aboutsummaryrefslogtreecommitdiff
path: root/arch/cris/arch-v10/lib/memset.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 15:20:36 -0700
committerLinus Torvalds <torvalds@ppc970.osdl.org>2005-04-16 15:20:36 -0700
commit1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 (patch)
tree0bba044c4ce775e45a88a51686b5d9f90697ea9d /arch/cris/arch-v10/lib/memset.c
downloadlinux-stericsson-1da177e4c3f41524e886b7f1b8a0c1fc7321cac2.tar.gz
Linux-2.6.12-rc2v2.6.12-rc2
Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip!
Diffstat (limited to 'arch/cris/arch-v10/lib/memset.c')
-rw-r--r--arch/cris/arch-v10/lib/memset.c252
1 files changed, 252 insertions, 0 deletions
diff --git a/arch/cris/arch-v10/lib/memset.c b/arch/cris/arch-v10/lib/memset.c
new file mode 100644
index 000000000000..82bb66839171
--- /dev/null
+++ b/arch/cris/arch-v10/lib/memset.c
@@ -0,0 +1,252 @@
+/*#************************************************************************#*/
+/*#-------------------------------------------------------------------------*/
+/*# */
+/*# 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. */
+/*# */
+/*#-------------------------------------------------------------------------*/
+
+#include <linux/types.h>
+
+/* 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 *pdst,
+ int c,
+ size_t 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 int 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,$r13\n\t"
+ "lslq 8,$r13\n\t"
+ "move.b %0,$r13\n\t"
+ "move.d $r13,%0\n\t"
+ "lslq 16,$r13\n\t"
+ "or.d $r13,%0"
+ : "=r" (lc) : "0" (lc) : "r13");
+
+ {
+ register char *dst __asm__ ("r13") = pdst;
+
+ /* This is NONPORTABLE, but since this whole routine is */
+ /* grossly nonportable that doesn't matter. */
+
+ if (((unsigned long) pdst & 3) != 0
+ /* Oops! n=0 must be a legal call, regardless of alignment. */
+ && n >= 3)
+ {
+ if ((unsigned long)dst & 1)
+ {
+ *dst = (char) lc;
+ n--;
+ dst++;
+ }
+
+ if ((unsigned long)dst & 2)
+ {
+ *(short *)dst = lc;
+ n -= 2;
+ dst += 2;
+ }
+ }
+
+ /* Now the fun part. For the threshold value of this, check the equation
+ above. */
+ /* Decide which copying method to use. */
+ if (n >= ZERO_BLOCK_SIZE)
+ {
+ /* For large copies we use 'movem' */
+
+ /* It is not optimal to tell the compiler about clobbering any
+ registers; that will move the saving/restoring of those registers
+ to the function prologue/epilogue, and make non-movem sizes
+ suboptimal.
+
+ This method is not foolproof; it assumes that the "asm reg"
+ declarations at the beginning of the function really are used
+ here (beware: they may be moved to temporary registers).
+ This way, we do not have to save/move the registers around into
+ temporaries; we can safely use them straight away.
+
+ If you want to check that the allocation was right; then
+ check the equalities in the first comment. It should say
+ "r13=r13, r12=r12, r11=r11" */
+ __asm__ volatile ("
+ ;; Check that the following is true (same register names on
+ ;; both sides of equal sign, as in r8=r8):
+ ;; %0=r13, %1=r12, %4=r11
+ ;;
+ ;; Save the registers we'll clobber in the movem process
+ ;; on the stack. Don't mention them to gcc, it will only be
+ ;; upset.
+ subq 11*4,$sp
+ movem $r10,[$sp]
+
+ move.d $r11,$r0
+ move.d $r11,$r1
+ move.d $r11,$r2
+ move.d $r11,$r3
+ move.d $r11,$r4
+ move.d $r11,$r5
+ move.d $r11,$r6
+ move.d $r11,$r7
+ move.d $r11,$r8
+ move.d $r11,$r9
+ move.d $r11,$r10
+
+ ;; Now we've got this:
+ ;; r13 - dst
+ ;; r12 - n
+
+ ;; Update n for the first loop
+ subq 12*4,$r12
+0:
+ subq 12*4,$r12
+ bge 0b
+ movem $r11,[$r13+]
+
+ addq 12*4,$r12 ;; compensate for last loop underflowing n
+
+ ;; Restore registers from stack
+ movem [$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() */