--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+ # 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
--- /dev/null
+; HP-PA-1.1 __mpn_mul_1 -- Multiply a limb vector with a limb and store
+; the result in a second limb vector.
+
+; Copyright (C) 1992, 1993, 1994 Free Software Foundation, Inc.
+
+; This file is part of the GNU MP Library.
+
+; The GNU MP Library is free software; you can redistribute it and/or modify
+; it under the terms of the GNU Library General Public License as published by
+; the Free Software Foundation; either version 2 of the License, or (at your
+; option) any later version.
+
+; The GNU MP Library is distributed in the hope that it will be useful, but
+; WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+; or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Library General Public
+; License for more details.
+
+; You should have received a copy of the GNU Library General Public License
+; along with the GNU MP Library; see the file COPYING.LIB. If not, write to
+; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+
+
+; INPUT PARAMETERS
+; res_ptr r26
+; s1_ptr r25
+; size r24
+; s2_limb r23
+
+; This runs at 9 cycles/limb on a PA7000. With the used instructions, it can
+; not become faster due to data cache contention after a store. On the
+; PA7100 it runs at 7 cycles/limb, and that can not be improved either, since
+; only the xmpyu does not need the integer pipeline, so the only dual-issue
+; we will get are addc+xmpyu. Unrolling would not help either CPU.
+
+; We could use fldds to read two limbs at a time from the S1 array, and that
+; could bring down the times to 8.5 and 6.5 cycles/limb for the PA7000 and
+; PA7100, respectively. We don't do that since it does not seem worth the
+; (alignment) troubles...
+
+; At least the PA7100 is rumored to be able to deal with cache-misses
+; without stalling instruction issue. If this is true, and the cache is
+; actually also lockup-free, we should use a deeper software pipeline, and
+; load from S1 very early! (The loads and stores to -12(sp) will surely be
+; in the cache.)
+
+ .code
+ .export __mpn_mul_1
+__mpn_mul_1
+ .proc
+ .callinfo frame=64,no_calls
+ .entry
+
+ ldo 64(%r30),%r30
+ fldws,ma 4(%r25),%fr5
+ stw %r23,-16(%r30) ; move s2_limb ...
+ addib,= -1,%r24,L$just_one_limb
+ fldws -16(%r30),%fr4 ; ... into fr4
+ add %r0,%r0,%r0 ; clear carry
+ xmpyu %fr4,%fr5,%fr6
+ fldws,ma 4(%r25),%fr7
+ fstds %fr6,-16(%r30)
+ xmpyu %fr4,%fr7,%fr8
+ ldw -12(%r30),%r19 ; least significant limb in product
+ ldw -16(%r30),%r28
+
+ fstds %fr8,-16(%r30)
+ addib,= -1,%r24,L$end
+ ldw -12(%r30),%r1
+
+; Main loop
+L$loop fldws,ma 4(%r25),%fr5
+ stws,ma %r19,4(%r26)
+ addc %r28,%r1,%r19
+ xmpyu %fr4,%fr5,%fr6
+ ldw -16(%r30),%r28
+ fstds %fr6,-16(%r30)
+ addib,<> -1,%r24,L$loop
+ ldw -12(%r30),%r1
+
+L$end stws,ma %r19,4(%r26)
+ addc %r28,%r1,%r19
+ ldw -16(%r30),%r28
+ stws,ma %r19,4(%r26)
+ addc %r0,%r28,%r28
+ bv 0(%r2)
+ ldo -64(%r30),%r30
+
+L$just_one_limb
+ xmpyu %fr4,%fr5,%fr6
+ fstds %fr6,-16(%r30)
+ ldw -16(%r30),%r28
+ ldo -64(%r30),%r30
+ bv 0(%r2)
+ fstws %fr6R,0(%r26)
+
+ .exit
+ .procend
--- /dev/null
+; HP-PA-1.1 __mpn_submul_1 -- Multiply a limb vector with a limb and
+; subtract the result from a second limb vector.
+
+; Copyright (C) 1992, 1993, 1994 Free Software Foundation, Inc.
+
+; This file is part of the GNU MP Library.
+
+; The GNU MP Library is free software; you can redistribute it and/or modify
+; it under the terms of the GNU Library General Public License as published by
+; the Free Software Foundation; either version 2 of the License, or (at your
+; option) any later version.
+
+; The GNU MP Library is distributed in the hope that it will be useful, but
+; WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+; or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Library General Public
+; License for more details.
+
+; You should have received a copy of the GNU Library General Public License
+; along with the GNU MP Library; see the file COPYING.LIB. If not, write to
+; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+
+
+; INPUT PARAMETERS
+; res_ptr r26
+; s1_ptr r25
+; size r24
+; s2_limb r23
+
+; This runs at 12 cycles/limb on a PA7000. With the used instructions, it
+; can not become faster due to data cache contention after a store. On the
+; PA7100 it runs at 11 cycles/limb, and that can not be improved either,
+; since only the xmpyu does not need the integer pipeline, so the only
+; dual-issue we will get are addc+xmpyu. Unrolling could gain a cycle/limb
+; on the PA7100.
+
+; There are some ideas described in mul_1.s that applies to this code too.
+
+; It seems possible to make this run as fast as __mpn_addmul_1, if we use
+; sub,>>= %r29,%r19,%r22
+; addi 1,%r28,%r28
+; but that requires reworking the hairy software pipeline...
+
+ .code
+ .export __mpn_submul_1
+__mpn_submul_1
+ .proc
+ .callinfo frame=64,no_calls
+ .entry
+
+ ldo 64(%r30),%r30
+ fldws,ma 4(%r25),%fr5
+ stw %r23,-16(%r30) ; move s2_limb ...
+ addib,= -1,%r24,L$just_one_limb
+ fldws -16(%r30),%fr4 ; ... into fr4
+ add %r0,%r0,%r0 ; clear carry
+ xmpyu %fr4,%fr5,%fr6
+ fldws,ma 4(%r25),%fr7
+ fstds %fr6,-16(%r30)
+ xmpyu %fr4,%fr7,%fr8
+ ldw -12(%r30),%r19 ; least significant limb in product
+ ldw -16(%r30),%r28
+
+ fstds %fr8,-16(%r30)
+ addib,= -1,%r24,L$end
+ ldw -12(%r30),%r1
+
+; Main loop
+L$loop ldws 0(%r26),%r29
+ fldws,ma 4(%r25),%fr5
+ sub %r29,%r19,%r22
+ add %r22,%r19,%r0
+ stws,ma %r22,4(%r26)
+ addc %r28,%r1,%r19
+ xmpyu %fr4,%fr5,%fr6
+ ldw -16(%r30),%r28
+ fstds %fr6,-16(%r30)
+ addc %r0,%r28,%r28
+ addib,<> -1,%r24,L$loop
+ ldw -12(%r30),%r1
+
+L$end ldw 0(%r26),%r29
+ sub %r29,%r19,%r22
+ add %r22,%r19,%r0
+ stws,ma %r22,4(%r26)
+ addc %r28,%r1,%r19
+ ldw -16(%r30),%r28
+ ldws 0(%r26),%r29
+ addc %r0,%r28,%r28
+ sub %r29,%r19,%r22
+ add %r22,%r19,%r0
+ stws,ma %r22,4(%r26)
+ addc %r0,%r28,%r28
+ bv 0(%r2)
+ ldo -64(%r30),%r30
+
+L$just_one_limb
+ xmpyu %fr4,%fr5,%fr6
+ ldw 0(%r26),%r29
+ fstds %fr6,-16(%r30)
+ ldw -12(%r30),%r1
+ ldw -16(%r30),%r28
+ sub %r29,%r1,%r22
+ add %r22,%r1,%r0
+ stw %r22,0(%r26)
+ addc %r0,%r28,%r28
+ bv 0(%r2)
+ ldo -64(%r30),%r30
+
+ .exit
+ .procend
--- /dev/null
+; HP-PA __udiv_qrnnd division support, used from longlong.h.
+; This version runs fast on PA 7000 and later.
+
+; Copyright (C) 1993, 1994 Free Software Foundation, Inc.
+
+; This file is part of the GNU MP Library.
+
+; The GNU MP Library is free software; you can redistribute it and/or modify
+; it under the terms of the GNU Library General Public License as published by
+; the Free Software Foundation; either version 2 of the License, or (at your
+; option) any later version.
+
+; The GNU MP Library is distributed in the hope that it will be useful, but
+; WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+; or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Library General Public
+; License for more details.
+
+; You should have received a copy of the GNU Library General Public License
+; along with the GNU MP Library; see the file COPYING.LIB. If not, write to
+; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+
+
+; INPUT PARAMETERS
+; rem_ptr gr26
+; n1 gr25
+; n0 gr24
+; d gr23
+
+ .code
+L$0000 .word 0x43f00000
+ .word 0x0
+ .export __udiv_qrnnd
+__udiv_qrnnd
+ .proc
+ .callinfo frame=64,no_calls
+ .entry
+ ldo 64(%r30),%r30
+
+ stws %r25,-16(0,%r30) ; n_hi
+ stws %r24,-12(0,%r30) ; n_lo
+ ldil L'L$0000,%r19
+ ldo R'L$0000(%r19),%r19
+ fldds -16(0,%r30),%fr5
+ stws %r23,-12(0,%r30)
+ comib,<= 0,%r25,L$1
+ fcnvxf,dbl,dbl %fr5,%fr5
+ fldds 0(0,%r19),%fr4
+ fadd,dbl %fr4,%fr5,%fr5
+L$1
+ fcpy,sgl %fr0,%fr6L
+ fldws -12(0,%r30),%fr6R
+ fcnvxf,dbl,dbl %fr6,%fr4
+
+ fdiv,dbl %fr5,%fr4,%fr5
+
+ fcnvfx,dbl,dbl %fr5,%fr4
+ fstws %fr4R,-16(%r30)
+ xmpyu %fr4R,%fr6R,%fr6
+ ldws -16(%r30),%r28
+ fstds %fr6,-16(0,%r30)
+ ldws -12(0,%r30),%r21
+ ldws -16(0,%r30),%r20
+ sub %r24,%r21,%r22
+ subb %r25,%r20,%r19
+ comib,= 0,%r19,L$2
+ ldo -64(%r30),%r30
+
+ add %r22,%r23,%r22
+ ldo -1(%r28),%r28
+L$2 bv 0(%r2)
+ stws %r22,0(0,%r26)
+
+ .exit
+ .procend
--- /dev/null
+# IBM POWER __mpn_add_n -- Add two limb vectors of equal, non-zero length.
+
+# 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 r3
+# s1_ptr r4
+# s2_ptr r5
+# size r6
+
+ .toc
+ .extern __mpn_add_n[DS]
+ .extern .__mpn_add_n
+.csect [PR]
+ .align 2
+ .globl __mpn_add_n
+ .globl .__mpn_add_n
+ .csect __mpn_add_n[DS]
+__mpn_add_n:
+ .long .__mpn_add_n, TOC[tc0], 0
+ .csect [PR]
+.__mpn_add_n:
+ mtctr 6 # copy size into CTR
+ l 8,0(4) # load least significant s1 limb
+ l 0,0(5) # load least significant s2 limb
+ cal 3,-4(3) # offset res_ptr, it's updated before used
+ a 7,0,8 # add least significant limbs, set cy
+ 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
+ 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
+ lil 3,0 # load cy into ...
+ aze 3,3 # ... return value register
+ br
--- /dev/null
+# IBM POWER __mpn_addmul_1 -- Multiply a limb vector with a limb and add
+# the result to 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 r3
+# s1_ptr r4
+# size r5
+# s2_limb r6
+
+# The RS/6000 has no unsigned 32x32->64 bit multiplication instruction. To
+# obtain that operation, we have to use the 32x32->64 signed multiplication
+# instruction, and add the appropriate compensation to the high limb of the
+# result. We add the multiplicand if the multiplier has its most significant
+# bit set, and we add the multiplier if the multiplicand has its most
+# significant bit set. We need to preserve the carry flag between each
+# iteration, so we have to compute the compensation carefully (the natural,
+# srai+and doesn't work). Since the POWER architecture has a branch unit
+# we can branch in zero cycles, so that's how we perform the additions.
+
+ .toc
+ .csect .__mpn_addmul_1[PR]
+ .align 2
+ .globl __mpn_addmul_1
+ .globl .__mpn_addmul_1
+ .csect __mpn_addmul_1[DS]
+__mpn_addmul_1:
+ .long .__mpn_addmul_1[PR], TOC[tc0], 0
+ .csect .__mpn_addmul_1[PR]
+.__mpn_addmul_1:
+
+ cal 3,-4(3)
+ l 0,0(4)
+ cmpi 0,6,0
+ mtctr 5
+ mul 9,0,6
+ srai 7,0,31
+ and 7,7,6
+ mfmq 8
+ cax 9,9,7
+ l 7,4(3)
+ a 8,8,7 # add res_limb
+ blt Lneg
+Lpos: bdz Lend
+
+Lploop: lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 10,0,6
+ mfmq 0
+ ae 8,0,9 # low limb + old_cy_limb + old cy
+ l 7,4(3)
+ aze 10,10 # propagate cy to new cy_limb
+ a 8,8,7 # add res_limb
+ bge Lp0
+ cax 10,10,6 # adjust high limb for negative limb from s1
+Lp0: bdz Lend0
+ lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 9,0,6
+ mfmq 0
+ ae 8,0,10
+ l 7,4(3)
+ aze 9,9
+ a 8,8,7
+ bge Lp1
+ cax 9,9,6 # adjust high limb for negative limb from s1
+Lp1: bdn Lploop
+
+ b Lend
+
+Lneg: cax 9,9,0
+ bdz Lend
+Lnloop: lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 10,0,6
+ mfmq 7
+ ae 8,7,9
+ l 7,4(3)
+ ae 10,10,0 # propagate cy to new cy_limb
+ a 8,8,7 # add res_limb
+ bge Ln0
+ cax 10,10,6 # adjust high limb for negative limb from s1
+Ln0: bdz Lend0
+ lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 9,0,6
+ mfmq 7
+ ae 8,7,10
+ l 7,4(3)
+ ae 9,9,0 # propagate cy to new cy_limb
+ a 8,8,7 # add res_limb
+ bge Ln1
+ cax 9,9,6 # adjust high limb for negative limb from s1
+Ln1: bdn Lnloop
+ b Lend
+
+Lend0: cal 9,0(10)
+Lend: st 8,4(3)
+ aze 3,9
+ br
--- /dev/null
+# IBM POWER __mpn_lshift --
+
+# 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 r3
+# s_ptr r4
+# size r5
+# cnt r6
+
+ .toc
+ .extern __mpn_lshift[DS]
+ .extern .__mpn_lshift
+.csect [PR]
+ .align 2
+ .globl __mpn_lshift
+ .globl .__mpn_lshift
+ .csect __mpn_lshift[DS]
+__mpn_lshift:
+ .long .__mpn_lshift, TOC[tc0], 0
+ .csect [PR]
+.__mpn_lshift:
+ sli 0,5,2
+ cax 9,3,0
+ cax 4,4,0
+ sfi 8,6,32
+ mtctr 5 # put limb count in CTR loop register
+ lu 0,-4(4) # read most significant limb
+ sre 3,0,8 # compute carry out limb, and init MQ register
+ bdz Lend2 # if just one limb, skip loop
+ lu 0,-4(4) # read 2:nd most significant limb
+ sreq 7,0,8 # compute most significant limb of result
+ bdz Lend # if just two limb, skip loop
+Loop: lu 0,-4(4) # load next lower limb
+ stu 7,-4(9) # store previous result during read latency
+ sreq 7,0,8 # compute result limb
+ bdn Loop # loop back until CTR is zero
+Lend: stu 7,-4(9) # store 2:nd least significant limb
+Lend2: sle 7,0,6 # compute least significant limb
+ st 7,-4(9) # store it" \
+ br
--- /dev/null
+# IBM POWER __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 r3
+# s1_ptr r4
+# size r5
+# s2_limb r6
+
+# The RS/6000 has no unsigned 32x32->64 bit multiplication instruction. To
+# obtain that operation, we have to use the 32x32->64 signed multiplication
+# instruction, and add the appropriate compensation to the high limb of the
+# result. We add the multiplicand if the multiplier has its most significant
+# bit set, and we add the multiplier if the multiplicand has its most
+# significant bit set. We need to preserve the carry flag between each
+# iteration, so we have to compute the compensation carefully (the natural,
+# srai+and doesn't work). Since the POWER architecture has a branch unit
+# we can branch in zero cycles, so that's how we perform the additions.
+
+ .toc
+ .csect .__mpn_mul_1[PR]
+ .align 2
+ .globl __mpn_mul_1
+ .globl .__mpn_mul_1
+ .csect __mpn_mul_1[DS]
+__mpn_mul_1:
+ .long .__mpn_mul_1[PR], TOC[tc0], 0
+ .csect .__mpn_mul_1[PR]
+.__mpn_mul_1:
+
+ cal 3,-4(3)
+ l 0,0(4)
+ cmpi 0,6,0
+ mtctr 5
+ mul 9,0,6
+ srai 7,0,31
+ and 7,7,6
+ mfmq 8
+ ai 0,0,0 # reset carry
+ cax 9,9,7
+ blt Lneg
+Lpos: bdz Lend
+Lploop: lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 10,0,6
+ mfmq 0
+ ae 8,0,9
+ bge Lp0
+ cax 10,10,6 # adjust high limb for negative limb from s1
+Lp0: bdz Lend0
+ lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 9,0,6
+ mfmq 0
+ ae 8,0,10
+ bge Lp1
+ cax 9,9,6 # adjust high limb for negative limb from s1
+Lp1: bdn Lploop
+ b Lend
+
+Lneg: cax 9,9,0
+ bdz Lend
+Lnloop: lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 10,0,6
+ cax 10,10,0 # adjust high limb for negative s2_limb
+ mfmq 0
+ ae 8,0,9
+ bge Ln0
+ cax 10,10,6 # adjust high limb for negative limb from s1
+Ln0: bdz Lend0
+ lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 9,0,6
+ cax 9,9,0 # adjust high limb for negative s2_limb
+ mfmq 0
+ ae 8,0,10
+ bge Ln1
+ cax 9,9,6 # adjust high limb for negative limb from s1
+Ln1: bdn Lnloop
+ b Lend
+
+Lend0: cal 9,0(10)
+Lend: st 8,4(3)
+ aze 3,9
+ br
--- /dev/null
+# IBM POWER __mpn_rshift --
+
+# 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 r3
+# s_ptr r4
+# size r5
+# cnt r6
+
+ .toc
+ .extern __mpn_rshift[DS]
+ .extern .__mpn_rshift
+.csect [PR]
+ .align 2
+ .globl __mpn_rshift
+ .globl .__mpn_rshift
+ .csect __mpn_rshift[DS]
+__mpn_rshift:
+ .long .__mpn_rshift, TOC[tc0], 0
+ .csect [PR]
+.__mpn_rshift:
+ sfi 8,6,32
+ mtctr 5 # put limb count in CTR loop register
+ l 0,0(4) # read least significant limb
+ ai 9,3,-4 # adjust res_ptr since it's offset in the stu:s
+ sle 3,0,8 # compute carry limb, and init MQ register
+ bdz Lend2 # if just one limb, skip loop
+ lu 0,4(4) # read 2:nd least significant limb
+ sleq 7,0,8 # compute least significant limb of result
+ bdz Lend # if just two limb, skip loop
+Loop: lu 0,4(4) # load next higher limb
+ stu 7,4(9) # store previous result during read latency
+ sleq 7,0,8 # compute result limb
+ bdn Loop # loop back until CTR is zero
+Lend: stu 7,4(9) # store 2:nd most significant limb
+Lend2: sre 7,0,6 # compute most significant limb
+ st 7,4(9) # store it" \
+ br
--- /dev/null
+# IBM POWER __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 r3
+# s1_ptr r4
+# s2_ptr r5
+# size r6
+
+ .toc
+ .extern __mpn_sub_n[DS]
+ .extern .__mpn_sub_n
+.csect [PR]
+ .align 2
+ .globl __mpn_sub_n
+ .globl .__mpn_sub_n
+ .csect __mpn_sub_n[DS]
+__mpn_sub_n:
+ .long .__mpn_sub_n, TOC[tc0], 0
+ .csect [PR]
+.__mpn_sub_n:
+ mtctr 6 # copy size into CTR
+ l 8,0(4) # load least significant s1 limb
+ l 0,0(5) # load least significant s2 limb
+ cal 3,-4(3) # offset res_ptr, it's updated before used
+ sf 7,0,8 # add least significant limbs, set cy
+ 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
+ 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
+ sfe 3,0,0 # load !cy into ...
+ sfi 3,3,0 # ... return value register
+ br
--- /dev/null
+# IBM POWER __mpn_submul_1 -- Multiply a limb vector with a limb and subtract
+# the result from 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 r3
+# s1_ptr r4
+# size r5
+# s2_limb r6
+
+# The RS/6000 has no unsigned 32x32->64 bit multiplication instruction. To
+# obtain that operation, we have to use the 32x32->64 signed multiplication
+# instruction, and add the appropriate compensation to the high limb of the
+# result. We add the multiplicand if the multiplier has its most significant
+# bit set, and we add the multiplier if the multiplicand has its most
+# significant bit set. We need to preserve the carry flag between each
+# iteration, so we have to compute the compensation carefully (the natural,
+# srai+and doesn't work). Since the POWER architecture has a branch unit
+# we can branch in zero cycles, so that's how we perform the additions.
+
+ .toc
+ .csect .__mpn_submul_1[PR]
+ .align 2
+ .globl __mpn_submul_1
+ .globl .__mpn_submul_1
+ .csect __mpn_submul_1[DS]
+__mpn_submul_1:
+ .long .__mpn_submul_1[PR], TOC[tc0], 0
+ .csect .__mpn_submul_1[PR]
+.__mpn_submul_1:
+
+ cal 3,-4(3)
+ l 0,0(4)
+ cmpi 0,6,0
+ mtctr 5
+ mul 9,0,6
+ srai 7,0,31
+ and 7,7,6
+ mfmq 11
+ cax 9,9,7
+ l 7,4(3)
+ sf 8,11,7 # add res_limb
+ a 11,8,11 # invert cy (r11 is junk)
+ blt Lneg
+Lpos: bdz Lend
+
+Lploop: lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 10,0,6
+ mfmq 0
+ ae 11,0,9 # low limb + old_cy_limb + old cy
+ l 7,4(3)
+ aze 10,10 # propagate cy to new cy_limb
+ sf 8,11,7 # add res_limb
+ a 11,8,11 # invert cy (r11 is junk)
+ bge Lp0
+ cax 10,10,6 # adjust high limb for negative limb from s1
+Lp0: bdz Lend0
+ lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 9,0,6
+ mfmq 0
+ ae 11,0,10
+ l 7,4(3)
+ aze 9,9
+ sf 8,11,7
+ a 11,8,11 # invert cy (r11 is junk)
+ bge Lp1
+ cax 9,9,6 # adjust high limb for negative limb from s1
+Lp1: bdn Lploop
+
+ b Lend
+
+Lneg: cax 9,9,0
+ bdz Lend
+Lnloop: lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 10,0,6
+ mfmq 7
+ ae 11,7,9
+ l 7,4(3)
+ ae 10,10,0 # propagate cy to new cy_limb
+ sf 8,11,7 # add res_limb
+ a 11,8,11 # invert cy (r11 is junk)
+ bge Ln0
+ cax 10,10,6 # adjust high limb for negative limb from s1
+Ln0: bdz Lend0
+ lu 0,4(4)
+ stu 8,4(3)
+ cmpi 0,0,0
+ mul 9,0,6
+ mfmq 7
+ ae 11,7,10
+ l 7,4(3)
+ ae 9,9,0 # propagate cy to new cy_limb
+ sf 8,11,7 # add res_limb
+ a 11,8,11 # invert cy (r11 is junk)
+ bge Ln1
+ cax 9,9,6 # adjust high limb for negative limb from s1
+Ln1: bdn Lnloop
+ b Lend
+
+Lend0: cal 9,0(10)
+Lend: st 8,4(3)
+ aze 3,9
+ br
projects / kopensolaris-gnu / glibc.git / commitdiff