LXR linux/arch/alpha/lib/divide.S

   1/*
   2 * arch/alpha/lib/divide.S
   3 *
   4 * (C) 1995 Linus Torvalds
   5 *
   6 * Alpha division..
   7 */
   8
   9/*
  10 * The alpha chip doesn't provide hardware division, so we have to do it
  11 * by hand.  The compiler expects the functions
  12 *
  13 *      __divqu: 64-bit unsigned long divide
  14 *      __remqu: 64-bit unsigned long remainder
  15 *      __divqs/__remqs: signed 64-bit
  16 *      __divlu/__remlu: unsigned 32-bit
  17 *      __divls/__remls: signed 32-bit
  18 *
  19 * These are not normal C functions: instead of the normal
  20 * calling sequence, these expect their arguments in registers
  21 * $24 and $25, and return the result in $27. Register $28 may
  22 * be clobbered (assembly temporary), anything else must be saved. 
  23 *
  24 * In short: painful.
  25 *
  26 * This is a rather simple bit-at-a-time algorithm: it's very good
  27 * at dividing random 64-bit numbers, but the more usual case where
  28 * the divisor is small is handled better by the DEC algorithm
  29 * using lookup tables. This uses much less memory, though, and is
  30 * nicer on the cache.. Besides, I don't know the copyright status
  31 * of the DEC code.
  32 */
  33
  34/*
  35 * My temporaries:
  36 *      $0 - current bit
  37 *      $1 - shifted divisor
  38 *      $2 - modulus/quotient
  39 *
  40 *      $23 - return address
  41 *      $24 - dividend
  42 *      $25 - divisor
  43 *
  44 *      $27 - quotient/modulus
  45 *      $28 - compare status
  46 */
  47
  48#include <asm/export.h>
  49#define halt .long 0
  50
  51/*
  52 * Select function type and registers
  53 */
  54#define mask    $0
  55#define divisor $1
  56#define compare $28
  57#define tmp1    $3
  58#define tmp2    $4
  59
  60#ifdef DIV
  61#define DIV_ONLY(x,y...) x,##y
  62#define MOD_ONLY(x,y...)
  63#define func(x) __div##x
  64#define modulus $2
  65#define quotient $27
  66#define GETSIGN(x) xor $24,$25,x
  67#define STACK 48
  68#else
  69#define DIV_ONLY(x,y...)
  70#define MOD_ONLY(x,y...) x,##y
  71#define func(x) __rem##x
  72#define modulus $27
  73#define quotient $2
  74#define GETSIGN(x) bis $24,$24,x
  75#define STACK 32
  76#endif
  77
  78/*
  79 * For 32-bit operations, we need to extend to 64-bit
  80 */
  81#ifdef INTSIZE
  82#define ufunction func(lu)
  83#define sfunction func(l)
  84#define LONGIFY(x) zapnot x,15,x
  85#define SLONGIFY(x) addl x,0,x
  86#else
  87#define ufunction func(qu)
  88#define sfunction func(q)
  89#define LONGIFY(x)
  90#define SLONGIFY(x)
  91#endif
  92
  93.set noat
  94.align  3
  95.globl  ufunction
  96.ent    ufunction
  97ufunction:
  98        subq    $30,STACK,$30
  99        .frame  $30,STACK,$23
 100        .prologue 0
 101
 1027:      stq     $1, 0($30)
 103        bis     $25,$25,divisor
 104        stq     $2, 8($30)
 105        bis     $24,$24,modulus
 106        stq     $0,16($30)
 107        bis     $31,$31,quotient
 108        LONGIFY(divisor)
 109        stq     tmp1,24($30)
 110        LONGIFY(modulus)
 111        bis     $31,1,mask
 112        DIV_ONLY(stq tmp2,32($30))
 113        beq     divisor, 9f                     /* div by zero */
 114
 115#ifdef INTSIZE
 116        /*
 117         * shift divisor left, using 3-bit shifts for
 118         * 32-bit divides as we can't overflow. Three-bit
 119         * shifts will result in looping three times less
 120         * here, but can result in two loops more later.
 121         * Thus using a large shift isn't worth it (and
 122         * s8add pairs better than a sll..)
 123         */
 1241:      cmpult  divisor,modulus,compare
 125        s8addq  divisor,$31,divisor
 126        s8addq  mask,$31,mask
 127        bne     compare,1b
 128#else
 1291:      cmpult  divisor,modulus,compare
 130        blt     divisor, 2f
 131        addq    divisor,divisor,divisor
 132        addq    mask,mask,mask
 133        bne     compare,1b
 134        unop
 135#endif
 136
 137        /* ok, start to go right again.. */
 1382:      DIV_ONLY(addq quotient,mask,tmp2)
 139        srl     mask,1,mask
 140        cmpule  divisor,modulus,compare
 141        subq    modulus,divisor,tmp1
 142        DIV_ONLY(cmovne compare,tmp2,quotient)
 143        srl     divisor,1,divisor
 144        cmovne  compare,tmp1,modulus
 145        bne     mask,2b
 146
 1479:      ldq     $1, 0($30)
 148        ldq     $2, 8($30)
 149        ldq     $0,16($30)
 150        ldq     tmp1,24($30)
 151        DIV_ONLY(ldq tmp2,32($30))
 152        addq    $30,STACK,$30
 153        ret     $31,($23),1
 154        .end    ufunction
 155EXPORT_SYMBOL(ufunction)
 156
 157/*
 158 * Uhh.. Ugly signed division. I'd rather not have it at all, but
 159 * it's needed in some circumstances. There are different ways to
 160 * handle this, really. This does:
 161 *      -a / b = a / -b = -(a / b)
 162 *      -a % b = -(a % b)
 163 *      a % -b = a % b
 164 * which is probably not the best solution, but at least should
 165 * have the property that (x/y)*y + (x%y) = x.
 166 */
 167.align 3
 168.globl  sfunction
 169.ent    sfunction
 170sfunction:
 171        subq    $30,STACK,$30
 172        .frame  $30,STACK,$23
 173        .prologue 0
 174        bis     $24,$25,$28
 175        SLONGIFY($28)
 176        bge     $28,7b
 177        stq     $24,0($30)
 178        subq    $31,$24,$28
 179        stq     $25,8($30)
 180        cmovlt  $24,$28,$24     /* abs($24) */
 181        stq     $23,16($30)
 182        subq    $31,$25,$28
 183        stq     tmp1,24($30)
 184        cmovlt  $25,$28,$25     /* abs($25) */
 185        unop
 186        bsr     $23,ufunction
 187        ldq     $24,0($30)
 188        ldq     $25,8($30)
 189        GETSIGN($28)
 190        subq    $31,$27,tmp1
 191        SLONGIFY($28)
 192        ldq     $23,16($30)
 193        cmovlt  $28,tmp1,$27
 194        ldq     tmp1,24($30)
 195        addq    $30,STACK,$30
 196        ret     $31,($23),1
 197        .end    sfunction
 198EXPORT_SYMBOL(sfunction)
 199