deps/lightning/lib/jit_arm.c

   1 /*
   2  * Copyright (C) 2012-2023  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 stack_framesize                 48
  25
  26 #define jit_arg_reg_p(i)                ((i) >= 0 && (i) < 4)
  27 #define jit_arg_f_reg_p(i)              ((i) >= 0 && (i) < 16)
  28 #define jit_arg_d_reg_p(i)              ((i) >= 0 && (i) < 15)
  29
  30 #define arm_patch_node                  0x80000000
  31 #define arm_patch_word                  0x40000000
  32 #define arm_patch_jump                  0x20000000
  33 #define arm_patch_load                  0x10000000
  34 #define arm_patch_call                  0x08000000
  35
  36 #define jit_fpr_p(rn)                   ((rn) > 15)
  37
  38 #define arg_base()                      (stack_framesize - 16)
  39 #define arg_offset(n)                                                   \
  40     ((n) < 4 ? arg_base() + ((n) << 2) : (n))
  41
  42 /* Assume functions called never match jit instruction set, that is
  43  * libc, gmp, mpfr, etc functions are in thumb mode and jit is in
  44  * arm mode, what may cause a crash upon return of that function
  45  * if generating jit for a relative jump.
  46  */
  47 #define jit_exchange_p()                jit_cpu.exchange
  48
  49 /* FIXME is it really required to not touch _R10? */
  50
  51 #define CHECK_REG_ARGS()                                                \
  52     do {                                                                \
  53         if (!_jitc->function->save_reg_args)                            \
  54             _jitc->again = _jitc->function->save_reg_args = 1;          \
  55     } while (0)
  56
  57 #define CHECK_SWF_OFFSET()                                              \
  58     do {                                                                \
  59         if (!_jitc->function->swf_offset) {                             \
  60             _jitc->again = _jitc->function->save_reg_args =             \
  61                 _jitc->function->swf_offset = 1;                        \
  62             _jitc->function->self.aoff = -64;                           \
  63         }                                                               \
  64     } while (0)
  65
  66 #define CHECK_RETURN()                                                  \
  67     do {                                                                \
  68         if (!_jitc->function->need_frame &&                             \
  69             !_jitc->function->need_return)                              \
  70             _jitc->again = _jitc->function->need_return = 1;            \
  71     } while (0)
  72
  73 /*
  74  * Types
  75  */
  76 typedef union _jit_thumb_t {
  77     jit_int32_t         i;
  78     jit_int16_t         s[2];
  79 } jit_thumb_t;
  80
  81 typedef jit_pointer_t   jit_va_list;
  82
  83 /*
  84  * Prototypes
  85  */
  86 #define jit_make_arg(node,code)         _jit_make_arg(_jit,node,code)
  87 static jit_node_t *_jit_make_arg(jit_state_t*,jit_node_t*,jit_code_t);
  88 #define jit_make_arg_f(node)            _jit_make_arg_f(_jit,node)
  89 static jit_node_t *_jit_make_arg_f(jit_state_t*,jit_node_t*);
  90 #define jit_make_arg_d(node)            _jit_make_arg_d(_jit,node)
  91 static jit_node_t *_jit_make_arg_d(jit_state_t*,jit_node_t*);
  92 #define jit_get_reg_pair()              _jit_get_reg_pair(_jit)
  93 static jit_int32_t _jit_get_reg_pair(jit_state_t*);
  94 #define jit_unget_reg_pair(rn)          _jit_unget_reg_pair(_jit,rn)
  95 static void _jit_unget_reg_pair(jit_state_t*,jit_int32_t);
  96 # define must_align_p(node)             _must_align_p(_jit, node)
  97 static jit_bool_t _must_align_p(jit_state_t*,jit_node_t*);
  98 #define load_const(uniq,r0,i0)          _load_const(_jit,uniq,r0,i0)
  99 static void _load_const(jit_state_t*,jit_bool_t,jit_int32_t,jit_word_t);
 100 #define flush_consts()                  _flush_consts(_jit)
 101 static void _flush_consts(jit_state_t*);
 102 #define invalidate_consts()             _invalidate_consts(_jit)
 103 static void _invalidate_consts(jit_state_t*);
 104 #define compute_framesize()             _compute_framesize(_jit)
 105 static void _compute_framesize(jit_state_t*);
 106 #define patch(instr, node, kind)        _patch(_jit, instr, node, kind)
 107 static void _patch(jit_state_t*,jit_word_t,jit_node_t*,jit_int32_t);
 108
 109 #if defined(__GNUC__)
 110 /* libgcc */
 111 extern void __clear_cache(void *, void *);
 112 #endif
 113
 114 #define PROTO                           1
 115 #  include "jit_rewind.c"
 116 #  include "jit_arm-cpu.c"
 117 #  include "jit_arm-swf.c"
 118 #  include "jit_arm-vfp.c"
 119 #  include "jit_fallback.c"
 120 #undef PROTO
 121
 122 /*
 123  * Initialization
 124  */
 125 jit_cpu_t               jit_cpu;
 126 jit_register_t          _rvs[] = {
 127     { rc(gpr) | 0x0c,                   "ip" },
 128     { rc(sav) | rc(gpr) | 0x04,         "r4" },
 129     { rc(sav) | rc(gpr) | 0x05,         "r5" },
 130     { rc(sav) | rc(gpr) | 0x06,         "r6" },
 131     { rc(sav) | rc(gpr) | 0x07,         "r7" },
 132     { rc(sav) | rc(gpr) | 0x08,         "r8" },
 133     { rc(sav) | rc(gpr) | 0x09,         "r9" },
 134     { rc(sav) | 0x0a,                   "sl" },
 135     { rc(sav) | 0x0b,                   "fp" },
 136     { rc(sav) | 0x0d,                   "sp" },
 137     { rc(sav) | 0x0e,                   "lr" },
 138     { 0x0f,                             "pc" },
 139     { rc(arg) | rc(gpr) | 0x03,         "r3" },
 140     { rc(arg) | rc(gpr) | 0x02,         "r2" },
 141     { rc(arg) | rc(gpr) | 0x01,         "r1" },
 142     { rc(arg) | rc(gpr) | 0x00,         "r0" },
 143     { rc(fpr) | 0x20,                   "d8" },
 144     { 0x21,                             "s17" },
 145     { rc(fpr) | 0x22,                   "d9" },
 146     { 0x23,                             "s19" },
 147     { rc(fpr) | 0x24,                   "d10" },
 148     { 0x25,                             "s21" },
 149     { rc(fpr) | 0x26,                   "d11" },
 150     { 0x27,                             "s23" },
 151     { rc(fpr) | 0x28,                   "d12" },
 152     { 0x29,                             "s25" },
 153     { rc(fpr) | 0x2a,                   "d13" },
 154     { 0x2b,                             "s27" },
 155     { rc(fpr) | 0x2c,                   "d14" },
 156     { 0x2d,                             "s29" },
 157     { rc(fpr) | 0x2e,                   "d15" },
 158     { 0x2f,                             "s31" },
 159     { rc(arg) | 0x1f,                   "s15" },
 160     { rc(arg)|rc(sft)|rc(fpr)|0x1e,     "d7" },
 161     { rc(arg) | 0x1d,                   "s13" },
 162     { rc(arg)|rc(sft)|rc(fpr)|0x1c,     "d6" },
 163     { rc(arg) | 0x1b,                   "s11" },
 164     { rc(arg)|rc(sft)|rc(fpr)|0x1a,     "d5" },
 165     { rc(arg) | 0x19,                   "s9" },
 166     { rc(arg)|rc(sft)|rc(fpr)|0x18,     "d4" },
 167     { rc(arg) | 0x17,                   "s7" },
 168     { rc(arg)|rc(sft)|rc(fpr)|0x16,     "d3" },
 169     { rc(arg) | 0x15,                   "s5" },
 170     { rc(arg)|rc(sft)|rc(fpr)|0x14,     "d2" },
 171     { rc(arg) | 0x13,                   "s3" },
 172     { rc(arg)|rc(sft)|rc(fpr)|0x12,     "d1" },
 173     { rc(arg) | 0x11,                   "s1" },
 174     { rc(arg)|rc(sft)|rc(fpr)|0x10,     "d0" },
 175     { _NOREG,                           "<none>" },
 176 };
 177
 178 static jit_int32_t iregs[] = {
 179     _R4, _R5, _R6, _R7, _R8, _R9,
 180 };
 181
 182 /*
 183  * Implementation
 184  */
 185 void
 186 jit_get_cpu(void)
 187 {
 188 #if defined(__linux__)
 189     FILE        *fp;
 190     char        *ptr;
 191     char         buf[128];
 192
 193     if ((fp = fopen("/proc/cpuinfo", "r")) != NULL) {
 194         while (fgets(buf, sizeof(buf), fp)) {
 195             if (strncmp(buf, "CPU architecture:", 17) == 0) {
 196                 jit_cpu.version = strtol(buf + 17, &ptr, 10);
 197                 while (*ptr) {
 198                     if (*ptr == 'T' || *ptr == 't') {
 199                         ++ptr;
 200                         jit_cpu.thumb = 1;
 201                     }
 202                     else if (*ptr == 'E' || *ptr == 'e') {
 203                         jit_cpu.extend = 1;
 204                         ++ptr;
 205                     }
 206                     else
 207                         ++ptr;
 208                 }
 209             }
 210             else if (strncmp(buf, "Features\t:", 10) == 0) {
 211                 if ((ptr = strstr(buf + 10, "vfpv")))
 212                     jit_cpu.vfp = strtol(ptr + 4, NULL, 0);
 213                 if ((ptr = strstr(buf + 10, "neon")))
 214                     jit_cpu.neon = 1;
 215                 if ((ptr = strstr(buf + 10, "thumb")))
 216                     jit_cpu.thumb = 1;
 217             }
 218         }
 219         fclose(fp);
 220     }
 221 #endif
 222 #if defined(__ARM_PCS_VFP)
 223     if (!jit_cpu.vfp)
 224         jit_cpu.vfp = 3;
 225     if (!jit_cpu.version)
 226         jit_cpu.version = 7;
 227     jit_cpu.abi = 1;
 228 #endif
 229 #if defined(__thumb2__)
 230     jit_cpu.thumb = 1;
 231 #endif
 232     /* armv6t2 todo (software float and thumb2) */
 233     if (!jit_cpu.vfp && jit_cpu.thumb)
 234         jit_cpu.thumb = 0;
 235     /* FIXME need test environments for the below. For the moment just
 236      * be very conservative */
 237     /* force generation of code assuming jit and function libraries called
 238      * instruction set do not match */
 239     jit_cpu.exchange = 1;
 240     /* do not generate hardware integer division by default */
 241     if (jit_cpu.version == 7)
 242         jit_cpu.extend = 0;
 243
 244     /* By default generate extra instructions for unaligned load/store. */
 245     jit_cpu.unaligned = 1;
 246     /* Linux should only not handle unaligned vfp load/store */
 247     jit_cpu.vfp_unaligned = 1;
 248 }
 249
 250 void
 251 _jit_init(jit_state_t *_jit)
 252 {
 253     jit_int32_t         regno;
 254     static jit_bool_t   first = 1;
 255
 256     _jitc->reglen = jit_size(_rvs) - 1;
 257     if (first) {
 258         /* jit_get_cpu() should have been already called, and only once */
 259         if (!jit_cpu.vfp) {
 260             /* cause register to never be allocated, because simple
 261              * software float only allocates stack space for 8 slots  */
 262             for (regno = _D8; regno < _D7; regno++)
 263                 _rvs[regno].spec = 0;
 264         }
 265         if (!jit_cpu.abi) {
 266             for (regno = _S15; regno <= _D0; regno++)
 267                 _rvs[regno].spec &= ~rc(arg);
 268         }
 269         first = 0;
 270     }
 271 }
 272
 273 void
 274 _jit_prolog(jit_state_t *_jit)
 275 {
 276     jit_int32_t          offset;
 277
 278     if (_jitc->function)
 279         jit_epilog();
 280     assert(jit_regset_cmp_ui(&_jitc->regarg, 0) == 0);
 281     jit_regset_set_ui(&_jitc->regsav, 0);
 282     offset = _jitc->functions.offset;
 283     if (offset >= _jitc->functions.length) {
 284         jit_realloc((jit_pointer_t *)&_jitc->functions.ptr,
 285                     _jitc->functions.length * sizeof(jit_function_t),
 286                     (_jitc->functions.length + 16) * sizeof(jit_function_t));
 287         _jitc->functions.length += 16;
 288     }
 289     _jitc->function = _jitc->functions.ptr + _jitc->functions.offset++;
 290     _jitc->function->self.size = stack_framesize;
 291     _jitc->function->self.argi = _jitc->function->self.argf =
 292         _jitc->function->self.alen = _jitc->function->self.aoff = 0;
 293     _jitc->function->swf_offset = _jitc->function->save_reg_args =
 294         _jitc->function->need_return = 0;
 295     _jitc->function->self.call = jit_call_default;
 296     jit_alloc((jit_pointer_t *)&_jitc->function->regoff,
 297               _jitc->reglen * sizeof(jit_int32_t));
 298
 299     /* _no_link here does not mean the jit_link() call can be removed
 300      * by rewriting as:
 301      * _jitc->function->prolog = jit_new_node(jit_code_prolog);
 302      */
 303     _jitc->function->prolog = jit_new_node_no_link(jit_code_prolog);
 304     jit_link(_jitc->function->prolog);
 305     _jitc->function->prolog->w.w = offset;
 306     _jitc->function->epilog = jit_new_node_no_link(jit_code_epilog);
 307     /*  u:      label value
 308      *  v:      offset in blocks vector
 309      *  w:      offset in functions vector
 310      */
 311     _jitc->function->epilog->w.w = offset;
 312
 313     jit_regset_new(&_jitc->function->regset);
 314 }
 315
 316 jit_int32_t
 317 _jit_allocai(jit_state_t *_jit, jit_int32_t length)
 318 {
 319     assert(_jitc->function);
 320     if (jit_swf_p())
 321         CHECK_SWF_OFFSET();
 322     jit_check_frame();
 323     switch (length) {
 324         case 0: case 1:                                         break;
 325         case 2:         _jitc->function->self.aoff &= -2;       break;
 326         case 3: case 4: _jitc->function->self.aoff &= -4;       break;
 327         default:        _jitc->function->self.aoff &= -8;       break;
 328     }
 329     _jitc->function->self.aoff -= length;
 330     if (!_jitc->realize) {
 331         jit_inc_synth_ww(allocai, _jitc->function->self.aoff, length);
 332         jit_dec_synth();
 333     }
 334     return (_jitc->function->self.aoff);
 335 }
 336
 337 void
 338 _jit_allocar(jit_state_t *_jit, jit_int32_t u, jit_int32_t v)
 339 {
 340     jit_int32_t          reg;
 341     assert(_jitc->function);
 342     jit_inc_synth_ww(allocar, u, v);
 343     if (!_jitc->function->allocar) {
 344         _jitc->function->aoffoff = jit_allocai(sizeof(jit_int32_t));
 345         _jitc->function->allocar = 1;
 346     }
 347     reg = jit_get_reg(jit_class_gpr);
 348     jit_negr(reg, v);
 349     jit_andi(reg, reg, -8);
 350     jit_ldxi_i(u, JIT_FP, _jitc->function->aoffoff);
 351     jit_addr(u, u, reg);
 352     jit_addr(JIT_SP, JIT_SP, reg);
 353     jit_stxi_i(_jitc->function->aoffoff, JIT_FP, u);
 354     jit_unget_reg(reg);
 355     jit_dec_synth();
 356 }
 357
 358 void
 359 _jit_ret(jit_state_t *_jit)
 360 {
 361     jit_node_t          *instr;
 362     assert(_jitc->function);
 363     jit_inc_synth(ret);
 364     /* jump to epilog */
 365     instr = jit_jmpi();
 366     jit_patch_at(instr, _jitc->function->epilog);
 367     jit_dec_synth();
 368 }
 369
 370 void
 371 _jit_retr(jit_state_t *_jit, jit_int32_t u, jit_code_t code)
 372 {
 373     jit_code_inc_synth_w(code, u);
 374     jit_movr(JIT_RET, u);
 375     jit_ret();
 376     jit_dec_synth();
 377 }
 378
 379 void
 380 _jit_reti(jit_state_t *_jit, jit_word_t u, jit_code_t code)
 381 {
 382     jit_code_inc_synth_w(code, u);
 383     jit_movi(JIT_RET, u);
 384     jit_ret();
 385     jit_dec_synth();
 386 }
 387
 388 void
 389 _jit_retr_f(jit_state_t *_jit, jit_int32_t u)
 390 {
 391     jit_inc_synth_w(retr_f, u);
 392     if (jit_cpu.abi) {
 393         if (u != JIT_FRET)
 394             jit_movr_f(JIT_FRET, u);
 395         else
 396             jit_live(JIT_FRET);
 397     }
 398     else {
 399         if (u != JIT_RET)
 400             jit_movr_f_w(JIT_RET, u);
 401         else
 402             jit_live(JIT_RET);
 403     }
 404     jit_ret();
 405     jit_dec_synth();
 406 }
 407
 408 void
 409 _jit_reti_f(jit_state_t *_jit, jit_float32_t u)
 410 {
 411     jit_inc_synth_f(reti_f, u);
 412     if (jit_cpu.abi)
 413         jit_movi_f(JIT_FRET, u);
 414     else
 415         jit_movi_f_w(JIT_RET, u);
 416     jit_ret();
 417     jit_dec_synth();
 418 }
 419
 420 void
 421 _jit_retr_d(jit_state_t *_jit, jit_int32_t u)
 422 {
 423     jit_inc_synth_w(retr_d, u);
 424     if (jit_cpu.abi) {
 425         if (u != JIT_FRET)
 426             jit_movr_d(JIT_FRET, u);
 427         else
 428             jit_live(JIT_FRET);
 429     }
 430     else {
 431         if (u != JIT_RET)
 432             jit_movr_d_ww(JIT_RET, _R1, u);
 433         else
 434             jit_live(JIT_RET);
 435     }
 436     jit_ret();
 437     jit_dec_synth();
 438 }
 439
 440 void
 441 _jit_reti_d(jit_state_t *_jit, jit_float64_t u)
 442 {
 443     jit_inc_synth_d(reti_d, u);
 444     if (jit_cpu.abi)
 445         jit_movi_d(JIT_FRET, u);
 446     else
 447         jit_movi_d_ww(JIT_RET, _R1, u);
 448     jit_ret();
 449     jit_dec_synth();
 450 }
 451
 452 void
 453 _jit_epilog(jit_state_t *_jit)
 454 {
 455     assert(_jitc->function);
 456     assert(_jitc->function->epilog->next == NULL);
 457     jit_link(_jitc->function->epilog);
 458     _jitc->function = NULL;
 459 }
 460
 461 jit_bool_t
 462 _jit_arg_register_p(jit_state_t *_jit, jit_node_t *u)
 463 {
 464     if (!(u->code >= jit_code_arg_c && u->code <= jit_code_arg)) {
 465         if (u->code == jit_code_arg_f) {
 466             if (jit_cpu.abi)
 467                 return (jit_arg_f_reg_p(u->u.w));
 468         }
 469         else {
 470             assert(u->code == jit_code_arg_d);
 471             if (jit_cpu.abi)
 472                 return (jit_arg_d_reg_p(u->u.w));
 473         }
 474     }
 475     return (jit_arg_reg_p(u->u.w));
 476 }
 477
 478 static jit_node_t *
 479 _jit_make_arg(jit_state_t *_jit, jit_node_t *node, jit_code_t code)
 480 {
 481     jit_int32_t          offset;
 482     if (jit_arg_reg_p(_jitc->function->self.argi))
 483         offset = _jitc->function->self.argi++;
 484     else {
 485         offset = _jitc->function->self.size;
 486         _jitc->function->self.size += sizeof(jit_word_t);
 487     }
 488     if (node == (jit_node_t *)0)
 489         node = jit_new_node(code);
 490     else
 491         link_node(node);
 492     node->u.w = offset;
 493     node->v.w = ++_jitc->function->self.argn;
 494     jit_link_prolog();
 495     return (node);
 496 }
 497
 498 jit_node_t *
 499 _jit_make_arg_f(jit_state_t *_jit, jit_node_t *node)
 500 {
 501     jit_int32_t          offset;
 502     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 503         if (jit_arg_f_reg_p(_jitc->function->self.argf)) {
 504             offset = _jitc->function->self.argf++;
 505             goto done;
 506         }
 507     }
 508     else {
 509         if (jit_arg_reg_p(_jitc->function->self.argi)) {
 510             offset = _jitc->function->self.argi++;
 511             goto done;
 512         }
 513     }
 514     offset = _jitc->function->self.size;
 515     _jitc->function->self.size += sizeof(jit_float32_t);
 516 done:
 517     if (node == (jit_node_t *)0)
 518         node = jit_new_node(jit_code_arg_f);
 519     else
 520         link_node(node);
 521     node->u.w = offset;
 522     node->v.w = ++_jitc->function->self.argn;
 523     jit_link_prolog();
 524     return (node);
 525 }
 526
 527 jit_node_t *
 528 _jit_make_arg_d(jit_state_t *_jit, jit_node_t *node)
 529 {
 530     jit_int32_t          offset;
 531     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 532         if (jit_arg_d_reg_p(_jitc->function->self.argf)) {
 533             if (_jitc->function->self.argf & 1)
 534                 ++_jitc->function->self.argf;
 535             offset = _jitc->function->self.argf;
 536             _jitc->function->self.argf += 2;
 537             goto done;
 538         }
 539     }
 540     else {
 541         if (_jitc->function->self.argi & 1)
 542             ++_jitc->function->self.argi;
 543         if (jit_arg_reg_p(_jitc->function->self.argi)) {
 544             offset = _jitc->function->self.argi;
 545             _jitc->function->self.argi += 2;
 546             goto done;
 547         }
 548     }
 549     if (_jitc->function->self.size & 7)
 550         _jitc->function->self.size += 4;
 551     offset = _jitc->function->self.size;
 552     _jitc->function->self.size += sizeof(jit_float64_t);
 553 done:
 554     if (node == (jit_node_t *)0)
 555         node = jit_new_node(jit_code_arg_d);
 556     else
 557         link_node(node);
 558     node->u.w = offset;
 559     node->v.w = ++_jitc->function->self.argn;
 560     jit_link_prolog();
 561     return (node);
 562 }
 563
 564 void
 565 _jit_ellipsis(jit_state_t *_jit)
 566 {
 567     if (_jitc->prepare) {
 568         assert(!(_jitc->function->call.call & jit_call_varargs));
 569         _jitc->function->call.call |= jit_call_varargs;
 570         if (jit_cpu.abi && _jitc->function->call.argf)
 571             rewind_prepare();
 572     }
 573     else {
 574         assert(!(_jitc->function->self.call & jit_call_varargs));
 575         _jitc->function->self.call |= jit_call_varargs;
 576         CHECK_REG_ARGS();
 577         if (jit_cpu.abi &&  _jitc->function->self.argf)
 578             rewind_prolog();
 579         /* First 4 stack addresses need to be spilled r0-r3 */
 580         if (jit_arg_reg_p(_jitc->function->self.argi))
 581             _jitc->function->vagp = _jitc->function->self.argi * 4;
 582         else
 583             _jitc->function->vagp = 16;
 584     }
 585     jit_inc_synth(ellipsis);
 586     if (_jitc->prepare)
 587         jit_link_prepare();
 588     else
 589         jit_link_prolog();
 590     jit_dec_synth();
 591 }
 592
 593 void
 594 _jit_va_push(jit_state_t *_jit, jit_int32_t u)
 595 {
 596     jit_inc_synth_w(va_push, u);
 597     jit_pushargr(u);
 598     jit_dec_synth();
 599 }
 600
 601 jit_node_t *
 602 _jit_arg(jit_state_t *_jit, jit_code_t code)
 603 {
 604     assert(_jitc->function);
 605     assert(!(_jitc->function->self.call & jit_call_varargs));
 606 #if STRONG_TYPE_CHECKING
 607     assert(code >= jit_code_arg_c && code <= jit_code_arg);
 608 #endif
 609     return (jit_make_arg((jit_node_t*)0, code));
 610 }
 611
 612 jit_node_t *
 613 _jit_arg_f(jit_state_t *_jit)
 614 {
 615     assert(_jitc->function);
 616     assert(!(_jitc->function->self.call & jit_call_varargs));
 617     return (jit_make_arg_f((jit_node_t*)0));
 618 }
 619
 620 jit_node_t *
 621 _jit_arg_d(jit_state_t *_jit)
 622 {
 623     assert(_jitc->function);
 624     assert(!(_jitc->function->self.call & jit_call_varargs));
 625     return (jit_make_arg_d((jit_node_t*)0));
 626 }
 627
 628 void
 629 _jit_getarg_c(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 630 {
 631     jit_node_t          *node = NULL;
 632     assert_arg_type(v->code, jit_code_arg_c);
 633     jit_inc_synth_wp(getarg_c, u, v);
 634     if (jit_swf_p())
 635         node = jit_ldxi_c(u, JIT_FP, arg_offset(v->u.w));
 636     else if (jit_arg_reg_p(v->u.w))
 637         jit_extr_c(u, JIT_RA0 - v->u.w);
 638     else
 639         node = jit_ldxi_c(u, JIT_FP, v->u.w);
 640     if (node) {
 641         CHECK_REG_ARGS();
 642         jit_link_alist(node);
 643         jit_check_frame();
 644     }
 645     jit_dec_synth();
 646 }
 647
 648 void
 649 _jit_getarg_uc(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 650 {
 651     jit_node_t          *node = NULL;
 652     assert_arg_type(v->code, jit_code_arg_c);
 653     jit_inc_synth_wp(getarg_uc, u, v);
 654     if (jit_swf_p())
 655         node = jit_ldxi_uc(u, JIT_FP, arg_offset(v->u.w));
 656     else if (jit_arg_reg_p(v->u.w))
 657         jit_extr_uc(u, JIT_RA0 - v->u.w);
 658     else
 659         node = jit_ldxi_uc(u, JIT_FP, v->u.w);
 660     if (node) {
 661         CHECK_REG_ARGS();
 662         jit_link_alist(node);
 663         jit_check_frame();
 664     }
 665     jit_dec_synth();
 666 }
 667
 668 void
 669 _jit_getarg_s(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 670 {
 671     jit_node_t          *node = NULL;
 672     assert_arg_type(v->code, jit_code_arg_s);
 673     jit_inc_synth_wp(getarg_s, u, v);
 674     if (jit_swf_p())
 675         node = jit_ldxi_s(u, JIT_FP, arg_offset(v->u.w));
 676     else if (jit_arg_reg_p(v->u.w))
 677         jit_extr_s(u, JIT_RA0 - v->u.w);
 678     else
 679         node = jit_ldxi_s(u, JIT_FP, v->u.w);
 680     if (node) {
 681         CHECK_REG_ARGS();
 682         jit_link_alist(node);
 683         jit_check_frame();
 684     }
 685     jit_dec_synth();
 686 }
 687
 688 void
 689 _jit_getarg_us(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 690 {
 691     jit_node_t          *node = NULL;
 692     assert_arg_type(v->code, jit_code_arg_s);
 693     jit_inc_synth_wp(getarg_us, u, v);
 694     if (jit_swf_p())
 695         node = jit_ldxi_us(u, JIT_FP, arg_offset(v->u.w));
 696     else if (jit_arg_reg_p(v->u.w))
 697         jit_extr_us(u, JIT_RA0 - v->u.w);
 698     else
 699         node = jit_ldxi_us(u, JIT_FP, v->u.w);
 700     if (node) {
 701         CHECK_REG_ARGS();
 702         jit_link_alist(node);
 703         jit_check_frame();
 704     }
 705     jit_dec_synth();
 706 }
 707
 708 void
 709 _jit_getarg_i(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 710 {
 711     jit_node_t          *node = NULL;
 712     assert_arg_type(v->code, jit_code_arg_i);
 713     jit_inc_synth_wp(getarg_i, u, v);
 714     if (jit_swf_p())
 715         node = jit_ldxi_i(u, JIT_FP, arg_offset(v->u.w));
 716     else if (jit_arg_reg_p(v->u.w))
 717         jit_movr(u, JIT_RA0 - v->u.w);
 718     else
 719         node = jit_ldxi_i(u, JIT_FP, v->u.w);
 720     if (node) {
 721         CHECK_REG_ARGS();
 722         jit_link_alist(node);
 723         jit_check_frame();
 724     }
 725     jit_dec_synth();
 726 }
 727
 728 void
 729 _jit_putargr(jit_state_t *_jit, jit_int32_t u, jit_node_t *v, jit_code_t code)
 730 {
 731     jit_node_t          *node = NULL;
 732     assert_putarg_type(code, v->code);
 733     jit_code_inc_synth_wp(code, u, v);
 734     if (jit_swf_p())
 735         node = jit_stxi(arg_offset(v->u.w), JIT_FP, u);
 736     else if (jit_arg_reg_p(v->u.w))
 737         jit_movr(JIT_RA0 - v->u.w, u);
 738     else
 739         node = jit_stxi(v->u.w, JIT_FP, u);
 740     if (node) {
 741         CHECK_REG_ARGS();
 742         jit_link_alist(node);
 743         jit_check_frame();
 744     }
 745     jit_dec_synth();
 746 }
 747
 748 void
 749 _jit_putargi(jit_state_t *_jit, jit_word_t u, jit_node_t *v, jit_code_t code)
 750 {
 751     jit_int32_t          regno;
 752     jit_node_t          *node = NULL;
 753     assert_putarg_type(code, v->code);
 754     jit_code_inc_synth_wp(code, u, v);
 755     if (jit_swf_p()) {
 756         regno = jit_get_reg(jit_class_gpr);
 757         jit_movi(regno, u);
 758         node = jit_stxi(arg_offset(v->u.w), JIT_FP, regno);
 759         jit_unget_reg(regno);
 760     }
 761     else if (jit_arg_reg_p(v->u.w))
 762         jit_movi(JIT_RA0 - v->u.w, u);
 763     else {
 764         regno = jit_get_reg(jit_class_gpr);
 765         jit_movi(regno, u);
 766         node = jit_stxi(v->u.w, JIT_FP, regno);
 767         jit_unget_reg(regno);
 768     }
 769     if (node) {
 770         CHECK_REG_ARGS();
 771         jit_link_alist(node);
 772         jit_check_frame();
 773     }
 774     jit_dec_synth();
 775 }
 776
 777 void
 778 _jit_getarg_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 779 {
 780     jit_node_t          *node = NULL;
 781     assert(v->code == jit_code_arg_f);
 782     jit_inc_synth_wp(getarg_f, u, v);
 783     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 784         if (jit_arg_f_reg_p(v->u.w))
 785             jit_movr_f(u, JIT_FA0 - v->u.w);
 786         else
 787             node = jit_ldxi_f(u, JIT_FP, v->u.w);
 788     }
 789     else if (jit_swf_p())
 790         node = jit_ldxi_f(u, JIT_FP, arg_offset(v->u.w));
 791     else {
 792         if (jit_arg_reg_p(v->u.w))
 793             jit_movr_w_f(u, JIT_RA0 - v->u.w);
 794         else
 795             node = jit_ldxi_f(u, JIT_FP, v->u.w);
 796     }
 797     if (node) {
 798         CHECK_REG_ARGS();
 799         jit_link_alist(node);
 800         jit_check_frame();
 801     }
 802     jit_dec_synth();
 803 }
 804
 805 void
 806 _jit_putargr_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 807 {
 808     jit_node_t          *node = NULL;
 809     assert(v->code == jit_code_arg_f);
 810     jit_inc_synth_wp(putargr_f, u, v);
 811     if (jit_cpu.abi) {
 812         if (jit_arg_f_reg_p(v->u.w))
 813             jit_movr_f(JIT_FA0 - v->u.w, u);
 814         else
 815             node = jit_stxi_f(v->u.w, JIT_FP, u);
 816     }
 817     else if (jit_swf_p())
 818         node = jit_stxi_f(arg_offset(v->u.w), JIT_FP, u);
 819     else {
 820         if (jit_arg_reg_p(v->u.w))
 821             jit_movr_f_w(JIT_RA0 - v->u.w, u);
 822         else
 823             node = jit_stxi_f(v->u.w, JIT_FP, u);
 824     }
 825     if (node) {
 826         CHECK_REG_ARGS();
 827         jit_link_alist(node);
 828         jit_check_frame();
 829     }
 830     jit_dec_synth();
 831 }
 832
 833 void
 834 _jit_putargi_f(jit_state_t *_jit, jit_float32_t u, jit_node_t *v)
 835 {
 836     jit_int32_t          regno;
 837     jit_node_t          *node = NULL;
 838     assert(v->code == jit_code_arg_f);
 839     jit_inc_synth_fp(putargi_f, u, v);
 840     if (jit_cpu.abi) {
 841         if (jit_arg_f_reg_p(v->u.w))
 842             jit_movi_f(JIT_FA0 - v->u.w, u);
 843         else {
 844             regno = jit_get_reg(jit_class_fpr);
 845             jit_movi_f(regno, u);
 846             node = jit_stxi_f(v->u.w, JIT_FP, regno);
 847             jit_unget_reg(regno);
 848         }
 849     }
 850     else if (jit_swf_p()) {
 851         regno = jit_get_reg(jit_class_fpr);
 852         jit_movi_f(regno, u);
 853         node = jit_stxi_f(arg_offset(v->u.w), JIT_FP, regno);
 854         jit_unget_reg(regno);
 855     }
 856     else {
 857         regno = jit_get_reg(jit_class_fpr);
 858         jit_movi_f(regno, u);
 859         if (jit_arg_reg_p(v->u.w))
 860             jit_movr_f_w(JIT_RA0 - v->u.w, regno);
 861         else
 862             node = jit_stxi_f(v->u.w, JIT_FP, regno);
 863         jit_unget_reg(regno);
 864     }
 865     if (node) {
 866         CHECK_REG_ARGS();
 867         jit_link_alist(node);
 868         jit_check_frame();
 869     }
 870     jit_dec_synth();
 871 }
 872
 873 void
 874 _jit_getarg_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 875 {
 876     jit_node_t          *node = NULL;
 877     assert(v->code == jit_code_arg_d);
 878     jit_inc_synth_wp(getarg_d, u, v);
 879     if (jit_cpu.abi && !(_jitc->function->self.call & jit_call_varargs)) {
 880         if (jit_arg_f_reg_p(v->u.w))
 881             jit_movr_d(u, JIT_FA0 - v->u.w);
 882         else
 883             node = jit_ldxi_d(u, JIT_FP, v->u.w);
 884     }
 885     else if (jit_swf_p())
 886         node = jit_ldxi_d(u, JIT_FP, arg_offset(v->u.w));
 887     else {
 888         if (jit_arg_reg_p(v->u.w))
 889             jit_movr_ww_d(u, JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1));
 890         else
 891             node = jit_ldxi_d(u, JIT_FP, v->u.w);
 892     }
 893     if (node) {
 894         CHECK_REG_ARGS();
 895         jit_link_alist(node);
 896         jit_check_frame();
 897     }
 898     jit_dec_synth();
 899 }
 900
 901 void
 902 _jit_putargr_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 903 {
 904     jit_node_t          *node = NULL;
 905     assert(v->code == jit_code_arg_d);
 906     jit_inc_synth_wp(putargr_d, u, v);
 907     if (jit_cpu.abi) {
 908         if (jit_arg_f_reg_p(v->u.w))
 909             jit_movr_d(JIT_FA0 - v->u.w, u);
 910         else
 911             node = jit_stxi_d(v->u.w, JIT_FP, u);
 912     }
 913     else if (jit_swf_p())
 914         node = jit_stxi_d(arg_offset(v->u.w), JIT_FP, u);
 915     else {
 916         if (jit_arg_reg_p(v->u.w))
 917             jit_movr_d_ww(JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1), u);
 918         else
 919             node = jit_stxi_d(v->u.w, JIT_FP, u);
 920     }
 921     if (node) {
 922         CHECK_REG_ARGS();
 923         jit_link_alist(node);
 924         jit_check_frame();
 925     }
 926     jit_dec_synth();
 927 }
 928
 929 void
 930 _jit_putargi_d(jit_state_t *_jit, jit_float64_t u, jit_node_t *v)
 931 {
 932     jit_int32_t          regno;
 933     jit_node_t          *node = NULL;
 934     assert(v->code == jit_code_arg_d);
 935     jit_inc_synth_dp(putargi_d, u, v);
 936     if (jit_cpu.abi) {
 937         if (jit_arg_f_reg_p(v->u.w))
 938             jit_movi_d(JIT_FA0 - v->u.w, u);
 939         else {
 940             regno = jit_get_reg(jit_class_fpr);
 941             jit_movi_d(regno, u);
 942             node = jit_stxi_d(v->u.w, JIT_FP, regno);
 943             jit_unget_reg(regno);
 944         }
 945     }
 946     else if (jit_swf_p()) {
 947         regno = jit_get_reg(jit_class_fpr);
 948         jit_movi_d(regno, u);
 949         node = jit_stxi_d(arg_offset(v->u.w), JIT_FP, regno);
 950         jit_unget_reg(regno);
 951     }
 952     else {
 953         regno = jit_get_reg(jit_class_fpr);
 954         jit_movi_d(regno, u);
 955         if (jit_arg_reg_p(v->u.w))
 956             jit_movr_d_ww(JIT_RA0 - v->u.w, JIT_RA0 - (v->u.w + 1), regno);
 957         else
 958             node = jit_stxi_d(v->u.w, JIT_FP, regno);
 959         jit_unget_reg(regno);
 960     }
 961     if (node) {
 962         CHECK_REG_ARGS();
 963         jit_link_alist(node);
 964         jit_check_frame();
 965     }
 966     jit_dec_synth();
 967 }
 968
 969 void
 970 _jit_pushargr(jit_state_t *_jit, jit_int32_t u, jit_code_t code)
 971 {
 972     assert(_jitc->function);
 973     jit_code_inc_synth_w(code, u);
 974     jit_link_prepare();
 975     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 976         jit_movr(JIT_RA0 - _jitc->function->call.argi, u);
 977         ++_jitc->function->call.argi;
 978     }
 979     else {
 980         jit_stxi(_jitc->function->call.size, JIT_SP, u);
 981         _jitc->function->call.size += sizeof(jit_word_t);
 982     }
 983     jit_dec_synth();
 984 }
 985
 986 void
 987 _jit_pushargi(jit_state_t *_jit, jit_word_t u, jit_code_t code)
 988 {
 989     jit_int32_t          regno;
 990     assert(_jitc->function);
 991     jit_code_inc_synth_w(code, u);
 992     jit_link_prepare();
 993     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 994         jit_movi(JIT_RA0 - _jitc->function->call.argi, u);
 995         ++_jitc->function->call.argi;
 996     }
 997     else {
 998         regno = jit_get_reg(jit_class_gpr);
 999         jit_movi(regno, u);
1000         jit_stxi(_jitc->function->call.size, JIT_SP, regno);
1001         jit_unget_reg(regno);
1002         _jitc->function->call.size += sizeof(jit_word_t);
1003     }
1004     jit_dec_synth();
1005 }
1006
1007 void
1008 _jit_pushargr_f(jit_state_t *_jit, jit_int32_t u)
1009 {
1010     assert(_jitc->function);
1011     jit_inc_synth_w(pushargr_f, u);
1012     jit_link_prepare();
1013     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
1014         if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
1015             jit_movr_f(JIT_FA0 - _jitc->function->call.argf, u);
1016             ++_jitc->function->call.argf;
1017             goto done;
1018         }
1019     }
1020     else {
1021         if (jit_arg_reg_p(_jitc->function->call.argi)) {
1022             jit_movr_f_w(JIT_RA0 - _jitc->function->call.argi, u);
1023             ++_jitc->function->call.argi;
1024             goto done;
1025         }
1026     }
1027     jit_stxi_f(_jitc->function->call.size, JIT_SP, u);
1028     _jitc->function->call.size += sizeof(jit_word_t);
1029 done:
1030     jit_dec_synth();
1031 }
1032
1033 void
1034 _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
1035 {
1036     jit_int32_t         regno;
1037     assert(_jitc->function);
1038     jit_inc_synth_f(pushargi_f, u);
1039     jit_link_prepare();
1040     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
1041         if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
1042             /* cannot jit_movi_f in the argument register because
1043              * float arguments are packed, and that would cause
1044              * either an assertion in debug mode, or overwritting
1045              * two registers */
1046             regno = jit_get_reg(jit_class_fpr);
1047             jit_movi_f(regno, u);
1048             jit_movr_f(JIT_FA0 - _jitc->function->call.argf, regno);
1049             jit_unget_reg(regno);
1050             ++_jitc->function->call.argf;
1051             goto done;
1052         }
1053     }
1054     else {
1055         if (jit_arg_reg_p(_jitc->function->call.argi)) {
1056             jit_movi_f_w(JIT_RA0 - _jitc->function->call.argi, u);
1057             ++_jitc->function->call.argi;
1058             goto done;
1059         }
1060     }
1061     regno = jit_get_reg(jit_class_fpr);
1062     jit_movi_f(regno, u);
1063     jit_stxi_f(_jitc->function->call.size, JIT_SP, regno);
1064     jit_unget_reg(regno);
1065     _jitc->function->call.size += sizeof(jit_word_t);
1066 done:
1067     jit_dec_synth();
1068 }
1069
1070 void
1071 _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
1072 {
1073     assert(_jitc->function);
1074     jit_inc_synth_w(pushargr_d, u);
1075     jit_link_prepare();
1076     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
1077         if (jit_arg_d_reg_p(_jitc->function->call.argf)) {
1078             if (_jitc->function->call.argf & 1)
1079                 ++_jitc->function->call.argf;
1080             jit_movr_d(JIT_FA0 - _jitc->function->call.argf, u);
1081             _jitc->function->call.argf += 2;
1082             goto done;
1083         }
1084     }
1085     else {
1086         if (_jitc->function->call.argi & 1)
1087             ++_jitc->function->call.argi;
1088         if (jit_arg_reg_p(_jitc->function->call.argi)) {
1089             jit_movr_d_ww(JIT_RA0 - _jitc->function->call.argi,
1090                           JIT_RA0 - (_jitc->function->call.argi + 1),
1091                           u);
1092             _jitc->function->call.argi += 2;
1093             goto done;
1094         }
1095     }
1096     if (_jitc->function->call.size & 7)
1097         _jitc->function->call.size += 4;
1098     jit_stxi_d(_jitc->function->call.size, JIT_SP, u);
1099     _jitc->function->call.size += sizeof(jit_float64_t);
1100 done:
1101     jit_dec_synth();
1102 }
1103
1104 void
1105 _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
1106 {
1107     jit_int32_t         regno;
1108     assert(_jitc->function);
1109     jit_inc_synth_d(pushargi_d, u);
1110     jit_link_prepare();
1111     if (jit_cpu.abi && !(_jitc->function->call.call & jit_call_varargs)) {
1112         if (jit_arg_d_reg_p(_jitc->function->call.argf)) {
1113             if (_jitc->function->call.argf & 1)
1114                 ++_jitc->function->call.argf;
1115             jit_movi_d(JIT_FA0 - _jitc->function->call.argf, u);
1116             _jitc->function->call.argf += 2;
1117             goto done;
1118         }
1119     }
1120     else {
1121         if (_jitc->function->call.argi & 1)
1122             ++_jitc->function->call.argi;
1123         if (jit_arg_reg_p(_jitc->function->call.argi)) {
1124             jit_movi_d_ww(JIT_RA0 - _jitc->function->call.argi,
1125                           JIT_RA0 - (_jitc->function->call.argi + 1),
1126                           u);
1127             _jitc->function->call.argi += 2;
1128             goto done;
1129         }
1130     }
1131     if (_jitc->function->call.size & 7)
1132         _jitc->function->call.size += 4;
1133     regno = jit_get_reg(jit_class_fpr);
1134     jit_movi_d(regno, u);
1135     jit_stxi_d(_jitc->function->call.size, JIT_SP, regno);
1136     jit_unget_reg(regno);
1137     _jitc->function->call.size += sizeof(jit_float64_t);
1138 done:
1139     jit_dec_synth();
1140 }
1141
1142 jit_bool_t
1143 _jit_regarg_p(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno)
1144 {
1145     jit_int32_t         spec;
1146     spec = jit_class(_rvs[regno].spec);
1147     if (spec & jit_class_arg) {
1148         regno = JIT_RA0 - regno;
1149         if (regno >= 0 && regno < node->v.w)
1150             return (1);
1151         if (jit_cpu.abi && spec & jit_class_fpr) {
1152             regno = JIT_FA0 - regno;
1153             if (regno >= 0 && regno < node->w.w)
1154                 return (1);
1155         }
1156     }
1157
1158     return (0);
1159 }
1160
1161 void
1162 _jit_finishr(jit_state_t *_jit, jit_int32_t r0)
1163 {
1164     jit_node_t          *node;
1165     assert(_jitc->function);
1166     jit_inc_synth_w(finishr, r0);
1167     if (_jitc->function->self.alen < _jitc->function->call.size)
1168         _jitc->function->self.alen = _jitc->function->call.size;
1169     node = jit_callr(r0);
1170     node->v.w = _jitc->function->self.argi;
1171     node->w.w = _jitc->function->call.argf;
1172     _jitc->function->call.argi = _jitc->function->call.argf =
1173         _jitc->function->call.size = 0;
1174     _jitc->prepare = 0;
1175     jit_dec_synth();
1176 }
1177
1178 jit_node_t *
1179 _jit_finishi(jit_state_t *_jit, jit_pointer_t i0)
1180 {
1181     jit_node_t          *node;
1182     assert(_jitc->function);
1183     jit_inc_synth_w(finishi, (jit_word_t)i0);
1184     if (_jitc->function->self.alen < _jitc->function->call.size)
1185         _jitc->function->self.alen = _jitc->function->call.size;
1186     node = jit_calli(i0);
1187     node->v.w = _jitc->function->call.argi;
1188     node->w.w = _jitc->function->call.argf;
1189     _jitc->function->call.argi = _jitc->function->call.argf =
1190         _jitc->function->call.size = 0;
1191     _jitc->prepare = 0;
1192     jit_dec_synth();
1193     return (node);
1194 }
1195
1196 void
1197 _jit_retval_c(jit_state_t *_jit, jit_int32_t r0)
1198 {
1199     jit_inc_synth_w(retval_c, r0);
1200     jit_extr_c(r0, JIT_RET);
1201     jit_dec_synth();
1202 }
1203
1204 void
1205 _jit_retval_uc(jit_state_t *_jit, jit_int32_t r0)
1206 {
1207     jit_inc_synth_w(retval_uc, r0);
1208     jit_extr_uc(r0, JIT_RET);
1209     jit_dec_synth();
1210 }
1211
1212 void
1213 _jit_retval_s(jit_state_t *_jit, jit_int32_t r0)
1214 {
1215     jit_inc_synth_w(retval_s, r0);
1216     jit_extr_s(r0, JIT_RET);
1217     jit_dec_synth();
1218 }
1219
1220 void
1221 _jit_retval_us(jit_state_t *_jit, jit_int32_t r0)
1222 {
1223     jit_inc_synth_w(retval_us, r0);
1224     jit_extr_us(r0, JIT_RET);
1225     jit_dec_synth();
1226 }
1227
1228 void
1229 _jit_retval_i(jit_state_t *_jit, jit_int32_t r0)
1230 {
1231     jit_inc_synth_w(retval_i, r0);
1232     if (r0 != JIT_RET)
1233         jit_movr(r0, JIT_RET);
1234     jit_dec_synth();
1235 }
1236
1237 void
1238 _jit_retval_f(jit_state_t *_jit, jit_int32_t r0)
1239 {
1240     jit_inc_synth_w(retval_f, r0);
1241     if (jit_cpu.abi) {
1242         if (r0 != JIT_FRET)
1243             jit_movr_f(r0, JIT_FRET);
1244     }
1245     else if (r0 != JIT_RET)
1246         jit_movr_w_f(r0, JIT_RET);
1247     jit_dec_synth();
1248 }
1249
1250 void
1251 _jit_retval_d(jit_state_t *_jit, jit_int32_t r0)
1252 {
1253     jit_inc_synth_w(retval_d, r0);
1254     if (jit_cpu.abi) {
1255         if (r0 != JIT_FRET)
1256             jit_movr_d(r0, JIT_FRET);
1257     }
1258     else if (r0 != JIT_RET)
1259         jit_movr_ww_d(r0, JIT_RET, _R1);
1260     jit_dec_synth();
1261 }
1262
1263 jit_pointer_t
1264 _emit_code(jit_state_t *_jit)
1265 {
1266     jit_node_t          *node;
1267     jit_node_t          *temp;
1268     jit_word_t           word;
1269     jit_int32_t          value;
1270     jit_int32_t          offset;
1271     struct {
1272         jit_node_t      *node;
1273         jit_uint8_t     *data;
1274         jit_word_t       word;
1275         jit_function_t   func;
1276 #if DEVEL_DISASSEMBLER
1277         jit_word_t       prevw;
1278 #endif
1279         jit_uword_t      thumb;
1280 #if DISASSEMBLER
1281         jit_int32_t      info_offset;
1282 #endif
1283         jit_int32_t      const_offset;
1284         jit_int32_t      patch_offset;
1285     } undo;
1286 #if DEVEL_DISASSEMBLER
1287     jit_word_t           prevw;
1288 #endif
1289
1290     _jitc->function = NULL;
1291     _jitc->thumb = 0;
1292
1293     jit_reglive_setup();
1294
1295     _jitc->consts.data = NULL;
1296     _jitc->consts.offset = _jitc->consts.length = 0;
1297
1298     undo.word = 0;
1299     undo.node = NULL;
1300     undo.data = NULL;
1301     undo.thumb = 0;
1302 #if DISASSEMBLER
1303     undo.info_offset =
1304 #endif
1305         undo.const_offset = undo.patch_offset = 0;
1306 #  define assert_data(node)             /**/
1307 #define case_rr(name, type)                                             \
1308             case jit_code_##name##r##type:                              \
1309                 name##r##type(rn(node->u.w), rn(node->v.w));            \
1310                 break
1311 #define case_rw(name, type)                                             \
1312             case jit_code_##name##i##type:                              \
1313                 name##i##type(rn(node->u.w), node->v.w);                \
1314                 break
1315 #define case_vv(name, type)                                             \
1316             case jit_code_##name##r##type:                              \
1317                 if (jit_swf_p())                                        \
1318                     swf_##name##r##type(rn(node->u.w), rn(node->v.w));  \
1319                 else                                                    \
1320                     vfp_##name##r##type(rn(node->u.w), rn(node->v.w));  \
1321                 break
1322 #define case_vw(name, type)                                             \
1323             case jit_code_##name##i##type:                              \
1324                 if (jit_swf_p())                                        \
1325                     swf_##name##i##type(rn(node->u.w), node->v.w);      \
1326                 else                                                    \
1327                     vfp_##name##i##type(rn(node->u.w), node->v.w);      \
1328                 break
1329 #define case_wr(name, type)                                             \
1330             case jit_code_##name##i##type:                              \
1331                 name##i##type(node->u.w, rn(node->v.w));                \
1332                 break
1333 #define case_wv(name, type)                                             \
1334             case jit_code_##name##i##type:                              \
1335                 if (jit_swf_p())                                        \
1336                     swf_##name##i##type(node->u.w, rn(node->v.w));      \
1337                 else                                                    \
1338                     vfp_##name##i##type(node->u.w, rn(node->v.w));      \
1339                 break
1340 #define case_rrr(name, type)                                            \
1341             case jit_code_##name##r##type:                              \
1342                 name##r##type(rn(node->u.w),                            \
1343                               rn(node->v.w), rn(node->w.w));            \
1344                 break
1345 #define case_rqr(name, type)                                            \
1346             case jit_code_##name##r##type:                              \
1347                 if (jit_swf_p())                                        \
1348                     swf_##name##r##type(rn(node->u.w), rn(node->v.q.l), \
1349                                         rn(node->v.q.h), rn(node->w.w));\
1350                 else                                                    \
1351                     vfp_##name##r##type(rn(node->u.w), rn(node->v.q.l), \
1352                                         rn(node->v.q.h), rn(node->w.w));\
1353             case jit_code_##name##i##type:                              \
1354                 break
1355 #define case_rrrr(name, type)                                           \
1356             case jit_code_##name##r##type:                              \
1357                 name##r##type(rn(node->u.q.l), rn(node->u.q.h),         \
1358                               rn(node->v.w), rn(node->w.w));            \
1359                 break
1360 #define case_vvv(name, type)                                            \
1361             case jit_code_##name##r##type:                              \
1362                 if (jit_swf_p())                                        \
1363                     swf_##name##r##type(rn(node->u.w),                  \
1364                                         rn(node->v.w), rn(node->w.w));  \
1365                 else                                                    \
1366                     vfp_##name##r##type(rn(node->u.w),                  \
1367                                         rn(node->v.w), rn(node->w.w));  \
1368                 break
1369 #define case_rrw(name, type)                                            \
1370             case jit_code_##name##i##type:                              \
1371                 name##i##type(rn(node->u.w), rn(node->v.w), node->w.w); \
1372                 break
1373 #define case_rrrw(name, type)                                           \
1374             case jit_code_##name##i##type:                              \
1375                 name##i##type(rn(node->u.q.l), rn(node->u.q.h),         \
1376                               rn(node->v.w), node->w.w);                \
1377                 break
1378 #define case_vvw(name, type)                                            \
1379             case jit_code_##name##i##type:                              \
1380                 if (jit_swf_p())                                        \
1381                     swf_##name##i##type(rn(node->u.w),                  \
1382                                         rn(node->v.w), node->w.w);      \
1383                 else                                                    \
1384                     vfp_##name##i##type(rn(node->u.w),                  \
1385                                         rn(node->v.w), node->w.w);      \
1386                 break
1387 #define case_vvf(name)                                                  \
1388             case jit_code_##name##i_f:                                  \
1389                 assert_data(node);                                      \
1390                 if (jit_swf_p())                                        \
1391                     swf_##name##i_f(rn(node->u.w), rn(node->v.w),       \
1392                                     node->w.f);                         \
1393                 else                                                    \
1394                     vfp_##name##i_f(rn(node->u.w), rn(node->v.w),       \
1395                                     node->w.f);                         \
1396                 break
1397 #define case_vvd(name)                                                  \
1398             case jit_code_##name##i_d:                                  \
1399                 assert_data(node);                                      \
1400                 if (jit_swf_p())                                        \
1401                     swf_##name##i_d(rn(node->u.w), rn(node->v.w),       \
1402                                     node->w.d);                         \
1403                 else                                                    \
1404                     vfp_##name##i_d(rn(node->u.w), rn(node->v.w),       \
1405                                     node->w.d);                         \
1406                 break
1407 #define case_wrr(name, type)                                            \
1408             case jit_code_##name##i##type:                              \
1409                 name##i##type(node->u.w, rn(node->v.w), rn(node->w.w)); \
1410                 break
1411 #define case_wvv(name, type)                                            \
1412             case jit_code_##name##i##type:                              \
1413                 if (jit_swf_p())                                        \
1414                     swf_##name##i##type(node->u.w,                      \
1415                                         rn(node->v.w), rn(node->w.w));  \
1416                 else                                                    \
1417                     vfp_##name##i##type(node->u.w,                      \
1418                                         rn(node->v.w), rn(node->w.w));  \
1419                 break
1420 #define case_brr(name, type)                                            \
1421             case jit_code_##name##r##type:                              \
1422                 temp = node->u.n;                                       \
1423                 assert(temp->code == jit_code_label ||                  \
1424                        temp->code == jit_code_epilog);                  \
1425                 if (temp->flag & jit_flag_patch)                        \
1426                     name##r##type(temp->u.w, rn(node->v.w),             \
1427                                   rn(node->w.w));                       \
1428                 else {                                                  \
1429                     word = name##r##type(_jit->pc.w,                    \
1430                                          rn(node->v.w), rn(node->w.w)); \
1431                     patch(word, node, arm_patch_jump);                  \
1432                 }                                                       \
1433                 break
1434 #define case_bvv(name, type)                                            \
1435             case jit_code_##name##r##type:                              \
1436                 temp = node->u.n;                                       \
1437                 assert(temp->code == jit_code_label ||                  \
1438                        temp->code == jit_code_epilog);                  \
1439                 if (temp->flag & jit_flag_patch) {                      \
1440                     if (jit_swf_p())                                    \
1441                         swf_##name##r##type(temp->u.w, rn(node->v.w),   \
1442                                             rn(node->w.w));             \
1443                     else                                                \
1444                         vfp_##name##r##type(temp->u.w, rn(node->v.w),   \
1445                                             rn(node->w.w));             \
1446                 }                                                       \
1447                 else {                                                  \
1448                     if (jit_swf_p())                                    \
1449                         word = swf_##name##r##type(_jit->pc.w,          \
1450                                                    rn(node->v.w),       \
1451                                                    rn(node->w.w));      \
1452                     else                                                \
1453                         word = vfp_##name##r##type(_jit->pc.w,          \
1454                                                    rn(node->v.w),       \
1455                                                    rn(node->w.w));      \
1456                     patch(word, node, arm_patch_jump);                  \
1457                 }                                                       \
1458                 break
1459 #define case_brw(name, type)                                            \
1460             case jit_code_##name##i##type:                              \
1461                 temp = node->u.n;                                       \
1462                 assert(temp->code == jit_code_label ||                  \
1463                        temp->code == jit_code_epilog);                  \
1464                 if (temp->flag & jit_flag_patch)                        \
1465                     name##i##type(temp->u.w,                            \
1466                                   rn(node->v.w), node->w.w);            \
1467                 else {                                                  \
1468                     word = name##i##type(_jit->pc.w,                    \
1469                                          rn(node->v.w), node->w.w);     \
1470                     patch(word, node, arm_patch_jump);                  \
1471                 }                                                       \
1472                 break;
1473 #define case_bvf(name)                                                  \
1474             case jit_code_##name##i_f:                                  \
1475                 temp = node->u.n;                                       \
1476                 assert(temp->code == jit_code_label ||                  \
1477                        temp->code == jit_code_epilog);                  \
1478                 if (temp->flag & jit_flag_patch) {                      \
1479                     if (jit_swf_p())                                    \
1480                         swf_##name##i_f(temp->u.w, rn(node->v.w),       \
1481                                         node->w.f);                     \
1482                     else                                                \
1483                         vfp_##name##i_f(temp->u.w, rn(node->v.w),       \
1484                                         node->w.f);                     \
1485                 }                                                       \
1486                 else {                                                  \
1487                     if (jit_swf_p())                                    \
1488                         word = swf_##name##i_f(_jit->pc.w,              \
1489                                                rn(node->v.w),           \
1490                                                node->w.f);              \
1491                     else                                                \
1492                         word = vfp_##name##i_f(_jit->pc.w,              \
1493                                                rn(node->v.w),           \
1494                                                node->w.f);              \
1495                     patch(word, node, arm_patch_jump);                  \
1496                 }                                                       \
1497                 break
1498 #define case_bvd(name)                                                  \
1499             case jit_code_##name##i_d:                                  \
1500                 temp = node->u.n;                                       \
1501                 assert(temp->code == jit_code_label ||                  \
1502                        temp->code == jit_code_epilog);                  \
1503                 if (temp->flag & jit_flag_patch) {                      \
1504                     if (jit_swf_p())                                    \
1505                         swf_##name##i_d(temp->u.w, rn(node->v.w),       \
1506                                         node->w.d);                     \
1507                     else                                                \
1508                         vfp_##name##i_d(temp->u.w, rn(node->v.w),       \
1509                                         node->w.d);                     \
1510                 }                                                       \
1511                 else {                                                  \
1512                     if (jit_swf_p())                                    \
1513                         word = swf_##name##i_d(_jit->pc.w,              \
1514                                                rn(node->v.w),           \
1515                                                node->w.d);              \
1516                     else                                                \
1517                         word = vfp_##name##i_d(_jit->pc.w,              \
1518                                                rn(node->v.w),           \
1519                                                node->w.d);              \
1520                     patch(word, node, arm_patch_jump);                  \
1521                 }                                                       \
1522                 break
1523 #if DEVEL_DISASSEMBLER
1524     prevw = _jit->pc.w;
1525 #endif
1526     for (node = _jitc->head; node; node = node->next) {
1527         if (_jit->pc.uc >= _jitc->code.end)
1528             return (NULL);
1529
1530 #if DEVEL_DISASSEMBLER
1531         node->offset = (jit_uword_t)_jit->pc.w - (jit_uword_t)prevw;
1532         prevw = _jit->pc.w;
1533 #endif
1534         value = jit_classify(node->code);
1535         jit_regarg_set(node, value);
1536         switch (node->code) {
1537             case jit_code_align:
1538                 /* Must align to a power of two */
1539                 assert(!(node->u.w & (node->u.w - 1)));
1540                 if ((word = _jit->pc.w & (node->u.w - 1)))
1541                     nop(node->u.w - word);
1542                 break;
1543             case jit_code_skip:
1544                 if (jit_thumb_p())
1545                     nop((node->u.w + 1) & ~1);
1546                 else
1547                     nop((node->u.w + 3) & ~3);
1548                 break;
1549             case jit_code_note:         case jit_code_name:
1550                 if (must_align_p(node->next))
1551                     nop(2);
1552                 node->u.w = _jit->pc.w;
1553                 break;
1554             case jit_code_label:
1555                 if (must_align_p(node->next))
1556                     nop(2);
1557                 /* remember label is defined */
1558                 node->flag |= jit_flag_patch;
1559                 node->u.w = _jit->pc.w;
1560                 break;
1561                 case_rrr(add,);
1562                 case_rrw(add,);
1563                 case_rrr(addc,);
1564                 case_rrw(addc,);
1565                 case_rrr(addx,);
1566                 case_rrw(addx,);
1567                 case_rrr(sub,);
1568                 case_rrw(sub,);
1569                 case_rrr(subc,);
1570                 case_rrw(subc,);
1571                 case_rrr(subx,);
1572                 case_rrw(subx,);
1573                 case_rrw(rsb,);
1574                 case_rrr(mul,);
1575                 case_rrw(mul,);
1576                 case_rrr(hmul,);
1577                 case_rrw(hmul,);
1578                 case_rrr(hmul, _u);
1579                 case_rrw(hmul, _u);
1580                 case_rrrr(qmul,);
1581                 case_rrrw(qmul,);
1582                 case_rrrr(qmul, _u);
1583                 case_rrrw(qmul, _u);
1584                 case_rrr(div,);
1585                 case_rrw(div,);
1586                 case_rrr(div, _u);
1587                 case_rrw(div, _u);
1588                 case_rrrr(qdiv,);
1589                 case_rrrw(qdiv,);
1590                 case_rrrr(qdiv, _u);
1591                 case_rrrw(qdiv, _u);
1592                 case_rrr(rem,);
1593                 case_rrw(rem,);
1594                 case_rrr(rem, _u);
1595                 case_rrw(rem, _u);
1596                 case_rrr(lsh,);
1597                 case_rrw(lsh,);
1598 #define qlshr(r0, r1, r2, r3)   fallback_qlshr(r0, r1, r2, r3)
1599 #define qlshi(r0, r1, r2, i0)   fallback_qlshi(r0, r1, r2, i0)
1600 #define qlshr_u(r0, r1, r2, r3) fallback_qlshr_u(r0, r1, r2, r3)
1601 #define qlshi_u(r0, r1, r2, i0) fallback_qlshi_u(r0, r1, r2, i0)
1602                 case_rrrr(qlsh,);
1603                 case_rrrw(qlsh,);
1604                 case_rrrr(qlsh, _u);
1605                 case_rrrw(qlsh, _u);
1606                 case_rrr(rsh,);
1607                 case_rrw(rsh,);
1608                 case_rrr(rsh, _u);
1609                 case_rrw(rsh, _u);
1610 #define qrshr(r0, r1, r2, r3)   fallback_qrshr(r0, r1, r2, r3)
1611 #define qrshi(r0, r1, r2, i0)   fallback_qrshi(r0, r1, r2, i0)
1612 #define qrshr_u(r0, r1, r2, r3) fallback_qrshr_u(r0, r1, r2, r3)
1613 #define qrshi_u(r0, r1, r2, i0) fallback_qrshi_u(r0, r1, r2, i0)
1614                 case_rrrr(qrsh,);
1615                 case_rrrw(qrsh,);
1616                 case_rrrr(qrsh, _u);
1617                 case_rrrw(qrsh, _u);
1618                 case_rrr(lrot,);
1619                 case_rrw(lrot,);
1620                 case_rrr(rrot,);
1621                 case_rrw(rrot,);
1622                 case_rr(neg,);
1623                 case_rr(com,);
1624                 case_rr(clo,);
1625                 case_rr(clz,);
1626                 case_rr(cto,);
1627                 case_rr(ctz,);
1628                 case_rr(rbit,);
1629                 case_rr(popcnt,);
1630                 case_rrr(and,);
1631                 case_rrw(and,);
1632                 case_rrr(or,);
1633                 case_rrw(or,);
1634                 case_rrr(xor,);
1635                 case_rrw(xor,);
1636                 case_vv(trunc, _f_i);
1637                 case_vv(trunc, _d_i);
1638                 case_rr(ld, _c);
1639                 case_rw(ld, _c);
1640                 case_rr(ld, _uc);
1641                 case_rw(ld, _uc);
1642                 case_rr(ld, _s);
1643                 case_rw(ld, _s);
1644                 case_rr(ld, _us);
1645                 case_rw(ld, _us);
1646                 case_rr(ld, _i);
1647                 case_rw(ld, _i);
1648                 case_rrr(ldx, _c);
1649                 case_rrw(ldx, _c);
1650                 case_rrr(ldx, _uc);
1651                 case_rrw(ldx, _uc);
1652                 case_rrr(ldx, _s);
1653                 case_rrw(ldx, _s);
1654                 case_rrr(ldx, _us);
1655                 case_rrw(ldx, _us);
1656                 case_rrr(ldx, _i);
1657                 case_rrw(ldx, _i);
1658             case jit_code_unldr:
1659                 unldr(rn(node->u.w), rn(node->v.w), node->w.w);
1660                 break;
1661             case jit_code_unldi:
1662                 unldi(rn(node->u.w), node->v.w, node->w.w);
1663                 break;
1664             case jit_code_unldr_u:
1665                 unldr_u(rn(node->u.w), rn(node->v.w), node->w.w);
1666                 break;
1667             case jit_code_unldi_u:
1668                 unldi_u(rn(node->u.w), node->v.w, node->w.w);
1669                 break;
1670                 case_rr(st, _c);
1671                 case_wr(st, _c);
1672                 case_rr(st, _s);
1673                 case_wr(st, _s);
1674                 case_rr(st, _i);
1675                 case_wr(st, _i);
1676                 case_rrr(stx, _c);
1677                 case_wrr(stx, _c);
1678                 case_rrr(stx, _s);
1679                 case_wrr(stx, _s);
1680                 case_rrr(stx, _i);
1681                 case_wrr(stx, _i);
1682             case jit_code_unstr:
1683                 unstr(rn(node->u.w), rn(node->v.w), node->w.w);
1684                 break;
1685             case jit_code_unsti:
1686                 unsti(node->u.w, rn(node->v.w), node->w.w);
1687                 break;
1688                 case_rr(hton, _us);
1689                 case_rr(hton, _ui);
1690                 case_rr(bswap, _us);
1691                 case_rr(bswap, _ui);
1692             case jit_code_extr:
1693                 extr(rn(node->u.w), rn(node->v.w), node->w.q.l, node->w.q.h);
1694                 break;
1695             case jit_code_extr_u:
1696                 extr_u(rn(node->u.w), rn(node->v.w), node->w.q.l, node->w.q.h);
1697                 break;
1698             case jit_code_depr:
1699                 depr(rn(node->u.w), rn(node->v.w), node->w.q.l, node->w.q.h);
1700                 break;
1701             case jit_code_depi:
1702                 depi(rn(node->u.w), node->v.w, node->w.q.l, node->w.q.h);
1703                 break;
1704                 case_rr(ext, _c);
1705                 case_rr(ext, _uc);
1706                 case_rr(ext, _s);
1707                 case_rr(ext, _us);
1708             case jit_code_casr:
1709                 casr(rn(node->u.w), rn(node->v.w),
1710                      rn(node->w.q.l), rn(node->w.q.h));
1711                 break;
1712             case jit_code_casi:
1713                 casi(rn(node->u.w), node->v.w,
1714                      rn(node->w.q.l), rn(node->w.q.h));
1715                 break;
1716                 case_rr(mov,);
1717                 case_rrr(movn,);
1718                 case_rrr(movz,);
1719             case jit_code_movi:
1720                 if (node->flag & jit_flag_node) {
1721                     temp = node->v.n;
1722                     if (temp->code == jit_code_data ||
1723                         (temp->code == jit_code_label &&
1724                          (temp->flag & jit_flag_patch)))
1725                         movi(rn(node->u.w), temp->u.w);
1726                     else {
1727                         assert(temp->code == jit_code_label ||
1728                                temp->code == jit_code_epilog);
1729                         word = movi_p(rn(node->u.w), temp->u.w);
1730                         patch(word, node, arm_patch_word);
1731                     }
1732                 }
1733                 else
1734                     movi(rn(node->u.w), node->v.w);
1735                 break;
1736                 case_rrr(lt,);
1737                 case_rrw(lt,);
1738                 case_rrr(lt, _u);
1739                 case_rrw(lt, _u);
1740                 case_rrr(le,);
1741                 case_rrw(le,);
1742                 case_rrr(le, _u);
1743                 case_rrw(le, _u);
1744                 case_rrr(eq,);
1745                 case_rrw(eq,);
1746                 case_rrr(ge,);
1747                 case_rrw(ge,);
1748                 case_rrr(ge, _u);
1749                 case_rrw(ge, _u);
1750                 case_rrr(gt,);
1751                 case_rrw(gt,);
1752                 case_rrr(gt, _u);
1753                 case_rrw(gt, _u);
1754                 case_rrr(ne,);
1755                 case_rrw(ne,);
1756                 case_brr(blt,);
1757                 case_brw(blt,);
1758                 case_brr(blt, _u);
1759                 case_brw(blt, _u);
1760                 case_brr(ble,);
1761                 case_brw(ble,);
1762                 case_brr(ble, _u);
1763                 case_brw(ble, _u);
1764                 case_brr(beq,);
1765                 case_brw(beq,);
1766                 case_brr(bge,);
1767                 case_brw(bge,);
1768                 case_brr(bge, _u);
1769                 case_brw(bge, _u);
1770                 case_brr(bgt,);
1771                 case_brw(bgt,);
1772                 case_brr(bgt, _u);
1773                 case_brw(bgt, _u);
1774                 case_brr(bne,);
1775                 case_brw(bne,);
1776                 case_brr(boadd,);
1777                 case_brw(boadd,);
1778                 case_brr(boadd, _u);
1779                 case_brw(boadd, _u);
1780                 case_brr(bxadd,);
1781                 case_brw(bxadd,);
1782                 case_brr(bxadd, _u);
1783                 case_brw(bxadd, _u);
1784                 case_brr(bosub,);
1785                 case_brw(bosub,);
1786                 case_brr(bosub, _u);
1787                 case_brw(bosub, _u);
1788                 case_brr(bxsub,);
1789                 case_brw(bxsub,);
1790                 case_brr(bxsub, _u);
1791                 case_brw(bxsub, _u);
1792                 case_brr(bms,);
1793                 case_brw(bms,);
1794                 case_brr(bmc,);
1795                 case_brw(bmc,);
1796                 case_vvv(add, _f);
1797                 case_vvf(add);
1798                 case_vvv(sub, _f);
1799                 case_vvf(sub);
1800                 case_vvf(rsb);
1801                 case_vvv(mul, _f);
1802                 case_vvf(mul);
1803                 case_vvv(div, _f);
1804                 case_vvf(div);
1805                 case_vv(abs, _f);
1806                 case_vv(neg, _f);
1807                 case_vv(sqrt, _f);
1808                 case_rqr(fma, _f);
1809                 case_rqr(fms, _f);
1810                 case_rqr(fnma, _f);
1811                 case_rqr(fnms, _f);
1812                 case_vv(ext, _f);
1813                 case_vv(ld, _f);
1814                 case_vw(ld, _f);
1815                 case_vvv(ldx, _f);
1816                 case_vvw(ldx, _f);
1817             case jit_code_unldr_x:
1818                 if (jit_swf_p())
1819                     swf_unldr_x(rn(node->u.w), rn(node->v.w), node->w.w);
1820                 else
1821                     vfp_unldr_x(rn(node->u.w), rn(node->v.w), node->w.w);
1822                 break;
1823             case jit_code_unldi_x:
1824                 if (jit_swf_p())
1825                     swf_unldi_x(rn(node->u.w), node->v.w, node->w.w);
1826                 else
1827                     vfp_unldi_x(rn(node->u.w), node->v.w, node->w.w);
1828                 break;
1829                 case_vv(st, _f);
1830                 case_wv(st, _f);
1831                 case_vvv(stx, _f);
1832                 case_wvv(stx, _f);
1833             case jit_code_unstr_x:
1834                 if (jit_swf_p())
1835                     swf_unstr_x(rn(node->u.w), rn(node->v.w), node->w.w);
1836                 else
1837                     vfp_unstr_x(rn(node->u.w), rn(node->v.w), node->w.w);
1838                 break;
1839             case jit_code_unsti_x:
1840                 if (jit_swf_p())
1841                     swf_unsti_x(node->u.w, rn(node->v.w), node->w.w);
1842                 else
1843                     vfp_unsti_x(node->u.w, rn(node->v.w), node->w.w);
1844                 break;
1845                 case_vv(mov, _f);
1846             case jit_code_movi_f:
1847                 assert_data(node);
1848                 if (jit_swf_p())
1849                     swf_movi_f(rn(node->u.w), node->v.f);
1850                 else
1851                     vfp_movi_f(rn(node->u.w), node->v.f);
1852                 break;
1853                 case_vv(ext, _d_f);
1854                 case_vvv(lt, _f);
1855                 case_vvf(lt);
1856                 case_vvv(le, _f);
1857                 case_vvf(le);
1858                 case_vvv(eq, _f);
1859                 case_vvf(eq);
1860                 case_vvv(ge, _f);
1861                 case_vvf(ge);
1862                 case_vvv(gt, _f);
1863                 case_vvf(gt);
1864                 case_vvv(ne, _f);
1865                 case_vvf(ne);
1866                 case_vvv(unlt, _f);
1867                 case_vvf(unlt);
1868                 case_vvv(unle, _f);
1869                 case_vvf(unle);
1870                 case_vvv(uneq, _f);
1871                 case_vvf(uneq);
1872                 case_vvv(unge, _f);
1873                 case_vvf(unge);
1874                 case_vvv(ungt, _f);
1875                 case_vvf(ungt);
1876                 case_vvv(ltgt, _f);
1877                 case_vvf(ltgt);
1878                 case_vvv(ord, _f);
1879                 case_vvf(ord);
1880                 case_vvv(unord, _f);
1881                 case_vvf(unord);
1882                 case_bvv(blt, _f);
1883                 case_bvf(blt);
1884                 case_bvv(ble, _f);
1885                 case_bvf(ble);
1886                 case_bvv(beq, _f);
1887                 case_bvf(beq);
1888                 case_bvv(bge, _f);
1889                 case_bvf(bge);
1890                 case_bvv(bgt, _f);
1891                 case_bvf(bgt);
1892                 case_bvv(bne, _f);
1893                 case_bvf(bne);
1894                 case_bvv(bunlt, _f);
1895                 case_bvf(bunlt);
1896                 case_bvv(bunle, _f);
1897                 case_bvf(bunle);
1898                 case_bvv(buneq, _f);
1899                 case_bvf(buneq);
1900                 case_bvv(bunge, _f);
1901                 case_bvf(bunge);
1902                 case_bvv(bungt, _f);
1903                 case_bvf(bungt);
1904                 case_bvv(bltgt, _f);
1905                 case_bvf(bltgt);
1906                 case_bvv(bord, _f);
1907                 case_bvf(bord);
1908                 case_bvv(bunord, _f);
1909                 case_bvf(bunord);
1910                 case_vvv(add, _d);
1911                 case_vvd(add);
1912                 case_vvv(sub, _d);
1913                 case_vvd(sub);
1914                 case_vvd(rsb);
1915                 case_vvv(mul, _d);
1916                 case_vvd(mul);
1917                 case_vvv(div, _d);
1918                 case_vvd(div);
1919                 case_vv(abs, _d);
1920                 case_vv(neg, _d);
1921                 case_vv(sqrt, _d);
1922                 case_rqr(fma, _d);
1923                 case_rqr(fms, _d);
1924                 case_rqr(fnma, _d);
1925                 case_rqr(fnms, _d);
1926                 case_vv(ext, _d);
1927                 case_vv(ld, _d);
1928                 case_vw(ld, _d);
1929                 case_vvv(ldx, _d);
1930                 case_vvw(ldx, _d);
1931                 case_vv(st, _d);
1932                 case_wv(st, _d);
1933                 case_vvv(stx, _d);
1934                 case_wvv(stx, _d);
1935                 case_vv(mov, _d);
1936             case jit_code_movi_d:
1937                 assert_data(node);
1938                 if (jit_swf_p())
1939                     swf_movi_d(rn(node->u.w), node->v.d);
1940                 else
1941                     vfp_movi_d(rn(node->u.w), node->v.d);
1942                 break;
1943                 case_vv(ext, _f_d);
1944                 case_vvv(lt, _d);
1945                 case_vvd(lt);
1946                 case_vvv(le, _d);
1947                 case_vvd(le);
1948                 case_vvv(eq, _d);
1949                 case_vvd(eq);
1950                 case_vvv(ge, _d);
1951                 case_vvd(ge);
1952                 case_vvv(gt, _d);
1953                 case_vvd(gt);
1954                 case_vvv(ne, _d);
1955                 case_vvd(ne);
1956                 case_vvv(unlt, _d);
1957                 case_vvd(unlt);
1958                 case_vvv(unle, _d);
1959                 case_vvd(unle);
1960                 case_vvv(uneq, _d);
1961                 case_vvd(uneq);
1962                 case_vvv(unge, _d);
1963                 case_vvd(unge);
1964                 case_vvv(ungt, _d);
1965                 case_vvd(ungt);
1966                 case_vvv(ltgt, _d);
1967                 case_vvd(ltgt);
1968                 case_vvv(ord, _d);
1969                 case_vvd(ord);
1970                 case_vvv(unord, _d);
1971                 case_vvd(unord);
1972                 case_bvv(blt, _d);
1973                 case_bvd(blt);
1974                 case_bvv(ble, _d);
1975                 case_bvd(ble);
1976                 case_bvv(beq, _d);
1977                 case_bvd(beq);
1978                 case_bvv(bge, _d);
1979                 case_bvd(bge);
1980                 case_bvv(bgt, _d);
1981                 case_bvd(bgt);
1982                 case_bvv(bne, _d);
1983                 case_bvd(bne);
1984                 case_bvv(bunlt, _d);
1985                 case_bvd(bunlt);
1986                 case_bvv(bunle, _d);
1987                 case_bvd(bunle);
1988                 case_bvv(buneq, _d);
1989                 case_bvd(buneq);
1990                 case_bvv(bunge, _d);
1991                 case_bvd(bunge);
1992                 case_bvv(bungt, _d);
1993                 case_bvd(bungt);
1994                 case_bvv(bltgt, _d);
1995                 case_bvd(bltgt);
1996                 case_bvv(bord, _d);
1997                 case_bvd(bord);
1998                 case_bvv(bunord, _d);
1999                 case_bvd(bunord);
2000             case jit_code_jmpr:
2001                 jit_check_frame();
2002                 jmpr(rn(node->u.w));
2003                 flush_consts();
2004                 break;
2005             case jit_code_jmpi:
2006                 if (node->flag & jit_flag_node) {
2007                     temp = node->u.n;
2008                     assert(temp->code == jit_code_label ||
2009                            temp->code == jit_code_epilog);
2010                     if (temp->flag & jit_flag_patch)
2011                         jmpi(temp->u.w);
2012                     else {
2013                         word = _jit->code.length -
2014                             (_jit->pc.uc - _jit->code.ptr);
2015                         if (jit_thumb_p())      word >>= 1;
2016                         else                    word >>= 2;
2017                         word -= 2;
2018                         value = _s24P(word);
2019                         word = jmpi_p(_jit->pc.w, value);
2020                         patch(word, node, value ?
2021                               arm_patch_jump : arm_patch_word);
2022                     }
2023                 }
2024                 else {
2025                     jit_check_frame();
2026                     jmpi(node->u.w);
2027                 }
2028                 flush_consts();
2029                 break;
2030             case jit_code_callr:
2031                 jit_check_frame();
2032                 callr(rn(node->u.w));
2033                 break;
2034             case jit_code_calli:
2035                 if (node->flag & jit_flag_node) {
2036                     CHECK_RETURN();
2037                     temp = node->u.n;
2038                     assert(temp->code == jit_code_label ||
2039                            temp->code == jit_code_epilog);
2040                     if (temp->flag & jit_flag_patch)
2041                         calli(temp->u.w, 0);
2042                     else {
2043                         word = _jit->code.length -
2044                             (_jit->pc.uc - _jit->code.ptr);
2045                         if (jit_exchange_p())
2046                             word -= 8;
2047                         if (jit_thumb_p())      word >>= 1;
2048                         else                    word >>= 2;
2049                         word -= 2;
2050                         value = _s24P(word);
2051                         word = calli_p(_jit->pc.w, value);
2052                         patch(word, node, value ?
2053                               arm_patch_call : arm_patch_word);
2054                     }
2055                 }
2056                 else {
2057                     jit_check_frame();
2058                     calli(node->u.w, jit_exchange_p());
2059                 }
2060                 break;
2061             case jit_code_prolog:
2062                 _jitc->function = _jitc->functions.ptr + node->w.w;
2063                 undo.node = node;
2064                 undo.word = _jit->pc.w;
2065                 memcpy(&undo.func, _jitc->function, sizeof(undo.func));
2066 #if DEVEL_DISASSEMBLER
2067                 undo.prevw = prevw;
2068 #endif
2069                 undo.data = _jitc->consts.data;
2070                 undo.thumb = _jitc->thumb;
2071                 undo.const_offset = _jitc->consts.offset;
2072                 undo.patch_offset = _jitc->patches.offset;
2073 #if DISASSEMBLER
2074                 if (_jitc->data_info.ptr)
2075                     undo.info_offset = _jitc->data_info.offset;
2076 #endif
2077             restart_function:
2078                 _jitc->again = 0;
2079                 compute_framesize();
2080                 patch_alist(0);
2081                 prolog(node);
2082                 break;
2083             case jit_code_epilog:
2084                 assert(_jitc->function == _jitc->functions.ptr + node->w.w);
2085                 if (_jitc->again) {
2086                     for (temp = undo.node->next;
2087                          temp != node; temp = temp->next) {
2088                         if (temp->code == jit_code_label ||
2089                             temp->code == jit_code_epilog)
2090                             temp->flag &= ~jit_flag_patch;
2091                     }
2092                     temp->flag &= ~jit_flag_patch;
2093                     node = undo.node;
2094                     _jit->pc.w = undo.word;
2095                     /* undo.func.self.aoff and undo.func.regset should not
2096                      * be undone, as they will be further updated, and are
2097                      * the reason of the undo. */
2098                     undo.func.self.aoff = _jitc->function->frame +
2099                         _jitc->function->self.aoff;
2100                     undo.func.need_frame = _jitc->function->need_frame;
2101                     undo.func.need_return = _jitc->function->need_return;
2102                     jit_regset_set(&undo.func.regset, &_jitc->function->regset);
2103                     /* allocar information also does not need to be undone */
2104                     undo.func.aoffoff = _jitc->function->aoffoff;
2105                     undo.func.allocar = _jitc->function->allocar;
2106                     /* swf_offset and check_reg_args must also not be undone */
2107                     undo.func.swf_offset = _jitc->function->swf_offset;
2108                     undo.func.save_reg_args = _jitc->function->save_reg_args;
2109                     memcpy(_jitc->function, &undo.func, sizeof(undo.func));
2110 #if DEVEL_DISASSEMBLER
2111                     prevw = undo.prevw;
2112 #endif
2113                     invalidate_consts();
2114                     _jitc->consts.data = undo.data;
2115                     _jitc->thumb = undo.thumb;
2116                     _jitc->consts.offset = undo.const_offset;
2117                     _jitc->patches.offset = undo.patch_offset;
2118 #if DISASSEMBLER
2119                     if (_jitc->data_info.ptr)
2120                         _jitc->data_info.offset = undo.info_offset;
2121 #endif
2122                     patch_alist(1);
2123                     goto restart_function;
2124                 }
2125                 /* remember label is defined */
2126                 node->flag |= jit_flag_patch;
2127                 node->u.w = _jit->pc.w;
2128                 epilog(node);
2129                 _jitc->function = NULL;
2130                 flush_consts();
2131                 break;
2132             case jit_code_movr_w_f:
2133                 if (jit_swf_p())
2134                     swf_movr_w_f(rn(node->u.w), rn(node->v.w));
2135                 else
2136                     vfp_movr_w_f(rn(node->u.w), rn(node->v.w));
2137                 break;
2138             case jit_code_movr_f_w:
2139                 if (jit_swf_p())
2140                     swf_movr_f_w(rn(node->u.w), rn(node->v.w));
2141                 else
2142                     vfp_movr_f_w(rn(node->u.w), rn(node->v.w));
2143                 break;
2144             case jit_code_movi_f_w:
2145                 assert_data(node);
2146                 movi_f_w(rn(node->u.w), node->v.f);
2147                 break;
2148             case jit_code_movi_w_f:
2149                 if (jit_swf_p())
2150                     swf_movi_w_f(rn(node->u.w), node->v.w);
2151                 else
2152                     vfp_movi_w_f(rn(node->u.w), node->v.w);
2153                 break;
2154             case jit_code_movr_ww_d:
2155                 if (jit_swf_p())
2156                     swf_movr_ww_d(rn(node->u.w), rn(node->v.w), rn(node->w.w));
2157                 else
2158                     vfp_movr_ww_d(rn(node->u.w), rn(node->v.w), rn(node->w.w));
2159                 break;
2160             case jit_code_movr_d_ww:
2161                 if (jit_swf_p())
2162                     swf_movr_d_ww(rn(node->u.w), rn(node->v.w), rn(node->w.w));
2163                 else
2164                     vfp_movr_d_ww(rn(node->u.w), rn(node->v.w), rn(node->w.w));
2165                 break;
2166             case jit_code_movi_d_ww:
2167                 movi_d_ww(rn(node->u.w), rn(node->v.w), node->w.d);
2168                 break;
2169             case jit_code_movi_ww_d:
2170                 if (jit_swf_p())
2171                     swf_movi_ww_d(rn(node->u.w), node->v.w, node->w.w);
2172                 else
2173                     vfp_movi_ww_d(rn(node->u.w), node->v.w, node->w.w);
2174                 break;
2175             case jit_code_va_start:
2176                 vastart(rn(node->u.w));
2177                 break;
2178             case jit_code_va_arg:
2179                 vaarg(rn(node->u.w), rn(node->v.w));
2180                 break;
2181             case jit_code_va_arg_d:
2182                 if (jit_swf_p())
2183                     swf_vaarg_d(rn(node->u.w), rn(node->v.w));
2184                 else
2185                     vfp_vaarg_d(rn(node->u.w), rn(node->v.w));
2186                 break;
2187             case jit_code_live:                 case jit_code_ellipsis:
2188             case jit_code_va_push:
2189             case jit_code_allocai:              case jit_code_allocar:
2190             case jit_code_arg_c:                case jit_code_arg_s:
2191             case jit_code_arg_i:
2192             case jit_code_arg_f:                case jit_code_arg_d:
2193             case jit_code_va_end:
2194             case jit_code_ret:
2195             case jit_code_retr_c:               case jit_code_reti_c:
2196             case jit_code_retr_uc:              case jit_code_reti_uc:
2197             case jit_code_retr_s:               case jit_code_reti_s:
2198             case jit_code_retr_us:              case jit_code_reti_us:
2199             case jit_code_retr_i:               case jit_code_reti_i:
2200             case jit_code_retr_f:               case jit_code_reti_f:
2201             case jit_code_retr_d:               case jit_code_reti_d:
2202             case jit_code_getarg_c:             case jit_code_getarg_uc:
2203             case jit_code_getarg_s:             case jit_code_getarg_us:
2204             case jit_code_getarg_i:
2205             case jit_code_getarg_f:             case jit_code_getarg_d:
2206             case jit_code_putargr_c:            case jit_code_putargi_c:
2207             case jit_code_putargr_uc:           case  jit_code_putargi_uc:
2208             case jit_code_putargr_s:            case jit_code_putargi_s:
2209             case jit_code_putargr_us:           case jit_code_putargi_us:
2210             case jit_code_putargr_i:            case jit_code_putargi_i:
2211             case jit_code_putargr_f:            case jit_code_putargi_f:
2212             case jit_code_putargr_d:            case jit_code_putargi_d:
2213             case jit_code_pushargr_c:           case jit_code_pushargi_c:
2214             case jit_code_pushargr_uc:          case jit_code_pushargi_uc:
2215             case jit_code_pushargr_s:           case jit_code_pushargi_s:
2216             case jit_code_pushargr_us:          case jit_code_pushargi_us:
2217             case jit_code_pushargr_i:           case jit_code_pushargi_i:
2218             case jit_code_pushargr_f:           case jit_code_pushargi_f:
2219             case jit_code_pushargr_d:           case jit_code_pushargi_d:
2220             case jit_code_retval_c:             case jit_code_retval_uc:
2221             case jit_code_retval_s:             case jit_code_retval_us:
2222             case jit_code_retval_i:
2223             case jit_code_retval_f:             case jit_code_retval_d:
2224             case jit_code_prepare:
2225             case jit_code_finishr:              case jit_code_finishi:
2226             case jit_code_negi_f:               case jit_code_absi_f:
2227             case jit_code_sqrti_f:              case jit_code_negi_d:
2228             case jit_code_absi_d:               case jit_code_sqrti_d:
2229                 break;
2230             case jit_code_negi:
2231                 negi(rn(node->u.w), node->v.w);
2232                 break;
2233             case jit_code_comi:
2234                 comi(rn(node->u.w), node->v.w);
2235                 break;
2236             case jit_code_exti_c:
2237                 exti_c(rn(node->u.w), node->v.w);
2238                 break;
2239             case jit_code_exti_uc:
2240                 exti_uc(rn(node->u.w), node->v.w);
2241                 break;
2242             case jit_code_exti_s:
2243                 exti_s(rn(node->u.w), node->v.w);
2244                 break;
2245             case jit_code_exti_us:
2246                 exti_us(rn(node->u.w), node->v.w);
2247                 break;
2248             case jit_code_bswapi_us:
2249                 bswapi_us(rn(node->u.w), node->v.w);
2250                 break;
2251             case jit_code_bswapi_ui:
2252                 bswapi_ui(rn(node->u.w), node->v.w);
2253                 break;
2254             case jit_code_htoni_us:
2255                 htoni_us(rn(node->u.w), node->v.w);
2256                 break;
2257             case jit_code_htoni_ui:
2258                 htoni_ui(rn(node->u.w), node->v.w);
2259                 break;
2260             case jit_code_cloi:
2261                 cloi(rn(node->u.w), node->v.w);
2262                 break;
2263             case jit_code_clzi:
2264                 clzi(rn(node->u.w), node->v.w);
2265                 break;
2266             case jit_code_ctoi:
2267                 ctoi(rn(node->u.w), node->v.w);
2268                 break;
2269             case jit_code_ctzi:
2270                 ctzi(rn(node->u.w), node->v.w);
2271                 break;
2272             case jit_code_rbiti:
2273                 rbiti(rn(node->u.w), node->v.w);
2274                 break;
2275             case jit_code_popcnti:
2276                 popcnti(rn(node->u.w), node->v.w);
2277                 break;
2278             case jit_code_exti:
2279                 exti(rn(node->u.w), node->v.w, node->w.q.l, node->w.q.h);
2280                 break;
2281             case jit_code_exti_u:
2282                 exti_u(rn(node->u.w), node->v.w, node->w.q.l, node->w.q.h);
2283                 break;
2284             default:
2285                 abort();
2286         }
2287         jit_regarg_clr(node, value);
2288         assert(_jitc->regarg == 0 && _jitc->synth == 0);
2289         /* update register live state */
2290         jit_reglive(node);
2291
2292 #if defined JIT_INSTR_MAX
2293         word = 4096 - JIT_INSTR_MAX;
2294 #else
2295         word = 3968;
2296 #endif
2297         /* longest sequence should be 64 bytes, but preventively
2298          * do not let it go past 256 remaining bytes before a flush */
2299         if (word > 3968)
2300             word = 3968;
2301         if (_jitc->consts.length &&
2302             (_jit->pc.uc - _jitc->consts.data >= word ||
2303              (jit_uword_t)_jit->pc.uc -
2304              (jit_uword_t)_jitc->consts.patches[0] >= word)) {
2305             if (node->next &&
2306                 node->next->code != jit_code_jmpi &&
2307                 node->next->code != jit_code_jmpr &&
2308                 node->next->code != jit_code_epilog) {
2309                 /* insert a jump, flush constants and continue */
2310                 word = _jit->pc.w;
2311                 assert(!jit_thumb_p());
2312                 B(0);
2313                 flush_consts();
2314                 patch_at(arm_patch_jump, word, _jit->pc.w);
2315             }
2316         }
2317     }
2318 #undef case_bvd
2319 #undef case_bvf
2320 #undef case_brw
2321 #undef case_bvv
2322 #undef case_brr
2323 #undef case_wvv
2324 #undef case_wrr
2325 #undef case_vvd
2326 #undef case_vvf
2327 #undef case_vvw
2328 #undef case_rrw
2329 #undef case_vvv
2330 #undef case_rrr
2331 #undef case_wv
2332 #undef case_wr
2333 #undef case_vw
2334 #undef case_vv
2335 #undef case_rw
2336 #undef case_rr
2337
2338     flush_consts();
2339     for (offset = 0; offset < _jitc->patches.offset; offset++) {
2340         assert(_jitc->patches.ptr[offset].kind & arm_patch_node);
2341         node = _jitc->patches.ptr[offset].node;
2342         word = _jitc->patches.ptr[offset].inst;
2343         if (!jit_thumb_p() &&
2344             (node->code == jit_code_movi ||
2345              (node->code == jit_code_calli &&
2346               (_jitc->patches.ptr[offset].kind & ~arm_patch_node) ==
2347               arm_patch_word))) {
2348             /* calculate where to patch word */
2349             value = *(jit_int32_t *)word;
2350             assert((value & 0x0f700000) == ARM_LDRI);
2351             /* offset may become negative (-4) if last instruction
2352              * before unconditional branch and data following
2353              * FIXME can this cause issues in the preprocessor prefetch
2354              * or something else? should not, as the constants are after
2355              * an unconditional jump */
2356             if (value & ARM_P)  value =   value & 0x00000fff;
2357             else                value = -(value & 0x00000fff);
2358             word = word + 8 + value;
2359         }
2360         value = node->code == jit_code_movi ? node->v.n->u.w : node->u.n->u.w;
2361         patch_at(_jitc->patches.ptr[offset].kind & ~arm_patch_node, word, value);
2362     }
2363
2364     jit_flush(_jit->code.ptr, _jit->pc.uc);
2365
2366     return (_jit->code.ptr);
2367 }
2368
2369 #define CODE                            1
2370 #  include "jit_rewind.c"
2371 #  include "jit_arm-cpu.c"
2372 #  include "jit_arm-swf.c"
2373 #  include "jit_arm-vfp.c"
2374 #  include "jit_fallback.c"
2375 #undef CODE
2376
2377 void
2378 jit_flush(void *fptr, void *tptr)
2379 {
2380 #if defined(__GNUC__)
2381     jit_uword_t         i, f, t, s;
2382
2383     s = sysconf(_SC_PAGE_SIZE);
2384     f = (jit_uword_t)fptr & -s;
2385     t = (((jit_uword_t)tptr) + s - 1) & -s;
2386     for (i = f; i < t; i += s)
2387         __clear_cache((void *)i, (void *)(i + s));
2388 #endif
2389 }
2390
2391 void
2392 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
2393 {
2394     ldxi_i(rn(r0), rn(r1), i0);
2395 }
2396
2397 void
2398 _emit_stxi(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
2399 {
2400     stxi_i(i0, rn(r0), rn(r1));
2401 }
2402
2403 void
2404 _emit_ldxi_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
2405 {
2406     if (jit_swf_p())
2407         swf_ldxi_d(rn(r0), rn(r1), i0);
2408     else
2409         vfp_ldxi_d(rn(r0), rn(r1), i0);
2410 }
2411
2412 void
2413 _emit_stxi_d(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
2414 {
2415     if (jit_swf_p())
2416         swf_stxi_d(i0, rn(r0), rn(r1));
2417     else
2418         vfp_stxi_d(i0, rn(r0), rn(r1));
2419 }
2420
2421 static jit_int32_t
2422 _jit_get_reg_pair(jit_state_t *_jit)
2423 {
2424     /*   bypass jit_get_reg() with argument or'ed with jit_class_chk
2425      * and try to find an consecutive, even free register pair, or
2426      * return JIT_NOREG if fail, as the cost of spills is greater
2427      * than splitting a double load/store in two operations. */
2428     if (jit_reg_free_p(_R0) && jit_reg_free_p(_R1)) {
2429         jit_regset_setbit(&_jitc->regarg, _R0);
2430         jit_regset_setbit(&_jitc->regarg, _R1);
2431         return (_R0);
2432     }
2433     if (jit_reg_free_p(_R2) && jit_reg_free_p(_R3)) {
2434         jit_regset_setbit(&_jitc->regarg, _R2);
2435         jit_regset_setbit(&_jitc->regarg, _R3);
2436         return (_R2);
2437     }
2438     if (jit_reg_free_p(_R4) && jit_reg_free_p(_R5)) {
2439         jit_regset_setbit(&_jitc->regarg, _R4);
2440         jit_regset_setbit(&_jitc->regarg, _R5);
2441         return (_R4);
2442     }
2443     if (jit_reg_free_p(_R6) && jit_reg_free_p(_R7)) {
2444         jit_regset_setbit(&_jitc->regarg, _R6);
2445         jit_regset_setbit(&_jitc->regarg, _R7);
2446         return (_R6);
2447     }
2448     if (jit_reg_free_p(_R8) && jit_reg_free_p(_R9)) {
2449         jit_regset_setbit(&_jitc->regarg, _R8);
2450         jit_regset_setbit(&_jitc->regarg, _R9);
2451         return (_R8);
2452     }
2453     return (JIT_NOREG);
2454 }
2455
2456 static void
2457 _jit_unget_reg_pair(jit_state_t *_jit, jit_int32_t reg)
2458 {
2459     jit_unget_reg(reg);
2460     switch (reg) {
2461         case _R0:       jit_unget_reg(_R1);     break;
2462         case _R2:       jit_unget_reg(_R3);     break;
2463         case _R4:       jit_unget_reg(_R5);     break;
2464         case _R6:       jit_unget_reg(_R7);     break;
2465         case _R8:       jit_unget_reg(_R9);     break;
2466         default:        abort();
2467     }
2468 }
2469
2470 /*   A prolog must be aligned at mod 4 bytes boundary.
2471  *   This condition was not being required to be tested by
2472  * accident previously, but with the jit_frame and jit_tramp
2473  * code it is required */
2474 static jit_bool_t
2475 _must_align_p(jit_state_t *_jit, jit_node_t *node)
2476 {
2477     if (jit_thumb_p() && (_jit->pc.w & 3)) {
2478         for (; node; node = node->next) {
2479             switch (node->code) {
2480                 case jit_code_note:
2481                 case jit_code_name:
2482                 case jit_code_label:
2483                     break;
2484                 case jit_code_prolog:
2485                     return (1);
2486                 default:
2487                     return (0);
2488             }
2489         }
2490     }
2491     return (0);
2492 }
2493
2494 static void
2495 _load_const(jit_state_t *_jit, jit_bool_t uniq, jit_int32_t r0, jit_word_t i0)
2496 {
2497     jit_word_t           w;
2498     jit_word_t           d;
2499     jit_word_t           base;
2500     jit_int32_t         *data;
2501     jit_int32_t          size;
2502     jit_int32_t          offset;
2503
2504     assert(!jit_thumb_p());
2505     if (!uniq) {
2506         /* use zero, a valid directly encoded immediate, to avoid the
2507          * need of a bitmask to know what offsets will be patched, so
2508          * that comparison will always fail for constants that cannot
2509          * be encoded */
2510         assert(i0 != 0);
2511
2512         /* Actually, code is (currently at least) not self modifying,
2513          * so, any value reachable backwards is valid as a constant. */
2514
2515         /* FIXME a quickly updateable/mutable hash table could be
2516          * better here, but most times only a few comparisons
2517          * should be done
2518          */
2519
2520         /* search in previous constant pool */
2521         if ((data = (jit_int32_t *)_jitc->consts.data)) {
2522             w = (jit_word_t)data;
2523             /* maximum backwards offset */
2524             base = (_jit->pc.w + 8) - 4092;
2525             if (base <= w)
2526                 /* can scan all possible available backward constants */
2527                 base = 0;
2528             else
2529                 base = (base - w) >> 2;
2530             size = _jitc->consts.size >> 2;
2531             for (offset = size - 1; offset >= base; offset--) {
2532                 if (data[offset] == i0) {
2533                     w = (jit_word_t)(data + offset);
2534                     d = (_jit->pc.w + 8) - w;
2535                     LDRIN(r0, _R15_REGNO, d);
2536                     return;
2537                 }
2538             }
2539         }
2540     }
2541     else
2542         assert(i0 == 0);
2543
2544     _jitc->consts.patches[_jitc->consts.offset++] = _jit->pc.w;
2545     /* (probably) positive forward offset */
2546     LDRI(r0, _R15_REGNO, 0);
2547
2548     if (!uniq) {
2549         /* search already requested values */
2550         for (offset = 0; offset < _jitc->consts.length; offset++) {
2551             if (_jitc->consts.values[offset] == i0) {
2552                 _jitc->consts.patches[_jitc->consts.offset++] = offset;
2553                 return;
2554             }
2555         }
2556     }
2557
2558 #if DEBUG
2559     /* cannot run out of space because of limited range
2560      * but assert anyway to catch logic errors */
2561     assert(_jitc->consts.length < 1024);
2562     assert(_jitc->consts.offset < 2048);
2563 #endif
2564     _jitc->consts.patches[_jitc->consts.offset++] = _jitc->consts.length;
2565     _jitc->consts.values[_jitc->consts.length++] = i0;
2566 }
2567
2568 static void
2569 _flush_consts(jit_state_t *_jit)
2570 {
2571     jit_word_t           word;
2572     jit_int32_t          offset;
2573
2574     /* if no forward constants */
2575     if (!_jitc->consts.length)
2576         return;
2577     assert(!jit_thumb_p());
2578     word = _jit->pc.w;
2579     _jitc->consts.data = _jit->pc.uc;
2580     _jitc->consts.size = _jitc->consts.length << 2;
2581     /* FIXME check will not overrun, otherwise, need to reallocate
2582      * code buffer and start over */
2583     jit_memcpy(_jitc->consts.data, _jitc->consts.values, _jitc->consts.size);
2584     _jit->pc.w += _jitc->consts.size;
2585
2586 #if DISASSEMBLER
2587     if (_jitc->data_info.ptr) {
2588         if (_jitc->data_info.offset >= _jitc->data_info.length) {
2589             jit_realloc((jit_pointer_t *)&_jitc->data_info.ptr,
2590                         _jitc->data_info.length * sizeof(jit_data_info_t),
2591                         (_jitc->data_info.length + 1024) *
2592                         sizeof(jit_data_info_t));
2593             _jitc->data_info.length += 1024;
2594         }
2595         _jitc->data_info.ptr[_jitc->data_info.offset].code = word;
2596         _jitc->data_info.ptr[_jitc->data_info.offset].length = _jitc->consts.size;
2597         ++_jitc->data_info.offset;
2598     }
2599 #endif
2600
2601     for (offset = 0; offset < _jitc->consts.offset; offset += 2)
2602         patch_at(arm_patch_load, _jitc->consts.patches[offset],
2603                  word + (_jitc->consts.patches[offset + 1] << 2));
2604     _jitc->consts.length = _jitc->consts.offset = 0;
2605 }
2606
2607 /* to be called if needing to start over a function */
2608 static void
2609 _invalidate_consts(jit_state_t *_jit)
2610 {
2611     /* if no forward constants */
2612     if (_jitc->consts.length)
2613         _jitc->consts.length = _jitc->consts.offset = 0;
2614 }
2615
2616 static void
2617 _compute_framesize(jit_state_t *_jit)
2618 {
2619     jit_int32_t         reg;
2620     _jitc->framesize = sizeof(jit_word_t) * 2;  /* lr+fp */
2621     for (reg = 0; reg < jit_size(iregs); reg++)
2622         if (jit_regset_tstbit(&_jitc->function->regset, iregs[reg]))
2623             _jitc->framesize += sizeof(jit_word_t);
2624
2625     if (_jitc->function->save_reg_args)
2626         _jitc->framesize += 16;
2627
2628     /* Make sure functions called have a 8 byte aligned stack */
2629     _jitc->framesize = (_jitc->framesize + 7) & -8;
2630 }
2631
2632 static void
2633 _patch(jit_state_t *_jit, jit_word_t instr, jit_node_t *node, jit_int32_t kind)
2634 {
2635     jit_int32_t          flag;
2636
2637     assert(node->flag & jit_flag_node);
2638     if (node->code == jit_code_movi)
2639         flag = node->v.n->flag;
2640     else
2641         flag = node->u.n->flag;
2642     assert(!(flag & jit_flag_patch));
2643     kind |= arm_patch_node;
2644     if (_jitc->patches.offset >= _jitc->patches.length) {
2645         jit_realloc((jit_pointer_t *)&_jitc->patches.ptr,
2646                     _jitc->patches.length * sizeof(jit_patch_t),
2647                     (_jitc->patches.length + 1024) * sizeof(jit_patch_t));
2648         _jitc->patches.length += 1024;
2649     }
2650     _jitc->patches.ptr[_jitc->patches.offset].kind = kind;
2651     _jitc->patches.ptr[_jitc->patches.offset].inst = instr;
2652     _jitc->patches.ptr[_jitc->patches.offset].node = node;
2653     ++_jitc->patches.offset;
2654 }