CRIS/CRISv32 architecture specific memcpy, memmove and memset.

4 files changed, 698 insertions, 0 deletions

diff --git a/libc/string/cris/memcopy.h b/libc/string/cris/memcopy.h
new file mode 100644
index 000000000..0eae998da
--- /dev/null
+++ b/libc/string/cris/memcopy.h
@@ -0,0 +1,62 @@
+/* Copyright (C) 2001 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   Modified for use in uClibc (C) 2007 Axis Communications AB.
+   Minimal modifications: include path name and #undef of WORD_COPY_FWD/BWD
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 of the
+   License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+#include "../generic/memcopy.h"
+
+/* We override the word-copying macros, partly because misalignment in one
+   pointer isn't cause for a special function, partly because we want to
+   get rid of all the static functions in generic/memcopy.c; these macros 
+   are only used in memmove.c since we have arch-specific mempcpy, memcpy and
+   memset.  */
+
+#undef OP_T_THRES
+#define OP_T_THRES OPSIZ
+
+#undef WORD_COPY_FWD
+#define WORD_COPY_FWD(dst_bp, src_bp, nbytes_left, nbytes)		\
+  do									\
+    {									\
+      unsigned long enddst_bp = dst_bp + nbytes - (nbytes % OPSIZ);	\
+      nbytes_left = (nbytes % OPSIZ);					\
+      while (dst_bp < (unsigned long) enddst_bp)			\
+	{								\
+	  op_t x = *(op_t *) src_bp;					\
+	  src_bp += sizeof x;						\
+	  *(op_t *) dst_bp = x;						\
+	  dst_bp += sizeof x;						\
+	}								\
+    } while (0)
+
+#undef WORD_COPY_BWD
+#define WORD_COPY_BWD(dst_bp, src_bp, nbytes_left, nbytes)		\
+  do									\
+    {									\
+      unsigned long enddst_bp = dst_bp - nbytes + (nbytes % OPSIZ);	\
+      nbytes_left = (nbytes % OPSIZ);					\
+      while (dst_bp > enddst_bp)					\
+	{								\
+	  op_t x;							\
+	  src_bp -= sizeof x;						\
+	  x = *(op_t *) src_bp;						\
+	  dst_bp -= sizeof x;						\
+	  *(op_t *) dst_bp = x;						\
+	}								\
+    } while (0)
diff --git a/libc/string/cris/memcpy.c b/libc/string/cris/memcpy.c
new file mode 100644
index 000000000..a85108109
--- /dev/null
+++ b/libc/string/cris/memcpy.c
@@ -0,0 +1,264 @@
+/* Copyright (C) 2001, 2003 Free Software Foundation, Inc.
+   Copyright (C) 1994, 1995, 2000 Axis Communications AB.
+
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 of the
+   License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+/*#************************************************************************#*/
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# FUNCTION NAME: memcpy()                                                 */
+/*#                                                                         */
+/*# PARAMETERS:  void* dst;   Destination address.                          */
+/*#              void* src;   Source address.                               */
+/*#              int   len;   Number of bytes to copy.                      */
+/*#                                                                         */
+/*# RETURNS:     dst.                                                       */
+/*#                                                                         */
+/*# DESCRIPTION: Copies len bytes of memory from src to dst.  No guarantees */
+/*#              about copying of overlapping memory areas. This routine is */
+/*#              very sensitive to compiler changes in register allocation. */
+/*#              Should really be rewritten to avoid this problem.          */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# HISTORY                                                                 */
+/*#                                                                         */
+/*# DATE      NAME            CHANGES                                       */
+/*# ----      ----            -------                                       */
+/*# 941007    Kenny R         Creation                                      */
+/*# 941011    Kenny R         Lots of optimizations and inlining.           */
+/*# 941129    Ulf A           Adapted for use in libc.                      */
+/*# 950216    HP              N==0 forgotten if non-aligned src/dst.        */
+/*#                           Added some optimizations.                     */
+/*# 001025    HP              Make src and dst char *.  Align dst to	    */
+/*#			      dword, not just word-if-both-src-and-dst-	    */
+/*#			      are-misaligned.				    */
+/*# 070806    RW              Modified for uClibc                           */
+/*#                           (__arch_v32 -> __CONFIG_CRISV32__,            */
+/*#                           include features.h to reach it.)              */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+
+#include <features.h>
+
+#ifdef __CONFIG_CRISV32__
+/* For CRISv32, movem is very cheap.  */
+#define MEMCPY_BLOCK_THRESHOLD (44)
+#else
+/* Break even between movem and move16 is at 38.7*2, but modulo 44. */
+#define MEMCPY_BLOCK_THRESHOLD (44*2)
+#endif
+
+void *memcpy(void *, const void *, unsigned int);
+
+libc_hidden_proto(memcpy)
+void *memcpy(void *pdst,
+             const void *psrc,
+             unsigned int pn)
+{
+  /* Ok.  Now we want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this.
+      As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop).
+
+     If gcc was allright, it really would need no temporaries, and no
+     stack space to save stuff on. */
+
+#ifndef MEMPCPY
+  register void *return_dst __asm__ ("r10") = pdst;
+#else
+  /* FIXME: Use R10 for something.  */
+# define return_dst dst
+#endif
+
+  register char *dst __asm__ ("r13") = pdst;
+  register char *src __asm__ ("r11") = (char *) psrc;
+  register int n __asm__ ("r12") = pn;
+  
+ 
+  /* When src is aligned but not dst, this makes a few extra needless
+     cycles.  I believe it would take as many to check that the
+     re-alignment was unnecessary.  */
+  if (((unsigned long) dst & 3) != 0
+      /* Don't align if we wouldn't copy more than a few bytes; so we
+	 don't have to check further for overflows.  */
+      && n >= 3)
+  {
+    if ((unsigned long) dst & 1)
+    {
+      n--;
+      *(char*)dst = *(char*)src;
+      src++;
+      dst++;
+    }
+
+    if ((unsigned long) dst & 2)
+    {
+      n -= 2;
+      *(short*)dst = *(short*)src;
+      src += 2;
+      dst += 2;
+    }
+  }
+
+  /* Decide which copying method to use. */
+  if (n >= MEMCPY_BLOCK_THRESHOLD)
+  {
+    /* For large copies we use 'movem' */
+
+  /* It is not optimal to tell the compiler about clobbering any
+     registers; that will move the saving/restoring of those registers
+     to the function prologue/epilogue, and make non-movem sizes
+     suboptimal.
+
+      This method is not foolproof; it assumes that the "register asm"
+     declarations at the beginning of the function really are used
+     here (beware: they may be moved to temporary registers).
+      This way, we do not have to save/move the registers around into
+     temporaries; we can safely use them straight away.  */
+    __asm__ volatile ("\
+	.syntax no_register_prefix					\n\
+									\n\
+        ;; Check that the register asm declaration got right.		\n\
+        ;; The GCC manual explicitly says TRT will happen.		\n\
+	.ifnc %0-%1-%2,$r13-$r11-$r12					\n\
+	.err								\n\
+	.endif								\n\
+									\n\
+	;; Save the registers we'll use in the movem process		\n\
+	;; on the stack.						\n\
+	subq 	11*4,sp							\n\
+	movem	r10,[sp]						\n\
+									\n\
+        ;; Now we've got this:						\n\
+	;; r11 - src							\n\
+	;; r13 - dst							\n\
+	;; r12 - n							\n\
+									\n\
+        ;; Update n for the first loop					\n\
+        subq    44,r12							\n\
+0:									\n\
+	movem	[r11+],r10						\n\
+        subq   44,r12							\n\
+        bge     0b							\n\
+	movem	r10,[r13+]						\n\
+									\n\
+        addq   44,r12  ;; compensate for last loop underflowing n	\n\
+									\n\
+	;; Restore registers from stack					\n\
+        movem [sp+],r10"
+
+     /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n) 
+     /* Inputs */ : "0" (dst), "1" (src), "2" (n));
+  }
+
+  /* Either we directly starts copying, using dword copying
+     in a loop, or we copy as much as possible with 'movem' 
+     and then the last block (<44 bytes) is copied here.
+     This will work since 'movem' will have updated src,dst,n. */
+
+  while ( n >= 16 )
+  {
+    *((long*)dst)++ = *((long*)src)++;
+    *((long*)dst)++ = *((long*)src)++;
+    *((long*)dst)++ = *((long*)src)++;
+    *((long*)dst)++ = *((long*)src)++;
+    n -= 16;
+  }
+
+  /* A switch() is definitely the fastest although it takes a LOT of code.
+   * Particularly if you inline code this.
+   */
+  switch (n)
+  {
+    case 0:
+      break;
+    case 1:
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+    case 2:
+      *((short*)dst)++ = *((short*)src)++;
+      break;
+    case 3:
+      *((short*)dst)++ = *((short*)src)++;
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+    case 4:
+      *((long*)dst)++ = *((long*)src)++;
+      break;
+    case 5:
+      *((long*)dst)++ = *((long*)src)++;
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+    case 6:
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      break;
+    case 7:
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+    case 8:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      break;
+    case 9:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+    case 10:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      break;
+    case 11:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+    case 12:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      break;
+    case 13:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+    case 14:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      break;
+    case 15:
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((long*)dst)++ = *((long*)src)++;
+      *((short*)dst)++ = *((short*)src)++;
+      *((char*)dst)++ = *((char*)src)++;
+      break;
+  }
+
+  return return_dst; /* destination pointer. */
+} /* memcpy() */
+libc_hidden_def(memcpy)
diff --git a/libc/string/cris/memmove.c b/libc/string/cris/memmove.c
new file mode 100644
index 000000000..437637078
--- /dev/null
+++ b/libc/string/cris/memmove.c
@@ -0,0 +1,101 @@
+/* Taken from generic/memmove.c; trivially modified to work with 
+   arch-specific memcopy.h for Cris.
+
+   Copy memory to memory until the specified number of bytes
+   has been copied.  Overlap is handled correctly.
+   Copyright (C) 1991, 1995, 1996, 1997, 2003 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+   Contributed by Torbjorn Granlund (tege@sics.se).
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, write to the Free
+   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+   02111-1307 USA.  */
+
+#include <string.h>
+
+#include "memcopy.h"
+#include "../generic/pagecopy.h"
+
+libc_hidden_proto(memmove)
+void *memmove (void *dest, const void *src, size_t len)
+{
+  unsigned long int dstp = (long int) dest;
+  unsigned long int srcp = (long int) src;
+
+  /* This test makes the forward copying code be used whenever possible.
+     Reduces the working set.  */
+  if (dstp - srcp >= len)	/* *Unsigned* compare!  */
+    {
+#if 1
+#warning REMINDER: Cris arch-opt memmove assumes memcpy does forward copying!
+      memcpy(dest, src, len);
+#else
+      /* Copy from the beginning to the end.  */
+
+      /* If there not too few bytes to copy, use word copy.  */
+      if (len >= OP_T_THRES)
+	{
+	  /* Copy just a few bytes to make DSTP aligned.  */
+	  len -= (-dstp) % OPSIZ;
+	  BYTE_COPY_FWD (dstp, srcp, (-dstp) % OPSIZ);
+
+	  /* Copy whole pages from SRCP to DSTP by virtual address
+	     manipulation, as much as possible.  */
+
+	  PAGE_COPY_FWD_MAYBE (dstp, srcp, len, len);
+
+	  /* Copy from SRCP to DSTP taking advantage of the known
+	     alignment of DSTP.  Number of bytes remaining is put
+	     in the third argument, i.e. in LEN.  This number may
+	     vary from machine to machine.  */
+
+	  WORD_COPY_FWD (dstp, srcp, len, len);
+
+	  /* Fall out and copy the tail.  */
+	}
+
+      /* There are just a few bytes to copy.  Use byte memory operations.  */
+      BYTE_COPY_FWD (dstp, srcp, len);
+#endif
+    }
+  else
+    {
+      /* Copy from the end to the beginning.  */
+      srcp += len;
+      dstp += len;
+
+      /* If there not too few bytes to copy, use word copy.  */
+      if (len >= OP_T_THRES)
+	{
+	  /* Copy just a few bytes to make DSTP aligned.  */
+	  len -= dstp % OPSIZ;
+	  BYTE_COPY_BWD (dstp, srcp, dstp % OPSIZ);
+
+	  /* Copy from SRCP to DSTP taking advantage of the known
+	     alignment of DSTP.  Number of bytes remaining is put
+	     in the third argument, i.e. in LEN.  This number may
+	     vary from machine to machine.  */
+
+	  WORD_COPY_BWD (dstp, srcp, len, len);
+
+	  /* Fall out and copy the tail.  */
+	}
+
+      /* There are just a few bytes to copy.  Use byte memory operations.  */
+      BYTE_COPY_BWD (dstp, srcp, len);
+    }
+
+  return (dest);
+}
+libc_hidden_def(memmove)
diff --git a/libc/string/cris/memset.c b/libc/string/cris/memset.c
new file mode 100644
index 000000000..7e71bc50f
--- /dev/null
+++ b/libc/string/cris/memset.c
@@ -0,0 +1,271 @@
+/* Copyright (C) 2001, 2003 Free Software Foundation, Inc.
+   Copyright (C) 1999, 2000 Axis Communications AB.
+
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Library General Public License as
+   published by the Free Software Foundation; either version 2 of the
+   License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Library General Public License for more details.
+
+   You should have received a copy of the GNU Library General Public
+   License along with the GNU C Library; see the file COPYING.LIB.  If not,
+   write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+/*#************************************************************************#*/
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# FUNCTION NAME: memset()                                                 */
+/*#                                                                         */
+/*# PARAMETERS:  void* dst;   Destination address.                          */
+/*#              int     c;   Value of byte to write.                       */
+/*#              int   len;   Number of bytes to write.                     */
+/*#                                                                         */
+/*# RETURNS:     dst.                                                       */
+/*#                                                                         */
+/*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */
+/*#              Framework taken from memcpy.  This routine is              */
+/*#              very sensitive to compiler changes in register allocation. */
+/*#              Should really be rewritten to avoid this problem.          */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+/*#                                                                         */
+/*# HISTORY                                                                 */
+/*#                                                                         */
+/*# DATE      NAME            CHANGES                                       */
+/*# ----      ----            -------                                       */
+/*# 990713    HP              Tired of watching this function (or           */
+/*#                           really, the nonoptimized generic              */
+/*#                           implementation) take up 90% of simulator      */
+/*#                           output.  Measurements needed.                 */
+/*#                                                                         */
+/*#-------------------------------------------------------------------------*/
+
+/* No, there's no macro saying 12*4, since it is "hard" to get it into
+   the asm in a good way.  Thus better to expose the problem everywhere.
+   */
+
+/* Assuming 1 cycle per dword written or read (ok, not really true), and
+   one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1)
+   so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */
+
+#define ZERO_BLOCK_SIZE (1*12*4)
+
+void *memset(void *, int, unsigned long);
+
+libc_hidden_proto(memset)
+void *memset(void *pdst,
+             int c,
+             unsigned long plen)
+{
+  /* Ok.  Now we want the parameters put in special registers.
+     Make sure the compiler is able to make something useful of this. */
+
+  register char *return_dst __asm__ ("r10") = pdst;
+  register long n __asm__ ("r12") = plen;
+  register int lc __asm__ ("r11") = c;
+
+  /* Most apps use memset sanely.  Only those memsetting about 3..4
+     bytes or less get penalized compared to the generic implementation
+     - and that's not really sane use. */
+
+  /* Ugh.  This is fragile at best.  Check with newer GCC releases, if
+     they compile cascaded "x |= x << 8" sanely! */
+  __asm__("movu.b %0,$r13 \n\
+	   lslq 8,$r13    \n\
+	   move.b %0,$r13 \n\
+	   move.d $r13,%0 \n\
+	   lslq 16,$r13   \n\
+	   or.d $r13,%0"
+          : "=r" (lc) : "0" (lc) : "r13");
+
+  {
+    register char *dst __asm__ ("r13") = pdst;
+ 
+  if (((unsigned long) pdst & 3) != 0
+     /* Oops! n=0 must be a legal call, regardless of alignment. */
+      && n >= 3)
+  {
+    if ((unsigned long)dst & 1)
+    {
+      *dst = (char) lc;
+      n--;
+      dst++;
+    }
+
+    if ((unsigned long)dst & 2)
+    {
+      *(short *)dst = lc;
+      n -= 2;
+      dst += 2;
+    }
+  }
+
+  /* Now the fun part.  For the threshold value of this, check the equation
+     above. */
+  /* Decide which copying method to use. */
+  if (n >= ZERO_BLOCK_SIZE)
+  {
+    /* For large copies we use 'movem' */
+
+  /* It is not optimal to tell the compiler about clobbering any
+     registers; that will move the saving/restoring of those registers
+     to the function prologue/epilogue, and make non-movem sizes
+     suboptimal.
+
+      This method is not foolproof; it assumes that the "asm reg"
+     declarations at the beginning of the function really are used
+     here (beware: they may be moved to temporary registers).
+      This way, we do not have to save/move the registers around into
+     temporaries; we can safely use them straight away.  */
+    __asm__ volatile ("								\n\
+	.syntax no_register_prefix						\n\
+										\n\
+        ;; Check that the register asm declaration got right.			\n\
+        ;; The GCC manual explicitly says there's no warranty for that (too).	\n\
+	.ifnc %0-%1-%4,$r13-$r12-$r11						\n\
+	.err									\n\
+	.endif									\n\
+										\n\
+	;; Save the registers we'll clobber in the movem process		\n\
+	;; on the stack.  Don't mention them to gcc, it will only be		\n\
+	;; upset.								\n\
+	subq 	11*4,sp								\n\
+        movem   r10,[sp]							\n\
+										\n\
+        move.d  r11,r0								\n\
+        move.d  r11,r1								\n\
+        move.d  r11,r2								\n\
+        move.d  r11,r3								\n\
+        move.d  r11,r4								\n\
+        move.d  r11,r5								\n\
+        move.d  r11,r6								\n\
+        move.d  r11,r7								\n\
+        move.d  r11,r8								\n\
+        move.d  r11,r9								\n\
+        move.d  r11,r10								\n\
+										\n\
+        ;; Now we've got this:							\n\
+	;; r13 - dst								\n\
+	;; r12 - n								\n\
+										\n\
+        ;; Update n for the first loop						\n\
+        subq    12*4,r12							\n\
+0:										\n\
+        subq   12*4,r12								\n\
+        bge     0b								\n\
+	movem	r11,[r13+]							\n\
+										\n\
+        addq   12*4,r12  ;; compensate for last loop underflowing n		\n\
+										\n\
+	;; Restore registers from stack						\n\
+        movem [sp+],r10" 
+
+     /* Outputs */ : "=r" (dst), "=r" (n)
+     /* Inputs */ : "0" (dst), "1" (n), "r" (lc));
+    
+  }
+
+    /* Either we directly starts copying, using dword copying
+       in a loop, or we copy as much as possible with 'movem' 
+       and then the last block (<44 bytes) is copied here.
+       This will work since 'movem' will have updated src,dst,n. */
+
+    while ( n >= 16 )
+    {
+      *((long*)dst)++ = lc;
+      *((long*)dst)++ = lc;
+      *((long*)dst)++ = lc;
+      *((long*)dst)++ = lc;
+      n -= 16;
+    }
+
+    /* A switch() is definitely the fastest although it takes a LOT of code.
+     * Particularly if you inline code this.
+     */
+    switch (n)
+    {
+      case 0:
+        break;
+      case 1:
+        *(char*)dst = (char) lc;
+        break;
+      case 2:
+        *(short*)dst = (short) lc;
+        break;
+      case 3:
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 4:
+        *((long*)dst)++ = lc;
+        break;
+      case 5:
+        *((long*)dst)++ = lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 6:
+        *((long*)dst)++ = lc;
+        *(short*)dst = (short) lc;
+        break;
+      case 7:
+        *((long*)dst)++ = lc;
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 8:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        break;
+      case 9:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 10:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(short*)dst = (short) lc;
+        break;
+      case 11:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 12:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        break;
+      case 13:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(char*)dst = (char) lc;
+        break;
+      case 14:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *(short*)dst = (short) lc;
+        break;
+      case 15:
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((long*)dst)++ = lc;
+        *((short*)dst)++ = (short) lc;
+        *(char*)dst = (char) lc;
+        break;
+    }
+  }
+
+  return return_dst; /* destination pointer. */
+} /* memset() */
+libc_hidden_def(memset)