root/arch/alpha/lib/ev6-divide.S

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /*
   3  * arch/alpha/lib/ev6-divide.S
   4  *
   5  * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
   6  *
   7  * Alpha division..
   8  */
   9 
  10 /*
  11  * The alpha chip doesn't provide hardware division, so we have to do it
  12  * by hand.  The compiler expects the functions
  13  *
  14  *      __divqu: 64-bit unsigned long divide
  15  *      __remqu: 64-bit unsigned long remainder
  16  *      __divqs/__remqs: signed 64-bit
  17  *      __divlu/__remlu: unsigned 32-bit
  18  *      __divls/__remls: signed 32-bit
  19  *
  20  * These are not normal C functions: instead of the normal
  21  * calling sequence, these expect their arguments in registers
  22  * $24 and $25, and return the result in $27. Register $28 may
  23  * be clobbered (assembly temporary), anything else must be saved. 
  24  *
  25  * In short: painful.
  26  *
  27  * This is a rather simple bit-at-a-time algorithm: it's very good
  28  * at dividing random 64-bit numbers, but the more usual case where
  29  * the divisor is small is handled better by the DEC algorithm
  30  * using lookup tables. This uses much less memory, though, and is
  31  * nicer on the cache.. Besides, I don't know the copyright status
  32  * of the DEC code.
  33  */
  34 
  35 /*
  36  * My temporaries:
  37  *      $0 - current bit
  38  *      $1 - shifted divisor
  39  *      $2 - modulus/quotient
  40  *
  41  *      $23 - return address
  42  *      $24 - dividend
  43  *      $25 - divisor
  44  *
  45  *      $27 - quotient/modulus
  46  *      $28 - compare status
  47  *
  48  * Much of the information about 21264 scheduling/coding comes from:
  49  *      Compiler Writer's Guide for the Alpha 21264
  50  *      abbreviated as 'CWG' in other comments here
  51  *      ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
  52  * Scheduling notation:
  53  *      E       - either cluster
  54  *      U       - upper subcluster; U0 - subcluster U0; U1 - subcluster U1
  55  *      L       - lower subcluster; L0 - subcluster L0; L1 - subcluster L1
  56  * Try not to change the actual algorithm if possible for consistency.
  57  */
  58 
  59 #include <asm/export.h>
  60 #define halt .long 0
  61 
  62 /*
  63  * Select function type and registers
  64  */
  65 #define mask    $0
  66 #define divisor $1
  67 #define compare $28
  68 #define tmp1    $3
  69 #define tmp2    $4
  70 
  71 #ifdef DIV
  72 #define DIV_ONLY(x,y...) x,##y
  73 #define MOD_ONLY(x,y...)
  74 #define func(x) __div##x
  75 #define modulus $2
  76 #define quotient $27
  77 #define GETSIGN(x) xor $24,$25,x
  78 #define STACK 48
  79 #else
  80 #define DIV_ONLY(x,y...)
  81 #define MOD_ONLY(x,y...) x,##y
  82 #define func(x) __rem##x
  83 #define modulus $27
  84 #define quotient $2
  85 #define GETSIGN(x) bis $24,$24,x
  86 #define STACK 32
  87 #endif
  88 
  89 /*
  90  * For 32-bit operations, we need to extend to 64-bit
  91  */
  92 #ifdef INTSIZE
  93 #define ufunction func(lu)
  94 #define sfunction func(l)
  95 #define LONGIFY(x) zapnot x,15,x
  96 #define SLONGIFY(x) addl x,0,x
  97 #else
  98 #define ufunction func(qu)
  99 #define sfunction func(q)
 100 #define LONGIFY(x)
 101 #define SLONGIFY(x)
 102 #endif
 103 
 104 .set noat
 105 .align  4
 106 .globl  ufunction
 107 .ent    ufunction
 108 ufunction:
 109         subq    $30,STACK,$30           # E :
 110         .frame  $30,STACK,$23
 111         .prologue 0
 112 
 113 7:      stq     $1, 0($30)              # L :
 114         bis     $25,$25,divisor         # E :
 115         stq     $2, 8($30)              # L : L U L U
 116 
 117         bis     $24,$24,modulus         # E :
 118         stq     $0,16($30)              # L :
 119         bis     $31,$31,quotient        # E :
 120         LONGIFY(divisor)                # E : U L L U
 121 
 122         stq     tmp1,24($30)            # L :
 123         LONGIFY(modulus)                # E :
 124         bis     $31,1,mask              # E :
 125         DIV_ONLY(stq tmp2,32($30))      # L : L U U L
 126 
 127         beq     divisor, 9f                     /* div by zero */
 128         /*
 129          * In spite of the DIV_ONLY being either a non-instruction
 130          * or an actual stq, the addition of the .align directive
 131          * below ensures that label 1 is going to be nicely aligned
 132          */
 133 
 134         .align  4
 135 #ifdef INTSIZE
 136         /*
 137          * shift divisor left, using 3-bit shifts for
 138          * 32-bit divides as we can't overflow. Three-bit
 139          * shifts will result in looping three times less
 140          * here, but can result in two loops more later.
 141          * Thus using a large shift isn't worth it (and
 142          * s8add pairs better than a sll..)
 143          */
 144 1:      cmpult  divisor,modulus,compare # E :
 145         s8addq  divisor,$31,divisor     # E :
 146         s8addq  mask,$31,mask           # E :
 147         bne     compare,1b              # U : U L U L
 148 #else
 149 1:      cmpult  divisor,modulus,compare # E :
 150         nop                             # E :
 151         nop                             # E :
 152         blt     divisor, 2f             # U : U L U L
 153 
 154         addq    divisor,divisor,divisor # E :
 155         addq    mask,mask,mask          # E :
 156         unop                            # E :
 157         bne     compare,1b              # U : U L U L
 158 #endif
 159 
 160         /* ok, start to go right again.. */
 161 2:
 162         /*
 163          * Keep things nicely bundled... use a nop instead of not
 164          * having an instruction for DIV_ONLY
 165          */
 166 #ifdef DIV
 167         DIV_ONLY(addq quotient,mask,tmp2) # E :
 168 #else
 169         nop                             # E :
 170 #endif
 171         srl     mask,1,mask             # U :
 172         cmpule  divisor,modulus,compare # E :
 173         subq    modulus,divisor,tmp1    # E :
 174 
 175 #ifdef DIV
 176         DIV_ONLY(cmovne compare,tmp2,quotient)  # E : Latency 2, extra map slot
 177         nop                             # E : as part of the cmovne
 178         srl     divisor,1,divisor       # U :
 179         nop                             # E : L U L U
 180 
 181         nop                             # E :
 182         cmovne  compare,tmp1,modulus    # E : Latency 2, extra map slot
 183         nop                             # E : as part of the cmovne
 184         bne     mask,2b                 # U : U L U L
 185 #else
 186         srl     divisor,1,divisor       # U :
 187         cmovne  compare,tmp1,modulus    # E : Latency 2, extra map slot
 188         nop                             # E : as part of the cmovne
 189         bne     mask,2b                 # U : U L L U
 190 #endif
 191 
 192 9:      ldq     $1, 0($30)              # L :
 193         ldq     $2, 8($30)              # L :
 194         nop                             # E :
 195         nop                             # E : U U L L
 196 
 197         ldq     $0,16($30)              # L :
 198         ldq     tmp1,24($30)            # L :
 199         nop                             # E :
 200         nop                             # E :
 201 
 202 #ifdef DIV
 203         DIV_ONLY(ldq tmp2,32($30))      # L :
 204 #else
 205         nop                             # E :
 206 #endif
 207         addq    $30,STACK,$30           # E :
 208         ret     $31,($23),1             # L0 : L U U L
 209         .end    ufunction
 210 EXPORT_SYMBOL(ufunction)
 211 
 212 /*
 213  * Uhh.. Ugly signed division. I'd rather not have it at all, but
 214  * it's needed in some circumstances. There are different ways to
 215  * handle this, really. This does:
 216  *      -a / b = a / -b = -(a / b)
 217  *      -a % b = -(a % b)
 218  *      a % -b = a % b
 219  * which is probably not the best solution, but at least should
 220  * have the property that (x/y)*y + (x%y) = x.
 221  */
 222 .align 4
 223 .globl  sfunction
 224 .ent    sfunction
 225 sfunction:
 226         subq    $30,STACK,$30           # E :
 227         .frame  $30,STACK,$23
 228         .prologue 0
 229         bis     $24,$25,$28             # E :
 230         SLONGIFY($28)                   # E :
 231         bge     $28,7b                  # U :
 232 
 233         stq     $24,0($30)              # L :
 234         subq    $31,$24,$28             # E :
 235         stq     $25,8($30)              # L :
 236         nop                             # E : U L U L
 237 
 238         cmovlt  $24,$28,$24     /* abs($24) */ # E : Latency 2, extra map slot
 239         nop                             # E : as part of the cmov
 240         stq     $23,16($30)             # L :
 241         subq    $31,$25,$28             # E : U L U L
 242 
 243         stq     tmp1,24($30)            # L :
 244         cmovlt  $25,$28,$25     /* abs($25) */ # E : Latency 2, extra map slot
 245         nop                             # E :
 246         bsr     $23,ufunction           # L0: L U L U
 247 
 248         ldq     $24,0($30)              # L :
 249         ldq     $25,8($30)              # L :
 250         GETSIGN($28)                    # E :
 251         subq    $31,$27,tmp1            # E : U U L L
 252 
 253         SLONGIFY($28)                   # E :
 254         ldq     $23,16($30)             # L :
 255         cmovlt  $28,tmp1,$27            # E : Latency 2, extra map slot
 256         nop                             # E : U L L U : as part of the cmov
 257 
 258         ldq     tmp1,24($30)            # L :
 259         nop                             # E : as part of the cmov
 260         addq    $30,STACK,$30           # E :
 261         ret     $31,($23),1             # L0 : L U U L
 262         .end    sfunction
 263 EXPORT_SYMBOL(sfunction)