deps/lightning/lib/jit_arm.c

   1 /*
   2  * Copyright (C) 2012-2019  Free Software Foundation, Inc.
   3  *
   4  * This file is part of GNU lightning.
   5  *
   6  * GNU lightning is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU Lesser General Public License as published
   8  * by the Free Software Foundation; either version 3, or (at your option)
   9  * any later version.
  10  *
  11  * GNU lightning is distributed in the hope that it will be useful, but
  12  * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  13  * or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  14  * License for more details.
  15  *
  16  * Authors:
  17  *      Paulo Cesar Pereira de Andrade
  18  */
  19
  20 #if defined(__linux__)
  21 #  include <stdio.h>
  22 #endif
  23
  24 #define jit_arg_reg_p(i)                ((i) >= 0 && (i) < 4)
  25 #define jit_arg_f_reg_p(i)              ((i) >= 0 && (i) < 16)
  26 #define jit_arg_d_reg_p(i)              ((i) >= 0 && (i) < 15)
  27
  28 #define arm_patch_node                  0x80000000
  29 #define arm_patch_word                  0x40000000
  30 #define arm_patch_jump                  0x20000000
  31 #define arm_patch_load                  0x00000000
  32
  33 #define jit_fpr_p(rn)                   ((rn) > 15)
  34
  35 #define arg_base()                                                      \
  36     (stack_framesize - 16 + (jit_cpu.abi ? 64 : 0))
  37 #define arg_offset(n)                                                   \
  38     ((n) < 4 ? arg_base() + ((n) << 2) : (n))
  39
  40 /* Assume functions called never match jit instruction set, that is
  41  * libc, gmp, mpfr, etc functions are in thumb mode and jit is in
  42  * arm mode, what may cause a crash upon return of that function
  43  * if generating jit for a relative jump.
  44  */
  45 #define jit_exchange_p()                1
  46
  47 /* FIXME is it really required to not touch _R10? */
  48
  49 /*
  50  * Types
  51  */
  52 typedef union _jit_thumb_t {
  53     jit_int32_t         i;
  54     jit_int16_t         s[2];
  55 } jit_thumb_t;
  56
  57 typedef jit_pointer_t   jit_va_list;
  58
  59 /*
  60  * Prototypes
  61  */
  62 #define jit_make_arg(node)              _jit_make_arg(_jit,node)
  63 static jit_node_t *_jit_make_arg(jit_state_t*,jit_node_t*);
  64 #define jit_make_arg_f(node)            _jit_make_arg_f(_jit,node)
  65 static jit_node_t *_jit_make_arg_f(jit_state_t*,jit_node_t*);
  66 #define jit_make_arg_d(node)            _jit_make_arg_d(_jit,node)
  67 static jit_node_t *_jit_make_arg_d(jit_state_t*,jit_node_t*);
  68 #define jit_get_reg_pair()              _jit_get_reg_pair(_jit)
  69 static jit_int32_t _jit_get_reg_pair(jit_state_t*);
  70 #define jit_unget_reg_pair(rn)          _jit_unget_reg_pair(_jit,rn)
  71 static void _jit_unget_reg_pair(jit_state_t*,jit_int32_t);
  72 # define must_align_p(node)             _must_align_p(_jit, node)
  73 static jit_bool_t _must_align_p(jit_state_t*,jit_node_t*);
  74 #define load_const(uniq,r0,i0)          _load_const(_jit,uniq,r0,i0)
  75 static void _load_const(jit_state_t*,jit_bool_t,jit_int32_t,jit_word_t);
  76 #define flush_consts()                  _flush_consts(_jit)
  77 static void _flush_consts(jit_state_t*);
  78 #define invalidate_consts()             _invalidate_consts(_jit)
  79 static void _invalidate_consts(jit_state_t*);
  80 #define patch(instr, node)              _patch(_jit, instr, node)
  81 static void _patch(jit_state_t*,jit_word_t,jit_node_t*);
  82
  83 #if defined(__GNUC__)
  84 /* libgcc */
  85 extern void __clear_cache(void *, void *);
  86 #endif
  87
  88 #define PROTO                           1
  89 #  include "jit_rewind.c"
  90 #  include "jit_arm-cpu.c"
  91 #  include "jit_arm-swf.c"
  92 #  include "jit_arm-vfp.c"
  93 #undef PROTO
  94
  95 /*
  96  * Initialization
  97  */
  98 jit_cpu_t               jit_cpu;
  99 jit_register_t          _rvs[] = {
 100     { rc(gpr) | 0x0c,                   "ip" },
 101     { rc(sav) | rc(gpr) | 0x04,         "r4" },
 102     { rc(sav) | rc(gpr) | 0x05,         "r5" },
 103     { rc(sav) | rc(gpr) | 0x06,         "r6" },
 104     { rc(sav) | rc(gpr) | 0x07,         "r7" },
 105     { rc(sav) | rc(gpr) | 0x08,         "r8" },
 106     { rc(sav) | rc(gpr) | 0x09,         "r9" },
 107     { rc(sav) | 0x0a,                   "sl" },
 108     { rc(sav) | 0x0b,                   "fp" },
 109     { rc(sav) | 0x0d,                   "sp" },
 110     { rc(sav) | 0x0e,                   "lr" },
 111     { 0x0f,                             "pc" },
 112     { rc(arg) | rc(gpr) | 0x03,         "r3" },
 113     { rc(arg) | rc(gpr) | 0x02,         "r2" },
 114     { rc(arg) | rc(gpr) | 0x01,         "r1" },
 115     { rc(arg) | rc(gpr) | 0x00,         "r0" },
 116     { rc(fpr) | 0x20,                   "d8" },
 117     { 0x21,                             "s17" },
 118     { rc(fpr) | 0x22,                   "d9" },
 119     { 0x23,                             "s19" },
 120     { rc(fpr) | 0x24,                   "d10" },
 121     { 0x25,                             "s21" },
 122     { rc(fpr) | 0x26,                   "d11" },
 123     { 0x27,                             "s23" },
 124     { rc(fpr) | 0x28,                   "d12" },
 125     { 0x29,                             "s25" },
 126     { rc(fpr) | 0x2a,                   "d13" },
 127     { 0x2b,                             "s27" },
 128     { rc(fpr) | 0x2c,                   "d14" },
 129     { 0x2d,                             "s29" },
 130     { rc(fpr) | 0x2e,                   "d15" },
 131     { 0x2f,                             "s31" },
 132     { rc(arg) | 0x1f,                   "s15" },
 133     { rc(arg)|rc(sft)|rc(fpr)|0x1e,     "d7" },
 134     { rc(arg) | 0x1d,                   "s13" },
 135     { rc(arg)|rc(sft)|rc(fpr)|0x1c,     "d6" },
 136     { rc(arg) | 0x1b,                   "s11" },
 137     { rc(arg)|rc(sft)|rc(fpr)|0x1a,     "d5" },
 138     { rc(arg) | 0x19,                   "s9" },
 139     { rc(arg)|rc(sft)|rc(fpr)|0x18,     "d4" },
 140     { rc(arg) | 0x17,                   "s7" },
 141     { rc(arg)|rc(sft)|rc(fpr)|0x16,     "d3" },
 142     { rc(arg) | 0x15,                   "s5" },
 143     { rc(arg)|rc(sft)|rc(fpr)|0x14,     "d2" },
 144     { rc(arg) | 0x13,                   "s3" },
 145     { rc(arg)|rc(sft)|rc(fpr)|0x12,     "d1" },
 146     { rc(arg) | 0x11,                   "s1" },
 147     { rc(arg)|rc(sft)|rc(fpr)|0x10,     "d0" },
 148     { _NOREG,                           "<none>" },
 149 };
 150
 151 /*
 152  * Implementation
 153  */
 154 void
 155 jit_get_cpu(void)
 156 {
 157 #if defined(__linux__)
 158     FILE        *fp;
 159     char        *ptr;
 160     char         buf[128];
 161
 162     if ((fp = fopen("/proc/cpuinfo", "r")) != NULL) {
 163         while (fgets(buf, sizeof(buf), fp)) {
 164             if (strncmp(buf, "CPU architecture:", 17) == 0) {
 165                 jit_cpu.version = strtol(buf + 17, &ptr, 10);
 166                 while (*ptr) {
 167                     if (*ptr == 'T' || *ptr == 't') {
 168                         ++ptr;
 169                         jit_cpu.thumb = 1;
 170                     }
 171                     else if (*ptr == 'E' || *ptr == 'e') {
 172                         jit_cpu.extend = 1;
 173                         ++ptr;
 174                     }
 175                     else
 176                         ++ptr;
 177                 }
 178             }
 179             else if (strncmp(buf, "Features\t:", 10) == 0) {
 180                 if ((ptr = strstr(buf + 10, "vfpv")))
 181                     jit_cpu.vfp = strtol(ptr + 4, NULL, 0);
 182                 if ((ptr = strstr(buf + 10, "neon")))
 183                     jit_cpu.neon = 1;
 184                 if ((ptr = strstr(buf + 10, "thumb")))
 185                     jit_cpu.thumb = 1;
 186             }
 187         }
 188         fclose(fp);
 189     }
 190 #endif
 191 #if defined(__ARM_PCS_VFP)
 192     if (!jit_cpu.vfp)
 193         jit_cpu.vfp = 3;
 194     if (!jit_cpu.version)
 195         jit_cpu.version = 7;
 196     jit_cpu.abi = 1;
 197 #endif
 198 #if defined(__thumb2__)
 199     jit_cpu.thumb = 1;
 200 #endif
 201     /* armv6t2 todo (software float and thumb2) */
 202     if (!jit_cpu.vfp && jit_cpu.thumb)
 203         jit_cpu.thumb = 0;
 204 }
 205
 206 void
 207 _jit_init(jit_state_t *_jit)
 208 {
 209     jit_int32_t         regno;
 210     static jit_bool_t   first = 1;
 211
 212     _jitc->reglen = jit_size(_rvs) - 1;
 213     if (first) {
 214         /* jit_get_cpu() should have been already called, and only once */
 215         if (!jit_cpu.vfp) {
 216             /* cause register to never be allocated, because simple
 217              * software float only allocates stack space for 8 slots  */
 218             for (regno = _D8; regno < _D7; regno++)
 219                 _rvs[regno].spec = 0;
 220         }
 221         if (!jit_cpu.abi) {
 222             for (regno = _S15; regno <= _D0; regno++)
 223                 _rvs[regno].spec &= ~rc(arg);
 224         }
 225         first = 0;
 226     }
 227 }
 228
 229 void
 230 _jit_prolog(jit_state_t *_jit)
 231 {
 232     jit_int32_t          offset;
 233
 234     if (_jitc->function)
 235         jit_epilog();
 236     assert(jit_regset_cmp_ui(&_jitc->regarg, 0) == 0);
 237     jit_regset_set_ui(&_jitc->regsav, 0);
 238     offset = _jitc->functions.offset;
 239     if (offset >= _jitc->functions.length) {
 240         jit_realloc((jit_pointer_t *)&_jitc->functions.ptr,
 241                     _jitc->functions.length * sizeof(jit_function_t),
 242                     (_jitc->functions.length + 16) * sizeof(jit_function_t));
 243         _jitc->functions.length += 16;
 244     }
 245     _jitc->function = _jitc->functions.ptr + _jitc->functions.offset++;
 246     _jitc->function->self.size = stack_framesize;
 247     if (jit_cpu.abi)
 248         _jitc->function->self.size += 64;
 249     _jitc->function->self.argi = _jitc->function->self.argf =
 250         _jitc->function->self.alen = 0;
 251     if (jit_swf_p())
 252         /* 8 soft float registers */
 253         _jitc->function->self.aoff = -64;
 254     else
 255         _jitc->function->self.aoff = 0;
 256     _jitc->function->self.call = jit_call_default;
 257     jit_alloc((jit_pointer_t *)&_jitc->function->regoff,
 258               _jitc->reglen * sizeof(jit_int32_t));
 259
 260     /* _no_link here does not mean the jit_link() call can be removed
 261      * by rewriting as:
 262      * _jitc->function->prolog = jit_new_node(jit_code_prolog);
 263      */
 264     _jitc->function->prolog = jit_new_node_no_link(jit_code_prolog);
 265     jit_link(_jitc->function->prolog);
 266     _jitc->function->prolog->w.w = offset;
 267     _jitc->function->epilog = jit_new_node_no_link(jit_code_epilog);
 268     /*  u:      label value
 269      *  v:      offset in blocks vector
 270      *  w:      offset in functions vector
 271      */
 272     _jitc->function->epilog->w.w = offset;
 273
 274     jit_regset_new(&_jitc->function->regset);
 275 }
 276
 277 jit_int32_t
 278 _jit_allocai(jit_state_t *_jit, jit_int32_t length)
 279 {
 280     assert(_jitc->function);
 281     switch (length) {
 282         case 0: case 1:                                         break;
 283         case 2:         _jitc->function->self.aoff &= -2;       break;
 284         case 3: case 4: _jitc->function->self.aoff &= -4;       break;
 285         default:        _jitc->function->self.aoff &= -8;       break;
 286     }
 287     _jitc->function->self.aoff -= length;
 288     if (!_jitc->realize) {
 289         jit_inc_synth_ww(allocai, _jitc->function->self.aoff, length);
 290         jit_dec_synth();
 291     }
 292     return (_jitc->function->self.aoff);
 293 }
 294
 295 void
 296 _jit_allocar(jit_state_t *_jit, jit_int32_t u, jit_int32_t v)
 297 {
 298     jit_int32_t          reg;
 299     assert(_jitc->function);
 300     jit_inc_synth_ww(allocar, u, v);
 301     if (!_jitc->function->allocar) {
 302         _jitc->function->aoffoff = jit_allocai(sizeof(jit_int32_t));
 303         _jitc->function->allocar = 1;
 304     }
 305     reg = jit_get_reg(jit_class_gpr);
 306     jit_negr(reg, v);
 307     jit_andi(reg, reg, -8);
 308     jit_ldxi_i(u, JIT_FP, _jitc->function->aoffoff);
 309     jit_addr(u, u, reg);
 310     jit_addr(JIT_SP, JIT_SP, reg);
 311     jit_stxi_i(_jitc->function->aoffoff, JIT_FP, u);
 312     jit_unget_reg(reg);
 313     jit_dec_synth();
 314 }
 315
 316 void
 317 _jit_ret(jit_state_t *_jit)
 318 {
 319     jit_node_t          *instr;
 320     assert(_jitc->function);
 321     jit_inc_synth(ret);
 322     /* jump to epilog */
 323     instr = jit_jmpi();
 324     jit_patch_at(instr, _jitc->function->epilog);
 325     jit_dec_synth();
 326 }
 327
 328 void
 329 _jit_retr(jit_state_t *_jit, jit_int32_t u)
 330 {
 331     jit_inc_synth_w(retr, u);
 332     if (JIT_RET != u)
 333         jit_movr(JIT_RET, u);
 334     jit_live(JIT_RET);
 335     jit_ret();
 336     jit_dec_synth();
 337 }
 338
 339 void
 340 _jit_reti(jit_state_t *_jit, jit_word_t u)
 341 {
 342     jit_inc_synth_w(reti, u);
 343     jit_movi(JIT_RET, u);
 344     jit_ret();
 345     jit_dec_synth();
 346 }
 347
 348 void
 349 _jit_retr_f(jit_state_t *_jit, jit_int32_t u)
 350 {
 351     jit_inc_synth_w(retr_f, u);
 352     if (jit_cpu.abi) {
 353         if (u != JIT_FRET)
 354             jit_movr_f(JIT_FRET, u);
 355         else
 356             jit_live(JIT_FRET);
 357     }
 358     else {
 359         if (u != JIT_RET)
 360             jit_movr_f_w(JIT_RET, u);
 361         else
 362             jit_live(JIT_RET);
 363     }
 364     jit_ret();
 365     jit_dec_synth();
 366 }
 367
 368 void
 369 _jit_reti_f(jit_state_t *_jit, jit_float32_t u)
 370 {
 371     jit_inc_synth_f(reti_f, u);
 372     if (jit_cpu.abi)
 373         jit_movi_f(JIT_FRET, u);
 374     else
 375         jit_movi_f_w(JIT_RET, u);
 376     jit_ret();
 377     jit_dec_synth();
 378 }
 379
 380 void
 381 _jit_retr_d(jit_state_t *_jit, jit_int32_t u)
 382 {
 383     jit_inc_synth_w(retr_d, u);
 384     if (jit_cpu.abi) {
 385         if (u != JIT_FRET)
 386             jit_movr_d(JIT_FRET, u);
 387         else
 388             jit_live(JIT_FRET);
 389     }
 390     else {
 391         if (u != JIT_RET)
 392             jit_movr_d_ww(JIT_RET, _R1, u);
 393         else
 394             jit_live(JIT_RET);
 395     }
 396     jit_ret();
 397     jit_dec_synth();
 398 }
 399
 400 void
 401 _jit_reti_d(jit_state_t *_jit, jit_float64_t u)
 402 {
 403     jit_inc_synth_d(reti_d, u);
 404     if (jit_cpu.abi)
 405         jit_movi_d(JIT_FRET, u);
 406     else
 407         jit_movi_d_ww(JIT_RET, _R1, u);
 408     jit_ret();
 409     jit_dec_synth();
 410 }
 411
 412 void
 413 _jit_epilog(jit_state_t *_jit)
 414 {
 415     assert(_jitc->function);
 416     assert(_jitc->function->epilog->next == NULL);
 417     jit_link(_jitc->function->epilog);
 418     _jitc->function = NULL;
 419 }
 420
 421 jit_bool_t
 422 _jit_arg_register_p(jit_state_t *_jit, jit_node_t *u)
 423 {
 424     if (u->code != jit_code_arg) {
 425         if (u->code == jit_code_arg_f) {
 426             if (jit_cpu.abi)
 427                 return (jit_arg_f_reg_p(u->u.w));
 428         }
 429         else {
 430             assert(u->code == jit_code_arg_d);
 431             if (jit_cpu.abi)
 432                 return (jit_arg_d_reg_p(u->u.w));
 433         }
 434     }
 435     return (jit_arg_reg_p(u->u.w));
 436 }
 437
 438 static jit_node_t *
 439 _jit_make_arg(jit_state_t *_jit, jit_node_t *node)
 440 {
 441     jit_int32_t          offset;
 442     if (jit_arg_reg_p(_jitc->function->self.argi))
 443         offset = _jitc->function->self.argi++;
 444     else {
 445         offset = _jitc->function->self.size;
 446         _jitc->function->self.size += sizeof(jit_word_t);
 447     }
 448     if (node == (jit_node_t *)0)
 449         node = jit_new_node(jit_code_arg);
 450     else
 451         link_node(node);
 452     node->u.w = offset;
 453     node->v.w = ++_jitc->function->self.argn;
 454     jit_link_prolog();
 455     return (node);
 456 }
 457
 458 jit_node_t *
 459 _jit_make_arg_f(jit_state_t *_jit, jit_node_t *node)
 460 {
 461     jit_int32_t          offset;
 462     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 463         if (jit_arg_f_reg_p(_jitc->function->self.argf)) {
 464             offset = _jitc->function->self.argf++;
 465             goto done;
 466         }
 467     }
 468     else {
 469         if (jit_arg_reg_p(_jitc->function->self.argi)) {
 470             offset = _jitc->function->self.argi++;
 471             goto done;
 472         }
 473     }
 474     offset = _jitc->function->self.size;
 475     _jitc->function->self.size += sizeof(jit_float32_t);
 476 done:
 477     if (node == (jit_node_t *)0)
 478         node = jit_new_node(jit_code_arg_f);
 479     else
 480         link_node(node);
 481     node->u.w = offset;
 482     node->v.w = ++_jitc->function->self.argn;
 483     jit_link_prolog();
 484     return (node);
 485 }
 486
 487 jit_node_t *
 488 _jit_make_arg_d(jit_state_t *_jit, jit_node_t *node)
 489 {
 490     jit_int32_t          offset;
 491     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 492         if (jit_arg_d_reg_p(_jitc->function->self.argf)) {
 493             if (_jitc->function->self.argf & 1)
 494                 ++_jitc->function->self.argf;
 495             offset = _jitc->function->self.argf;
 496             _jitc->function->self.argf += 2;
 497             goto done;
 498         }
 499     }
 500     else {
 501         if (_jitc->function->self.argi & 1)
 502             ++_jitc->function->self.argi;
 503         if (jit_arg_reg_p(_jitc->function->self.argi)) {
 504             offset = _jitc->function->self.argi;
 505             _jitc->function->self.argi += 2;
 506             goto done;
 507         }
 508     }
 509     if (_jitc->function->self.size & 7)
 510         _jitc->function->self.size += 4;
 511     offset = _jitc->function->self.size;
 512     _jitc->function->self.size += sizeof(jit_float64_t);
 513 done:
 514     if (node == (jit_node_t *)0)
 515         node = jit_new_node(jit_code_arg_d);
 516     else
 517         link_node(node);
 518     node->u.w = offset;
 519     node->v.w = ++_jitc->function->self.argn;
 520     jit_link_prolog();
 521     return (node);
 522 }
 523
 524 void
 525 _jit_ellipsis(jit_state_t *_jit)
 526 {
 527     if (_jitc->prepare) {
 528         assert(!(_jitc->function->call.call & jit_call_varargs));
 529         _jitc->function->call.call |= jit_call_varargs;
 530         if (jit_cpu.abi && _jitc->function->call.argf)
 531             rewind_prepare();
 532     }
 533     else {
 534         assert(!(_jitc->function->self.call & jit_call_varargs));
 535         _jitc->function->self.call |= jit_call_varargs;
 536         if (jit_cpu.abi &&  _jitc->function->self.argf)
 537             rewind_prolog();
 538         /* First 4 stack addresses are always spilled r0-r3 */
 539         if (jit_arg_reg_p(_jitc->function->self.argi))
 540             _jitc->function->vagp = _jitc->function->self.argi * 4;
 541         else
 542             _jitc->function->vagp = 16;
 543     }
 544     jit_inc_synth(ellipsis);
 545     if (_jitc->prepare)
 546         jit_link_prepare();
 547     else
 548         jit_link_prolog();
 549     jit_dec_synth();
 550 }
 551
 552 void
 553 _jit_va_push(jit_state_t *_jit, jit_int32_t u)
 554 {
 555     jit_inc_synth_w(va_push, u);
 556     jit_pushargr(u);
 557     jit_dec_synth();
 558 }
 559
 560 jit_node_t *
 561 _jit_arg(jit_state_t *_jit)
 562 {
 563     assert(_jitc->function);
 564     return (jit_make_arg((jit_node_t*)0));
 565 }
 566
 567 jit_node_t *
 568 _jit_arg_f(jit_state_t *_jit)
 569 {
 570     assert(_jitc->function);
 571     return (jit_make_arg_f((jit_node_t*)0));
 572 }
 573
 574 jit_node_t *
 575 _jit_arg_d(jit_state_t *_jit)
 576 {
 577     assert(_jitc->function);
 578     return (jit_make_arg_d((jit_node_t*)0));
 579 }
 580
 581 void
 582 _jit_getarg_c(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 583 {
 584     assert(v->code == jit_code_arg);
 585     jit_inc_synth_wp(getarg_c, u, v);
 586     if (jit_swf_p())
 587         jit_ldxi_c(u, JIT_FP, arg_offset(v->u.w));
 588     else if (jit_arg_reg_p(v->u.w))
 589         jit_extr_c(u, JIT_RA0 - v->u.w);
 590     else
 591         jit_ldxi_c(u, JIT_FP, v->u.w);
 592     jit_dec_synth();
 593 }
 594
 595 void
 596 _jit_getarg_uc(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 597 {
 598     assert(v->code == jit_code_arg);
 599     jit_inc_synth_wp(getarg_uc, u, v);
 600     if (jit_swf_p())
 601         jit_ldxi_uc(u, JIT_FP, arg_offset(v->u.w));
 602     else if (jit_arg_reg_p(v->u.w))
 603         jit_extr_uc(u, JIT_RA0 - v->u.w);
 604     else
 605         jit_ldxi_uc(u, JIT_FP, v->u.w);
 606     jit_dec_synth();
 607 }
 608
 609 void
 610 _jit_getarg_s(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 611 {
 612     assert(v->code == jit_code_arg);
 613     jit_inc_synth_wp(getarg_s, u, v);
 614     if (jit_swf_p())
 615         jit_ldxi_s(u, JIT_FP, arg_offset(v->u.w));
 616     else if (jit_arg_reg_p(v->u.w))
 617         jit_extr_s(u, JIT_RA0 - v->u.w);
 618     else
 619         jit_ldxi_s(u, JIT_FP, v->u.w);
 620     jit_dec_synth();
 621 }
 622
 623 void
 624 _jit_getarg_us(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 625 {
 626     assert(v->code == jit_code_arg);
 627     jit_inc_synth_wp(getarg_us, u, v);
 628     if (jit_swf_p())
 629         jit_ldxi_us(u, JIT_FP, arg_offset(v->u.w));
 630     else if (jit_arg_reg_p(v->u.w))
 631         jit_extr_us(u, JIT_RA0 - v->u.w);
 632     else
 633         jit_ldxi_us(u, JIT_FP, v->u.w);
 634     jit_dec_synth();
 635 }
 636
 637 void
 638 _jit_getarg_i(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 639 {
 640     assert(v->code == jit_code_arg);
 641     jit_inc_synth_wp(getarg_i, u, v);
 642     if (jit_swf_p())
 643         jit_ldxi_i(u, JIT_FP, arg_offset(v->u.w));
 644     else if (jit_arg_reg_p(v->u.w))
 645         jit_movr(u, JIT_RA0 - v->u.w);
 646     else
 647         jit_ldxi_i(u, JIT_FP, v->u.w);
 648     jit_dec_synth();
 649 }
 650
 651 void
 652 _jit_putargr(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 653 {
 654     assert(v->code == jit_code_arg);
 655     jit_inc_synth_wp(putargr, u, v);
 656     if (jit_swf_p())
 657         jit_stxi(arg_offset(v->u.w), JIT_FP, u);
 658     else if (jit_arg_reg_p(v->u.w))
 659         jit_movr(JIT_RA0 - v->u.w, u);
 660     else
 661         jit_stxi(v->u.w, JIT_FP, u);
 662     jit_dec_synth();
 663 }
 664
 665 void
 666 _jit_putargi(jit_state_t *_jit, jit_word_t u, jit_node_t *v)
 667 {
 668     jit_int32_t         regno;
 669     assert(v->code == jit_code_arg);
 670     jit_inc_synth_wp(putargi, u, v);
 671     if (jit_swf_p()) {
 672         regno = jit_get_reg(jit_class_gpr);
 673         jit_movi(regno, u);
 674         jit_stxi(arg_offset(v->u.w), JIT_FP, regno);
 675         jit_unget_reg(regno);
 676     }
 677     else if (jit_arg_reg_p(v->u.w))
 678         jit_movi(JIT_RA0 - v->u.w, u);
 679     else {
 680         regno = jit_get_reg(jit_class_gpr);
 681         jit_movi(regno, u);
 682         jit_stxi(v->u.w, JIT_FP, regno);
 683         jit_unget_reg(regno);
 684     }
 685     jit_dec_synth();
 686 }
 687
 688 void
 689 _jit_getarg_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 690 {
 691     assert(v->code == jit_code_arg_f);
 692     jit_inc_synth_wp(getarg_f, u, v);
 693     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 694         if (jit_arg_f_reg_p(v->u.w))
 695             jit_movr_f(u, JIT_FA0 - v->u.w);
 696         else
 697             jit_ldxi_f(u, JIT_FP, v->u.w);
 698     }
 699     else if (jit_swf_p())
 700         jit_ldxi_f(u, JIT_FP, arg_offset(v->u.w));
 701     else {
 702         if (jit_arg_reg_p(v->u.w))
 703             jit_movr_w_f(u, JIT_RA0 - v->u.w);
 704         else
 705             jit_ldxi_f(u, JIT_FP, v->u.w);
 706     }
 707     jit_dec_synth();
 708 }
 709
 710 void
 711 _jit_putargr_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 712 {
 713     assert(v->code == jit_code_arg_f);
 714     jit_inc_synth_wp(putargr_f, u, v);
 715     if (jit_cpu.abi) {
 716         if (jit_arg_f_reg_p(v->u.w))
 717             jit_movr_f(JIT_FA0 - v->u.w, u);
 718         else
 719             jit_stxi_f(v->u.w, JIT_FP, u);
 720     }
 721     else if (jit_swf_p())
 722         jit_stxi_f(arg_offset(v->u.w), JIT_FP, u);
 723     else {
 724         if (jit_arg_reg_p(v->u.w))
 725             jit_movr_f_w(JIT_RA0 - v->u.w, u);
 726         else
 727             jit_stxi_f(v->u.w, JIT_FP, u);
 728     }
 729     jit_dec_synth();
 730 }
 731
 732 void
 733 _jit_putargi_f(jit_state_t *_jit, jit_float32_t u, jit_node_t *v)
 734 {
 735     jit_int32_t         regno;
 736     assert(v->code == jit_code_arg_f);
 737     jit_inc_synth_fp(putargi_f, u, v);
 738     if (jit_cpu.abi) {
 739         if (jit_arg_f_reg_p(v->u.w))
 740             jit_movi_f(JIT_FA0 - v->u.w, u);
 741         else {
 742             regno = jit_get_reg(jit_class_fpr);
 743             jit_movi_f(regno, u);
 744             jit_stxi_f(v->u.w, JIT_FP, regno);
 745             jit_unget_reg(regno);
 746         }
 747     }
 748     else if (jit_swf_p()) {
 749         regno = jit_get_reg(jit_class_fpr);
 750         jit_movi_f(regno, u);
 751         jit_stxi_f(arg_offset(v->u.w), JIT_FP, regno);
 752         jit_unget_reg(regno);
 753     }
 754     else {
 755         regno = jit_get_reg(jit_class_fpr);
 756         jit_movi_f(regno, u);
 757         if (jit_arg_reg_p(v->u.w))
 758             jit_movr_f_w(JIT_RA0 - v->u.w, regno);
 759         else
 760             jit_stxi_f(v->u.w, JIT_FP, regno);
 761         jit_unget_reg(regno);
 762     }
 763     jit_dec_synth();
 764 }
 765
 766 void
 767 _jit_getarg_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 768 {
 769     assert(v->code == jit_code_arg_d);
 770     jit_inc_synth_wp(getarg_d, u, v);
 771     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 772         if (jit_arg_f_reg_p(v->u.w))
 773             jit_movr_d(u, JIT_FA0 - v->u.w);
 774         else
 775             jit_ldxi_d(u, JIT_FP, v->u.w);
 776     }
 777     else if (jit_swf_p())
 778         jit_ldxi_d(u, JIT_FP, arg_offset(v->u.w));
 779     else {
 780         if (jit_arg_reg_p(v->u.w))
 781             jit_movr_ww_d(u, JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1));
 782         else
 783             jit_ldxi_d(u, JIT_FP, v->u.w);
 784     }
 785     jit_dec_synth();
 786 }
 787
 788 void
 789 _jit_putargr_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 790 {
 791     assert(v->code == jit_code_arg_d);
 792     jit_inc_synth_wp(putargr_d, u, v);
 793     if (jit_cpu.abi) {
 794         if (jit_arg_f_reg_p(v->u.w))
 795             jit_movr_d(JIT_FA0 - v->u.w, u);
 796         else
 797             jit_stxi_d(v->u.w, JIT_FP, u);
 798     }
 799     else if (jit_swf_p())
 800         jit_stxi_d(arg_offset(v->u.w), JIT_FP, u);
 801     else {
 802         if (jit_arg_reg_p(v->u.w))
 803             jit_movr_d_ww(JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1), u);
 804         else
 805             jit_stxi_d(v->u.w, JIT_FP, u);
 806     }
 807     jit_dec_synth();
 808 }
 809
 810 void
 811 _jit_putargi_d(jit_state_t *_jit, jit_float64_t u, jit_node_t *v)
 812 {
 813     jit_int32_t         regno;
 814     assert(v->code == jit_code_arg_d);
 815     jit_inc_synth_dp(putargi_d, u, v);
 816     if (jit_cpu.abi) {
 817         if (jit_arg_f_reg_p(v->u.w))
 818             jit_movi_d(JIT_FA0 - v->u.w, u);
 819         else {
 820             regno = jit_get_reg(jit_class_fpr);
 821             jit_movi_d(regno, u);
 822             jit_stxi_d(v->u.w, JIT_FP, regno);
 823             jit_unget_reg(regno);
 824         }
 825     }
 826     else if (jit_swf_p()) {
 827         regno = jit_get_reg(jit_class_fpr);
 828         jit_movi_d(regno, u);
 829         jit_stxi_d(arg_offset(v->u.w), JIT_FP, regno);
 830         jit_unget_reg(regno);
 831     }
 832     else {
 833         regno = jit_get_reg(jit_class_fpr);
 834         jit_movi_d(regno, u);
 835         if (jit_arg_reg_p(v->u.w))
 836             jit_movr_d_ww(JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1), regno);
 837         else
 838             jit_stxi_d(v->u.w, JIT_FP, regno);
 839         jit_unget_reg(regno);
 840     }
 841     jit_dec_synth();
 842 }
 843
 844 void
 845 _jit_pushargr(jit_state_t *_jit, jit_int32_t u)
 846 {
 847     assert(_jitc->function);
 848     jit_inc_synth_w(pushargr, u);
 849     jit_link_prepare();
 850     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 851         jit_movr(JIT_RA0 - _jitc->function->call.argi, u);
 852         ++_jitc->function->call.argi;
 853     }
 854     else {
 855         jit_stxi(_jitc->function->call.size, JIT_SP, u);
 856         _jitc->function->call.size += sizeof(jit_word_t);
 857     }
 858     jit_dec_synth();
 859 }
 860
 861 void
 862 _jit_pushargi(jit_state_t *_jit, jit_word_t u)
 863 {
 864     jit_int32_t          regno;
 865     assert(_jitc->function);
 866     jit_inc_synth_w(pushargi, u);
 867     jit_link_prepare();
 868     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 869         jit_movi(JIT_RA0 - _jitc->function->call.argi, u);
 870         ++_jitc->function->call.argi;
 871     }
 872     else {
 873         regno = jit_get_reg(jit_class_gpr);
 874         jit_movi(regno, u);
 875         jit_stxi(_jitc->function->call.size, JIT_SP, regno);
 876         jit_unget_reg(regno);
 877         _jitc->function->call.size += sizeof(jit_word_t);
 878     }
 879     jit_dec_synth();
 880 }
 881
 882 void
 883 _jit_pushargr_f(jit_state_t *_jit, jit_int32_t u)
 884 {
 885     assert(_jitc->function);
 886     jit_inc_synth_w(pushargr_f, u);
 887     jit_link_prepare();
 888     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
 889         if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 890             jit_movr_f(JIT_FA0 - _jitc->function->call.argf, u);
 891             ++_jitc->function->call.argf;
 892             goto done;
 893         }
 894     }
 895     else {
 896         if (jit_arg_reg_p(_jitc->function->call.argi)) {
 897             jit_movr_f_w(JIT_RA0 - _jitc->function->call.argi, u);
 898             ++_jitc->function->call.argi;
 899             goto done;
 900         }
 901     }
 902     jit_stxi_f(_jitc->function->call.size, JIT_SP, u);
 903     _jitc->function->call.size += sizeof(jit_word_t);
 904 done:
 905     jit_dec_synth();
 906 }
 907
 908 void
 909 _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
 910 {
 911     jit_int32_t         regno;
 912     assert(_jitc->function);
 913     jit_inc_synth_f(pushargi_f, u);
 914     jit_link_prepare();
 915     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
 916         if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 917             /* cannot jit_movi_f in the argument register because
 918              * float arguments are packed, and that would cause
 919              * either an assertion in debug mode, or overwritting
 920              * two registers */
 921             regno = jit_get_reg(jit_class_fpr);
 922             jit_movi_f(regno, u);
 923             jit_movr_f(JIT_FA0 - _jitc->function->call.argf, regno);
 924             jit_unget_reg(regno);
 925             ++_jitc->function->call.argf;
 926             goto done;
 927         }
 928     }
 929     else {
 930         if (jit_arg_reg_p(_jitc->function->call.argi)) {
 931             jit_movi_f_w(JIT_RA0 - _jitc->function->call.argi, u);
 932             ++_jitc->function->call.argi;
 933             goto done;
 934         }
 935     }
 936     regno = jit_get_reg(jit_class_fpr);
 937     jit_movi_f(regno, u);
 938     jit_stxi_f(_jitc->function->call.size, JIT_SP, regno);
 939     jit_unget_reg(regno);
 940     _jitc->function->call.size += sizeof(jit_word_t);
 941 done:
 942     jit_dec_synth();
 943 }
 944
 945 void
 946 _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
 947 {
 948     assert(_jitc->function);
 949     jit_inc_synth_w(pushargr_d, u);
 950     jit_link_prepare();
 951     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
 952         if (jit_arg_d_reg_p(_jitc->function->call.argf)) {
 953             if (_jitc->function->call.argf & 1)
 954                 ++_jitc->function->call.argf;
 955             jit_movr_d(JIT_FA0 - _jitc->function->call.argf, u);
 956             _jitc->function->call.argf += 2;
 957             goto done;
 958         }
 959     }
 960     else {
 961         if (_jitc->function->call.argi & 1)
 962             ++_jitc->function->call.argi;
 963         if (jit_arg_reg_p(_jitc->function->call.argi)) {
 964             jit_movr_d_ww(JIT_RA0 - _jitc->function->call.argi,
 965                           JIT_RA0 - (_jitc->function->call.argi + 1),
 966                           u);
 967             _jitc->function->call.argi += 2;
 968             goto done;
 969         }
 970     }
 971     if (_jitc->function->call.size & 7)
 972         _jitc->function->call.size += 4;
 973     jit_stxi_d(_jitc->function->call.size, JIT_SP, u);
 974     _jitc->function->call.size += sizeof(jit_float64_t);
 975 done:
 976     jit_dec_synth();
 977 }
 978
 979 void
 980 _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
 981 {
 982     jit_int32_t         regno;
 983     assert(_jitc->function);
 984     jit_inc_synth_d(pushargi_d, u);
 985     jit_link_prepare();
 986     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
 987         if (jit_arg_d_reg_p(_jitc->function->call.argf)) {
 988             if (_jitc->function->call.argf & 1)
 989                 ++_jitc->function->call.argf;
 990             jit_movi_d(JIT_FA0 - _jitc->function->call.argf, u);
 991             _jitc->function->call.argf += 2;
 992             goto done;
 993         }
 994     }
 995     else {
 996         if (_jitc->function->call.argi & 1)
 997             ++_jitc->function->call.argi;
 998         if (jit_arg_reg_p(_jitc->function->call.argi)) {
 999             jit_movi_d_ww(JIT_RA0 - _jitc->function->call.argi,
1000                           JIT_RA0 - (_jitc->function->call.argi + 1),
1001                           u);
1002             _jitc->function->call.argi += 2;
1003             goto done;
1004         }
1005     }
1006     if (_jitc->function->call.size & 7)
1007         _jitc->function->call.size += 4;
1008     regno = jit_get_reg(jit_class_fpr);
1009     jit_movi_d(regno, u);
1010     jit_stxi_d(_jitc->function->call.size, JIT_SP, regno);
1011     jit_unget_reg(regno);
1012     _jitc->function->call.size += sizeof(jit_float64_t);
1013 done:
1014     jit_dec_synth();
1015 }
1016
1017 jit_bool_t
1018 _jit_regarg_p(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno)
1019 {
1020     jit_int32_t         spec;
1021     spec = jit_class(_rvs[regno].spec);
1022     if (spec & jit_class_arg) {
1023         regno = JIT_RA0 - regno;
1024         if (regno >= 0 && regno < node->v.w)
1025             return (1);
1026         if (jit_cpu.abi && spec & jit_class_fpr) {
1027             regno = JIT_FA0 - regno;
1028             if (regno >= 0 && regno < node->w.w)
1029                 return (1);
1030         }
1031     }
1032
1033     return (0);
1034 }
1035
1036 void
1037 _jit_finishr(jit_state_t *_jit, jit_int32_t r0)
1038 {
1039     jit_node_t          *node;
1040     assert(_jitc->function);
1041     jit_inc_synth_w(finishr, r0);
1042     if (_jitc->function->self.alen < _jitc->function->call.size)
1043         _jitc->function->self.alen = _jitc->function->call.size;
1044     node = jit_callr(r0);
1045     node->v.w = _jitc->function->self.argi;
1046     node->w.w = _jitc->function->call.argf;
1047     _jitc->function->call.argi = _jitc->function->call.argf =
1048         _jitc->function->call.size = 0;
1049     _jitc->prepare = 0;
1050     jit_dec_synth();
1051 }
1052
1053 jit_node_t *
1054 _jit_finishi(jit_state_t *_jit, jit_pointer_t i0)
1055 {
1056     jit_node_t          *node;
1057     assert(_jitc->function);
1058     jit_inc_synth_w(finishi, (jit_word_t)i0);
1059     if (_jitc->function->self.alen < _jitc->function->call.size)
1060         _jitc->function->self.alen = _jitc->function->call.size;
1061     node = jit_calli(i0);
1062     node->v.w = _jitc->function->call.argi;
1063     node->w.w = _jitc->function->call.argf;
1064     _jitc->function->call.argi = _jitc->function->call.argf =
1065         _jitc->function->call.size = 0;
1066     _jitc->prepare = 0;
1067     jit_dec_synth();
1068     return (node);
1069 }
1070
1071 void
1072 _jit_retval_c(jit_state_t *_jit, jit_int32_t r0)
1073 {
1074     jit_inc_synth_w(retval_c, r0);
1075     jit_extr_c(r0, JIT_RET);
1076     jit_dec_synth();
1077 }
1078
1079 void
1080 _jit_retval_uc(jit_state_t *_jit, jit_int32_t r0)
1081 {
1082     jit_inc_synth_w(retval_uc, r0);
1083     jit_extr_uc(r0, JIT_RET);
1084     jit_dec_synth();
1085 }
1086
1087 void
1088 _jit_retval_s(jit_state_t *_jit, jit_int32_t r0)
1089 {
1090     jit_inc_synth_w(retval_s, r0);
1091     jit_extr_s(r0, JIT_RET);
1092     jit_dec_synth();
1093 }
1094
1095 void
1096 _jit_retval_us(jit_state_t *_jit, jit_int32_t r0)
1097 {
1098     jit_inc_synth_w(retval_us, r0);
1099     jit_extr_us(r0, JIT_RET);
1100     jit_dec_synth();
1101 }
1102
1103 void
1104 _jit_retval_i(jit_state_t *_jit, jit_int32_t r0)
1105 {
1106     jit_inc_synth_w(retval_i, r0);
1107     if (r0 != JIT_RET)
1108         jit_movr(r0, JIT_RET);
1109     jit_dec_synth();
1110 }
1111
1112 void
1113 _jit_retval_f(jit_state_t *_jit, jit_int32_t r0)
1114 {
1115     jit_inc_synth_w(retval_f, r0);
1116     if (jit_cpu.abi) {
1117         if (r0 != JIT_FRET)
1118             jit_movr_f(r0, JIT_FRET);
1119     }
1120     else if (r0 != JIT_RET)
1121         jit_movr_w_f(r0, JIT_RET);
1122     jit_dec_synth();
1123 }
1124
1125 void
1126 _jit_retval_d(jit_state_t *_jit, jit_int32_t r0)
1127 {
1128     jit_inc_synth_w(retval_d, r0);
1129     if (jit_cpu.abi) {
1130         if (r0 != JIT_FRET)
1131             jit_movr_d(r0, JIT_FRET);
1132     }
1133     else if (r0 != JIT_RET)
1134         jit_movr_ww_d(r0, JIT_RET, _R1);
1135     jit_dec_synth();
1136 }
1137
1138 jit_pointer_t
1139 _emit_code(jit_state_t *_jit)
1140 {
1141     jit_node_t          *node;
1142     jit_node_t          *temp;
1143     jit_word_t           word;
1144     jit_int32_t          value;
1145     jit_int32_t          offset;
1146     struct {
1147         jit_node_t      *node;
1148         jit_uint8_t     *data;
1149         jit_word_t       word;
1150 #if DEVEL_DISASSEMBLER
1151         jit_word_t       prevw;
1152 #endif
1153         jit_uword_t      thumb;
1154 #if DISASSEMBLER
1155         jit_int32_t      info_offset;
1156 #endif
1157         jit_int32_t      const_offset;
1158         jit_int32_t      patch_offset;
1159     } undo;
1160 #if DEVEL_DISASSEMBLER
1161     jit_word_t           prevw;
1162 #endif
1163
1164     _jitc->function = NULL;
1165     _jitc->thumb = 0;
1166
1167     jit_reglive_setup();
1168
1169     _jitc->consts.data = NULL;
1170     _jitc->consts.offset = _jitc->consts.length = 0;
1171
1172     undo.word = 0;
1173     undo.node = NULL;
1174     undo.data = NULL;
1175     undo.thumb = 0;
1176 #if DISASSEMBLER
1177     undo.info_offset =
1178 #endif
1179         undo.const_offset = undo.patch_offset = 0;
1180 #  define assert_data(node)             /**/
1181 #define case_rr(name, type)                                             \
1182             case jit_code_##name##r##type:                              \
1183                 name##r##type(rn(node->u.w), rn(node->v.w));            \
1184                 break
1185 #define case_rw(name, type)                                             \
1186             case jit_code_##name##i##type:                              \
1187                 name##i##type(rn(node->u.w), node->v.w);                \
1188                 break
1189 #define case_vv(name, type)                                             \
1190             case jit_code_##name##r##type:                              \
1191                 if (jit_swf_p())                                        \
1192                     swf_##name##r##type(rn(node->u.w), rn(node->v.w));  \
1193                 else                                                    \
1194                     vfp_##name##r##type(rn(node->u.w), rn(node->v.w));  \
1195                 break
1196 #define case_vw(name, type)                                             \
1197             case jit_code_##name##i##type:                              \
1198                 if (jit_swf_p())                                        \
1199                     swf_##name##i##type(rn(node->u.w), node->v.w);      \
1200                 else                                                    \
1201                     vfp_##name##i##type(rn(node->u.w), node->v.w);      \
1202                 break
1203 #define case_wr(name, type)                                             \
1204             case jit_code_##name##i##type:                              \
1205                 name##i##type(node->u.w, rn(node->v.w));                \
1206                 break
1207 #define case_wv(name, type)                                             \
1208             case jit_code_##name##i##type:                              \
1209                 if (jit_swf_p())                                        \
1210                     swf_##name##i##type(node->u.w, rn(node->v.w));      \
1211                 else                                                    \
1212                     vfp_##name##i##type(node->u.w, rn(node->v.w));      \
1213                 break
1214 #define case_rrr(name, type)                                            \
1215             case jit_code_##name##r##type:                              \
1216                 name##r##type(rn(node->u.w),                            \
1217                               rn(node->v.w), rn(node->w.w));            \
1218                 break
1219 #define case_rrrr(name, type)                                           \
1220             case jit_code_##name##r##type:                              \
1221                 name##r##type(rn(node->u.q.l), rn(node->u.q.h),         \
1222                               rn(node->v.w), rn(node->w.w));            \
1223                 break
1224 #define case_vvv(name, type)                                            \
1225             case jit_code_##name##r##type:                              \
1226                 if (jit_swf_p())                                        \
1227                     swf_##name##r##type(rn(node->u.w),                  \
1228                                         rn(node->v.w), rn(node->w.w));  \
1229                 else                                                    \
1230                     vfp_##name##r##type(rn(node->u.w),                  \
1231                                         rn(node->v.w), rn(node->w.w));  \
1232                 break
1233 #define case_rrw(name, type)                                            \
1234             case jit_code_##name##i##type:                              \
1235                 name##i##type(rn(node->u.w), rn(node->v.w), node->w.w); \
1236                 break
1237 #define case_rrrw(name, type)                                           \
1238             case jit_code_##name##i##type:                              \
1239                 name##i##type(rn(node->u.q.l), rn(node->u.q.h),         \
1240                               rn(node->v.w), node->w.w);                \
1241                 break
1242 #define case_vvw(name, type)                                            \
1243             case jit_code_##name##i##type:                              \
1244                 if (jit_swf_p())                                        \
1245                     swf_##name##i##type(rn(node->u.w),                  \
1246                                         rn(node->v.w), node->w.w);      \
1247                 else                                                    \
1248                     vfp_##name##i##type(rn(node->u.w),                  \
1249                                         rn(node->v.w), node->w.w);      \
1250                 break
1251 #define case_vvf(name)                                                  \
1252             case jit_code_##name##i_f:                                  \
1253                 assert_data(node);                                      \
1254                 if (jit_swf_p())                                        \
1255                     swf_##name##i_f(rn(node->u.w), rn(node->v.w),       \
1256                                     node->w.f);                         \
1257                 else                                                    \
1258                     vfp_##name##i_f(rn(node->u.w), rn(node->v.w),       \
1259                                     node->w.f);                         \
1260                 break
1261 #define case_vvd(name)                                                  \
1262             case jit_code_##name##i_d:                                  \
1263                 assert_data(node);                                      \
1264                 if (jit_swf_p())                                        \
1265                     swf_##name##i_d(rn(node->u.w), rn(node->v.w),       \
1266                                     node->w.d);                         \
1267                 else                                                    \
1268                     vfp_##name##i_d(rn(node->u.w), rn(node->v.w),       \
1269                                     node->w.d);                         \
1270                 break
1271 #define case_wrr(name, type)                                            \
1272             case jit_code_##name##i##type:                              \
1273                 name##i##type(node->u.w, rn(node->v.w), rn(node->w.w)); \
1274                 break
1275 #define case_wvv(name, type)                                            \
1276             case jit_code_##name##i##type:                              \
1277                 if (jit_swf_p())                                        \
1278                     swf_##name##i##type(node->u.w,                      \
1279                                         rn(node->v.w), rn(node->w.w));  \
1280                 else                                                    \
1281                     vfp_##name##i##type(node->u.w,                      \
1282                                         rn(node->v.w), rn(node->w.w));  \
1283                 break
1284 #define case_brr(name, type)                                            \
1285             case jit_code_##name##r##type:                              \
1286                 temp = node->u.n;                                       \
1287                 assert(temp->code == jit_code_label ||                  \
1288                        temp->code == jit_code_epilog);                  \
1289                 if (temp->flag & jit_flag_patch)                        \
1290                     name##r##type(temp->u.w, rn(node->v.w),             \
1291                                   rn(node->w.w));                       \
1292                 else {                                                  \
1293                     word = name##r##type(_jit->pc.w,                    \
1294                                          rn(node->v.w), rn(node->w.w)); \
1295                     patch(word, node);                                  \
1296                 }                                                       \
1297                 break
1298 #define case_bvv(name, type)                                            \
1299             case jit_code_##name##r##type:                              \
1300                 temp = node->u.n;                                       \
1301                 assert(temp->code == jit_code_label ||                  \
1302                        temp->code == jit_code_epilog);                  \
1303                 if (temp->flag & jit_flag_patch) {                      \
1304                     if (jit_swf_p())                                    \
1305                         swf_##name##r##type(temp->u.w, rn(node->v.w),   \
1306                                             rn(node->w.w));             \
1307                     else                                                \
1308                         vfp_##name##r##type(temp->u.w, rn(node->v.w),   \
1309                                             rn(node->w.w));             \
1310                 }                                                       \
1311                 else {                                                  \
1312                     if (jit_swf_p())                                    \
1313                         word = swf_##name##r##type(_jit->pc.w,          \
1314                                                    rn(node->v.w),       \
1315                                                    rn(node->w.w));      \
1316                     else                                                \
1317                         word = vfp_##name##r##type(_jit->pc.w,          \
1318                                                    rn(node->v.w),       \
1319                                                    rn(node->w.w));      \
1320                     patch(word, node);                                  \
1321                 }                                                       \
1322                 break
1323 #define case_brw(name, type)                                            \
1324             case jit_code_##name##i##type:                              \
1325                 temp = node->u.n;                                       \
1326                 assert(temp->code == jit_code_label ||                  \
1327                        temp->code == jit_code_epilog);                  \
1328                 if (temp->flag & jit_flag_patch)                        \
1329                     name##i##type(temp->u.w,                            \
1330                                   rn(node->v.w), node->w.w);            \
1331                 else {                                                  \
1332                     word = name##i##type(_jit->pc.w,                    \
1333                                          rn(node->v.w), node->w.w);     \
1334                     patch(word, node);                                  \
1335                 }                                                       \
1336                 break;
1337 #define case_bvf(name)                                                  \
1338             case jit_code_##name##i_f:                                  \
1339                 temp = node->u.n;                                       \
1340                 assert(temp->code == jit_code_label ||                  \
1341                        temp->code == jit_code_epilog);                  \
1342                 if (temp->flag & jit_flag_patch) {                      \
1343                     if (jit_swf_p())                                    \
1344                         swf_##name##i_f(temp->u.w, rn(node->v.w),       \
1345                                         node->w.f);                     \
1346                     else                                                \
1347                         vfp_##name##i_f(temp->u.w, rn(node->v.w),       \
1348                                         node->w.f);                     \
1349                 }                                                       \
1350                 else {                                                  \
1351                     if (jit_swf_p())                                    \
1352                         word = swf_##name##i_f(_jit->pc.w,              \
1353                                                rn(node->v.w),           \
1354                                                node->w.f);              \
1355                     else                                                \
1356                         word = vfp_##name##i_f(_jit->pc.w,              \
1357                                                rn(node->v.w),           \
1358                                                node->w.f);              \
1359                     patch(word, node);                                  \
1360                 }                                                       \
1361                 break
1362 #define case_bvd(name)                                                  \
1363             case jit_code_##name##i_d:                                  \
1364                 temp = node->u.n;                                       \
1365                 assert(temp->code == jit_code_label ||                  \
1366                        temp->code == jit_code_epilog);                  \
1367                 if (temp->flag & jit_flag_patch) {                      \
1368                     if (jit_swf_p())                                    \
1369                         swf_##name##i_d(temp->u.w, rn(node->v.w),       \
1370                                         node->w.d);                     \
1371                     else                                                \
1372                         vfp_##name##i_d(temp->u.w, rn(node->v.w),       \
1373                                         node->w.d);                     \
1374                 }                                                       \
1375                 else {                                                  \
1376                     if (jit_swf_p())                                    \
1377                         word = swf_##name##i_d(_jit->pc.w,              \
1378                                                rn(node->v.w),           \
1379                                                node->w.d);              \
1380                     else                                                \
1381                         word = vfp_##name##i_d(_jit->pc.w,              \
1382                                                rn(node->v.w),           \
1383                                                node->w.d);              \
1384                     patch(word, node);                                  \
1385                 }                                                       \
1386                 break
1387 #if DEVEL_DISASSEMBLER
1388     prevw = _jit->pc.w;
1389 #endif
1390     for (node = _jitc->head; node; node = node->next) {
1391         if (_jit->pc.uc >= _jitc->code.end)
1392             return (NULL);
1393
1394 #if DEVEL_DISASSEMBLER
1395         node->offset = (jit_uword_t)_jit->pc.w - (jit_uword_t)prevw;
1396         prevw = _jit->pc.w;
1397 #endif
1398         value = jit_classify(node->code);
1399         jit_regarg_set(node, value);
1400         switch (node->code) {
1401             case jit_code_align:
1402                 assert(!(node->u.w & (node->u.w - 1)) &&
1403                        node->u.w <= sizeof(jit_word_t));
1404                 if (node->u.w == sizeof(jit_word_t) &&
1405                     (word = _jit->pc.w & (sizeof(jit_word_t) - 1)))
1406                     nop(sizeof(jit_word_t) - word);
1407                 break;
1408             case jit_code_note:         case jit_code_name:
1409                 if (must_align_p(node->next))
1410                     nop(2);
1411                 node->u.w = _jit->pc.w;
1412                 break;
1413             case jit_code_label:
1414                 if (must_align_p(node->next))
1415                     nop(2);
1416                 /* remember label is defined */
1417                 node->flag |= jit_flag_patch;
1418                 node->u.w = _jit->pc.w;
1419                 break;
1420                 case_rrr(add,);
1421                 case_rrw(add,);
1422                 case_rrr(addc,);
1423                 case_rrw(addc,);
1424                 case_rrr(addx,);
1425                 case_rrw(addx,);
1426                 case_rrr(sub,);
1427                 case_rrw(sub,);
1428                 case_rrr(subc,);
1429                 case_rrw(subc,);
1430                 case_rrr(subx,);
1431                 case_rrw(subx,);
1432                 case_rrw(rsb,);
1433                 case_rrr(mul,);
1434                 case_rrw(mul,);
1435                 case_rrrr(qmul,);
1436                 case_rrrw(qmul,);
1437                 case_rrrr(qmul, _u);
1438                 case_rrrw(qmul, _u);
1439                 case_rrr(div,);
1440                 case_rrw(div,);
1441                 case_rrr(div, _u);
1442                 case_rrw(div, _u);
1443                 case_rrrr(qdiv,);
1444                 case_rrrw(qdiv,);
1445                 case_rrrr(qdiv, _u);
1446                 case_rrrw(qdiv, _u);
1447                 case_rrr(rem,);
1448                 case_rrw(rem,);
1449                 case_rrr(rem, _u);
1450                 case_rrw(rem, _u);
1451                 case_rrr(lsh,);
1452                 case_rrw(lsh,);
1453                 case_rrr(rsh,);
1454                 case_rrw(rsh,);
1455                 case_rrr(rsh, _u);
1456                 case_rrw(rsh, _u);
1457                 case_rr(neg,);
1458                 case_rr(com,);
1459                 case_rrr(and,);
1460                 case_rrw(and,);
1461                 case_rrr(or,);
1462                 case_rrw(or,);
1463                 case_rrr(xor,);
1464                 case_rrw(xor,);
1465                 case_vv(trunc, _f_i);
1466                 case_vv(trunc, _d_i);
1467                 case_rr(ld, _c);
1468                 case_rw(ld, _c);
1469                 case_rr(ld, _uc);
1470                 case_rw(ld, _uc);
1471                 case_rr(ld, _s);
1472                 case_rw(ld, _s);
1473                 case_rr(ld, _us);
1474                 case_rw(ld, _us);
1475                 case_rr(ld, _i);
1476                 case_rw(ld, _i);
1477                 case_rrr(ldx, _c);
1478                 case_rrw(ldx, _c);
1479                 case_rrr(ldx, _uc);
1480                 case_rrw(ldx, _uc);
1481                 case_rrr(ldx, _s);
1482                 case_rrw(ldx, _s);
1483                 case_rrr(ldx, _us);
1484                 case_rrw(ldx, _us);
1485                 case_rrr(ldx, _i);
1486                 case_rrw(ldx, _i);
1487                 case_rr(st, _c);
1488                 case_wr(st, _c);
1489                 case_rr(st, _s);
1490                 case_wr(st, _s);
1491                 case_rr(st, _i);
1492                 case_wr(st, _i);
1493                 case_rrr(stx, _c);
1494                 case_wrr(stx, _c);
1495                 case_rrr(stx, _s);
1496                 case_wrr(stx, _s);
1497                 case_rrr(stx, _i);
1498                 case_wrr(stx, _i);
1499                 case_rr(hton, _us);
1500                 case_rr(hton, _ui);
1501                 case_rr(bswap, _us);
1502                 case_rr(bswap, _ui);
1503                 case_rr(ext, _c);
1504                 case_rr(ext, _uc);
1505                 case_rr(ext, _s);
1506                 case_rr(ext, _us);
1507                 case_rr(mov,);
1508                 case_rrr(movn,);
1509                 case_rrr(movz,);
1510             case jit_code_movi:
1511                 if (node->flag & jit_flag_node) {
1512                     temp = node->v.n;
1513                     if (temp->code == jit_code_data ||
1514                         (temp->code == jit_code_label &&
1515                          (temp->flag & jit_flag_patch)))
1516                         movi(rn(node->u.w), temp->u.w);
1517                     else {
1518                         assert(temp->code == jit_code_label ||
1519                                temp->code == jit_code_epilog);
1520                         word = movi_p(rn(node->u.w), temp->u.w);
1521                         patch(word, node);
1522                     }
1523                 }
1524                 else
1525                     movi(rn(node->u.w), node->v.w);
1526                 break;
1527                 case_rrr(lt,);
1528                 case_rrw(lt,);
1529                 case_rrr(lt, _u);
1530                 case_rrw(lt, _u);
1531                 case_rrr(le,);
1532                 case_rrw(le,);
1533                 case_rrr(le, _u);
1534                 case_rrw(le, _u);
1535                 case_rrr(eq,);
1536                 case_rrw(eq,);
1537                 case_rrr(ge,);
1538                 case_rrw(ge,);
1539                 case_rrr(ge, _u);
1540                 case_rrw(ge, _u);
1541                 case_rrr(gt,);
1542                 case_rrw(gt,);
1543                 case_rrr(gt, _u);
1544                 case_rrw(gt, _u);
1545                 case_rrr(ne,);
1546                 case_rrw(ne,);
1547                 case_brr(blt,);
1548                 case_brw(blt,);
1549                 case_brr(blt, _u);
1550                 case_brw(blt, _u);
1551                 case_brr(ble,);
1552                 case_brw(ble,);
1553                 case_brr(ble, _u);
1554                 case_brw(ble, _u);
1555                 case_brr(beq,);
1556                 case_brw(beq,);
1557                 case_brr(bge,);
1558                 case_brw(bge,);
1559                 case_brr(bge, _u);
1560                 case_brw(bge, _u);
1561                 case_brr(bgt,);
1562                 case_brw(bgt,);
1563                 case_brr(bgt, _u);
1564                 case_brw(bgt, _u);
1565                 case_brr(bne,);
1566                 case_brw(bne,);
1567                 case_brr(boadd,);
1568                 case_brw(boadd,);
1569                 case_brr(boadd, _u);
1570                 case_brw(boadd, _u);
1571                 case_brr(bxadd,);
1572                 case_brw(bxadd,);
1573                 case_brr(bxadd, _u);
1574                 case_brw(bxadd, _u);
1575                 case_brr(bosub,);
1576                 case_brw(bosub,);
1577                 case_brr(bosub, _u);
1578                 case_brw(bosub, _u);
1579                 case_brr(bxsub,);
1580                 case_brw(bxsub,);
1581                 case_brr(bxsub, _u);
1582                 case_brw(bxsub, _u);
1583                 case_brr(bms,);
1584                 case_brw(bms,);
1585                 case_brr(bmc,);
1586                 case_brw(bmc,);
1587                 case_vvv(add, _f);
1588                 case_vvf(add);
1589                 case_vvv(sub, _f);
1590                 case_vvf(sub);
1591                 case_vvf(rsb);
1592                 case_vvv(mul, _f);
1593                 case_vvf(mul);
1594                 case_vvv(div, _f);
1595                 case_vvf(div);
1596                 case_vv(abs, _f);
1597                 case_vv(neg, _f);
1598                 case_vv(sqrt, _f);
1599                 case_vv(ext, _f);
1600                 case_vv(ld, _f);
1601                 case_vw(ld, _f);
1602                 case_vvv(ldx, _f);
1603                 case_vvw(ldx, _f);
1604                 case_vv(st, _f);
1605                 case_wv(st, _f);
1606                 case_vvv(stx, _f);
1607                 case_wvv(stx, _f);
1608                 case_vv(mov, _f);
1609             case jit_code_movi_f:
1610                 assert_data(node);
1611                 if (jit_swf_p())
1612                     swf_movi_f(rn(node->u.w), node->v.f);
1613                 else
1614                     vfp_movi_f(rn(node->u.w), node->v.f);
1615                 break;
1616                 case_vv(ext, _d_f);
1617                 case_vvv(lt, _f);
1618                 case_vvf(lt);
1619                 case_vvv(le, _f);
1620                 case_vvf(le);
1621                 case_vvv(eq, _f);
1622                 case_vvf(eq);
1623                 case_vvv(ge, _f);
1624                 case_vvf(ge);
1625                 case_vvv(gt, _f);
1626                 case_vvf(gt);
1627                 case_vvv(ne, _f);
1628                 case_vvf(ne);
1629                 case_vvv(unlt, _f);
1630                 case_vvf(unlt);
1631                 case_vvv(unle, _f);
1632                 case_vvf(unle);
1633                 case_vvv(uneq, _f);
1634                 case_vvf(uneq);
1635                 case_vvv(unge, _f);
1636                 case_vvf(unge);
1637                 case_vvv(ungt, _f);
1638                 case_vvf(ungt);
1639                 case_vvv(ltgt, _f);
1640                 case_vvf(ltgt);
1641                 case_vvv(ord, _f);
1642                 case_vvf(ord);
1643                 case_vvv(unord, _f);
1644                 case_vvf(unord);
1645                 case_bvv(blt, _f);
1646                 case_bvf(blt);
1647                 case_bvv(ble, _f);
1648                 case_bvf(ble);
1649                 case_bvv(beq, _f);
1650                 case_bvf(beq);
1651                 case_bvv(bge, _f);
1652                 case_bvf(bge);
1653                 case_bvv(bgt, _f);
1654                 case_bvf(bgt);
1655                 case_bvv(bne, _f);
1656                 case_bvf(bne);
1657                 case_bvv(bunlt, _f);
1658                 case_bvf(bunlt);
1659                 case_bvv(bunle, _f);
1660                 case_bvf(bunle);
1661                 case_bvv(buneq, _f);
1662                 case_bvf(buneq);
1663                 case_bvv(bunge, _f);
1664                 case_bvf(bunge);
1665                 case_bvv(bungt, _f);
1666                 case_bvf(bungt);
1667                 case_bvv(bltgt, _f);
1668                 case_bvf(bltgt);
1669                 case_bvv(bord, _f);
1670                 case_bvf(bord);
1671                 case_bvv(bunord, _f);
1672                 case_bvf(bunord);
1673                 case_vvv(add, _d);
1674                 case_vvd(add);
1675                 case_vvv(sub, _d);
1676                 case_vvd(sub);
1677                 case_vvd(rsb);
1678                 case_vvv(mul, _d);
1679                 case_vvd(mul);
1680                 case_vvv(div, _d);
1681                 case_vvd(div);
1682                 case_vv(abs, _d);
1683                 case_vv(neg, _d);
1684                 case_vv(sqrt, _d);
1685                 case_vv(ext, _d);
1686                 case_vv(ld, _d);
1687                 case_vw(ld, _d);
1688                 case_vvv(ldx, _d);
1689                 case_vvw(ldx, _d);
1690                 case_vv(st, _d);
1691                 case_wv(st, _d);
1692                 case_vvv(stx, _d);
1693                 case_wvv(stx, _d);
1694                 case_vv(mov, _d);
1695             case jit_code_movi_d:
1696                 assert_data(node);
1697                 if (jit_swf_p())
1698                     swf_movi_d(rn(node->u.w), node->v.d);
1699                 else
1700                     vfp_movi_d(rn(node->u.w), node->v.d);
1701                 break;
1702                 case_vv(ext, _f_d);
1703                 case_vvv(lt, _d);
1704                 case_vvd(lt);
1705                 case_vvv(le, _d);
1706                 case_vvd(le);
1707                 case_vvv(eq, _d);
1708                 case_vvd(eq);
1709                 case_vvv(ge, _d);
1710                 case_vvd(ge);
1711                 case_vvv(gt, _d);
1712                 case_vvd(gt);
1713                 case_vvv(ne, _d);
1714                 case_vvd(ne);
1715                 case_vvv(unlt, _d);
1716                 case_vvd(unlt);
1717                 case_vvv(unle, _d);
1718                 case_vvd(unle);
1719                 case_vvv(uneq, _d);
1720                 case_vvd(uneq);
1721                 case_vvv(unge, _d);
1722                 case_vvd(unge);
1723                 case_vvv(ungt, _d);
1724                 case_vvd(ungt);
1725                 case_vvv(ltgt, _d);
1726                 case_vvd(ltgt);
1727                 case_vvv(ord, _d);
1728                 case_vvd(ord);
1729                 case_vvv(unord, _d);
1730                 case_vvd(unord);
1731                 case_bvv(blt, _d);
1732                 case_bvd(blt);
1733                 case_bvv(ble, _d);
1734                 case_bvd(ble);
1735                 case_bvv(beq, _d);
1736                 case_bvd(beq);
1737                 case_bvv(bge, _d);
1738                 case_bvd(bge);
1739                 case_bvv(bgt, _d);
1740                 case_bvd(bgt);
1741                 case_bvv(bne, _d);
1742                 case_bvd(bne);
1743                 case_bvv(bunlt, _d);
1744                 case_bvd(bunlt);
1745                 case_bvv(bunle, _d);
1746                 case_bvd(bunle);
1747                 case_bvv(buneq, _d);
1748                 case_bvd(buneq);
1749                 case_bvv(bunge, _d);
1750                 case_bvd(bunge);
1751                 case_bvv(bungt, _d);
1752                 case_bvd(bungt);
1753                 case_bvv(bltgt, _d);
1754                 case_bvd(bltgt);
1755                 case_bvv(bord, _d);
1756                 case_bvd(bord);
1757                 case_bvv(bunord, _d);
1758                 case_bvd(bunord);
1759             case jit_code_jmpr:
1760                 jmpr(rn(node->u.w));
1761                 flush_consts();
1762                 break;
1763             case jit_code_jmpi:
1764                 if (node->flag & jit_flag_node) {
1765                     temp = node->u.n;
1766                     assert(temp->code == jit_code_label ||
1767                            temp->code == jit_code_epilog);
1768                     if (temp->flag & jit_flag_patch)
1769                         jmpi(temp->u.w);
1770                     else {
1771                         word = jmpi_p(_jit->pc.w, 1);
1772                         patch(word, node);
1773                     }
1774                 }
1775                 else
1776                     jmpi(node->u.w);
1777                 flush_consts();
1778                 break;
1779             case jit_code_callr:
1780                 callr(rn(node->u.w));
1781                 break;
1782             case jit_code_calli:
1783                 if (node->flag & jit_flag_node) {
1784                     temp = node->u.n;
1785                     assert(temp->code == jit_code_label ||
1786                            temp->code == jit_code_epilog);
1787                     if (temp->flag & jit_flag_patch)
1788                         calli(temp->u.w);
1789                     else {
1790                         word = calli_p(_jit->pc.w);
1791                         patch(word, node);
1792                     }
1793                 }
1794                 else
1795                     calli(node->u.w);
1796                 break;
1797             case jit_code_prolog:
1798                 _jitc->function = _jitc->functions.ptr + node->w.w;
1799                 undo.node = node;
1800                 undo.word = _jit->pc.w;
1801 #if DEVEL_DISASSEMBLER
1802                 undo.prevw = prevw;
1803 #endif
1804                 undo.data = _jitc->consts.data;
1805                 undo.thumb = _jitc->thumb;
1806                 undo.const_offset = _jitc->consts.offset;
1807                 undo.patch_offset = _jitc->patches.offset;
1808 #if DISASSEMBLER
1809                 if (_jitc->data_info.ptr)
1810                     undo.info_offset = _jitc->data_info.offset;
1811 #endif
1812             restart_function:
1813                 _jitc->again = 0;
1814                 prolog(node);
1815                 break;
1816             case jit_code_epilog:
1817                 assert(_jitc->function == _jitc->functions.ptr + node->w.w);
1818                 if (_jitc->again) {
1819                     for (temp = undo.node->next;
1820                          temp != node; temp = temp->next) {
1821                         if (temp->code == jit_code_label ||
1822                             temp->code == jit_code_epilog)
1823                             temp->flag &= ~jit_flag_patch;
1824                     }
1825                     temp->flag &= ~jit_flag_patch;
1826                     node = undo.node;
1827                     _jit->pc.w = undo.word;
1828 #if DEVEL_DISASSEMBLER
1829                     prevw = undo.prevw;
1830 #endif
1831                     invalidate_consts();
1832                     _jitc->consts.data = undo.data;
1833                     _jitc->thumb = undo.thumb;
1834                     _jitc->consts.offset = undo.const_offset;
1835                     _jitc->patches.offset = undo.patch_offset;
1836 #if DISASSEMBLER
1837                     if (_jitc->data_info.ptr)
1838                         _jitc->data_info.offset = undo.info_offset;
1839 #endif
1840                     goto restart_function;
1841                 }
1842                 /* remember label is defined */
1843                 node->flag |= jit_flag_patch;
1844                 node->u.w = _jit->pc.w;
1845                 epilog(node);
1846                 _jitc->function = NULL;
1847                 flush_consts();
1848                 break;
1849             case jit_code_movr_w_f:
1850                 if (jit_swf_p())
1851                     swf_movr_f(rn(node->u.w), rn(node->v.w));
1852                 else
1853                     vfp_movr_f(rn(node->u.w), rn(node->v.w));
1854                 break;
1855             case jit_code_movr_f_w:
1856                 if (jit_swf_p())
1857                     swf_movr_f(rn(node->u.w), rn(node->v.w));
1858                 else
1859                     vfp_movr_f(rn(node->u.w), rn(node->v.w));
1860                 break;
1861             case jit_code_movi_f_w:
1862                 assert_data(node);
1863                 if (jit_swf_p())
1864                     swf_movi_f(rn(node->u.w), node->v.f);
1865                 else
1866                     vfp_movi_f(rn(node->u.w), node->v.f);
1867                 break;
1868             case jit_code_movr_ww_d:
1869                 if (jit_swf_p())
1870                     swf_movr_d(rn(node->u.w), rn(node->v.w));
1871                 else
1872                     vfp_movr_d(rn(node->u.w), rn(node->v.w));
1873                 break;
1874             case jit_code_movr_d_ww:
1875                 if (jit_swf_p())
1876                     swf_movr_d(rn(node->u.w), rn(node->w.w));
1877                 else
1878                     vfp_movr_d(rn(node->u.w), rn(node->w.w));
1879                 break;
1880             case jit_code_movi_d_ww:
1881                 assert_data(node);
1882                 if (jit_swf_p())
1883                     swf_movi_d(rn(node->u.w), node->w.d);
1884                 else
1885                     vfp_movi_d(rn(node->u.w), node->w.d);
1886                 break;
1887             case jit_code_va_start:
1888                 vastart(rn(node->u.w));
1889                 break;
1890             case jit_code_va_arg:
1891                 vaarg(rn(node->u.w), rn(node->v.w));
1892                 break;
1893             case jit_code_va_arg_d:
1894                 if (jit_swf_p())
1895                     swf_vaarg_d(rn(node->u.w), rn(node->v.w));
1896                 else
1897                     vfp_vaarg_d(rn(node->u.w), rn(node->v.w));
1898                 break;
1899             case jit_code_live:                 case jit_code_ellipsis:
1900             case jit_code_va_push:
1901             case jit_code_allocai:              case jit_code_allocar:
1902             case jit_code_arg:
1903             case jit_code_arg_f:                case jit_code_arg_d:
1904             case jit_code_va_end:
1905             case jit_code_ret:
1906             case jit_code_retr:                 case jit_code_reti:
1907             case jit_code_retr_f:               case jit_code_reti_f:
1908             case jit_code_retr_d:               case jit_code_reti_d:
1909             case jit_code_getarg_c:             case jit_code_getarg_uc:
1910             case jit_code_getarg_s:             case jit_code_getarg_us:
1911             case jit_code_getarg_i:
1912             case jit_code_getarg_f:             case jit_code_getarg_d:
1913             case jit_code_putargr:              case jit_code_putargi:
1914             case jit_code_putargr_f:            case jit_code_putargi_f:
1915             case jit_code_putargr_d:            case jit_code_putargi_d:
1916             case jit_code_pushargr:             case jit_code_pushargi:
1917             case jit_code_pushargr_f:           case jit_code_pushargi_f:
1918             case jit_code_pushargr_d:           case jit_code_pushargi_d:
1919             case jit_code_retval_c:             case jit_code_retval_uc:
1920             case jit_code_retval_s:             case jit_code_retval_us:
1921             case jit_code_retval_i:
1922             case jit_code_retval_f:             case jit_code_retval_d:
1923             case jit_code_prepare:
1924             case jit_code_finishr:              case jit_code_finishi:
1925                 break;
1926             default:
1927                 abort();
1928         }
1929         jit_regarg_clr(node, value);
1930         assert(_jitc->regarg == 0 && _jitc->synth == 0);
1931         /* update register live state */
1932         jit_reglive(node);
1933
1934         if (_jitc->consts.length &&
1935             (_jit->pc.uc - _jitc->consts.data >= 3968 ||
1936              (jit_uword_t)_jit->pc.uc -
1937              (jit_uword_t)_jitc->consts.patches[0] >= 3968)) {
1938             /* longest sequence should be 64 bytes, but preventively
1939              * do not let it go past 128 remaining bytes before a flush */
1940             if (node->next &&
1941                 node->next->code != jit_code_jmpi &&
1942                 node->next->code != jit_code_jmpr &&
1943                 node->next->code != jit_code_epilog) {
1944                 /* insert a jump, flush constants and continue */
1945                 word = _jit->pc.w;
1946                 assert(!jit_thumb_p());
1947                 B(0);
1948                 flush_consts();
1949                 patch_at(arm_patch_jump, word, _jit->pc.w);
1950             }
1951         }
1952     }
1953 #undef case_bvd
1954 #undef case_bvf
1955 #undef case_brw
1956 #undef case_bvv
1957 #undef case_brr
1958 #undef case_wvv
1959 #undef case_wrr
1960 #undef case_vvd
1961 #undef case_vvf
1962 #undef case_vvw
1963 #undef case_rrw
1964 #undef case_vvv
1965 #undef case_rrr
1966 #undef case_wv
1967 #undef case_wr
1968 #undef case_vw
1969 #undef case_vv
1970 #undef case_rw
1971 #undef case_rr
1972
1973     flush_consts();
1974     for (offset = 0; offset < _jitc->patches.offset; offset++) {
1975         assert(_jitc->patches.ptr[offset].kind & arm_patch_node);
1976         node = _jitc->patches.ptr[offset].node;
1977         word = _jitc->patches.ptr[offset].inst;
1978         if (!jit_thumb_p() &&
1979             (node->code == jit_code_movi || node->code == jit_code_calli)) {
1980             /* calculate where to patch word */
1981             value = *(jit_int32_t *)word;
1982             assert((value & 0x0f700000) == ARM_LDRI);
1983             /* offset may become negative (-4) if last instruction
1984              * before unconditional branch and data following
1985              * FIXME can this cause issues in the preprocessor prefetch
1986              * or something else? should not, as the constants are after
1987              * an unconditional jump */
1988             if (value & ARM_P)  value =   value & 0x00000fff;
1989             else                value = -(value & 0x00000fff);
1990             word = word + 8 + value;
1991         }
1992         value = node->code == jit_code_movi ? node->v.n->u.w : node->u.n->u.w;
1993         patch_at(_jitc->patches.ptr[offset].kind & ~arm_patch_node, word, value);
1994     }
1995
1996     jit_flush(_jit->code.ptr, _jit->pc.uc);
1997
1998     return (_jit->code.ptr);
1999 }
2000
2001 #define CODE                            1
2002 #  include "jit_rewind.c"
2003 #  include "jit_arm-cpu.c"
2004 #  include "jit_arm-swf.c"
2005 #  include "jit_arm-vfp.c"
2006 #undef CODE
2007
2008 void
2009 jit_flush(void *fptr, void *tptr)
2010 {
2011 #if defined(__GNUC__)
2012     jit_uword_t         i, f, t, s;
2013
2014     s = sysconf(_SC_PAGE_SIZE);
2015     f = (jit_uword_t)fptr & -s;
2016     t = (((jit_uword_t)tptr) + s - 1) & -s;
2017     for (i = f; i < t; i += s)
2018         __clear_cache((void *)i, (void *)(i + s));
2019 #endif
2020 }
2021
2022 void
2023 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
2024 {
2025     ldxi_i(rn(r0), rn(r1), i0);
2026 }
2027
2028 void
2029 _emit_stxi(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
2030 {
2031     stxi_i(i0, rn(r0), rn(r1));
2032 }
2033
2034 void
2035 _emit_ldxi_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
2036 {
2037     if (jit_swf_p())
2038         swf_ldxi_d(rn(r0), rn(r1), i0);
2039     else
2040         vfp_ldxi_d(rn(r0), rn(r1), i0);
2041 }
2042
2043 void
2044 _emit_stxi_d(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
2045 {
2046     if (jit_swf_p())
2047         swf_stxi_d(i0, rn(r0), rn(r1));
2048     else
2049         vfp_stxi_d(i0, rn(r0), rn(r1));
2050 }
2051
2052 static jit_int32_t
2053 _jit_get_reg_pair(jit_state_t *_jit)
2054 {
2055     /*   bypass jit_get_reg() with argument or'ed with jit_class_chk
2056      * and try to find an consecutive, even free register pair, or
2057      * return JIT_NOREG if fail, as the cost of spills is greater
2058      * than splitting a double load/store in two operations. */
2059     if (jit_reg_free_p(_R0) && jit_reg_free_p(_R1)) {
2060         jit_regset_setbit(&_jitc->regarg, _R0);
2061         jit_regset_setbit(&_jitc->regarg, _R1);
2062         return (_R0);
2063     }
2064     if (jit_reg_free_p(_R2) && jit_reg_free_p(_R3)) {
2065         jit_regset_setbit(&_jitc->regarg, _R2);
2066         jit_regset_setbit(&_jitc->regarg, _R3);
2067         return (_R2);
2068     }
2069     if (jit_reg_free_p(_R4) && jit_reg_free_p(_R5)) {
2070         jit_regset_setbit(&_jitc->regarg, _R4);
2071         jit_regset_setbit(&_jitc->regarg, _R5);
2072         return (_R4);
2073     }
2074     if (jit_reg_free_p(_R6) && jit_reg_free_p(_R7)) {
2075         jit_regset_setbit(&_jitc->regarg, _R6);
2076         jit_regset_setbit(&_jitc->regarg, _R7);
2077         return (_R6);
2078     }
2079     if (jit_reg_free_p(_R8) && jit_reg_free_p(_R9)) {
2080         jit_regset_setbit(&_jitc->regarg, _R8);
2081         jit_regset_setbit(&_jitc->regarg, _R9);
2082         return (_R8);
2083     }
2084     return (JIT_NOREG);
2085 }
2086
2087 static void
2088 _jit_unget_reg_pair(jit_state_t *_jit, jit_int32_t reg)
2089 {
2090     jit_unget_reg(reg);
2091     switch (reg) {
2092         case _R0:       jit_unget_reg(_R1);     break;
2093         case _R2:       jit_unget_reg(_R3);     break;
2094         case _R4:       jit_unget_reg(_R5);     break;
2095         case _R6:       jit_unget_reg(_R7);     break;
2096         case _R8:       jit_unget_reg(_R9);     break;
2097         default:        abort();
2098     }
2099 }
2100
2101 /*   A prolog must be aligned at mod 4 bytes boundary.
2102  *   This condition was not being required to be tested by
2103  * accident previously, but with the jit_frame and jit_tramp
2104  * code it is required */
2105 static jit_bool_t
2106 _must_align_p(jit_state_t *_jit, jit_node_t *node)
2107 {
2108     if (jit_thumb_p() && (_jit->pc.w & 3)) {
2109         for (; node; node = node->next) {
2110             switch (node->code) {
2111                 case jit_code_note:
2112                 case jit_code_name:
2113                 case jit_code_label:
2114                     break;
2115                 case jit_code_prolog:
2116                     return (1);
2117                 default:
2118                     return (0);
2119             }
2120         }
2121     }
2122     return (0);
2123 }
2124
2125 static void
2126 _load_const(jit_state_t *_jit, jit_bool_t uniq, jit_int32_t r0, jit_word_t i0)
2127 {
2128     jit_word_t           w;
2129     jit_word_t           d;
2130     jit_word_t           base;
2131     jit_int32_t         *data;
2132     jit_int32_t          size;
2133     jit_int32_t          offset;
2134
2135     assert(!jit_thumb_p());
2136     if (!uniq) {
2137         /* use zero, a valid directly encoded immediate, to avoid the
2138          * need of a bitmask to know what offsets will be patched, so
2139          * that comparison will always fail for constants that cannot
2140          * be encoded */
2141         assert(i0 != 0);
2142
2143         /* Actually, code is (currently at least) not self modifying,
2144          * so, any value reachable backwards is valid as a constant. */
2145
2146         /* FIXME a quickly updateable/mutable hash table could be
2147          * better here, but most times only a few comparisons
2148          * should be done
2149          */
2150
2151         /* search in previous constant pool */
2152         if ((data = (jit_int32_t *)_jitc->consts.data)) {
2153             w = (jit_word_t)data;
2154             /* maximum backwards offset */
2155             base = (_jit->pc.w + 8) - 4092;
2156             if (base <= w)
2157                 /* can scan all possible available backward constants */
2158                 base = 0;
2159             else
2160                 base = (base - w) >> 2;
2161             size = _jitc->consts.size >> 2;
2162             for (offset = size - 1; offset >= base; offset--) {
2163                 if (data[offset] == i0) {
2164                     w = (jit_word_t)(data + offset);
2165                     d = (_jit->pc.w + 8) - w;
2166                     LDRIN(r0, _R15_REGNO, d);
2167                     return;
2168                 }
2169             }
2170         }
2171     }
2172     else
2173         assert(i0 == 0);
2174
2175     _jitc->consts.patches[_jitc->consts.offset++] = _jit->pc.w;
2176     /* (probably) positive forward offset */
2177     LDRI(r0, _R15_REGNO, 0);
2178
2179     if (!uniq) {
2180         /* search already requested values */
2181         for (offset = 0; offset < _jitc->consts.length; offset++) {
2182             if (_jitc->consts.values[offset] == i0) {
2183                 _jitc->consts.patches[_jitc->consts.offset++] = offset;
2184                 return;
2185             }
2186         }
2187     }
2188
2189 #if DEBUG
2190     /* cannot run out of space because of limited range
2191      * but assert anyway to catch logic errors */
2192     assert(_jitc->consts.length < 1024);
2193     assert(_jitc->consts.offset < 2048);
2194 #endif
2195     _jitc->consts.patches[_jitc->consts.offset++] = _jitc->consts.length;
2196     _jitc->consts.values[_jitc->consts.length++] = i0;
2197 }
2198
2199 static void
2200 _flush_consts(jit_state_t *_jit)
2201 {
2202     jit_word_t           word;
2203     jit_int32_t          offset;
2204
2205     /* if no forward constants */
2206     if (!_jitc->consts.length)
2207         return;
2208     assert(!jit_thumb_p());
2209     word = _jit->pc.w;
2210     _jitc->consts.data = _jit->pc.uc;
2211     _jitc->consts.size = _jitc->consts.length << 2;
2212     /* FIXME check will not overrun, otherwise, need to reallocate
2213      * code buffer and start over */
2214     jit_memcpy(_jitc->consts.data, _jitc->consts.values, _jitc->consts.size);
2215     _jit->pc.w += _jitc->consts.size;
2216
2217 #if DISASSEMBLER
2218     if (_jitc->data_info.ptr) {
2219         if (_jitc->data_info.offset >= _jitc->data_info.length) {
2220             jit_realloc((jit_pointer_t *)&_jitc->data_info.ptr,
2221                         _jitc->data_info.length * sizeof(jit_data_info_t),
2222                         (_jitc->data_info.length + 1024) *
2223                         sizeof(jit_data_info_t));
2224             _jitc->data_info.length += 1024;
2225         }
2226         _jitc->data_info.ptr[_jitc->data_info.offset].code = word;
2227         _jitc->data_info.ptr[_jitc->data_info.offset].length = _jitc->consts.size;
2228         ++_jitc->data_info.offset;
2229     }
2230 #endif
2231
2232     for (offset = 0; offset < _jitc->consts.offset; offset += 2)
2233         patch_at(arm_patch_load, _jitc->consts.patches[offset],
2234                  word + (_jitc->consts.patches[offset + 1] << 2));
2235     _jitc->consts.length = _jitc->consts.offset = 0;
2236 }
2237
2238 /* to be called if needing to start over a function */
2239 static void
2240 _invalidate_consts(jit_state_t *_jit)
2241 {
2242     /* if no forward constants */
2243     if (_jitc->consts.length)
2244         _jitc->consts.length = _jitc->consts.offset = 0;
2245 }
2246
2247 static void
2248 _patch(jit_state_t *_jit, jit_word_t instr, jit_node_t *node)
2249 {
2250     jit_int32_t          flag;
2251     jit_int32_t          kind;
2252
2253     assert(node->flag & jit_flag_node);
2254     if (node->code == jit_code_movi) {
2255         flag = node->v.n->flag;
2256         kind = arm_patch_word;
2257     }
2258     else {
2259         flag = node->u.n->flag;
2260         if (node->code == jit_code_calli ||
2261             (node->code == jit_code_jmpi && !(node->flag & jit_flag_node)))
2262             kind = arm_patch_word;
2263         else
2264             kind = arm_patch_jump;
2265     }
2266     assert(!(flag & jit_flag_patch));
2267     kind |= arm_patch_node;
2268     if (_jitc->patches.offset >= _jitc->patches.length) {
2269         jit_realloc((jit_pointer_t *)&_jitc->patches.ptr,
2270                     _jitc->patches.length * sizeof(jit_patch_t),
2271                     (_jitc->patches.length + 1024) * sizeof(jit_patch_t));
2272         _jitc->patches.length += 1024;
2273     }
2274     _jitc->patches.ptr[_jitc->patches.offset].kind = kind;
2275     _jitc->patches.ptr[_jitc->patches.offset].inst = instr;
2276     _jitc->patches.ptr[_jitc->patches.offset].node = node;
2277     ++_jitc->patches.offset;
2278 }