Sat Oct 14 02:52:36 1995 Ulrich Drepper <drepper@ipd.info.uni-karlsruhe.de>

	* malloc/malloc.c (_malloc_internal): Performance fix.  Move
	if statement out of loop.

	* stdio/_itoa.c, stdio/_itoa.h: Complete rewrite.  Much faster
	implementation using GMP functions.  Contributed by
	Torbjorn Granlund and Ulrich Drepper.

	* stdio/test_rdwr.c: Include <errno.h>.

	* sysdeps/i386/i586/Implies: New file.

	New highly optimized string functions for i[345]86.
	* sysdeps/i386/memchr.S, sysdeps/i386/memcmp.S: New files.
        * sysdeps/i386/stpcpy.S, sysdeps/i386/stpncpy.S: New files.
        * sysdeps/i386/strchr.S, sysdeps/i386/strcspn.S: New files.
        * sysdeps/i386/strpbrk.S, sysdeps/i386/strrchr.S: New files.
        * sysdeps/i386/strspn.S, sysdeps/i386/i486/strcat.S: New files.
        * sysdeps/i386/i486/strlen.S, sysdeps/i386/i586/strchr.S: New files.
        * sysdeps/i386/i586/strlen.S: New file.
	* sysdeps/i386/memchr.c: Removed.  There is now an assembler version.

	* sysdeps/i386/i586/memcopy.h (WORD_COPY_BWD): Parameters did
	not correspond to used values.

	* sysdeps/unix/sysv/linux/nfs/nfs.h: New file.  Simply a wrapper
        around a kernel header file.
	* sysdeps/unix/sysv/linux/Dist: Add it.
	* sysdeps/unix/sysv/linux/Makefile [$(subdir)=sunrpc] (headers):
	Likewise.

	* sysdeps/unix/sysv/linux/local_lim.h: Rewrite.  Instead of
        defining ourself we use a kernel header file.

	* sysdeps/unix/sysv/linux/i386/sysdep.h (DO_CALL): Optimize system
        call handler for i586.

	* sysdeps/unix/sysv/linux/sys/param.h: Add copyright and clean up.
Sat Oct 14 02:52:36 1995  Ulrich Drepper  <drepper@ipd.info.uni-karlsruhe.de>

	* malloc/malloc.c (_malloc_internal): Performance fix.  Move
	if statement out of loop.

	* stdio/_itoa.c, stdio/_itoa.h: Complete rewrite.  Much faster
	implementation using GMP functions.  Contributed by
	Torbjorn Granlund and Ulrich Drepper.

	* stdio/test_rdwr.c: Include <errno.h>.

	* sysdeps/i386/i586/Implies: New file.

	New highly optimized string functions for i[345]86.
	* sysdeps/i386/memchr.S, sysdeps/i386/memcmp.S: New files.
        * sysdeps/i386/stpcpy.S, sysdeps/i386/stpncpy.S: New files.
        * sysdeps/i386/strchr.S, sysdeps/i386/strcspn.S: New files.
        * sysdeps/i386/strpbrk.S, sysdeps/i386/strrchr.S: New files.
        * sysdeps/i386/strspn.S, sysdeps/i386/i486/strcat.S: New files.
        * sysdeps/i386/i486/strlen.S, sysdeps/i386/i586/strchr.S: New files.
        * sysdeps/i386/i586/strlen.S: New file.
	* sysdeps/i386/memchr.c: Removed.  There is now an assembler version.

	* sysdeps/i386/i586/memcopy.h (WORD_COPY_BWD): Parameters did
	not correspond to used values.

	* sysdeps/unix/sysv/linux/nfs/nfs.h: New file.  Simply a wrapper
        around a kernel header file.
	* sysdeps/unix/sysv/linux/Dist: Add it.
	* sysdeps/unix/sysv/linux/Makefile [$(subdir)=sunrpc] (headers):
	Likewise.

	* sysdeps/unix/sysv/linux/local_lim.h: Rewrite.  Instead of
        defining ourself we use a kernel header file.

	* sysdeps/unix/sysv/linux/i386/sysdep.h (DO_CALL): Optimize system
        call handler for i586.

	* sysdeps/unix/sysv/linux/sys/param.h: Add copyright and clean up.

4 files changed, 382 insertions, 0 deletions

diff --git a/sysdeps/hppa/hppa1.1/addmul_1.s b/sysdeps/hppa/hppa1.1/addmul_1.s
new file mode 100644
index 0000000000..a9dfdd1c28
--- /dev/null
+++ b/sysdeps/hppa/hppa1.1/addmul_1.s
@@ -0,0 +1,101 @@
+; HP-PA-1.1 __mpn_addmul_1 -- Multiply a limb vector with a limb and
+; add the result to a second limb vector.
+
+; Copyright (C) 1992, 1993, 1994 Free Software Foundation, Inc.
+
+; This file is part of the GNU MP Library.
+
+; The GNU MP 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 MP 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 MP Library; see the file COPYING.LIB.  If not, write to
+; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+
+
+; INPUT PARAMETERS
+; res_ptr	r26
+; s1_ptr	r25
+; size		r24
+; s2_limb	r23
+
+; This runs at 11 cycles/limb on a PA7000.  With the used instructions, it
+; can not become faster due to data cache contention after a store.  On the
+; PA7100 it runs at 10 cycles/limb, and that can not be improved either,
+; since only the xmpyu does not need the integer pipeline, so the only
+; dual-issue we will get are addc+xmpyu.  Unrolling could gain a cycle/limb
+; on the PA7100.
+
+; There are some ideas described in mul_1.s that applies to this code too.
+
+	.code
+	.export		__mpn_addmul_1
+__mpn_addmul_1
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+
+	ldo		64(%r30),%r30
+	fldws,ma	4(%r25),%fr5
+	stw		%r23,-16(%r30)		; move s2_limb ...
+	addib,=		-1,%r24,L$just_one_limb
+	 fldws		-16(%r30),%fr4		; ... into fr4
+	add		%r0,%r0,%r0		; clear carry
+	xmpyu		%fr4,%fr5,%fr6
+	fldws,ma	4(%r25),%fr7
+	fstds		%fr6,-16(%r30)
+	xmpyu		%fr4,%fr7,%fr8
+	ldw		-12(%r30),%r19		; least significant limb in product
+	ldw		-16(%r30),%r28
+
+	fstds		%fr8,-16(%r30)
+	addib,=		-1,%r24,L$end
+	 ldw		-12(%r30),%r1
+
+; Main loop
+L$loop	ldws		0(%r26),%r29
+	fldws,ma	4(%r25),%fr5
+	add		%r29,%r19,%r19
+	stws,ma		%r19,4(%r26)
+	addc		%r28,%r1,%r19
+	xmpyu		%fr4,%fr5,%fr6
+	ldw		-16(%r30),%r28
+	fstds		%fr6,-16(%r30)
+	addc		%r0,%r28,%r28
+	addib,<>	-1,%r24,L$loop
+	 ldw		-12(%r30),%r1
+
+L$end	ldw		0(%r26),%r29
+	add		%r29,%r19,%r19
+	stws,ma		%r19,4(%r26)
+	addc		%r28,%r1,%r19
+	ldw		-16(%r30),%r28
+	ldws		0(%r26),%r29
+	addc		%r0,%r28,%r28
+	add		%r29,%r19,%r19
+	stws,ma		%r19,4(%r26)
+	addc		%r0,%r28,%r28
+	bv		0(%r2)
+	 ldo		-64(%r30),%r30
+
+L$just_one_limb
+	xmpyu		%fr4,%fr5,%fr6
+	ldw		0(%r26),%r29
+	fstds		%fr6,-16(%r30)
+	ldw		-12(%r30),%r1
+	ldw		-16(%r30),%r28
+	add		%r29,%r1,%r19
+	stw		%r19,0(%r26)
+	addc		%r0,%r28,%r28
+	bv		0(%r2)
+	 ldo		-64(%r30),%r30
+
+	.exit
+	.procend
diff --git a/sysdeps/hppa/hppa1.1/mul_1.s b/sysdeps/hppa/hppa1.1/mul_1.s
new file mode 100644
index 0000000000..ebf0778b90
--- /dev/null
+++ b/sysdeps/hppa/hppa1.1/mul_1.s
@@ -0,0 +1,97 @@
+; HP-PA-1.1 __mpn_mul_1 -- Multiply a limb vector with a limb and store
+; the result in a second limb vector.
+
+; Copyright (C) 1992, 1993, 1994 Free Software Foundation, Inc.
+
+; This file is part of the GNU MP Library.
+
+; The GNU MP 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 MP 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 MP Library; see the file COPYING.LIB.  If not, write to
+; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+
+
+; INPUT PARAMETERS
+; res_ptr	r26
+; s1_ptr	r25
+; size		r24
+; s2_limb	r23
+
+; This runs at 9 cycles/limb on a PA7000.  With the used instructions, it can
+; not become faster due to data cache contention after a store.  On the
+; PA7100 it runs at 7 cycles/limb, and that can not be improved either, since
+; only the xmpyu does not need the integer pipeline, so the only dual-issue
+; we will get are addc+xmpyu.  Unrolling would not help either CPU.
+
+; We could use fldds to read two limbs at a time from the S1 array, and that
+; could bring down the times to 8.5 and 6.5 cycles/limb for the PA7000 and
+; PA7100, respectively.  We don't do that since it does not seem worth the
+; (alignment) troubles...
+
+; At least the PA7100 is rumored to be able to deal with cache-misses
+; without stalling instruction issue.  If this is true, and the cache is
+; actually also lockup-free, we should use a deeper software pipeline, and
+; load from S1 very early!  (The loads and stores to -12(sp) will surely be
+; in the cache.)
+
+	.code
+	.export		__mpn_mul_1
+__mpn_mul_1
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+
+	ldo		64(%r30),%r30
+	fldws,ma	4(%r25),%fr5
+	stw		%r23,-16(%r30)		; move s2_limb ...
+	addib,=		-1,%r24,L$just_one_limb
+	 fldws		-16(%r30),%fr4		; ... into fr4
+	add		%r0,%r0,%r0		; clear carry
+	xmpyu		%fr4,%fr5,%fr6
+	fldws,ma	4(%r25),%fr7
+	fstds	 	%fr6,-16(%r30)
+	xmpyu		%fr4,%fr7,%fr8
+	ldw		-12(%r30),%r19		; least significant limb in product
+	ldw		-16(%r30),%r28
+
+	fstds		%fr8,-16(%r30)
+	addib,=		-1,%r24,L$end
+	 ldw		-12(%r30),%r1
+
+; Main loop
+L$loop	fldws,ma	4(%r25),%fr5
+	stws,ma		%r19,4(%r26)
+	addc		%r28,%r1,%r19
+	xmpyu		%fr4,%fr5,%fr6
+	ldw		-16(%r30),%r28
+	fstds		%fr6,-16(%r30)
+	addib,<>	-1,%r24,L$loop
+	 ldw		-12(%r30),%r1
+
+L$end	stws,ma		%r19,4(%r26)
+	addc		%r28,%r1,%r19
+	ldw		-16(%r30),%r28
+	stws,ma		%r19,4(%r26)
+	addc		%r0,%r28,%r28
+	bv		0(%r2)
+	 ldo		-64(%r30),%r30
+
+L$just_one_limb
+	xmpyu		%fr4,%fr5,%fr6
+	fstds		%fr6,-16(%r30)
+	ldw		-16(%r30),%r28
+	ldo		-64(%r30),%r30
+	bv		0(%r2)
+	 fstws		%fr6R,0(%r26)
+
+	.exit
+	.procend
diff --git a/sysdeps/hppa/hppa1.1/submul_1.s b/sysdeps/hppa/hppa1.1/submul_1.s
new file mode 100644
index 0000000000..44cabf4690
--- /dev/null
+++ b/sysdeps/hppa/hppa1.1/submul_1.s
@@ -0,0 +1,110 @@
+; HP-PA-1.1 __mpn_submul_1 -- Multiply a limb vector with a limb and
+; subtract the result from a second limb vector.
+
+; Copyright (C) 1992, 1993, 1994 Free Software Foundation, Inc.
+
+; This file is part of the GNU MP Library.
+
+; The GNU MP 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 MP 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 MP Library; see the file COPYING.LIB.  If not, write to
+; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+
+
+; INPUT PARAMETERS
+; res_ptr	r26
+; s1_ptr	r25
+; size		r24
+; s2_limb	r23
+
+; This runs at 12 cycles/limb on a PA7000.  With the used instructions, it
+; can not become faster due to data cache contention after a store.  On the
+; PA7100 it runs at 11 cycles/limb, and that can not be improved either,
+; since only the xmpyu does not need the integer pipeline, so the only
+; dual-issue we will get are addc+xmpyu.  Unrolling could gain a cycle/limb
+; on the PA7100.
+
+; There are some ideas described in mul_1.s that applies to this code too.
+
+; It seems possible to make this run as fast as __mpn_addmul_1, if we use
+; 	sub,>>=	%r29,%r19,%r22
+;	addi	1,%r28,%r28
+; but that requires reworking the hairy software pipeline...
+
+	.code
+	.export		__mpn_submul_1
+__mpn_submul_1
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+
+	ldo		64(%r30),%r30
+	fldws,ma	4(%r25),%fr5
+	stw		%r23,-16(%r30)		; move s2_limb ...
+	addib,=		-1,%r24,L$just_one_limb
+	 fldws		-16(%r30),%fr4		; ... into fr4
+	add		%r0,%r0,%r0		; clear carry
+	xmpyu		%fr4,%fr5,%fr6
+	fldws,ma	4(%r25),%fr7
+	fstds		%fr6,-16(%r30)
+	xmpyu		%fr4,%fr7,%fr8
+	ldw		-12(%r30),%r19		; least significant limb in product
+	ldw		-16(%r30),%r28
+
+	fstds		%fr8,-16(%r30)
+	addib,=		-1,%r24,L$end
+	 ldw		-12(%r30),%r1
+
+; Main loop
+L$loop	ldws		0(%r26),%r29
+	fldws,ma	4(%r25),%fr5
+	sub		%r29,%r19,%r22
+	add		%r22,%r19,%r0
+	stws,ma		%r22,4(%r26)
+	addc		%r28,%r1,%r19
+	xmpyu		%fr4,%fr5,%fr6
+	ldw		-16(%r30),%r28
+	fstds		%fr6,-16(%r30)
+	addc		%r0,%r28,%r28
+	addib,<>	-1,%r24,L$loop
+	 ldw		-12(%r30),%r1
+
+L$end	ldw		0(%r26),%r29
+	sub		%r29,%r19,%r22
+	add		%r22,%r19,%r0
+	stws,ma		%r22,4(%r26)
+	addc		%r28,%r1,%r19
+	ldw		-16(%r30),%r28
+	ldws		0(%r26),%r29
+	addc		%r0,%r28,%r28
+	sub		%r29,%r19,%r22
+	add		%r22,%r19,%r0
+	stws,ma		%r22,4(%r26)
+	addc		%r0,%r28,%r28
+	bv		0(%r2)
+	 ldo		-64(%r30),%r30
+
+L$just_one_limb
+	xmpyu		%fr4,%fr5,%fr6
+	ldw		0(%r26),%r29
+	fstds		%fr6,-16(%r30)
+	ldw		-12(%r30),%r1
+	ldw		-16(%r30),%r28
+	sub		%r29,%r1,%r22
+	add		%r22,%r1,%r0
+	stw		%r22,0(%r26)
+	addc		%r0,%r28,%r28
+	bv		0(%r2)
+	 ldo		-64(%r30),%r30
+
+	.exit
+	.procend
diff --git a/sysdeps/hppa/hppa1.1/udiv_qrnnd.s b/sysdeps/hppa/hppa1.1/udiv_qrnnd.s
new file mode 100644
index 0000000000..4ffef3a4fb
--- /dev/null
+++ b/sysdeps/hppa/hppa1.1/udiv_qrnnd.s
@@ -0,0 +1,74 @@
+; HP-PA  __udiv_qrnnd division support, used from longlong.h.
+; This version runs fast on PA 7000 and later.
+
+; Copyright (C) 1993, 1994 Free Software Foundation, Inc.
+
+; This file is part of the GNU MP Library.
+
+; The GNU MP 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 MP 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 MP Library; see the file COPYING.LIB.  If not, write to
+; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+
+
+; INPUT PARAMETERS
+; rem_ptr	gr26
+; n1		gr25
+; n0		gr24
+; d		gr23
+
+	.code
+L$0000	.word		0x43f00000
+	.word		0x0
+	.export		__udiv_qrnnd
+__udiv_qrnnd
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+	ldo		64(%r30),%r30
+
+	stws		%r25,-16(0,%r30)	; n_hi
+	stws		%r24,-12(0,%r30)	; n_lo
+	ldil		L'L$0000,%r19
+	ldo		R'L$0000(%r19),%r19
+	fldds		-16(0,%r30),%fr5
+	stws		%r23,-12(0,%r30)
+	comib,<=	0,%r25,L$1
+	fcnvxf,dbl,dbl	%fr5,%fr5
+	fldds		0(0,%r19),%fr4
+	fadd,dbl	%fr4,%fr5,%fr5
+L$1
+	fcpy,sgl	%fr0,%fr6L
+	fldws		-12(0,%r30),%fr6R
+	fcnvxf,dbl,dbl	%fr6,%fr4
+
+	fdiv,dbl	%fr5,%fr4,%fr5
+
+	fcnvfx,dbl,dbl	%fr5,%fr4
+	fstws		%fr4R,-16(%r30)
+	xmpyu		%fr4R,%fr6R,%fr6
+	ldws		-16(%r30),%r28
+	fstds		%fr6,-16(0,%r30)
+	ldws		-12(0,%r30),%r21
+	ldws		-16(0,%r30),%r20
+	sub		%r24,%r21,%r22
+	subb		%r25,%r20,%r19
+	comib,=		0,%r19,L$2
+	ldo		-64(%r30),%r30
+
+	add		%r22,%r23,%r22
+	ldo		-1(%r28),%r28
+L$2	bv		0(%r2)
+	stws		%r22,0(0,%r26)
+
+	.exit
+	.procend