added assembler stuff for hppa

12 files changed, 894 insertions, 4 deletions

diff --git a/mpi/config.links b/mpi/config.links
index cf370e054..4c7cbf259 100644
--- a/mpi/config.links
+++ b/mpi/config.links
@@ -37,7 +37,7 @@ case "${target}" in
 	;;
     hppa7000*-*-*)
 	echo '/* configured for HPPA (pa7000) */' >>./mpi/asm-syntax.h
-	path="hppa1_1 hppa"
+	path="hppa1.1 hppa"
 	mpi_extra_modules="udiv-qrnnd"
 	;;
     hppa1.0*-*-*)
@@ -47,7 +47,7 @@ case "${target}" in
 	;;
     hppa*-*-*)	# assume pa7100
 	echo '/* configured for HPPA (pa7100) */' >>./mpi/asm-syntax.h
-	path="pa7100 hppa1_1 hppa"
+	path="pa7100 hppa1.1 hppa"
 	mpi_extra_modules="udiv-qrnnd"
 	;;
     sparc9*-*-* | sparc64*-*-* | ultrasparc*-*-*)
diff --git a/mpi/hppa/README b/mpi/hppa/README
new file mode 100644
index 000000000..5a2d5fd97
--- /dev/null
+++ b/mpi/hppa/README
@@ -0,0 +1,84 @@
+This directory contains mpn functions for various HP PA-RISC chips.  Code
+that runs faster on the PA7100 and later implementations, is in the pa7100
+directory.
+
+RELEVANT OPTIMIZATION ISSUES
+
+  Load and Store timing
+
+On the PA7000 no memory instructions can issue the two cycles after a store.
+For the PA7100, this is reduced to one cycle.
+
+The PA7100 has a lookup-free cache, so it helps to schedule loads and the
+dependent instruction really far from each other.
+
+STATUS
+
+1. mpn_mul_1 could be improved to 6.5 cycles/limb on the PA7100, using the
+   instructions bwlow (but some sw pipelining is needed to avoid the
+   xmpyu-fstds delay):
+
+	fldds	s1_ptr
+
+	xmpyu
+	fstds	N(%r30)
+	xmpyu
+	fstds	N(%r30)
+
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+
+	addc
+	stws	res_ptr
+	addc
+	stws	res_ptr
+
+	addib	Loop
+
+2. mpn_addmul_1 could be improved from the current 10 to 7.5 cycles/limb
+   (asymptotically) on the PA7100, using the instructions below.  With proper
+   sw pipelining and the unrolling level below, the speed becomes 8
+   cycles/limb.
+
+	fldds	s1_ptr
+	fldds	s1_ptr
+
+	xmpyu
+	fstds	N(%r30)
+	xmpyu
+	fstds	N(%r30)
+	xmpyu
+	fstds	N(%r30)
+	xmpyu
+	fstds	N(%r30)
+
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+	ldws	N(%r30)
+	addc
+	addc
+	addc
+	addc
+	addc	%r0,%r0,cy-limb
+
+	ldws	res_ptr
+	ldws	res_ptr
+	ldws	res_ptr
+	ldws	res_ptr
+	add
+	stws	res_ptr
+	addc
+	stws	res_ptr
+	addc
+	stws	res_ptr
+	addc
+	stws	res_ptr
+
+	addib
diff --git a/mpi/hppa/distfiles b/mpi/hppa/distfiles
index 7ca77f801..83c59ea62 100644
--- a/mpi/hppa/distfiles
+++ b/mpi/hppa/distfiles
@@ -1,4 +1,6 @@
-
-mpih-add1.S
+README
 udiv-qrnnd.S
+mpih-add1.S
+mpih-sub1.S
+mpih-shift.S
 
diff --git a/mpi/hppa/mpih-shift.S b/mpi/hppa/mpih-shift.S
new file mode 100644
index 000000000..5039c7381
--- /dev/null
+++ b/mpi/hppa/mpih-shift.S
@@ -0,0 +1,124 @@
+/* hppa   rshift, lshift
+ *	Copyright (C) 1992, 1994 Free Software Foundation, Inc.
+ *	Copyright (c) 1997 by Werner Koch (dd9jn)
+ *
+ * This file is part of G10.
+ *
+ * G10 is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * G10 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ *	 Actually it's the same code with only minor changes in the
+ *	 way the data is stored; this is to support the abstraction
+ *	 of an optional secure memory allocation which may be used
+ *	 to avoid revealing of sensitive data due to paging etc.
+ *	 The GNU MP Library itself is published under the LGPL;
+ *	 however I decided to publish this code under the plain GPL.
+ */
+
+
+
+/*******************
+ * mpi_limb_t
+ * mpihelp_lshift( mpi_ptr_t wp,	(gr26)
+ *		   mpi_ptr_t up,	(gr25)
+ *		   mpi_size_t usize,	(gr24)
+ *		   unsigned cnt)	(gr23)
+ */
+
+	.code
+	.export 	mpihelp_lshift
+mpihelp_lshift
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+
+	sh2add		%r24,%r25,%r25
+	sh2add		%r24,%r26,%r26
+	ldws,mb 	-4(0,%r25),%r22
+	subi		32,%r23,%r1
+	mtsar		%r1
+	addib,= 	-1,%r24,L$0004
+	vshd		%r0,%r22,%r28		; compute carry out limb
+	ldws,mb 	-4(0,%r25),%r29
+	addib,= 	-1,%r24,L$0002
+	vshd		%r22,%r29,%r20
+
+L$loop	ldws,mb 	-4(0,%r25),%r22
+	stws,mb 	%r20,-4(0,%r26)
+	addib,= 	-1,%r24,L$0003
+	vshd		%r29,%r22,%r20
+	ldws,mb 	-4(0,%r25),%r29
+	stws,mb 	%r20,-4(0,%r26)
+	addib,<>	-1,%r24,L$loop
+	vshd		%r22,%r29,%r20
+
+L$0002	stws,mb 	%r20,-4(0,%r26)
+	vshd		%r29,%r0,%r20
+	bv		0(%r2)
+	stw		%r20,-4(0,%r26)
+L$0003	stws,mb 	%r20,-4(0,%r26)
+L$0004	vshd		%r22,%r0,%r20
+	bv		0(%r2)
+	stw		%r20,-4(0,%r26)
+
+	.exit
+	.procend
+
+
+
+/*******************
+ * mpi_limb_t
+ * mpihelp_rshift( mpi_ptr_t wp,       (gr26)
+ *		   mpi_ptr_t up,       (gr25)
+ *		   mpi_size_t usize,   (gr24)
+ *		   unsigned cnt)       (gr23)
+ */
+
+	.code
+	.export 	mpihelp_rshift
+mpihelp_rshift
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+
+	ldws,ma 	4(0,%r25),%r22
+	mtsar		%r23
+	addib,= 	-1,%r24,L$r004
+	vshd		%r22,%r0,%r28		; compute carry out limb
+	ldws,ma 	4(0,%r25),%r29
+	addib,= 	-1,%r24,L$r002
+	vshd		%r29,%r22,%r20
+
+L$roop	ldws,ma 	4(0,%r25),%r22
+	stws,ma 	%r20,4(0,%r26)
+	addib,= 	-1,%r24,L$r003
+	vshd		%r22,%r29,%r20
+	ldws,ma 	4(0,%r25),%r29
+	stws,ma 	%r20,4(0,%r26)
+	addib,<>	-1,%r24,L$roop
+	vshd		%r29,%r22,%r20
+
+L$r002	stws,ma 	%r20,4(0,%r26)
+	vshd		%r0,%r29,%r20
+	bv		0(%r2)
+	stw		%r20,0(0,%r26)
+L$r003	stws,ma 	%r20,4(0,%r26)
+L$r004	vshd		%r0,%r22,%r20
+	bv		0(%r2)
+	stw		%r20,0(0,%r26)
+
+	.exit
+	.procend
+
diff --git a/mpi/hppa/mpih-sub1.S b/mpi/hppa/mpih-sub1.S
new file mode 100644
index 000000000..361cc6832
--- /dev/null
+++ b/mpi/hppa/mpih-sub1.S
@@ -0,0 +1,77 @@
+/* hppa   sub_n -- Sub two limb vectors of the same length > 0 and store
+ *		   sum in a third limb vector.
+ *	Copyright (C) 1992, 1994 Free Software Foundation, Inc.
+ *	Copyright (c) 1998 by Werner Koch (dd9jn)
+ *
+ * This file is part of G10.
+ *
+ * G10 is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * G10 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ *	 Actually it's the same code with only minor changes in the
+ *	 way the data is stored; this is to support the abstraction
+ *	 of an optional secure memory allocation which may be used
+ *	 to avoid revealing of sensitive data due to paging etc.
+ *	 The GNU MP Library itself is published under the LGPL;
+ *	 however I decided to publish this code under the plain GPL.
+ */
+
+
+#include "sysdep.h"
+#include "asm-syntax.h"
+
+
+/*******************
+ *  mpi_limb_t
+ *  mpihelp_sub_n( mpi_ptr_t res_ptr,	(gr26)
+ *		   mpi_ptr_t s1_ptr,	(gr25)
+ *		   mpi_ptr_t s2_ptr,	(gr24)
+ *		   mpi_size_t size)	(gr23)
+ *
+ * One might want to unroll this as for other processors, but it turns
+ * out that the data cache contention after a store makes such
+ * unrolling useless.  We can't come under 5 cycles/limb anyway.
+ */
+
+
+	.code
+	.export 	mpihelp_sub_n
+mpihelp_sub_n
+	.proc
+	.callinfo	frame=0,no_calls
+	.entry
+
+	ldws,ma 	4(0,%r25),%r20
+	ldws,ma 	4(0,%r24),%r19
+
+	addib,= 	-1,%r23,L$end	; check for (SIZE == 1)
+	 sub		%r20,%r19,%r28	; subtract first limbs ignoring cy
+
+L$loop	ldws,ma 	4(0,%r25),%r20
+	ldws,ma 	4(0,%r24),%r19
+	stws,ma 	%r28,4(0,%r26)
+	addib,<>	-1,%r23,L$loop
+	 subb		%r20,%r19,%r28
+
+L$end	stws		%r28,0(0,%r26)
+	addc		%r0,%r0,%r28
+	bv		0(%r2)
+	 subi		1,%r28,%r28
+
+	.exit
+	.procend
+
+
+
diff --git a/mpi/hppa1.1/distfiles b/mpi/hppa1.1/distfiles
new file mode 100644
index 000000000..d68227ac7
--- /dev/null
+++ b/mpi/hppa1.1/distfiles
@@ -0,0 +1,5 @@
+udiv-qrnnd.S
+mpih-mul1.S
+mpih-mul2.S
+mpih-mul3.S
+
diff --git a/mpi/hppa1.1/mpih-mul1.S b/mpi/hppa1.1/mpih-mul1.S
new file mode 100644
index 000000000..d68b4eb6f
--- /dev/null
+++ b/mpi/hppa1.1/mpih-mul1.S
@@ -0,0 +1,112 @@
+/* hppa1.1  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.
+ *	Copyright (c) 1998 by Werner Koch (dd9jn)
+ *
+ * This file is part of G10.
+ *
+ * G10 is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * G10 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ *	 Actually it's the same code with only minor changes in the
+ *	 way the data is stored; this is to support the abstraction
+ *	 of an optional secure memory allocation which may be used
+ *	 to avoid revealing of sensitive data due to paging etc.
+ *	 The GNU MP Library itself is published under the LGPL;
+ *	 however I decided to publish this code under the plain GPL.
+ */
+
+
+/*******************
+ * mpi_limb_t
+ * mpihelp_mul_1( mpi_ptr_t res_ptr,	(r26)
+ *		  mpi_ptr_t s1_ptr,	(r25)
+ *		  mpi_size_t s1_size,	(r24)
+ *		  mpi_limb_t 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 	mpihelp_mul_1
+mpihelp_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/mpi/hppa1.1/mpih-mul2.S b/mpi/hppa1.1/mpih-mul2.S
new file mode 100644
index 000000000..654eb074f
--- /dev/null
+++ b/mpi/hppa1.1/mpih-mul2.S
@@ -0,0 +1,114 @@
+/* hppa1.1   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.
+ *	Copyright (c) 1997 by Werner Koch (dd9jn)
+ *
+ * This file is part of G10.
+ *
+ * G10 is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * G10 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ *	 Actually it's the same code with only minor changes in the
+ *	 way the data is stored; this is to support the abstraction
+ *	 of an optional secure memory allocation which may be used
+ *	 to avoid revealing of sensitive data due to paging etc.
+ *	 The GNU MP Library itself is published under the LGPL;
+ *	 however I decided to publish this code under the plain GPL.
+ */
+
+
+/*******************
+ * mpi_limb_t
+ * mpihelp_addmul_1( mpi_ptr_t res_ptr,      (r26)
+ *		     mpi_ptr_t s1_ptr,	     (r25)
+ *		     mpi_size_t s1_size,     (r24)
+ *		     mpi_limb_t 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 mul1.S that applies to this code too.
+ */
+
+	.code
+	.export 	mpihelp_addmul_1
+mpihelp_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/mpi/hppa1.1/mpih-mul3.S b/mpi/hppa1.1/mpih-mul3.S
new file mode 100644
index 000000000..87f9e6a6b
--- /dev/null
+++ b/mpi/hppa1.1/mpih-mul3.S
@@ -0,0 +1,124 @@
+/* hppa1.1   submul_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.
+ *	Copyright (c) 1997 by Werner Koch (dd9jn)
+ *
+ * This file is part of G10.
+ *
+ * G10 is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * G10 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ *	 Actually it's the same code with only minor changes in the
+ *	 way the data is stored; this is to support the abstraction
+ *	 of an optional secure memory allocation which may be used
+ *	 to avoid revealing of sensitive data due to paging etc.
+ *	 The GNU MP Library itself is published under the LGPL;
+ *	 however I decided to publish this code under the plain GPL.
+ */
+
+
+/*******************
+ * mpi_limb_t
+ * mpihelp_submul_1( mpi_ptr_t res_ptr,      (r26)
+ *		     mpi_ptr_t s1_ptr,	     (r25)
+ *		     mpi_size_t s1_size,     (r24)
+ *		     mpi_limb_t 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 mul1.S that applies to this code too.
+ *
+ * It seems possible to make this run as fast as addmul_1, if we use
+ *	 sub,>>= %r29,%r19,%r22
+ *	 addi	 1,%r28,%r28
+ * but that requires reworking the hairy software pipeline...
+ */
+
+
+	.code
+	.export 	mpihelp_submul_1
+mpihelp_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/mpi/hppa1.1/udiv-qrnnd.S b/mpi/hppa1.1/udiv-qrnnd.S
new file mode 100644
index 000000000..5e88f5c26
--- /dev/null
+++ b/mpi/hppa1.1/udiv-qrnnd.S
@@ -0,0 +1,88 @@
+/* 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.
+ *	Copyright (c) 1997 by Werner Koch (dd9jn)
+ *
+ * This file is part of G10.
+ *
+ * G10 is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * G10 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ *	 Actually it's the same code with only minor changes in the
+ *	 way the data is stored; this is to support the abstraction
+ *	 of an optional secure memory allocation which may be used
+ *	 to avoid revealing of sensitive data due to paging etc.
+ *	 The GNU MP Library itself is published under the LGPL;
+ *	 however I decided to publish this code under the plain GPL.
+ */
+
+
+
+/* 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
+
+
diff --git a/mpi/pa7100/distfiles b/mpi/pa7100/distfiles
new file mode 100644
index 000000000..d468fa20a
--- /dev/null
+++ b/mpi/pa7100/distfiles
@@ -0,0 +1,2 @@
+mpih-shift.S
+
diff --git a/mpi/pa7100/mpih-shift.S b/mpi/pa7100/mpih-shift.S
new file mode 100644
index 000000000..daff03f3f
--- /dev/null
+++ b/mpi/pa7100/mpih-shift.S
@@ -0,0 +1,158 @@
+/* hppa   rshift, lshift
+ *	  optimized for the PA7100, where is runs at 3.25 cycles/limb
+ *	Copyright (C) 1992, 1994 Free Software Foundation, Inc.
+ *	Copyright (c) 1997 by Werner Koch (dd9jn)
+ *
+ * This file is part of G10.
+ *
+ * G10 is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * G10 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
+ *
+ * Note: This code is heavily based on the GNU MP Library.
+ *	 Actually it's the same code with only minor changes in the
+ *	 way the data is stored; this is to support the abstraction
+ *	 of an optional secure memory allocation which may be used
+ *	 to avoid revealing of sensitive data due to paging etc.
+ *	 The GNU MP Library itself is published under the LGPL;
+ *	 however I decided to publish this code under the plain GPL.
+ */
+
+
+
+/*******************
+ * mpi_limb_t
+ * mpihelp_lshift( mpi_ptr_t wp,	(gr26)
+ *		   mpi_ptr_t up,	(gr25)
+ *		   mpi_size_t usize,	(gr24)
+ *		   unsigned cnt)	(gr23)
+ */
+
+	.code
+	.export 	mpihelp_lshift
+mpihelp_lshift
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+
+	sh2add		%r24,%r25,%r25
+	sh2add		%r24,%r26,%r26
+	ldws,mb 	-4(0,%r25),%r22
+	subi		32,%r23,%r1
+	mtsar		%r1
+	addib,= 	-1,%r24,L$0004
+	vshd		%r0,%r22,%r28		; compute carry out limb
+	ldws,mb 	-4(0,%r25),%r29
+	addib,<=	-5,%r24,L$rest
+	vshd		%r22,%r29,%r20
+
+L$loop	ldws,mb 	-4(0,%r25),%r22
+	stws,mb 	%r20,-4(0,%r26)
+	vshd		%r29,%r22,%r20
+	ldws,mb 	-4(0,%r25),%r29
+	stws,mb 	%r20,-4(0,%r26)
+	vshd		%r22,%r29,%r20
+	ldws,mb 	-4(0,%r25),%r22
+	stws,mb 	%r20,-4(0,%r26)
+	vshd		%r29,%r22,%r20
+	ldws,mb 	-4(0,%r25),%r29
+	stws,mb 	%r20,-4(0,%r26)
+	addib,> 	-4,%r24,L$loop
+	vshd		%r22,%r29,%r20
+
+L$rest	addib,= 	4,%r24,L$end1
+	nop
+L$eloop ldws,mb 	-4(0,%r25),%r22
+	stws,mb 	%r20,-4(0,%r26)
+	addib,<=	-1,%r24,L$end2
+	vshd		%r29,%r22,%r20
+	ldws,mb 	-4(0,%r25),%r29
+	stws,mb 	%r20,-4(0,%r26)
+	addib,> 	-1,%r24,L$eloop
+	vshd		%r22,%r29,%r20
+
+L$end1	stws,mb 	%r20,-4(0,%r26)
+	vshd		%r29,%r0,%r20
+	bv		0(%r2)
+	stw		%r20,-4(0,%r26)
+L$end2	stws,mb 	%r20,-4(0,%r26)
+L$0004	vshd		%r22,%r0,%r20
+	bv		0(%r2)
+	stw		%r20,-4(0,%r26)
+
+	.exit
+	.procend
+
+
+
+/*******************
+ * mpi_limb_t
+ * mpihelp_rshift( mpi_ptr_t wp,       (gr26)
+ *		   mpi_ptr_t up,       (gr25)
+ *		   mpi_size_t usize,   (gr24)
+ *		   unsigned cnt)       (gr23)
+ */
+
+	.code
+	.export 	mpihelp_rshift
+mpihelp_rshift
+	.proc
+	.callinfo	frame=64,no_calls
+	.entry
+
+	ldws,ma 	4(0,%r25),%r22
+	mtsar		%r23
+	addib,= 	-1,%r24,L$r004
+	vshd		%r22,%r0,%r28		; compute carry out limb
+	ldws,ma 	4(0,%r25),%r29
+	addib,<=	-5,%r24,L$rrest
+	vshd		%r29,%r22,%r20
+
+L$roop	ldws,ma 	4(0,%r25),%r22
+	stws,ma 	%r20,4(0,%r26)
+	vshd		%r22,%r29,%r20
+	ldws,ma 	4(0,%r25),%r29
+	stws,ma 	%r20,4(0,%r26)
+	vshd		%r29,%r22,%r20
+	ldws,ma 	4(0,%r25),%r22
+	stws,ma 	%r20,4(0,%r26)
+	vshd		%r22,%r29,%r20
+	ldws,ma 	4(0,%r25),%r29
+	stws,ma 	%r20,4(0,%r26)
+	addib,> 	-4,%r24,L$roop
+	vshd		%r29,%r22,%r20
+
+L$rrest addib,= 	4,%r24,L$rend1
+	nop
+L$eroop ldws,ma 	4(0,%r25),%r22
+	stws,ma 	%r20,4(0,%r26)
+	addib,<=	-1,%r24,L$rend2
+	vshd		%r22,%r29,%r20
+	ldws,ma 	4(0,%r25),%r29
+	stws,ma 	%r20,4(0,%r26)
+	addib,> 	-1,%r24,L$eroop
+	vshd		%r29,%r22,%r20
+
+L$rend1  stws,ma	 %r20,4(0,%r26)
+	vshd		%r0,%r29,%r20
+	bv		0(%r2)
+	stw		%r20,0(0,%r26)
+L$rend2  stws,ma	 %r20,4(0,%r26)
+L$r004	vshd		%r0,%r22,%r20
+	bv		0(%r2)
+	stw		%r20,0(0,%r26)
+
+	.exit
+	.procend
+
+