5 files changed, 143 insertions, 6 deletions
diff --git a/arch/sh/include/asm/fpu.h b/arch/sh/include/asm/fpu.h
index 1d3aee0..bfd78e1 100644
--- a/arch/sh/include/asm/fpu.h
+++ b/arch/sh/include/asm/fpu.h
@@ -19,6 +19,7 @@ static inline void grab_fpu(struct pt_regs *regs)
 struct task_struct;
 
 extern void save_fpu(struct task_struct *__tsk, struct pt_regs *regs);
+void fpu_state_restore(struct pt_regs *regs);
 #else
 
 #define release_fpu(regs)	do { } while (0)
@@ -44,6 +45,8 @@ static inline void unlazy_fpu(struct task_struct *tsk, struct pt_regs *regs)
 	preempt_disable();
 	if (test_tsk_thread_flag(tsk, TIF_USEDFPU))
 		save_fpu(tsk, regs);
+	else
+		tsk->fpu_counter = 0;
 	preempt_enable();
 }
 
diff --git a/arch/sh/kernel/cpu/sh4/fpu.c b/arch/sh/kernel/cpu/sh4/fpu.c
index e3ea541..d79226f 100644
--- a/arch/sh/kernel/cpu/sh4/fpu.c
+++ b/arch/sh/kernel/cpu/sh4/fpu.c
@@ -483,18 +483,18 @@ BUILD_TRAP_HANDLER(fpu_error)
 	force_sig(SIGFPE, tsk);
 }
 
-BUILD_TRAP_HANDLER(fpu_state_restore)
+void fpu_state_restore(struct pt_regs *regs)
 {
 	struct task_struct *tsk = current;
-	TRAP_HANDLER_DECL;
 
 	grab_fpu(regs);
-	if (!user_mode(regs)) {
+	if (unlikely(!user_mode(regs))) {
 		printk(KERN_ERR "BUG: FPU is used in kernel mode.\n");
+		BUG();
 		return;
 	}
 
-	if (used_math()) {
+	if (likely(used_math())) {
 		/* Using the FPU again.  */
 		restore_fpu(tsk);
 	} else {
@@ -503,4 +503,12 @@ BUILD_TRAP_HANDLER(fpu_state_restore)
 		set_used_math();
 	}
 	set_tsk_thread_flag(tsk, TIF_USEDFPU);
+	tsk->fpu_counter++;
+}
+
+BUILD_TRAP_HANDLER(fpu_state_restore)
+{
+	TRAP_HANDLER_DECL;
+
+	fpu_state_restore(regs);
 }
diff --git a/arch/sh/kernel/process_32.c b/arch/sh/kernel/process_32.c
index a40342b..7733f5f 100644
--- a/arch/sh/kernel/process_32.c
+++ b/arch/sh/kernel/process_32.c
@@ -290,8 +290,14 @@ static void ubc_set_tracing(int asid, unsigned long pc)
 __notrace_funcgraph struct task_struct *
 __switch_to(struct task_struct *prev, struct task_struct *next)
 {
+	struct thread_struct *next_t = &next->thread;
+
 #if defined(CONFIG_SH_FPU)
 	unlazy_fpu(prev, task_pt_regs(prev));
+
+	/* we're going to use this soon, after a few expensive things */
+	if (next->fpu_counter > 5)
+		prefetch(&next_t->fpu.hard);
 #endif
 
 #ifdef CONFIG_MMU
@@ -323,6 +329,16 @@ __switch_to(struct task_struct *prev, struct task_struct *next)
 #endif
 	}
 
+#if defined(CONFIG_SH_FPU)
+	/* If the task has used fpu the last 5 timeslices, just do a full
+	 * restore of the math state immediately to avoid the trap; the
+	 * chances of needing FPU soon are obviously high now
+	 */
+	if (next->fpu_counter > 5) {
+		fpu_state_restore(task_pt_regs(next));
+	}
+#endif
+
 	return prev;
 }
 
diff --git a/arch/sh/lib/Makefile b/arch/sh/lib/Makefile
index a969b47..dab4d21 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 0000000..1a6e32c
--- /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