target-arm/helper-a64.c

   1 /*
   2  *  AArch64 specific helpers
   3  *
   4  *  Copyright (c) 2013 Alexander Graf <agraf@suse.de>
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #include "qemu/osdep.h"
  21 #include "cpu.h"
  22 #include "exec/gdbstub.h"
  23 #include "exec/helper-proto.h"
  24 #include "qemu/host-utils.h"
  25 #include "sysemu/sysemu.h"
  26 #include "qemu/bitops.h"
  27 #include "internals.h"
  28 #include "qemu/crc32c.h"
  29 #include <zlib.h> /* For crc32 */
  30
  31 /* C2.4.7 Multiply and divide */
  32 /* special cases for 0 and LLONG_MIN are mandated by the standard */
  33 uint64_t HELPER(udiv64)(uint64_t num, uint64_t den)
  34 {
  35     if (den == 0) {
  36         return 0;
  37     }
  38     return num / den;
  39 }
  40
  41 int64_t HELPER(sdiv64)(int64_t num, int64_t den)
  42 {
  43     if (den == 0) {
  44         return 0;
  45     }
  46     if (num == LLONG_MIN && den == -1) {
  47         return LLONG_MIN;
  48     }
  49     return num / den;
  50 }
  51
  52 uint64_t HELPER(clz64)(uint64_t x)
  53 {
  54     return clz64(x);
  55 }
  56
  57 uint64_t HELPER(cls64)(uint64_t x)
  58 {
  59     return clrsb64(x);
  60 }
  61
  62 uint32_t HELPER(cls32)(uint32_t x)
  63 {
  64     return clrsb32(x);
  65 }
  66
  67 uint32_t HELPER(clz32)(uint32_t x)
  68 {
  69     return clz32(x);
  70 }
  71
  72 uint64_t HELPER(rbit64)(uint64_t x)
  73 {
  74     return revbit64(x);
  75 }
  76
  77 /* Convert a softfloat float_relation_ (as returned by
  78  * the float*_compare functions) to the correct ARM
  79  * NZCV flag state.
  80  */
  81 static inline uint32_t float_rel_to_flags(int res)
  82 {
  83     uint64_t flags;
  84     switch (res) {
  85     case float_relation_equal:
  86         flags = PSTATE_Z | PSTATE_C;
  87         break;
  88     case float_relation_less:
  89         flags = PSTATE_N;
  90         break;
  91     case float_relation_greater:
  92         flags = PSTATE_C;
  93         break;
  94     case float_relation_unordered:
  95     default:
  96         flags = PSTATE_C | PSTATE_V;
  97         break;
  98     }
  99     return flags;
 100 }
 101
 102 uint64_t HELPER(vfp_cmps_a64)(float32 x, float32 y, void *fp_status)
 103 {
 104     return float_rel_to_flags(float32_compare_quiet(x, y, fp_status));
 105 }
 106
 107 uint64_t HELPER(vfp_cmpes_a64)(float32 x, float32 y, void *fp_status)
 108 {
 109     return float_rel_to_flags(float32_compare(x, y, fp_status));
 110 }
 111
 112 uint64_t HELPER(vfp_cmpd_a64)(float64 x, float64 y, void *fp_status)
 113 {
 114     return float_rel_to_flags(float64_compare_quiet(x, y, fp_status));
 115 }
 116
 117 uint64_t HELPER(vfp_cmped_a64)(float64 x, float64 y, void *fp_status)
 118 {
 119     return float_rel_to_flags(float64_compare(x, y, fp_status));
 120 }
 121
 122 float32 HELPER(vfp_mulxs)(float32 a, float32 b, void *fpstp)
 123 {
 124     float_status *fpst = fpstp;
 125
 126     a = float32_squash_input_denormal(a, fpst);
 127     b = float32_squash_input_denormal(b, fpst);
 128
 129     if ((float32_is_zero(a) && float32_is_infinity(b)) ||
 130         (float32_is_infinity(a) && float32_is_zero(b))) {
 131         /* 2.0 with the sign bit set to sign(A) XOR sign(B) */
 132         return make_float32((1U << 30) |
 133                             ((float32_val(a) ^ float32_val(b)) & (1U << 31)));
 134     }
 135     return float32_mul(a, b, fpst);
 136 }
 137
 138 float64 HELPER(vfp_mulxd)(float64 a, float64 b, void *fpstp)
 139 {
 140     float_status *fpst = fpstp;
 141
 142     a = float64_squash_input_denormal(a, fpst);
 143     b = float64_squash_input_denormal(b, fpst);
 144
 145     if ((float64_is_zero(a) && float64_is_infinity(b)) ||
 146         (float64_is_infinity(a) && float64_is_zero(b))) {
 147         /* 2.0 with the sign bit set to sign(A) XOR sign(B) */
 148         return make_float64((1ULL << 62) |
 149                             ((float64_val(a) ^ float64_val(b)) & (1ULL << 63)));
 150     }
 151     return float64_mul(a, b, fpst);
 152 }
 153
 154 uint64_t HELPER(simd_tbl)(CPUARMState *env, uint64_t result, uint64_t indices,
 155                           uint32_t rn, uint32_t numregs)
 156 {
 157     /* Helper function for SIMD TBL and TBX. We have to do the table
 158      * lookup part for the 64 bits worth of indices we're passed in.
 159      * result is the initial results vector (either zeroes for TBL
 160      * or some guest values for TBX), rn the register number where
 161      * the table starts, and numregs the number of registers in the table.
 162      * We return the results of the lookups.
 163      */
 164     int shift;
 165
 166     for (shift = 0; shift < 64; shift += 8) {
 167         int index = extract64(indices, shift, 8);
 168         if (index < 16 * numregs) {
 169             /* Convert index (a byte offset into the virtual table
 170              * which is a series of 128-bit vectors concatenated)
 171              * into the correct vfp.regs[] element plus a bit offset
 172              * into that element, bearing in mind that the table
 173              * can wrap around from V31 to V0.
 174              */
 175             int elt = (rn * 2 + (index >> 3)) % 64;
 176             int bitidx = (index & 7) * 8;
 177             uint64_t val = extract64(env->vfp.regs[elt], bitidx, 8);
 178
 179             result = deposit64(result, shift, 8, val);
 180         }
 181     }
 182     return result;
 183 }
 184
 185 /* 64bit/double versions of the neon float compare functions */
 186 uint64_t HELPER(neon_ceq_f64)(float64 a, float64 b, void *fpstp)
 187 {
 188     float_status *fpst = fpstp;
 189     return -float64_eq_quiet(a, b, fpst);
 190 }
 191
 192 uint64_t HELPER(neon_cge_f64)(float64 a, float64 b, void *fpstp)
 193 {
 194     float_status *fpst = fpstp;
 195     return -float64_le(b, a, fpst);
 196 }
 197
 198 uint64_t HELPER(neon_cgt_f64)(float64 a, float64 b, void *fpstp)
 199 {
 200     float_status *fpst = fpstp;
 201     return -float64_lt(b, a, fpst);
 202 }
 203
 204 /* Reciprocal step and sqrt step. Note that unlike the A32/T32
 205  * versions, these do a fully fused multiply-add or
 206  * multiply-add-and-halve.
 207  */
 208 #define float32_two make_float32(0x40000000)
 209 #define float32_three make_float32(0x40400000)
 210 #define float32_one_point_five make_float32(0x3fc00000)
 211
 212 #define float64_two make_float64(0x4000000000000000ULL)
 213 #define float64_three make_float64(0x4008000000000000ULL)
 214 #define float64_one_point_five make_float64(0x3FF8000000000000ULL)
 215
 216 float32 HELPER(recpsf_f32)(float32 a, float32 b, void *fpstp)
 217 {
 218     float_status *fpst = fpstp;
 219
 220     a = float32_squash_input_denormal(a, fpst);
 221     b = float32_squash_input_denormal(b, fpst);
 222
 223     a = float32_chs(a);
 224     if ((float32_is_infinity(a) && float32_is_zero(b)) ||
 225         (float32_is_infinity(b) && float32_is_zero(a))) {
 226         return float32_two;
 227     }
 228     return float32_muladd(a, b, float32_two, 0, fpst);
 229 }
 230
 231 float64 HELPER(recpsf_f64)(float64 a, float64 b, void *fpstp)
 232 {
 233     float_status *fpst = fpstp;
 234
 235     a = float64_squash_input_denormal(a, fpst);
 236     b = float64_squash_input_denormal(b, fpst);
 237
 238     a = float64_chs(a);
 239     if ((float64_is_infinity(a) && float64_is_zero(b)) ||
 240         (float64_is_infinity(b) && float64_is_zero(a))) {
 241         return float64_two;
 242     }
 243     return float64_muladd(a, b, float64_two, 0, fpst);
 244 }
 245
 246 float32 HELPER(rsqrtsf_f32)(float32 a, float32 b, void *fpstp)
 247 {
 248     float_status *fpst = fpstp;
 249
 250     a = float32_squash_input_denormal(a, fpst);
 251     b = float32_squash_input_denormal(b, fpst);
 252
 253     a = float32_chs(a);
 254     if ((float32_is_infinity(a) && float32_is_zero(b)) ||
 255         (float32_is_infinity(b) && float32_is_zero(a))) {
 256         return float32_one_point_five;
 257     }
 258     return float32_muladd(a, b, float32_three, float_muladd_halve_result, fpst);
 259 }
 260
 261 float64 HELPER(rsqrtsf_f64)(float64 a, float64 b, void *fpstp)
 262 {
 263     float_status *fpst = fpstp;
 264
 265     a = float64_squash_input_denormal(a, fpst);
 266     b = float64_squash_input_denormal(b, fpst);
 267
 268     a = float64_chs(a);
 269     if ((float64_is_infinity(a) && float64_is_zero(b)) ||
 270         (float64_is_infinity(b) && float64_is_zero(a))) {
 271         return float64_one_point_five;
 272     }
 273     return float64_muladd(a, b, float64_three, float_muladd_halve_result, fpst);
 274 }
 275
 276 /* Pairwise long add: add pairs of adjacent elements into
 277  * double-width elements in the result (eg _s8 is an 8x8->16 op)
 278  */
 279 uint64_t HELPER(neon_addlp_s8)(uint64_t a)
 280 {
 281     uint64_t nsignmask = 0x0080008000800080ULL;
 282     uint64_t wsignmask = 0x8000800080008000ULL;
 283     uint64_t elementmask = 0x00ff00ff00ff00ffULL;
 284     uint64_t tmp1, tmp2;
 285     uint64_t res, signres;
 286
 287     /* Extract odd elements, sign extend each to a 16 bit field */
 288     tmp1 = a & elementmask;
 289     tmp1 ^= nsignmask;
 290     tmp1 |= wsignmask;
 291     tmp1 = (tmp1 - nsignmask) ^ wsignmask;
 292     /* Ditto for the even elements */
 293     tmp2 = (a >> 8) & elementmask;
 294     tmp2 ^= nsignmask;
 295     tmp2 |= wsignmask;
 296     tmp2 = (tmp2 - nsignmask) ^ wsignmask;
 297
 298     /* calculate the result by summing bits 0..14, 16..22, etc,
 299      * and then adjusting the sign bits 15, 23, etc manually.
 300      * This ensures the addition can't overflow the 16 bit field.
 301      */
 302     signres = (tmp1 ^ tmp2) & wsignmask;
 303     res = (tmp1 & ~wsignmask) + (tmp2 & ~wsignmask);
 304     res ^= signres;
 305
 306     return res;
 307 }
 308
 309 uint64_t HELPER(neon_addlp_u8)(uint64_t a)
 310 {
 311     uint64_t tmp;
 312
 313     tmp = a & 0x00ff00ff00ff00ffULL;
 314     tmp += (a >> 8) & 0x00ff00ff00ff00ffULL;
 315     return tmp;
 316 }
 317
 318 uint64_t HELPER(neon_addlp_s16)(uint64_t a)
 319 {
 320     int32_t reslo, reshi;
 321
 322     reslo = (int32_t)(int16_t)a + (int32_t)(int16_t)(a >> 16);
 323     reshi = (int32_t)(int16_t)(a >> 32) + (int32_t)(int16_t)(a >> 48);
 324
 325     return (uint32_t)reslo | (((uint64_t)reshi) << 32);
 326 }
 327
 328 uint64_t HELPER(neon_addlp_u16)(uint64_t a)
 329 {
 330     uint64_t tmp;
 331
 332     tmp = a & 0x0000ffff0000ffffULL;
 333     tmp += (a >> 16) & 0x0000ffff0000ffffULL;
 334     return tmp;
 335 }
 336
 337 /* Floating-point reciprocal exponent - see FPRecpX in ARM ARM */
 338 float32 HELPER(frecpx_f32)(float32 a, void *fpstp)
 339 {
 340     float_status *fpst = fpstp;
 341     uint32_t val32, sbit;
 342     int32_t exp;
 343
 344     if (float32_is_any_nan(a)) {
 345         float32 nan = a;
 346         if (float32_is_signaling_nan(a)) {
 347             float_raise(float_flag_invalid, fpst);
 348             nan = float32_maybe_silence_nan(a);
 349         }
 350         if (fpst->default_nan_mode) {
 351             nan = float32_default_nan;
 352         }
 353         return nan;
 354     }
 355
 356     val32 = float32_val(a);
 357     sbit = 0x80000000ULL & val32;
 358     exp = extract32(val32, 23, 8);
 359
 360     if (exp == 0) {
 361         return make_float32(sbit | (0xfe << 23));
 362     } else {
 363         return make_float32(sbit | (~exp & 0xff) << 23);
 364     }
 365 }
 366
 367 float64 HELPER(frecpx_f64)(float64 a, void *fpstp)
 368 {
 369     float_status *fpst = fpstp;
 370     uint64_t val64, sbit;
 371     int64_t exp;
 372
 373     if (float64_is_any_nan(a)) {
 374         float64 nan = a;
 375         if (float64_is_signaling_nan(a)) {
 376             float_raise(float_flag_invalid, fpst);
 377             nan = float64_maybe_silence_nan(a);
 378         }
 379         if (fpst->default_nan_mode) {
 380             nan = float64_default_nan;
 381         }
 382         return nan;
 383     }
 384
 385     val64 = float64_val(a);
 386     sbit = 0x8000000000000000ULL & val64;
 387     exp = extract64(float64_val(a), 52, 11);
 388
 389     if (exp == 0) {
 390         return make_float64(sbit | (0x7feULL << 52));
 391     } else {
 392         return make_float64(sbit | (~exp & 0x7ffULL) << 52);
 393     }
 394 }
 395
 396 float32 HELPER(fcvtx_f64_to_f32)(float64 a, CPUARMState *env)
 397 {
 398     /* Von Neumann rounding is implemented by using round-to-zero
 399      * and then setting the LSB of the result if Inexact was raised.
 400      */
 401     float32 r;
 402     float_status *fpst = &env->vfp.fp_status;
 403     float_status tstat = *fpst;
 404     int exflags;
 405
 406     set_float_rounding_mode(float_round_to_zero, &tstat);
 407     set_float_exception_flags(0, &tstat);
 408     r = float64_to_float32(a, &tstat);
 409     r = float32_maybe_silence_nan(r);
 410     exflags = get_float_exception_flags(&tstat);
 411     if (exflags & float_flag_inexact) {
 412         r = make_float32(float32_val(r) | 1);
 413     }
 414     exflags |= get_float_exception_flags(fpst);
 415     set_float_exception_flags(exflags, fpst);
 416     return r;
 417 }
 418
 419 /* 64-bit versions of the CRC helpers. Note that although the operation
 420  * (and the prototypes of crc32c() and crc32() mean that only the bottom
 421  * 32 bits of the accumulator and result are used, we pass and return
 422  * uint64_t for convenience of the generated code. Unlike the 32-bit
 423  * instruction set versions, val may genuinely have 64 bits of data in it.
 424  * The upper bytes of val (above the number specified by 'bytes') must have
 425  * been zeroed out by the caller.
 426  */
 427 uint64_t HELPER(crc32_64)(uint64_t acc, uint64_t val, uint32_t bytes)
 428 {
 429     uint8_t buf[8];
 430
 431     stq_le_p(buf, val);
 432
 433     /* zlib crc32 converts the accumulator and output to one's complement.  */
 434     return crc32(acc ^ 0xffffffff, buf, bytes) ^ 0xffffffff;
 435 }
 436
 437 uint64_t HELPER(crc32c_64)(uint64_t acc, uint64_t val, uint32_t bytes)
 438 {
 439     uint8_t buf[8];
 440
 441     stq_le_p(buf, val);
 442
 443     /* Linux crc32c converts the output to one's complement.  */
 444     return crc32c(acc, buf, bytes) ^ 0xffffffff;
 445 }