_space_handle_t, bus_size_t, bus_size_t));
 
+/* read multiple raw */
+void		__C(CHIP,_mem_read_raw_multi_2) __P((void *,
+		    bus_space_handle_t, bus_size_t, u_int8_t *, bus_size_t));
+void		__C(CHIP,_mem_read_raw_multi_4) __P((void *,
+		    bus_space_handle_t, bus_size_t, u_int8_t *, bus_size_t));
+void		__C(CHIP,_mem_read_raw_multi_8) __P((void *,
+		    bus_space_handle_t, bus_size_t, u_int8_t *, bus_size_t));
+
+/* write multiple raw */
+void		__C(CHIP,_mem_write_raw_multi_2) __P((void *,
+		    bus_space_handle_t, bus_size_t, const u_int8_t *,
+		    bus_size_t));
+void		__C(CHIP,_mem_write_raw_multi_4) __P((void *,
+		    bus_space_handle_t, bus_size_t, const u_int8_t *,
+		    bus_size_t));
+void		__C(CHIP,_mem_write_raw_multi_8) __P((void *,
+		    bus_space_handle_t, bus_size_t, const u_int8_t *,
+		    bus_size_t));
+
 static long
     __C(CHIP,_dmem_ex_storage)[EXTENT_FIXED_STORAGE_SIZE(8) / sizeof(long)];
 static long
@@ -226,6 +245,16 @@ static struct alpha_bus_space __C(CHIP,_mem_space) = {
 	__C(CHIP,_mem_copy_2),
 	__C(CHIP,_mem_copy_4),
 	__C(CHIP,_mem_copy_8),
+
+	/* read multiple raw */
+	__C(CHIP,_mem_read_raw_multi_2),
+	__C(CHIP,_mem_read_raw_multi_4),
+	__C(CHIP,_mem_read_raw_multi_8),
+	
+	/* write multiple raw*/
+	__C(CHIP,_mem_write_raw_multi_2),
+	__C(CHIP,_mem_write_raw_multi_4),
+	__C(CHIP,_mem_write_raw_multi_8),
 };
 
 bus_space_tag_t
@@ -983,3 +1012,55 @@ CHIP_mem_copy_N(1)
 CHIP_mem_copy_N(2)
 CHIP_mem_copy_N(4)
 CHIP_mem_copy_N(8)
+
+#define CHIP_mem_read_raw_multi_N(BYTES,TYPE)				\
+void									\
+__C(__C(CHIP,_mem_read_raw_multi_),BYTES)(v, h, o, a, c)			\
+	void *v;							\
+	bus_space_handle_t h;						\
+	bus_size_t o, c;						\
+	u_int8_t *a;							\
+{									\
+	TYPE temp;							\
+	int i;								\
+									\
+	while (c > 0) {							\
+		__C(CHIP,_mem_barrier)(v, h, o, BYTES, BUS_BARRIER_READ); \
+		temp = __C(__C(CHIP,_mem_read_),BYTES)(v, h, o);	\
+		for (i = 0; i < BYTES; i++) {				\
+			*a++ = temp & 0xff;				\
+			temp >>= 8;					\
+		}							\
+		c -= BYTES;						\
+	}								\
+}
+CHIP_mem_read_raw_multi_N(2,u_int16_t)
+CHIP_mem_read_raw_multi_N(4,u_int32_t)
+CHIP_mem_read_raw_multi_N(8,u_int64_t)
+
+#define CHIP_mem_write_raw_multi_N(BYTES,TYPE)				\
+void									\
+__C(__C(CHIP,_mem_write_raw_multi_),BYTES)(v, h, o, a, c)		\
+	void *v;							\
+	bus_space_handle_t h;						\
+	bus_size_t o, c;						\
+	const u_int8_t *a;						\
+{									\
+	TYPE temp;							\
+	int i;								\
+									\
+	while (c > 0) {							\
+		temp = 0;						\
+		for (i = BYTES - 1; i >= 0; i--) {			\
+			temp <<= 8;					\
+			temp |= *(a + i);				\
+		}							\
+		__C(__C(CHIP,_mem_write_),BYTES)(v, h, o, temp);	\
+		__C(CHIP,_mem_barrier)(v, h, o, BYTES, BUS_BARRIER_WRITE); \
+		c -= BYTES;						\
+		a += BYTES;						\
+	}								\
+}
+CHIP_mem_write_raw_multi_N(2,u_int16_t)
+CHIP_mem_write_raw_multi_N(4,u_int32_t)
+CHIP_mem_write_raw_multi_N(8,u_int64_t)
diff --git a/sys/arch/alpha/tc/tc_bus_mem.c b/sys/arch/alpha/tc/tc_bus_mem.c
index 4d901513410..79740b2dc08 100644
--- a/sys/arch/alpha/tc/tc_bus_mem.c
+++ b/sys/arch/alpha/tc/tc_bus_mem.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: tc_bus_mem.c,v 1.6 1997/01/24 19:58:19 niklas Exp $	*/
+/*	$OpenBSD: tc_bus_mem.c,v 1.7 1997/04/02 22:08:10 niklas Exp $	*/
 /*	$NetBSD: tc_bus_mem.c,v 1.13 1996/12/02 22:19:34 cgd Exp $	*/
 
 /*
@@ -145,6 +145,22 @@ void		tc_mem_copy_4 __P((void *, bus_space_handle_t,
 void		tc_mem_copy_8 __P((void *, bus_space_handle_t,
 		    bus_size_t, bus_space_handle_t, bus_size_t, bus_size_t));
 
+/* read multiple raw */
+void		tc_mem_read_raw_multi_2 __P((void *, bus_space_handle_t,
+		    bus_size_t, u_int8_t *, bus_size_t));
+void		tc_mem_read_raw_multi_4 __P((void *, bus_space_handle_t,
+		    bus_size_t, u_int8_t *, bus_size_t));
+void		tc_mem_read_raw_multi_8 __P((void *, bus_space_handle_t,
+		    bus_size_t, u_int8_t *, bus_size_t));
+
+/* write multiple raw */
+void		tc_mem_write_raw_multi_2 __P((void *, bus_space_handle_t,
+		    bus_size_t, const u_int8_t *, bus_size_t));
+void		tc_mem_write_raw_multi_4 __P((void *, bus_space_handle_t,
+		    bus_size_t, const u_int8_t *, bus_size_t));
+void		tc_mem_write_raw_multi_8 __P((void *, bus_space_handle_t,
+		    bus_size_t, const u_int8_t *, bus_size_t));
+
 static struct alpha_bus_space tc_mem_space = {
 	/* cookie */
 	NULL,
@@ -214,6 +230,16 @@ static struct alpha_bus_space tc_mem_space = {
 	tc_mem_copy_2,
 	tc_mem_copy_4,
 	tc_mem_copy_8,
+
+	/* write multiple raw */
+	tc_mem_read_raw_multi_2,
+	tc_mem_read_raw_multi_4,
+	tc_mem_read_raw_multi_8,
+	
+	/* write multiple raw*/
+	tc_mem_write_raw_multi_2,
+	tc_mem_write_raw_multi_4,
+	tc_mem_write_raw_multi_8,
 };
 
 bus_space_tag_t
@@ -612,3 +638,55 @@ tc_mem_copy_N(1)
 tc_mem_copy_N(2)
 tc_mem_copy_N(4)
 tc_mem_copy_N(8)
+
+#define tc_mem_read_raw_multi_N(BYTES,TYPE)				\
+void									\
+__C(tc_mem_read_raw_multi_,BYTES)(v, h, o, a, c)			\
+	void *v;							\
+	bus_space_handle_t h;						\
+	bus_size_t o, c;						\
+	u_int8_t *a;							\
+{									\
+	TYPE temp;							\
+	int i;								\
+									\
+	while (c > 0) {							\
+		tc_mem_barrier(v, h, o, BYTES, BUS_BARRIER_READ);	\
+		temp = __C(tc_mem_read_,BYTES)(v, h, o);		\
+		for (i = 0; i < BYTES; i++) {				\
+			*a++ = temp & 0xff;				\
+			temp >>= 8;					\
+		}							\
+		c -= BYTES;						\
+	}								\
+}
+tc_mem_read_raw_multi_N(2,u_int16_t)
+tc_mem_read_raw_multi_N(4,u_int32_t)
+tc_mem_read_raw_multi_N(8,u_int64_t)
+
+#define tc_mem_write_raw_multi_N(BYTES,TYPE)				\
+void									\
+__C(tc_mem_write_raw_multi_,BYTES)(v, h, o, a, c)		\
+	void *v;							\
+	bus_space_handle_t h;						\
+	bus_size_t o, c;						\
+	const u_int8_t *a;						\
+{									\
+	TYPE temp;							\
+	int i;								\
+									\
+	while (c > 0) {							\
+		temp = 0;						\
+		for (i = BYTES - 1; i >= 0; i--) {			\
+			temp <<= 8;					\
+			temp |= *(a + i);				\
+		}							\
+		__C(tc_mem_write_,BYTES)(v, h, o, temp);		\
+		tc_mem_barrier(v, h, o, BYTES, BUS_BARRIER_WRITE);	\
+		c -= BYTES;						\
+		a += BYTES;						\
+	}								\
+}
+tc_mem_write_raw_multi_N(2,u_int16_t)
+tc_mem_write_raw_multi_N(4,u_int32_t)
+tc_mem_write_raw_multi_N(8,u_int64_t)