about summary refs log tree commit diff
path: root/sysdeps/alpha/memchr.S
diff options
context:
space:
mode:
Diffstat (limited to 'sysdeps/alpha/memchr.S')
-rw-r--r--sysdeps/alpha/memchr.S163
1 files changed, 163 insertions, 0 deletions
diff --git a/sysdeps/alpha/memchr.S b/sysdeps/alpha/memchr.S
new file mode 100644
index 0000000000..118a1f13d1
--- /dev/null
+++ b/sysdeps/alpha/memchr.S
@@ -0,0 +1,163 @@
+/* Copyright (C) 1996 Free Software Foundation, Inc.
+   Contributed by David Mosberger (davidm@cs.arizona.edu).
+
+This file is part of the GNU C Library.
+
+The GNU C Library is free software; you can redistribute it and/or
+modify it under the terms of the GNU Library General Public License as
+published by the Free Software Foundation; either version 2 of the
+License, or (at your option) any later version.
+
+The GNU C Library is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+Library General Public License for more details.
+
+You should have received a copy of the GNU Library General Public
+License along with the GNU C Library; see the file COPYING.LIB.  If
+not, write to the Free Software Foundation, Inc., 675 Mass Ave,
+Cambridge, MA 02139, USA.  */
+
+/* Finds characters in a memory area.  Optimized for the Alpha
+architecture:
+
+      - memory accessed as aligned quadwords only
+      - uses cmpbge to compare 8 bytes in parallel
+      - does binary search to find 0 byte in last
+        quadword (HAKMEM needed 12 instructions to
+        do this instead of the 9 instructions that
+        binary search needs).
+
+For correctness consider that:
+
+      - only minimum number of quadwords may be accessed
+      - the third argument is an unsigned long
+*/
+
+#include <sysdep.h>
+#ifdef __linux__
+# include <alpha/regdef.h>
+#else
+#include <regdef.h>
+#endif
+
+        .set noreorder
+        .set noat
+
+ENTRY(memchr)
+	beq	a2, not_found
+        ldq_u   t0, 0(a0)       # load first quadword (a0 may be misaligned)
+	addq	a0, a2, t4
+	and	a1, 0xff, a1	# a1 = 00000000000000ch
+	ldq_u	t5, -1(t4)
+	sll	a1,  8, t1	# t1 = 000000000000ch00
+	cmpult	a2, 9, t3
+	or	t1, a1, a1	# a1 = 000000000000chch
+	sll	a1, 16, t1	# t1 = 00000000chch0000
+        lda     t2, -1(zero)
+	or	t1, a1, a1	# a1 = 00000000chchchch
+	sll	a1, 32, t1	# t1 = chchchch00000000
+	extql	t0, a0, t6
+	or	t1, a1, a1	# a1 = chchchchchchchch
+
+	beq	t3, first_quad
+
+	extqh	t5, a0, t5
+	mov	a0, v0
+	or	t6, t5, t0	# t0 = quadword starting at a0
+
+	#
+	# Deal with the case where at most 8 bytes remain to be searched
+	# in t0.  E.g.:
+	#	a2 = 6
+	#	t0 = ????c6c5c4c3c2c1
+last_quad:
+	negq	a2, t5
+	srl	t2, t5, t5	# t5 = mask of a2 bits set
+        xor	a1, t0, t0
+        cmpbge  zero, t0, t1
+	and	t1, t5, t1
+        beq     t1, not_found
+
+found_it:
+	# now, determine which byte matched:
+        negq    t1, t2
+        and     t1, t2, t1
+
+        and     t1, 0x0f, t0
+        addq    v0, 4, t2
+        cmoveq  t0, t2, v0
+
+        and     t1, 0x33, t0
+        addq    v0, 2, t2
+        cmoveq  t0, t2, v0
+
+        and     t1, 0x55, t0
+        addq    v0, 1, t2
+        cmoveq  t0, t2, v0
+
+done:	ret
+
+
+	#
+	# Deal with the case where a2 > 8 bytes remain to be
+	# searched.  a0 may not be aligned.
+	#
+first_quad:
+	andnot	a0, 0x7, v0
+        insqh   t2, a0, t1	# t1 = 0000ffffffffffff (a0<0:2> ff bytes)
+        xor	t0, a1, t0
+	or	t0, t1, t0	# t0 = ====ffffffffffff
+        cmpbge  zero, t0, t1
+        bne     t1, found_it
+
+	/* at least one byte left to process */
+
+	ldq	t0, 8(v0)
+	addq	v0, 8, v0
+	/*
+	 * Make a2 point to last quad to be accessed (the
+	 * last quad may or may not be partial).
+	 */
+	subq	t4, 1, a2
+	andnot	a2, 0x7, a2
+	cmpult	v0, a2, t1
+	beq	t1, final
+
+	/* at least two quads remain to be accessed */
+
+	subq	a2, v0, t3	# t3 <- number of quads to be processed in loop
+	and	t3, 8, t3	# odd number of quads?
+	bne	t3, odd_quad_count
+
+	/* at least three quads remain to be accessed */
+
+	mov	t0, t3		# move prefetched value into correct register
+
+	.align	3
+unrolled_loop:
+	ldq	t0, 8(v0)	# prefetch t0
+	xor	a1, t3, t1
+	cmpbge	zero, t1, t1
+	bne	t1, found_it
+
+	addq	v0, 8, v0
+odd_quad_count:
+	xor	a1, t0, t1
+	ldq	t3, 8(v0)	# prefetch t3
+	cmpbge	zero, t1, t1
+	bne	t1, found_it
+
+	addq	v0, 8, v0
+	cmpult	v0, a2, t5
+	bne	t5, unrolled_loop
+
+	mov	t3, t0		# move prefetched value into t0
+final:	subq	t4, v0, a2	# a2 <- number of bytes left to do
+	bne	a2, last_quad
+
+not_found:
+	mov	zero, v0
+	ret
+
+        .end    memchr