aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/alpha
diff options
context:
space:
mode:
authorRoland McGrath <roland@gnu.org>1995-10-16 01:37:51 +0000
committerRoland McGrath <roland@gnu.org>1995-10-16 01:37:51 +0000
commit8f5ca04bc7fd53741d80117df992995ace8f6d2d (patch)
treee39c13fc198b22ec55647259a8080051988e8c69 /sysdeps/alpha
parent5d82cf5c55f56ae10d3b0a205d1fcc7de1cf56a0 (diff)
downloadglibc-8f5ca04bc7fd53741d80117df992995ace8f6d2d.zip
glibc-8f5ca04bc7fd53741d80117df992995ace8f6d2d.tar.gz
glibc-8f5ca04bc7fd53741d80117df992995ace8f6d2d.tar.bz2
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.
Diffstat (limited to 'sysdeps/alpha')
-rw-r--r--sysdeps/alpha/add_n.s119
-rw-r--r--sysdeps/alpha/addmul_1.s100
-rw-r--r--sysdeps/alpha/alphaev5/add_n.s118
-rw-r--r--sysdeps/alpha/alphaev5/lshift.s175
-rw-r--r--sysdeps/alpha/alphaev5/rshift.s173
-rw-r--r--sysdeps/alpha/lshift.s108
-rw-r--r--sysdeps/alpha/mul_1.s84
-rw-r--r--sysdeps/alpha/rshift.s106
-rw-r--r--sysdeps/alpha/sub_n.s119
-rw-r--r--sysdeps/alpha/submul_1.s100
-rw-r--r--sysdeps/alpha/udiv_qrnnd.S2
11 files changed, 1203 insertions, 1 deletions
diff --git a/sysdeps/alpha/add_n.s b/sysdeps/alpha/add_n.s
new file mode 100644
index 0000000..e1ad460
--- /dev/null
+++ b/sysdeps/alpha/add_n.s
@@ -0,0 +1,119 @@
+ # Alpha __mpn_add_n -- Add two limb vectors of the same length > 0 and
+ # store sum in a third limb vector.
+
+ # 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 $16
+ # s1_ptr $17
+ # s2_ptr $18
+ # size $19
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_add_n
+ .ent __mpn_add_n
+__mpn_add_n:
+ .frame $30,0,$26,0
+
+ ldq $3,0($17)
+ ldq $4,0($18)
+
+ subq $19,1,$19
+ and $19,4-1,$2 # number of limbs in first loop
+ bis $31,$31,$0
+ beq $2,.L0 # if multiple of 4 limbs, skip first loop
+
+ subq $19,$2,$19
+
+.Loop0: subq $2,1,$2
+ ldq $5,8($17)
+ addq $4,$0,$4
+ ldq $6,8($18)
+ cmpult $4,$0,$1
+ addq $3,$4,$4
+ cmpult $4,$3,$0
+ stq $4,0($16)
+ or $0,$1,$0
+
+ addq $17,8,$17
+ addq $18,8,$18
+ bis $5,$5,$3
+ bis $6,$6,$4
+ addq $16,8,$16
+ bne $2,.Loop0
+
+.L0: beq $19,.Lend
+
+ .align 3
+.Loop: subq $19,4,$19
+
+ ldq $5,8($17)
+ addq $4,$0,$4
+ ldq $6,8($18)
+ cmpult $4,$0,$1
+ addq $3,$4,$4
+ cmpult $4,$3,$0
+ stq $4,0($16)
+ or $0,$1,$0
+
+ ldq $3,16($17)
+ addq $6,$0,$6
+ ldq $4,16($18)
+ cmpult $6,$0,$1
+ addq $5,$6,$6
+ cmpult $6,$5,$0
+ stq $6,8($16)
+ or $0,$1,$0
+
+ ldq $5,24($17)
+ addq $4,$0,$4
+ ldq $6,24($18)
+ cmpult $4,$0,$1
+ addq $3,$4,$4
+ cmpult $4,$3,$0
+ stq $4,16($16)
+ or $0,$1,$0
+
+ ldq $3,32($17)
+ addq $6,$0,$6
+ ldq $4,32($18)
+ cmpult $6,$0,$1
+ addq $5,$6,$6
+ cmpult $6,$5,$0
+ stq $6,24($16)
+ or $0,$1,$0
+
+ addq $17,32,$17
+ addq $18,32,$18
+ addq $16,32,$16
+ bne $19,.Loop
+
+.Lend: addq $4,$0,$4
+ cmpult $4,$0,$1
+ addq $3,$4,$4
+ cmpult $4,$3,$0
+ stq $4,0($16)
+ or $0,$1,$0
+ ret $31,($26),1
+
+ .end __mpn_add_n
diff --git a/sysdeps/alpha/addmul_1.s b/sysdeps/alpha/addmul_1.s
new file mode 100644
index 0000000..46d277d
--- /dev/null
+++ b/sysdeps/alpha/addmul_1.s
@@ -0,0 +1,100 @@
+ # Alpha 21064 __mpn_addmul_1 -- Multiply a limb vector with a limb and add
+ # the result to a second limb vector.
+
+ # Copyright (C) 1992, 1994, 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 r16
+ # s1_ptr r17
+ # size r18
+ # s2_limb r19
+
+ # This code runs at 42 cycles/limb on the 21064.
+
+ # To improve performance for long multiplications, we would use
+ # 'fetch' for S1 and 'fetch_m' for RES. It's not obvious how to use
+ # these instructions without slowing down the general code: 1. We can
+ # only have two prefetches in operation at any time in the Alpha
+ # architecture. 2. There will seldom be any special alignment
+ # between RES_PTR and S1_PTR. Maybe we can simply divide the current
+ # loop into an inner and outer loop, having the inner loop handle
+ # exactly one prefetch block?
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_addmul_1
+ .ent __mpn_addmul_1 2
+__mpn_addmul_1:
+ .frame $30,0,$26
+
+ ldq $2,0($17) # $2 = s1_limb
+ addq $17,8,$17 # s1_ptr++
+ subq $18,1,$18 # size--
+ mulq $2,$19,$3 # $3 = prod_low
+ ldq $5,0($16) # $5 = *res_ptr
+ umulh $2,$19,$0 # $0 = prod_high
+ beq $18,Lend1 # jump if size was == 1
+ ldq $2,0($17) # $2 = s1_limb
+ addq $17,8,$17 # s1_ptr++
+ subq $18,1,$18 # size--
+ addq $5,$3,$3
+ cmpult $3,$5,$4
+ stq $3,0($16)
+ addq $16,8,$16 # res_ptr++
+ beq $18,Lend2 # jump if size was == 2
+
+ .align 3
+Loop: mulq $2,$19,$3 # $3 = prod_low
+ ldq $5,0($16) # $5 = *res_ptr
+ addq $4,$0,$0 # cy_limb = cy_limb + 'cy'
+ subq $18,1,$18 # size--
+ umulh $2,$19,$4 # $4 = cy_limb
+ ldq $2,0($17) # $2 = s1_limb
+ addq $17,8,$17 # s1_ptr++
+ addq $3,$0,$3 # $3 = cy_limb + prod_low
+ cmpult $3,$0,$0 # $0 = carry from (cy_limb + prod_low)
+ addq $5,$3,$3
+ cmpult $3,$5,$5
+ stq $3,0($16)
+ addq $16,8,$16 # res_ptr++
+ addq $5,$0,$0 # combine carries
+ bne $18,Loop
+
+Lend2: mulq $2,$19,$3 # $3 = prod_low
+ ldq $5,0($16) # $5 = *res_ptr
+ addq $4,$0,$0 # cy_limb = cy_limb + 'cy'
+ umulh $2,$19,$4 # $4 = cy_limb
+ addq $3,$0,$3 # $3 = cy_limb + prod_low
+ cmpult $3,$0,$0 # $0 = carry from (cy_limb + prod_low)
+ addq $5,$3,$3
+ cmpult $3,$5,$5
+ stq $3,0($16)
+ addq $5,$0,$0 # combine carries
+ addq $4,$0,$0 # cy_limb = prod_high + cy
+ ret $31,($26),1
+Lend1: addq $5,$3,$3
+ cmpult $3,$5,$5
+ stq $3,0($16)
+ addq $0,$5,$0
+ ret $31,($26),1
+
+ .end __mpn_addmul_1
diff --git a/sysdeps/alpha/alphaev5/add_n.s b/sysdeps/alpha/alphaev5/add_n.s
new file mode 100644
index 0000000..2aaf041
--- /dev/null
+++ b/sysdeps/alpha/alphaev5/add_n.s
@@ -0,0 +1,118 @@
+ # Alpha __mpn_add_n -- Add two limb vectors of the same length > 0 and
+ # store sum in a third limb vector.
+
+ # 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 $16
+ # s1_ptr $17
+ # s2_ptr $18
+ # size $19
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_add_n
+ .ent __mpn_add_n
+__mpn_add_n:
+ .frame $30,0,$26,0
+
+ ldq $3,0($17)
+ ldq $4,0($18)
+
+ subq $19,1,$19
+ and $19,4-1,$2 # number of limbs in first loop
+ bis $31,$31,$0
+ beq $2,.L0 # if multiple of 4 limbs, skip first loop
+
+ subq $19,$2,$19
+
+.Loop0: subq $2,1,$2
+ ldq $5,8($17)
+ addq $4,$0,$4
+ ldq $6,8($18)
+ cmpult $4,$0,$1
+ addq $3,$4,$4
+ cmpult $4,$3,$0
+ stq $4,0($16)
+ or $0,$1,$0
+
+ addq $17,8,$17
+ addq $18,8,$18
+ bis $5,$5,$3
+ bis $6,$6,$4
+ addq $16,8,$16
+ bne $2,.Loop0
+
+.L0: beq $19,.Lend
+
+ .align 4
+.Loop: subq $19,4,$19
+ unop
+
+ ldq $6,8($18)
+ addq $4,$0,$0
+ ldq $5,8($17)
+ cmpult $0,$4,$1
+ ldq $4,16($18)
+ addq $3,$0,$20
+ cmpult $20,$3,$0
+ ldq $3,16($17)
+ or $0,$1,$0
+ addq $6,$0,$0
+ cmpult $0,$6,$1
+ ldq $6,24($18)
+ addq $5,$0,$21
+ cmpult $21,$5,$0
+ ldq $5,24($17)
+ or $0,$1,$0
+ addq $4,$0,$0
+ cmpult $0,$4,$1
+ ldq $4,32($18)
+ addq $3,$0,$22
+ cmpult $22,$3,$0
+ ldq $3,32($17)
+ or $0,$1,$0
+ addq $6,$0,$0
+ cmpult $0,$6,$1
+ addq $5,$0,$23
+ cmpult $23,$5,$0
+ or $0,$1,$0
+
+ stq $20,0($16)
+ stq $21,8($16)
+ stq $22,16($16)
+ stq $23,24($16)
+
+ addq $17,32,$17
+ addq $18,32,$18
+ addq $16,32,$16
+ bne $19,.Loop
+
+.Lend: addq $4,$0,$4
+ cmpult $4,$0,$1
+ addq $3,$4,$4
+ cmpult $4,$3,$0
+ stq $4,0($16)
+ or $0,$1,$0
+ ret $31,($26),1
+
+ .end __mpn_add_n
diff --git a/sysdeps/alpha/alphaev5/lshift.s b/sysdeps/alpha/alphaev5/lshift.s
new file mode 100644
index 0000000..fdb0895
--- /dev/null
+++ b/sysdeps/alpha/alphaev5/lshift.s
@@ -0,0 +1,175 @@
+ # Alpha EV5 __mpn_lshift --
+
+ # Copyright (C) 1994, 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 r16
+ # s1_ptr r17
+ # size r18
+ # cnt r19
+
+ # This code runs at 4.25 cycles/limb on the EV5.
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_lshift
+ .ent __mpn_lshift
+__mpn_lshift:
+ .frame $30,0,$26,0
+
+ s8addq $18,$17,$17 # make r17 point at end of s1
+ ldq $4,-8($17) # load first limb
+ subq $31,$19,$20
+ s8addq $18,$16,$16 # make r16 point at end of RES
+ subq $18,1,$18
+ and $18,4-1,$28 # number of limbs in first loop
+ srl $4,$20,$0 # compute function result
+
+ beq $28,L0
+ subq $18,$28,$18
+
+ .align 3
+Loop0: ldq $3,-16($17)
+ subq $16,8,$16
+ sll $4,$19,$5
+ subq $17,8,$17
+ subq $28,1,$28
+ srl $3,$20,$6
+ or $3,$3,$4
+ or $5,$6,$8
+ stq $8,0($16)
+ bne $28,Loop0
+
+L0: sll $4,$19,$24
+ beq $18,Lend
+ # warm up phase 1
+ ldq $1,-16($17)
+ subq $18,4,$18
+ ldq $2,-24($17)
+ ldq $3,-32($17)
+ ldq $4,-40($17)
+ beq $18,Lcool1
+ # warm up phase 2
+ srl $1,$20,$7
+ sll $1,$19,$21
+ srl $2,$20,$8
+ ldq $1,-48($17)
+ sll $2,$19,$22
+ ldq $2,-56($17)
+ srl $3,$20,$5
+ or $7,$24,$7
+ sll $3,$19,$23
+ or $8,$21,$8
+ srl $4,$20,$6
+ ldq $3,-64($17)
+ sll $4,$19,$24
+ ldq $4,-72($17)
+ subq $18,4,$18
+ beq $18,Lcool1
+ .align 4
+ # main loop
+Loop: stq $7,-8($16)
+ or $5,$22,$5
+ stq $8,-16($16)
+ or $6,$23,$6
+
+ srl $1,$20,$7
+ subq $18,4,$18
+ sll $1,$19,$21
+ unop # ldq $31,-96($17)
+
+ srl $2,$20,$8
+ ldq $1,-80($17)
+ sll $2,$19,$22
+ ldq $2,-88($17)
+
+ stq $5,-24($16)
+ or $7,$24,$7
+ stq $6,-32($16)
+ or $8,$21,$8
+
+ srl $3,$20,$5
+ unop # ldq $31,-96($17)
+ sll $3,$19,$23
+ subq $16,32,$16
+
+ srl $4,$20,$6
+ ldq $3,-96($17
+ sll $4,$19,$24
+ ldq $4,-104($17)
+
+ subq $17,32,$17
+ bne $18,Loop
+ unop
+ unop
+ # cool down phase 2/1
+Lcool1: stq $7,-8($16)
+ or $5,$22,$5
+ stq $8,-16($16)
+ or $6,$23,$6
+ srl $1,$20,$7
+ sll $1,$19,$21
+ srl $2,$20,$8
+ sll $2,$19,$22
+ stq $5,-24($16)
+ or $7,$24,$7
+ stq $6,-32($16)
+ or $8,$21,$8
+ srl $3,$20,$5
+ sll $3,$19,$23
+ srl $4,$20,$6
+ sll $4,$19,$24
+ # cool down phase 2/2
+ stq $7,-40($16)
+ or $5,$22,$5
+ stq $8,-48($16)
+ or $6,$23,$6
+ stq $5,-56($16)
+ stq $6,-64($16)
+ # cool down phase 2/3
+ stq $24,-72($16)
+ ret $31,($26),1
+
+ # cool down phase 1/1
+Lcool1: srl $1,$20,$7
+ sll $1,$19,$21
+ srl $2,$20,$8
+ sll $2,$19,$22
+ srl $3,$20,$5
+ or $7,$24,$7
+ sll $3,$19,$23
+ or $8,$21,$8
+ srl $4,$20,$6
+ sll $4,$19,$24
+ # cool down phase 1/2
+ stq $7,-8($16)
+ or $5,$22,$5
+ stq $8,-16($16)
+ or $6,$23,$6
+ stq $5,-24($16)
+ stq $6,-32($16)
+ stq $24,-40($16)
+ ret $31,($26),1
+
+Lend stq $24,-8($16)
+ ret $31,($26),1
+ .end __mpn_lshift
diff --git a/sysdeps/alpha/alphaev5/rshift.s b/sysdeps/alpha/alphaev5/rshift.s
new file mode 100644
index 0000000..1da9960
--- /dev/null
+++ b/sysdeps/alpha/alphaev5/rshift.s
@@ -0,0 +1,173 @@
+ # Alpha EV5 __mpn_rshift --
+
+ # Copyright (C) 1994, 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 r16
+ # s1_ptr r17
+ # size r18
+ # cnt r19
+
+ # This code runs at 4.25 cycles/limb on the EV5.
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_rshift
+ .ent __mpn_rshift
+__mpn_rshift:
+ .frame $30,0,$26,0
+
+ ldq $4,0($17) # load first limb
+ subq $31,$19,$20
+ subq $18,1,$18
+ and $18,4-1,$28 # number of limbs in first loop
+ sll $4,$20,$0 # compute function result
+
+ beq $28,L0
+ subq $18,$28,$18
+
+ .align 3
+Loop0: ldq $3,8($17)
+ addq $16,8,$16
+ srl $4,$19,$5
+ addq $17,8,$17
+ subq $28,1,$28
+ sll $3,$20,$6
+ or $3,$3,$4
+ or $5,$6,$8
+ stq $8,-8($16)
+ bne $28,Loop0
+
+L0: srl $4,$19,$24
+ beq $18,Lend
+ # warm up phase 1
+ ldq $1,8($17)
+ subq $18,4,$18
+ ldq $2,16($17)
+ ldq $3,24($17)
+ ldq $4,32($17)
+ beq $18,Lcool1
+ # warm up phase 2
+ sll $1,$20,$7
+ srl $1,$19,$21
+ sll $2,$20,$8
+ ldq $1,40($17)
+ srl $2,$19,$22
+ ldq $2,48($17)
+ sll $3,$20,$5
+ or $7,$24,$7
+ srl $3,$19,$23
+ or $8,$21,$8
+ sll $4,$20,$6
+ ldq $3,56($17)
+ srl $4,$19,$24
+ ldq $4,64($17)
+ subq $18,4,$18
+ beq $18,Lcool2
+ .align 4
+ # main loop
+Loop: stq $7,0($16)
+ or $5,$22,$5
+ stq $8,8($16)
+ or $6,$23,$6
+
+ sll $1,$20,$7
+ subq $18,4,$18
+ srl $1,$19,$21
+ unop # ldq $31,-96($17)
+
+ sll $2,$20,$8
+ ldq $1,72($17)
+ srl $2,$19,$22
+ ldq $2,80($17)
+
+ stq $5,16($16)
+ or $7,$24,$7
+ stq $6,24($16)
+ or $8,$21,$8
+
+ sll $3,$20,$5
+ unop # ldq $31,-96($17)
+ srl $3,$19,$23
+ addq $16,32,$16
+
+ sll $4,$20,$6
+ ldq $3,88($17)
+ srl $4,$19,$24
+ ldq $4,96($17)
+
+ addq $17,32,$17
+ bne $18,Loop
+ unop
+ unop
+ # cool down phase 2/1
+Lcool2: stq $7,0($16)
+ or $5,$22,$5
+ stq $8,8($16)
+ or $6,$23,$6
+ sll $1,$20,$7
+ srl $1,$19,$21
+ sll $2,$20,$8
+ srl $2,$19,$22
+ stq $5,16($16)
+ or $7,$24,$7
+ stq $6,24($16)
+ or $8,$21,$8
+ sll $3,$20,$5
+ srl $3,$19,$23
+ sll $4,$20,$6
+ srl $4,$19,$24
+ # cool down phase 2/2
+ stq $7,32($16)
+ or $5,$22,$5
+ stq $8,40($16)
+ or $6,$23,$6
+ stq $5,48($16)
+ stq $6,56($16)
+ # cool down phase 2/3
+ stq $24,64($16)
+ ret $31,($26),1
+
+ # cool down phase 1/1
+Lcool1: sll $1,$20,$7
+ srl $1,$19,$21
+ sll $2,$20,$8
+ srl $2,$19,$22
+ sll $3,$20,$5
+ or $7,$24,$7
+ srl $3,$19,$23
+ or $8,$21,$8
+ sll $4,$20,$6
+ srl $4,$19,$24
+ # cool down phase 1/2
+ stq $7,0($16)
+ or $5,$22,$5
+ stq $8,8($16)
+ or $6,$23,$6
+ stq $5,16($16)
+ stq $6,24($16)
+ stq $24,32($16)
+ ret $31,($26),1
+
+Lend: stq $24,0($16)
+ ret $31,($26),1
+ .end __mpn_rshift
diff --git a/sysdeps/alpha/lshift.s b/sysdeps/alpha/lshift.s
new file mode 100644
index 0000000..c284349
--- /dev/null
+++ b/sysdeps/alpha/lshift.s
@@ -0,0 +1,108 @@
+ # Alpha 21064 __mpn_lshift --
+
+ # Copyright (C) 1994, 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 r16
+ # s1_ptr r17
+ # size r18
+ # cnt r19
+
+ # This code runs at 4.8 cycles/limb on the 21064. With infinite unrolling,
+ # it would take 4 cycles/limb. It should be possible to get down to 3
+ # cycles/limb since both ldq and stq can be paired with the other used
+ # instructions. But there are many restrictions in the 21064 pipeline that
+ # makes it hard, if not impossible, to get down to 3 cycles/limb:
+
+ # 1. ldq has a 3 cycle delay, srl and sll have a 2 cycle delay.
+ # 2. Only aligned instruction pairs can be paired.
+ # 3. The store buffer or silo might not be able to deal with the bandwidth.
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_lshift
+ .ent __mpn_lshift
+__mpn_lshift:
+ .frame $30,0,$26,0
+
+ s8addq $18,$17,$17 # make r17 point at end of s1
+ ldq $4,-8($17) # load first limb
+ subq $17,8,$17
+ subq $31,$19,$7
+ s8addq $18,$16,$16 # make r16 point at end of RES
+ subq $18,1,$18
+ and $18,4-1,$20 # number of limbs in first loop
+ srl $4,$7,$0 # compute function result
+
+ beq $20,L0
+ subq $18,$20,$18
+
+ .align 3
+Loop0:
+ ldq $3,-8($17)
+ subq $16,8,$16
+ subq $17,8,$17
+ subq $20,1,$20
+ sll $4,$19,$5
+ srl $3,$7,$6
+ bis $3,$3,$4
+ bis $5,$6,$8
+ stq $8,0($16)
+ bne $20,Loop0
+
+L0: beq $18,Lend
+
+ .align 3
+Loop: ldq $3,-8($17)
+ subq $16,32,$16
+ subq $18,4,$18
+ sll $4,$19,$5
+ srl $3,$7,$6
+
+ ldq $4,-16($17)
+ sll $3,$19,$1
+ bis $5,$6,$8
+ stq $8,24($16)
+ srl $4,$7,$2
+
+ ldq $3,-24($17)
+ sll $4,$19,$5
+ bis $1,$2,$8
+ stq $8,16($16)
+ srl $3,$7,$6
+
+ ldq $4,-32($17)
+ sll $3,$19,$1
+ bis $5,$6,$8
+ stq $8,8($16)
+ srl $4,$7,$2
+
+ subq $17,32,$17
+ bis $1,$2,$8
+ stq $8,0($16)
+
+ bgt $18,Loop
+
+Lend: sll $4,$19,$8
+ stq $8,-8($16)
+ ret $31,($26),1
+ .end __mpn_lshift
diff --git a/sysdeps/alpha/mul_1.s b/sysdeps/alpha/mul_1.s
new file mode 100644
index 0000000..3ef194d
--- /dev/null
+++ b/sysdeps/alpha/mul_1.s
@@ -0,0 +1,84 @@
+ # Alpha 21064 __mpn_mul_1 -- Multiply a limb vector with a limb and store
+ # the result 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 r16
+ # s1_ptr r17
+ # size r18
+ # s2_limb r19
+
+ # This code runs at 42 cycles/limb on the EV4 and 18 cycles/limb on the EV5.
+
+ # To improve performance for long multiplications, we would use
+ # 'fetch' for S1 and 'fetch_m' for RES. It's not obvious how to use
+ # these instructions without slowing down the general code: 1. We can
+ # only have two prefetches in operation at any time in the Alpha
+ # architecture. 2. There will seldom be any special alignment
+ # between RES_PTR and S1_PTR. Maybe we can simply divide the current
+ # loop into an inner and outer loop, having the inner loop handle
+ # exactly one prefetch block?
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_mul_1
+ .ent __mpn_mul_1 2
+__mpn_mul_1:
+ .frame $30,0,$26
+
+ ldq $2,0($17) # $2 = s1_limb
+ subq $18,1,$18 # size--
+ mulq $2,$19,$3 # $3 = prod_low
+ bic $31,$31,$4 # clear cy_limb
+ umulh $2,$19,$0 # $0 = prod_high
+ beq $18,Lend1 # jump if size was == 1
+ ldq $2,8($17) # $2 = s1_limb
+ subq $18,1,$18 # size--
+ stq $3,0($16)
+ beq $18,Lend2 # jump if size was == 2
+
+ .align 3
+Loop: mulq $2,$19,$3 # $3 = prod_low
+ addq $4,$0,$0 # cy_limb = cy_limb + 'cy'
+ subq $18,1,$18 # size--
+ umulh $2,$19,$4 # $4 = cy_limb
+ ldq $2,16($17) # $2 = s1_limb
+ addq $17,8,$17 # s1_ptr++
+ addq $3,$0,$3 # $3 = cy_limb + prod_low
+ stq $3,8($16)
+ cmpult $3,$0,$0 # $0 = carry from (cy_limb + prod_low)
+ addq $16,8,$16 # res_ptr++
+ bne $18,Loop
+
+Lend2: mulq $2,$19,$3 # $3 = prod_low
+ addq $4,$0,$0 # cy_limb = cy_limb + 'cy'
+ umulh $2,$19,$4 # $4 = cy_limb
+ addq $3,$0,$3 # $3 = cy_limb + prod_low
+ cmpult $3,$0,$0 # $0 = carry from (cy_limb + prod_low)
+ stq $3,8($16)
+ addq $4,$0,$0 # cy_limb = prod_high + cy
+ ret $31,($26),1
+Lend1: stq $3,0($16)
+ ret $31,($26),1
+
+ .end __mpn_mul_1
diff --git a/sysdeps/alpha/rshift.s b/sysdeps/alpha/rshift.s
new file mode 100644
index 0000000..74eab04
--- /dev/null
+++ b/sysdeps/alpha/rshift.s
@@ -0,0 +1,106 @@
+ # Alpha 21064 __mpn_rshift --
+
+ # Copyright (C) 1994, 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 r16
+ # s1_ptr r17
+ # size r18
+ # cnt r19
+
+ # This code runs at 4.8 cycles/limb on the 21064. With infinite unrolling,
+ # it would take 4 cycles/limb. It should be possible to get down to 3
+ # cycles/limb since both ldq and stq can be paired with the other used
+ # instructions. But there are many restrictions in the 21064 pipeline that
+ # makes it hard, if not impossible, to get down to 3 cycles/limb:
+
+ # 1. ldq has a 3 cycle delay, srl and sll have a 2 cycle delay.
+ # 2. Only aligned instruction pairs can be paired.
+ # 3. The store buffer or silo might not be able to deal with the bandwidth.
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_rshift
+ .ent __mpn_rshift
+__mpn_rshift:
+ .frame $30,0,$26,0
+
+ ldq $4,0($17) # load first limb
+ addq $17,8,$17
+ subq $31,$19,$7
+ subq $18,1,$18
+ and $18,4-1,$20 # number of limbs in first loop
+ sll $4,$7,$0 # compute function result
+
+ beq $20,L0
+ subq $18,$20,$18
+
+ .align 3
+Loop0:
+ ldq $3,0($17)
+ addq $16,8,$16
+ addq $17,8,$17
+ subq $20,1,$20
+ srl $4,$19,$5
+ sll $3,$7,$6
+ bis $3,$3,$4
+ bis $5,$6,$8
+ stq $8,-8($16)
+ bne $20,Loop0
+
+L0: beq $18,Lend
+
+ .align 3
+Loop: ldq $3,0($17)
+ addq $16,32,$16
+ subq $18,4,$18
+ srl $4,$19,$5
+ sll $3,$7,$6
+
+ ldq $4,8($17)
+ srl $3,$19,$1
+ bis $5,$6,$8
+ stq $8,-32($16)
+ sll $4,$7,$2
+
+ ldq $3,16($17)
+ srl $4,$19,$5
+ bis $1,$2,$8
+ stq $8,-24($16)
+ sll $3,$7,$6
+
+ ldq $4,24($17)
+ srl $3,$19,$1
+ bis $5,$6,$8
+ stq $8,-16($16)
+ sll $4,$7,$2
+
+ addq $17,32,$17
+ bis $1,$2,$8
+ stq $8,-8($16)
+
+ bgt $18,Loop
+
+Lend: srl $4,$19,$8
+ stq $8,0($16)
+ ret $31,($26),1
+ .end __mpn_rshift
diff --git a/sysdeps/alpha/sub_n.s b/sysdeps/alpha/sub_n.s
new file mode 100644
index 0000000..5200025
--- /dev/null
+++ b/sysdeps/alpha/sub_n.s
@@ -0,0 +1,119 @@
+ # Alpha __mpn_sub_n -- Subtract two limb vectors of the same length > 0 and
+ # store difference in a third limb vector.
+
+ # 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 $16
+ # s1_ptr $17
+ # s2_ptr $18
+ # size $19
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_sub_n
+ .ent __mpn_sub_n
+__mpn_sub_n:
+ .frame $30,0,$26,0
+
+ ldq $3,0($17)
+ ldq $4,0($18)
+
+ subq $19,1,$19
+ and $19,4-1,$2 # number of limbs in first loop
+ bis $31,$31,$0
+ beq $2,.L0 # if multiple of 4 limbs, skip first loop
+
+ subq $19,$2,$19
+
+.Loop0: subq $2,1,$2
+ ldq $5,8($17)
+ addq $4,$0,$4
+ ldq $6,8($18)
+ cmpult $4,$0,$1
+ subq $3,$4,$4
+ cmpult $3,$4,$0
+ stq $4,0($16)
+ or $0,$1,$0
+
+ addq $17,8,$17
+ addq $18,8,$18
+ bis $5,$5,$3
+ bis $6,$6,$4
+ addq $16,8,$16
+ bne $2,.Loop0
+
+.L0: beq $19,.Lend
+
+ .align 3
+.Loop: subq $19,4,$19
+
+ ldq $5,8($17)
+ addq $4,$0,$4
+ ldq $6,8($18)
+ cmpult $4,$0,$1
+ subq $3,$4,$4
+ cmpult $3,$4,$0
+ stq $4,0($16)
+ or $0,$1,$0
+
+ ldq $3,16($17)
+ addq $6,$0,$6
+ ldq $4,16($18)
+ cmpult $6,$0,$1
+ subq $5,$6,$6
+ cmpult $5,$6,$0
+ stq $6,8($16)
+ or $0,$1,$0
+
+ ldq $5,24($17)
+ addq $4,$0,$4
+ ldq $6,24($18)
+ cmpult $4,$0,$1
+ subq $3,$4,$4
+ cmpult $3,$4,$0
+ stq $4,16($16)
+ or $0,$1,$0
+
+ ldq $3,32($17)
+ addq $6,$0,$6
+ ldq $4,32($18)
+ cmpult $6,$0,$1
+ subq $5,$6,$6
+ cmpult $5,$6,$0
+ stq $6,24($16)
+ or $0,$1,$0
+
+ addq $17,32,$17
+ addq $18,32,$18
+ addq $16,32,$16
+ bne $19,.Loop
+
+.Lend: addq $4,$0,$4
+ cmpult $4,$0,$1
+ subq $3,$4,$4
+ cmpult $3,$4,$0
+ stq $4,0($16)
+ or $0,$1,$0
+ ret $31,($26),1
+
+ .end __mpn_sub_n
diff --git a/sysdeps/alpha/submul_1.s b/sysdeps/alpha/submul_1.s
new file mode 100644
index 0000000..acaa11c
--- /dev/null
+++ b/sysdeps/alpha/submul_1.s
@@ -0,0 +1,100 @@
+ # Alpha 21064 __mpn_submul_1 -- Multiply a limb vector with a limb and
+ # subtract the result from a second limb vector.
+
+ # Copyright (C) 1992, 1994, 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 r16
+ # s1_ptr r17
+ # size r18
+ # s2_limb r19
+
+ # This code runs at 42 cycles/limb on the 21064.
+
+ # To improve performance for long multiplications, we would use
+ # 'fetch' for S1 and 'fetch_m' for RES. It's not obvious how to use
+ # these instructions without slowing down the general code: 1. We can
+ # only have two prefetches in operation at any time in the Alpha
+ # architecture. 2. There will seldom be any special alignment
+ # between RES_PTR and S1_PTR. Maybe we can simply divide the current
+ # loop into an inner and outer loop, having the inner loop handle
+ # exactly one prefetch block?
+
+ .set noreorder
+ .set noat
+.text
+ .align 3
+ .globl __mpn_submul_1
+ .ent __mpn_submul_1 2
+__mpn_submul_1:
+ .frame $30,0,$26
+
+ ldq $2,0($17) # $2 = s1_limb
+ addq $17,8,$17 # s1_ptr++
+ subq $18,1,$18 # size--
+ mulq $2,$19,$3 # $3 = prod_low
+ ldq $5,0($16) # $5 = *res_ptr
+ umulh $2,$19,$0 # $0 = prod_high
+ beq $18,Lend1 # jump if size was == 1
+ ldq $2,0($17) # $2 = s1_limb
+ addq $17,8,$17 # s1_ptr++
+ subq $18,1,$18 # size--
+ subq $5,$3,$3
+ cmpult $5,$3,$4
+ stq $3,0($16)
+ addq $16,8,$16 # res_ptr++
+ beq $18,Lend2 # jump if size was == 2
+
+ .align 3
+Loop: mulq $2,$19,$3 # $3 = prod_low
+ ldq $5,0($16) # $5 = *res_ptr
+ addq $4,$0,$0 # cy_limb = cy_limb + 'cy'
+ subq $18,1,$18 # size--
+ umulh $2,$19,$4 # $4 = cy_limb
+ ldq $2,0($17) # $2 = s1_limb
+ addq $17,8,$17 # s1_ptr++
+ addq $3,$0,$3 # $3 = cy_limb + prod_low
+ cmpult $3,$0,$0 # $0 = carry from (cy_limb + prod_low)
+ subq $5,$3,$3
+ cmpult $5,$3,$5
+ stq $3,0($16)
+ addq $16,8,$16 # res_ptr++
+ addq $5,$0,$0 # combine carries
+ bne $18,Loop
+
+Lend2: mulq $2,$19,$3 # $3 = prod_low
+ ldq $5,0($16) # $5 = *res_ptr
+ addq $4,$0,$0 # cy_limb = cy_limb + 'cy'
+ umulh $2,$19,$4 # $4 = cy_limb
+ addq $3,$0,$3 # $3 = cy_limb + prod_low
+ cmpult $3,$0,$0 # $0 = carry from (cy_limb + prod_low)
+ subq $5,$3,$3
+ cmpult $5,$3,$5
+ stq $3,0($16)
+ addq $5,$0,$0 # combine carries
+ addq $4,$0,$0 # cy_limb = prod_high + cy
+ ret $31,($26),1
+Lend1: subq $5,$3,$3
+ cmpult $5,$3,$5
+ stq $3,0($16)
+ addq $0,$5,$0
+ ret $31,($26),1
+
+ .end __mpn_submul_1
diff --git a/sysdeps/alpha/udiv_qrnnd.S b/sysdeps/alpha/udiv_qrnnd.S
index 942d7a8..bafafd6 100644
--- a/sysdeps/alpha/udiv_qrnnd.S
+++ b/sysdeps/alpha/udiv_qrnnd.S
@@ -134,7 +134,7 @@ Loop2: cmplt n0,0,tmp
ret $31,($26),1
Odd:
- /* q' in n0. r' in n1. */
+ /* q' in n0. r' in n1 */
addq n1,n0,n1
cmpult n1,n0,tmp # tmp := carry from addq
beq tmp,LLp6