deps/lightning/lib/jit_aarch64.c

   1 /*
   2  * Copyright (C) 2013-2023  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 /* callee save
  21  * align16(lr+fp+x19+x2[0-8]+v8+v9+v1[0-15]) */
  22 #define stack_framesize                 160
  23
  24 #define jit_arg_reg_p(i)                ((i) >= 0 && (i) < 8)
  25 #define jit_arg_f_reg_p(i)              ((i) >= 0 && (i) < 8)
  26
  27 #if __APPLE__
  28 typedef jit_pointer_t jit_va_list_t;
  29 #else
  30 typedef struct jit_qreg {
  31     jit_float64_t       l;
  32     jit_float64_t       h;
  33 } jit_qreg_t;
  34
  35 #define va_gp_top_offset                offsetof(jit_va_list_t, q0)
  36 #define va_fp_top_offset                sizeof(jit_va_list_t)
  37 typedef struct jit_va_list {
  38     jit_pointer_t       stack;
  39     jit_pointer_t       gptop;
  40     jit_pointer_t       fptop;
  41     jit_int32_t         gpoff;
  42     jit_int32_t         fpoff;
  43
  44     jit_int64_t         x0;
  45     jit_int64_t         x1;
  46     jit_int64_t         x2;
  47     jit_int64_t         x3;
  48     jit_int64_t         x4;
  49     jit_int64_t         x5;
  50     jit_int64_t         x6;
  51     jit_int64_t         x7;
  52
  53     jit_qreg_t          q0;
  54     jit_qreg_t          q1;
  55     jit_qreg_t          q2;
  56     jit_qreg_t          q3;
  57     jit_qreg_t          q4;
  58     jit_qreg_t          q5;
  59     jit_qreg_t          q6;
  60     jit_qreg_t          q7;
  61 } jit_va_list_t;
  62 #endif
  63
  64 /*
  65  * Prototypes
  66  */
  67 #define compute_framesize()             _compute_framesize(_jit)
  68 static void _compute_framesize(jit_state_t*);
  69 #define patch(instr, node)              _patch(_jit, instr, node)
  70 static void _patch(jit_state_t*,jit_word_t,jit_node_t*);
  71
  72 /* libgcc */
  73 extern void __clear_cache(void *, void *);
  74
  75 #define PROTO                           1
  76 #  include "jit_aarch64-cpu.c"
  77 #  include "jit_aarch64-fpu.c"
  78 #undef PROTO
  79
  80 /*
  81  * Initialization
  82  */
  83 jit_register_t          _rvs[] = {
  84     { rc(gpr) | 0x08,                   "x8" },
  85 #if __APPLE__
  86     { 0x12,                             "x18" },
  87 #else
  88     { rc(gpr) | 0x12,                   "x18" },
  89 #endif
  90     { rc(gpr) | 0x11,                   "x17" },
  91     { rc(gpr) | 0x10,                   "x16" },
  92     { rc(gpr) | 0x09,                   "x9" },
  93     { rc(gpr) | 0x0a,                   "x10" },
  94     { rc(gpr) | 0x0b,                   "x11" },
  95     { rc(gpr) | 0x0c,                   "x12" },
  96     { rc(gpr) | 0x0d,                   "x13" },
  97     { rc(gpr) | 0x0e,                   "x14" },
  98     { rc(gpr) | 0x0f,                   "x15" },
  99     { rc(sav) | rc(gpr) | 0x13,         "x19" },
 100     { rc(sav) | rc(gpr) | 0x14,         "x20" },
 101     { rc(sav) | rc(gpr) | 0x15,         "x21" },
 102     { rc(sav) | rc(gpr) | 0x16,         "x22" },
 103     { rc(sav) | rc(gpr) | 0x17,         "x23" },
 104     { rc(sav) | rc(gpr) | 0x18,         "x24" },
 105     { rc(sav) | rc(gpr) | 0x19,         "x25" },
 106     { rc(sav) | rc(gpr) | 0x1a,         "x26" },
 107     { rc(sav) | rc(gpr) | 0x1b,         "x27" },
 108     { rc(sav) | rc(gpr) | 0x1c,         "x28" },
 109     { 0x1f,                             "sp" },
 110     { 0x1e,                             "lr" },
 111     { 0x1d,                             "fp" },
 112     { rc(arg) | rc(gpr) | 0x07,         "x7" },
 113     { rc(arg) | rc(gpr) | 0x06,         "x6" },
 114     { rc(arg) | rc(gpr) | 0x05,         "x5" },
 115     { rc(arg) | rc(gpr) | 0x04,         "x4" },
 116     { rc(arg) | rc(gpr) | 0x03,         "x3" },
 117     { rc(arg) | rc(gpr) | 0x02,         "x2" },
 118     { rc(arg) | rc(gpr) | 0x01,         "x1" },
 119     { rc(arg) | rc(gpr) | 0x00,         "x0" },
 120     { rc(fpr) | 0x1f,                   "v31" },
 121     { rc(fpr) | 0x1e,                   "v30" },
 122     { rc(fpr) | 0x1d,                   "v29" },
 123     { rc(fpr) | 0x1c,                   "v28" },
 124     { rc(fpr) | 0x1b,                   "v27" },
 125     { rc(fpr) | 0x1a,                   "v26" },
 126     { rc(fpr) | 0x19,                   "v25" },
 127     { rc(fpr) | 0x18,                   "v24" },
 128     { rc(fpr) | 0x17,                   "v23" },
 129     { rc(fpr) | 0x16,                   "v22" },
 130     { rc(fpr) | 0x15,                   "v21" },
 131     { rc(fpr) | 0x14,                   "v20" },
 132     { rc(fpr) | 0x13,                   "v19" },
 133     { rc(fpr) | 0x12,                   "v18" },
 134     { rc(fpr) | 0x11,                   "v17" },
 135     { rc(fpr) | 0x10,                   "v16" },
 136     { rc(sav) | rc(fpr) | 0x08,         "v8" },
 137     { rc(sav) | rc(fpr) | 0x09,         "v9" },
 138     { rc(sav) | rc(fpr) | 0x0a,         "v10" },
 139     { rc(sav) | rc(fpr) | 0x0b,         "v11" },
 140     { rc(sav) | rc(fpr) | 0x0c,         "v12" },
 141     { rc(sav) | rc(fpr) | 0x0d,         "v13" },
 142     { rc(sav) | rc(fpr) | 0x0e,         "v14" },
 143     { rc(sav) | rc(fpr) | 0x0f,         "v15" },
 144     { rc(arg) | rc(fpr) | 0x07,         "v7" },
 145     { rc(arg) | rc(fpr) | 0x06,         "v6" },
 146     { rc(arg) | rc(fpr) | 0x05,         "v5" },
 147     { rc(arg) | rc(fpr) | 0x04,         "v4" },
 148     { rc(arg) | rc(fpr) | 0x03,         "v3" },
 149     { rc(arg) | rc(fpr) | 0x02,         "v2" },
 150     { rc(arg) | rc(fpr) | 0x01,         "v1" },
 151     { rc(arg) | rc(fpr) | 0x00,         "v0" },
 152     { _NOREG,                           "<none>" },
 153 };
 154
 155 static jit_int32_t iregs[] = {
 156     _R19, _R20, _R21, _R22, _R23, _R24, _R25, _R26, _R27, _R28
 157 };
 158
 159 static jit_int32_t fregs[] = {
 160     _V8, _V9, _V10, _V11, _V12, _V13, _V14, _V15
 161 };
 162
 163 /*
 164  * Implementation
 165  */
 166 void
 167 jit_get_cpu(void)
 168 {
 169 }
 170
 171 void
 172 _jit_init(jit_state_t *_jit)
 173 {
 174     _jitc->reglen = jit_size(_rvs) - 1;
 175 }
 176
 177 void
 178 _jit_prolog(jit_state_t *_jit)
 179 {
 180     jit_int32_t          offset;
 181
 182     if (_jitc->function)
 183         jit_epilog();
 184     assert(jit_regset_cmp_ui(&_jitc->regarg, 0) == 0);
 185     jit_regset_set_ui(&_jitc->regsav, 0);
 186     offset = _jitc->functions.offset;
 187     if (offset >= _jitc->functions.length) {
 188         jit_realloc((jit_pointer_t *)&_jitc->functions.ptr,
 189                     _jitc->functions.length * sizeof(jit_function_t),
 190                     (_jitc->functions.length + 16) * sizeof(jit_function_t));
 191         _jitc->functions.length += 16;
 192     }
 193     _jitc->function = _jitc->functions.ptr + _jitc->functions.offset++;
 194     _jitc->function->self.size = stack_framesize;
 195     _jitc->function->self.argi = _jitc->function->self.argf =
 196         _jitc->function->self.alen = 0;
 197     _jitc->function->self.aoff = 0;
 198     _jitc->function->self.call = jit_call_default;
 199     jit_alloc((jit_pointer_t *)&_jitc->function->regoff,
 200               _jitc->reglen * sizeof(jit_int32_t));
 201
 202     /* _no_link here does not mean the jit_link() call can be removed
 203      * by rewriting as:
 204      * _jitc->function->prolog = jit_new_node(jit_code_prolog);
 205      */
 206     _jitc->function->prolog = jit_new_node_no_link(jit_code_prolog);
 207     jit_link(_jitc->function->prolog);
 208     _jitc->function->prolog->w.w = offset;
 209     _jitc->function->epilog = jit_new_node_no_link(jit_code_epilog);
 210     /*  u:      label value
 211      *  v:      offset in blocks vector
 212      *  w:      offset in functions vector
 213      */
 214     _jitc->function->epilog->w.w = offset;
 215
 216     jit_regset_new(&_jitc->function->regset);
 217 }
 218
 219 jit_int32_t
 220 _jit_allocai(jit_state_t *_jit, jit_int32_t length)
 221 {
 222     assert(_jitc->function);
 223     jit_check_frame();
 224     switch (length) {
 225         case 0: case 1:                                         break;
 226         case 2:         _jitc->function->self.aoff &= -2;       break;
 227         case 3: case 4: _jitc->function->self.aoff &= -4;       break;
 228         default:        _jitc->function->self.aoff &= -8;       break;
 229     }
 230     _jitc->function->self.aoff -= length;
 231     if (!_jitc->realize) {
 232         jit_inc_synth_ww(allocai, _jitc->function->self.aoff, length);
 233         jit_dec_synth();
 234     }
 235     return (_jitc->function->self.aoff);
 236 }
 237
 238 void
 239 _jit_allocar(jit_state_t *_jit, jit_int32_t u, jit_int32_t v)
 240 {
 241     jit_int32_t          r0, r1;
 242     assert(_jitc->function);
 243     jit_inc_synth_ww(allocar, u, v);
 244     if (!_jitc->function->allocar) {
 245         _jitc->function->aoffoff = jit_allocai(sizeof(jit_int32_t));
 246         _jitc->function->allocar = 1;
 247     }
 248     r0 = jit_get_reg(jit_class_gpr);
 249     jit_negr(r0, v);
 250     jit_andi(r0, r0, -16);
 251     jit_ldxi_i(u, JIT_FP, _jitc->function->aoffoff);
 252     jit_addr(u, u, r0);
 253     /* Cannot "addr sp, sp, reg" because in this context "sp" is "[w|x]zr",
 254      * the zero register */
 255 #if 0
 256     jit_addr(JIT_SP, JIT_SP, r0);
 257 #else
 258     r1 = jit_get_reg(jit_class_gpr);
 259     /* note that "mov r1, sp" does not work, but the proper encoding
 260      * can be triggered before actually emiting with "add r1, sp, 0" */
 261     jit_addi(r1, JIT_SP, 0);
 262     jit_addr(r1, r1, r0);
 263     jit_addi(JIT_SP, r1, 0);
 264     jit_unget_reg(r1);
 265 #endif
 266     jit_stxi_i(_jitc->function->aoffoff, JIT_FP, u);
 267     jit_unget_reg(r0);
 268     jit_dec_synth();
 269 }
 270
 271 void
 272 _jit_ret(jit_state_t *_jit)
 273 {
 274     jit_node_t          *instr;
 275     assert(_jitc->function);
 276     jit_inc_synth(ret);
 277     /* jump to epilog */
 278     instr = jit_jmpi();
 279     jit_patch_at(instr, _jitc->function->epilog);
 280     jit_dec_synth();
 281 }
 282
 283 void
 284 _jit_retr(jit_state_t *_jit, jit_int32_t u, jit_code_t code)
 285 {
 286     jit_code_inc_synth_w(code, u);
 287     jit_movr(JIT_RET, u);
 288     jit_ret();
 289     jit_dec_synth();
 290 }
 291
 292 void
 293 _jit_reti(jit_state_t *_jit, jit_word_t u, jit_code_t code)
 294 {
 295     jit_code_inc_synth_w(code, u);
 296     jit_movi(JIT_RET, u);
 297     jit_ret();
 298     jit_dec_synth();
 299 }
 300
 301 void
 302 _jit_retr_f(jit_state_t *_jit, jit_int32_t u)
 303 {
 304     jit_inc_synth_w(retr_f, u);
 305     if (u != JIT_FRET)
 306         jit_movr_f(JIT_FRET, u);
 307     else
 308         jit_live(JIT_FRET);
 309     jit_ret();
 310     jit_dec_synth();
 311 }
 312
 313 void
 314 _jit_reti_f(jit_state_t *_jit, jit_float32_t u)
 315 {
 316     jit_inc_synth_f(reti_f, u);
 317     jit_movi_f(JIT_FRET, u);
 318     jit_ret();
 319     jit_dec_synth();
 320 }
 321
 322 void
 323 _jit_retr_d(jit_state_t *_jit, jit_int32_t u)
 324 {
 325     jit_inc_synth_w(retr_d, u);
 326     if (u != JIT_FRET)
 327         jit_movr_d(JIT_FRET, u);
 328     else
 329         jit_live(JIT_FRET);
 330     jit_ret();
 331     jit_dec_synth();
 332 }
 333
 334 void
 335 _jit_reti_d(jit_state_t *_jit, jit_float64_t u)
 336 {
 337     jit_inc_synth_d(reti_d, u);
 338     jit_movi_d(JIT_FRET, u);
 339     jit_ret();
 340     jit_dec_synth();
 341 }
 342
 343 void
 344 _jit_epilog(jit_state_t *_jit)
 345 {
 346     assert(_jitc->function);
 347     assert(_jitc->function->epilog->next == NULL);
 348     jit_link(_jitc->function->epilog);
 349     _jitc->function = NULL;
 350 }
 351
 352 jit_bool_t
 353 _jit_arg_register_p(jit_state_t *_jit, jit_node_t *u)
 354 {
 355     if (u->code >= jit_code_arg_c && u->code <= jit_code_arg)
 356         return (jit_arg_reg_p(u->u.w));
 357     assert(u->code == jit_code_arg_f || u->code == jit_code_arg_d);
 358     return (jit_arg_f_reg_p(u->u.w));
 359 }
 360
 361 void
 362 _jit_ellipsis(jit_state_t *_jit)
 363 {
 364     jit_inc_synth(ellipsis);
 365     jit_check_frame();
 366     if (_jitc->prepare) {
 367         jit_link_prepare();
 368         assert(!(_jitc->function->call.call & jit_call_varargs));
 369         _jitc->function->call.call |= jit_call_varargs;
 370     }
 371     else {
 372         jit_link_prolog();
 373         assert(!(_jitc->function->self.call & jit_call_varargs));
 374         _jitc->function->self.call |= jit_call_varargs;
 375
 376 #if !__APPLE_
 377         /* Allocate va_list like object in the stack,
 378          * with enough space to save all argument
 379          * registers, and use fixed offsets for them. */
 380         _jitc->function->vaoff = jit_allocai(sizeof(jit_va_list_t));
 381
 382         /* Initialize gp offset in save area. */
 383         if (jit_arg_reg_p(_jitc->function->self.argi))
 384             _jitc->function->vagp = (8 - _jitc->function->self.argi) * -8;
 385         else
 386             _jitc->function->vagp = 0;
 387
 388         /* Initialize fp offset in save area. */
 389         if (jit_arg_f_reg_p(_jitc->function->self.argf))
 390             _jitc->function->vafp = (8 - _jitc->function->self.argf) * -16;
 391         else
 392             _jitc->function->vafp = 0;
 393 #endif
 394     }
 395     jit_dec_synth();
 396 }
 397
 398 void
 399 _jit_va_push(jit_state_t *_jit, jit_int32_t u)
 400 {
 401     jit_inc_synth_w(va_push, u);
 402     jit_pushargr(u);
 403     jit_dec_synth();
 404 }
 405
 406 jit_node_t *
 407 _jit_arg(jit_state_t *_jit, jit_code_t code)
 408 {
 409     jit_node_t          *node;
 410     jit_int32_t          offset;
 411     assert(_jitc->function);
 412     assert(!(_jitc->function->self.call & jit_call_varargs));
 413     if (jit_arg_reg_p(_jitc->function->self.argi))
 414         offset = _jitc->function->self.argi++;
 415     else {
 416 #if PACKED_STACK || STRONG_TYPE_CHECKING
 417         assert(code >= jit_code_arg_c && code <= jit_code_arg);
 418 #endif
 419 #if PACKED_STACK
 420         _jitc->function->self.size +=
 421             _jitc->function->self.size & ((1 << (code - jit_code_arg_c)) - 1);
 422 #endif
 423         offset = _jitc->function->self.size;
 424 #if PACKED_STACK
 425         _jitc->function->self.size += 1 << (code - jit_code_arg_c);
 426 #else
 427         _jitc->function->self.size += sizeof(jit_word_t);
 428 #endif
 429         jit_check_frame();
 430     }
 431     node = jit_new_node_ww(code, offset,
 432                            ++_jitc->function->self.argn);
 433     jit_link_prolog();
 434     return (node);
 435 }
 436
 437 jit_node_t *
 438 _jit_arg_f(jit_state_t *_jit)
 439 {
 440     jit_node_t          *node;
 441     jit_int32_t          offset;
 442     assert(_jitc->function);
 443     assert(!(_jitc->function->self.call & jit_call_varargs));
 444     if (jit_arg_f_reg_p(_jitc->function->self.argf))
 445         offset = _jitc->function->self.argf++;
 446     else {
 447 #if PACKED_STACK
 448         _jitc->function->self.size +=
 449             _jitc->function->self.size & (sizeof(jit_float32_t) - 1);
 450 #endif
 451         offset = _jitc->function->self.size;
 452 #if PACKED_STACK
 453         _jitc->function->self.size += sizeof(jit_float32_t);
 454 #else
 455         _jitc->function->self.size += sizeof(jit_word_t);
 456 #endif
 457         jit_check_frame();
 458     }
 459     node = jit_new_node_ww(jit_code_arg_f, offset,
 460                            ++_jitc->function->self.argn);
 461     jit_link_prolog();
 462     return (node);
 463 }
 464
 465 jit_node_t *
 466 _jit_arg_d(jit_state_t *_jit)
 467 {
 468     jit_node_t          *node;
 469     jit_int32_t          offset;
 470     assert(_jitc->function);
 471     assert(!(_jitc->function->self.call & jit_call_varargs));
 472     if (jit_arg_f_reg_p(_jitc->function->self.argf))
 473         offset = _jitc->function->self.argf++;
 474     else {
 475 #if PACKED_STACK
 476         _jitc->function->self.size +=
 477             _jitc->function->self.size & (sizeof(jit_float64_t) - 1);
 478 #endif
 479         offset = _jitc->function->self.size;
 480         _jitc->function->self.size += sizeof(jit_float64_t);
 481         jit_check_frame();
 482     }
 483     node = jit_new_node_ww(jit_code_arg_d, offset,
 484                            ++_jitc->function->self.argn);
 485     jit_link_prolog();
 486     return (node);
 487 }
 488
 489 void
 490 _jit_getarg_c(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 491 {
 492     assert_arg_type(v->code, jit_code_arg_c);
 493     jit_inc_synth_wp(getarg_c, u, v);
 494     if (jit_arg_reg_p(v->u.w)) {
 495 #if PACKED_STACK
 496         jit_movr(u, JIT_RA0 - v->u.w);
 497 #else
 498         jit_extr_c(u, JIT_RA0 - v->u.w);
 499 #endif
 500     }
 501     else {
 502         jit_node_t      *node = jit_ldxi_c(u, JIT_FP, v->u.w);
 503         jit_link_alist(node);
 504     }
 505     jit_dec_synth();
 506 }
 507
 508 void
 509 _jit_getarg_uc(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 510 {
 511     assert_arg_type(v->code, jit_code_arg_c);
 512     jit_inc_synth_wp(getarg_uc, u, v);
 513     if (jit_arg_reg_p(v->u.w)) {
 514 #if PACKED_STACK
 515         jit_movr(u, JIT_RA0 - v->u.w);
 516 #else
 517         jit_extr_uc(u, JIT_RA0 - v->u.w);
 518 #endif
 519     }
 520     else {
 521         jit_node_t      *node = jit_ldxi_uc(u, JIT_FP, v->u.w);
 522         jit_link_alist(node);
 523     }
 524     jit_dec_synth();
 525 }
 526
 527 void
 528 _jit_getarg_s(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 529 {
 530     assert_arg_type(v->code, jit_code_arg_s);
 531     jit_inc_synth_wp(getarg_s, u, v);
 532     if (jit_arg_reg_p(v->u.w)) {
 533 #if PACKED_STACK
 534         jit_movr(u, JIT_RA0 - v->u.w);
 535 #else
 536         jit_extr_s(u, JIT_RA0 - v->u.w);
 537 #endif
 538     }
 539     else {
 540         jit_node_t      *node = jit_ldxi_s(u, JIT_FP, v->u.w);
 541         jit_link_alist(node);
 542     }
 543     jit_dec_synth();
 544 }
 545
 546 void
 547 _jit_getarg_us(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 548 {
 549     assert_arg_type(v->code, jit_code_arg_s);
 550     jit_inc_synth_wp(getarg_us, u, v);
 551     if (jit_arg_reg_p(v->u.w)) {
 552 #if PACKED_STACK
 553         jit_movr(u, JIT_RA0 - v->u.w);
 554 #else
 555         jit_extr_us(u, JIT_RA0 - v->u.w);
 556 #endif
 557     }
 558     else {
 559         jit_node_t      *node = jit_ldxi_us(u, JIT_FP, v->u.w);
 560         jit_link_alist(node);
 561     }
 562     jit_dec_synth();
 563 }
 564
 565 void
 566 _jit_getarg_i(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 567 {
 568     assert_arg_type(v->code, jit_code_arg_i);
 569     jit_inc_synth_wp(getarg_i, u, v);
 570     if (jit_arg_reg_p(v->u.w)) {
 571 #if PACKED_STACK || __WORDSIZE == 32
 572         jit_movr(u, JIT_RA0 - v->u.w);
 573 #else
 574         jit_extr_i(u, JIT_RA0 - v->u.w);
 575 #endif
 576     }
 577     else {
 578         jit_node_t      *node = jit_ldxi_i(u, JIT_FP, v->u.w);
 579         jit_link_alist(node);
 580     }
 581     jit_dec_synth();
 582 }
 583
 584 #if __WORDSIZE == 64
 585 void
 586 _jit_getarg_ui(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 587 {
 588     assert_arg_type(v->code, jit_code_arg_i);
 589     jit_inc_synth_wp(getarg_ui, u, v);
 590     if (jit_arg_reg_p(v->u.w)) {
 591 #if PACKED_STACK
 592         jit_movr(u, JIT_RA0 - v->u.w);
 593 #else
 594         jit_extr_ui(u, JIT_RA0 - v->u.w);
 595 #endif
 596     }
 597     else {
 598         jit_node_t      *node = jit_ldxi_ui(u, JIT_FP, v->u.w);
 599         jit_link_alist(node);
 600     }
 601     jit_dec_synth();
 602 }
 603
 604 void
 605 _jit_getarg_l(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 606 {
 607     assert_arg_type(v->code, jit_code_arg_l);
 608     jit_inc_synth_wp(getarg_l, u, v);
 609     if (jit_arg_reg_p(v->u.w))
 610         jit_movr(u, JIT_RA0 - v->u.w);
 611     else {
 612         jit_node_t      *node = jit_ldxi_l(u, JIT_FP, v->u.w);
 613         jit_link_alist(node);
 614     }
 615     jit_dec_synth();
 616 }
 617 #endif
 618
 619 void
 620 _jit_putargr(jit_state_t *_jit, jit_int32_t u, jit_node_t *v, jit_code_t code)
 621 {
 622     assert_putarg_type(code, v->code);
 623     jit_code_inc_synth_wp(code, u, v);
 624     if (jit_arg_reg_p(v->u.w)) {
 625         jit_int32_t     regno = JIT_RA0 - v->u.w;
 626 #if PACKED_STACK
 627         switch (code) {
 628             case jit_code_putargr_c:    jit_extr_c(regno, u);   break;
 629             case jit_code_putargr_uc:   jit_extr_uc(regno, u);  break;
 630             case jit_code_putargr_s:    jit_extr_s(regno, u);   break;
 631             case jit_code_putargr_us:   jit_extr_us(regno, u);  break;
 632 #  if __WORDISZE == 32
 633             case jit_code_putargr_i:    jit_movr(regno, u);     break;
 634 #  else
 635             case jit_code_putargr_i:    jit_extr_i(regno, u);   break;
 636             case jit_code_putargr_ui:   jit_extr_ui(regno, u);  break;
 637             case jit_code_putargr_l:    jit_movr(regno, u);     break;
 638 #  endif
 639             default:                    abort();                break;
 640         }
 641 #else
 642         jit_movr(regno, u);
 643 #endif
 644     }
 645     else {
 646         jit_node_t      *node;
 647 #if PACKED_STACK
 648         switch (code) {
 649             case jit_code_putargr_c:    case jit_code_putargr_uc:
 650                 node = jit_stxi_c(v->u.w, JIT_FP, u);           break;
 651             case jit_code_putargr_s:    case jit_code_putargr_us:
 652                 node = jit_stxi_s(v->u.w, JIT_FP, u);           break;
 653 #  if __WORDSIZE == 32
 654             case jit_code_putargr_i:
 655                 node = jit_stxi(v->u.w, JIT_FP, u);             break;
 656 #  else
 657             case jit_code_putargr_i:    case jit_code_putargr_ui:
 658                 node = jit_stxi_i(v->u.w, JIT_FP, u);           break;
 659             case jit_code_putargr_l:
 660                 node = jit_stxi(v->u.w, JIT_FP, u);             break;
 661 #  endif
 662             default:                    abort();                break;
 663         }
 664 #else
 665         node = jit_stxi(v->u.w, JIT_FP, u);
 666 #endif
 667         jit_link_alist(node);
 668     }
 669     jit_dec_synth();
 670 }
 671
 672 void
 673 _jit_putargi(jit_state_t *_jit, jit_word_t u, jit_node_t *v, jit_code_t code)
 674 {
 675     jit_int32_t         regno;
 676     assert_putarg_type(code, v->code);
 677     jit_code_inc_synth_wp(code, u, v);
 678 #if PACKED_STACK
 679     switch (code) {
 680         case jit_code_putargi_c:        u = (jit_int8_t)u;      break;
 681         case jit_code_putargi_uc:       u = (jit_uint8_t)u;     break;
 682         case jit_code_putargi_s:        u = (jit_int16_t)u;     break;
 683         case jit_code_putargi_us:       u = (jit_uint16_t)u;    break;
 684 #  if __WORDSIZE == 32
 685         case jit_code_putargi_i:                                break;
 686 #  else
 687         case jit_code_putargi_i:        u = (jit_int32_t)u;     break;
 688         case jit_code_putargi_ui:       u = (jit_uint32_t)u;    break;
 689         case jit_code_putargi_l:                                break;
 690 #  endif
 691         default:                        abort();                break;
 692     }
 693 #endif
 694     if (jit_arg_reg_p(v->u.w))
 695         jit_movi(JIT_RA0 - v->u.w, u);
 696     else {
 697         jit_node_t      *node;
 698         regno = jit_get_reg(jit_class_gpr);
 699         jit_movi(regno, u);
 700 #if PACKED_STACK
 701         switch (code) {
 702             case jit_code_putargi_c:    case jit_code_putargi_uc:
 703                 node = jit_stxi_c(v->u.w, JIT_FP, regno);       break;
 704             case jit_code_putargi_s:    case jit_code_putargi_us:
 705                 node = jit_stxi_s(v->u.w, JIT_FP, regno);       break;
 706 #  if __WORDSIZE == 32
 707             case jit_code_putargi_i:
 708                 node = jit_stxi(v->u.w, JIT_FP, regno);         break;
 709 #  else
 710             case jit_code_putargi_i:    case jit_code_putargi_ui:
 711                 node = jit_stxi_i(v->u.w, JIT_FP, regno);       break;
 712             case jit_code_putargi_l:
 713                 node = jit_stxi(v->u.w, JIT_FP, regno);         break;
 714 #  endif
 715             default:                    abort();                break;
 716         }
 717 #else
 718         node = jit_stxi(v->u.w, JIT_FP, regno);
 719 #endif
 720         jit_link_alist(node);
 721         jit_unget_reg(regno);
 722     }
 723     jit_dec_synth();
 724 }
 725
 726 void
 727 _jit_getarg_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 728 {
 729     assert(v->code == jit_code_arg_f);
 730     jit_inc_synth_wp(getarg_f, u, v);
 731     if (jit_arg_reg_p(v->u.w))
 732         jit_movr_f(u, JIT_FA0 - v->u.w);
 733     else {
 734         jit_node_t      *node = jit_ldxi_f(u, JIT_FP, v->u.w);
 735         jit_link_alist(node);
 736     }
 737     jit_dec_synth();
 738 }
 739
 740 void
 741 _jit_putargr_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 742 {
 743     assert(v->code == jit_code_arg_f);
 744     jit_inc_synth_wp(putargr_f, u, v);
 745     if (jit_arg_f_reg_p(v->u.w))
 746         jit_movr_f(JIT_FA0 - v->u.w, u);
 747     else {
 748         jit_node_t      *node = jit_stxi_f(v->u.w, JIT_FP, u);
 749         jit_link_alist(node);
 750     }
 751     jit_dec_synth();
 752 }
 753
 754 void
 755 _jit_putargi_f(jit_state_t *_jit, jit_float32_t u, jit_node_t *v)
 756 {
 757     jit_int32_t         regno;
 758     assert(v->code == jit_code_arg_f);
 759     jit_inc_synth_fp(putargi_f, u, v);
 760     if (jit_arg_f_reg_p(v->u.w))
 761         jit_movi_f(JIT_FA0 - v->u.w, u);
 762     else {
 763         jit_node_t      *node;
 764         regno = jit_get_reg(jit_class_fpr);
 765         jit_movi_f(regno, u);
 766         node = jit_stxi_f(v->u.w, JIT_FP, regno);
 767         jit_link_alist(node);
 768         jit_unget_reg(regno);
 769     }
 770     jit_dec_synth();
 771 }
 772
 773 void
 774 _jit_getarg_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 775 {
 776     assert(v->code == jit_code_arg_d);
 777     jit_inc_synth_wp(getarg_d, u, v);
 778     if (jit_arg_f_reg_p(v->u.w))
 779         jit_movr_d(u, JIT_FA0 - v->u.w);
 780     else {
 781         jit_node_t      *node = jit_ldxi_d(u, JIT_FP, v->u.w);
 782         jit_link_alist(node);
 783     }
 784     jit_dec_synth();
 785 }
 786
 787 void
 788 _jit_putargr_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 789 {
 790     assert(v->code == jit_code_arg_d);
 791     jit_inc_synth_wp(putargr_d, u, v);
 792     if (jit_arg_reg_p(v->u.w))
 793         jit_movr_d(JIT_FA0 - v->u.w, u);
 794     else {
 795         jit_node_t      *node = jit_stxi_d(v->u.w, JIT_FP, u);
 796         jit_link_alist(node);
 797     }
 798     jit_dec_synth();
 799 }
 800
 801 void
 802 _jit_putargi_d(jit_state_t *_jit, jit_float64_t u, jit_node_t *v)
 803 {
 804     jit_int32_t         regno;
 805     assert(v->code == jit_code_arg_d);
 806     jit_inc_synth_dp(putargi_d, u, v);
 807     if (jit_arg_reg_p(v->u.w))
 808         jit_movi_d(JIT_FA0 - v->u.w, u);
 809     else {
 810         jit_node_t      *node;
 811         regno = jit_get_reg(jit_class_fpr);
 812         jit_movi_d(regno, u);
 813         node = jit_stxi_d(v->u.w, JIT_FP, regno);
 814         jit_link_alist(node);
 815         jit_unget_reg(regno);
 816     }
 817     jit_dec_synth();
 818 }
 819
 820 void
 821 _jit_pushargr(jit_state_t *_jit, jit_int32_t u, jit_code_t code)
 822 {
 823     assert(_jitc->function);
 824     jit_code_inc_synth_w(code, u);
 825     jit_link_prepare();
 826     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 827         jit_int32_t     regno = JIT_RA0 - _jitc->function->call.argi;
 828 #if PACKED_STACK
 829         switch (code) {
 830             case jit_code_pushargr_c:   jit_extr_c(regno, u);   break;
 831             case jit_code_pushargr_uc:  jit_extr_uc(regno, u);  break;
 832             case jit_code_pushargr_s:   jit_extr_s(regno, u);   break;
 833             case jit_code_pushargr_us:  jit_extr_us(regno, u);  break;
 834 #  if __WORDISZE == 32
 835             case jit_code_pushargr_i:   jit_movr(regno, u);     break;
 836 #  else
 837             case jit_code_pushargr_i:   jit_extr_i(regno, u);   break;
 838             case jit_code_pushargr_ui:  jit_extr_ui(regno, u);  break;
 839             case jit_code_pushargr_l:   jit_movr(regno, u);     break;
 840 #  endif
 841             default:                    abort();                break;
 842         }
 843 #else
 844         jit_movr(regno, u);
 845 #endif
 846 #if __APPLE__
 847         if (_jitc->function->call.call & jit_call_varargs) {
 848             assert(code == jit_code_pushargr);
 849             jit_stxi(_jitc->function->call.size, JIT_SP, u);
 850             _jitc->function->call.size += sizeof(jit_word_t);
 851         }
 852 #endif
 853         ++_jitc->function->call.argi;
 854     }
 855     else {
 856 #if PACKED_STACK
 857         _jitc->function->call.size +=
 858             _jitc->function->call.size &
 859                 ((1 << ((code - jit_code_pushargr_c) >> 2)) - 1);
 860         switch (code) {
 861             case jit_code_pushargr_c:   case jit_code_pushargr_uc:
 862                 jit_stxi_c(_jitc->function->call.size, JIT_SP, u);
 863                 break;
 864             case jit_code_pushargr_s:   case jit_code_pushargr_us:
 865                 jit_stxi_s(_jitc->function->call.size, JIT_SP, u);
 866                 break;
 867 #  if __WORDSIZE == 32
 868             case jit_code_pushargr_i:
 869                 jit_stxi(_jitc->function->call.size, JIT_SP, u);
 870                 break;
 871 #  else
 872             case jit_code_pushargr_i:   case jit_code_pushargr_ui:
 873                 jit_stxi_i(_jitc->function->call.size, JIT_SP, u);
 874                 break;
 875             case jit_code_pushargr_l:
 876                 jit_stxi(_jitc->function->call.size, JIT_SP, u);
 877                 break;
 878 #  endif
 879             default:
 880                 abort();
 881                 break;
 882         }
 883         _jitc->function->call.size += 1 << ((code - jit_code_pushargr_c) >> 2);
 884 #else
 885         jit_stxi(_jitc->function->call.size, JIT_SP, u);
 886         _jitc->function->call.size += sizeof(jit_word_t);
 887 #endif
 888         jit_check_frame();
 889     }
 890     jit_dec_synth();
 891 }
 892
 893 void
 894 _jit_pushargi(jit_state_t *_jit, jit_word_t u, jit_code_t code)
 895 {
 896     jit_int32_t          regno;
 897     assert(_jitc->function);
 898     jit_code_inc_synth_w(code, u);
 899     jit_link_prepare();
 900 #if PACKED_STACK
 901     switch (code) {
 902         case jit_code_pushargi_c:       u = (jit_int8_t)u;      break;
 903         case jit_code_pushargi_uc:      u = (jit_uint8_t)u;     break;
 904         case jit_code_pushargi_s:       u = (jit_int16_t)u;     break;
 905         case jit_code_pushargi_us:      u = (jit_uint16_t)u;    break;
 906 #  if __WORDSIZE == 32
 907         case jit_code_pushargi_i:                               break;
 908 #  else
 909         case jit_code_pushargi_i:       u = (jit_int32_t)u;     break;
 910         case jit_code_pushargi_ui:      u = (jit_uint32_t)u;    break;
 911         case jit_code_pushargi_l:                               break;
 912 #  endif
 913         default:                        abort();                break;
 914     }
 915 #endif
 916     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 917         regno = JIT_RA0 - _jitc->function->call.argi;
 918         jit_movi(regno, u);
 919 #if __APPLE__
 920         if (_jitc->function->call.call & jit_call_varargs) {
 921             assert(code == jit_code_pushargi);
 922             jit_stxi(_jitc->function->call.size, JIT_SP, regno);
 923             _jitc->function->call.size += sizeof(jit_word_t);
 924         }
 925 #endif
 926         ++_jitc->function->call.argi;
 927     }
 928     else {
 929         regno = jit_get_reg(jit_class_gpr);
 930         jit_movi(regno, u);
 931 #if PACKED_STACK
 932         _jitc->function->call.size +=
 933             _jitc->function->call.size &
 934                 ((1 << ((code - jit_code_pushargr_c) >> 2)) - 1);
 935         switch (code) {
 936             case jit_code_pushargi_c:   case jit_code_pushargi_uc:
 937                 jit_stxi_c(_jitc->function->call.size, JIT_SP, regno);
 938                 break;
 939             case jit_code_pushargi_s:   case jit_code_pushargi_us:
 940                 jit_stxi_s(_jitc->function->call.size, JIT_SP, regno);
 941                 break;
 942 #  if __WORDSIZE == 32
 943             case jit_code_pushargi_i:
 944                 jit_stxi(_jitc->function->call.size, JIT_SP, regno);
 945                 break;
 946 #  else
 947             case jit_code_pushargi_i:   case jit_code_pushargi_ui:
 948                 jit_stxi_i(_jitc->function->call.size, JIT_SP, regno);
 949                 break;
 950             case jit_code_pushargi_l:
 951                 jit_stxi(_jitc->function->call.size, JIT_SP, regno);
 952                 break;
 953 #  endif
 954             default:
 955                 abort();
 956                 break;
 957         }
 958         _jitc->function->call.size += 1 << ((code - jit_code_pushargr_c) >> 2);
 959 #else
 960         jit_stxi(_jitc->function->call.size, JIT_SP, regno);
 961         _jitc->function->call.size += sizeof(jit_word_t);
 962 #endif
 963         jit_unget_reg(regno);
 964         jit_check_frame();
 965     }
 966     jit_dec_synth();
 967 }
 968
 969 void
 970 _jit_pushargr_f(jit_state_t *_jit, jit_int32_t u)
 971 {
 972     assert(_jitc->function);
 973     jit_inc_synth_w(pushargr_f, u);
 974     jit_link_prepare();
 975     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 976         jit_movr_f(JIT_FA0 - _jitc->function->call.argf, u);
 977 #if __APPLE__
 978         if (_jitc->function->call.call & jit_call_varargs) {
 979             assert(sizeof(jit_float32_t) == sizeof(jit_word_t));
 980             jit_stxi_f(_jitc->function->call.size, JIT_SP,
 981                        JIT_FA0 - _jitc->function->call.argf);
 982             _jitc->function->call.size += sizeof(jit_word_t);
 983         }
 984 #endif
 985         ++_jitc->function->call.argf;
 986     }
 987     else {
 988 #if PACKED_STACK
 989         _jitc->function->call.size +=
 990             _jitc->function->call.size & (sizeof(jit_float32_t) - 1);
 991         jit_stxi_f(_jitc->function->call.size, JIT_SP, u);
 992         _jitc->function->call.size += sizeof(jit_float32_t);
 993 #else
 994         jit_stxi_f(_jitc->function->call.size, JIT_SP, u);
 995         _jitc->function->call.size += sizeof(jit_word_t);
 996 #endif
 997         jit_check_frame();
 998     }
 999     jit_dec_synth();
1000 }
1001
1002 void
1003 _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
1004 {
1005     jit_int32_t         regno;
1006     assert(_jitc->function);
1007     jit_inc_synth_f(pushargi_f, u);
1008     jit_link_prepare();
1009     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
1010         jit_movi_f(JIT_FA0 - _jitc->function->call.argf, u);
1011 #if __APPLE__
1012         if (_jitc->function->call.call & jit_call_varargs) {
1013             assert(sizeof(jit_float32_t) == sizeof(jit_word_t));
1014             jit_stxi_f(_jitc->function->call.size, JIT_SP,
1015                        JIT_FA0 - _jitc->function->call.argf);
1016             _jitc->function->call.size += sizeof(jit_word_t);
1017         }
1018 #endif
1019         ++_jitc->function->call.argf;
1020     }
1021     else {
1022         regno = jit_get_reg(jit_class_fpr);
1023         jit_movi_f(regno, u);
1024 #if PACKED_STACK
1025         _jitc->function->call.size +=
1026             _jitc->function->call.size & (sizeof(jit_float32_t) - 1);
1027         jit_stxi_f(_jitc->function->call.size, JIT_SP, regno);
1028         _jitc->function->call.size += sizeof(jit_float32_t);
1029 #else
1030         jit_stxi_f(_jitc->function->call.size, JIT_SP, regno);
1031         _jitc->function->call.size += sizeof(jit_word_t);
1032 #endif
1033         jit_unget_reg(regno);
1034         jit_check_frame();
1035     }
1036     jit_dec_synth();
1037 }
1038
1039 void
1040 _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
1041 {
1042     assert(_jitc->function);
1043     jit_inc_synth_w(pushargr_d, u);
1044     jit_link_prepare();
1045     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
1046         jit_movr_d(JIT_FA0 - _jitc->function->call.argf, u);
1047 #if __APPLE__
1048         if (_jitc->function->call.call & jit_call_varargs) {
1049             assert(sizeof(jit_float64_t) == sizeof(jit_word_t));
1050             jit_stxi_d(_jitc->function->call.size, JIT_SP,
1051                        JIT_FA0 - _jitc->function->call.argf);
1052             _jitc->function->call.size += sizeof(jit_float64_t);
1053         }
1054 #endif
1055         ++_jitc->function->call.argf;
1056     }
1057     else {
1058 #if PACKED_STACK
1059         _jitc->function->call.size +=
1060             _jitc->function->call.size & (sizeof(jit_float64_t) - 1);
1061 #endif
1062         jit_stxi_d(_jitc->function->call.size, JIT_SP, u);
1063         _jitc->function->call.size += sizeof(jit_float64_t);
1064         jit_check_frame();
1065     }
1066     jit_dec_synth();
1067 }
1068
1069 void
1070 _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
1071 {
1072     jit_int32_t         regno;
1073     assert(_jitc->function);
1074     jit_inc_synth_d(pushargi_d, u);
1075     jit_link_prepare();
1076     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
1077         jit_movi_d(JIT_FA0 - _jitc->function->call.argf, u);
1078 #if __APPLE__
1079         if (_jitc->function->call.call & jit_call_varargs) {
1080             assert(sizeof(jit_float64_t) == sizeof(jit_word_t));
1081             jit_stxi_d(_jitc->function->call.size, JIT_SP,
1082                        JIT_FA0 - _jitc->function->call.argf);
1083             _jitc->function->call.size += sizeof(jit_float64_t);
1084         }
1085 #endif
1086         ++_jitc->function->call.argf;
1087     }
1088     else {
1089         regno = jit_get_reg(jit_class_fpr);
1090         jit_movi_d(regno, u);
1091 #if PACKED_STACK
1092         _jitc->function->call.size +=
1093             _jitc->function->call.size & (sizeof(jit_float64_t) - 1);
1094 #endif
1095         jit_stxi_d(_jitc->function->call.size, JIT_SP, regno);
1096         jit_unget_reg(regno);
1097         _jitc->function->call.size += sizeof(jit_float64_t);
1098         jit_check_frame();
1099     }
1100     jit_dec_synth();
1101 }
1102
1103 jit_bool_t
1104 _jit_regarg_p(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno)
1105 {
1106     jit_int32_t         spec;
1107     spec = jit_class(_rvs[regno].spec);
1108     if (spec & jit_class_arg) {
1109         regno = JIT_RA0 - regno;
1110         if (regno >= 0 && regno < node->v.w)
1111             return (1);
1112         if (spec & jit_class_fpr) {
1113             regno = JIT_FA0 - regno;
1114             if (regno >= 0 && regno < node->w.w)
1115                 return (1);
1116         }
1117     }
1118
1119     return (0);
1120 }
1121
1122 void
1123 _jit_finishr(jit_state_t *_jit, jit_int32_t r0)
1124 {
1125     jit_node_t          *node;
1126     assert(_jitc->function);
1127     jit_check_frame();
1128     jit_inc_synth_w(finishr, r0);
1129 #if PACKED_STACK
1130     _jitc->function->call.size +=
1131         _jitc->function->call.size & (sizeof(jit_word_t) - 1);
1132 #endif
1133     if (_jitc->function->self.alen < _jitc->function->call.size)
1134         _jitc->function->self.alen = _jitc->function->call.size;
1135     node = jit_callr(r0);
1136     node->v.w = _jitc->function->self.argi;
1137     node->w.w = _jitc->function->call.argf;
1138     _jitc->function->call.argi = _jitc->function->call.argf =
1139         _jitc->function->call.size = 0;
1140     _jitc->prepare = 0;
1141     jit_dec_synth();
1142 }
1143
1144 jit_node_t *
1145 _jit_finishi(jit_state_t *_jit, jit_pointer_t i0)
1146 {
1147     jit_node_t          *node;
1148     assert(_jitc->function);
1149     jit_check_frame();
1150     jit_inc_synth_w(finishi, (jit_word_t)i0);
1151 #if PACKED_STACK
1152     _jitc->function->call.size +=
1153         _jitc->function->call.size & (sizeof(jit_word_t) - 1);
1154 #endif
1155     if (_jitc->function->self.alen < _jitc->function->call.size)
1156         _jitc->function->self.alen = _jitc->function->call.size;
1157     node = jit_calli(i0);
1158     node->v.w = _jitc->function->call.argi;
1159     node->w.w = _jitc->function->call.argf;
1160     _jitc->function->call.argi = _jitc->function->call.argf =
1161         _jitc->function->call.size = 0;
1162     _jitc->prepare = 0;
1163     jit_dec_synth();
1164     return (node);
1165 }
1166
1167 void
1168 _jit_retval_c(jit_state_t *_jit, jit_int32_t r0)
1169 {
1170     jit_inc_synth_w(retval_c, r0);
1171     jit_extr_c(r0, JIT_RET);
1172     jit_dec_synth();
1173 }
1174
1175 void
1176 _jit_retval_uc(jit_state_t *_jit, jit_int32_t r0)
1177 {
1178     jit_inc_synth_w(retval_uc, r0);
1179     jit_extr_uc(r0, JIT_RET);
1180     jit_dec_synth();
1181 }
1182
1183 void
1184 _jit_retval_s(jit_state_t *_jit, jit_int32_t r0)
1185 {
1186     jit_inc_synth_w(retval_s, r0);
1187     jit_extr_s(r0, JIT_RET);
1188     jit_dec_synth();
1189 }
1190
1191 void
1192 _jit_retval_us(jit_state_t *_jit, jit_int32_t r0)
1193 {
1194     jit_inc_synth_w(retval_us, r0);
1195     jit_extr_us(r0, JIT_RET);
1196     jit_dec_synth();
1197 }
1198
1199 void
1200 _jit_retval_i(jit_state_t *_jit, jit_int32_t r0)
1201 {
1202     jit_inc_synth_w(retval_i, r0);
1203 #if __WORDSIZE == 32
1204     jit_movr(r0, JIT_RET);
1205 #else
1206     jit_extr_i(r0, JIT_RET);
1207 #endif
1208     jit_dec_synth();
1209 }
1210
1211 #if __WORDSIZE == 64
1212 void
1213 _jit_retval_ui(jit_state_t *_jit, jit_int32_t r0)
1214 {
1215     jit_inc_synth_w(retval_ui, r0);
1216     jit_extr_ui(r0, JIT_RET);
1217     jit_dec_synth();
1218 }
1219
1220 void
1221 _jit_retval_l(jit_state_t *_jit, jit_int32_t r0)
1222 {
1223     jit_inc_synth_w(retval_l, r0);
1224     jit_movr(r0, JIT_RET);
1225     jit_dec_synth();
1226 }
1227 #endif
1228
1229 void
1230 _jit_retval_f(jit_state_t *_jit, jit_int32_t r0)
1231 {
1232     jit_inc_synth_w(retval_f, r0);
1233     if (r0 != JIT_FRET)
1234         jit_movr_f(r0, JIT_FRET);
1235     jit_dec_synth();
1236 }
1237
1238 void
1239 _jit_retval_d(jit_state_t *_jit, jit_int32_t r0)
1240 {
1241     jit_inc_synth_w(retval_d, r0);
1242     if (r0 != JIT_FRET)
1243         jit_movr_d(r0, JIT_FRET);
1244     jit_dec_synth();
1245 }
1246
1247 jit_pointer_t
1248 _emit_code(jit_state_t *_jit)
1249 {
1250     jit_node_t          *node;
1251     jit_node_t          *temp;
1252     jit_word_t           word;
1253     jit_word_t           value;
1254     jit_int32_t          offset;
1255     struct {
1256         jit_node_t      *node;
1257         jit_uint8_t     *data;
1258         jit_word_t       word;
1259         jit_function_t   func;
1260 #if DEVEL_DISASSEMBLER
1261         jit_word_t       prevw;
1262 #endif
1263         jit_int32_t      const_offset;
1264         jit_int32_t      patch_offset;
1265     } undo;
1266 #if DEVEL_DISASSEMBLER
1267     jit_word_t           prevw;
1268 #endif
1269
1270     _jitc->function = NULL;
1271
1272     jit_reglive_setup();
1273
1274     undo.word = 0;
1275     undo.node = NULL;
1276     undo.const_offset = undo.patch_offset = 0;
1277 #  define assert_data(node)             /**/
1278 #define case_rr(name, type)                                             \
1279             case jit_code_##name##r##type:                              \
1280                 name##r##type(rn(node->u.w), rn(node->v.w));            \
1281                 break
1282 #define case_rw(name, type)                                             \
1283             case jit_code_##name##i##type:                              \
1284                 name##i##type(rn(node->u.w), node->v.w);                \
1285                 break
1286 #define case_wr(name, type)                                             \
1287             case jit_code_##name##i##type:                              \
1288                 name##i##type(node->u.w, rn(node->v.w));                \
1289                 break
1290 #define case_rrr(name, type)                                            \
1291             case jit_code_##name##r##type:                              \
1292                 name##r##type(rn(node->u.w),                            \
1293                               rn(node->v.w), rn(node->w.w));            \
1294                 break
1295 #define case_rrrr(name, type)                                           \
1296             case jit_code_##name##r##type:                              \
1297                 name##r##type(rn(node->u.q.l), rn(node->u.q.h),         \
1298                               rn(node->v.w), rn(node->w.w));            \
1299                 break
1300 #define case_rqr(name, type)                                            \
1301             case jit_code_##name##r##type:                              \
1302                 name##r##type(rn(node->u.w), rn(node->v.q.l),           \
1303                               rn(node->v.q.h), rn(node->w.w));          \
1304             case jit_code_##name##i##type:                              \
1305                 break;
1306 #define case_rrw(name, type)                                            \
1307             case jit_code_##name##i##type:                              \
1308                 name##i##type(rn(node->u.w), rn(node->v.w), node->w.w); \
1309                 break
1310 #define case_rrrw(name, type)                                           \
1311             case jit_code_##name##i##type:                              \
1312                 name##i##type(rn(node->u.q.l), rn(node->u.q.h),         \
1313                               rn(node->v.w), node->w.w);                \
1314                 break
1315 #define case_rrf(name)                                                  \
1316             case jit_code_##name##i_f:                                  \
1317                 assert_data(node);                                      \
1318                 name##i_f(rn(node->u.w), rn(node->v.w), node->w.f);     \
1319                 break
1320 #define case_rrd(name)                                                  \
1321             case jit_code_##name##i_d:                                  \
1322                 assert_data(node);                                      \
1323                 name##i_d(rn(node->u.w), rn(node->v.w), node->w.d);     \
1324                 break
1325 #define case_wrr(name, type)                                            \
1326             case jit_code_##name##i##type:                              \
1327                 name##i##type(node->u.w, rn(node->v.w), rn(node->w.w)); \
1328                 break
1329 #define case_brr(name, type)                                            \
1330             case jit_code_##name##r##type:                              \
1331                 temp = node->u.n;                                       \
1332                 assert(temp->code == jit_code_label ||                  \
1333                        temp->code == jit_code_epilog);                  \
1334                 if (temp->flag & jit_flag_patch)                        \
1335                     name##r##type(temp->u.w, rn(node->v.w),             \
1336                                   rn(node->w.w));                       \
1337                 else {                                                  \
1338                     word = name##r##type(_jit->pc.w,                    \
1339                                          rn(node->v.w), rn(node->w.w)); \
1340                     patch(word, node);                                  \
1341                 }                                                       \
1342                 break
1343 #define case_brw(name, type)                                            \
1344             case jit_code_##name##i##type:                              \
1345                 temp = node->u.n;                                       \
1346                 assert(temp->code == jit_code_label ||                  \
1347                        temp->code == jit_code_epilog);                  \
1348                 if (temp->flag & jit_flag_patch)                        \
1349                     name##i##type(temp->u.w,                            \
1350                                   rn(node->v.w), node->w.w);            \
1351                 else {                                                  \
1352                     word = name##i##type(_jit->pc.w,                    \
1353                                          rn(node->v.w), node->w.w);     \
1354                     patch(word, node);                                  \
1355                 }                                                       \
1356                 break;
1357 #define case_brf(name)                                                  \
1358             case jit_code_##name##i_f:                                  \
1359                 temp = node->u.n;                                       \
1360                 assert(temp->code == jit_code_label ||                  \
1361                        temp->code == jit_code_epilog);                  \
1362                 if (temp->flag & jit_flag_patch)                        \
1363                     name##i_f(temp->u.w, rn(node->v.w), node->w.f);     \
1364                 else {                                                  \
1365                     word = name##i_f(_jit->pc.w, rn(node->v.w),         \
1366                                      node->w.f);                        \
1367                     patch(word, node);                                  \
1368                 }                                                       \
1369                 break
1370 #define case_brd(name)                                                  \
1371             case jit_code_##name##i_d:                                  \
1372                 temp = node->u.n;                                       \
1373                 assert(temp->code == jit_code_label ||                  \
1374                        temp->code == jit_code_epilog);                  \
1375                 if (temp->flag & jit_flag_patch)                        \
1376                     name##i_d(temp->u.w, rn(node->v.w), node->w.d);     \
1377                 else {                                                  \
1378                     word = name##i_d(_jit->pc.w, rn(node->v.w),         \
1379                                      node->w.d);                        \
1380                     patch(word, node);                                  \
1381                 }                                                       \
1382                 break
1383 #if DEVEL_DISASSEMBLER
1384     prevw = _jit->pc.w;
1385 #endif
1386     for (node = _jitc->head; node; node = node->next) {
1387         if (_jit->pc.uc >= _jitc->code.end)
1388             return (NULL);
1389
1390 #if DEVEL_DISASSEMBLER
1391         node->offset = (jit_uword_t)_jit->pc.w - (jit_uword_t)prevw;
1392         prevw = _jit->pc.w;
1393 #endif
1394         value = jit_classify(node->code);
1395         jit_regarg_set(node, value);
1396         switch (node->code) {
1397             case jit_code_align:
1398                 /* Must align to a power of two */
1399                 assert(!(node->u.w & (node->u.w - 1)));
1400                 if ((word = _jit->pc.w & (node->u.w - 1)))
1401                     nop(node->u.w - word);
1402                 break;
1403             case jit_code_skip:
1404                 nop((node->u.w + 3) & ~3);
1405                 break;
1406             case jit_code_note:         case jit_code_name:
1407                 node->u.w = _jit->pc.w;
1408                 break;
1409             case jit_code_label:
1410                 /* remember label is defined */
1411                 node->flag |= jit_flag_patch;
1412                 node->u.w = _jit->pc.w;
1413                 break;
1414                 case_rrr(add,);
1415                 case_rrw(add,);
1416                 case_rrr(addc,);
1417                 case_rrw(addc,);
1418                 case_rrr(addx,);
1419                 case_rrw(addx,);
1420                 case_rrr(sub,);
1421                 case_rrw(sub,);
1422                 case_rrr(subc,);
1423                 case_rrw(subc,);
1424                 case_rrr(subx,);
1425                 case_rrw(subx,);
1426                 case_rrw(rsb,);
1427                 case_rrr(mul,);
1428                 case_rrw(mul,);
1429                 case_rrr(hmul,);
1430                 case_rrw(hmul,);
1431                 case_rrr(hmul, _u);
1432                 case_rrw(hmul, _u);
1433                 case_rrrr(qmul,);
1434                 case_rrrw(qmul,);
1435                 case_rrrr(qmul, _u);
1436                 case_rrrw(qmul, _u);
1437                 case_rrr(div,);
1438                 case_rrw(div,);
1439                 case_rrr(div, _u);
1440                 case_rrw(div, _u);
1441                 case_rrrr(qdiv,);
1442                 case_rrrw(qdiv,);
1443                 case_rrrr(qdiv, _u);
1444                 case_rrrw(qdiv, _u);
1445                 case_rrr(rem,);
1446                 case_rrw(rem,);
1447                 case_rrr(rem, _u);
1448                 case_rrw(rem, _u);
1449                 case_rrr(lsh,);
1450                 case_rrw(lsh,);
1451                 case_rrrr(qlsh,);
1452                 case_rrrw(qlsh,);
1453                 case_rrrr(qlsh, _u);
1454                 case_rrrw(qlsh, _u);
1455                 case_rrr(rsh,);
1456                 case_rrw(rsh,);
1457                 case_rrr(rsh, _u);
1458                 case_rrw(rsh, _u);
1459                 case_rrrr(qrsh,);
1460                 case_rrrw(qrsh,);
1461                 case_rrrr(qrsh, _u);
1462                 case_rrrw(qrsh, _u);
1463                 case_rrr(lrot,);
1464                 case_rrw(lrot,);
1465                 case_rrr(rrot,);
1466                 case_rrw(rrot,);
1467                 case_rr(neg,);
1468                 case_rr(com,);
1469                 case_rr(clo,);
1470                 case_rr(clz,);
1471                 case_rr(cto,);
1472                 case_rr(ctz,);
1473                 case_rr(rbit,);
1474                 case_rr(popcnt,);
1475                 case_rrr(and,);
1476                 case_rrw(and,);
1477                 case_rrr(or,);
1478                 case_rrw(or,);
1479                 case_rrr(xor,);
1480                 case_rrw(xor,);
1481                 case_rr(trunc, _f_i);
1482                 case_rr(trunc, _d_i);
1483                 case_rr(trunc, _f_l);
1484                 case_rr(trunc, _d_l);
1485                 case_rr(ld, _c);
1486                 case_rw(ld, _c);
1487                 case_rr(ld, _uc);
1488                 case_rw(ld, _uc);
1489                 case_rr(ld, _s);
1490                 case_rw(ld, _s);
1491                 case_rr(ld, _us);
1492                 case_rw(ld, _us);
1493                 case_rr(ld, _i);
1494                 case_rw(ld, _i);
1495                 case_rr(ld, _ui);
1496                 case_rw(ld, _ui);
1497                 case_rr(ld, _l);
1498                 case_rw(ld, _l);
1499                 case_rrr(ldx, _c);
1500                 case_rrw(ldx, _c);
1501                 case_rrr(ldx, _uc);
1502                 case_rrw(ldx, _uc);
1503                 case_rrr(ldx, _s);
1504                 case_rrw(ldx, _s);
1505                 case_rrr(ldx, _us);
1506                 case_rrw(ldx, _us);
1507                 case_rrr(ldx, _i);
1508                 case_rrw(ldx, _i);
1509                 case_rrr(ldx, _ui);
1510                 case_rrw(ldx, _ui);
1511                 case_rrr(ldx, _l);
1512                 case_rrw(ldx, _l);
1513             case jit_code_unldr:
1514                 unldr(rn(node->u.w), rn(node->v.w), node->w.w);
1515                 break;
1516             case jit_code_unldi:
1517                 unldi(rn(node->u.w), node->v.w, node->w.w);
1518                 break;
1519             case jit_code_unldr_u:
1520                 unldr_u(rn(node->u.w), rn(node->v.w), node->w.w);
1521                 break;
1522             case jit_code_unldi_u:
1523                 unldi_u(rn(node->u.w), node->v.w, node->w.w);
1524                 break;
1525                 case_rr(st, _c);
1526                 case_wr(st, _c);
1527                 case_rr(st, _s);
1528                 case_wr(st, _s);
1529                 case_rr(st, _i);
1530                 case_wr(st, _i);
1531                 case_rr(st, _l);
1532                 case_wr(st, _l);
1533                 case_rrr(stx, _c);
1534                 case_wrr(stx, _c);
1535                 case_rrr(stx, _s);
1536                 case_wrr(stx, _s);
1537                 case_rrr(stx, _i);
1538                 case_wrr(stx, _i);
1539                 case_rrr(stx, _l);
1540                 case_wrr(stx, _l);
1541             case jit_code_unstr:
1542                 unstr(rn(node->u.w), rn(node->v.w), node->w.w);
1543                 break;
1544             case jit_code_unsti:
1545                 unsti(node->u.w, rn(node->v.w), node->w.w);
1546                 break;
1547                 case_rr(hton, _us);
1548                 case_rr(hton, _ui);
1549                 case_rr(hton, _ul);
1550                 case_rr(bswap, _us);
1551                 case_rr(bswap, _ui);
1552                 case_rr(bswap, _ul);
1553             case jit_code_extr:
1554                 extr(rn(node->u.w), rn(node->v.w), node->w.q.l, node->w.q.h);
1555                 break;
1556             case jit_code_extr_u:
1557                 extr_u(rn(node->u.w), rn(node->v.w), node->w.q.l, node->w.q.h);
1558                 break;
1559             case jit_code_depr:
1560                 depr(rn(node->u.w), rn(node->v.w), node->w.q.l, node->w.q.h);
1561                 break;
1562             case jit_code_depi:
1563                 depi(rn(node->u.w), node->v.w, node->w.q.l, node->w.q.h);
1564                 break;
1565                 case_rr(ext, _c);
1566                 case_rr(ext, _uc);
1567                 case_rr(ext, _s);
1568                 case_rr(ext, _us);
1569                 case_rr(ext, _i);
1570                 case_rr(ext, _ui);
1571             case jit_code_casr:
1572                 casr(rn(node->u.w), rn(node->v.w),
1573                      rn(node->w.q.l), rn(node->w.q.h));
1574                 break;
1575             case jit_code_casi:
1576                 casi(rn(node->u.w), node->v.w,
1577                      rn(node->w.q.l), rn(node->w.q.h));
1578                 break;
1579                 case_rr(mov,);
1580                 case_rrr(movn,);
1581                 case_rrr(movz,);
1582             case jit_code_movi:
1583                 if (node->flag & jit_flag_node) {
1584                     temp = node->v.n;
1585                     if (temp->code == jit_code_data ||
1586                         (temp->code == jit_code_label &&
1587                          (temp->flag & jit_flag_patch)))
1588                         movi(rn(node->u.w), temp->u.w);
1589                     else {
1590                         assert(temp->code == jit_code_label ||
1591                                temp->code == jit_code_epilog);
1592                         word = movi_p(rn(node->u.w), temp->u.w);
1593                         patch(word, node);
1594                     }
1595                 }
1596                 else
1597                     movi(rn(node->u.w), node->v.w);
1598                 break;
1599                 case_rrr(lt,);
1600                 case_rrw(lt,);
1601                 case_rrr(lt, _u);
1602                 case_rrw(lt, _u);
1603                 case_rrr(le,);
1604                 case_rrw(le,);
1605                 case_rrr(le, _u);
1606                 case_rrw(le, _u);
1607                 case_rrr(eq,);
1608                 case_rrw(eq,);
1609                 case_rrr(ge,);
1610                 case_rrw(ge,);
1611                 case_rrr(ge, _u);
1612                 case_rrw(ge, _u);
1613                 case_rrr(gt,);
1614                 case_rrw(gt,);
1615                 case_rrr(gt, _u);
1616                 case_rrw(gt, _u);
1617                 case_rrr(ne,);
1618                 case_rrw(ne,);
1619                 case_brr(blt,);
1620                 case_brw(blt,);
1621                 case_brr(blt, _u);
1622                 case_brw(blt, _u);
1623                 case_brr(ble,);
1624                 case_brw(ble,);
1625                 case_brr(ble, _u);
1626                 case_brw(ble, _u);
1627                 case_brr(beq,);
1628                 case_brw(beq,);
1629                 case_brr(bge,);
1630                 case_brw(bge,);
1631                 case_brr(bge, _u);
1632                 case_brw(bge, _u);
1633                 case_brr(bgt,);
1634                 case_brw(bgt,);
1635                 case_brr(bgt, _u);
1636                 case_brw(bgt, _u);
1637                 case_brr(bne,);
1638                 case_brw(bne,);
1639                 case_brr(boadd,);
1640                 case_brw(boadd,);
1641                 case_brr(boadd, _u);
1642                 case_brw(boadd, _u);
1643                 case_brr(bxadd,);
1644                 case_brw(bxadd,);
1645                 case_brr(bxadd, _u);
1646                 case_brw(bxadd, _u);
1647                 case_brr(bosub,);
1648                 case_brw(bosub,);
1649                 case_brr(bosub, _u);
1650                 case_brw(bosub, _u);
1651                 case_brr(bxsub,);
1652                 case_brw(bxsub,);
1653                 case_brr(bxsub, _u);
1654                 case_brw(bxsub, _u);
1655                 case_brr(bms,);
1656                 case_brw(bms,);
1657                 case_brr(bmc,);
1658                 case_brw(bmc,);
1659                 case_rrr(add, _f);
1660                 case_rrf(add);
1661                 case_rrr(sub, _f);
1662                 case_rrf(sub);
1663                 case_rrf(rsb);
1664                 case_rrr(mul, _f);
1665                 case_rrf(mul);
1666                 case_rrr(div, _f);
1667                 case_rrf(div);
1668                 case_rr(abs, _f);
1669                 case_rr(neg, _f);
1670                 case_rr(sqrt, _f);
1671                 case_rqr(fma, _f);
1672                 case_rqr(fms, _f);
1673                 case_rqr(fnma, _f);
1674                 case_rqr(fnms, _f);
1675                 case_rr(ext, _f);
1676                 case_rr(ld, _f);
1677                 case_rw(ld, _f);
1678                 case_rrr(ldx, _f);
1679                 case_rrw(ldx, _f);
1680             case jit_code_unldr_x:
1681                 unldr_x(rn(node->u.w), rn(node->v.w), node->w.w);
1682                 break;
1683             case jit_code_unldi_x:
1684                 unldi_x(rn(node->u.w), node->v.w, node->w.w);
1685                 break;
1686                 case_rr(st, _f);
1687                 case_wr(st, _f);
1688                 case_rrr(stx, _f);
1689                 case_wrr(stx, _f);
1690             case jit_code_unstr_x:
1691                 unstr_x(rn(node->u.w), rn(node->v.w), node->w.w);
1692                 break;
1693             case jit_code_unsti_x:
1694                 unsti_x(node->u.w, rn(node->v.w), node->w.w);
1695                 break;
1696                 case_rr(mov, _f);
1697             case jit_code_movi_f:
1698                 assert_data(node);
1699                 movi_f(rn(node->u.w), node->v.f);
1700                 break;
1701                 case_rr(ext, _d_f);
1702                 case_rrr(lt, _f);
1703                 case_rrf(lt);
1704                 case_rrr(le, _f);
1705                 case_rrf(le);
1706                 case_rrr(eq, _f);
1707                 case_rrf(eq);
1708                 case_rrr(ge, _f);
1709                 case_rrf(ge);
1710                 case_rrr(gt, _f);
1711                 case_rrf(gt);
1712                 case_rrr(ne, _f);
1713                 case_rrf(ne);
1714                 case_rrr(unlt, _f);
1715                 case_rrf(unlt);
1716                 case_rrr(unle, _f);
1717                 case_rrf(unle);
1718                 case_rrr(uneq, _f);
1719                 case_rrf(uneq);
1720                 case_rrr(unge, _f);
1721                 case_rrf(unge);
1722                 case_rrr(ungt, _f);
1723                 case_rrf(ungt);
1724                 case_rrr(ltgt, _f);
1725                 case_rrf(ltgt);
1726                 case_rrr(ord, _f);
1727                 case_rrf(ord);
1728                 case_rrr(unord, _f);
1729                 case_rrf(unord);
1730                 case_brr(blt, _f);
1731                 case_brf(blt);
1732                 case_brr(ble, _f);
1733                 case_brf(ble);
1734                 case_brr(beq, _f);
1735                 case_brf(beq);
1736                 case_brr(bge, _f);
1737                 case_brf(bge);
1738                 case_brr(bgt, _f);
1739                 case_brf(bgt);
1740                 case_brr(bne, _f);
1741                 case_brf(bne);
1742                 case_brr(bunlt, _f);
1743                 case_brf(bunlt);
1744                 case_brr(bunle, _f);
1745                 case_brf(bunle);
1746                 case_brr(buneq, _f);
1747                 case_brf(buneq);
1748                 case_brr(bunge, _f);
1749                 case_brf(bunge);
1750                 case_brr(bungt, _f);
1751                 case_brf(bungt);
1752                 case_brr(bltgt, _f);
1753                 case_brf(bltgt);
1754                 case_brr(bord, _f);
1755                 case_brf(bord);
1756                 case_brr(bunord, _f);
1757                 case_brf(bunord);
1758                 case_rrr(add, _d);
1759                 case_rrd(add);
1760                 case_rrr(sub, _d);
1761                 case_rrd(sub);
1762                 case_rrd(rsb);
1763                 case_rrr(mul, _d);
1764                 case_rrd(mul);
1765                 case_rrr(div, _d);
1766                 case_rrd(div);
1767                 case_rr(abs, _d);
1768                 case_rr(neg, _d);
1769                 case_rr(sqrt, _d);
1770                 case_rqr(fma, _d);
1771                 case_rqr(fms, _d);
1772                 case_rqr(fnma, _d);
1773                 case_rqr(fnms, _d);
1774                 case_rr(ext, _d);
1775                 case_rr(ld, _d);
1776                 case_rw(ld, _d);
1777                 case_rrr(ldx, _d);
1778                 case_rrw(ldx, _d);
1779                 case_rr(st, _d);
1780                 case_wr(st, _d);
1781                 case_rrr(stx, _d);
1782                 case_wrr(stx, _d);
1783                 case_rr(mov, _d);
1784             case jit_code_movi_d:
1785                 assert_data(node);
1786                 movi_d(rn(node->u.w), node->v.d);
1787                 break;
1788                 case_rr(ext, _f_d);
1789                 case_rrr(lt, _d);
1790                 case_rrd(lt);
1791                 case_rrr(le, _d);
1792                 case_rrd(le);
1793                 case_rrr(eq, _d);
1794                 case_rrd(eq);
1795                 case_rrr(ge, _d);
1796                 case_rrd(ge);
1797                 case_rrr(gt, _d);
1798                 case_rrd(gt);
1799                 case_rrr(ne, _d);
1800                 case_rrd(ne);
1801                 case_rrr(unlt, _d);
1802                 case_rrd(unlt);
1803                 case_rrr(unle, _d);
1804                 case_rrd(unle);
1805                 case_rrr(uneq, _d);
1806                 case_rrd(uneq);
1807                 case_rrr(unge, _d);
1808                 case_rrd(unge);
1809                 case_rrr(ungt, _d);
1810                 case_rrd(ungt);
1811                 case_rrr(ltgt, _d);
1812                 case_rrd(ltgt);
1813                 case_rrr(ord, _d);
1814                 case_rrd(ord);
1815                 case_rrr(unord, _d);
1816                 case_rrd(unord);
1817                 case_brr(blt, _d);
1818                 case_brd(blt);
1819                 case_brr(ble, _d);
1820                 case_brd(ble);
1821                 case_brr(beq, _d);
1822                 case_brd(beq);
1823                 case_brr(bge, _d);
1824                 case_brd(bge);
1825                 case_brr(bgt, _d);
1826                 case_brd(bgt);
1827                 case_brr(bne, _d);
1828                 case_brd(bne);
1829                 case_brr(bunlt, _d);
1830                 case_brd(bunlt);
1831                 case_brr(bunle, _d);
1832                 case_brd(bunle);
1833                 case_brr(buneq, _d);
1834                 case_brd(buneq);
1835                 case_brr(bunge, _d);
1836                 case_brd(bunge);
1837                 case_brr(bungt, _d);
1838                 case_brd(bungt);
1839                 case_brr(bltgt, _d);
1840                 case_brd(bltgt);
1841                 case_brr(bord, _d);
1842                 case_brd(bord);
1843                 case_brr(bunord, _d);
1844                 case_brd(bunord);
1845             case jit_code_jmpr:
1846                 jit_check_frame();
1847                 jmpr(rn(node->u.w));
1848                 break;
1849             case jit_code_jmpi:
1850                 if (node->flag & jit_flag_node) {
1851                     temp = node->u.n;
1852                     assert(temp->code == jit_code_label ||
1853                            temp->code == jit_code_epilog);
1854                     if (temp->flag & jit_flag_patch)
1855                         jmpi(temp->u.w);
1856                     else {
1857                         word = _jit->code.length -
1858                             (_jit->pc.uc - _jit->code.ptr);
1859                         if (s26_p(word))
1860                             word = jmpi(_jit->pc.w);
1861                         else
1862                             word = jmpi_p(_jit->pc.w);
1863                         patch(word, node);
1864                     }
1865                 }
1866                 else {
1867                     jit_check_frame();
1868                     jmpi(node->u.w);
1869                 }
1870                 break;
1871             case jit_code_callr:
1872                 jit_check_frame();
1873                 callr(rn(node->u.w));
1874                 break;
1875             case jit_code_calli:
1876                 jit_check_frame();
1877                 if (node->flag & jit_flag_node) {
1878                     temp = node->u.n;
1879                     assert(temp->code == jit_code_label ||
1880                            temp->code == jit_code_epilog);
1881                     if (temp->flag & jit_flag_patch)
1882                         calli(temp->u.w);
1883                     else {
1884                         word = _jit->code.length -
1885                             (_jit->pc.uc - _jit->code.ptr);
1886                         if (s26_p(word))
1887                             word = calli(_jit->pc.w);
1888                         else
1889                             word = calli_p(_jit->pc.w);
1890                         patch(word, node);
1891                     }
1892                 }
1893                 else
1894                     calli(node->u.w);
1895                 break;
1896             case jit_code_prolog:
1897                 _jitc->function = _jitc->functions.ptr + node->w.w;
1898                 undo.node = node;
1899                 undo.word = _jit->pc.w;
1900                 memcpy(&undo.func, _jitc->function, sizeof(undo.func));
1901 #if DEVEL_DISASSEMBLER
1902                 undo.prevw = prevw;
1903 #endif
1904                 undo.patch_offset = _jitc->patches.offset;
1905             restart_function:
1906                 compute_framesize();
1907                 patch_alist(0);
1908                 _jitc->again = 0;
1909                 prolog(node);
1910                 break;
1911             case jit_code_epilog:
1912                 assert(_jitc->function == _jitc->functions.ptr + node->w.w);
1913                 if (_jitc->again) {
1914                     for (temp = undo.node->next;
1915                          temp != node; temp = temp->next) {
1916                         if (temp->code == jit_code_label ||
1917                             temp->code == jit_code_epilog)
1918                             temp->flag &= ~jit_flag_patch;
1919                     }
1920                     temp->flag &= ~jit_flag_patch;
1921                     node = undo.node;
1922                     _jit->pc.w = undo.word;
1923                     /* undo.func.self.aoff and undo.func.regset should not
1924                      * be undone, as they will be further updated, and are
1925                      * the reason of the undo. */
1926                     undo.func.self.aoff = _jitc->function->frame +
1927                         _jitc->function->self.aoff;
1928                     undo.func.need_frame = _jitc->function->need_frame;
1929                     jit_regset_set(&undo.func.regset, &_jitc->function->regset);
1930                     /* allocar information also does not need to be undone */
1931                     undo.func.aoffoff = _jitc->function->aoffoff;
1932                     undo.func.allocar = _jitc->function->allocar;
1933                     memcpy(_jitc->function, &undo.func, sizeof(undo.func));
1934 #if DEVEL_DISASSEMBLER
1935                     prevw = undo.prevw;
1936 #endif
1937                     _jitc->patches.offset = undo.patch_offset;
1938                     patch_alist(1);
1939                     goto restart_function;
1940                 }
1941                 /* remember label is defined */
1942                 node->flag |= jit_flag_patch;
1943                 node->u.w = _jit->pc.w;
1944                 epilog(node);
1945                 _jitc->function = NULL;
1946                 break;
1947             case jit_code_movr_w_f:
1948                 movr_w_f(rn(node->u.w), rn(node->v.w));
1949                 break;
1950             case jit_code_movr_f_w:
1951                 movr_f_w(rn(node->u.w), rn(node->v.w));
1952                 break;
1953             case jit_code_movi_f_w:
1954                 movi_f_w(rn(node->u.w), node->v.f);
1955                 break;
1956             case jit_code_movi_w_f:
1957                 movi_w_f(rn(node->u.w), node->v.w);
1958                 break;
1959             case jit_code_movr_w_d:
1960                 movr_w_d(rn(node->u.w), rn(node->v.w));
1961                 break;
1962             case jit_code_movr_d_w:
1963                 movr_d_w(rn(node->u.w), rn(node->v.w));
1964                 break;
1965             case jit_code_movi_d_w:
1966                 movi_d_w(rn(node->u.w), node->v.d);
1967                 break;
1968             case jit_code_movi_w_d:
1969                 movi_w_d(rn(node->u.w), node->v.w);
1970                 break;
1971             case jit_code_va_start:
1972                 vastart(rn(node->u.w));
1973                 break;
1974             case jit_code_va_arg:
1975                 vaarg(rn(node->u.w), rn(node->v.w));
1976                 break;
1977             case jit_code_va_arg_d:
1978                 vaarg_d(rn(node->u.w), rn(node->v.w));
1979                 break;
1980             case jit_code_live:                 case jit_code_ellipsis:
1981             case jit_code_va_push:
1982             case jit_code_allocai:              case jit_code_allocar:
1983             case jit_code_arg_c:                case jit_code_arg_s:
1984             case jit_code_arg_i:
1985 #  if __WORDSIZE == 64
1986             case jit_code_arg_l:
1987 #  endif
1988             case jit_code_arg_f:                case jit_code_arg_d:
1989             case jit_code_va_end:
1990             case jit_code_ret:
1991             case jit_code_retr_c:               case jit_code_reti_c:
1992             case jit_code_retr_uc:              case jit_code_reti_uc:
1993             case jit_code_retr_s:               case jit_code_reti_s:
1994             case jit_code_retr_us:              case jit_code_reti_us:
1995             case jit_code_retr_i:               case jit_code_reti_i:
1996 #if __WORDSIZE == 64
1997             case jit_code_retr_ui:              case jit_code_reti_ui:
1998             case jit_code_retr_l:               case jit_code_reti_l:
1999 #endif
2000             case jit_code_retr_f:               case jit_code_reti_f:
2001             case jit_code_retr_d:               case jit_code_reti_d:
2002             case jit_code_getarg_c:             case jit_code_getarg_uc:
2003             case jit_code_getarg_s:             case jit_code_getarg_us:
2004             case jit_code_getarg_i:             case jit_code_getarg_ui:
2005             case jit_code_getarg_l:
2006             case jit_code_getarg_f:             case jit_code_getarg_d:
2007             case jit_code_putargr_c:            case jit_code_putargi_c:
2008             case jit_code_putargr_uc:           case jit_code_putargi_uc:
2009             case jit_code_putargr_s:            case jit_code_putargi_s:
2010             case jit_code_putargr_us:           case jit_code_putargi_us:
2011             case jit_code_putargr_i:            case jit_code_putargi_i:
2012 #if __WORDSIZE == 64
2013             case jit_code_putargr_ui:           case jit_code_putargi_ui:
2014             case jit_code_putargr_l:            case jit_code_putargi_l:
2015 #endif
2016             case jit_code_putargr_f:            case jit_code_putargi_f:
2017             case jit_code_putargr_d:            case jit_code_putargi_d:
2018             case jit_code_pushargr_c:           case jit_code_pushargi_c:
2019             case jit_code_pushargr_uc:          case jit_code_pushargi_uc:
2020             case jit_code_pushargr_s:           case jit_code_pushargi_s:
2021             case jit_code_pushargr_us:          case jit_code_pushargi_us:
2022             case jit_code_pushargr_i:           case jit_code_pushargi_i:
2023 #if __WORDSIZE == 64
2024             case jit_code_pushargr_ui:          case jit_code_pushargi_ui:
2025             case jit_code_pushargr_l:           case jit_code_pushargi_l:
2026 #endif
2027             case jit_code_pushargr_f:           case jit_code_pushargi_f:
2028             case jit_code_pushargr_d:           case jit_code_pushargi_d:
2029             case jit_code_retval_c:             case jit_code_retval_uc:
2030             case jit_code_retval_s:             case jit_code_retval_us:
2031             case jit_code_retval_i:
2032 #if __WORDSIZE == 64
2033             case jit_code_retval_ui:            case jit_code_retval_l:
2034 #endif
2035             case jit_code_retval_f:             case jit_code_retval_d:
2036             case jit_code_prepare:
2037             case jit_code_finishr:              case jit_code_finishi:
2038             case jit_code_negi_f:               case jit_code_absi_f:
2039             case jit_code_sqrti_f:              case jit_code_negi_d:
2040             case jit_code_absi_d:               case jit_code_sqrti_d:
2041                 break;
2042             case jit_code_negi:
2043                 negi(rn(node->u.w), node->v.w);
2044                 break;
2045             case jit_code_comi:
2046                 comi(rn(node->u.w), node->v.w);
2047                 break;
2048             case jit_code_exti_c:
2049                 exti_c(rn(node->u.w), node->v.w);
2050                 break;
2051             case jit_code_exti_uc:
2052                 exti_uc(rn(node->u.w), node->v.w);
2053                 break;
2054             case jit_code_exti_s:
2055                 exti_s(rn(node->u.w), node->v.w);
2056                 break;
2057             case jit_code_exti_us:
2058                 exti_us(rn(node->u.w), node->v.w);
2059                 break;
2060             case jit_code_bswapi_us:
2061                 bswapi_us(rn(node->u.w), node->v.w);
2062                 break;
2063             case jit_code_bswapi_ui:
2064                 bswapi_ui(rn(node->u.w), node->v.w);
2065                 break;
2066             case jit_code_htoni_us:
2067                 htoni_us(rn(node->u.w), node->v.w);
2068                 break;
2069             case jit_code_htoni_ui:
2070                 htoni_ui(rn(node->u.w), node->v.w);
2071                 break;
2072 #if __WORDSIZE == 64
2073             case jit_code_exti_i:
2074                 exti_i(rn(node->u.w), node->v.w);
2075                 break;
2076             case jit_code_exti_ui:
2077                 exti_ui(rn(node->u.w), node->v.w);
2078                 break;
2079             case jit_code_bswapi_ul:
2080                 bswapi_ul(rn(node->u.w), node->v.w);
2081                 break;
2082             case jit_code_htoni_ul:
2083                 htoni_ul(rn(node->u.w), node->v.w);
2084                 break;
2085 #endif
2086             case jit_code_cloi:
2087                 cloi(rn(node->u.w), node->v.w);
2088                 break;
2089             case jit_code_clzi:
2090                 clzi(rn(node->u.w), node->v.w);
2091                 break;
2092             case jit_code_ctoi:
2093                 ctoi(rn(node->u.w), node->v.w);
2094                 break;
2095             case jit_code_ctzi:
2096                 ctzi(rn(node->u.w), node->v.w);
2097                 break;
2098             case jit_code_rbiti:
2099                 rbiti(rn(node->u.w), node->v.w);
2100                 break;
2101             case jit_code_popcnti:
2102                 popcnti(rn(node->u.w), node->v.w);
2103                 break;
2104             case jit_code_exti:
2105                 exti(rn(node->u.w), node->v.w, node->w.q.l, node->w.q.h);
2106                 break;
2107             case jit_code_exti_u:
2108                 exti_u(rn(node->u.w), node->v.w, node->w.q.l, node->w.q.h);
2109                 break;
2110             default:
2111                 abort();
2112         }
2113         jit_regarg_clr(node, value);
2114         assert(_jitc->regarg == 0 && _jitc->synth == 0);
2115         /* update register live state */
2116         jit_reglive(node);
2117     }
2118 #undef case_brw
2119 #undef case_brr
2120 #undef case_wrr
2121 #undef case_rrw
2122 #undef case_rrr
2123 #undef case_wr
2124 #undef case_rw
2125 #undef case_rr
2126
2127     for (offset = 0; offset < _jitc->patches.offset; offset++) {
2128         node = _jitc->patches.ptr[offset].node;
2129         word = _jitc->patches.ptr[offset].inst;
2130         value = node->code == jit_code_movi ? node->v.n->u.w : node->u.n->u.w;
2131         patch_at(word, value);
2132     }
2133
2134     jit_flush(_jit->code.ptr, _jit->pc.uc);
2135
2136     return (_jit->code.ptr);
2137 }
2138
2139 #define CODE                            1
2140 #  include "jit_aarch64-cpu.c"
2141 #  include "jit_aarch64-fpu.c"
2142 #undef CODE
2143
2144 void
2145 jit_flush(void *fptr, void *tptr)
2146 {
2147 #if defined(__GNUC__)
2148     jit_word_t          f, t, s;
2149
2150     s = sysconf(_SC_PAGE_SIZE);
2151     f = (jit_word_t)fptr & -s;
2152     t = (((jit_word_t)tptr) + s - 1) & -s;
2153     __clear_cache((void *)f, (void *)t);
2154 #endif
2155 }
2156
2157 void
2158 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
2159 {
2160     ldxi(rn(r0), rn(r1), i0);
2161 }
2162
2163 void
2164 _emit_stxi(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
2165 {
2166     stxi(i0, rn(r0), rn(r1));
2167 }
2168
2169 void
2170 _emit_ldxi_d(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
2171 {
2172     ldxi_d(rn(r0), rn(r1), i0);
2173 }
2174
2175 void
2176 _emit_stxi_d(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
2177 {
2178     stxi_d(i0, rn(r0), rn(r1));
2179 }
2180
2181 static void
2182 _compute_framesize(jit_state_t *_jit)
2183 {
2184     jit_int32_t         reg;
2185     _jitc->framesize = 16;      /* ra+fp */
2186     for (reg = 0; reg < jit_size(iregs); reg++)
2187         if (jit_regset_tstbit(&_jitc->function->regset, iregs[reg]))
2188             _jitc->framesize += sizeof(jit_word_t);
2189
2190     for (reg = 0; reg < jit_size(fregs); reg++)
2191         if (jit_regset_tstbit(&_jitc->function->regset, fregs[reg]))
2192             _jitc->framesize += sizeof(jit_float64_t);
2193
2194     /* Make sure functions called have a 16 byte aligned stack */
2195     _jitc->framesize = (_jitc->framesize + 15) & -16;
2196 }
2197
2198 static void
2199 _patch(jit_state_t *_jit, jit_word_t instr, jit_node_t *node)
2200 {
2201     jit_int32_t          flag;
2202
2203     assert(node->flag & jit_flag_node);
2204     if (node->code == jit_code_movi)
2205         flag = node->v.n->flag;
2206     else
2207         flag = node->u.n->flag;
2208     assert(!(flag & jit_flag_patch));
2209     if (_jitc->patches.offset >= _jitc->patches.length) {
2210         jit_realloc((jit_pointer_t *)&_jitc->patches.ptr,
2211                     _jitc->patches.length * sizeof(jit_patch_t),
2212                     (_jitc->patches.length + 1024) * sizeof(jit_patch_t));
2213         _jitc->patches.length += 1024;
2214     }
2215     _jitc->patches.ptr[_jitc->patches.offset].inst = instr;
2216     _jitc->patches.ptr[_jitc->patches.offset].node = node;
2217     ++_jitc->patches.offset;
2218 }