1 dnl AMD K6 mpn_lshift -- mpn left shift.
3 dnl K6: 3.0 cycles/limb
6 dnl Copyright (C) 1999, 2000 Free Software Foundation, Inc.
8 dnl This file is part of the GNU MP Library.
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.
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.
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.
26 include(`../config.m4')
29 C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size,
32 C The loop runs at 3 cycles/limb, limited by decoding and by having 3 mmx
33 C instructions. This is despite every second fetch being unaligned.
36 defframe(PARAM_SHIFT,16)
37 defframe(PARAM_SIZE, 12)
38 defframe(PARAM_SRC, 8)
39 defframe(PARAM_DST, 4)
47 C The 1 limb case can be done without the push %ebx, but it's then
48 C still the same speed. The push is left as a free helping hand for
49 C the two_or_more code.
52 pushl %ebx FRAME_pushl()
57 movl PARAM_SHIFT, %ecx
60 movl (%ebx), %edx C src limb
63 shldl( %cl, %edx, %eax) C return value
67 movl %edx, (%ebx) C dst limb
73 ALIGN(16) C avoid offset 0x1f
74 nop C avoid bad cache line crossing
81 movl (%ebx,%eax,4), %edx C src high limb
84 movd PARAM_SHIFT, %mm6
85 addl $32, %ecx C 32-shift
93 C eax counter, size-1 to 1
102 movq -4(%ebx,%eax,4), %mm0
107 movd %mm0, 4(%ecx,%eax,4)