mpir/mpn/x86_64w/dive_1.asm

147 lines
3.9 KiB
NASM

; Copyright 1999, 2000, 2001, 2002 Free Software Foundation, Inc.
;
; Copyright 2008 Brian Gladman
;
; This file is part of the MPIR Library.
;
; The MPIR Library is free software; you can redistribute it and/or
; modify it under the terms of the GNU Lesser General Public License as
; published by the Free Software Foundation; either version 2.1 of the
; License, or (at your option) any later version.
;
; The MPIR 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
; Lesser General Public License for more details.
;
; You should have received a copy of the GNU Lesser General Public
; License along with the MPIR Library; see the file COPYING.LIB. If
; not, write to the Free Software Foundation, Inc., 51 Franklin Street,
; Fifth Floor, Boston, MA 02110-1301, USA.
;
; AMD64 mpn_divexact_1 -- mpn by limb exact division
;
; Calling interface:
;
; void mpn_divexact_1(
; mp_ptr dst, rcx
; mp_srcptr src, rdx
; mp_size_t size, r8
; mp_limb_t divisor r9
; )
;
; since the inverse takes a while to setup,plain division is used for small
; Multiplying works out faster for size>=3 when the divisor is odd or size>=4
; when the divisor is even.
;
; This is an SEH Frame Function with a leaf prologue
%include "yasm_mac.inc"
%define reg_save_list rsi, rdi
BITS 64
extern __gmp_modlimb_invert_table
LEAF_PROC mpn_divexact_1
movsxd r8, r8d
mov r10, rdx
mov rax, r9
and rax, byte 1
add rax, r8
cmp rax, byte 4
jae .1
xor rdx,rdx
.0: mov rax, [r10+r8*8-8]
div r9
mov [rcx+r8*8-8], rax
sub r8, 1
jnz .0
ret ; avoid single byte return
.1:
FRAME_PROC ?mpn_divexact, 0, reg_save_list
mov rsi, rdx ; src pointer
mov rdi, rcx ; dst pointer
bsf rcx, r9 ; remove powers of two
shr r9, cl
mov rax, r9
shr rax, 1
and rax, 127
lea rdx, [rel __gmp_modlimb_invert_table]
movzx rax, byte [rdx+rax]
; If f(x) = 0, then x[n+1] = x[n] - f(x) / f'(x) is Newton's iteration for a
; root. With f(x) = 1/x - v we obtain x[n + 1] = 2 * x[n] - v * x[n] * x[n]
; as an iteration for x = 1 / v. This provides quadratic convergence so
; that the number of bits of precision doubles on each iteration. The
; iteration starts with 8-bit precision.
lea edx, [rax+rax]
imul eax, eax
imul eax, r9d
sub edx, eax ; inv -> rdx (16-bit approx)
lea eax, [rdx+rdx]
imul edx, edx
imul edx, r9d
sub eax, edx ; inv -> rdx (32-bit approx)
lea rdx, [rax+rax]
imul rax, rax
imul rax, r9
sub rdx, rax ; inv -> rdx (64-bit approx)
lea rsi, [rsi+r8*8]
lea rdi, [rdi+r8*8]
neg r8
mov r10, rdx ; inverse multiplier -> r10
xor r11, r11
mov rax, [rsi+r8*8]
or rcx, rcx
mov rdx, [rsi+r8*8+8]
jz .3 ; if divisor is odd
shrd rax, rdx, cl
add r8, 1
jmp .5
xalign 16
.2: mul r9 ; divisor is odd
mov rax, [rsi+r8*8]
sub rdx, r11
sub rax, rdx
sbb r11, r11
.3: imul rax, r10
mov [rdi+r8*8], rax
add r8, 1
jnz .2
jmp .6
xalign 16
.4: mul r9 ; divisor is even
sub rdx, r11
mov rax, [rsi+r8*8-8]
mov r11, [rsi+r8*8]
shrd rax, r11, cl
sub rax, rdx
sbb r11, r11
.5: imul rax, r10
mov [rdi+r8*8-8],rax
add r8, 1
jnz .4
mul r9
mov rax, [rsi-8]
sub rdx, r11
shr rax, cl
sub rax, rdx
imul rax, r10
mov [rdi-8], rax
.6: END_PROC reg_save_list
end