Update.

2001-05-22  David Mosberger  <davidm@hpl.hp.com>

	* sysdeps/ia64/memmove.S: Increase MEMLAT from 6 to 21 for better
	performance.
	* sysdeps/ia64/memcpy.S: Likewise.

	* sysdeps/ia64/bcopy.S: New file.

	* sysdeps/ia64/bzero.S: New file (derived from memset.S).

2001-05-26  Ulrich Drepper  <drepper@redhat.com>

	* sysdeps/ia64/fpu/libm-test-ulps: Add deltas for tanh(-0.7).

5 files changed, 127 insertions, 12 deletions

diff --git a/sysdeps/ia64/bcopy.S b/sysdeps/ia64/bcopy.S
new file mode 100644
index 0000000000..fe90d2166d
--- /dev/null
+++ b/sysdeps/ia64/bcopy.S
@@ -0,0 +1,10 @@
+#include <sysdep.h>
+
+ENTRY(bcopy)
+	.regstk 3, 0, 0, 0
+	mov r8 = in0
+	mov in0 = in1
+	;;
+	mov in1 = r8
+	br.cond.sptk.many memmove
+END(bcopy)
diff --git a/sysdeps/ia64/bzero.S b/sysdeps/ia64/bzero.S
new file mode 100644
index 0000000000..f219f25646
--- /dev/null
+++ b/sysdeps/ia64/bzero.S
@@ -0,0 +1,94 @@
+/* Optimized version of the standard bzero() function.
+   This file is part of the GNU C Library.
+   Copyright (C) 2000, 2001 Free Software Foundation, Inc.
+   Contributed by Dan Pop <Dan.Pop@cern.ch>.
+
+   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., 59 Temple Place - Suite 330,
+   Boston, MA 02111-1307, USA.  */
+
+/* Return: dest
+  
+   Inputs:
+        in0:    dest
+        in1:    count
+
+   The algorithm is fairly straightforward: set byte by byte until we
+   we get to a word aligned address, then set word by word as much as
+   possible; the remaining few bytes are set one by one.  */
+
+#include <sysdep.h>
+#undef ret
+
+#define dest		in0
+#define	cnt		in1
+
+#define save_pfs 	loc0
+#define ptr1		loc1
+#define ptr2		loc2
+#define tmp		loc3
+#define	loopcnt		loc4
+#define save_lc		loc5
+
+ENTRY(bzero)
+	.prologue
+	alloc	save_pfs = ar.pfs, 2, 6, 0, 0	
+	.save ar.lc, save_lc
+	mov	save_lc = ar.lc
+	.body
+	mov	ret0 = dest
+	and	tmp = 7, dest
+	cmp.eq	p6, p0 = cnt, r0
+(p6)	br.cond.spnt .restore_and_exit ;;
+	mov	ptr1 = dest
+	sub	loopcnt = 8, tmp
+	cmp.gt	p6, p0 = 16, cnt
+(p6)	br.cond.spnt .set_few;;
+	cmp.eq	p6, p0 = tmp, r0
+(p6)	br.cond.sptk .dest_aligned
+	sub	cnt = cnt, loopcnt
+	adds	loopcnt = -1, loopcnt;;
+	mov	ar.lc = loopcnt;;	
+.l1:
+	st1	[ptr1] = r0, 1
+	br.cloop.dptk	.l1 ;;
+.dest_aligned:
+	adds	ptr2 = 8, ptr1
+	shr.u	loopcnt = cnt, 4 ;;	// loopcnt = cnt / 16
+	cmp.eq	p6, p0 = loopcnt, r0
+(p6)	br.cond.spnt	.one_more
+	and	cnt = 0xf, cnt		// compute the remaining cnt
+	adds	loopcnt = -1, loopcnt;;
+	mov     ar.lc = loopcnt;;	
+.l2:
+	st8	[ptr1] = r0, 16
+	st8	[ptr2] = r0, 16
+	br.cloop.dptk .l2
+	cmp.le	p6, p0 = 8, cnt	;;
+.one_more:
+(p6)	st8     [ptr1] = r0, 8
+(p6)	adds	cnt = -8, cnt ;;
+	cmp.eq	p6, p0 = cnt, r0
+(p6)	br.cond.spnt	.restore_and_exit
+.set_few:
+	adds	loopcnt = -1, cnt;;
+	mov	ar.lc = loopcnt;;
+.l3:	
+	st1     [ptr1] = r0, 1
+	br.cloop.dptk   .l3 ;;	
+.restore_and_exit:
+	mov	ar.lc = save_lc
+	mov	ar.pfs = save_pfs
+	br.ret.sptk.many b0					
+END(bzero)
diff --git a/sysdeps/ia64/fpu/libm-test-ulps b/sysdeps/ia64/fpu/libm-test-ulps
index 5df69269be..78629b1f36 100644
--- a/sysdeps/ia64/fpu/libm-test-ulps
+++ b/sysdeps/ia64/fpu/libm-test-ulps
@@ -974,6 +974,13 @@ double: 1
 idouble: 1
 ildouble: 521
 ldouble: 521
+Test "tanh (-0.7) == -0.60436777711716349631":
+float: 1
+ifloat: 1
+double: 1
+idouble: 1
+ildouble: 521
+ldouble: 521
 
 # tgamma
 Test "tgamma (-0.5) == -2 sqrt (pi)":
diff --git a/sysdeps/ia64/memcpy.S b/sysdeps/ia64/memcpy.S
index 5dcceaafa3..cd65deee0b 100644
--- a/sysdeps/ia64/memcpy.S
+++ b/sysdeps/ia64/memcpy.S
@@ -68,10 +68,12 @@
 		br.ctop.sptk .loop##shift ;				\
 		br.cond.sptk .cpyfew ; /* deal with the remaining bytes */
 
+#define MEMLAT	21
+#define Nrot	(((2*MEMLAT+3) + 7) & ~7)
+
 ENTRY(memcpy)
 	.prologue
-	alloc 	r2 = ar.pfs, 3, 16 - 3, 0, 16
-#include "softpipe.h"
+	alloc 	r2 = ar.pfs, 3, Nrot - 3, 0, Nrot
 	.rotr	r[MEMLAT + 2], q[MEMLAT + 1]
 	.rotp	p[MEMLAT + 2]
 	mov	ret0 = in0		// return value = dest
diff --git a/sysdeps/ia64/memmove.S b/sysdeps/ia64/memmove.S
index a3f7edef92..7b9fe2265d 100644
--- a/sysdeps/ia64/memmove.S
+++ b/sysdeps/ia64/memmove.S
@@ -26,7 +26,7 @@
         in2:    byte count
 
    The core of the function is the memcpy implementation used in memcpy.S.
-   When bytes have to be copied backwards, only the easy case, when 
+   When bytes have to be copied backwards, only the easy case, when
    all arguments are multiples of 8, is optimised.
 
    In this form, it assumes little endian mode.  For big endian mode,
@@ -67,10 +67,12 @@
 		br.ctop.sptk .loop##shift ;				\
 		br.cond.sptk .cpyfew ; /* deal with the remaining bytes */
 
+#define MEMLAT	21
+#define Nrot	(((2*MEMLAT+3) + 7) & ~7)
+
 ENTRY(memmove)
 	.prologue
-	alloc 	r2 = ar.pfs, 3, 29, 0, 32
-#include "softpipe.h"
+	alloc 	r2 = ar.pfs, 3, Nrot - 3, 0, Nrot
 	.rotr	r[MEMLAT + 2], q[MEMLAT + 1]
 	.rotp	p[MEMLAT + 2]
 	mov	ret0 = in0		// return value = dest
@@ -90,7 +92,7 @@ ENTRY(memmove)
 	and	tmp4 = 7, tmp3 		// tmp4 = (dest | src | len) & 7
 	cmp.le	p6, p0 = dest, src	// if dest <= src it's always safe
 (p6)	br.cond.spnt .forward		// to copy forward
-	add	tmp3 = src, len;;	
+	add	tmp3 = src, len;;
 	cmp.lt	p6, p0 = dest, tmp3	// if dest > src && dest < src + len
 (p6)	br.cond.spnt .backward		// we have to copy backward
 
@@ -113,7 +115,7 @@ ENTRY(memmove)
 (p6)	br.cond.spnt .restore_and_exit;;// the one-word special case
 	adds	adest = 8, dest		// set adest one word ahead of dest
 	adds	asrc = 8, src ;;	// set asrc one word ahead of src
-	nop.b	0			// get the "golden" alignment for 
+	nop.b	0			// get the "golden" alignment for
 	nop.b	0			// the next loop
 .l0:
 (p[0])		ld8	r[0] = [src], 16
@@ -139,8 +141,8 @@ ENTRY(memmove)
 .l1:					// copy -dest % 8 bytes
 	ld1	value = [src], 1	// value = *src++
 	;;
-	st1	[dest] = value, 1	// *dest++ = value  
-	br.cloop.dptk .l1	
+	st1	[dest] = value, 1	// *dest++ = value
+	br.cloop.dptk .l1
 .dest_aligned:
 	and	sh1 = 7, src 		// sh1 = src % 8
 	and	tmp2 = -8, len   	// tmp2 = len & -OPSIZ
@@ -148,7 +150,7 @@ ENTRY(memmove)
 	shr.u	loopcnt = len, 3	// loopcnt = len / 8
 	and	len = 7, len;;		// len = len % 8
 	adds	loopcnt = -1, loopcnt	// --loopcnt
-	addl	tmp4 = @ltoff(.table), gp 
+	addl	tmp4 = @ltoff(.table), gp
 	addl	tmp3 = @ltoff(.loop56), gp
 	mov     ar.ec = MEMLAT + 1	// set EC
 	mov     pr.rot = 1 << 16;;	// set rotating predicates
@@ -174,7 +176,7 @@ ENTRY(memmove)
 	LOOP(40)
 	LOOP(48)
 	LOOP(56)
-	
+
 .src_aligned:
 .l3:
 (p[0])		ld8	r[0] = [src], 8
@@ -220,7 +222,7 @@ ENTRY(memmove)
 	adds	src = -1, src		// src points to the last byte
 	adds	dest = -1, dest		// dest points to the last byte
 	adds	loopcnt = -1, len;;	// loopcnt = len - 1
-	mov	ar.lc = loopcnt;;	// set the loop counter	
+	mov	ar.lc = loopcnt;;	// set the loop counter
 .l6:
 (p[0])		ld1	r[0] = [src], -1
 (p[MEMLAT])	st1	[dest] = r[MEMLAT], -1