Updated from ../=mpn/gmp-1.910

14 files changed, 1165 insertions, 248 deletions

diff --git a/sysdeps/generic/mul_n.c b/sysdeps/generic/mul_n.c
index 7900988..e37c5d8 100644
--- a/sysdeps/generic/mul_n.c
+++ b/sysdeps/generic/mul_n.c
@@ -216,15 +216,7 @@ ____mpn_mul_n (prodp, up, vp, size, tspace)
 
       cy += __mpn_add_n (prodp + hsize, prodp + hsize, tspace, size);
       if (cy)
-	{
-	  if (cy > 0)
-	    __mpn_add_1 (prodp + hsize + size, prodp + hsize + size, hsize, cy);
-	  else
-	    {
-	      __mpn_sub_1 (prodp + hsize + size, prodp + hsize + size, hsize, cy);
-	      abort ();
-	    }
-	}
+	__mpn_add_1 (prodp + hsize + size, prodp + hsize + size, hsize, cy);
 
       MPN_COPY (prodp, tspace, hsize);
       cy = __mpn_add_n (prodp + hsize, prodp + hsize, tspace + hsize, hsize);
@@ -362,15 +354,7 @@ ____mpn_sqr_n (prodp, up, size, tspace)
 
       cy += __mpn_add_n (prodp + hsize, prodp + hsize, tspace, size);
       if (cy)
-	{
-	  if (cy > 0)
-	    __mpn_add_1 (prodp + hsize + size, prodp + hsize + size, hsize, cy);
-	  else
-	    {
-	      __mpn_sub_1 (prodp + hsize + size, prodp + hsize + size, hsize, cy);
-	      abort ();
-	    }
-	}
+	__mpn_add_1 (prodp + hsize + size, prodp + hsize + size, hsize, cy);
 
       MPN_COPY (prodp, tspace, hsize);
       cy = __mpn_add_n (prodp + hsize, prodp + hsize, tspace + hsize, hsize);
diff --git a/sysdeps/i386/i586/lshift.S b/sysdeps/i386/i586/lshift.S
index b9f8131..c41f74e 100644
--- a/sysdeps/i386/i586/lshift.S
+++ b/sysdeps/i386/i586/lshift.S
@@ -43,12 +43,15 @@ C_SYMBOL_NAME(__mpn_lshift:)
 	movl	28(%esp),%ebp		/* size */
 	movl	32(%esp),%ecx		/* cnt */
 
+/* We can use faster code for shift-by-1 under certain conditions.  */
 	cmp	$1,%ecx
 	jne	Lnormal
-	movl	%edi,%eax
-	subl	%esi,%eax
-	cmpl	%ebp,%eax
-	jnc	Lspecial
+	leal	4(%esi),%eax
+	cmpl	%edi,%eax
+	jnc	Lspecial		/* jump if s_ptr + 1 >= res_ptr */
+	leal	(%esi,%ebp,4),%eax
+	cmpl	%eax,%edi
+	jnc	Lspecial		/* jump if res_ptr >= s_ptr + size */
 
 Lnormal:
 	leal	-4(%edi,%ebp,4),%edi
diff --git a/sysdeps/i386/i586/rshift.S b/sysdeps/i386/i586/rshift.S
index 51cde8f..a820a79 100644
--- a/sysdeps/i386/i586/rshift.S
+++ b/sysdeps/i386/i586/rshift.S
@@ -43,12 +43,15 @@ C_SYMBOL_NAME(__mpn_rshift:)
 	movl	28(%esp),%ebp		/* size */
 	movl	32(%esp),%ecx		/* cnt */
 
+/* We can use faster code for shift-by-1 under certain conditions.  */
 	cmp	$1,%ecx
 	jne	Lnormal
-	movl	%edi,%eax
-	subl	%esi,%eax
-	cmpl	%ebp,%eax
-	jnc	Lspecial
+	leal	4(%edi),%eax
+	cmpl	%esi,%eax
+	jnc	Lspecial		/* jump if res_ptr + 1 >= s_ptr */
+	leal	(%edi,%ebp,4),%eax
+	cmpl	%eax,%esi
+	jnc	Lspecial		/* jump if s_ptr >= res_ptr + size */
 
 Lnormal:
 	movl	(%esi),%edx
diff --git a/sysdeps/m68k/add_n.S b/sysdeps/m68k/add_n.S
new file mode 100644
index 0000000..ea7a4458
--- /dev/null
+++ b/sysdeps/m68k/add_n.S
@@ -0,0 +1,76 @@
+/* mc68020 __mpn_add_n -- Add two limb vectors of the same length > 0 and store
+   sum in a third limb vector.
+
+Copyright (C) 1992, 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	(sp + 4)
+  s1_ptr	(sp + 8)
+  s2_ptr	(sp + 16)
+  size		(sp + 12)
+*/
+
+#include "asm-syntax.h"
+
+	TEXT
+	ALIGN
+	GLOBL	___mpn_add_n
+
+LAB(___mpn_add_n)
+/* Save used registers on the stack.  */
+	INSN2(move,l	,MEM_PREDEC(sp),d2)
+	INSN2(move,l	,MEM_PREDEC(sp),a2)
+
+/* Copy the arguments to registers.  Better use movem?  */
+	INSN2(move,l	,a2,MEM_DISP(sp,12))
+	INSN2(move,l	,a0,MEM_DISP(sp,16))
+	INSN2(move,l	,a1,MEM_DISP(sp,20))
+	INSN2(move,l	,d2,MEM_DISP(sp,24))
+
+	INSN2(eor,w	,d2,#1)
+	INSN2(lsr,l	,d2,#1)
+	bcc L1
+	INSN2(subq,l	,d2,#1)		/* clears cy as side effect */
+
+LAB(Loop)
+	INSN2(move,l	,d0,MEM_POSTINC(a0))
+	INSN2(move,l	,d1,MEM_POSTINC(a1))
+	INSN2(addx,l	,d0,d1)
+	INSN2(move,l	,MEM_POSTINC(a2),d0)
+LAB(L1)	INSN2(move,l	,d0,MEM_POSTINC(a0))
+	INSN2(move,l	,d1,MEM_POSTINC(a1))
+	INSN2(addx,l	,d0,d1)
+	INSN2(move,l	,MEM_POSTINC(a2),d0)
+
+	dbf d2,Loop			/* loop until 16 lsb of %4 == -1 */
+	INSN2(subx,l	,d0,d0)		/* d0 <= -cy; save cy as 0 or -1 in d0 */
+	INSN2(sub,l	,d2,#0x10000)
+	bcs L2
+	INSN2(add,l	,d0,d0)		/* restore cy */
+	bra Loop
+
+LAB(L2)
+	INSN1(neg,l	,d0)
+
+/* Restore used registers from stack frame.  */
+	INSN2(move,l	,a2,MEM_POSTINC(sp))
+	INSN2(move,l	,d2,MEM_POSTINC(sp))
+
+	rts
diff --git a/sysdeps/m68k/sub_n.S b/sysdeps/m68k/sub_n.S
new file mode 100644
index 0000000..19f0ec1
--- /dev/null
+++ b/sysdeps/m68k/sub_n.S
@@ -0,0 +1,76 @@
+/* mc68020 __mpn_sub_n -- Subtract two limb vectors of the same length > 0 and
+   store difference in a third limb vector.
+
+Copyright (C) 1992, 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	(sp + 4)
+  s1_ptr	(sp + 8)
+  s2_ptr	(sp + 16)
+  size		(sp + 12)
+*/
+
+#include "asm-syntax.h"
+
+	TEXT
+	ALIGN
+	GLOBL	___mpn_sub_n
+
+LAB(___mpn_sub_n)
+/* Save used registers on the stack.  */
+	INSN2(move,l	,MEM_PREDEC(sp),d2)
+	INSN2(move,l	,MEM_PREDEC(sp),a2)
+
+/* Copy the arguments to registers.  Better use movem?  */
+	INSN2(move,l	,a2,MEM_DISP(sp,12))
+	INSN2(move,l	,a0,MEM_DISP(sp,16))
+	INSN2(move,l	,a1,MEM_DISP(sp,20))
+	INSN2(move,l	,d2,MEM_DISP(sp,24))
+
+	INSN2(eor,w	,d2,#1)
+	INSN2(lsr,l	,d2,#1)
+	bcc L1
+	INSN2(subq,l	,d2,#1)		/* clears cy as side effect */
+
+LAB(Loop)
+	INSN2(move,l	,d0,MEM_POSTINC(a0))
+	INSN2(move,l	,d1,MEM_POSTINC(a1))
+	INSN2(subx,l	,d0,d1)
+	INSN2(move,l	,MEM_POSTINC(a2),d0)
+LAB(L1)	INSN2(move,l	,d0,MEM_POSTINC(a0))
+	INSN2(move,l	,d1,MEM_POSTINC(a1))
+	INSN2(subx,l	,d0,d1)
+	INSN2(move,l	,MEM_POSTINC(a2),d0)
+
+	dbf d2,Loop			/* loop until 16 lsb of %4 == -1 */
+	INSN2(subx,l	,d0,d0)		/* d0 <= -cy; save cy as 0 or -1 in d0 */
+	INSN2(sub,l	,d2,#0x10000)
+	bcs L2
+	INSN2(add,l	,d0,d0)		/* restore cy */
+	bra Loop
+
+LAB(L2)
+	INSN1(neg,l	,d0)
+
+/* Restore used registers from stack frame.  */
+	INSN2(move,l	,a2,MEM_POSTINC(sp))
+	INSN2(move,l	,d2,MEM_POSTINC(sp))
+
+	rts
diff --git a/sysdeps/m88k/add_n.s b/sysdeps/m88k/add_n.s
new file mode 100644
index 0000000..7e4cccc
--- /dev/null
+++ b/sysdeps/m88k/add_n.s
@@ -0,0 +1,103 @@
+; mc88100 __mpn_add -- Add two limb vectors of the same length > 0 and store
+; sum in a third limb vector.
+
+; Copyright (C) 1992, 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	r2
+; s1_ptr	r3
+; s2_ptr	r4
+; size		r5
+
+; This code has been optimized to run one instruction per clock, avoiding
+; load stalls and writeback contention.  As a result, the instruction
+; order is not always natural.
+
+; The speed is about 4.6 clocks/limb + 18 clocks/limb-vector on an 88100,
+; but on the 88110, it seems to run much slower, 6.6 clocks/limb.
+
+	text
+	align	 16
+	global	 ___mpn_add_n
+___mpn_add_n:
+	ld	r6,r3,0			; read first limb from s1_ptr
+	extu	r10,r5,3
+	ld	r7,r4,0			; read first limb from s2_ptr
+
+	subu.co	r5,r0,r5		; (clear carry as side effect)
+	mak	r5,r5,3<4>
+	bcnd	eq0,r5,Lzero
+
+	or	r12,r0,lo16(Lbase)
+	or.u	r12,r12,hi16(Lbase)
+	addu	r12,r12,r5		; r12 is address for entering in loop
+
+	extu	r5,r5,2			; divide by 4
+	subu	r2,r2,r5		; adjust res_ptr
+	subu	r3,r3,r5		; adjust s1_ptr
+	subu	r4,r4,r5		; adjust s2_ptr
+
+	or	r8,r6,r0
+
+	jmp.n	r12
+	 or	r9,r7,r0
+
+Loop:	addu	r3,r3,32
+	st	r8,r2,28
+	addu	r4,r4,32
+	ld	r6,r3,0
+	addu	r2,r2,32
+	ld	r7,r4,0
+Lzero:	subu	r10,r10,1		; add 0 + 8r limbs (adj loop cnt)
+Lbase:	ld	r8,r3,4
+	addu.cio r6,r6,r7
+	ld	r9,r4,4
+	st	r6,r2,0
+	ld	r6,r3,8			; add 7 + 8r limbs
+	addu.cio r8,r8,r9
+	ld	r7,r4,8
+	st	r8,r2,4
+	ld	r8,r3,12		; add 6 + 8r limbs
+	addu.cio r6,r6,r7
+	ld	r9,r4,12
+	st	r6,r2,8
+	ld	r6,r3,16		; add 5 + 8r limbs
+	addu.cio r8,r8,r9
+	ld	r7,r4,16
+	st	r8,r2,12
+	ld	r8,r3,20		; add 4 + 8r limbs
+	addu.cio r6,r6,r7
+	ld	r9,r4,20
+	st	r6,r2,16
+	ld	r6,r3,24		; add 3 + 8r limbs
+	addu.cio r8,r8,r9
+	ld	r7,r4,24
+	st	r8,r2,20
+	ld	r8,r3,28		; add 2 + 8r limbs
+	addu.cio r6,r6,r7
+	ld	r9,r4,28
+	st	r6,r2,24
+	bcnd.n	ne0,r10,Loop		; add 1 + 8r limbs
+	 addu.cio r8,r8,r9
+
+	st	r8,r2,28		; store most significant limb
+
+	jmp.n	 r1
+	 addu.ci r2,r0,r0		; return carry-out from most sign. limb
diff --git a/sysdeps/m88k/mul_1.s b/sysdeps/m88k/mul_1.s
new file mode 100644
index 0000000..35c238d
--- /dev/null
+++ b/sysdeps/m88k/mul_1.s
@@ -0,0 +1,128 @@
+; mc88100 __mpn_mul_1 -- Multiply a limb vector with a single limb and
+; store the product in a second limb vector.
+
+; Copyright (C) 1992, 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	r2
+; s1_ptr	r3
+; size		r4
+; s2_limb	r5
+
+; Common overhead is about 11 cycles/invocation.
+
+; The speed for S2_LIMB >= 0x10000 is approximately 21 cycles/limb.  (The
+; pipeline stalls 2 cycles due to WB contention.)
+
+; The speed for S2_LIMB < 0x10000 is approximately 16 cycles/limb.  (The
+; pipeline stalls 2 cycles due to WB contention and 1 cycle due to latency.)
+
+; To enhance speed:
+; 1. Unroll main loop 4-8 times.
+; 2. Schedule code to avoid WB contention.  It might be tempting to move the
+;    ld instruction in the loops down to save 2 cycles (less WB contention),
+;    but that looses because the ultimate value will be read from outside
+;    the allocated space.  But if we handle the ultimate multiplication in
+;    the tail, we can do this.
+; 3. Make the multiplication with less instructions.  I think the code for
+;    (S2_LIMB >= 0x10000) is not minimal.
+; With these techniques the (S2_LIMB >= 0x10000) case would run in 17 or
+; less cycles/limb; the (S2_LIMB < 0x10000) case would run in 11
+; cycles/limb.  (Assuming infinite unrolling.)
+
+	text
+	align	 16
+	global	 ___mpn_mul_1
+___mpn_mul_1:
+
+	; Make S1_PTR and RES_PTR point at the end of their blocks
+	; and negate SIZE.
+	lda	 r3,r3[r4]
+	lda	 r6,r2[r4]		; RES_PTR in r6 since r2 is retval
+	subu	 r4,r0,r4
+
+	addu.co	 r2,r0,r0		; r2 = cy = 0
+	ld	 r9,r3[r4]
+	mask	 r7,r5,0xffff		; r7 = lo(S2_LIMB)
+	extu	 r8,r5,16		; r8 = hi(S2_LIMB)
+	bcnd.n	 eq0,r8,Lsmall		; jump if (hi(S2_LIMB) == 0)
+	 subu	 r6,r6,4
+
+; General code for any value of S2_LIMB.
+
+	; Make a stack frame and save r25 and r26
+	subu	 r31,r31,16
+	st.d	 r25,r31,8
+
+	; Enter the loop in the middle
+	br.n	L1
+	addu	 r4,r4,1
+
+Loop:
+	ld	 r9,r3[r4]
+	st	 r26,r6[r4]
+; bcnd	ne0,r0,0			; bubble
+	addu	 r4,r4,1
+L1:	mul	 r26,r9,r5		; low word of product	mul_1	WB ld
+	mask	 r12,r9,0xffff		; r12 = lo(s1_limb)	mask_1
+	mul	 r11,r12,r7		; r11 =  prod_0		mul_2	WB mask_1
+	mul	 r10,r12,r8		; r10 = prod_1a		mul_3
+	extu	 r13,r9,16		; r13 = hi(s1_limb)	extu_1	WB mul_1
+	mul	 r12,r13,r7		; r12 = prod_1b		mul_4	WB extu_1
+	mul	 r25,r13,r8		; r25  = prod_2		mul_5	WB mul_2
+	extu	 r11,r11,16		; r11 = hi(prod_0)	extu_2	WB mul_3
+	addu	 r10,r10,r11		;			addu_1	WB extu_2
+; bcnd	ne0,r0,0			; bubble			WB addu_1
+	addu.co	 r10,r10,r12		;				WB mul_4
+	mask.u	 r10,r10,0xffff		; move the 16 most significant bits...
+	addu.ci	 r10,r10,r0		; ...to the low half of the word...
+	rot	 r10,r10,16		; ...and put carry in pos 16.
+	addu.co	 r26,r26,r2		; add old carry limb
+	bcnd.n	 ne0,r4,Loop
+	 addu.ci r2,r25,r10		; compute new carry limb
+
+	st	 r26,r6[r4]
+	ld.d	 r25,r31,8
+	jmp.n	 r1
+	 addu	 r31,r31,16
+
+; Fast code for S2_LIMB < 0x10000
+Lsmall:
+	; Enter the loop in the middle
+	br.n	SL1
+	addu	 r4,r4,1
+
+SLoop:
+	ld	 r9,r3[r4]		;
+	st	 r8,r6[r4]		;
+	addu	 r4,r4,1		;
+SL1:	mul	 r8,r9,r5		; low word of product
+	mask	 r12,r9,0xffff		; r12 = lo(s1_limb)
+	extu	 r13,r9,16		; r13 = hi(s1_limb)
+	mul	 r11,r12,r7		; r11 =  prod_0
+	mul	 r12,r13,r7		; r12 = prod_1b
+	addu.cio r8,r8,r2		; add old carry limb
+	extu	 r10,r11,16		; r11 = hi(prod_0)
+	addu	 r10,r10,r12		;
+	bcnd.n	 ne0,r4,SLoop
+	extu	 r2,r10,16		; r2 = new carry limb
+
+	jmp.n	 r1
+	st	 r8,r6[r4]
diff --git a/sysdeps/m88k/sub_n.s b/sysdeps/m88k/sub_n.s
new file mode 100644
index 0000000..3963cd5
--- /dev/null
+++ b/sysdeps/m88k/sub_n.s
@@ -0,0 +1,104 @@
+; mc88100 __mpn_sub -- Subtract two limb vectors of the same length > 0 and
+; store difference in a third limb vector.
+
+; Copyright (C) 1992, 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	r2
+; s1_ptr	r3
+; s2_ptr	r4
+; size		r5
+
+; This code has been optimized to run one instruction per clock, avoiding
+; load stalls and writeback contention.  As a result, the instruction
+; order is not always natural.
+
+; The speed is about 4.6 clocks/limb + 18 clocks/limb-vector on an 88100,
+; but on the 88110, it seems to run much slower, 6.6 clocks/limb.
+
+	text
+	align	 16
+	global	 ___mpn_sub_n
+___mpn_sub_n:
+	ld	r6,r3,0			; read first limb from s1_ptr
+	extu	r10,r5,3
+	ld	r7,r4,0			; read first limb from s2_ptr
+
+	subu.co	r5,r0,r5		; (clear carry as side effect)
+	mak	r5,r5,3<4>
+	bcnd	eq0,r5,Lzero
+
+	or	r12,r0,lo16(Lbase)
+	or.u	r12,r12,hi16(Lbase)
+	addu	r12,r12,r5		; r12 is address for entering in loop
+
+	extu	r5,r5,2			; divide by 4
+	subu	r2,r2,r5		; adjust res_ptr
+	subu	r3,r3,r5		; adjust s1_ptr
+	subu	r4,r4,r5		; adjust s2_ptr
+
+	or	r8,r6,r0
+
+	jmp.n	r12
+	 or	r9,r7,r0
+
+Loop:	addu	r3,r3,32
+	st	r8,r2,28
+	addu	r4,r4,32
+	ld	r6,r3,0
+	addu	r2,r2,32
+	ld	r7,r4,0
+Lzero:	subu	r10,r10,1		; subtract 0 + 8r limbs (adj loop cnt)
+Lbase:	ld	r8,r3,4
+	subu.cio r6,r6,r7
+	ld	r9,r4,4
+	st	r6,r2,0
+	ld	r6,r3,8			; subtract 7 + 8r limbs
+	subu.cio r8,r8,r9
+	ld	r7,r4,8
+	st	r8,r2,4
+	ld	r8,r3,12		; subtract 6 + 8r limbs
+	subu.cio r6,r6,r7
+	ld	r9,r4,12
+	st	r6,r2,8
+	ld	r6,r3,16		; subtract 5 + 8r limbs
+	subu.cio r8,r8,r9
+	ld	r7,r4,16
+	st	r8,r2,12
+	ld	r8,r3,20		; subtract 4 + 8r limbs
+	subu.cio r6,r6,r7
+	ld	r9,r4,20
+	st	r6,r2,16
+	ld	r6,r3,24		; subtract 3 + 8r limbs
+	subu.cio r8,r8,r9
+	ld	r7,r4,24
+	st	r8,r2,20
+	ld	r8,r3,28		; subtract 2 + 8r limbs
+	subu.cio r6,r6,r7
+	ld	r9,r4,28
+	st	r6,r2,24
+	bcnd.n	ne0,r10,Loop		; subtract 1 + 8r limbs
+	 subu.cio r8,r8,r9
+
+	st	r8,r2,28		; store most significant limb
+
+	addu.ci r2,r0,r0		; return carry-out from most sign. limb
+	jmp.n	 r1
+	 xor	r2,r2,1
diff --git a/sysdeps/rs6000/add_n.s b/sysdeps/rs6000/add_n.s
index 34ad9e1..7090cf1 100644
--- a/sysdeps/rs6000/add_n.s
+++ b/sysdeps/rs6000/add_n.s
@@ -45,7 +45,7 @@ __mpn_add_n:
 	bdz	Lend		# If done, skip loop
 Loop:	lu	8,4(4)		# load s1 limb and update s1_ptr
 	lu	0,4(5)		# load s2 limb and update s2_ptr
-	stu	7,4(3)		# store previous limb in load latecny slot
+	stu	7,4(3)		# store previous limb in load latency slot
 	ae	7,0,8		# add new limbs with cy, set cy
 	bdn	Loop		# decrement CTR and loop back
 Lend:	st	7,4(3)		# store ultimate result limb
diff --git a/sysdeps/rs6000/sub_n.s b/sysdeps/rs6000/sub_n.s
index 402fdce..40fe7d6 100644
--- a/sysdeps/rs6000/sub_n.s
+++ b/sysdeps/rs6000/sub_n.s
@@ -46,7 +46,7 @@ __mpn_sub_n:
 	bdz	Lend		# If done, skip loop
 Loop:	lu	8,4(4)		# load s1 limb and update s1_ptr
 	lu	0,4(5)		# load s2 limb and update s2_ptr
-	stu	7,4(3)		# store previous limb in load latecny slot
+	stu	7,4(3)		# store previous limb in load latency slot
 	sfe	7,0,8		# add new limbs with cy, set cy
 	bdn	Loop		# decrement CTR and loop back
 Lend:	st	7,4(3)		# store ultimate result limb
diff --git a/sysdeps/sparc/add_n.S b/sysdeps/sparc/add_n.S
index 13704d3..80c3b99 100644
--- a/sysdeps/sparc/add_n.S
+++ b/sysdeps/sparc/add_n.S
@@ -1,7 +1,7 @@
 ! sparc __mpn_add_n -- Add two limb vectors of the same length > 0 and store
 ! sum in a third limb vector.
 
-! Copyright (C) 1992, 1994, 1995 Free Software Foundation, Inc.
+! Copyright (C) 1995 Free Software Foundation, Inc.
 
 ! This file is part of the GNU MP Library.
 
@@ -21,10 +21,10 @@
 
 
 ! INPUT PARAMETERS
-! res_ptr	%o0
-! s1_ptr	%o1
-! s2_ptr	%o2
-! size		%o3
+#define res_ptr	%o0
+#define s1_ptr	%o1
+#define s2_ptr	%o2
+#define size	%o3
 
 #include "sysdep.h"
 
@@ -32,108 +32,192 @@
 	.align	4
 	.global	C_SYMBOL_NAME(__mpn_add_n)
 C_SYMBOL_NAME(__mpn_add_n):
-	ld	[%o1+0],%o4		! read first limb from s1_ptr
-	srl	%o3,4,%g1
-	ld	[%o2+0],%o5		! read first limb from s2_ptr
-
-	sub	%g0,%o3,%o3
-	andcc	%o3,(16-1),%o3
-	be	Lzero
-	 mov	%o4,%g2			! put first s1_limb in g2 too
-
-	sll	%o3,2,%o3		! multiply by 4
-	sub	%o0,%o3,%o0		! adjust res_ptr
-	sub	%o1,%o3,%o1		! adjust s1_ptr
-	sub	%o2,%o3,%o2		! adjust s2_ptr
-
-#if PIC
-	mov	%o7,%g4			! Save return address register
-	call	1f
-	add	%o7,Lbase-1f,%g3
-1:	mov	%g4,%o7			! Restore return address register
-#else
-	sethi	%hi(Lbase),%g3
-	or	%g3,%lo(Lbase),%g3
-#endif
-	sll	%o3,2,%o3		! multiply by 4
-	jmp	%g3+%o3
-	 mov	%o5,%g3			! put first s2_limb in g3 too
-
-Loop:	addxcc	%g2,%g3,%o3
-	add	%o1,64,%o1
-	st	%o3,[%o0+60]
-	add	%o2,64,%o2
-	ld	[%o1+0],%o4
-	add	%o0,64,%o0
-	ld	[%o2+0],%o5
-Lzero:	sub	%g1,1,%g1	! add 0 + 16r limbs (adjust loop counter)
-Lbase:	ld	[%o1+4],%g2
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+4],%g3
-	st	%o3,[%o0+0]
-	ld	[%o1+8],%o4	! add 15 + 16r limbs
-	addxcc	%g2,%g3,%o3
-	ld	[%o2+8],%o5
-	st	%o3,[%o0+4]
-	ld	[%o1+12],%g2	! add 14 + 16r limbs
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+12],%g3
-	st	%o3,[%o0+8]
-	ld	[%o1+16],%o4	! add 13 + 16r limbs
-	addxcc	%g2,%g3,%o3
-	ld	[%o2+16],%o5
-	st	%o3,[%o0+12]
-	ld	[%o1+20],%g2	! add 12 + 16r limbs
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+20],%g3
-	st	%o3,[%o0+16]
-	ld	[%o1+24],%o4	! add 11 + 16r limbs
-	addxcc	%g2,%g3,%o3
-	ld	[%o2+24],%o5
-	st	%o3,[%o0+20]
-	ld	[%o1+28],%g2	! add 10 + 16r limbs
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+28],%g3
-	st	%o3,[%o0+24]
-	ld	[%o1+32],%o4	! add 9 + 16r limbs
-	addxcc	%g2,%g3,%o3
-	ld	[%o2+32],%o5
-	st	%o3,[%o0+28]
-	ld	[%o1+36],%g2	! add 8 + 16r limbs
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+36],%g3
-	st	%o3,[%o0+32]
-	ld	[%o1+40],%o4	! add 7 + 16r limbs
-	addxcc	%g2,%g3,%o3
-	ld	[%o2+40],%o5
-	st	%o3,[%o0+36]
-	ld	[%o1+44],%g2	! add 6 + 16r limbs
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+44],%g3
-	st	%o3,[%o0+40]
-	ld	[%o1+48],%o4	! add 5 + 16r limbs
-	addxcc	%g2,%g3,%o3
-	ld	[%o2+48],%o5
-	st	%o3,[%o0+44]
-	ld	[%o1+52],%g2	! add 4 + 16r limbs
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+52],%g3
-	st	%o3,[%o0+48]
-	ld	[%o1+56],%o4	! add 3 + 16r limbs
-	addxcc	%g2,%g3,%o3
-	ld	[%o2+56],%o5
-	st	%o3,[%o0+52]
-	ld	[%o1+60],%g2	! add 2 + 16r limbs
-	addxcc	%o4,%o5,%o3
-	ld	[%o2+60],%g3
-	st	%o3,[%o0+56]
-	addx	%g0,%g0,%o4
-	tst	%g1
-	bne	Loop
-	 subcc	%g0,%o4,%g0	! restore cy (delay slot)
-
-	addxcc	%g2,%g3,%o3
-	st	%o3,[%o0+60]	! store most significant limb
-
-	retl
-	 addx	%g0,%g0,%o0	! return carry-out from most sign. limb
+	cmp	size,8
+	mov	0,%o4			! clear cy-save register
+	blt,a	Ltriv
+	addcc	size,-2,size
+	xor	s2_ptr,res_ptr,%g1
+	andcc	%g1,4,%g0
+	bne	L1			! branch if alignment differs
+	nop
+L0:	andcc	res_ptr,4,%g0		! res_ptr unaligned? Side effect: cy=0
+	beq	L_v1			! if no, branch
+	nop
+! **  V1a  **
+/* Add least significant limb separately to align res_ptr and s2_ptr */
+	ld	[s1_ptr],%g4
+	add	s1_ptr,4,s1_ptr
+	ld	[s2_ptr],%g2
+	add	s2_ptr,4,s2_ptr
+	add	size,-1,size
+	addcc	%g4,%g2,%o4
+	st	%o4,[res_ptr]
+	add	res_ptr,4,res_ptr
+
+L_v1:	ld	[s1_ptr+0],%g4
+	ld	[s1_ptr+4],%g1
+	ldd	[s2_ptr+0],%g2
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-10,size
+	blt	Lfin1
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 8 limbs until less than 8 limbs remain */
+Loop1:	addxcc	%g4,%g2,%o4
+	ld	[s1_ptr+8],%g4
+	addxcc	%g1,%g3,%o5
+	ld	[s1_ptr+12],%g1
+	ldd	[s2_ptr+8],%g2
+	std	%o4,[res_ptr+0]
+	addxcc	%g4,%g2,%o4
+	ld	[s1_ptr+16],%g4
+	addxcc	%g1,%g3,%o5
+	ld	[s1_ptr+20],%g1
+	ldd	[s2_ptr+16],%g2
+	std	%o4,[res_ptr+8]
+	addxcc	%g4,%g2,%o4
+	ld	[s1_ptr+24],%g4
+	addxcc	%g1,%g3,%o5
+	ld	[s1_ptr+28],%g1
+	ldd	[s2_ptr+24],%g2
+	std	%o4,[res_ptr+16]
+	addxcc	%g4,%g2,%o4
+	ld	[s1_ptr+32],%g4
+	addxcc	%g1,%g3,%o5
+	ld	[s1_ptr+36],%g1
+	ldd	[s2_ptr+32],%g2
+	std	%o4,[res_ptr+24]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-8,size
+	add	s1_ptr,32,s1_ptr
+	add	s2_ptr,32,s2_ptr
+	add	res_ptr,32,res_ptr
+	bge	Loop1
+	subcc	%g0,%o4,%g0		! restore cy
+
+Lfin1:	addcc	size,8-2,size
+	blt	Lend1
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 2 limbs until less than 2 limbs remain */
+Loop1b:	addxcc	%g4,%g2,%o4
+	ld	[s1_ptr+8],%g4
+	addxcc	%g1,%g3,%o5
+	ld	[s1_ptr+12],%g1
+	ldd	[s2_ptr+8],%g2
+	std	%o4,[res_ptr+0]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-2,size
+	add	s1_ptr,8,s1_ptr
+	add	s2_ptr,8,s2_ptr
+	add	res_ptr,8,res_ptr
+	bge	Loop1b
+	subcc	%g0,%o4,%g0		! restore cy
+Lend1:	addxcc	%g4,%g2,%o4
+	addxcc	%g1,%g3,%o5
+	std	%o4,[res_ptr+0]
+	addx	%g0,%g0,%o4		! save cy in register
+
+	andcc	size,1,%g0
+	be	Lret1
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add last limb */
+	ld	[s1_ptr+8],%g4
+	ld	[s2_ptr+8],%g2
+	addxcc	%g4,%g2,%o4
+	st	%o4,[res_ptr+8]
+
+Lret1:	retl
+	addx	%g0,%g0,%o0	! return carry-out from most sign. limb
+
+L1:	xor	s1_ptr,res_ptr,%g1
+	andcc	%g1,4,%g0
+	bne	L2
+	nop
+! **  V1b  **
+	mov	s2_ptr,%g1
+	mov	s1_ptr,s2_ptr
+	b	L0
+	mov	%g1,s1_ptr
+
+! **  V2  **
+/* If we come here, the alignment of s1_ptr and res_ptr as well as the
+   alignment of s2_ptr and res_ptr differ.  Since there are only two ways
+   things can be aligned (that we care about) we now know that the alignment
+   of s1_ptr and s2_ptr are the same.  */
+
+L2:	andcc	s1_ptr,4,%g0		! s1_ptr unaligned? Side effect: cy=0
+	beq	L_v2			! if no, branch
+	nop
+/* Add least significant limb separately to align res_ptr and s2_ptr */
+	ld	[s1_ptr],%g4
+	add	s1_ptr,4,s1_ptr
+	ld	[s2_ptr],%g2
+	add	s2_ptr,4,s2_ptr
+	add	size,-1,size
+	addcc	%g4,%g2,%o4
+	st	%o4,[res_ptr]
+	add	res_ptr,4,res_ptr
+
+L_v2:	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-8,size
+	blt	Lfin2
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 8 limbs until less than 8 limbs remain */
+Loop2:	ldd	[s1_ptr+0],%g2
+	ldd	[s2_ptr+0],%o4
+	addxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+0]
+	addxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+4]
+	ldd	[s1_ptr+8],%g2
+	ldd	[s2_ptr+8],%o4
+	addxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+8]
+	addxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+12]
+	ldd	[s1_ptr+16],%g2
+	ldd	[s2_ptr+16],%o4
+	addxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+16]
+	addxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+20]
+	ldd	[s1_ptr+24],%g2
+	ldd	[s2_ptr+24],%o4
+	addxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+24]
+	addxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+28]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-8,size
+	add	s1_ptr,32,s1_ptr
+	add	s2_ptr,32,s2_ptr
+	add	res_ptr,32,res_ptr
+	bge	Loop2
+	subcc	%g0,%o4,%g0		! restore cy
+
+Lfin2:	addcc	size,8-2,size
+Ltriv:	blt	Lend2
+	subcc	%g0,%o4,%g0		! restore cy
+Loop2b:	ldd	[s1_ptr+0],%g2
+	ldd	[s2_ptr+0],%o4
+	addxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+0]
+	addxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+4]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-2,size
+	add	s1_ptr,8,s1_ptr
+	add	s2_ptr,8,s2_ptr
+	add	res_ptr,8,res_ptr
+	bge	Loop2b
+	subcc	%g0,%o4,%g0		! restore cy
+Lend2:	andcc	size,1,%g0
+	be	Lret2
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add last limb */
+	ld	[s1_ptr],%g4
+	ld	[s2_ptr],%g2
+	addxcc	%g4,%g2,%o4
+	st	%o4,[res_ptr]
+
+Lret2:	retl
+	addx	%g0,%g0,%o0	! return carry-out from most sign. limb
diff --git a/sysdeps/sparc/lshift.S b/sysdeps/sparc/lshift.S
new file mode 100644
index 0000000..497272a
--- /dev/null
+++ b/sysdeps/sparc/lshift.S
@@ -0,0 +1,94 @@
+! sparc __mpn_lshift --
+
+! Copyright (C) 1995 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	%o0
+! src_ptr	%o1
+! size		%o2
+! cnt		%o3
+
+#include "sysdep.h"
+
+	.text
+	.align	4
+	.global	C_SYMBOL_NAME(__mpn_lshift)
+C_SYMBOL_NAME(__mpn_lshift):
+	sll	%o2,2,%g1
+	add	%o1,%g1,%o1	! make %o1 point at end of src
+	ld	[%o1-4],%g2	! load first limb
+	sub	%g0,%o3,%o5	! negate shift count
+	add	%o0,%g1,%o0	! make %o0 point at end of res
+	add	%o2,-1,%o2
+	andcc	%o2,4-1,%g4	! number of limbs in first loop
+	srl	%g2,%o5,%g1	! compute function result
+	beq	L0		! if multiple of 4 limbs, skip first loop
+	st	%g1,[%sp+80]
+
+	sub	%o2,%g4,%o2	! adjust count for main loop
+
+Loop0:	ld	[%o1-8],%g3
+	add	%o0,-4,%o0
+	add	%o1,-4,%o1
+	addcc	%g4,-1,%g4
+	sll	%g2,%o3,%o4
+	srl	%g3,%o5,%g1
+	mov	%g3,%g2
+	or	%o4,%g1,%o4
+	bne	Loop0
+	 st	%o4,[%o0+0]
+
+L0:	tst	%o2
+	beq	Lend
+	 nop
+
+Loop:	ld	[%o1-8],%g3
+	add	%o0,-16,%o0
+	addcc	%o2,-4,%o2
+	sll	%g2,%o3,%o4
+	srl	%g3,%o5,%g1
+
+	ld	[%o1-12],%g2
+	sll	%g3,%o3,%g4
+	or	%o4,%g1,%o4
+	st	%o4,[%o0+12]
+	srl	%g2,%o5,%g1
+
+	ld	[%o1-16],%g3
+	sll	%g2,%o3,%o4
+	or	%g4,%g1,%g4
+	st	%g4,[%o0+8]
+	srl	%g3,%o5,%g1
+
+	ld	[%o1-20],%g2
+	sll	%g3,%o3,%g4
+	or	%o4,%g1,%o4
+	st	%o4,[%o0+4]
+	srl	%g2,%o5,%g1
+
+	add	%o1,-16,%o1
+	or	%g4,%g1,%g4
+	bne	Loop
+	 st	%g4,[%o0+0]
+
+Lend:	sll	%g2,%o3,%g2
+	st	%g2,[%o0-4]
+	retl
+	ld	[%sp+80],%o0
diff --git a/sysdeps/sparc/rshift.S b/sysdeps/sparc/rshift.S
new file mode 100644
index 0000000..3428cfe
--- /dev/null
+++ b/sysdeps/sparc/rshift.S
@@ -0,0 +1,91 @@
+! sparc __mpn_rshift --
+
+! Copyright (C) 1995 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	%o0
+! src_ptr	%o1
+! size		%o2
+! cnt		%o3
+
+#include "sysdep.h"
+
+	.text
+	.align	4
+	.global	C_SYMBOL_NAME(__mpn_rshift)
+C_SYMBOL_NAME(__mpn_rshift):
+	ld	[%o1],%g2	! load first limb
+	sub	%g0,%o3,%o5	! negate shift count
+	add	%o2,-1,%o2
+	andcc	%o2,4-1,%g4	! number of limbs in first loop
+	sll	%g2,%o5,%g1	! compute function result
+	beq	L0		! if multiple of 4 limbs, skip first loop
+	st	%g1,[%sp+80]
+
+	sub	%o2,%g4,%o2	! adjust count for main loop
+
+Loop0:	ld	[%o1+4],%g3
+	add	%o0,4,%o0
+	add	%o1,4,%o1
+	addcc	%g4,-1,%g4
+	srl	%g2,%o3,%o4
+	sll	%g3,%o5,%g1
+	mov	%g3,%g2
+	or	%o4,%g1,%o4
+	bne	Loop0
+	 st	%o4,[%o0-4]
+
+L0:	tst	%o2
+	beq	Lend
+	 nop
+
+Loop:	ld	[%o1+4],%g3
+	add	%o0,16,%o0
+	addcc	%o2,-4,%o2
+	srl	%g2,%o3,%o4
+	sll	%g3,%o5,%g1
+
+	ld	[%o1+8],%g2
+	srl	%g3,%o3,%g4
+	or	%o4,%g1,%o4
+	st	%o4,[%o0-16]
+	sll	%g2,%o5,%g1
+
+	ld	[%o1+12],%g3
+	srl	%g2,%o3,%o4
+	or	%g4,%g1,%g4
+	st	%g4,[%o0-12]
+	sll	%g3,%o5,%g1
+
+	ld	[%o1+16],%g2
+	srl	%g3,%o3,%g4
+	or	%o4,%g1,%o4
+	st	%o4,[%o0-8]
+	sll	%g2,%o5,%g1
+
+	add	%o1,16,%o1
+	or	%g4,%g1,%g4
+	bne	Loop
+	 st	%g4,[%o0-4]
+
+Lend:	srl	%g2,%o3,%g2
+	st	%g2,[%o0-0]
+	retl
+	ld	[%sp+80],%o0
diff --git a/sysdeps/sparc/sub_n.S b/sysdeps/sparc/sub_n.S
index 6264344..2e217ed 100644
--- a/sysdeps/sparc/sub_n.S
+++ b/sysdeps/sparc/sub_n.S
@@ -1,7 +1,7 @@
 ! sparc __mpn_sub_n -- Subtract two limb vectors of the same length > 0 and
 ! store difference in a third limb vector.
 
-! Copyright (C) 1992, 1994, 1995 Free Software Foundation, Inc.
+! Copyright (C) 1995 Free Software Foundation, Inc.
 
 ! This file is part of the GNU MP Library.
 
@@ -21,10 +21,10 @@
 
 
 ! INPUT PARAMETERS
-! res_ptr	%o0
-! s1_ptr	%o1
-! s2_ptr	%o2
-! size		%o3
+#define res_ptr	%o0
+#define s1_ptr	%o1
+#define s2_ptr	%o2
+#define size	%o3
 
 #include "sysdep.h"
 
@@ -32,108 +32,279 @@
 	.align	4
 	.global	C_SYMBOL_NAME(__mpn_sub_n)
 C_SYMBOL_NAME(__mpn_sub_n):
-	ld	[%o1+0],%o4		! read first limb from s1_ptr
-	srl	%o3,4,%g1
-	ld	[%o2+0],%o5		! read first limb from s2_ptr
-
-	sub	%g0,%o3,%o3
-	andcc	%o3,(16-1),%o3
-	be	Lzero
-	 mov	%o4,%g2			! put first s1_limb in g2 too
-
-	sll	%o3,2,%o3		! multiply by 4
-	sub	%o0,%o3,%o0		! adjust res_ptr
-	sub	%o1,%o3,%o1		! adjust s1_ptr
-	sub	%o2,%o3,%o2		! adjust s2_ptr
-
-#if PIC
-	mov	%o7,%g4			! Save return address register
-	call	1f
-	add	%o7,Lbase-1f,%g3
-1:	mov	%g4,%o7			! Restore return address register
-#else
-	sethi	%hi(Lbase),%g3
-	or	%g3,%lo(Lbase),%g3
-#endif
-	sll	%o3,2,%o3		! multiply by 4
-	jmp	%g3+%o3
-	 mov	%o5,%g3			! put first s2_limb in g3 too
-
-Loop:	subxcc	%g2,%g3,%o3
-	add	%o1,64,%o1
-	st	%o3,[%o0+60]
-	add	%o2,64,%o2
-	ld	[%o1+0],%o4
-	add	%o0,64,%o0
-	ld	[%o2+0],%o5
-Lzero:	sub	%g1,1,%g1	! add 0 + 16r limbs (adjust loop counter)
-Lbase:	ld	[%o1+4],%g2
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+4],%g3
-	st	%o3,[%o0+0]
-	ld	[%o1+8],%o4	! add 15 + 16r limbs
-	subxcc	%g2,%g3,%o3
-	ld	[%o2+8],%o5
-	st	%o3,[%o0+4]
-	ld	[%o1+12],%g2	! add 14 + 16r limbs
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+12],%g3
-	st	%o3,[%o0+8]
-	ld	[%o1+16],%o4	! add 13 + 16r limbs
-	subxcc	%g2,%g3,%o3
-	ld	[%o2+16],%o5
-	st	%o3,[%o0+12]
-	ld	[%o1+20],%g2	! add 12 + 16r limbs
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+20],%g3
-	st	%o3,[%o0+16]
-	ld	[%o1+24],%o4	! add 11 + 16r limbs
-	subxcc	%g2,%g3,%o3
-	ld	[%o2+24],%o5
-	st	%o3,[%o0+20]
-	ld	[%o1+28],%g2	! add 10 + 16r limbs
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+28],%g3
-	st	%o3,[%o0+24]
-	ld	[%o1+32],%o4	! add 9 + 16r limbs
-	subxcc	%g2,%g3,%o3
-	ld	[%o2+32],%o5
-	st	%o3,[%o0+28]
-	ld	[%o1+36],%g2	! add 8 + 16r limbs
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+36],%g3
-	st	%o3,[%o0+32]
-	ld	[%o1+40],%o4	! add 7 + 16r limbs
-	subxcc	%g2,%g3,%o3
-	ld	[%o2+40],%o5
-	st	%o3,[%o0+36]
-	ld	[%o1+44],%g2	! add 6 + 16r limbs
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+44],%g3
-	st	%o3,[%o0+40]
-	ld	[%o1+48],%o4	! add 5 + 16r limbs
-	subxcc	%g2,%g3,%o3
-	ld	[%o2+48],%o5
-	st	%o3,[%o0+44]
-	ld	[%o1+52],%g2	! add 4 + 16r limbs
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+52],%g3
-	st	%o3,[%o0+48]
-	ld	[%o1+56],%o4	! add 3 + 16r limbs
-	subxcc	%g2,%g3,%o3
-	ld	[%o2+56],%o5
-	st	%o3,[%o0+52]
-	ld	[%o1+60],%g2	! add 2 + 16r limbs
-	subxcc	%o4,%o5,%o3
-	ld	[%o2+60],%g3
-	st	%o3,[%o0+56]
-	subx	%g0,%g0,%o4
-	tst	%g1
-	bne	Loop
-	 subcc	%g0,%o4,%g0	! restore cy (delay slot)
-
-	subxcc	%g2,%g3,%o3
-	st	%o3,[%o0+60]	! store most significant limb
-
-	retl
-	 addx	%g0,%g0,%o0	! return carry-out from most sign. limb
+	xor	s2_ptr,res_ptr,%g1
+	andcc	%g1,4,%g0
+	bne	L1			! branch if alignment differs
+	nop
+! **  V1a  **
+	andcc	res_ptr,4,%g0		! res_ptr unaligned? Side effect: cy=0
+	beq	L_v1			! if no, branch
+	nop
+/* Add least significant limb separately to align res_ptr and s2_ptr */
+	ld	[s1_ptr],%g4
+	add	s1_ptr,4,s1_ptr
+	ld	[s2_ptr],%g2
+	add	s2_ptr,4,s2_ptr
+	add	size,-1,size
+	subcc	%g4,%g2,%o4
+	st	%o4,[res_ptr]
+	add	res_ptr,4,res_ptr
+L_v1:	addx	%g0,%g0,%o4		! save cy in register
+	cmp	size,2			! if size < 2 ...
+	bl	Lend2			! ... branch to tail code
+	subcc	%g0,%o4,%g0		! restore cy
+
+	ld	[s1_ptr+0],%g4
+	addcc	size,-10,size
+	ld	[s1_ptr+4],%g1
+	ldd	[s2_ptr+0],%g2
+	blt	Lfin1
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 8 limbs until less than 8 limbs remain */
+Loop1:	subxcc	%g4,%g2,%o4
+	ld	[s1_ptr+8],%g4
+	subxcc	%g1,%g3,%o5
+	ld	[s1_ptr+12],%g1
+	ldd	[s2_ptr+8],%g2
+	std	%o4,[res_ptr+0]
+	subxcc	%g4,%g2,%o4
+	ld	[s1_ptr+16],%g4
+	subxcc	%g1,%g3,%o5
+	ld	[s1_ptr+20],%g1
+	ldd	[s2_ptr+16],%g2
+	std	%o4,[res_ptr+8]
+	subxcc	%g4,%g2,%o4
+	ld	[s1_ptr+24],%g4
+	subxcc	%g1,%g3,%o5
+	ld	[s1_ptr+28],%g1
+	ldd	[s2_ptr+24],%g2
+	std	%o4,[res_ptr+16]
+	subxcc	%g4,%g2,%o4
+	ld	[s1_ptr+32],%g4
+	subxcc	%g1,%g3,%o5
+	ld	[s1_ptr+36],%g1
+	ldd	[s2_ptr+32],%g2
+	std	%o4,[res_ptr+24]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-8,size
+	add	s1_ptr,32,s1_ptr
+	add	s2_ptr,32,s2_ptr
+	add	res_ptr,32,res_ptr
+	bge	Loop1
+	subcc	%g0,%o4,%g0		! restore cy
+
+Lfin1:	addcc	size,8-2,size
+	blt	Lend1
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 2 limbs until less than 2 limbs remain */
+Loope1:	subxcc	%g4,%g2,%o4
+	ld	[s1_ptr+8],%g4
+	subxcc	%g1,%g3,%o5
+	ld	[s1_ptr+12],%g1
+	ldd	[s2_ptr+8],%g2
+	std	%o4,[res_ptr+0]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-2,size
+	add	s1_ptr,8,s1_ptr
+	add	s2_ptr,8,s2_ptr
+	add	res_ptr,8,res_ptr
+	bge	Loope1
+	subcc	%g0,%o4,%g0		! restore cy
+Lend1:	subxcc	%g4,%g2,%o4
+	subxcc	%g1,%g3,%o5
+	std	%o4,[res_ptr+0]
+	addx	%g0,%g0,%o4		! save cy in register
+
+	andcc	size,1,%g0
+	be	Lret1
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add last limb */
+	ld	[s1_ptr+8],%g4
+	ld	[s2_ptr+8],%g2
+	subxcc	%g4,%g2,%o4
+	st	%o4,[res_ptr+8]
+
+Lret1:	retl
+	addx	%g0,%g0,%o0	! return carry-out from most sign. limb
+
+L1:	xor	s1_ptr,res_ptr,%g1
+	andcc	%g1,4,%g0
+	bne	L2
+	nop
+! **  V1b  **
+	andcc	res_ptr,4,%g0		! res_ptr unaligned? Side effect: cy=0
+	beq	L_v1b			! if no, branch
+	nop
+/* Add least significant limb separately to align res_ptr and s2_ptr */
+	ld	[s2_ptr],%g4
+	add	s2_ptr,4,s2_ptr
+	ld	[s1_ptr],%g2
+	add	s1_ptr,4,s1_ptr
+	add	size,-1,size
+	subcc	%g2,%g4,%o4
+	st	%o4,[res_ptr]
+	add	res_ptr,4,res_ptr
+L_v1b:	addx	%g0,%g0,%o4		! save cy in register
+	cmp	size,2			! if size < 2 ...
+	bl	Lend2			! ... branch to tail code
+	subcc	%g0,%o4,%g0		! restore cy
+
+	ld	[s2_ptr+0],%g4
+	addcc	size,-10,size
+	ld	[s2_ptr+4],%g1
+	ldd	[s1_ptr+0],%g2
+	blt	Lfin1b
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 8 limbs until less than 8 limbs remain */
+Loop1b:	subxcc	%g2,%g4,%o4
+	ld	[s2_ptr+8],%g4
+	subxcc	%g3,%g1,%o5
+	ld	[s2_ptr+12],%g1
+	ldd	[s1_ptr+8],%g2
+	std	%o4,[res_ptr+0]
+	subxcc	%g2,%g4,%o4
+	ld	[s2_ptr+16],%g4
+	subxcc	%g3,%g1,%o5
+	ld	[s2_ptr+20],%g1
+	ldd	[s1_ptr+16],%g2
+	std	%o4,[res_ptr+8]
+	subxcc	%g2,%g4,%o4
+	ld	[s2_ptr+24],%g4
+	subxcc	%g3,%g1,%o5
+	ld	[s2_ptr+28],%g1
+	ldd	[s1_ptr+24],%g2
+	std	%o4,[res_ptr+16]
+	subxcc	%g2,%g4,%o4
+	ld	[s2_ptr+32],%g4
+	subxcc	%g3,%g1,%o5
+	ld	[s2_ptr+36],%g1
+	ldd	[s1_ptr+32],%g2
+	std	%o4,[res_ptr+24]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-8,size
+	add	s1_ptr,32,s1_ptr
+	add	s2_ptr,32,s2_ptr
+	add	res_ptr,32,res_ptr
+	bge	Loop1b
+	subcc	%g0,%o4,%g0		! restore cy
+
+Lfin1b:	addcc	size,8-2,size
+	blt	Lend1b
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 2 limbs until less than 2 limbs remain */
+Loope1b:subxcc	%g2,%g4,%o4
+	ld	[s2_ptr+8],%g4
+	subxcc	%g3,%g1,%o5
+	ld	[s2_ptr+12],%g1
+	ldd	[s1_ptr+8],%g2
+	std	%o4,[res_ptr+0]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-2,size
+	add	s1_ptr,8,s1_ptr
+	add	s2_ptr,8,s2_ptr
+	add	res_ptr,8,res_ptr
+	bge	Loope1b
+	subcc	%g0,%o4,%g0		! restore cy
+Lend1b:	subxcc	%g2,%g4,%o4
+	subxcc	%g3,%g1,%o5
+	std	%o4,[res_ptr+0]
+	addx	%g0,%g0,%o4		! save cy in register
+
+	andcc	size,1,%g0
+	be	Lret1b
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add last limb */
+	ld	[s2_ptr+8],%g4
+	ld	[s1_ptr+8],%g2
+	subxcc	%g2,%g4,%o4
+	st	%o4,[res_ptr+8]
+
+Lret1b:	retl
+	addx	%g0,%g0,%o0	! return carry-out from most sign. limb
+
+! **  V2  **
+/* If we come here, the alignment of s1_ptr and res_ptr as well as the
+   alignment of s2_ptr and res_ptr differ.  Since there are only two ways
+   things can be aligned (that we care about) we now know that the alignment
+   of s1_ptr and s2_ptr are the same.  */
+
+L2:	cmp	size,1
+	be	Ljone
+	nop
+	andcc	s1_ptr,4,%g0		! s1_ptr unaligned? Side effect: cy=0
+	beq	L_v2			! if no, branch
+	nop
+/* Add least significant limb separately to align s1_ptr and s2_ptr */
+	ld	[s1_ptr],%g4
+	add	s1_ptr,4,s1_ptr
+	ld	[s2_ptr],%g2
+	add	s2_ptr,4,s2_ptr
+	add	size,-1,size
+	subcc	%g4,%g2,%o4
+	st	%o4,[res_ptr]
+	add	res_ptr,4,res_ptr
+
+L_v2:	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-8,size
+	blt	Lfin2
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add blocks of 8 limbs until less than 8 limbs remain */
+Loop2:	ldd	[s1_ptr+0],%g2
+	ldd	[s2_ptr+0],%o4
+	subxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+0]
+	subxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+4]
+	ldd	[s1_ptr+8],%g2
+	ldd	[s2_ptr+8],%o4
+	subxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+8]
+	subxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+12]
+	ldd	[s1_ptr+16],%g2
+	ldd	[s2_ptr+16],%o4
+	subxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+16]
+	subxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+20]
+	ldd	[s1_ptr+24],%g2
+	ldd	[s2_ptr+24],%o4
+	subxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+24]
+	subxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+28]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-8,size
+	add	s1_ptr,32,s1_ptr
+	add	s2_ptr,32,s2_ptr
+	add	res_ptr,32,res_ptr
+	bge	Loop2
+	subcc	%g0,%o4,%g0		! restore cy
+
+Lfin2:	addcc	size,8-2,size
+	blt	Lend2
+	subcc	%g0,%o4,%g0		! restore cy
+Loope2:	ldd	[s1_ptr+0],%g2
+	ldd	[s2_ptr+0],%o4
+	subxcc	%g2,%o4,%g2
+	st	%g2,[res_ptr+0]
+	subxcc	%g3,%o5,%g3
+	st	%g3,[res_ptr+4]
+	addx	%g0,%g0,%o4		! save cy in register
+	addcc	size,-2,size
+	add	s1_ptr,8,s1_ptr
+	add	s2_ptr,8,s2_ptr
+	add	res_ptr,8,res_ptr
+	bge	Loope2
+	subcc	%g0,%o4,%g0		! restore cy
+Lend2:	andcc	size,1,%g0
+	be	Lret2
+	subcc	%g0,%o4,%g0		! restore cy
+/* Add last limb */
+Ljone:	ld	[s1_ptr],%g4
+	ld	[s2_ptr],%g2
+	subxcc	%g4,%g2,%o4
+	st	%o4,[res_ptr]
+
+Lret2:	retl
+	addx	%g0,%g0,%o0	! return carry-out from most sign. limb