Return to diveby3.asm CVS log

Up to [local] / OpenXM_contrib / gmp / mpn / x86 / k6

Annotation of OpenXM_contrib/gmp/mpn/x86/k6/diveby3.asm, Revision 1.1.1.1

1.1       maekawa     1: dnl  AMD K6 mpn_divexact_by3 -- mpn division by 3, expecting no remainder.
                      2: dnl
                      3: dnl  K6: 11.0 cycles/limb
                      4:
                      5:
                      6: dnl  Copyright (C) 2000 Free Software Foundation, Inc.
                      7: dnl
                      8: dnl  This file is part of the GNU MP Library.
                      9: dnl
                     10: dnl  The GNU MP Library is free software; you can redistribute it and/or
                     11: dnl  modify it under the terms of the GNU Lesser General Public License as
                     12: dnl  published by the Free Software Foundation; either version 2.1 of the
                     13: dnl  License, or (at your option) any later version.
                     14: dnl
                     15: dnl  The GNU MP Library is distributed in the hope that it will be useful,
                     16: dnl  but WITHOUT ANY WARRANTY; without even the implied warranty of
                     17: dnl  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
                     18: dnl  Lesser General Public License for more details.
                     19: dnl
                     20: dnl  You should have received a copy of the GNU Lesser General Public
                     21: dnl  License along with the GNU MP Library; see the file COPYING.LIB.  If
                     22: dnl  not, write to the Free Software Foundation, Inc., 59 Temple Place -
                     23: dnl  Suite 330, Boston, MA 02111-1307, USA.
                     24:
                     25:
                     26: include(`../config.m4')
                     27:
                     28:
                     29: C mp_limb_t mpn_divexact_by3c (mp_ptr dst, mp_srcptr src, mp_size_t size,
                     30: C                              mp_limb_t carry);
                     31: C
                     32: C Using %esi in (%esi,%ecx,4) or 0(%esi,%ecx,4) addressing modes doesn't
                     33: C lead to vector decoding, unlike plain (%esi) does.
                     34:
                     35: defframe(PARAM_CARRY,16)
                     36: defframe(PARAM_SIZE, 12)
                     37: defframe(PARAM_SRC,   8)
                     38: defframe(PARAM_DST,   4)
                     39:
                     40: dnl  multiplicative inverse of 3, modulo 2^32
                     41: deflit(INVERSE_3, 0xAAAAAAAB)
                     42:
                     43:        .text
                     44:        ALIGN(32)
                     45:
                     46: PROLOGUE(mpn_divexact_by3c)
                     47: deflit(`FRAME',0)
                     48:
                     49:        movl    PARAM_SIZE, %ecx
                     50:        pushl   %esi            defframe_pushl(SAVE_ESI)
                     51:
                     52:        movl    PARAM_SRC, %esi
                     53:        pushl   %edi            defframe_pushl(SAVE_EDI)
                     54:
                     55:        movl    PARAM_DST, %edi
                     56:        pushl   %ebx            defframe_pushl(SAVE_EBX)
                     57:
                     58:        movl    PARAM_CARRY, %ebx
                     59:        leal    (%esi,%ecx,4), %esi
                     60:
                     61:        pushl   $3              defframe_pushl(VAR_THREE)
                     62:        leal    (%edi,%ecx,4), %edi
                     63:
                     64:        negl    %ecx
                     65:
                     66:
                     67:        C Need 32 alignment for claimed speed, to avoid the movl store
                     68:        C opcode/modrm crossing a cache line boundary
                     69:
                     70:        ALIGN(32)
                     71: L(top):
                     72:        C eax   scratch, low product
                     73:        C ebx   carry limb (0 to 3)
                     74:        C ecx   counter, limbs, negative
                     75:        C edx   scratch, high product
                     76:        C esi   &src[size]
                     77:        C edi   &dst[size]
                     78:        C ebp
                     79:        C
                     80:        C The 0(%esi,%ecx,4) form pads so the finishup "movl %ebx, %eax"
                     81:        C doesn't cross a 32 byte boundary, saving a couple of cycles
                     82:        C (that's a fixed couple, not per loop).
                     83:
                     84: Zdisp( movl,   0,(%esi,%ecx,4), %eax)
                     85:        subl    %ebx, %eax
                     86:
                     87:        setc    %bl
                     88:
                     89:        imull   $INVERSE_3, %eax
                     90:
                     91:        movl    %eax, (%edi,%ecx,4)
                     92:        addl    $2, %ecx
                     93:
                     94:        mull    VAR_THREE
                     95:
                     96:        addl    %edx, %ebx
                     97:        loop    L(top)
                     98:
                     99:
                    100:        movl    SAVE_ESI, %esi
                    101:        movl    %ebx, %eax
                    102:
                    103:        movl    SAVE_EBX, %ebx
                    104:
                    105:        movl    SAVE_EDI, %edi
                    106:        addl    $FRAME, %esp
                    107:
                    108:        ret
                    109:
                    110: EPILOGUE()