sh: Add SH-5 optimized memcpy()/memset()/strcpy()/strlen().
Adopted from the uClibc optimized string versions.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
diff --git a/arch/sh/lib64/strcpy.S b/arch/sh/lib64/strcpy.S
new file mode 100644
index 0000000..ea7c9c5
--- /dev/null
+++ b/arch/sh/lib64/strcpy.S
@@ -0,0 +1,97 @@
+/* Cloned and hacked for uClibc by Paul Mundt, December 2003 */
+/* Modified by SuperH, Inc. September 2003 */
+! Entry: arg0: destination
+! arg1: source
+! Exit: result: destination
+!
+! SH5 code Copyright 2002 SuperH Ltd.
+
+#if __BYTE_ORDER == __LITTLE_ENDIAN
+#define SHHI shlld
+#define SHLO shlrd
+#else
+#define SHHI shlrd
+#define SHLO shlld
+#endif
+
+ .section .text..SHmedia32,"ax"
+ .globl strcpy
+ .type strcpy, @function
+ .align 5
+
+strcpy:
+
+ pta/l shortstring,tr1
+ ldlo.q r3,0,r4
+ ptabs r18,tr4
+ shlli r3,3,r7
+ addi r2, 8, r0
+ mcmpeq.b r4,r63,r6
+ SHHI r6,r7,r6
+ bnei/u r6,0,tr1 // shortstring
+ pta/l no_lddst, tr2
+ ori r3,-8,r23
+ sub r2, r23, r0
+ sub r3, r2, r21
+ addi r21, 8, r20
+ ldx.q r0, r21, r5
+ pta/l loop, tr0
+ ori r2,-8,r22
+ mcmpeq.b r5, r63, r6
+ bgt/u r22, r23, tr2 // no_lddst
+
+ // r22 < r23 : Need to do a load from the destination.
+ // r22 == r23 : Doesn't actually need to load from destination,
+ // but still can be handled here.
+ ldlo.q r2, 0, r9
+ movi -1, r8
+ SHLO r8, r7, r8
+ mcmv r4, r8, r9
+ stlo.q r2, 0, r9
+ beqi/l r6, 0, tr0 // loop
+
+ add r5, r63, r4
+ addi r0, 8, r0
+ blink tr1, r63 // shortstring
+no_lddst:
+ // r22 > r23: note that for r22 == r23 the sthi.q would clobber
+ // bytes before the destination region.
+ stlo.q r2, 0, r4
+ SHHI r4, r7, r4
+ sthi.q r0, -1, r4
+ beqi/l r6, 0, tr0 // loop
+
+ add r5, r63, r4
+ addi r0, 8, r0
+shortstring:
+#if __BYTE_ORDER != __LITTLE_ENDIAN
+ pta/l shortstring2,tr1
+ byterev r4,r4
+#endif
+shortstring2:
+ st.b r0,-8,r4
+ andi r4,0xff,r5
+ shlri r4,8,r4
+ addi r0,1,r0
+ bnei/l r5,0,tr1
+ blink tr4,r63 // return
+
+ .balign 8
+loop:
+ stlo.q r0, 0, r5
+ ldx.q r0, r20, r4
+ addi r0, 16, r0
+ sthi.q r0, -9, r5
+ mcmpeq.b r4, r63, r6
+ bnei/u r6, 0, tr1 // shortstring
+ ldx.q r0, r21, r5
+ stlo.q r0, -8, r4
+ sthi.q r0, -1, r4
+ mcmpeq.b r5, r63, r6
+ beqi/l r6, 0, tr0 // loop
+
+ add r5, r63, r4
+ addi r0, 8, r0
+ blink tr1, r63 // shortstring
+
+ .size strcpy,.-strcpy