Return to copyd.asm CVS log

Up to [local] / OpenXM_contrib / gmp / mpn / x86 / pentium4

Annotation of OpenXM_contrib/gmp/mpn/x86/pentium4/copyd.asm, Revision 1.1.1.1

1.1       ohara       1: dnl  Pentium-4 mpn_copyd -- copy limb vector, decrementing.
                      2: dnl
                      3:
                      4: dnl  Copyright 1999, 2000, 2001 Free Software Foundation, Inc.
                      5: dnl
                      6: dnl  This file is part of the GNU MP Library.
                      7: dnl
                      8: dnl  The GNU MP Library is free software; you can redistribute it and/or
                      9: dnl  modify it under the terms of the GNU Lesser General Public License as
                     10: dnl  published by the Free Software Foundation; either version 2.1 of the
                     11: dnl  License, or (at your option) any later version.
                     12: dnl
                     13: dnl  The GNU MP Library is distributed in the hope that it will be useful,
                     14: dnl  but WITHOUT ANY WARRANTY; without even the implied warranty of
                     15: dnl  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
                     16: dnl  Lesser General Public License for more details.
                     17: dnl
                     18: dnl  You should have received a copy of the GNU Lesser General Public
                     19: dnl  License along with the GNU MP Library; see the file COPYING.LIB.  If
                     20: dnl  not, write to the Free Software Foundation, Inc., 59 Temple Place -
                     21: dnl  Suite 330, Boston, MA 02111-1307, USA.
                     22:
                     23:
                     24: dnl  The std/rep/movsl/cld is very slow for small blocks on pentium4.  Its
                     25: dnl  startup time seems to be about 165 cycles.  It then copies at a rate of
                     26: dnl  one limb per cycle.  We therefore fall back to an open-coded 2 c/l
                     27: dnl  copying loop for smaller sizes.
                     28:
                     29: dnl  Ultimately, we may want to use 64-bit movq or 128-bit movdqu in some
                     30: dnl  nifty unrolled arrangement.  Clearly, that could reach much higher
                     31: dnl  speeds, at least for large blocks.
                     32:
                     33: include(`../config.m4')
                     34:
                     35:
                     36: defframe(PARAM_SIZE, 12)
                     37: defframe(PARAM_SRC, 8)
                     38: defframe(PARAM_DST,  4)
                     39:
                     40:        TEXT
                     41:        ALIGN(8)
                     42:
                     43: PROLOGUE(mpn_copyd)
                     44: deflit(`FRAME',0)
                     45:
                     46:        movl    PARAM_SIZE, %ecx
                     47:
                     48:        movl    PARAM_SRC, %eax
                     49:        movl    PARAM_DST, %edx
                     50:        movl    %ebx, PARAM_SIZE
                     51:        addl    $-1, %ecx
                     52:        js      L(end)
                     53:
                     54: L(loop):
                     55:        movl    (%eax,%ecx,4), %ebx
                     56:        movl    %ebx, (%edx,%ecx,4)
                     57:        addl    $-1, %ecx
                     58:
                     59:        jns     L(loop)
                     60: L(end):
                     61:        movl    PARAM_SIZE, %ebx
                     62:        ret
                     63:
                     64: EPILOGUE()