Home | History | Annotate | Line # | Download | only in k6 
      1 dnl  AMD K6 mpn_modexact_1_odd -- exact division style remainder.
      2 
      3 dnl  Copyright 2000-2003, 2007 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 K6: 10.0 cycles/limb
     35 
     36 
     37 C mp_limb_t mpn_modexact_1_odd (mp_srcptr src, mp_size_t size,
     38 C                               mp_limb_t divisor);
     39 C mp_limb_t mpn_modexact_1c_odd (mp_srcptr src, mp_size_t size,
     40 C                                mp_limb_t divisor, mp_limb_t carry);
     41 C
     42 C A special case for high<divisor at the end measured only about 4 cycles
     43 C faster, and so isn't used.
     44 C
     45 C A special case for size==1 using a divl rather than the inverse measured
     46 C only about 5 cycles faster, and so isn't used.  When size==1 and
     47 C high<divisor it can skip a division and be a full 24 cycles faster, but
     48 C this isn't an important case.
     49 
     50 defframe(PARAM_CARRY,  16)
     51 defframe(PARAM_DIVISOR,12)
     52 defframe(PARAM_SIZE,   8)
     53 defframe(PARAM_SRC,    4)
     54 
     55 	TEXT
     56 
     57 	ALIGN(32)
     58 PROLOGUE(mpn_modexact_1c_odd)
     59 deflit(`FRAME',0)
     60 
     61 	movl	PARAM_DIVISOR, %ecx
     62 	pushl	%esi		FRAME_pushl()
     63 
     64 	movl	PARAM_CARRY, %edx
     65 	jmp	L(start_1c)
     66 
     67 EPILOGUE()
     68 
     69 
     70 	ALIGN(16)
     71 PROLOGUE(mpn_modexact_1_odd)
     72 deflit(`FRAME',0)
     73 
     74 	movl	PARAM_DIVISOR, %ecx
     75 	pushl	%esi		FRAME_pushl()
     76 
     77 	xorl	%edx, %edx
     78 L(start_1c):
     79 	pushl	%edi		FRAME_pushl()
     80 
     81 	shrl	%ecx			C d/2
     82 	movl	PARAM_DIVISOR, %esi
     83 
     84 	andl	$127, %ecx		C d/2, 7 bits
     85 	pushl	%ebp		FRAME_pushl()
     86 
     87 ifdef(`PIC',`
     88 	LEA(	binvert_limb_table, %edi)
     89 Zdisp(	movzbl,	0,(%ecx,%edi), %edi)		C inv 8 bits
     90 ',`
     91 	movzbl	binvert_limb_table(%ecx), %edi	C inv 8 bits
     92 ')
     93 	leal	(%edi,%edi), %ecx	C 2*inv
     94 
     95 	imull	%edi, %edi		C inv*inv
     96 
     97 	movl	PARAM_SRC, %eax
     98 	movl	PARAM_SIZE, %ebp
     99 
    100 	imull	%esi, %edi		C inv*inv*d
    101 
    102 	pushl	%ebx		FRAME_pushl()
    103 	leal	(%eax,%ebp,4), %ebx	C src end
    104 
    105 	subl	%edi, %ecx		C inv = 2*inv - inv*inv*d
    106 	leal	(%ecx,%ecx), %edi	C 2*inv
    107 
    108 	imull	%ecx, %ecx		C inv*inv
    109 
    110 	movl	(%eax), %eax		C src low limb
    111 	negl	%ebp			C -size
    112 
    113 	imull	%esi, %ecx		C inv*inv*d
    114 
    115 	subl	%ecx, %edi		C inv = 2*inv - inv*inv*d
    116 
    117 	ASSERT(e,`	C d*inv == 1 mod 2^GMP_LIMB_BITS
    118 	pushl	%eax
    119 	movl	%esi, %eax
    120 	imull	%edi, %eax
    121 	cmpl	$1, %eax
    122 	popl	%eax')
    123 
    124 	jmp	L(entry)
    125 
    126 
    127 C Rotating the mul to the top of the loop saves 1 cycle, presumably by
    128 C hiding the loop control under the imul latency.
    129 C
    130 C The run time is 10 cycles, but decoding is only 9 (and the dependent chain
    131 C only 8).  It's not clear how to get down to 9 cycles.
    132 C
    133 C The xor and rcl to handle the carry bit could be an sbb instead, with the
    134 C the carry bit add becoming a sub, but that doesn't save anything.
    135 
    136 L(top):
    137 	C eax	(low product)
    138 	C ebx	src end
    139 	C ecx	carry bit, 0 or 1
    140 	C edx	(high product, being carry limb)
    141 	C esi	divisor
    142 	C edi	inverse
    143 	C ebp	counter, limbs, negative
    144 
    145 	mull	%esi
    146 
    147 	movl	(%ebx,%ebp,4), %eax
    148 	addl	%ecx, %edx		C apply carry bit to carry limb
    149 
    150 L(entry):
    151 	xorl	%ecx, %ecx
    152 	subl	%edx, %eax		C apply carry limb
    153 
    154 	rcll	%ecx
    155 
    156 	imull	%edi, %eax
    157 
    158 	incl	%ebp
    159 	jnz	L(top)
    160 
    161 
    162 
    163 	popl	%ebx
    164 	popl	%ebp
    165 
    166 	mull	%esi
    167 
    168 	popl	%edi
    169 	popl	%esi
    170 
    171 	leal	(%ecx,%edx), %eax
    172 
    173 	ret
    174 
    175 EPILOGUE()
    176 ASM_END()
    177