deps/lightning/lib/jit_s390.c

   1 /*
   2  * Copyright (C) 2013-2022  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 #include <lightning.h>
  21 #include <lightning/jit_private.h>
  22
  23 #if __WORDSIZE == 32
  24 #  define NUM_FLOAT_REG_ARGS            2
  25 #else
  26 #  define NUM_FLOAT_REG_ARGS            4
  27 #endif
  28 #define jit_arg_reg_p(i)                ((i) >= 0 && (i) < 5)
  29 #define jit_arg_f_reg_p(i)              ((i) >= 0 && (i) < NUM_FLOAT_REG_ARGS)
  30
  31 /*
  32  * Types
  33  */
  34 typedef struct jit_va_list {
  35     /* The offsets are "1" based, as addresses are fixed in the
  36      * standard stack frame format. */
  37     jit_word_t          gpoff;
  38     jit_word_t          fpoff;
  39
  40     /* Easier when there is an explicitly defined type...
  41 (gdb) ptype ap
  42 type = struct __va_list_tag {
  43     long __gpr;
  44     long __fpr;
  45     void *__overflow_arg_area;
  46     void *__reg_save_area;
  47
  48     Note that gopff (__gpr) and fpoff (__fpr) are jit_word_t equivalent
  49     and, again, "1" (unit) based, so must be adjusted at va_arg time.
  50  */
  51     jit_pointer_t       over;
  52     jit_pointer_t       save;
  53
  54     /* For variadic functions, always allocate space to save callee
  55      * save fpr registers.
  56      * Note that s390 has a standard stack frame format that lightning
  57      * does not fully comply with, but for variadic functions it must,
  58      * for those (variadic) do not use the "empty" spaces for any
  59      * callee save fpr register, but save them after the va_list
  60      * space; and use the standard stack frame format, as required
  61      * by variadic functions (and have a compatible va_list pointer). */
  62     jit_float64_t       f8;
  63     jit_float64_t       f9;
  64     jit_float64_t       f10;
  65     jit_float64_t       f11;
  66     jit_float64_t       f12;
  67     jit_float64_t       f13;
  68     jit_float64_t       f14;
  69     jit_float64_t       f15;
  70 } jit_va_list_t;
  71
  72 /*
  73  * Prototypes
  74  */
  75 #define jit_get_reg_pair()              _jit_get_reg_pair(_jit)
  76 static jit_int32_t _jit_get_reg_pair(jit_state_t*);
  77 #define jit_unget_reg_pair(regno)       _jit_unget_reg_pair(_jit,regno)
  78 static void _jit_unget_reg_pair(jit_state_t*,jit_int32_t);
  79 #define jit_get_reg_but_zero(flags)     _jit_get_reg_but_zero(_jit,flags)
  80 static jit_int32_t _jit_get_reg_but_zero(jit_state_t*,jit_int32_t);
  81 #define jit_unget_reg_but_zero(reg)     jit_unget_reg(reg)
  82 #define patch(instr, node)              _patch(_jit, instr, node)
  83 static void _patch(jit_state_t*,jit_word_t,jit_node_t*);
  84
  85 /* libgcc */
  86 extern void __clear_cache(void *, void *);
  87
  88 #define PROTO                           1
  89 #  include "jit_s390-cpu.c"
  90 #  include "jit_s390-fpu.c"
  91 #undef PROTO
  92
  93 /*
  94  * Initialization
  95  */
  96 jit_register_t          _rvs[] = {
  97     { rc(gpr) | 0x0,                    "%r0" },
  98     { rc(gpr) | 0x1,                    "%r1" },
  99     { rc(gpr) | rc(sav) | 0xc,          "%r12" },
 100     { rc(gpr) | rc(sav) | 0xb,          "%r11" },
 101     { rc(gpr) | rc(sav) | 0xa,          "%r10" },
 102     { rc(gpr) | rc(sav) | 0x9,          "%r9" },
 103     { rc(gpr) | rc(sav) | 0x8,          "%r8" },
 104     { rc(gpr) | rc(sav) | 0x7,          "%r7" },
 105     { rc(gpr) | rc(arg) | rc(sav) | 0x6,"%r6" },
 106     { rc(gpr) | rc(arg) | 0x5,          "%r5" },
 107     { rc(gpr) | rc(arg) | 0x4,          "%r4" },
 108     { rc(gpr) | rc(arg) | 0x3,          "%r3" },
 109     { rc(gpr) | rc(arg) | 0x2,          "%r2" },
 110     { rc(sav) | 0xd,                    "%r13" },       /* used as JIT_FP */
 111     { 0xe,                              "%r14" },
 112     { rc(sav) | 0xf,                    "%r15" },
 113     { rc(fpr) | 0x1,                    "%f1" },
 114     { rc(fpr) | 0x3,                    "%f3" },
 115     { rc(fpr) | 0x5,                    "%f5" },
 116     { rc(fpr) | 0x7,                    "%f7" },
 117     { rc(fpr) | rc(sav) | 0xe,          "%f14" },
 118     /* Do not use as temporary to simplify stack layout */
 119     { 0xf,                              "%f15" },
 120     { rc(fpr) | rc(sav) | 0x8,          "%f8" },
 121     { rc(fpr) | rc(sav) | 0x9,          "%f9" },
 122     { rc(fpr) | rc(sav) | 0xa,          "%f10" },
 123     { rc(fpr) | rc(sav) | 0xb,          "%f11" },
 124     { rc(fpr) | rc(sav) | 0xc,          "%f12" },
 125     { rc(fpr) | rc(sav) | 0xd,          "%f13" },
 126     { rc(fpr) | rc(arg) | 0x6,          "%f6" },
 127     { rc(fpr) | rc(arg) | 0x4,          "%f4" },
 128     { rc(fpr) | rc(arg) | 0x2,          "%f2" },
 129     { rc(fpr) | rc(arg) | 0x0,          "%f0" },
 130     { _NOREG,                           "<none>" },
 131 };
 132
 133 /*
 134  * Implementation
 135  */
 136 void
 137 jit_get_cpu(void)
 138 {
 139 }
 140
 141 void
 142 _jit_init(jit_state_t *_jit)
 143 {
 144     _jitc->reglen = jit_size(_rvs) - 1;
 145 }
 146
 147 void
 148 _jit_prolog(jit_state_t *_jit)
 149 {
 150     jit_int32_t         offset;
 151
 152     if (_jitc->function)
 153         jit_epilog();
 154     assert(jit_regset_cmp_ui(&_jitc->regarg, 0) == 0);
 155     jit_regset_set_ui(&_jitc->regsav, 0);
 156     offset = _jitc->functions.offset;
 157     if (offset >= _jitc->functions.length) {
 158         jit_realloc((jit_pointer_t *)&_jitc->functions.ptr,
 159                     _jitc->functions.length * sizeof(jit_function_t),
 160                     (_jitc->functions.length + 16) * sizeof(jit_function_t));
 161         _jitc->functions.length += 16;
 162     }
 163     _jitc->function = _jitc->functions.ptr + _jitc->functions.offset++;
 164     _jitc->function->self.size = stack_framesize;
 165     _jitc->function->self.argi = _jitc->function->self.argf =
 166         _jitc->function->self.aoff = _jitc->function->self.alen = 0;
 167     /* preallocate 8 bytes if not using a constant data buffer */
 168     if (_jitc->no_data)
 169         _jitc->function->self.aoff = -8;
 170     _jitc->function->self.call = jit_call_default;
 171     jit_alloc((jit_pointer_t *)&_jitc->function->regoff,
 172               _jitc->reglen * sizeof(jit_int32_t));
 173
 174     /* _no_link here does not mean the jit_link() call can be removed
 175      * by rewriting as:
 176      * _jitc->function->prolog = jit_new_node(jit_code_prolog);
 177      */
 178     _jitc->function->prolog = jit_new_node_no_link(jit_code_prolog);
 179     jit_link(_jitc->function->prolog);
 180     _jitc->function->prolog->w.w = offset;
 181     _jitc->function->epilog = jit_new_node_no_link(jit_code_epilog);
 182     /*  u:      label value
 183      *  v:      offset in blocks vector
 184      *  w:      offset in functions vector
 185      */
 186     _jitc->function->epilog->w.w = offset;
 187
 188     jit_regset_new(&_jitc->function->regset);
 189 }
 190
 191 jit_int32_t
 192 _jit_allocai(jit_state_t *_jit, jit_int32_t length)
 193 {
 194     assert(_jitc->function);
 195     switch (length) {
 196         case 0: case 1:                                         break;
 197         case 2:         _jitc->function->self.aoff &= -2;       break;
 198         case 3: case 4: _jitc->function->self.aoff &= -4;       break;
 199         default:        _jitc->function->self.aoff &= -8;       break;
 200     }
 201     _jitc->function->self.aoff -= length;
 202     if (!_jitc->realize) {
 203         jit_inc_synth_ww(allocai, _jitc->function->self.aoff, length);
 204         jit_dec_synth();
 205     }
 206     return (_jitc->function->self.aoff);
 207 }
 208
 209 void
 210 _jit_allocar(jit_state_t *_jit, jit_int32_t u, jit_int32_t v)
 211 {
 212     jit_int32_t          reg;
 213     assert(_jitc->function);
 214     jit_inc_synth_ww(allocar, u, v);
 215     if (!_jitc->function->allocar) {
 216         _jitc->function->aoffoff = jit_allocai(sizeof(jit_int32_t));
 217         _jitc->function->allocar = 1;
 218     }
 219     reg = jit_get_reg(jit_class_gpr);
 220     jit_negr(reg, v);
 221     jit_andi(reg, reg, -8);
 222     jit_ldxi_i(u, JIT_FP, _jitc->function->aoffoff);
 223     jit_addr(u, u, reg);
 224     jit_addr(JIT_SP, JIT_SP, reg);
 225     jit_stxi_i(_jitc->function->aoffoff, JIT_FP, u);
 226     jit_unget_reg(reg);
 227     jit_dec_synth();
 228 }
 229
 230 void
 231 _jit_ret(jit_state_t *_jit)
 232 {
 233     jit_node_t          *instr;
 234     assert(_jitc->function);
 235     jit_inc_synth(ret);
 236     /* jump to epilog */
 237     instr = jit_jmpi();
 238     jit_patch_at(instr, _jitc->function->epilog);
 239     jit_dec_synth();
 240 }
 241
 242 void
 243 _jit_retr(jit_state_t *_jit, jit_int32_t u)
 244 {
 245     jit_inc_synth_w(retr, u);
 246     jit_movr(JIT_RET, u);
 247     jit_ret();
 248     jit_dec_synth();
 249 }
 250
 251 void
 252 _jit_reti(jit_state_t *_jit, jit_word_t u)
 253 {
 254     jit_inc_synth_w(reti, u);
 255     jit_movi(JIT_RET, u);
 256     jit_ret();
 257     jit_dec_synth();
 258 }
 259
 260 void
 261 _jit_retr_f(jit_state_t *_jit, jit_int32_t u)
 262 {
 263     jit_inc_synth_w(retr_f, u);
 264     jit_movr_f(JIT_FRET, u);
 265     jit_ret();
 266     jit_dec_synth();
 267 }
 268
 269 void
 270 _jit_reti_f(jit_state_t *_jit, jit_float32_t u)
 271 {
 272     jit_inc_synth_f(reti_f, u);
 273     jit_movi_f(JIT_FRET, u);
 274     jit_ret();
 275     jit_dec_synth();
 276 }
 277
 278 void
 279 _jit_retr_d(jit_state_t *_jit, jit_int32_t u)
 280 {
 281     jit_inc_synth_w(retr_d, u);
 282     jit_movr_d(JIT_FRET, u);
 283     jit_ret();
 284     jit_dec_synth();
 285 }
 286
 287 void
 288 _jit_reti_d(jit_state_t *_jit, jit_float64_t u)
 289 {
 290     jit_inc_synth_d(reti_d, u);
 291     jit_movi_d(JIT_FRET, u);
 292     jit_ret();
 293     jit_dec_synth();
 294 }
 295
 296 void
 297 _jit_epilog(jit_state_t *_jit)
 298 {
 299     assert(_jitc->function);
 300     assert(_jitc->function->epilog->next == NULL);
 301     jit_link(_jitc->function->epilog);
 302     _jitc->function = NULL;
 303 }
 304
 305 jit_bool_t
 306 _jit_arg_register_p(jit_state_t *_jit, jit_node_t *u)
 307 {
 308     if (u->code == jit_code_arg)
 309         return (jit_arg_reg_p(u->u.w));
 310     assert(u->code == jit_code_arg_f || u->code == jit_code_arg_d);
 311     return (jit_arg_f_reg_p(u->u.w));
 312 }
 313
 314 void
 315 _jit_ellipsis(jit_state_t *_jit)
 316 {
 317     jit_inc_synth(ellipsis);
 318     if (_jitc->prepare) {
 319         jit_link_prepare();
 320         assert(!(_jitc->function->call.call & jit_call_varargs));
 321         _jitc->function->call.call |= jit_call_varargs;
 322     }
 323     else {
 324         jit_link_prolog();
 325         assert(!(_jitc->function->self.call & jit_call_varargs));
 326         _jitc->function->self.call |= jit_call_varargs;
 327
 328         /* Allocate va_list like object in the stack. */
 329         _jitc->function->vaoff = jit_allocai(sizeof(jit_va_list_t));
 330
 331         /* Initialize gp offset in save area. */
 332         if (jit_arg_reg_p(_jitc->function->self.argi))
 333             _jitc->function->vagp = _jitc->function->self.argi;
 334         else
 335             _jitc->function->vagp = 5;
 336
 337         /* Initialize fp offset in save area. */
 338         if (jit_arg_f_reg_p(_jitc->function->self.argf))
 339             _jitc->function->vafp = _jitc->function->self.argf;
 340         else
 341             _jitc->function->vafp = NUM_FLOAT_REG_ARGS;
 342     }
 343     jit_dec_synth();
 344 }
 345
 346 void
 347 _jit_va_push(jit_state_t *_jit, jit_int32_t u)
 348 {
 349     jit_inc_synth_w(va_push, u);
 350     jit_pushargr(u);
 351     jit_dec_synth();
 352 }
 353
 354 jit_node_t *
 355 _jit_arg(jit_state_t *_jit)
 356 {
 357     jit_node_t          *node;
 358     jit_int32_t          offset;
 359     assert(_jitc->function);
 360     if (jit_arg_reg_p(_jitc->function->self.argi))
 361         offset = _jitc->function->self.argi++;
 362     else {
 363         offset = _jitc->function->self.size;
 364         _jitc->function->self.size += sizeof(jit_word_t);
 365     }
 366     node = jit_new_node_ww(jit_code_arg, offset,
 367                            ++_jitc->function->self.argn);
 368     jit_link_prolog();
 369     return (node);
 370 }
 371
 372 jit_node_t *
 373 _jit_arg_f(jit_state_t *_jit)
 374 {
 375     jit_node_t          *node;
 376     jit_int32_t          offset;
 377     assert(_jitc->function);
 378     if (jit_arg_f_reg_p(_jitc->function->self.argf))
 379         offset = _jitc->function->self.argf++;
 380     else {
 381         offset = _jitc->function->self.size;
 382         _jitc->function->self.size += sizeof(jit_word_t);
 383     }
 384     node = jit_new_node_ww(jit_code_arg_f, offset,
 385                            ++_jitc->function->self.argn);
 386     jit_link_prolog();
 387     return (node);
 388 }
 389
 390 jit_node_t *
 391 _jit_arg_d(jit_state_t *_jit)
 392 {
 393     jit_node_t          *node;
 394     jit_int32_t          offset;
 395     assert(_jitc->function);
 396     if (jit_arg_f_reg_p(_jitc->function->self.argf))
 397         offset = _jitc->function->self.argf++;
 398     else {
 399         offset = _jitc->function->self.size;
 400         _jitc->function->self.size += sizeof(jit_float64_t);
 401     }
 402     node = jit_new_node_ww(jit_code_arg_d, offset,
 403                            ++_jitc->function->self.argn);
 404     jit_link_prolog();
 405     return (node);
 406 }
 407
 408 void
 409 _jit_getarg_c(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 410 {
 411     assert(v->code == jit_code_arg);
 412     jit_inc_synth_wp(getarg_c, u, v);
 413     if (jit_arg_reg_p(v->u.w))
 414         jit_extr_c(u, _R2 - v->u.w);
 415     else
 416         jit_ldxi_c(u, JIT_FP,
 417                    v->u.w + (__WORDSIZE >> 3) - sizeof(jit_int8_t));
 418     jit_dec_synth();
 419 }
 420
 421 void
 422 _jit_getarg_uc(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 423 {
 424     assert(v->code == jit_code_arg);
 425     jit_inc_synth_wp(getarg_uc, u, v);
 426     if (jit_arg_reg_p(v->u.w))
 427         jit_extr_uc(u, _R2 - v->u.w);
 428     else
 429         jit_ldxi_uc(u, JIT_FP,
 430                     v->u.w + (__WORDSIZE >> 3) - sizeof(jit_uint8_t));
 431     jit_dec_synth();
 432 }
 433
 434 void
 435 _jit_getarg_s(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 436 {
 437     assert(v->code == jit_code_arg);
 438     jit_inc_synth_wp(getarg_s, u, v);
 439     if (jit_arg_reg_p(v->u.w))
 440         jit_extr_s(u, _R2 - v->u.w);
 441     else
 442         jit_ldxi_s(u, JIT_FP,
 443                    v->u.w + (__WORDSIZE >> 3) - sizeof(jit_int16_t));
 444     jit_dec_synth();
 445 }
 446
 447 void
 448 _jit_getarg_us(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 449 {
 450     assert(v->code == jit_code_arg);
 451     jit_inc_synth_wp(getarg_us, u, v);
 452     if (jit_arg_reg_p(v->u.w))
 453         jit_extr_us(u, _R2 - v->u.w);
 454     else
 455         jit_ldxi_us(u, JIT_FP,
 456                     v->u.w + (__WORDSIZE >> 3) - sizeof(jit_uint16_t));
 457     jit_dec_synth();
 458 }
 459
 460 void
 461 _jit_getarg_i(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 462 {
 463     assert(v->code == jit_code_arg);
 464     jit_inc_synth_wp(getarg_i, u, v);
 465     if (jit_arg_reg_p(v->u.w)) {
 466 #if __WORDSIZE == 32
 467         jit_movr(u, _R2 - v->u.w);
 468 #else
 469         jit_extr_i(u, _R2 - v->u.w);
 470 #endif
 471     }
 472     else
 473         jit_ldxi_i(u, JIT_FP,
 474                    v->u.w + (__WORDSIZE >> 3) - sizeof(jit_int32_t));
 475     jit_dec_synth();
 476 }
 477
 478 #if __WORDSIZE == 64
 479 void
 480 _jit_getarg_ui(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 481 {
 482     assert(v->code == jit_code_arg);
 483     jit_inc_synth_wp(getarg_ui, u, v);
 484     if (jit_arg_reg_p(v->u.w))
 485         jit_extr_ui(u, _R2 - v->u.w);
 486     else
 487         jit_ldxi_ui(u, JIT_FP,
 488                     v->u.w + (__WORDSIZE >> 3) - sizeof(jit_uint32_t));
 489     jit_dec_synth();
 490 }
 491
 492 void
 493 _jit_getarg_l(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 494 {
 495     assert(v->code == jit_code_arg);
 496     jit_inc_synth_wp(getarg_l, u, v);
 497     if (jit_arg_reg_p(v->u.w))
 498         jit_movr(u, _R2 - v->u.w);
 499     else
 500         jit_ldxi_l(u, JIT_FP, v->u.w);
 501     jit_dec_synth();
 502 }
 503 #endif
 504
 505 void
 506 _jit_putargr(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 507 {
 508     assert(v->code == jit_code_arg);
 509     jit_inc_synth_wp(putargr, u, v);
 510     if (jit_arg_reg_p(v->u.w))
 511         jit_movr(_R2 - v->u.w, u);
 512     else
 513         jit_stxi(v->u.w, JIT_FP, u);
 514     jit_dec_synth();
 515 }
 516
 517 void
 518 _jit_putargi(jit_state_t *_jit, jit_word_t u, jit_node_t *v)
 519 {
 520     jit_int32_t         regno;
 521     assert(v->code == jit_code_arg);
 522     jit_inc_synth_wp(putargi, u, v);
 523     if (jit_arg_reg_p(v->u.w))
 524         jit_movi(_R2 - v->u.w, u);
 525     else {
 526         regno = jit_get_reg(jit_class_gpr);
 527         jit_movi(regno, u);
 528         jit_stxi(v->u.w, JIT_FP, regno);
 529         jit_unget_reg(regno);
 530     }
 531     jit_dec_synth();
 532 }
 533
 534 void
 535 _jit_getarg_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 536 {
 537     assert(v->code == jit_code_arg_f);
 538     jit_inc_synth_wp(getarg_f, u, v);
 539     if (jit_arg_f_reg_p(v->u.w))
 540         jit_movr_f(u, _F0 - v->u.w);
 541     else
 542         jit_ldxi_f(u, JIT_FP,
 543                    v->u.w
 544 #if __WORDSIZE == 64
 545                    + (__WORDSIZE >> 3) - sizeof(jit_float32_t)
 546 #endif
 547                    );
 548     jit_dec_synth();
 549 }
 550
 551 void
 552 _jit_putargr_f(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 553 {
 554     assert(v->code == jit_code_arg_f);
 555     jit_inc_synth_wp(putargr_f, u, v);
 556     if (jit_arg_f_reg_p(v->u.w))
 557         jit_movr_f(_F0 - v->u.w, u);
 558     else
 559         jit_stxi_f(v->u.w
 560 #if __WORDSIZE == 64
 561                    + (__WORDSIZE >> 3) - sizeof(jit_float32_t)
 562 #endif
 563                    , JIT_FP, u);
 564     jit_dec_synth();
 565 }
 566
 567 void
 568 _jit_putargi_f(jit_state_t *_jit, jit_float32_t u, jit_node_t *v)
 569 {
 570     jit_int32_t         regno;
 571     assert(v->code == jit_code_arg_f);
 572     jit_inc_synth_fp(putargi_f, u, v);
 573     if (jit_arg_f_reg_p(v->u.w))
 574         jit_movi_f(_F0 - v->u.w, u);
 575     else {
 576         regno = jit_get_reg(jit_class_fpr);
 577         jit_movi_f(regno, u);
 578         jit_stxi_f(v->u.w
 579 #if __WORDSIZE == 64
 580                    + (__WORDSIZE >> 3) - sizeof(jit_float32_t)
 581 #endif
 582                    , JIT_FP, regno);
 583         jit_unget_reg(regno);
 584     }
 585     jit_dec_synth();
 586 }
 587
 588 void
 589 _jit_getarg_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 590 {
 591     assert(v->code == jit_code_arg_d);
 592     jit_inc_synth_wp(getarg_d, u, v);
 593     if (jit_arg_f_reg_p(v->u.w))
 594         jit_movr_d(u, _F0 - v->u.w);
 595     else
 596         jit_ldxi_d(u, JIT_FP, v->u.w);
 597     jit_dec_synth();
 598 }
 599
 600 void
 601 _jit_putargr_d(jit_state_t *_jit, jit_int32_t u, jit_node_t *v)
 602 {
 603     assert(v->code == jit_code_arg_d);
 604     jit_inc_synth_wp(putargr_d, u, v);
 605     if (jit_arg_f_reg_p(v->u.w))
 606         jit_movr_d(_F0 - v->u.w, u);
 607     else
 608         jit_stxi_d(v->u.w, JIT_FP, u);
 609     jit_dec_synth();
 610 }
 611
 612 void
 613 _jit_putargi_d(jit_state_t *_jit, jit_float64_t u, jit_node_t *v)
 614 {
 615     jit_int32_t         regno;
 616     assert(v->code == jit_code_arg_d);
 617     jit_inc_synth_dp(putargi_d, u, v);
 618     if (jit_arg_f_reg_p(v->u.w))
 619         jit_movi_d(_F0 - v->u.w, u);
 620     else {
 621         regno = jit_get_reg(jit_class_fpr);
 622         jit_movi_d(regno, u);
 623         jit_stxi_d(v->u.w, JIT_FP, regno);
 624         jit_unget_reg(regno);
 625     }
 626     jit_dec_synth();
 627 }
 628
 629 void
 630 _jit_pushargr(jit_state_t *_jit, jit_int32_t u)
 631 {
 632     assert(_jitc->function);
 633     jit_inc_synth_w(pushargr, u);
 634     jit_link_prepare();
 635     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 636         jit_movr(_R2 - _jitc->function->call.argi, u);
 637         ++_jitc->function->call.argi;
 638     }
 639     else {
 640         jit_stxi(_jitc->function->call.size + stack_framesize, JIT_SP, u);
 641         _jitc->function->call.size += sizeof(jit_word_t);
 642     }
 643     jit_dec_synth();
 644 }
 645
 646 void
 647 _jit_pushargi(jit_state_t *_jit, jit_word_t u)
 648 {
 649     jit_int32_t          regno;
 650     assert(_jitc->function);
 651     jit_inc_synth_w(pushargi, u);
 652     jit_link_prepare();
 653     if (jit_arg_reg_p(_jitc->function->call.argi)) {
 654         jit_movi(_R2 - _jitc->function->call.argi, u);
 655         ++_jitc->function->call.argi;
 656     }
 657     else {
 658         regno = jit_get_reg(jit_class_gpr);
 659         jit_movi(regno, u);
 660         jit_stxi(_jitc->function->call.size + stack_framesize, JIT_SP, regno);
 661         jit_unget_reg(regno);
 662         _jitc->function->call.size += sizeof(jit_word_t);
 663     }
 664     jit_dec_synth();
 665 }
 666
 667 void
 668 _jit_pushargr_f(jit_state_t *_jit, jit_int32_t u)
 669 {
 670     assert(_jitc->function);
 671     jit_inc_synth_w(pushargr_f, u);
 672     jit_link_prepare();
 673     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 674         jit_movr_f(_F0 - _jitc->function->call.argf, u);
 675         ++_jitc->function->call.argf;
 676     }
 677     else {
 678         jit_stxi_f(_jitc->function->call.size + stack_framesize
 679 #if __WORDSIZE == 64
 680                    + (__WORDSIZE >> 3) - sizeof(jit_float32_t)
 681 #endif
 682                    , JIT_SP, u);
 683         _jitc->function->call.size += sizeof(jit_word_t);
 684     }
 685     jit_dec_synth();
 686 }
 687
 688 void
 689 _jit_pushargi_f(jit_state_t *_jit, jit_float32_t u)
 690 {
 691     jit_int32_t         regno;
 692     assert(_jitc->function);
 693     jit_inc_synth_f(pushargi_f, u);
 694     jit_link_prepare();
 695     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 696         jit_movi_f(_F0 - _jitc->function->call.argf, u);
 697         ++_jitc->function->call.argf;
 698     }
 699     else {
 700         regno = jit_get_reg(jit_class_fpr);
 701         jit_movi_f(regno, u);
 702         jit_stxi_f(_jitc->function->call.size + stack_framesize
 703 #if __WORDSIZE == 64
 704                    + (__WORDSIZE >> 3) - sizeof(jit_float32_t)
 705 #endif
 706                    , JIT_SP, regno);
 707         jit_unget_reg(regno);
 708         _jitc->function->call.size += sizeof(jit_word_t);
 709     }
 710     jit_dec_synth();
 711 }
 712
 713 void
 714 _jit_pushargr_d(jit_state_t *_jit, jit_int32_t u)
 715 {
 716     assert(_jitc->function);
 717     jit_inc_synth_w(pushargr_d, u);
 718     jit_link_prepare();
 719     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 720         jit_movr_d(_F0 - _jitc->function->call.argf, u);
 721         ++_jitc->function->call.argf;
 722     }
 723     else {
 724         jit_stxi_d(_jitc->function->call.size + stack_framesize, JIT_SP, u);
 725         _jitc->function->call.size += sizeof(jit_float64_t);
 726     }
 727     jit_dec_synth();
 728 }
 729
 730 void
 731 _jit_pushargi_d(jit_state_t *_jit, jit_float64_t u)
 732 {
 733     jit_int32_t         regno;
 734     assert(_jitc->function);
 735     jit_inc_synth_d(pushargi_d, u);
 736     jit_link_prepare();
 737     if (jit_arg_f_reg_p(_jitc->function->call.argf)) {
 738         jit_movi_d(_F0 - _jitc->function->call.argf, u);
 739         ++_jitc->function->call.argf;
 740     }
 741     else {
 742         regno = jit_get_reg(jit_class_fpr);
 743         jit_movi_d(regno, u);
 744         jit_stxi_d(_jitc->function->call.size + stack_framesize, JIT_SP, regno);
 745         jit_unget_reg(regno);
 746         _jitc->function->call.size += sizeof(jit_float64_t);
 747     }
 748     jit_dec_synth();
 749 }
 750
 751 jit_bool_t
 752 _jit_regarg_p(jit_state_t *_jit, jit_node_t *node, jit_int32_t regno)
 753 {
 754     jit_int32_t         spec;
 755     spec = jit_class(_rvs[regno].spec);
 756     if (spec & jit_class_arg) {
 757         regno = _R2 - regno;
 758         if (regno >= 0 && regno < node->v.w)
 759             return (1);
 760         if (spec & jit_class_fpr) {
 761             regno = _F0 - regno;
 762             if (regno >= 0 && regno < node->w.w)
 763                 return (1);
 764         }
 765     }
 766     return (0);
 767 }
 768
 769 void
 770 _jit_finishr(jit_state_t *_jit, jit_int32_t r0)
 771 {
 772     jit_node_t          *call;
 773     assert(_jitc->function);
 774     jit_inc_synth_w(finishr, r0);
 775     if (_jitc->function->self.alen < _jitc->function->call.size)
 776         _jitc->function->self.alen = _jitc->function->call.size;
 777     call = jit_callr(r0);
 778     call->v.w = _jitc->function->call.argi;
 779     call->w.w = _jitc->function->call.argf;
 780     _jitc->function->call.argi = _jitc->function->call.argf =
 781         _jitc->function->call.size = 0;
 782     _jitc->prepare = 0;
 783     jit_dec_synth();
 784 }
 785
 786 jit_node_t *
 787 _jit_finishi(jit_state_t *_jit, jit_pointer_t i0)
 788 {
 789     jit_node_t          *node;
 790     assert(_jitc->function);
 791     jit_inc_synth_w(finishi, (jit_word_t)i0);
 792     if (_jitc->function->self.alen < _jitc->function->call.size)
 793         _jitc->function->self.alen = _jitc->function->call.size;
 794     node = jit_calli(i0);
 795     node->v.w = _jitc->function->call.argi;
 796     node->w.w = _jitc->function->call.argf;
 797     _jitc->function->call.argi = _jitc->function->call.argf =
 798         _jitc->function->call.size = 0;
 799     _jitc->prepare = 0;
 800     jit_dec_synth();
 801     return (node);
 802 }
 803
 804 void
 805 _jit_retval_c(jit_state_t *_jit, jit_int32_t r0)
 806 {
 807     jit_inc_synth_w(retval_c, r0);
 808     jit_extr_c(r0, JIT_RET);
 809     jit_dec_synth();
 810 }
 811
 812 void
 813 _jit_retval_uc(jit_state_t *_jit, jit_int32_t r0)
 814 {
 815     jit_inc_synth_w(retval_uc, r0);
 816     jit_extr_uc(r0, JIT_RET);
 817     jit_dec_synth();
 818 }
 819
 820 void
 821 _jit_retval_s(jit_state_t *_jit, jit_int32_t r0)
 822 {
 823     jit_inc_synth_w(retval_s, r0);
 824     jit_extr_s(r0, JIT_RET);
 825     jit_dec_synth();
 826 }
 827
 828 void
 829 _jit_retval_us(jit_state_t *_jit, jit_int32_t r0)
 830 {
 831     jit_inc_synth_w(retval_us, r0);
 832     jit_extr_us(r0, JIT_RET);
 833     jit_dec_synth();
 834 }
 835
 836 void
 837 _jit_retval_i(jit_state_t *_jit, jit_int32_t r0)
 838 {
 839     jit_inc_synth_w(retval_i, r0);
 840 #if __WORDSIZE == 64
 841     jit_extr_i(r0, JIT_RET);
 842 #else
 843     jit_movr(r0, JIT_RET);
 844 #endif
 845     jit_dec_synth();
 846 }
 847
 848 #if __WORDSIZE == 64
 849 void
 850 _jit_retval_ui(jit_state_t *_jit, jit_int32_t r0)
 851 {
 852     jit_inc_synth_w(retval_ui, r0);
 853     jit_extr_ui(r0, JIT_RET);
 854     jit_dec_synth();
 855 }
 856
 857 void
 858 _jit_retval_l(jit_state_t *_jit, jit_int32_t r0)
 859 {
 860     jit_inc_synth_w(retval_l, r0);
 861     jit_movr(r0, JIT_RET);
 862     jit_dec_synth();
 863 }
 864 #endif
 865
 866 void
 867 _jit_retval_f(jit_state_t *_jit, jit_int32_t r0)
 868 {
 869     jit_inc_synth_w(retval_f, r0);
 870     jit_movr_f(r0, JIT_FRET);
 871     jit_dec_synth();
 872 }
 873
 874 void
 875 _jit_retval_d(jit_state_t *_jit, jit_int32_t r0)
 876 {
 877     jit_inc_synth_w(retval_d, r0);
 878     jit_movr_d(r0, JIT_FRET);
 879     jit_dec_synth();
 880 }
 881
 882 jit_pointer_t
 883 _emit_code(jit_state_t *_jit)
 884 {
 885     jit_node_t          *node;
 886     jit_node_t          *temp;
 887     jit_word_t           word;
 888     jit_int32_t          value;
 889     jit_int32_t          offset;
 890     struct {
 891         jit_node_t      *node;
 892         jit_word_t       word;
 893 #if DEVEL_DISASSEMBLER
 894         jit_word_t       prevw;
 895 #endif
 896         jit_int32_t      patch_offset;
 897     } undo;
 898 #if DEVEL_DISASSEMBLER
 899     jit_word_t           prevw;
 900 #endif
 901
 902     _jitc->function = NULL;
 903
 904     jit_reglive_setup();
 905
 906     undo.word = 0;
 907     undo.node = NULL;
 908     undo.patch_offset = 0;
 909
 910 #define assert_data(node)               /**/
 911 #define case_rr(name, type)                                             \
 912             case jit_code_##name##r##type:                              \
 913                 name##r##type(rn(node->u.w), rn(node->v.w));            \
 914                 break
 915 #define case_rw(name, type)                                             \
 916             case jit_code_##name##i##type:                              \
 917                 name##i##type(rn(node->u.w), node->v.w);                \
 918                 break
 919 #define case_wr(name, type)                                             \
 920             case jit_code_##name##i##type:                              \
 921                 name##i##type(node->u.w, rn(node->v.w));                \
 922                 break
 923 #define case_rrr(name, type)                                            \
 924             case jit_code_##name##r##type:                              \
 925                 name##r##type(rn(node->u.w),                            \
 926                               rn(node->v.w), rn(node->w.w));            \
 927                 break
 928 #define case_rrrr(name, type)                                           \
 929             case jit_code_##name##r##type:                              \
 930                 name##r##type(rn(node->u.q.l), rn(node->u.q.h),         \
 931                               rn(node->v.w), rn(node->w.w));            \
 932                 break
 933 #define case_rrw(name, type)                                            \
 934             case jit_code_##name##i##type:                              \
 935                 name##i##type(rn(node->u.w), rn(node->v.w), node->w.w); \
 936                 break
 937 #define case_rrrw(name, type)                                           \
 938             case jit_code_##name##i##type:                              \
 939                 name##i##type(rn(node->u.q.l), rn(node->u.q.h),         \
 940                               rn(node->v.w), node->w.w);                \
 941                 break
 942 #define case_rrf(name)                                                  \
 943             case jit_code_##name##i_f:                                  \
 944                 assert_data(node);                                      \
 945                 name##i_f(rn(node->u.w), rn(node->v.w),                 \
 946                           (jit_float32_t *)node->w.n->u.w);             \
 947                 break
 948 #define case_rrd(name)                                                  \
 949             case jit_code_##name##i_d:                                  \
 950                 assert_data(node);                                      \
 951                 name##i_d(rn(node->u.w), rn(node->v.w),                 \
 952                           (jit_float64_t *)node->w.n->u.w);             \
 953                 break
 954 #define case_wrr(name, type)                                            \
 955             case jit_code_##name##i##type:                              \
 956                 name##i##type(node->u.w, rn(node->v.w), rn(node->w.w)); \
 957                 break
 958 #define case_brr(name, type)                                            \
 959             case jit_code_##name##r##type:                              \
 960                 temp = node->u.n;                                       \
 961                 assert(temp->code == jit_code_label ||                  \
 962                        temp->code == jit_code_epilog);                  \
 963                 if (temp->flag & jit_flag_patch)                        \
 964                     name##r##type(temp->u.w, rn(node->v.w),             \
 965                                   rn(node->w.w));                       \
 966                 else {                                                  \
 967                     word = name##r##type##_p(_jit->pc.w,                \
 968                                              rn(node->v.w),             \
 969                                              rn(node->w.w));            \
 970                     patch(word, node);                                  \
 971                 }                                                       \
 972                 break
 973 #define case_brw(name, type)                                            \
 974             case jit_code_##name##i##type:                              \
 975                 temp = node->u.n;                                       \
 976                 assert(temp->code == jit_code_label ||                  \
 977                        temp->code == jit_code_epilog);                  \
 978                 if (temp->flag & jit_flag_patch)                        \
 979                     name##i##type(temp->u.w,                            \
 980                                   rn(node->v.w), node->w.w);            \
 981                 else {                                                  \
 982                     word = name##i##type##_p(_jit->pc.w,                \
 983                                              rn(node->v.w), node->w.w); \
 984                     patch(word, node);                                  \
 985                 }                                                       \
 986                 break;
 987 #define case_brf(name)                                                  \
 988             case jit_code_##name##i_f:                                  \
 989                 temp = node->u.n;                                       \
 990                 assert(temp->code == jit_code_label ||                  \
 991                        temp->code == jit_code_epilog);                  \
 992                 if (temp->flag & jit_flag_patch)                        \
 993                     name##i_f(temp->u.w, rn(node->v.w),                 \
 994                               (jit_float32_t *)node->w.n->u.w);         \
 995                 else {                                                  \
 996                     word = name##i_f_p(_jit->pc.w, rn(node->v.w),       \
 997                                        (jit_float32_t *)node->w.n->u.w);\
 998                     patch(word, node);                                  \
 999                 }                                                       \
1000                 break
1001 #define case_brd(name)                                                  \
1002             case jit_code_##name##i_d:                                  \
1003                 temp = node->u.n;                                       \
1004                 assert(temp->code == jit_code_label ||                  \
1005                        temp->code == jit_code_epilog);                  \
1006                 if (temp->flag & jit_flag_patch)                        \
1007                     name##i_d(temp->u.w, rn(node->v.w),                 \
1008                               (jit_float64_t *)node->w.n->u.w);         \
1009                 else {                                                  \
1010                     word = name##i_d_p(_jit->pc.w, rn(node->v.w),       \
1011                                        (jit_float64_t *)node->w.n->u.w);\
1012                     patch(word, node);                                  \
1013                 }                                                       \
1014                 break
1015 #if DEVEL_DISASSEMBLER
1016     prevw = _jit->pc.w;
1017 #endif
1018     for (node = _jitc->head; node; node = node->next) {
1019         if (_jit->pc.uc >= _jitc->code.end)
1020             return (NULL);
1021
1022 #if DEVEL_DISASSEMBLER
1023         node->offset = (jit_uword_t)_jit->pc.w - (jit_uword_t)prevw;
1024         prevw = _jit->pc.w;
1025 #endif
1026         value = jit_classify(node->code);
1027         jit_regarg_set(node, value);
1028         switch (node->code) {
1029             case jit_code_align:
1030                 /* Must align to a power of two */
1031                 assert(!(node->u.w & (node->u.w - 1)));
1032                 if ((word = _jit->pc.w & (node->u.w - 1)))
1033                     nop(node->u.w - word);
1034                 break;
1035             case jit_code_note:         case jit_code_name:
1036                 node->u.w = _jit->pc.w;
1037                 break;
1038             case jit_code_label:
1039                 if ((node->link || (node->flag & jit_flag_use)) &&
1040                     (word = _jit->pc.w & 3))
1041                     nop(4 - word);
1042                 /* remember label is defined */
1043                 node->flag |= jit_flag_patch;
1044                 node->u.w = _jit->pc.w;
1045                 break;
1046                 case_rrr(add,);
1047                 case_rrw(add,);
1048                 case_rrr(addc,);
1049                 case_rrw(addc,);
1050                 case_rrr(addx,);
1051                 case_rrw(addx,);
1052                 case_rrr(sub,);
1053                 case_rrw(sub,);
1054                 case_rrr(subc,);
1055                 case_rrw(subc,);
1056                 case_rrr(subx,);
1057                 case_rrw(subx,);
1058                 case_rrw(rsb,);
1059                 case_rrr(mul,);
1060                 case_rrw(mul,);
1061                 case_rrrr(qmul,);
1062                 case_rrrw(qmul,);
1063                 case_rrrr(qmul, _u);
1064                 case_rrrw(qmul, _u);
1065                 case_rrr(div,);
1066                 case_rrw(div,);
1067                 case_rrr(div, _u);
1068                 case_rrw(div, _u);
1069                 case_rrr(rem,);
1070                 case_rrw(rem,);
1071                 case_rrr(rem, _u);
1072                 case_rrw(rem, _u);
1073                 case_rrrr(qdiv,);
1074                 case_rrrw(qdiv,);
1075                 case_rrrr(qdiv, _u);
1076                 case_rrrw(qdiv, _u);
1077                 case_rrr(lsh,);
1078                 case_rrw(lsh,);
1079                 case_rrr(rsh,);
1080                 case_rrw(rsh,);
1081                 case_rrr(rsh, _u);
1082                 case_rrw(rsh, _u);
1083                 case_rr(neg,);
1084                 case_rr(com,);
1085                 case_rrr(and,);
1086                 case_rrw(and,);
1087                 case_rrr(or,);
1088                 case_rrw(or,);
1089                 case_rrr(xor,);
1090                 case_rrw(xor,);
1091                 case_rr(trunc, _f_i);
1092                 case_rr(trunc, _d_i);
1093 #if __WORDSIZE == 64
1094                 case_rr(trunc, _f_l);
1095                 case_rr(trunc, _d_l);
1096 #endif
1097                 case_rr(ld, _c);
1098                 case_rw(ld, _c);
1099                 case_rr(ld, _uc);
1100                 case_rw(ld, _uc);
1101                 case_rr(ld, _s);
1102                 case_rw(ld, _s);
1103                 case_rr(ld, _us);
1104                 case_rw(ld, _us);
1105                 case_rr(ld, _i);
1106                 case_rw(ld, _i);
1107 #if __WORDSIZE == 64
1108                 case_rr(ld, _ui);
1109                 case_rw(ld, _ui);
1110                 case_rr(ld, _l);
1111                 case_rw(ld, _l);
1112 #endif
1113                 case_rrr(ldx, _c);
1114                 case_rrw(ldx, _c);
1115                 case_rrr(ldx, _uc);
1116                 case_rrw(ldx, _uc);
1117                 case_rrr(ldx, _s);
1118                 case_rrw(ldx, _s);
1119                 case_rrr(ldx, _us);
1120                 case_rrw(ldx, _us);
1121                 case_rrr(ldx, _i);
1122                 case_rrw(ldx, _i);
1123 #if __WORDSIZE == 64
1124                 case_rrr(ldx, _ui);
1125                 case_rrw(ldx, _ui);
1126                 case_rrr(ldx, _l);
1127                 case_rrw(ldx, _l);
1128 #endif
1129                 case_rr(st, _c);
1130                 case_wr(st, _c);
1131                 case_rr(st, _s);
1132                 case_wr(st, _s);
1133                 case_rr(st, _i);
1134                 case_wr(st, _i);
1135 #if __WORDSIZE == 64
1136                 case_rr(st, _l);
1137                 case_wr(st, _l);
1138 #endif
1139                 case_rrr(stx, _c);
1140                 case_wrr(stx, _c);
1141                 case_rrr(stx, _s);
1142                 case_wrr(stx, _s);
1143                 case_rrr(stx, _i);
1144                 case_wrr(stx, _i);
1145 #if __WORDSIZE == 64
1146                 case_rrr(stx, _l);
1147                 case_wrr(stx, _l);
1148 #endif
1149                 case_rr(hton, _us);
1150                 case_rr(hton, _ui);
1151 #if __WORDSIZE == 64
1152                 case_rr(hton, _ul);
1153 #endif
1154                 case_rr(bswap, _us);
1155                 case_rr(bswap, _ui);
1156 #if __WORDSIZE == 64
1157                 case_rr(bswap, _ul);
1158 #endif
1159                 case_rr(ext, _c);
1160                 case_rr(ext, _uc);
1161                 case_rr(ext, _s);
1162                 case_rr(ext, _us);
1163 #if __WORDSIZE == 64
1164                 case_rr(ext, _i);
1165                 case_rr(ext, _ui);
1166 #endif
1167             case jit_code_casr:
1168                 casr(rn(node->u.w), rn(node->v.w),
1169                      rn(node->w.q.l), rn(node->w.q.h));
1170                 break;
1171             case jit_code_casi:
1172                 casi(rn(node->u.w), node->v.w,
1173                      rn(node->w.q.l), rn(node->w.q.h));
1174                 break;
1175                 case_rrr(movn,);
1176                 case_rrr(movz,);
1177                 case_rr(mov,);
1178             case jit_code_movi:
1179                 if (node->flag & jit_flag_node) {
1180                     temp = node->v.n;
1181                     if (temp->code == jit_code_data ||
1182                         (temp->code == jit_code_label &&
1183                          (temp->flag & jit_flag_patch)))
1184                         movi(rn(node->u.w), temp->u.w);
1185                     else {
1186                         assert(temp->code == jit_code_label ||
1187                                temp->code == jit_code_epilog);
1188                         word = movi_p(rn(node->u.w), temp->u.w);
1189                         patch(word, node);
1190                     }
1191                 }
1192                 else
1193                     movi(rn(node->u.w), node->v.w);
1194                 break;
1195                 case_rrr(lt,);
1196                 case_rrw(lt,);
1197                 case_rrr(lt, _u);
1198                 case_rrw(lt, _u);
1199                 case_rrr(le,);
1200                 case_rrw(le,);
1201                 case_rrr(le, _u);
1202                 case_rrw(le, _u);
1203                 case_rrr(eq,);
1204                 case_rrw(eq,);
1205                 case_rrr(ge,);
1206                 case_rrw(ge,);
1207                 case_rrr(ge, _u);
1208                 case_rrw(ge, _u);
1209                 case_rrr(gt,);
1210                 case_rrw(gt,);
1211                 case_rrr(gt, _u);
1212                 case_rrw(gt, _u);
1213                 case_rrr(ne,);
1214                 case_rrw(ne,);
1215                 case_brr(blt,);
1216                 case_brw(blt,);
1217                 case_brr(blt, _u);
1218                 case_brw(blt, _u);
1219                 case_brr(ble,);
1220                 case_brw(ble,);
1221                 case_brr(ble, _u);
1222                 case_brw(ble, _u);
1223                 case_brr(beq,);
1224                 case_brw(beq,);
1225                 case_brr(bge,);
1226                 case_brw(bge,);
1227                 case_brr(bge, _u);
1228                 case_brw(bge, _u);
1229                 case_brr(bgt,);
1230                 case_brw(bgt,);
1231                 case_brr(bgt, _u);
1232                 case_brw(bgt, _u);
1233                 case_brr(bne,);
1234                 case_brw(bne,);
1235                 case_brr(boadd,);
1236                 case_brw(boadd,);
1237                 case_brr(boadd, _u);
1238                 case_brw(boadd, _u);
1239                 case_brr(bxadd,);
1240                 case_brw(bxadd,);
1241                 case_brr(bxadd, _u);
1242                 case_brw(bxadd, _u);
1243                 case_brr(bosub,);
1244                 case_brw(bosub,);
1245                 case_brr(bosub, _u);
1246                 case_brw(bosub, _u);
1247                 case_brr(bxsub,);
1248                 case_brw(bxsub,);
1249                 case_brr(bxsub, _u);
1250                 case_brw(bxsub, _u);
1251                 case_brr(bms,);
1252                 case_brw(bms,);
1253                 case_brr(bmc,);
1254                 case_brw(bmc,);
1255                 case_rrr(add, _f);
1256                 case_rrf(add);
1257                 case_rrr(sub, _f);
1258                 case_rrf(sub);
1259                 case_rrf(rsb);
1260                 case_rrr(mul, _f);
1261                 case_rrf(mul);
1262                 case_rrr(div, _f);
1263                 case_rrf(div);
1264                 case_rr(abs, _f);
1265                 case_rr(neg, _f);
1266                 case_rr(sqrt, _f);
1267                 case_rr(ext, _f);
1268                 case_rr(ld, _f);
1269                 case_rw(ld, _f);
1270                 case_rrr(ldx, _f);
1271                 case_rrw(ldx, _f);
1272                 case_rr(st, _f);
1273                 case_wr(st, _f);
1274                 case_rrr(stx, _f);
1275                 case_wrr(stx, _f);
1276                 case_rr(mov, _f);
1277             case jit_code_movi_f:
1278                 assert_data(node);
1279                 movi_f(rn(node->u.w), (jit_float32_t *)node->v.n->u.w);
1280                 break;
1281                 case_rr(ext, _d_f);
1282                 case_rrr(lt, _f);
1283                 case_rrf(lt);
1284                 case_rrr(le, _f);
1285                 case_rrf(le);
1286                 case_rrr(eq, _f);
1287                 case_rrf(eq);
1288                 case_rrr(ge, _f);
1289                 case_rrf(ge);
1290                 case_rrr(gt, _f);
1291                 case_rrf(gt);
1292                 case_rrr(ne, _f);
1293                 case_rrf(ne);
1294                 case_rrr(unlt, _f);
1295                 case_rrf(unlt);
1296                 case_rrr(unle, _f);
1297                 case_rrf(unle);
1298                 case_rrr(uneq, _f);
1299                 case_rrf(uneq);
1300                 case_rrr(unge, _f);
1301                 case_rrf(unge);
1302                 case_rrr(ungt, _f);
1303                 case_rrf(ungt);
1304                 case_rrr(ltgt, _f);
1305                 case_rrf(ltgt);
1306                 case_rrr(ord, _f);
1307                 case_rrf(ord);
1308                 case_rrr(unord, _f);
1309                 case_rrf(unord);
1310                 case_brr(blt, _f);
1311                 case_brf(blt);
1312                 case_brr(ble, _f);
1313                 case_brf(ble);
1314                 case_brr(beq, _f);
1315                 case_brf(beq);
1316                 case_brr(bge, _f);
1317                 case_brf(bge);
1318                 case_brr(bgt, _f);
1319                 case_brf(bgt);
1320                 case_brr(bne, _f);
1321                 case_brf(bne);
1322                 case_brr(bunlt, _f);
1323                 case_brf(bunlt);
1324                 case_brr(bunle, _f);
1325                 case_brf(bunle);
1326                 case_brr(buneq, _f);
1327                 case_brf(buneq);
1328                 case_brr(bunge, _f);
1329                 case_brf(bunge);
1330                 case_brr(bungt, _f);
1331                 case_brf(bungt);
1332                 case_brr(bltgt, _f);
1333                 case_brf(bltgt);
1334                 case_brr(bord, _f);
1335                 case_brf(bord);
1336                 case_brr(bunord, _f);
1337                 case_brf(bunord);
1338                 case_rrr(add, _d);
1339                 case_rrd(add);
1340                 case_rrr(sub, _d);
1341                 case_rrd(sub);
1342                 case_rrd(rsb);
1343                 case_rrr(mul, _d);
1344                 case_rrd(mul);
1345                 case_rrr(div, _d);
1346                 case_rrd(div);
1347                 case_rr(abs, _d);
1348                 case_rr(neg, _d);
1349                 case_rr(sqrt, _d);
1350                 case_rr(ext, _d);
1351                 case_rr(ld, _d);
1352                 case_rw(ld, _d);
1353                 case_rrr(ldx, _d);
1354                 case_rrw(ldx, _d);
1355                 case_rr(st, _d);
1356                 case_wr(st, _d);
1357                 case_rrr(stx, _d);
1358                 case_wrr(stx, _d);
1359                 case_rr(mov, _d);
1360             case jit_code_movi_d:
1361                 assert_data(node);
1362                 movi_d(rn(node->u.w), (jit_float64_t *)node->v.n->u.w);
1363                 break;
1364                 case_rr(ext, _f_d);
1365                 case_rrr(lt, _d);
1366                 case_rrd(lt);
1367                 case_rrr(le, _d);
1368                 case_rrd(le);
1369                 case_rrr(eq, _d);
1370                 case_rrd(eq);
1371                 case_rrr(ge, _d);
1372                 case_rrd(ge);
1373                 case_rrr(gt, _d);
1374                 case_rrd(gt);
1375                 case_rrr(ne, _d);
1376                 case_rrd(ne);
1377                 case_rrr(unlt, _d);
1378                 case_rrd(unlt);
1379                 case_rrr(unle, _d);
1380                 case_rrd(unle);
1381                 case_rrr(uneq, _d);
1382                 case_rrd(uneq);
1383                 case_rrr(unge, _d);
1384                 case_rrd(unge);
1385                 case_rrr(ungt, _d);
1386                 case_rrd(ungt);
1387                 case_rrr(ltgt, _d);
1388                 case_rrd(ltgt);
1389                 case_rrr(ord, _d);
1390                 case_rrd(ord);
1391                 case_rrr(unord, _d);
1392                 case_rrd(unord);
1393                 case_brr(blt, _d);
1394                 case_brd(blt);
1395                 case_brr(ble, _d);
1396                 case_brd(ble);
1397                 case_brr(beq, _d);
1398                 case_brd(beq);
1399                 case_brr(bge, _d);
1400                 case_brd(bge);
1401                 case_brr(bgt, _d);
1402                 case_brd(bgt);
1403                 case_brr(bne, _d);
1404                 case_brd(bne);
1405                 case_brr(bunlt, _d);
1406                 case_brd(bunlt);
1407                 case_brr(bunle, _d);
1408                 case_brd(bunle);
1409                 case_brr(buneq, _d);
1410                 case_brd(buneq);
1411                 case_brr(bunge, _d);
1412                 case_brd(bunge);
1413                 case_brr(bungt, _d);
1414                 case_brd(bungt);
1415                 case_brr(bltgt, _d);
1416                 case_brd(bltgt);
1417                 case_brr(bord, _d);
1418                 case_brd(bord);
1419                 case_brr(bunord, _d);
1420                 case_brd(bunord);
1421             case jit_code_jmpr:
1422                 jmpr(rn(node->u.w));
1423                 break;
1424             case jit_code_jmpi:
1425                 if (node->flag & jit_flag_node) {
1426                     temp = node->u.n;
1427                     assert(temp->code == jit_code_label ||
1428                            temp->code == jit_code_epilog);
1429                     if (temp->flag & jit_flag_patch)
1430                         jmpi(temp->u.w);
1431                     else {
1432                         word = jmpi_p(_jit->pc.w);
1433                         patch(word, node);
1434                     }
1435                 }
1436                 else
1437                     jmpi(node->u.w);
1438                 break;
1439             case jit_code_callr:
1440                 callr(rn(node->u.w));
1441                 break;
1442             case jit_code_calli:
1443                 if (node->flag & jit_flag_node) {
1444                     temp = node->u.n;
1445                     assert(temp->code == jit_code_label ||
1446                            temp->code == jit_code_epilog);
1447                     if (temp->flag & jit_flag_patch)
1448                         calli(temp->u.w);
1449                     else {
1450                         word = calli_p(_jit->pc.w);
1451                         patch(word, node);
1452                     }
1453                 }
1454                 else
1455                     calli(node->u.w);
1456                 break;
1457             case jit_code_prolog:
1458                 _jitc->function = _jitc->functions.ptr + node->w.w;
1459                 undo.node = node;
1460                 undo.word = _jit->pc.w;
1461 #if DEVEL_DISASSEMBLER
1462                 undo.prevw = prevw;
1463 #endif
1464                 undo.patch_offset = _jitc->patches.offset;
1465             restart_function:
1466                 _jitc->again = 0;
1467                 prolog(node);
1468                 break;
1469             case jit_code_epilog:
1470                 assert(_jitc->function == _jitc->functions.ptr + node->w.w);
1471                 if (_jitc->again) {
1472                     for (temp = undo.node->next;
1473                          temp != node; temp = temp->next) {
1474                         if (temp->code == jit_code_label ||
1475                             temp->code == jit_code_epilog)
1476                             temp->flag &= ~jit_flag_patch;
1477                     }
1478                     temp->flag &= ~jit_flag_patch;
1479                     node = undo.node;
1480                     _jit->pc.w = undo.word;
1481 #if DEVEL_DISASSEMBLER
1482                     prevw = undo.prevw;
1483 #endif
1484                     _jitc->patches.offset = undo.patch_offset;
1485                     goto restart_function;
1486                 }
1487                 if (node->link && (word = _jit->pc.w & 3))
1488                     nop(4 - word);
1489                 /* remember label is defined */
1490                 node->flag |= jit_flag_patch;
1491                 node->u.w = _jit->pc.w;
1492                 epilog(node);
1493                 _jitc->function = NULL;
1494                 break;
1495             case jit_code_va_start:
1496                 vastart(rn(node->u.w));
1497                 break;
1498             case jit_code_va_arg:
1499                 vaarg(rn(node->u.w), rn(node->v.w));
1500                 break;
1501             case jit_code_va_arg_d:
1502                 vaarg_d(rn(node->u.w), rn(node->v.w));
1503                 break;
1504             case jit_code_live:                 case jit_code_ellipsis:
1505             case jit_code_va_push:
1506             case jit_code_allocai:              case jit_code_allocar:
1507             case jit_code_arg:
1508             case jit_code_arg_f:                case jit_code_arg_d:
1509             case jit_code_va_end:
1510             case jit_code_ret:
1511             case jit_code_retr:                 case jit_code_reti:
1512             case jit_code_retr_f:               case jit_code_reti_f:
1513             case jit_code_retr_d:               case jit_code_reti_d:
1514             case jit_code_getarg_c:             case jit_code_getarg_uc:
1515             case jit_code_getarg_s:             case jit_code_getarg_us:
1516             case jit_code_getarg_i:
1517 #if __WORDSIZE == 64
1518             case jit_code_getarg_ui:            case jit_code_getarg_l:
1519 #endif
1520             case jit_code_getarg_f:             case jit_code_getarg_d:
1521             case jit_code_putargr:              case jit_code_putargi:
1522             case jit_code_putargr_f:            case jit_code_putargi_f:
1523             case jit_code_putargr_d:            case jit_code_putargi_d:
1524             case jit_code_pushargr:             case jit_code_pushargi:
1525             case jit_code_pushargr_f:           case jit_code_pushargi_f:
1526             case jit_code_pushargr_d:           case jit_code_pushargi_d:
1527             case jit_code_retval_c:             case jit_code_retval_uc:
1528             case jit_code_retval_s:             case jit_code_retval_us:
1529             case jit_code_retval_i:
1530 #if __WORDSIZE == 64
1531             case jit_code_retval_ui:            case jit_code_retval_l:
1532 #endif
1533             case jit_code_retval_f:             case jit_code_retval_d:
1534             case jit_code_prepare:
1535             case jit_code_finishr:              case jit_code_finishi:
1536                 break;
1537             default:
1538                 abort();
1539         }
1540         jit_regarg_clr(node, value);
1541         assert(_jitc->regarg == 0 && _jitc->synth == 0);
1542         /* update register live state */
1543         jit_reglive(node);
1544     }
1545 #undef case_brw
1546 #undef case_brr
1547 #undef case_wrr
1548 #undef case_rrw
1549 #undef case_rrr
1550 #undef case_wr
1551 #undef case_rw
1552 #undef case_rr
1553
1554     for (offset = 0; offset < _jitc->patches.offset; offset++) {
1555         node = _jitc->patches.ptr[offset].node;
1556         word = node->code == jit_code_movi ? node->v.n->u.w : node->u.n->u.w;
1557         patch_at(_jitc->patches.ptr[offset].inst, word);
1558     }
1559
1560     jit_flush(_jit->code.ptr, _jit->pc.uc);
1561
1562     return (_jit->code.ptr);
1563 }
1564
1565 #define CODE                            1
1566 #  include "jit_s390-cpu.c"
1567 #  include "jit_s390-fpu.c"
1568 #undef CODE
1569
1570 void
1571 jit_flush(void *fptr, void *tptr)
1572 {
1573 #if defined(__GNUC__)
1574     jit_word_t          f, t, s;
1575
1576     s = sysconf(_SC_PAGE_SIZE);
1577     f = (jit_word_t)fptr & -s;
1578     t = (((jit_word_t)tptr) + s - 1) & -s;
1579     __clear_cache((void *)f, (void *)t);
1580 #endif
1581 }
1582
1583 void
1584 _emit_ldxi(jit_state_t *_jit, jit_gpr_t r0, jit_gpr_t r1, jit_word_t i0)
1585 {
1586     ldxi(rn(r0), rn(r1), i0);
1587 }
1588
1589 void
1590 _emit_stxi(jit_state_t *_jit, jit_word_t i0, jit_gpr_t r0, jit_gpr_t r1)
1591 {
1592     stxi(i0, rn(r0), rn(r1));
1593 }
1594
1595 void
1596 _emit_ldxi_d(jit_state_t *_jit, jit_fpr_t r0, jit_gpr_t r1, jit_word_t i0)
1597 {
1598     ldxi_d(rn(r0), rn(r1), i0);
1599 }
1600
1601 void
1602 _emit_stxi_d(jit_state_t *_jit, jit_word_t i0, jit_gpr_t r0, jit_fpr_t r1)
1603 {
1604     stxi_d(i0, rn(r0), rn(r1));
1605 }
1606
1607 static jit_int32_t
1608 _jit_get_reg_pair(jit_state_t *_jit)
1609 {
1610     jit_int32_t         r1, r2;
1611     /* Try to find a register pair for use with operations that
1612      * require a odd based register pair. Search for the best
1613      * match to avoid spills or at least a valid operation.
1614      */
1615
1616     /* Try non callee save first */
1617     if (jit_reg_free_p(_R0) && jit_reg_free_p(_R1))
1618         r1 = _R0, r2 = _R1;
1619     else if (jit_reg_free_p(_R2) && jit_reg_free_p(_R3))
1620         r1 = _R2, r2 = _R3;
1621     else if (jit_reg_free_p(_R4) && jit_reg_free_p(_R5))
1622         r1 = _R4, r2 = _R5;
1623     /* Try callee save registers */
1624     else if (jit_reg_free_p(_R10) && jit_reg_free_p(_R11))
1625         r1 = _R10, r2 = _R11;
1626     else if (jit_reg_free_p(_R8) && jit_reg_free_p(_R9))
1627         r1 = _R8, r2 = _R9;
1628     else if (jit_reg_free_p(_R6) && jit_reg_free_p(_R7))
1629         r1 = _R6, r2 = _R7;
1630
1631     /* We *must* find a register pair */
1632     else if (jit_reg_free_if_spill_p(_R0) && jit_reg_free_if_spill_p(_R1))
1633         r1 = _R0, r2 = _R1;
1634     else if (jit_reg_free_if_spill_p(_R2) && jit_reg_free_if_spill_p(_R3))
1635         r1 = _R2, r2 = _R3;
1636     else if (jit_reg_free_if_spill_p(_R4) && jit_reg_free_if_spill_p(_R5))
1637         r1 = _R4, r2 = _R5;
1638     else if (jit_reg_free_if_spill_p(_R10) && jit_reg_free_if_spill_p(_R11))
1639         r1 = _R10, r2 = _R11;
1640     else if (jit_reg_free_if_spill_p(_R8) && jit_reg_free_if_spill_p(_R9))
1641         r1 = _R8, r2 = _R9;
1642     else if (jit_reg_free_if_spill_p(_R6) && jit_reg_free_if_spill_p(_R7))
1643         r1 = _R6, r2 = _R7;
1644     else
1645         /* Do not jit_get_reg() all registers to avoid it */
1646         abort();
1647
1648     (void)jit_get_reg(jit_class_gpr|jit_class_named|r1);
1649     (void)jit_get_reg(jit_class_gpr|jit_class_named|r2);
1650
1651     return (r1);
1652 }
1653
1654 static void
1655 _jit_unget_reg_pair(jit_state_t *_jit, jit_int32_t reg)
1656 {
1657     jit_int32_t         r1, r2;
1658     r1 = reg;
1659     switch (r1) {
1660         case _R0:       r2 = _R1;       break;
1661         case _R2:       r2 = _R3;       break;
1662         case _R4:       r2 = _R5;       break;
1663         case _R6:       r2 = _R7;       break;
1664         case _R8:       r2 = _R9;       break;
1665         case _R10:      r2 = _R11;      break;
1666         default:        abort();
1667     }
1668     jit_unget_reg(r1);
1669     jit_unget_reg(r2);
1670 }
1671
1672 static jit_int32_t
1673 _jit_get_reg_but_zero(jit_state_t *_jit, jit_int32_t flags)
1674 {
1675     jit_int32_t         reg;
1676     reg = jit_get_reg(jit_class_gpr);
1677     if (reg == _R0) {
1678         reg = jit_get_reg(jit_class_gpr|flags);
1679         jit_unget_reg(_R0);
1680     }
1681     return (reg);
1682 }
1683
1684 static void
1685 _patch(jit_state_t *_jit, jit_word_t instr, jit_node_t *node)
1686 {
1687     jit_int32_t         flag;
1688
1689     assert(node->flag & jit_flag_node);
1690     if (node->code == jit_code_movi)
1691         flag = node->v.n->flag;
1692     else
1693         flag = node->u.n->flag;
1694     assert(!(flag & jit_flag_patch));
1695     if (_jitc->patches.offset >= _jitc->patches.length) {
1696         jit_realloc((jit_pointer_t *)&_jitc->patches.ptr,
1697                     _jitc->patches.length * sizeof(jit_patch_t),
1698                     (_jitc->patches.length + 1024) * sizeof(jit_patch_t));
1699         _jitc->patches.length += 1024;
1700     }
1701     _jitc->patches.ptr[_jitc->patches.offset].inst = instr;
1702     _jitc->patches.ptr[_jitc->patches.offset].node = node;
1703     ++_jitc->patches.offset;
1704 }