1 files changed, 138 insertions, 0 deletions
diff --git a/arch/sparc64/lib/blockops.S b/arch/sparc64/lib/blockops.S
new file mode 100644
index 000000000..b3f06c18d
--- /dev/null
+++ b/arch/sparc64/lib/blockops.S
@@ -0,0 +1,138 @@
+/* $Id: blockops.S,v 1.5 1997/03/26 18:34:28 jj Exp $
+ * arch/sparc64/lib/blockops.S: UltraSparc block zero optimized routines.
+ *
+ * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
+ * Copyright (C) 1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
+ */
+
+#include <asm/asi.h>
+
+	/* Zero out 256 bytes of memory at (buf + offset). */
+#define BLAST_BLOCK(buf, offset)				\
+	stda		%f48, [buf + offset + 0x00] %asi;	\
+	stda		%f48, [buf + offset + 0x40] %asi;	\
+	stda		%f48, [buf + offset + 0x80] %asi;	\
+	stda		%f48, [buf + offset + 0xc0] %asi;
+
+	/* Copy 256 bytes of memory at (src + offset) to
+	 * (dst + offset).
+	 */
+#define MIRROR_BLOCK(dst, src, offset, sync)			\
+	ldda		[src + offset + 0x000] %asi, %f0;	\
+	ldda		[src + offset + 0x040] %asi, %f16;	\
+	ldda		[src + offset + 0x080] %asi, %f32;	\
+	ldda		[src + offset + 0x0c0] %asi, %f48;	\
+	membar		sync;					\
+	stda		%f0, [dst + offset + 0x000] %asi;	\
+	stda		%f16, [dst + offset + 0x040] %asi;	\
+	stda		%f32, [dst + offset + 0x080] %asi;	\
+	stda		%f48, [dst + offset + 0x0c0] %asi;
+
+	.text
+	.align	4
+
+	.globl		bzero_2page, bzero_1page
+bzero_2page:
+	/* %o0 = buf */
+	mov		%o0, %o1
+	wr		%g0, ASI_BLK_P, %asi
+	mov		0x10, %g2
+
+	membar		#Sync|#StoreLoad
+
+	fzero		%f48
+	fzero		%f50
+	fzero		%f52
+	fzero		%f54
+	fzero		%f56
+	fzero		%f58
+	fzero		%f60
+	fzero		%f62
+1:
+	BLAST_BLOCK(%o0, 0x000)
+	BLAST_BLOCK(%o0, 0x100)
+	BLAST_BLOCK(%o0, 0x200)
+	BLAST_BLOCK(%o0, 0x300)
+	subcc		%g2, 1, %g2
+	bne,pt		%icc, 1b
+	 add		%o0, 0x400, %o0
+
+	membar		#Sync|#LoadStore|#StoreStore
+
+	retl
+	 mov		%o1, %o0
+
+bzero_1page:
+	/* %o0 = buf */
+	mov		%o0, %o1
+	wr		%g0, ASI_BLK_P, %asi
+	mov		0x08, %g2
+	membar		#Sync|#StoreLoad
+	fzero		%f48
+	fzero		%f50
+	fzero		%f52
+	fzero		%f54
+	fzero		%f56
+	fzero		%f58
+	fzero		%f60
+	fzero		%f62
+1:
+	BLAST_BLOCK(%o0, 0x000)
+	BLAST_BLOCK(%o0, 0x100)
+	BLAST_BLOCK(%o0, 0x200)
+	BLAST_BLOCK(%o0, 0x300)
+	subcc		%g2, 1, %g2
+	bne,pt		%icc, 1b
+	 add		%o0, 0x400, %o0
+
+	membar		#Sync|#LoadStore|#StoreStore
+
+	retl
+	 mov		%o1, %o0
+
+	.globl		__bfill64
+__bfill64:
+	/* %o0 = buf */
+	stx		%o1, [%sp + 0x7ff + 128]
+	wr		%g0, ASI_BLK_P, %asi
+	mov		0x08, %g2
+	ldd		[%sp + 0x7ff + 128], %f48
+	membar		#Sync|#StoreLoad
+	fmovd		%f48, %f50
+	fmovd		%f48, %f52
+	fmovd		%f48, %f54
+	fmovd		%f48, %f56
+	fmovd		%f48, %f58
+	fmovd		%f48, %f60
+	fmovd		%f48, %f62
+1:
+	BLAST_BLOCK(%o0, 0x000)
+	BLAST_BLOCK(%o0, 0x100)
+	BLAST_BLOCK(%o0, 0x200)
+	BLAST_BLOCK(%o0, 0x300)
+	subcc		%g2, 1, %g2
+	bne,pt		%icc, 1b
+	 add		%o0, 0x400, %o0
+
+	retl
+	 membar		#Sync|#LoadStore|#StoreStore
+
+	.globl		__copy_1page
+__copy_1page:
+	/* %o0 = dst, %o1 = src */
+	or		%g0, 0x08, %g1
+	wr		%g0, ASI_BLK_P, %asi
+	membar		#Sync|#StoreLoad
+1:
+	MIRROR_BLOCK(%o0, %o1, 0x000, #Sync)
+	MIRROR_BLOCK(%o0, %o1, 0x100, #Sync)
+	MIRROR_BLOCK(%o0, %o1, 0x200, #Sync)
+	MIRROR_BLOCK(%o0, %o1, 0x300, #Sync)
+	subcc		%g1, 1, %g1
+	add		%o0, 0x400, %o0
+	bne,pt		%icc, 1b
+	 add		%o1, 0x400, %o1
+
+	retl
+	 membar		#Sync|#LoadStore|#StoreStore
+