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 1.1.1.2

1.1       maekawa     1: dnl  AMD K6 mpn_divexact_by3 -- mpn division by 3, expecting no remainder.
                      2:
1.1.1.2 ! ohara       3: dnl  Copyright 2000, 2001, 2002 Free Software Foundation, Inc.
1.1       maekawa     4: dnl
                      5: dnl  This file is part of the GNU MP Library.
                      6: dnl
                      7: dnl  The GNU MP Library is free software; you can redistribute it and/or
                      8: dnl  modify it under the terms of the GNU Lesser General Public License as
                      9: dnl  published by the Free Software Foundation; either version 2.1 of the
                     10: dnl  License, or (at your option) any later version.
                     11: dnl
                     12: dnl  The GNU MP Library is distributed in the hope that it will be useful,
                     13: dnl  but WITHOUT ANY WARRANTY; without even the implied warranty of
                     14: dnl  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
                     15: dnl  Lesser General Public License for more details.
                     16: dnl
                     17: dnl  You should have received a copy of the GNU Lesser General Public
                     18: dnl  License along with the GNU MP Library; see the file COPYING.LIB.  If
                     19: dnl  not, write to the Free Software Foundation, Inc., 59 Temple Place -
                     20: dnl  Suite 330, Boston, MA 02111-1307, USA.
                     21:
                     22: include(`../config.m4')
                     23:
                     24:
1.1.1.2 ! ohara      25: C K6: 11.0 cycles/limb
        !            26:
        !            27:
1.1       maekawa    28: C mp_limb_t mpn_divexact_by3c (mp_ptr dst, mp_srcptr src, mp_size_t size,
                     29: C                              mp_limb_t carry);
                     30: C
                     31: C Using %esi in (%esi,%ecx,4) or 0(%esi,%ecx,4) addressing modes doesn't
                     32: C lead to vector decoding, unlike plain (%esi) does.
                     33:
                     34: defframe(PARAM_CARRY,16)
                     35: defframe(PARAM_SIZE, 12)
                     36: defframe(PARAM_SRC,   8)
                     37: defframe(PARAM_DST,   4)
                     38:
                     39: dnl  multiplicative inverse of 3, modulo 2^32
                     40: deflit(INVERSE_3, 0xAAAAAAAB)
                     41:
1.1.1.2 ! ohara      42:        TEXT
1.1       maekawa    43:        ALIGN(32)
                     44:
                     45: PROLOGUE(mpn_divexact_by3c)
                     46: deflit(`FRAME',0)
                     47:
                     48:        movl    PARAM_SIZE, %ecx
                     49:        pushl   %esi            defframe_pushl(SAVE_ESI)
                     50:
                     51:        movl    PARAM_SRC, %esi
                     52:        pushl   %edi            defframe_pushl(SAVE_EDI)
                     53:
                     54:        movl    PARAM_DST, %edi
                     55:        pushl   %ebx            defframe_pushl(SAVE_EBX)
                     56:
                     57:        movl    PARAM_CARRY, %ebx
                     58:        leal    (%esi,%ecx,4), %esi
                     59:
                     60:        pushl   $3              defframe_pushl(VAR_THREE)
                     61:        leal    (%edi,%ecx,4), %edi
                     62:
                     63:        negl    %ecx
                     64:
                     65:
                     66:        C Need 32 alignment for claimed speed, to avoid the movl store
                     67:        C opcode/modrm crossing a cache line boundary
                     68:
                     69:        ALIGN(32)
                     70: L(top):
                     71:        C eax   scratch, low product
1.1.1.2 ! ohara      72:        C ebx   carry limb (0 to 2)
1.1       maekawa    73:        C ecx   counter, limbs, negative
                     74:        C edx   scratch, high product
                     75:        C esi   &src[size]
                     76:        C edi   &dst[size]
                     77:        C ebp
                     78:        C
                     79:        C The 0(%esi,%ecx,4) form pads so the finishup "movl %ebx, %eax"
                     80:        C doesn't cross a 32 byte boundary, saving a couple of cycles
                     81:        C (that's a fixed couple, not per loop).
                     82:
                     83: Zdisp( movl,   0,(%esi,%ecx,4), %eax)
                     84:        subl    %ebx, %eax
                     85:
                     86:        setc    %bl
                     87:
1.1.1.2 ! ohara      88:        imull   $INVERSE_3, %eax, %eax
1.1       maekawa    89:
                     90:        movl    %eax, (%edi,%ecx,4)
                     91:        addl    $2, %ecx
                     92:
                     93:        mull    VAR_THREE
                     94:
                     95:        addl    %edx, %ebx
                     96:        loop    L(top)
                     97:
                     98:
                     99:        movl    SAVE_ESI, %esi
                    100:        movl    %ebx, %eax
                    101:
                    102:        movl    SAVE_EBX, %ebx
                    103:
                    104:        movl    SAVE_EDI, %edi
                    105:        addl    $FRAME, %esp
                    106:
                    107:        ret
                    108:
                    109: EPILOGUE()