sh: Optimised memset for SH4

Optimised version of memset for the SH4 which uses movca.l.

Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>

2 files changed, 112 insertions, 2 deletions

diff --git a/arch/sh/lib/Makefile b/arch/sh/lib/Makefile
index a969b47c5463..dab4d2129812 100644
--- a/arch/sh/lib/Makefile
+++ b/arch/sh/lib/Makefile
@@ -2,7 +2,7 @@
 # Makefile for SuperH-specific library files..
 #
 
-lib-y  = delay.o memset.o memmove.o memchr.o \
+lib-y  = delay.o memmove.o memchr.o \
          checksum.o strlen.o div64.o div64-generic.o
 
 # Extracted from libgcc
@@ -23,8 +23,11 @@ obj-y				+= io.o
 memcpy-y                        := memcpy.o
 memcpy-$(CONFIG_CPU_SH4)        := memcpy-sh4.o
 
+memset-y                        := memset.o
+memset-$(CONFIG_CPU_SH4)        := memset-sh4.o
+
 lib-$(CONFIG_MMU)               += copy_page.o __clear_user.o
 lib-$(CONFIG_MCOUNT)            += mcount.o
-lib-y                           += $(memcpy-y) $(udivsi3-y)
+lib-y                           += $(memcpy-y) $(memset-y) $(udivsi3-y)
 
 EXTRA_CFLAGS += -Werror
diff --git a/arch/sh/lib/memset-sh4.S b/arch/sh/lib/memset-sh4.S
new file mode 100644
index 000000000000..1a6e32cc4e4d
--- /dev/null
+++ b/arch/sh/lib/memset-sh4.S
@@ -0,0 +1,107 @@
+/*
+ * "memset" implementation for SH4
+ *
+ * Copyright (C) 1999  Niibe Yutaka
+ * Copyright (c) 2009  STMicroelectronics Limited
+ * Author: Stuart Menefy <stuart.menefy:st.com>
+ */
+
+/*
+ *            void *memset(void *s, int c, size_t n);
+ */
+
+#include <linux/linkage.h>
+
+ENTRY(memset)
+        mov     #12,r0
+        add     r6,r4
+        cmp/gt  r6,r0
+        bt/s    40f             ! if it's too small, set a byte at once
+         mov    r4,r0
+        and     #3,r0
+        cmp/eq  #0,r0
+        bt/s    2f              ! It's aligned
+         sub    r0,r6
+1:
+        dt      r0
+        bf/s    1b
+         mov.b  r5,@-r4
+2:                              ! make VVVV
+        extu.b  r5,r5
+        swap.b  r5,r0           !   V0
+        or      r0,r5           !   VV
+        swap.w  r5,r0           ! VV00
+        or      r0,r5           ! VVVV
+
+        ! Check if enough bytes need to be copied to be worth the big loop
+        mov     #0x40, r0       ! (MT)
+        cmp/gt  r6,r0           ! (MT)  64 > len => slow loop
+
+        bt/s    22f
+         mov    r6,r0
+
+        ! align the dst to the cache block size if necessary
+        mov     r4, r3
+        mov     #~(0x1f), r1
+
+        and     r3, r1
+        cmp/eq  r3, r1
+
+        bt/s    11f             ! dst is already aligned
+         sub    r1, r3          ! r3-r1 -> r3
+        shlr2   r3              ! number of loops
+
+10:     mov.l   r5,@-r4
+        dt      r3
+        bf/s    10b
+         add    #-4, r6
+
+11:     ! dst is 32byte aligned
+        mov     r6,r2
+        mov     #-5,r0
+        shld    r0,r2           ! number of loops
+
+        add     #-32, r4
+        mov     r5, r0
+12:
+        movca.l r0,@r4
+        mov.l   r5,@(4, r4)
+        mov.l   r5,@(8, r4)
+        mov.l   r5,@(12,r4)
+        mov.l   r5,@(16,r4)
+        mov.l   r5,@(20,r4)
+        add     #-0x20, r6
+        mov.l   r5,@(24,r4)
+        dt      r2
+        mov.l   r5,@(28,r4)
+        bf/s    12b
+         add    #-32, r4
+
+        add     #32, r4
+        mov     #8, r0
+        cmp/ge  r0, r6
+        bf      40f
+
+        mov     r6,r0
+22:
+        shlr2   r0
+        shlr    r0              ! r0 = r6 >> 3
+3:
+        dt      r0
+        mov.l   r5,@-r4         ! set 8-byte at once
+        bf/s    3b
+         mov.l  r5,@-r4
+        !
+        mov     #7,r0
+        and     r0,r6
+
+        ! fill bytes (length may be zero)
+40:     tst     r6,r6
+        bt      5f
+4:
+        dt      r6
+        bf/s    4b
+         mov.b  r5,@-r4
+5:
+        rts
+         mov    r4,r0