Home | History | Annotate | Line # | Download | only in s390_64 
      1 dnl  S/390-64 mpn_mul_basecase.
      2 
      3 dnl  Copyright 2011 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 C            cycles/limb
     34 C z900		 ?
     35 C z990		23
     36 C z9		 ?
     37 C z10		28
     38 C z196		 ?
     39 
     40 C TODO
     41 C  * Perhaps add special case for un <= 2.
     42 C  * Replace loops by faster code.  The mul_1 and addmul_1 loops could be sped
     43 C    up by about 10%.
     44 
     45 C INPUT PARAMETERS
     46 define(`rp',	`%r2')
     47 define(`up',	`%r3')
     48 define(`un',	`%r4')
     49 define(`vp',	`%r5')
     50 define(`vn',	`%r6')
     51 
     52 define(`zero',	`%r8')
     53 
     54 ASM_START()
     55 PROLOGUE(mpn_mul_basecase)
     56 	cghi	un, 2
     57 	jhe	L(ge2)
     58 
     59 C un = vn = 1
     60 	lg	%r1, 0(vp)
     61 	mlg	%r0, 0(up)
     62 	stg	%r1, 0(rp)
     63 	stg	%r0, 8(rp)
     64 	br	%r14
     65 
     66 L(ge2):	C jne	L(gen)
     67 
     68 
     69 L(gen):
     70 C mul_1 =======================================================================
     71 
     72 	stmg	%r6, %r12, 48(%r15)
     73 	lghi	zero, 0
     74 	aghi	un, -1
     75 
     76 	lg	%r7, 0(vp)
     77 	lg	%r11, 0(up)
     78 	lghi	%r12, 8			C init index register
     79 	mlgr	%r10, %r7
     80 	lgr	%r9, un
     81 	stg	%r11, 0(rp)
     82 	cr	%r15, %r15		C clear carry flag
     83 
     84 L(tm):	lg	%r1, 0(%r12,up)
     85 	mlgr	%r0, %r7
     86 	alcgr	%r1, %r10
     87 	lgr	%r10, %r0		C copy high part to carry limb
     88 	stg	%r1, 0(%r12,rp)
     89 	la	%r12, 8(%r12)
     90 	brctg	%r9, L(tm)
     91 
     92 	alcgr	%r0, zero
     93 	stg	%r0, 0(%r12,rp)
     94 
     95 C addmul_1 loop ===============================================================
     96 
     97 	aghi	vn, -1
     98 	je	L(outer_end)
     99 L(outer_loop):
    100 
    101 	la	rp, 8(rp)		C rp += 1
    102 	la	vp, 8(vp)		C up += 1
    103 	lg	%r7, 0(vp)
    104 	lg	%r11, 0(up)
    105 	lghi	%r12, 8			C init index register
    106 	mlgr	%r10, %r7
    107 	lgr	%r9, un
    108 	alg	%r11, 0(rp)
    109 	stg	%r11, 0(rp)
    110 
    111 L(tam):	lg	%r1, 0(%r12,up)
    112 	lg	%r11, 0(%r12,rp)
    113 	mlgr	%r0, %r7
    114 	alcgr	%r1, %r11
    115 	alcgr	%r0, zero
    116 	algr	%r1, %r10
    117 	lgr	%r10, %r0
    118 	stg	%r1, 0(%r12,rp)
    119 	la	%r12, 8(%r12)
    120 	brctg	%r9, L(tam)
    121 
    122 	alcgr	%r0, zero
    123 	stg	%r0, 0(%r12,rp)
    124 
    125 	brctg	vn, L(outer_loop)
    126 L(outer_end):
    127 
    128 	lmg	%r6, %r12, 48(%r15)
    129 	br	%r14
    130 EPILOGUE()
    131