Home | History | Annotate | Line # | Download | only in sse2 
      1 dnl  Intel Pentium-4 mpn_cnd_add_n -- mpn addition.
      2 
      3 dnl  Copyright 2001, 2002, 2013 Free Software Foundation, Inc.
      4 
      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 modify
      8 dnl  it under the terms of either:
      9 dnl
     10 dnl    * the GNU Lesser General Public License as published by the Free
     11 dnl      Software Foundation; either version 3 of the License, or (at your
     12 dnl      option) any later version.
     13 dnl
     14 dnl  or
     15 dnl
     16 dnl    * the GNU General Public License as published by the Free Software
     17 dnl      Foundation; either version 2 of the License, or (at your option) any
     18 dnl      later version.
     19 dnl
     20 dnl  or both in parallel, as here.
     21 dnl
     22 dnl  The GNU MP Library is distributed in the hope that it will be useful, but
     23 dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
     24 dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
     25 dnl  for more details.
     26 dnl
     27 dnl  You should have received copies of the GNU General Public License and the
     28 dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
     29 dnl  see https://www.gnu.org/licenses/.
     30 
     31 include(`../config.m4')
     32 
     33 
     34 C			    cycles/limb
     35 C P6 model 0-8,10-12		 -
     36 C P6 model 9   (Banias)		 ?
     37 C P6 model 13  (Dothan)		 4.67
     38 C P4 model 0-1 (Willamette)	 ?
     39 C P4 model 2   (Northwood)	 5
     40 C P4 model 3-4 (Prescott)	 5.25
     41 
     42 defframe(PARAM_SIZE, 20)
     43 defframe(PARAM_SRC2, 16)
     44 defframe(PARAM_SRC1, 12)
     45 defframe(PARAM_DST,  8)
     46 defframe(PARAM_CND,  4)
     47 
     48 dnl  re-use parameter space
     49 define(SAVE_EBX,`PARAM_SRC1')
     50 
     51 define(`cnd', `%mm3')
     52 
     53 	TEXT
     54 	ALIGN(8)
     55 
     56 	ALIGN(8)
     57 PROLOGUE(mpn_cnd_add_n)
     58 deflit(`FRAME',0)
     59 	pxor	%mm0, %mm0
     60 
     61 	mov	PARAM_CND, %eax
     62 	neg	%eax
     63 	sbb	%eax, %eax
     64 	movd	%eax, cnd
     65 
     66 	mov	PARAM_SRC1, %eax
     67 	mov	%ebx, SAVE_EBX
     68 	mov	PARAM_SRC2, %ebx
     69 	mov	PARAM_DST, %edx
     70 	mov	PARAM_SIZE, %ecx
     71 
     72 	lea	(%eax,%ecx,4), %eax	C src1 end
     73 	lea	(%ebx,%ecx,4), %ebx	C src2 end
     74 	lea	(%edx,%ecx,4), %edx	C dst end
     75 	neg	%ecx			C -size
     76 
     77 L(top):	movd	(%ebx,%ecx,4), %mm2
     78 	movd	(%eax,%ecx,4), %mm1
     79 	pand	cnd, %mm2
     80 	paddq	%mm2, %mm1
     81 
     82 	paddq	%mm1, %mm0
     83 	movd	%mm0, (%edx,%ecx,4)
     84 
     85 	psrlq	$32, %mm0
     86 
     87 	add	$1, %ecx
     88 	jnz	L(top)
     89 
     90 	movd	%mm0, %eax
     91 	mov	SAVE_EBX, %ebx
     92 	emms
     93 	ret
     94 
     95 EPILOGUE()
     96