src/libcompiler_builtins/compiler-rt/lib/builtins/arm/udivsi3.S

   1 /*===-- udivsi3.S - 32-bit unsigned integer divide ------------------------===//
   2  *
   3  *                     The LLVM Compiler Infrastructure
   4  *
   5  * This file is dual licensed under the MIT and the University of Illinois Open
   6  * Source Licenses. See LICENSE.TXT for details.
   7  *
   8  *===----------------------------------------------------------------------===//
   9  *
  10  * This file implements the __udivsi3 (32-bit unsigned integer divide)
  11  * function for the ARM 32-bit architecture.
  12  *
  13  *===----------------------------------------------------------------------===*/
  14
  15 #include "../assembly.h"
  16
  17         .syntax unified
  18         .text
  19
  20 DEFINE_CODE_STATE
  21
  22         .p2align 2
  23 DEFINE_AEABI_FUNCTION_ALIAS(__aeabi_uidiv, __udivsi3)
  24
  25 @ unsigned int __udivsi3(unsigned int divident, unsigned int divisor)
  26 @   Calculate and return the quotient of the (unsigned) division.
  27
  28 DEFINE_COMPILERRT_FUNCTION(__udivsi3)
  29 #if __ARM_ARCH_EXT_IDIV__
  30         tst     r1, r1
  31         beq     LOCAL_LABEL(divby0)
  32         udiv    r0, r0, r1
  33         bx      lr
  34
  35 LOCAL_LABEL(divby0):
  36         mov     r0, #0
  37 #  ifdef __ARM_EABI__
  38         b       __aeabi_idiv0
  39 #  else
  40         JMP(lr)
  41 #  endif
  42
  43 #else /* ! __ARM_ARCH_EXT_IDIV__ */
  44         cmp     r1, #1
  45         bcc     LOCAL_LABEL(divby0)
  46 #if defined(USE_THUMB_1)
  47         bne LOCAL_LABEL(num_neq_denom)
  48         JMP(lr)
  49 LOCAL_LABEL(num_neq_denom):
  50 #else
  51         IT(eq)
  52         JMPc(lr, eq)
  53 #endif
  54         cmp     r0, r1
  55 #if defined(USE_THUMB_1)
  56         bhs LOCAL_LABEL(num_ge_denom)
  57         movs r0, #0
  58         JMP(lr)
  59 LOCAL_LABEL(num_ge_denom):
  60 #else
  61         ITT(cc)
  62         movcc   r0, #0
  63         JMPc(lr, cc)
  64 #endif
  65
  66         /*
  67          * Implement division using binary long division algorithm.
  68          *
  69          * r0 is the numerator, r1 the denominator.
  70          *
  71          * The code before JMP computes the correct shift I, so that
  72          * r0 and (r1 << I) have the highest bit set in the same position.
  73          * At the time of JMP, ip := .Ldiv0block - 12 * I.
  74          * This depends on the fixed instruction size of block.
  75          * For ARM mode, this is 12 Bytes, for THUMB mode 14 Bytes.
  76          *
  77          * block(shift) implements the test-and-update-quotient core.
  78          * It assumes (r0 << shift) can be computed without overflow and
  79          * that (r0 << shift) < 2 * r1. The quotient is stored in r3.
  80          */
  81
  82 #  if defined(__ARM_FEATURE_CLZ)
  83         clz     ip, r0
  84         clz     r3, r1
  85         /* r0 >= r1 implies clz(r0) <= clz(r1), so ip <= r3. */
  86         sub     r3, r3, ip
  87 #    if defined(USE_THUMB_2)
  88         adr     ip, LOCAL_LABEL(div0block) + 1
  89         sub     ip, ip, r3, lsl #1
  90 #    else
  91         adr     ip, LOCAL_LABEL(div0block)
  92 #    endif
  93         sub     ip, ip, r3, lsl #2
  94         sub     ip, ip, r3, lsl #3
  95         mov     r3, #0
  96         bx      ip
  97 #  else /* No CLZ Feature */
  98 #    if defined(USE_THUMB_2)
  99 #    error THUMB mode requires CLZ or UDIV
 100 #    endif
 101 #    if defined(USE_THUMB_1)
 102 #      define BLOCK_SIZE 10
 103 #    else
 104 #      define BLOCK_SIZE 12
 105 #    endif
 106
 107         mov     r2, r0
 108 #    if defined(USE_THUMB_1)
 109         mov ip, r0
 110         adr r0, LOCAL_LABEL(div0block)
 111         adds r0, #1
 112 #    else
 113         adr     ip, LOCAL_LABEL(div0block)
 114 #    endif
 115         lsrs    r3, r2, #16
 116         cmp     r3, r1
 117 #    if defined(USE_THUMB_1)
 118         blo LOCAL_LABEL(skip_16)
 119         movs r2, r3
 120         subs r0, r0, #(16 * BLOCK_SIZE)
 121 LOCAL_LABEL(skip_16):
 122 #    else
 123         movhs   r2, r3
 124         subhs   ip, ip, #(16 * BLOCK_SIZE)
 125 #    endif
 126
 127         lsrs    r3, r2, #8
 128         cmp     r3, r1
 129 #    if defined(USE_THUMB_1)
 130         blo LOCAL_LABEL(skip_8)
 131         movs r2, r3
 132         subs r0, r0, #(8 * BLOCK_SIZE)
 133 LOCAL_LABEL(skip_8):
 134 #    else
 135         movhs   r2, r3
 136         subhs   ip, ip, #(8 * BLOCK_SIZE)
 137 #    endif
 138
 139         lsrs    r3, r2, #4
 140         cmp     r3, r1
 141 #    if defined(USE_THUMB_1)
 142         blo LOCAL_LABEL(skip_4)
 143         movs r2, r3
 144         subs r0, r0, #(4 * BLOCK_SIZE)
 145 LOCAL_LABEL(skip_4):
 146 #    else
 147         movhs   r2, r3
 148         subhs   ip, #(4 * BLOCK_SIZE)
 149 #    endif
 150
 151         lsrs    r3, r2, #2
 152         cmp     r3, r1
 153 #    if defined(USE_THUMB_1)
 154         blo LOCAL_LABEL(skip_2)
 155         movs r2, r3
 156         subs r0, r0, #(2 * BLOCK_SIZE)
 157 LOCAL_LABEL(skip_2):
 158 #    else
 159         movhs   r2, r3
 160         subhs   ip, ip, #(2 * BLOCK_SIZE)
 161 #    endif
 162
 163         /* Last block, no need to update r2 or r3. */
 164 #    if defined(USE_THUMB_1)
 165         lsrs r3, r2, #1
 166         cmp r3, r1
 167         blo LOCAL_LABEL(skip_1)
 168         subs r0, r0, #(1 * BLOCK_SIZE)
 169 LOCAL_LABEL(skip_1):
 170         movs r2, r0
 171         mov r0, ip
 172         movs r3, #0
 173         JMP (r2)
 174
 175 #    else
 176         cmp     r1, r2, lsr #1
 177         subls   ip, ip, #(1 * BLOCK_SIZE)
 178
 179         movs    r3, #0
 180
 181         JMP(ip)
 182 #    endif
 183 #  endif /* __ARM_FEATURE_CLZ */
 184
 185
 186 #define IMM     #
 187         /* due to the range limit of branch in Thumb1, we have to place the
 188                  block closer */
 189 LOCAL_LABEL(divby0):
 190         movs    r0, #0
 191 #      if defined(__ARM_EABI__)
 192         push {r7, lr}
 193         bl      __aeabi_idiv0 // due to relocation limit, can't use b.
 194         pop  {r7, pc}
 195 #      else
 196         JMP(lr)
 197 #      endif
 198
 199
 200 #if defined(USE_THUMB_1)
 201 #define block(shift)                                                           \
 202         lsls r2, r1, IMM shift;                                                      \
 203         cmp r0, r2;                                                                  \
 204         blo LOCAL_LABEL(block_skip_##shift);                                         \
 205         subs r0, r0, r2;                                                             \
 206         LOCAL_LABEL(block_skip_##shift) :;                                           \
 207         adcs r3, r3 /* same as ((r3 << 1) | Carry). Carry is set if r0 >= r2. */
 208
 209         /* TODO: if current location counter is not not word aligned, we don't
 210                  need the .p2align and nop */
 211         /* Label div0block must be word-aligned. First align block 31 */
 212         .p2align 2
 213         nop /* Padding to align div0block as 31 blocks = 310 bytes */
 214
 215 #else
 216 #define block(shift)                                                           \
 217         cmp     r0, r1, lsl IMM shift;                                         \
 218         ITT(hs);                                                               \
 219         WIDE(addhs)     r3, r3, IMM (1 << shift);                              \
 220         WIDE(subhs)     r0, r0, r1, lsl IMM shift
 221 #endif
 222
 223         block(31)
 224         block(30)
 225         block(29)
 226         block(28)
 227         block(27)
 228         block(26)
 229         block(25)
 230         block(24)
 231         block(23)
 232         block(22)
 233         block(21)
 234         block(20)
 235         block(19)
 236         block(18)
 237         block(17)
 238         block(16)
 239         block(15)
 240         block(14)
 241         block(13)
 242         block(12)
 243         block(11)
 244         block(10)
 245         block(9)
 246         block(8)
 247         block(7)
 248         block(6)
 249         block(5)
 250         block(4)
 251         block(3)
 252         block(2)
 253         block(1)
 254 LOCAL_LABEL(div0block):
 255         block(0)
 256
 257         mov     r0, r3
 258         JMP(lr)
 259 #endif /* __ARM_ARCH_EXT_IDIV__ */
 260
 261 END_COMPILERRT_FUNCTION(__udivsi3)
 262
 263 NO_EXEC_STACK_DIRECTIVE
 264