dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc.
dnl This file is part of the GNU MP Library.
dnl
dnl The GNU MP Library is free software; you can redistribute it and/or modify
dnl it under the terms of either:
dnl
dnl * the GNU Lesser General Public License as published by the Free
dnl Software Foundation; either version 3 of the License, or (at your
dnl option) any later version.
dnl
dnl or
dnl
dnl * the GNU General Public License as published by the Free Software
dnl Foundation; either version 2 of the License, or (at your option) any
dnl later version.
dnl
dnl or both in parallel, as here.
dnl
dnl The GNU MP Library is distributed in the hope that it will be useful, but
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
dnl for more details.
dnl
dnl You should have received copies of the GNU General Public License and the
dnl GNU Lesser General Public License along with the GNU MP Library. If not,
dnl see https://www.gnu.org/licenses/.
include(`../config.m4')
C K6-2: 1.75 cycles/limb
C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size,
C unsigned shift);
C
dnl used after src has been fetched
define(VAR_RETVAL,`PARAM_SRC')
dnl minimum 9, because unrolled loop can't handle less
deflit(UNROLL_THRESHOLD, 9)
TEXT ALIGN(32)
PROLOGUE(mpn_lshift)
deflit(`FRAME',0)
C The 1 limb case can be done without the push %ebx, but it's then
C still the same speed. The push is left as a free helping hand for
C the two_or_more code.
movl PARAM_SIZE, %eax
pushl %ebx FRAME_pushl()
movl PARAM_SRC, %ebx
decl %eax
movl PARAM_SHIFT, %ecx jnz L(two_or_more)
movl (%ebx), %edx C src limb
movl PARAM_DST, %ebx
shldl( %cl, %edx, %eax) C return value
shll %cl, %edx
movl %edx, (%ebx) C dst limb
popl %ebx
ret
C ----------------------------------------------------------------------------- ALIGN(16) C avoid offset 0x1f
L(two_or_more):
C eax size-1
C ebx src
C ecx shift
C edx
movl (%ebx,%eax,4), %edx C src high limb
negl %ecx
movd PARAM_SHIFT, %mm6
addl $32, %ecx C 32-shift
shrl %cl, %edx
cmpl $UNROLL_THRESHOLD-1, %eax
movl %edx, VAR_RETVAL
jae L(unroll)
movd %ecx, %mm7
movl %eax, %ecx
movl PARAM_DST, %eax
L(simple):
C eax dst
C ebx src
C ecx counter, size-1 to 1
C edx retval
C
C mm0 scratch
C mm6 shift
C mm7 32-shift
C ----------------------------------------------------------------------------- ALIGN(16)
L(unroll):
C eax size-1
C ebx src
C ecx 32-shift
C edx retval (but instead VAR_RETVAL is used)
C
C mm6 shift
addl $32, %ecx
movl PARAM_DST, %edx
movd %ecx, %mm7
subl $7, %eax C size-8
leal (%edx,%eax,4), %ecx C alignment of dst
movq 32-8(%ebx,%eax,4), %mm2 C src high qword
testb $4, %cl
jz L(dst_aligned)
psllq %mm6, %mm2
psrlq $32, %mm2
decl %eax
movd %mm2, 32(%edx,%eax,4) C dst high limb
movq 32-8(%ebx,%eax,4), %mm2 C new src high qword
L(dst_aligned):
movq 32-16(%ebx,%eax,4), %mm0 C src second highest qword
C This loop is the important bit, the rest is just support for it.
C Four src limbs are held at the start, and four more will be read.
C Four dst limbs will be written. This schedule seems necessary for
C full speed.
C
C The use of size-8 lets the loop stop when %eax goes negative and
C leaves -4 to -1 which can be tested with test $1 and $2.
L(top):
C eax counter, size-8 step by -4 until <0
C ebx src
C ecx
C edx dst
C
C mm0 src next qword
C mm1 scratch
C mm2 src prev qword
C mm6 shift
C mm7 64-shift
psllq %mm6, %mm2
subl $4, %eax
movq %mm0, %mm1
psrlq %mm7, %mm0
por %mm0, %mm2
movq 24(%ebx,%eax,4), %mm0
psllq %mm6, %mm1
movq %mm2, 40(%edx,%eax,4)
movq %mm0, %mm2
psrlq %mm7, %mm0
por %mm0, %mm1
movq 16(%ebx,%eax,4), %mm0
movq %mm1, 32(%edx,%eax,4)
jnc L(top)
C Now have four limbs in mm2 (prev) and mm0 (next), plus eaxmod 4.
C
C 8(%ebx) is the next source, and 24(%edx) is the next destination.
C %eax is between -4 and -1, representing respectively 0 to 3 extra
C limbs that must be read.
testl $2, %eax C testl to avoid bad cache line crossing jz L(finish_nottwo)
C Two more limbs: lshift mm2, OR it with rshifted mm0, mm0 becomes
C new mm2 and a new mm0 is loaded.
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.