mono/mini/mini-ia64.c

   1 /*
   2  * mini-ia64.c: IA64 backend for the Mono code generator
   3  *
   4  * Authors:
   5  *   Zoltan Varga (vargaz@gmail.com)
   6  *
   7  * (C) 2003 Ximian, Inc.
   8  */
   9 #include "mini.h"
  10 #include <string.h>
  11 #include <math.h>
  12 #include <unistd.h>
  13 #include <sys/mman.h>
  14
  15 #ifdef __INTEL_COMPILER
  16 #include <ia64intrin.h>
  17 #endif
  18
  19 #include <mono/metadata/appdomain.h>
  20 #include <mono/metadata/debug-helpers.h>
  21 #include <mono/metadata/threads.h>
  22 #include <mono/metadata/profiler-private.h>
  23 #include <mono/utils/mono-math.h>
  24 #include <mono/utils/mono-hwcap-ia64.h>
  25
  26 #include "trace.h"
  27 #include "mini-ia64.h"
  28 #include "cpu-ia64.h"
  29 #include "jit-icalls.h"
  30 #include "ir-emit.h"
  31
  32 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
  33
  34 #define IS_IMM32(val) ((((guint64)val) >> 32) == 0)
  35
  36 /*
  37  * IA64 register usage:
  38  * - local registers are used for global register allocation
  39  * - r8..r11, r14..r30 is used for local register allocation
  40  * - r31 is a scratch register used within opcode implementations
  41  * - FIXME: Use out registers as well
  42  * - the first three locals are used for saving ar.pfst, b0, and sp
  43  * - compare instructions allways set p6 and p7
  44  */
  45
  46 /*
  47  * There are a lot of places where generated code is disassembled/patched.
  48  * The automatic bundling of instructions done by the code generation macros
  49  * could complicate things, so it is best to call
  50  * ia64_codegen_set_one_ins_per_bundle () at those places.
  51  */
  52
  53 #define ARGS_OFFSET 16
  54
  55 #define GP_SCRATCH_REG 31
  56 #define GP_SCRATCH_REG2 30
  57 #define FP_SCRATCH_REG 32
  58 #define FP_SCRATCH_REG2 33
  59
  60 #define LOOP_ALIGNMENT 8
  61 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
  62
  63 static const char* gregs [] = {
  64         "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7", "r8", "r9",
  65         "r10", "r11", "r12", "r13", "r14", "r15", "r16", "r17", "r18", "r19",
  66         "r20", "r21", "r22", "r23", "r24", "r25", "r26", "r27", "r28", "r29",
  67         "r30", "r31", "r32", "r33", "r34", "r35", "r36", "r37", "r38", "r39",
  68         "r40", "r41", "r42", "r43", "r44", "r45", "r46", "r47", "r48", "r49",
  69         "r50", "r51", "r52", "r53", "r54", "r55", "r56", "r57", "r58", "r59",
  70         "r60", "r61", "r62", "r63", "r64", "r65", "r66", "r67", "r68", "r69",
  71         "r70", "r71", "r72", "r73", "r74", "r75", "r76", "r77", "r78", "r79",
  72         "r80", "r81", "r82", "r83", "r84", "r85", "r86", "r87", "r88", "r89",
  73         "r90", "r91", "r92", "r93", "r94", "r95", "r96", "r97", "r98", "r99",
  74         "r100", "r101", "r102", "r103", "r104", "r105", "r106", "r107", "r108", "r109",
  75         "r110", "r111", "r112", "r113", "r114", "r115", "r116", "r117", "r118", "r119",
  76         "r120", "r121", "r122", "r123", "r124", "r125", "r126", "r127"
  77 };
  78
  79 const char*
  80 mono_arch_regname (int reg)
  81 {
  82         if (reg < 128)
  83                 return gregs [reg];
  84         else
  85                 return "unknown";
  86 }
  87
  88 static const char* fregs [] = {
  89         "f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "f9",
  90         "f10", "f11", "f12", "f13", "f14", "f15", "f16", "f17", "f18", "f19",
  91         "f20", "f21", "f22", "f23", "f24", "f25", "f26", "f27", "f28", "f29",
  92         "f30", "f31", "f32", "f33", "f34", "f35", "f36", "f37", "f38", "f39",
  93         "f40", "f41", "f42", "f43", "f44", "f45", "f46", "f47", "f48", "f49",
  94         "f50", "f51", "f52", "f53", "f54", "f55", "f56", "f57", "f58", "f59",
  95         "f60", "f61", "f62", "f63", "f64", "f65", "f66", "f67", "f68", "f69",
  96         "f70", "f71", "f72", "f73", "f74", "f75", "f76", "f77", "f78", "f79",
  97         "f80", "f81", "f82", "f83", "f84", "f85", "f86", "f87", "f88", "f89",
  98         "f90", "f91", "f92", "f93", "f94", "f95", "f96", "f97", "f98", "f99",
  99         "f100", "f101", "f102", "f103", "f104", "f105", "f106", "f107", "f108", "f109",
 100         "f110", "f111", "f112", "f113", "f114", "f115", "f116", "f117", "f118", "f119",
 101         "f120", "f121", "f122", "f123", "f124", "f125", "f126", "f127"
 102 };
 103
 104 const char*
 105 mono_arch_fregname (int reg)
 106 {
 107         if (reg < 128)
 108                 return fregs [reg];
 109         else
 110                 return "unknown";
 111 }
 112
 113 static gboolean
 114 debug_ins_sched (void)
 115 {
 116 #if 0
 117         return mono_debug_count ();
 118 #else
 119         return TRUE;
 120 #endif
 121 }
 122
 123 static gboolean
 124 debug_omit_fp (void)
 125 {
 126 #if 0
 127         return mono_debug_count ();
 128 #else
 129         return TRUE;
 130 #endif
 131 }
 132
 133 static void
 134 ia64_patch (unsigned char* code, gpointer target);
 135
 136 typedef enum {
 137         ArgInIReg,
 138         ArgInFloatReg,
 139         ArgInFloatRegR4,
 140         ArgOnStack,
 141         ArgValuetypeAddrInIReg,
 142         ArgAggregate,
 143         ArgSingleHFA,
 144         ArgDoubleHFA,
 145         ArgNone
 146 } ArgStorage;
 147
 148 typedef enum {
 149         AggregateNormal,
 150         AggregateSingleHFA,
 151         AggregateDoubleHFA
 152 } AggregateType;
 153
 154 typedef struct {
 155         gint16 offset;
 156         gint8  reg;
 157         ArgStorage storage;
 158
 159         /* Only if storage == ArgAggregate */
 160         int nregs, nslots;
 161         AggregateType atype;
 162 } ArgInfo;
 163
 164 typedef struct {
 165         int nargs;
 166         guint32 stack_usage;
 167         guint32 reg_usage;
 168         guint32 freg_usage;
 169         gboolean need_stack_align;
 170         gboolean vtype_retaddr;
 171         /* The index of the vret arg in the argument list */
 172         int vret_arg_index;
 173         ArgInfo ret;
 174         ArgInfo sig_cookie;
 175         ArgInfo args [1];
 176 } CallInfo;
 177
 178 #define DEBUG(a) if (cfg->verbose_level > 1) a
 179
 180 #define PARAM_REGS 8
 181
 182 static void inline
 183 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
 184 {
 185     ainfo->offset = *stack_size;
 186
 187     if (*gr >= PARAM_REGS) {
 188                 ainfo->storage = ArgOnStack;
 189                 (*stack_size) += sizeof (gpointer);
 190     }
 191     else {
 192                 ainfo->storage = ArgInIReg;
 193                 ainfo->reg = *gr;
 194                 *(gr) += 1;
 195     }
 196 }
 197
 198 #define FLOAT_PARAM_REGS 8
 199
 200 static void inline
 201 add_float (guint32 *gr, guint32 *fr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
 202 {
 203     ainfo->offset = *stack_size;
 204
 205     if (*gr >= PARAM_REGS) {
 206                 ainfo->storage = ArgOnStack;
 207                 (*stack_size) += sizeof (gpointer);
 208     }
 209     else {
 210                 ainfo->storage = is_double ? ArgInFloatReg : ArgInFloatRegR4;
 211                 ainfo->reg = 8 + *fr;
 212                 (*fr) += 1;
 213                 (*gr) += 1;
 214     }
 215 }
 216
 217 static void
 218 add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
 219                gboolean is_return,
 220                guint32 *gr, guint32 *fr, guint32 *stack_size)
 221 {
 222         guint32 size, i;
 223         MonoClass *klass;
 224         MonoMarshalType *info;
 225         gboolean is_hfa = TRUE;
 226         guint32 hfa_type = 0;
 227
 228         klass = mono_class_from_mono_type (type);
 229         if (type->type == MONO_TYPE_TYPEDBYREF)
 230                 size = 3 * sizeof (gpointer);
 231         else if (sig->pinvoke)
 232                 size = mono_type_native_stack_size (&klass->byval_arg, NULL);
 233         else
 234                 size = mini_type_stack_size (&klass->byval_arg, NULL);
 235
 236         if (!sig->pinvoke || (size == 0)) {
 237                 /* Allways pass in memory */
 238                 ainfo->offset = *stack_size;
 239                 *stack_size += ALIGN_TO (size, 8);
 240                 ainfo->storage = ArgOnStack;
 241
 242                 return;
 243         }
 244
 245         /* Determine whenever it is a HFA (Homogeneous Floating Point Aggregate) */
 246         info = mono_marshal_load_type_info (klass);
 247         g_assert (info);
 248         for (i = 0; i < info->num_fields; ++i) {
 249                 guint32 ftype = info->fields [i].field->type->type;
 250                 if (!(info->fields [i].field->type->byref) &&
 251                         ((ftype == MONO_TYPE_R4) || (ftype == MONO_TYPE_R8))) {
 252                         if (hfa_type == 0)
 253                                 hfa_type = ftype;
 254                         else if (hfa_type != ftype)
 255                                 is_hfa = FALSE;
 256                 }
 257                 else
 258                         is_hfa = FALSE;
 259         }
 260         if (hfa_type == 0)
 261                 is_hfa = FALSE;
 262
 263         ainfo->storage = ArgAggregate;
 264         ainfo->atype = AggregateNormal;
 265
 266         if (is_hfa) {
 267                 ainfo->atype = hfa_type == MONO_TYPE_R4 ? AggregateSingleHFA : AggregateDoubleHFA;
 268                 if (is_return) {
 269                         if (info->num_fields <= 8) {
 270                                 ainfo->reg = 8;
 271                                 ainfo->nregs = info->num_fields;
 272                                 ainfo->nslots = ainfo->nregs;
 273                                 return;
 274                         }
 275                         /* Fall through */
 276                 }
 277                 else {
 278                         if ((*fr) + info->num_fields > 8)
 279                                 NOT_IMPLEMENTED;
 280
 281                         ainfo->reg = 8 + (*fr);
 282                         ainfo->nregs = info->num_fields;
 283                         ainfo->nslots = ainfo->nregs;
 284                         (*fr) += info->num_fields;
 285                         if (ainfo->atype == AggregateSingleHFA) {
 286                                 /*
 287                                  * FIXME: Have to keep track of the parameter slot number, which is
 288                                  * not the same as *gr.
 289                                  */
 290                                 (*gr) += ALIGN_TO (info->num_fields, 2) / 2;
 291                         } else {
 292                                 (*gr) += info->num_fields;
 293                         }
 294                         return;
 295                 }
 296         }
 297
 298         /* This also handles returning of TypedByRef used by some icalls */
 299         if (is_return) {
 300                 if (size <= 32) {
 301                         ainfo->reg = IA64_R8;
 302                         ainfo->nregs = (size + 7) / 8;
 303                         ainfo->nslots = ainfo->nregs;
 304                         return;
 305                 }
 306                 NOT_IMPLEMENTED;
 307         }
 308
 309         ainfo->reg = (*gr);
 310         ainfo->offset = *stack_size;
 311         ainfo->nslots = (size + 7) / 8;
 312
 313         if (((*gr) + ainfo->nslots) <= 8) {
 314                 /* Fits entirely in registers */
 315                 ainfo->nregs = ainfo->nslots;
 316                 (*gr) += ainfo->nregs;
 317                 return;
 318         }
 319
 320         ainfo->nregs = 8 - (*gr);
 321         (*gr) = 8;
 322         (*stack_size) += (ainfo->nslots - ainfo->nregs) * 8;
 323 }
 324
 325 /*
 326  * get_call_info:
 327  *
 328  *  Obtain information about a call according to the calling convention.
 329  * For IA64, see the "Itanium Software Conventions and Runtime Architecture
 330  * Gude" document for more information.
 331  */
 332 static CallInfo*
 333 get_call_info (MonoCompile *cfg, MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
 334 {
 335         guint32 i, gr, fr, pstart;
 336         MonoType *ret_type;
 337         int n = sig->hasthis + sig->param_count;
 338         guint32 stack_size = 0;
 339         CallInfo *cinfo;
 340
 341         if (mp)
 342                 cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
 343         else
 344                 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
 345
 346         gr = 0;
 347         fr = 0;
 348
 349         /* return value */
 350         {
 351                 ret_type = mini_get_underlying_type (sig->ret);
 352                 switch (ret_type->type) {
 353                 case MONO_TYPE_BOOLEAN:
 354                 case MONO_TYPE_I1:
 355                 case MONO_TYPE_U1:
 356                 case MONO_TYPE_I2:
 357                 case MONO_TYPE_U2:
 358                 case MONO_TYPE_CHAR:
 359                 case MONO_TYPE_I4:
 360                 case MONO_TYPE_U4:
 361                 case MONO_TYPE_I:
 362                 case MONO_TYPE_U:
 363                 case MONO_TYPE_PTR:
 364                 case MONO_TYPE_FNPTR:
 365                 case MONO_TYPE_CLASS:
 366                 case MONO_TYPE_OBJECT:
 367                 case MONO_TYPE_SZARRAY:
 368                 case MONO_TYPE_ARRAY:
 369                 case MONO_TYPE_STRING:
 370                         cinfo->ret.storage = ArgInIReg;
 371                         cinfo->ret.reg = IA64_R8;
 372                         break;
 373                 case MONO_TYPE_U8:
 374                 case MONO_TYPE_I8:
 375                         cinfo->ret.storage = ArgInIReg;
 376                         cinfo->ret.reg = IA64_R8;
 377                         break;
 378                 case MONO_TYPE_R4:
 379                 case MONO_TYPE_R8:
 380                         cinfo->ret.storage = ArgInFloatReg;
 381                         cinfo->ret.reg = 8;
 382                         break;
 383                 case MONO_TYPE_GENERICINST:
 384                         if (!mono_type_generic_inst_is_valuetype (ret_type)) {
 385                                 cinfo->ret.storage = ArgInIReg;
 386                                 cinfo->ret.reg = IA64_R8;
 387                                 break;
 388                         }
 389                         /* Fall through */
 390                 case MONO_TYPE_VALUETYPE:
 391                 case MONO_TYPE_TYPEDBYREF: {
 392                         guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
 393
 394                         if (sig->ret->byref) {
 395                                 /* This seems to happen with ldfld wrappers */
 396                                 cinfo->ret.storage = ArgInIReg;
 397                         } else {
 398                                 add_valuetype (sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
 399                                 if (cinfo->ret.storage == ArgOnStack) {
 400                                         /* The caller passes the address where the value is stored */
 401                                         cinfo->vtype_retaddr = TRUE;
 402                                 }
 403                         }
 404                         break;
 405                 }
 406                 case MONO_TYPE_VOID:
 407                         cinfo->ret.storage = ArgNone;
 408                         break;
 409                 default:
 410                         g_error ("Can't handle as return value 0x%x", sig->ret->type);
 411                 }
 412         }
 413
 414         pstart = 0;
 415         /*
 416          * To simplify get_this_arg_reg () and LLVM integration, emit the vret arg after
 417          * the first argument, allowing 'this' to be always passed in the first arg reg.
 418          * Also do this if the first argument is a reference type, since virtual calls
 419          * are sometimes made using calli without sig->hasthis set, like in the delegate
 420          * invoke wrappers.
 421          */
 422         if (cinfo->vtype_retaddr && !is_pinvoke && (sig->hasthis || (sig->param_count > 0 && MONO_TYPE_IS_REFERENCE (mini_get_underlying_type (sig->params [0]))))) {
 423                 if (sig->hasthis) {
 424                         add_general (&gr, &stack_size, cinfo->args + 0);
 425                 } else {
 426                         add_general (&gr, &stack_size, &cinfo->args [sig->hasthis + 0]);
 427                         pstart = 1;
 428                 }
 429                 add_general (&gr, &stack_size, &cinfo->ret);
 430                 if (cinfo->ret.storage == ArgInIReg)
 431                         cinfo->ret.storage = ArgValuetypeAddrInIReg;
 432                 cinfo->vret_arg_index = 1;
 433         } else {
 434                 /* this */
 435                 if (sig->hasthis)
 436                         add_general (&gr, &stack_size, cinfo->args + 0);
 437
 438                 if (cinfo->vtype_retaddr) {
 439                         add_general (&gr, &stack_size, &cinfo->ret);
 440                         if (cinfo->ret.storage == ArgInIReg)
 441                                 cinfo->ret.storage = ArgValuetypeAddrInIReg;
 442                 }
 443         }
 444
 445         if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
 446                 gr = PARAM_REGS;
 447                 fr = FLOAT_PARAM_REGS;
 448
 449                 /* Emit the signature cookie just before the implicit arguments */
 450                 add_general (&gr, &stack_size, &cinfo->sig_cookie);
 451         }
 452
 453         for (i = pstart; i < sig->param_count; ++i) {
 454                 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
 455                 MonoType *ptype;
 456
 457                 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
 458                         /* We allways pass the sig cookie on the stack for simplicity */
 459                         /*
 460                          * Prevent implicit arguments + the sig cookie from being passed
 461                          * in registers.
 462                          */
 463                         gr = PARAM_REGS;
 464                         fr = FLOAT_PARAM_REGS;
 465
 466                         /* Emit the signature cookie just before the implicit arguments */
 467                         add_general (&gr, &stack_size, &cinfo->sig_cookie);
 468                 }
 469
 470                 if (sig->params [i]->byref) {
 471                         add_general (&gr, &stack_size, ainfo);
 472                         continue;
 473                 }
 474                 ptype = mini_get_underlying_type (sig->params [i]);
 475                 switch (ptype->type) {
 476                 case MONO_TYPE_BOOLEAN:
 477                 case MONO_TYPE_I1:
 478                 case MONO_TYPE_U1:
 479                         add_general (&gr, &stack_size, ainfo);
 480                         break;
 481                 case MONO_TYPE_I2:
 482                 case MONO_TYPE_U2:
 483                 case MONO_TYPE_CHAR:
 484                         add_general (&gr, &stack_size, ainfo);
 485                         break;
 486                 case MONO_TYPE_I4:
 487                 case MONO_TYPE_U4:
 488                         add_general (&gr, &stack_size, ainfo);
 489                         break;
 490                 case MONO_TYPE_I:
 491                 case MONO_TYPE_U:
 492                 case MONO_TYPE_PTR:
 493                 case MONO_TYPE_FNPTR:
 494                 case MONO_TYPE_CLASS:
 495                 case MONO_TYPE_OBJECT:
 496                 case MONO_TYPE_STRING:
 497                 case MONO_TYPE_SZARRAY:
 498                 case MONO_TYPE_ARRAY:
 499                         add_general (&gr, &stack_size, ainfo);
 500                         break;
 501                 case MONO_TYPE_GENERICINST:
 502                         if (!mono_type_generic_inst_is_valuetype (ptype)) {
 503                                 add_general (&gr, &stack_size, ainfo);
 504                                 break;
 505                         }
 506                         /* Fall through */
 507                 case MONO_TYPE_VALUETYPE:
 508                 case MONO_TYPE_TYPEDBYREF:
 509                         /* FIXME: */
 510                         /* We allways pass valuetypes on the stack */
 511                         add_valuetype (sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
 512                         break;
 513                 case MONO_TYPE_U8:
 514                 case MONO_TYPE_I8:
 515                         add_general (&gr, &stack_size, ainfo);
 516                         break;
 517                 case MONO_TYPE_R4:
 518                         add_float (&gr, &fr, &stack_size, ainfo, FALSE);
 519                         break;
 520                 case MONO_TYPE_R8:
 521                         add_float (&gr, &fr, &stack_size, ainfo, TRUE);
 522                         break;
 523                 default:
 524                         g_assert_not_reached ();
 525                 }
 526         }
 527
 528         if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
 529                 gr = PARAM_REGS;
 530                 fr = FLOAT_PARAM_REGS;
 531
 532                 /* Emit the signature cookie just before the implicit arguments */
 533                 add_general (&gr, &stack_size, &cinfo->sig_cookie);
 534         }
 535
 536         cinfo->stack_usage = stack_size;
 537         cinfo->reg_usage = gr;
 538         cinfo->freg_usage = fr;
 539         return cinfo;
 540 }
 541
 542 /*
 543  * mono_arch_get_argument_info:
 544  * @csig:  a method signature
 545  * @param_count: the number of parameters to consider
 546  * @arg_info: an array to store the result infos
 547  *
 548  * Gathers information on parameters such as size, alignment and
 549  * padding. arg_info should be large enought to hold param_count + 1 entries.
 550  *
 551  * Returns the size of the argument area on the stack.
 552  */
 553 int
 554 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
 555 {
 556         int k;
 557         CallInfo *cinfo = get_call_info (NULL, NULL, csig, FALSE);
 558         guint32 args_size = cinfo->stack_usage;
 559
 560         /* The arguments are saved to a stack area in mono_arch_instrument_prolog */
 561         if (csig->hasthis) {
 562                 arg_info [0].offset = 0;
 563         }
 564
 565         for (k = 0; k < param_count; k++) {
 566                 arg_info [k + 1].offset = ((k + csig->hasthis) * 8);
 567                 /* FIXME: */
 568                 arg_info [k + 1].size = 0;
 569         }
 570
 571         g_free (cinfo);
 572
 573         return args_size;
 574 }
 575
 576 /*
 577  * Initialize the cpu to execute managed code.
 578  */
 579 void
 580 mono_arch_cpu_init (void)
 581 {
 582 }
 583
 584 /*
 585  * Initialize architecture specific code.
 586  */
 587 void
 588 mono_arch_init (void)
 589 {
 590 }
 591
 592 /*
 593  * Cleanup architecture specific code.
 594  */
 595 void
 596 mono_arch_cleanup (void)
 597 {
 598 }
 599
 600 /*
 601  * This function returns the optimizations supported on this cpu.
 602  */
 603 guint32
 604 mono_arch_cpu_optimizations (guint32 *exclude_mask)
 605 {
 606         *exclude_mask = 0;
 607
 608         return 0;
 609 }
 610
 611 /*
 612  * This function test for all SIMD functions supported.
 613  *
 614  * Returns a bitmask corresponding to all supported versions.
 615  *
 616  */
 617 guint32
 618 mono_arch_cpu_enumerate_simd_versions (void)
 619 {
 620         /* SIMD is currently unimplemented */
 621         return 0;
 622 }
 623
 624 GList *
 625 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
 626 {
 627         GList *vars = NULL;
 628         int i;
 629         MonoMethodSignature *sig;
 630         MonoMethodHeader *header;
 631         CallInfo *cinfo;
 632
 633         header = cfg->header;
 634
 635         sig = mono_method_signature (cfg->method);
 636
 637         cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
 638
 639         for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
 640                 MonoInst *ins = cfg->args [i];
 641
 642                 ArgInfo *ainfo = &cinfo->args [i];
 643
 644                 if (ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT))
 645                         continue;
 646
 647                 if (ainfo->storage == ArgInIReg) {
 648                         /* The input registers are non-volatile */
 649                         ins->opcode = OP_REGVAR;
 650                         ins->dreg = 32 + ainfo->reg;
 651                 }
 652         }
 653
 654         for (i = 0; i < cfg->num_varinfo; i++) {
 655                 MonoInst *ins = cfg->varinfo [i];
 656                 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
 657
 658                 /* unused vars */
 659                 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
 660                         continue;
 661
 662                 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
 663                     (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
 664                         continue;
 665
 666                 if (mono_is_regsize_var (ins->inst_vtype)) {
 667                         g_assert (MONO_VARINFO (cfg, i)->reg == -1);
 668                         g_assert (i == vmv->idx);
 669                         vars = g_list_prepend (vars, vmv);
 670                 }
 671         }
 672
 673         vars = mono_varlist_sort (cfg, vars, 0);
 674
 675         return vars;
 676 }
 677
 678 static void
 679 mono_ia64_alloc_stacked_registers (MonoCompile *cfg)
 680 {
 681         CallInfo *cinfo;
 682         guint32 reserved_regs;
 683         MonoMethodHeader *header;
 684
 685         if (cfg->arch.reg_local0 > 0)
 686                 /* Already done */
 687                 return;
 688
 689         cinfo = get_call_info (cfg, cfg->mempool, mono_method_signature (cfg->method), FALSE);
 690
 691         header = cfg->header;
 692
 693         /* Some registers are reserved for use by the prolog/epilog */
 694         reserved_regs = header->num_clauses ? 4 : 3;
 695
 696         if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
 697                 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
 698                 /* One registers is needed by instrument_epilog to save the return value */
 699                 reserved_regs ++;
 700                 if (cinfo->reg_usage < 2)
 701                         /* Number of arguments passed to function call in instrument_prolog */
 702                         cinfo->reg_usage = 2;
 703         }
 704
 705         cfg->arch.reg_in0 = 32;
 706         cfg->arch.reg_local0 = cfg->arch.reg_in0 + cinfo->reg_usage + reserved_regs;
 707         cfg->arch.reg_out0 = cfg->arch.reg_local0 + 16;
 708
 709         cfg->arch.reg_saved_ar_pfs = cfg->arch.reg_local0 - 1;
 710         cfg->arch.reg_saved_b0 = cfg->arch.reg_local0 - 2;
 711         cfg->arch.reg_fp = cfg->arch.reg_local0 - 3;
 712
 713         /*
 714          * Frames without handlers save sp to fp, frames with handlers save it into
 715          * a dedicated register.
 716          */
 717         if (header->num_clauses)
 718                 cfg->arch.reg_saved_sp = cfg->arch.reg_local0 - 4;
 719         else
 720                 cfg->arch.reg_saved_sp = cfg->arch.reg_fp;
 721
 722         if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
 723                 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
 724                 cfg->arch.reg_saved_return_val = cfg->arch.reg_local0 - reserved_regs;
 725         }
 726
 727         /*
 728          * Need to allocate at least 2 out register for use by OP_THROW / the system
 729          * exception throwing code.
 730          */
 731         cfg->arch.n_out_regs = MAX (cfg->arch.n_out_regs, 2);
 732 }
 733
 734 GList *
 735 mono_arch_get_global_int_regs (MonoCompile *cfg)
 736 {
 737         GList *regs = NULL;
 738         int i;
 739
 740         mono_ia64_alloc_stacked_registers (cfg);
 741
 742         for (i = cfg->arch.reg_local0; i < cfg->arch.reg_out0; ++i) {
 743                 /* FIXME: regmask */
 744                 g_assert (i < 64);
 745                 regs = g_list_prepend (regs, (gpointer)(gssize)(i));
 746         }
 747
 748         return regs;
 749 }
 750
 751 /*
 752  * mono_arch_regalloc_cost:
 753  *
 754  *  Return the cost, in number of memory references, of the action of
 755  * allocating the variable VMV into a register during global register
 756  * allocation.
 757  */
 758 guint32
 759 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
 760 {
 761         /* FIXME: Increase costs linearly to avoid using all local registers */
 762
 763         return 0;
 764 }
 765
 766 void
 767 mono_arch_allocate_vars (MonoCompile *cfg)
 768 {
 769         MonoMethodSignature *sig;
 770         MonoMethodHeader *header;
 771         MonoInst *inst;
 772         int i, offset;
 773         guint32 locals_stack_size, locals_stack_align;
 774         gint32 *offsets;
 775         CallInfo *cinfo;
 776
 777         header = cfg->header;
 778
 779         sig = mono_method_signature (cfg->method);
 780
 781         cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
 782
 783         /*
 784          * Determine whenever the frame pointer can be eliminated.
 785          * FIXME: Remove some of the restrictions.
 786          */
 787         cfg->arch.omit_fp = TRUE;
 788
 789         if (!debug_omit_fp ())
 790                 cfg->arch.omit_fp = FALSE;
 791
 792         if (cfg->flags & MONO_CFG_HAS_ALLOCA)
 793                 cfg->arch.omit_fp = FALSE;
 794         if (header->num_clauses)
 795                 cfg->arch.omit_fp = FALSE;
 796         if (cfg->param_area)
 797                 cfg->arch.omit_fp = FALSE;
 798         if ((sig->ret->type != MONO_TYPE_VOID) && (cinfo->ret.storage == ArgAggregate))
 799                 cfg->arch.omit_fp = FALSE;
 800         if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
 801                 cfg->arch.omit_fp = FALSE;
 802         for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
 803                 ArgInfo *ainfo = &cinfo->args [i];
 804
 805                 if (ainfo->storage == ArgOnStack) {
 806                         /*
 807                          * The stack offset can only be determined when the frame
 808                          * size is known.
 809                          */
 810                         cfg->arch.omit_fp = FALSE;
 811                 }
 812         }
 813
 814         mono_ia64_alloc_stacked_registers (cfg);
 815
 816         /*
 817          * We use the ABI calling conventions for managed code as well.
 818          * Exception: valuetypes are never passed or returned in registers.
 819          */
 820
 821         if (cfg->arch.omit_fp) {
 822                 cfg->flags |= MONO_CFG_HAS_SPILLUP;
 823                 cfg->frame_reg = IA64_SP;
 824                 offset = ARGS_OFFSET;
 825         }
 826         else {
 827                 /* Locals are allocated backwards from %fp */
 828                 cfg->frame_reg = cfg->arch.reg_fp;
 829                 offset = 0;
 830         }
 831
 832         if (cfg->method->save_lmf) {
 833                 /* No LMF on IA64 */
 834         }
 835
 836         if (sig->ret->type != MONO_TYPE_VOID) {
 837                 switch (cinfo->ret.storage) {
 838                 case ArgInIReg:
 839                         cfg->ret->opcode = OP_REGVAR;
 840                         cfg->ret->inst_c0 = cinfo->ret.reg;
 841                         break;
 842                 case ArgInFloatReg:
 843                         cfg->ret->opcode = OP_REGVAR;
 844                         cfg->ret->inst_c0 = cinfo->ret.reg;
 845                         break;
 846                 case ArgValuetypeAddrInIReg:
 847                         cfg->vret_addr->opcode = OP_REGVAR;
 848                         cfg->vret_addr->dreg = cfg->arch.reg_in0 + cinfo->ret.reg;
 849                         break;
 850                 case ArgAggregate:
 851                         /* Allocate a local to hold the result, the epilog will copy it to the correct place */
 852                         if (cfg->arch.omit_fp)
 853                                 g_assert_not_reached ();
 854                         offset = ALIGN_TO (offset, 8);
 855                         offset += cinfo->ret.nslots * 8;
 856                         cfg->ret->opcode = OP_REGOFFSET;
 857                         cfg->ret->inst_basereg = cfg->frame_reg;
 858                         cfg->ret->inst_offset = - offset;
 859                         break;
 860                 default:
 861                         g_assert_not_reached ();
 862                 }
 863                 cfg->ret->dreg = cfg->ret->inst_c0;
 864         }
 865
 866         /* Allocate locals */
 867         offsets = mono_allocate_stack_slots (cfg, cfg->arch.omit_fp ? FALSE : TRUE, &locals_stack_size, &locals_stack_align);
 868         if (locals_stack_align) {
 869                 offset = ALIGN_TO (offset, locals_stack_align);
 870         }
 871         for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
 872                 if (offsets [i] != -1) {
 873                         MonoInst *inst = cfg->varinfo [i];
 874                         inst->opcode = OP_REGOFFSET;
 875                         inst->inst_basereg = cfg->frame_reg;
 876                         if (cfg->arch.omit_fp)
 877                                 inst->inst_offset = (offset + offsets [i]);
 878                         else
 879                                 inst->inst_offset = - (offset + offsets [i]);
 880                         // printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
 881                 }
 882         }
 883         offset += locals_stack_size;
 884
 885         if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
 886                 if (cfg->arch.omit_fp)
 887                         g_assert_not_reached ();
 888                 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
 889                 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
 890         }
 891
 892         for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
 893                 inst = cfg->args [i];
 894                 if (inst->opcode != OP_REGVAR) {
 895                         ArgInfo *ainfo = &cinfo->args [i];
 896                         gboolean inreg = TRUE;
 897                         MonoType *arg_type;
 898
 899                         if (sig->hasthis && (i == 0))
 900                                 arg_type = &mono_defaults.object_class->byval_arg;
 901                         else
 902                                 arg_type = sig->params [i - sig->hasthis];
 903
 904                         /* FIXME: VOLATILE is only set if the liveness pass runs */
 905                         if (inst->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
 906                                 inreg = FALSE;
 907
 908                         inst->opcode = OP_REGOFFSET;
 909
 910                         switch (ainfo->storage) {
 911                         case ArgInIReg:
 912                                 inst->opcode = OP_REGVAR;
 913                                 inst->dreg = cfg->arch.reg_in0 + ainfo->reg;
 914                                 break;
 915                         case ArgInFloatReg:
 916                         case ArgInFloatRegR4:
 917                                 /*
 918                                  * Since float regs are volatile, we save the arguments to
 919                                  * the stack in the prolog.
 920                                  */
 921                                 inreg = FALSE;
 922                                 break;
 923                         case ArgOnStack:
 924                                 if (cfg->arch.omit_fp)
 925                                         g_assert_not_reached ();
 926                                 inst->opcode = OP_REGOFFSET;
 927                                 inst->inst_basereg = cfg->frame_reg;
 928                                 inst->inst_offset = ARGS_OFFSET + ainfo->offset;
 929                                 break;
 930                         case ArgAggregate:
 931                                 inreg = FALSE;
 932                                 break;
 933                         default:
 934                                 NOT_IMPLEMENTED;
 935                         }
 936
 937                         if (!inreg && (ainfo->storage != ArgOnStack)) {
 938                                 guint32 size = 0;
 939
 940                                 inst->opcode = OP_REGOFFSET;
 941                                 inst->inst_basereg = cfg->frame_reg;
 942                                 /* These arguments are saved to the stack in the prolog */
 943                                 switch (ainfo->storage) {
 944                                 case ArgAggregate:
 945                                         if (ainfo->atype == AggregateSingleHFA)
 946                                                 size = ainfo->nslots * 4;
 947                                         else
 948                                                 size = ainfo->nslots * 8;
 949                                         break;
 950                                 default:
 951                                         size = sizeof (gpointer);
 952                                         break;
 953                                 }
 954
 955                                 offset = ALIGN_TO (offset, sizeof (gpointer));
 956
 957                                 if (cfg->arch.omit_fp) {
 958                                         inst->inst_offset = offset;
 959                                         offset += size;
 960                                 } else {
 961                                         offset += size;
 962                                         inst->inst_offset = - offset;
 963                                 }
 964                         }
 965                 }
 966         }
 967
 968         /*
 969          * FIXME: This doesn't work because some variables are allocated during local
 970          * regalloc.
 971          */
 972         /*
 973         if (cfg->arch.omit_fp && offset == 16)
 974                 offset = 0;
 975         */
 976
 977         cfg->stack_offset = offset;
 978 }
 979
 980 void
 981 mono_arch_create_vars (MonoCompile *cfg)
 982 {
 983         MonoMethodSignature *sig;
 984         CallInfo *cinfo;
 985
 986         sig = mono_method_signature (cfg->method);
 987
 988         cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
 989
 990         if (cinfo->ret.storage == ArgAggregate)
 991                 cfg->ret_var_is_local = TRUE;
 992         if (cinfo->ret.storage == ArgValuetypeAddrInIReg) {
 993                 cfg->vret_addr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_ARG);
 994                 if (G_UNLIKELY (cfg->verbose_level > 1)) {
 995                         printf ("vret_addr = ");
 996                         mono_print_ins (cfg->vret_addr);
 997                 }
 998         }
 999 }
1000
1001 static void
1002 add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, ArgStorage storage, int reg, MonoInst *tree)
1003 {
1004         MonoInst *arg;
1005
1006         MONO_INST_NEW (cfg, arg, OP_NOP);
1007         arg->sreg1 = tree->dreg;
1008
1009         switch (storage) {
1010         case ArgInIReg:
1011                 arg->opcode = OP_MOVE;
1012                 arg->dreg = mono_alloc_ireg (cfg);
1013
1014                 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, FALSE);
1015                 break;
1016         case ArgInFloatReg:
1017                 arg->opcode = OP_FMOVE;
1018                 arg->dreg = mono_alloc_freg (cfg);
1019
1020                 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, TRUE);
1021                 break;
1022         case ArgInFloatRegR4:
1023                 arg->opcode = OP_FCONV_TO_R4;
1024                 arg->dreg = mono_alloc_freg (cfg);
1025
1026                 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, TRUE);
1027                 break;
1028         default:
1029                 g_assert_not_reached ();
1030         }
1031
1032         MONO_ADD_INS (cfg->cbb, arg);
1033 }
1034
1035 static void
1036 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
1037 {
1038         MonoMethodSignature *tmp_sig;
1039
1040         /* Emit the signature cookie just before the implicit arguments */
1041         MonoInst *sig_arg;
1042         /* FIXME: Add support for signature tokens to AOT */
1043         cfg->disable_aot = TRUE;
1044
1045         g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1046
1047         /*
1048          * mono_ArgIterator_Setup assumes the signature cookie is
1049          * passed first and all the arguments which were before it are
1050          * passed on the stack after the signature. So compensate by
1051          * passing a different signature.
1052          */
1053         tmp_sig = mono_metadata_signature_dup (call->signature);
1054         tmp_sig->param_count -= call->signature->sentinelpos;
1055         tmp_sig->sentinelpos = 0;
1056         memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
1057
1058         MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
1059         sig_arg->dreg = mono_alloc_ireg (cfg);
1060         sig_arg->inst_p0 = tmp_sig;
1061         MONO_ADD_INS (cfg->cbb, sig_arg);
1062
1063         MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI8_MEMBASE_REG, IA64_SP, 16 + cinfo->sig_cookie.offset, sig_arg->dreg);
1064 }
1065
1066 void
1067 mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
1068 {
1069         MonoInst *in;
1070         MonoMethodSignature *sig;
1071         int i, n, stack_size;
1072         CallInfo *cinfo;
1073         ArgInfo *ainfo;
1074
1075         stack_size = 0;
1076
1077         mono_ia64_alloc_stacked_registers (cfg);
1078
1079         sig = call->signature;
1080         n = sig->param_count + sig->hasthis;
1081
1082         cinfo = get_call_info (cfg, cfg->mempool, sig, sig->pinvoke);
1083
1084         if (cinfo->ret.storage == ArgAggregate) {
1085                 MonoInst *vtarg;
1086                 MonoInst *local;
1087
1088                 /*
1089                  * The valuetype is in registers after the call, need to be copied
1090                  * to the stack. Save the address to a local here, so the call
1091                  * instruction can access it.
1092                  */
1093                 local = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
1094                 local->flags |= MONO_INST_VOLATILE;
1095                 cfg->arch.ret_var_addr_local = local;
1096
1097                 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
1098                 vtarg->sreg1 = call->vret_var->dreg;
1099                 vtarg->dreg = local->dreg;
1100                 MONO_ADD_INS (cfg->cbb, vtarg);
1101         }
1102
1103         if (cinfo->ret.storage == ArgValuetypeAddrInIReg) {
1104                 add_outarg_reg (cfg, call, ArgInIReg, cfg->arch.reg_out0 + cinfo->ret.reg, call->vret_var);
1105         }
1106
1107         for (i = 0; i < n; ++i) {
1108                 MonoType *arg_type;
1109
1110                 ainfo = cinfo->args + i;
1111
1112                 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
1113                         /* Emit the signature cookie just before the implicit arguments */
1114                         emit_sig_cookie (cfg, call, cinfo);
1115                 }
1116
1117                 in = call->args [i];
1118
1119                 if (sig->hasthis && (i == 0))
1120                         arg_type = &mono_defaults.object_class->byval_arg;
1121                 else
1122                         arg_type = sig->params [i - sig->hasthis];
1123
1124                 if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(arg_type))) {
1125                         guint32 align;
1126                         guint32 size;
1127
1128                         if (arg_type->type == MONO_TYPE_TYPEDBYREF) {
1129                                 size = sizeof (MonoTypedRef);
1130                                 align = sizeof (gpointer);
1131                         }
1132                         else if (sig->pinvoke)
1133                                 size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
1134                         else {
1135                                 /*
1136                                  * Other backends use mono_type_stack_size (), but that
1137                                  * aligns the size to 8, which is larger than the size of
1138                                  * the source, leading to reads of invalid memory if the
1139                                  * source is at the end of address space.
1140                                  */
1141                                 size = mono_class_value_size (in->klass, &align);
1142                         }
1143
1144                         if (size > 0) {
1145                                 MonoInst *arg;
1146
1147                                 MONO_INST_NEW (cfg, arg, OP_OUTARG_VT);
1148                                 arg->sreg1 = in->dreg;
1149                                 arg->klass = in->klass;
1150                                 arg->backend.size = size;
1151                                 arg->inst_p0 = call;
1152                                 arg->inst_p1 = mono_mempool_alloc (cfg->mempool, sizeof (ArgInfo));
1153                                 memcpy (arg->inst_p1, ainfo, sizeof (ArgInfo));
1154
1155                                 MONO_ADD_INS (cfg->cbb, arg);
1156                         }
1157                 }
1158                 else {
1159                         switch (ainfo->storage) {
1160                         case ArgInIReg:
1161                                 add_outarg_reg (cfg, call, ainfo->storage, cfg->arch.reg_out0 + ainfo->reg, in);
1162                                 break;
1163                         case ArgInFloatReg:
1164                         case ArgInFloatRegR4:
1165                                 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
1166                                 break;
1167                         case ArgOnStack:
1168                                 if (arg_type->type == MONO_TYPE_R4 && !arg_type->byref)
1169                                         MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, IA64_SP, 16 + ainfo->offset, in->dreg);
1170                                 else if (arg_type->type == MONO_TYPE_R8 && !arg_type->byref)
1171                                         MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, IA64_SP, 16 + ainfo->offset, in->dreg);
1172                                 else
1173                                         MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI8_MEMBASE_REG, IA64_SP, 16 + ainfo->offset, in->dreg);
1174                                 break;
1175                         default:
1176                                 g_assert_not_reached ();
1177                         }
1178                 }
1179         }
1180
1181         /* Handle the case where there are no implicit arguments */
1182         if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos)) {
1183                 emit_sig_cookie (cfg, call, cinfo);
1184         }
1185
1186         call->stack_usage = cinfo->stack_usage;
1187         cfg->arch.n_out_regs = MAX (cfg->arch.n_out_regs, cinfo->reg_usage);
1188 }
1189
1190 void
1191 mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
1192 {
1193         MonoCallInst *call = (MonoCallInst*)ins->inst_p0;
1194         ArgInfo *ainfo = (ArgInfo*)ins->inst_p1;
1195         int size = ins->backend.size;
1196
1197         if (ainfo->storage == ArgAggregate) {
1198                 MonoInst *load, *store;
1199                 int i, slot;
1200
1201                 /*
1202                  * Part of the structure is passed in registers.
1203                  */
1204                 for (i = 0; i < ainfo->nregs; ++i) {
1205                         slot = ainfo->reg + i;
1206
1207                         if (ainfo->atype == AggregateSingleHFA) {
1208                                 MONO_INST_NEW (cfg, load, OP_LOADR4_MEMBASE);
1209                                 load->inst_basereg = src->dreg;
1210                                 load->inst_offset = i * 4;
1211                                 load->dreg = mono_alloc_freg (cfg);
1212
1213                                 mono_call_inst_add_outarg_reg (cfg, call, load->dreg, ainfo->reg + i, TRUE);
1214                         } else if (ainfo->atype == AggregateDoubleHFA) {
1215                                 MONO_INST_NEW (cfg, load, OP_LOADR8_MEMBASE);
1216                                 load->inst_basereg = src->dreg;
1217                                 load->inst_offset = i * 8;
1218                                 load->dreg = mono_alloc_freg (cfg);
1219
1220                                 mono_call_inst_add_outarg_reg (cfg, call, load->dreg, ainfo->reg + i, TRUE);
1221                         } else {
1222                                 MONO_INST_NEW (cfg, load, OP_LOADI8_MEMBASE);
1223                                 load->inst_basereg = src->dreg;
1224                                 load->inst_offset = i * 8;
1225                                 load->dreg = mono_alloc_ireg (cfg);
1226
1227                                 mono_call_inst_add_outarg_reg (cfg, call, load->dreg, cfg->arch.reg_out0 + ainfo->reg + i, FALSE);
1228                         }
1229                         MONO_ADD_INS (cfg->cbb, load);
1230                 }
1231
1232                 /*
1233                  * Part of the structure is passed on the stack.
1234                  */
1235                 for (i = ainfo->nregs; i < ainfo->nslots; ++i) {
1236                         slot = ainfo->reg + i;
1237
1238                         MONO_INST_NEW (cfg, load, OP_LOADI8_MEMBASE);
1239                         load->inst_basereg = src->dreg;
1240                         load->inst_offset = i * sizeof (gpointer);
1241                         load->dreg = mono_alloc_preg (cfg);
1242                         MONO_ADD_INS (cfg->cbb, load);
1243
1244                         MONO_INST_NEW (cfg, store, OP_STOREI8_MEMBASE_REG);
1245                         store->sreg1 = load->dreg;
1246                         store->inst_destbasereg = IA64_SP;
1247                         store->inst_offset = 16 + ainfo->offset + (slot - 8) * 8;
1248                         MONO_ADD_INS (cfg->cbb, store);
1249                 }
1250         } else {
1251                 mini_emit_memcpy (cfg, IA64_SP, 16 + ainfo->offset, src->dreg, 0, size, 4);
1252         }
1253 }
1254
1255 void
1256 mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
1257 {
1258         CallInfo *cinfo = get_call_info (cfg, cfg->mempool, mono_method_signature (method), FALSE);
1259
1260         switch (cinfo->ret.storage) {
1261         case ArgInIReg:
1262                 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg);
1263                 break;
1264         case ArgInFloatReg:
1265                 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
1266                 break;
1267         default:
1268                 g_assert_not_reached ();
1269         }
1270 }
1271
1272 void
1273 mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
1274 {
1275 }
1276
1277 void
1278 mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
1279 {
1280         MonoInst *ins, *n, *last_ins = NULL;
1281         ins = bb->code;
1282
1283         MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
1284                 switch (ins->opcode) {
1285                 case OP_MOVE:
1286                 case OP_FMOVE:
1287                         /*
1288                          * Removes:
1289                          *
1290                          * OP_MOVE reg, reg
1291                          */
1292                         if (ins->dreg == ins->sreg1) {
1293                                 MONO_DELETE_INS (bb, ins);
1294                                 continue;
1295                         }
1296                         /*
1297                          * Removes:
1298                          *
1299                          * OP_MOVE sreg, dreg
1300                          * OP_MOVE dreg, sreg
1301                          */
1302                         if (last_ins && last_ins->opcode == OP_MOVE &&
1303                             ins->sreg1 == last_ins->dreg &&
1304                             ins->dreg == last_ins->sreg1) {
1305                                 MONO_DELETE_INS (bb, ins);
1306                                 continue;
1307                         }
1308                         break;
1309                 case OP_MUL_IMM:
1310                 case OP_IMUL_IMM:
1311                         /* remove unnecessary multiplication with 1 */
1312                         if (ins->inst_imm == 1) {
1313                                 if (ins->dreg != ins->sreg1) {
1314                                         ins->opcode = OP_MOVE;
1315                                 } else {
1316                                         MONO_DELETE_INS (bb, ins);
1317                                         continue;
1318                                 }
1319                         }
1320                         break;
1321                 }
1322
1323                 last_ins = ins;
1324                 ins = ins->next;
1325         }
1326         bb->last_ins = last_ins;
1327 }
1328
1329 int cond_to_ia64_cmp [][3] = {
1330         {OP_IA64_CMP_EQ, OP_IA64_CMP4_EQ, OP_IA64_FCMP_EQ},
1331         {OP_IA64_CMP_NE, OP_IA64_CMP4_NE, OP_IA64_FCMP_NE},
1332         {OP_IA64_CMP_LE, OP_IA64_CMP4_LE, OP_IA64_FCMP_LE},
1333         {OP_IA64_CMP_GE, OP_IA64_CMP4_GE, OP_IA64_FCMP_GE},
1334         {OP_IA64_CMP_LT, OP_IA64_CMP4_LT, OP_IA64_FCMP_LT},
1335         {OP_IA64_CMP_GT, OP_IA64_CMP4_GT, OP_IA64_FCMP_GT},
1336         {OP_IA64_CMP_LE_UN, OP_IA64_CMP4_LE_UN, OP_IA64_FCMP_LE_UN},
1337         {OP_IA64_CMP_GE_UN, OP_IA64_CMP4_GE_UN, OP_IA64_FCMP_GE_UN},
1338         {OP_IA64_CMP_LT_UN, OP_IA64_CMP4_LT_UN, OP_IA64_FCMP_LT_UN},
1339         {OP_IA64_CMP_GT_UN, OP_IA64_CMP4_GT_UN, OP_IA64_FCMP_GT_UN}
1340 };
1341
1342 static int
1343 opcode_to_ia64_cmp (int opcode, int cmp_opcode)
1344 {
1345         return cond_to_ia64_cmp [mono_opcode_to_cond (opcode)][mono_opcode_to_type (opcode, cmp_opcode)];
1346 }
1347
1348 int cond_to_ia64_cmp_imm [][3] = {
1349         {OP_IA64_CMP_EQ_IMM, OP_IA64_CMP4_EQ_IMM, 0},
1350         {OP_IA64_CMP_NE_IMM, OP_IA64_CMP4_NE_IMM, 0},
1351         {OP_IA64_CMP_GE_IMM, OP_IA64_CMP4_GE_IMM, 0},
1352         {OP_IA64_CMP_LE_IMM, OP_IA64_CMP4_LE_IMM, 0},
1353         {OP_IA64_CMP_GT_IMM, OP_IA64_CMP4_GT_IMM, 0},
1354         {OP_IA64_CMP_LT_IMM, OP_IA64_CMP4_LT_IMM, 0},
1355         {OP_IA64_CMP_GE_UN_IMM, OP_IA64_CMP4_GE_UN_IMM, 0},
1356         {OP_IA64_CMP_LE_UN_IMM, OP_IA64_CMP4_LE_UN_IMM, 0},
1357         {OP_IA64_CMP_GT_UN_IMM, OP_IA64_CMP4_GT_UN_IMM, 0},
1358         {OP_IA64_CMP_LT_UN_IMM, OP_IA64_CMP4_LT_UN_IMM, 0},
1359 };
1360
1361 static int
1362 opcode_to_ia64_cmp_imm (int opcode, int cmp_opcode)
1363 {
1364         /* The condition needs to be reversed */
1365         return cond_to_ia64_cmp_imm [mono_opcode_to_cond (opcode)][mono_opcode_to_type (opcode, cmp_opcode)];
1366 }
1367
1368 #define NEW_INS(cfg,dest,op) do {       \
1369                 (dest) = mono_mempool_alloc0 ((cfg)->mempool, sizeof (MonoInst));       \
1370                 (dest)->opcode = (op);  \
1371         mono_bblock_insert_after_ins (bb, last_ins, (dest)); \
1372         last_ins = (dest); \
1373         } while (0)
1374
1375 /*
1376  * mono_arch_lowering_pass:
1377  *
1378  *  Converts complex opcodes into simpler ones so that each IR instruction
1379  * corresponds to one machine instruction.
1380  */
1381 void
1382 mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
1383 {
1384         MonoInst *ins, *n, *next, *temp, *temp2, *temp3, *last_ins = NULL;
1385         ins = bb->code;
1386
1387         MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
1388                 switch (ins->opcode) {
1389                 case OP_STOREI1_MEMBASE_IMM:
1390                 case OP_STOREI2_MEMBASE_IMM:
1391                 case OP_STOREI4_MEMBASE_IMM:
1392                 case OP_STOREI8_MEMBASE_IMM:
1393                 case OP_STORE_MEMBASE_IMM:
1394                         /* There are no store_membase instructions on ia64 */
1395                         if (ins->inst_offset == 0) {
1396                                 temp2 = NULL;
1397                         } else if (ia64_is_imm14 (ins->inst_offset)) {
1398                                 NEW_INS (cfg, temp2, OP_ADD_IMM);
1399                                 temp2->sreg1 = ins->inst_destbasereg;
1400                                 temp2->inst_imm = ins->inst_offset;
1401                                 temp2->dreg = mono_alloc_ireg (cfg);
1402                         }
1403                         else {
1404                                 NEW_INS (cfg, temp, OP_I8CONST);
1405                                 temp->inst_c0 = ins->inst_offset;
1406                                 temp->dreg = mono_alloc_ireg (cfg);
1407
1408                                 NEW_INS (cfg, temp2, OP_LADD);
1409                                 temp2->sreg1 = ins->inst_destbasereg;
1410                                 temp2->sreg2 = temp->dreg;
1411                                 temp2->dreg = mono_alloc_ireg (cfg);
1412                         }
1413
1414                         switch (ins->opcode) {
1415                         case OP_STOREI1_MEMBASE_IMM:
1416                                 ins->opcode = OP_STOREI1_MEMBASE_REG;
1417                                 break;
1418                         case OP_STOREI2_MEMBASE_IMM:
1419                                 ins->opcode = OP_STOREI2_MEMBASE_REG;
1420                                 break;
1421                         case OP_STOREI4_MEMBASE_IMM:
1422                                 ins->opcode = OP_STOREI4_MEMBASE_REG;
1423                                 break;
1424                         case OP_STOREI8_MEMBASE_IMM:
1425                         case OP_STORE_MEMBASE_IMM:
1426                                 ins->opcode = OP_STOREI8_MEMBASE_REG;
1427                                 break;
1428                         default:
1429                                 g_assert_not_reached ();
1430                         }
1431
1432                         if (ins->inst_imm == 0)
1433                                 ins->sreg1 = IA64_R0;
1434                         else {
1435                                 NEW_INS (cfg, temp3, OP_I8CONST);
1436                                 temp3->inst_c0 = ins->inst_imm;
1437                                 temp3->dreg = mono_alloc_ireg (cfg);
1438                                 ins->sreg1 = temp3->dreg;
1439                         }
1440
1441                         ins->inst_offset = 0;
1442                         if (temp2)
1443                                 ins->inst_destbasereg = temp2->dreg;
1444                         break;
1445                 case OP_STOREI1_MEMBASE_REG:
1446                 case OP_STOREI2_MEMBASE_REG:
1447                 case OP_STOREI4_MEMBASE_REG:
1448                 case OP_STOREI8_MEMBASE_REG:
1449                 case OP_STORER4_MEMBASE_REG:
1450                 case OP_STORER8_MEMBASE_REG:
1451                 case OP_STORE_MEMBASE_REG:
1452                         /* There are no store_membase instructions on ia64 */
1453                         if (ins->inst_offset == 0) {
1454                                 break;
1455                         }
1456                         else if (ia64_is_imm14 (ins->inst_offset)) {
1457                                 NEW_INS (cfg, temp2, OP_ADD_IMM);
1458                                 temp2->sreg1 = ins->inst_destbasereg;
1459                                 temp2->inst_imm = ins->inst_offset;
1460                                 temp2->dreg = mono_alloc_ireg (cfg);
1461                         }
1462                         else {
1463                                 NEW_INS (cfg, temp, OP_I8CONST);
1464                                 temp->inst_c0 = ins->inst_offset;
1465                                 temp->dreg = mono_alloc_ireg (cfg);
1466                                 NEW_INS (cfg, temp2, OP_LADD);
1467                                 temp2->sreg1 = ins->inst_destbasereg;
1468                                 temp2->sreg2 = temp->dreg;
1469                                 temp2->dreg = mono_alloc_ireg (cfg);
1470                         }
1471
1472                         ins->inst_offset = 0;
1473                         ins->inst_destbasereg = temp2->dreg;
1474                         break;
1475                 case OP_LOADI1_MEMBASE:
1476                 case OP_LOADU1_MEMBASE:
1477                 case OP_LOADI2_MEMBASE:
1478                 case OP_LOADU2_MEMBASE:
1479                 case OP_LOADI4_MEMBASE:
1480                 case OP_LOADU4_MEMBASE:
1481                 case OP_LOADI8_MEMBASE:
1482                 case OP_LOAD_MEMBASE:
1483                 case OP_LOADR4_MEMBASE:
1484                 case OP_LOADR8_MEMBASE:
1485                 case OP_ATOMIC_EXCHANGE_I4:
1486                 case OP_ATOMIC_EXCHANGE_I8:
1487                 case OP_ATOMIC_ADD_I4:
1488                 case OP_ATOMIC_ADD_I8:
1489                 case OP_ATOMIC_ADD_IMM_I4:
1490                 case OP_ATOMIC_ADD_IMM_I8:
1491                         /* There are no membase instructions on ia64 */
1492                         if (ins->inst_offset == 0) {
1493                                 break;
1494                         }
1495                         else if (ia64_is_imm14 (ins->inst_offset)) {
1496                                 NEW_INS (cfg, temp2, OP_ADD_IMM);
1497                                 temp2->sreg1 = ins->inst_basereg;
1498                                 temp2->inst_imm = ins->inst_offset;
1499                                 temp2->dreg = mono_alloc_ireg (cfg);
1500                         }
1501                         else {
1502                                 NEW_INS (cfg, temp, OP_I8CONST);
1503                                 temp->inst_c0 = ins->inst_offset;
1504                                 temp->dreg = mono_alloc_ireg (cfg);
1505                                 NEW_INS (cfg, temp2, OP_LADD);
1506                                 temp2->sreg1 = ins->inst_basereg;
1507                                 temp2->sreg2 = temp->dreg;
1508                                 temp2->dreg = mono_alloc_ireg (cfg);
1509                         }
1510
1511                         ins->inst_offset = 0;
1512                         ins->inst_basereg = temp2->dreg;
1513                         break;
1514                 case OP_ADD_IMM:
1515                 case OP_IADD_IMM:
1516                 case OP_LADD_IMM:
1517                 case OP_ISUB_IMM:
1518                 case OP_LSUB_IMM:
1519                 case OP_AND_IMM:
1520                 case OP_IAND_IMM:
1521                 case OP_LAND_IMM:
1522                 case OP_IOR_IMM:
1523                 case OP_LOR_IMM:
1524                 case OP_IXOR_IMM:
1525                 case OP_LXOR_IMM:
1526                 case OP_SHL_IMM:
1527                 case OP_SHR_IMM:
1528                 case OP_ISHL_IMM:
1529                 case OP_LSHL_IMM:
1530                 case OP_ISHR_IMM:
1531                 case OP_LSHR_IMM:
1532                 case OP_ISHR_UN_IMM:
1533                 case OP_LSHR_UN_IMM: {
1534                         gboolean is_imm = FALSE;
1535                         gboolean switched = FALSE;
1536
1537                         if (ins->opcode == OP_AND_IMM && ins->inst_imm == 255) {
1538                                 ins->opcode = OP_ZEXT_I1;
1539                                 break;
1540                         }
1541
1542                         switch (ins->opcode) {
1543                         case OP_ADD_IMM:
1544                         case OP_IADD_IMM:
1545                         case OP_LADD_IMM:
1546                                 is_imm = ia64_is_imm14 (ins->inst_imm);
1547                                 switched = TRUE;
1548                                 break;
1549                         case OP_ISUB_IMM:
1550                         case OP_LSUB_IMM:
1551                                 is_imm = ia64_is_imm14 (- (ins->inst_imm));
1552                                 if (is_imm) {
1553                                         /* A = B - IMM -> A = B + (-IMM) */
1554                                         ins->inst_imm = - ins->inst_imm;
1555                                         ins->opcode = OP_IADD_IMM;
1556                                 }
1557                                 switched = TRUE;
1558                                 break;
1559                         case OP_IAND_IMM:
1560                         case OP_IOR_IMM:
1561                         case OP_IXOR_IMM:
1562                         case OP_AND_IMM:
1563                         case OP_LAND_IMM:
1564                         case OP_LOR_IMM:
1565                         case OP_LXOR_IMM:
1566                                 is_imm = ia64_is_imm8 (ins->inst_imm);
1567                                 switched = TRUE;
1568                                 break;
1569                         case OP_SHL_IMM:
1570                         case OP_SHR_IMM:
1571                         case OP_ISHL_IMM:
1572                         case OP_LSHL_IMM:
1573                         case OP_ISHR_IMM:
1574                         case OP_LSHR_IMM:
1575                         case OP_ISHR_UN_IMM:
1576                         case OP_LSHR_UN_IMM:
1577                                 is_imm = (ins->inst_imm >= 0) && (ins->inst_imm < 64);
1578                                 break;
1579                         default:
1580                                 break;
1581                         }
1582
1583                         if (is_imm) {
1584                                 if (switched)
1585                                         ins->sreg2 = ins->sreg1;
1586                                 break;
1587                         }
1588
1589                         ins->opcode = mono_op_imm_to_op (ins->opcode);
1590
1591                         if (ins->inst_imm == 0)
1592                                 ins->sreg2 = IA64_R0;
1593                         else {
1594                                 NEW_INS (cfg, temp, OP_I8CONST);
1595                                 temp->inst_c0 = ins->inst_imm;
1596                                 temp->dreg = mono_alloc_ireg (cfg);
1597                                 ins->sreg2 = temp->dreg;
1598                         }
1599                         break;
1600                 }
1601                 case OP_COMPARE_IMM:
1602                 case OP_ICOMPARE_IMM:
1603                 case OP_LCOMPARE_IMM: {
1604                         /* Instead of compare+b<cond>, ia64 has compare<cond>+br */
1605                         gboolean imm;
1606                         CompRelation cond;
1607
1608                         next = ins->next;
1609
1610                         /* Branch opts can eliminate the branch */
1611                         if (!next || (!(MONO_IS_COND_BRANCH_OP (next) || MONO_IS_COND_EXC (next) || MONO_IS_SETCC (next)))) {
1612                                 NULLIFY_INS (ins);
1613                                 break;
1614                         }
1615
1616                         /*
1617                          * The compare_imm instructions have switched up arguments, and
1618                          * some of them take an imm between -127 and 128.
1619                          */
1620                         next = ins->next;
1621                         cond = mono_opcode_to_cond (next->opcode);
1622                         if ((cond == CMP_LT) || (cond == CMP_GE))
1623                                 imm = ia64_is_imm8 (ins->inst_imm - 1);
1624                         else if ((cond == CMP_LT_UN) || (cond == CMP_GE_UN))
1625                                 imm = ia64_is_imm8 (ins->inst_imm - 1) && (ins->inst_imm > 0);
1626                         else
1627                                 imm = ia64_is_imm8 (ins->inst_imm);
1628
1629                         if (imm) {
1630                                 ins->opcode = opcode_to_ia64_cmp_imm (next->opcode, ins->opcode);
1631                                 ins->sreg2 = ins->sreg1;
1632                         }
1633                         else {
1634                                 ins->opcode = opcode_to_ia64_cmp (next->opcode, ins->opcode);
1635
1636                                 if (ins->inst_imm == 0)
1637                                         ins->sreg2 = IA64_R0;
1638                                 else {
1639                                         NEW_INS (cfg, temp, OP_I8CONST);
1640                                         temp->inst_c0 = ins->inst_imm;
1641                                         temp->dreg = mono_alloc_ireg (cfg);
1642                                         ins->sreg2 = temp->dreg;
1643                                 }
1644                         }
1645
1646                         if (MONO_IS_COND_BRANCH_OP (next)) {
1647                                 next->opcode = OP_IA64_BR_COND;
1648                                 next->inst_target_bb = next->inst_true_bb;
1649                         } else if (MONO_IS_COND_EXC (next)) {
1650                                 next->opcode = OP_IA64_COND_EXC;
1651                         } else if (MONO_IS_SETCC (next)) {
1652                                 next->opcode = OP_IA64_CSET;
1653                         } else {
1654                                 printf ("%s\n", mono_inst_name (next->opcode));
1655                                 NOT_IMPLEMENTED;
1656                         }
1657
1658                         break;
1659                 }
1660                 case OP_COMPARE:
1661                 case OP_ICOMPARE:
1662                 case OP_LCOMPARE:
1663                 case OP_FCOMPARE: {
1664                         /* Instead of compare+b<cond>, ia64 has compare<cond>+br */
1665
1666                         next = ins->next;
1667
1668                         /* Branch opts can eliminate the branch */
1669                         if (!next || (!(MONO_IS_COND_BRANCH_OP (next) || MONO_IS_COND_EXC (next) || MONO_IS_SETCC (next)))) {
1670                                 NULLIFY_INS (ins);
1671                                 break;
1672                         }
1673
1674                         ins->opcode = opcode_to_ia64_cmp (next->opcode, ins->opcode);
1675
1676                         if (MONO_IS_COND_BRANCH_OP (next)) {
1677                                 next->opcode = OP_IA64_BR_COND;
1678                                 next->inst_target_bb = next->inst_true_bb;
1679                         } else if (MONO_IS_COND_EXC (next)) {
1680                                 next->opcode = OP_IA64_COND_EXC;
1681                         } else if (MONO_IS_SETCC (next)) {
1682                                 next->opcode = OP_IA64_CSET;
1683                         } else {
1684                                 printf ("%s\n", mono_inst_name (next->opcode));
1685                                 NOT_IMPLEMENTED;
1686                         }
1687
1688                         break;
1689                 }
1690                 case OP_FCEQ:
1691                 case OP_FCGT:
1692                 case OP_FCGT_UN:
1693                 case OP_FCLT:
1694                 case OP_FCLT_UN:
1695                         /* The front end removes the fcompare, so introduce it again */
1696                         NEW_INS (cfg, temp, opcode_to_ia64_cmp (ins->opcode, OP_FCOMPARE));
1697                         temp->sreg1 = ins->sreg1;
1698                         temp->sreg2 = ins->sreg2;
1699
1700                         ins->opcode = OP_IA64_CSET;
1701                         MONO_INST_NULLIFY_SREGS (ins);
1702                         break;
1703                 case OP_MUL_IMM:
1704                 case OP_LMUL_IMM:
1705                 case OP_IMUL_IMM: {
1706                         int i, sum_reg;
1707                         gboolean found = FALSE;
1708                         int shl_op = ins->opcode == OP_IMUL_IMM ? OP_ISHL_IMM : OP_SHL_IMM;
1709
1710                         /* First the easy cases */
1711                         if (ins->inst_imm == 1) {
1712                                 ins->opcode = OP_MOVE;
1713                                 break;
1714                         }
1715                         for (i = 1; i < 64; ++i)
1716                                 if (ins->inst_imm == (((gint64)1) << i)) {
1717                                         ins->opcode = shl_op;
1718                                         ins->inst_imm = i;
1719                                         found = TRUE;
1720                                         break;
1721                                 }
1722
1723                         /* This could be optimized */
1724                         if (!found) {
1725                                 sum_reg = 0;
1726                                 for (i = 0; i < 64; ++i) {
1727                                         if (ins->inst_imm & (((gint64)1) << i)) {
1728                                                 NEW_INS (cfg, temp, shl_op);
1729                                                 temp->dreg = mono_alloc_ireg (cfg);
1730                                                 temp->sreg1 = ins->sreg1;
1731                                                 temp->inst_imm = i;
1732
1733                                                 if (sum_reg == 0)
1734                                                         sum_reg = temp->dreg;
1735                                                 else {
1736                                                         NEW_INS (cfg, temp2, OP_LADD);
1737                                                         temp2->dreg = mono_alloc_ireg (cfg);
1738                                                         temp2->sreg1 = sum_reg;
1739                                                         temp2->sreg2 = temp->dreg;
1740                                                         sum_reg = temp2->dreg;
1741                                                 }
1742                                         }
1743                                 }
1744                                 ins->opcode = OP_MOVE;
1745                                 ins->sreg1 = sum_reg;
1746                         }
1747                         break;
1748                 }
1749                 case OP_LCONV_TO_OVF_U4:
1750                         NEW_INS (cfg, temp, OP_IA64_CMP4_LT);
1751                         temp->sreg1 = ins->sreg1;
1752                         temp->sreg2 = IA64_R0;
1753
1754                         NEW_INS (cfg, temp, OP_IA64_COND_EXC);
1755                         temp->inst_p1 = (char*)"OverflowException";
1756
1757                         ins->opcode = OP_MOVE;
1758                         break;
1759                 case OP_LCONV_TO_OVF_I4_UN:
1760                         NEW_INS (cfg, temp, OP_ICONST);
1761                         temp->inst_c0 = 0x7fffffff;
1762                         temp->dreg = mono_alloc_ireg (cfg);
1763
1764                         NEW_INS (cfg, temp2, OP_IA64_CMP4_GT_UN);
1765                         temp2->sreg1 = ins->sreg1;
1766                         temp2->sreg2 = temp->dreg;
1767
1768                         NEW_INS (cfg, temp, OP_IA64_COND_EXC);
1769                         temp->inst_p1 = (char*)"OverflowException";
1770
1771                         ins->opcode = OP_MOVE;
1772                         break;
1773                 case OP_FCONV_TO_I4:
1774                 case OP_FCONV_TO_I2:
1775                 case OP_FCONV_TO_U2:
1776                 case OP_FCONV_TO_I1:
1777                 case OP_FCONV_TO_U1:
1778                         NEW_INS (cfg, temp, OP_FCONV_TO_I8);
1779                         temp->sreg1 = ins->sreg1;
1780                         temp->dreg = ins->dreg;
1781
1782                         switch (ins->opcode) {
1783                         case OP_FCONV_TO_I4:
1784                                 ins->opcode = OP_SEXT_I4;
1785                                 break;
1786                         case OP_FCONV_TO_I2:
1787                                 ins->opcode = OP_SEXT_I2;
1788                                 break;
1789                         case OP_FCONV_TO_U2:
1790                                 ins->opcode = OP_ZEXT_I4;
1791                                 break;
1792                         case OP_FCONV_TO_I1:
1793                                 ins->opcode = OP_SEXT_I1;
1794                                 break;
1795                         case OP_FCONV_TO_U1:
1796                                 ins->opcode = OP_ZEXT_I1;
1797                                 break;
1798                         default:
1799                                 g_assert_not_reached ();
1800                         }
1801                         ins->sreg1 = ins->dreg;
1802                         break;
1803                 default:
1804                         break;
1805                 }
1806                 last_ins = ins;
1807                 ins = ins->next;
1808         }
1809         bb->last_ins = last_ins;
1810
1811         bb->max_vreg = cfg->next_vreg;
1812 }
1813
1814 /*
1815  * emit_load_volatile_arguments:
1816  *
1817  *  Load volatile arguments from the stack to the original input registers.
1818  * Required before a tail call.
1819  */
1820 static Ia64CodegenState
1821 emit_load_volatile_arguments (MonoCompile *cfg, Ia64CodegenState code)
1822 {
1823         MonoMethod *method = cfg->method;
1824         MonoMethodSignature *sig;
1825         MonoInst *ins;
1826         CallInfo *cinfo;
1827         guint32 i;
1828
1829         /* FIXME: Generate intermediate code instead */
1830
1831         sig = mono_method_signature (method);
1832
1833         cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
1834
1835         /* This is the opposite of the code in emit_prolog */
1836         for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1837                 ArgInfo *ainfo = cinfo->args + i;
1838                 gint32 stack_offset;
1839                 MonoType *arg_type;
1840
1841                 ins = cfg->args [i];
1842
1843                 if (sig->hasthis && (i == 0))
1844                         arg_type = &mono_defaults.object_class->byval_arg;
1845                 else
1846                         arg_type = sig->params [i - sig->hasthis];
1847
1848                 arg_type = mini_get_underlying_type (arg_type);
1849
1850                 stack_offset = ainfo->offset + ARGS_OFFSET;
1851
1852                 /* Save volatile arguments to the stack */
1853                 if (ins->opcode != OP_REGVAR) {
1854                         switch (ainfo->storage) {
1855                         case ArgInIReg:
1856                         case ArgInFloatReg:
1857                                 /* FIXME: big offsets */
1858                                 g_assert (ins->opcode == OP_REGOFFSET);
1859                                 ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_basereg);
1860                                 if (arg_type->byref)
1861                                         ia64_ld8 (code, cfg->arch.reg_in0 + ainfo->reg, GP_SCRATCH_REG);
1862                                 else {
1863                                         switch (arg_type->type) {
1864                                         case MONO_TYPE_R4:
1865                                                 ia64_ldfs (code, ainfo->reg, GP_SCRATCH_REG);
1866                                                 break;
1867                                         case MONO_TYPE_R8:
1868                                                 ia64_ldfd (code, ainfo->reg, GP_SCRATCH_REG);
1869                                                 break;
1870                                         default:
1871                                                 ia64_ld8 (code, cfg->arch.reg_in0 + ainfo->reg, GP_SCRATCH_REG);
1872                                                 break;
1873                                         }
1874                                 }
1875                                 break;
1876                         case ArgOnStack:
1877                                 break;
1878                         default:
1879                                 NOT_IMPLEMENTED;
1880                         }
1881                 }
1882
1883                 if (ins->opcode == OP_REGVAR) {
1884                         /* Argument allocated to (non-volatile) register */
1885                         switch (ainfo->storage) {
1886                         case ArgInIReg:
1887                                 if (ins->dreg != cfg->arch.reg_in0 + ainfo->reg)
1888                                         ia64_mov (code, cfg->arch.reg_in0 + ainfo->reg, ins->dreg);
1889                                 break;
1890                         case ArgOnStack:
1891                                 ia64_adds_imm (code, GP_SCRATCH_REG, 16 + ainfo->offset, cfg->frame_reg);
1892                                 ia64_st8 (code, GP_SCRATCH_REG, ins->dreg);
1893                                 break;
1894                         default:
1895                                 NOT_IMPLEMENTED;
1896                         }
1897                 }
1898         }
1899
1900         return code;
1901 }
1902
1903 static Ia64CodegenState
1904 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, Ia64CodegenState code)
1905 {
1906         CallInfo *cinfo;
1907         int i;
1908
1909         /* Move return value to the target register */
1910         switch (ins->opcode) {
1911         case OP_VOIDCALL:
1912         case OP_VOIDCALL_REG:
1913         case OP_VOIDCALL_MEMBASE:
1914                 break;
1915         case OP_CALL:
1916         case OP_CALL_REG:
1917         case OP_CALL_MEMBASE:
1918         case OP_LCALL:
1919         case OP_LCALL_REG:
1920         case OP_LCALL_MEMBASE:
1921                 g_assert (ins->dreg == IA64_R8);
1922                 break;
1923         case OP_FCALL:
1924         case OP_FCALL_REG:
1925         case OP_FCALL_MEMBASE:
1926                 g_assert (ins->dreg == 8);
1927                 if (((MonoCallInst*)ins)->signature->ret->type == MONO_TYPE_R4)
1928                         ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
1929                 break;
1930         case OP_VCALL:
1931         case OP_VCALL_REG:
1932         case OP_VCALL_MEMBASE:
1933         case OP_VCALL2:
1934         case OP_VCALL2_REG:
1935         case OP_VCALL2_MEMBASE: {
1936                 ArgStorage storage;
1937
1938                 cinfo = get_call_info (cfg, cfg->mempool, ((MonoCallInst*)ins)->signature, FALSE);
1939                 storage = cinfo->ret.storage;
1940
1941                 if (storage == ArgAggregate) {
1942                         MonoInst *local = (MonoInst*)cfg->arch.ret_var_addr_local;
1943
1944                         /* Load address of stack space allocated for the return value */
1945                         ia64_movl (code, GP_SCRATCH_REG, local->inst_offset);
1946                         ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, local->inst_basereg);
1947                         ia64_ld8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG);
1948
1949                         for (i = 0; i < cinfo->ret.nregs; ++i) {
1950                                 switch (cinfo->ret.atype) {
1951                                 case AggregateNormal:
1952                                         ia64_st8_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 8, 0);
1953                                         break;
1954                                 case AggregateSingleHFA:
1955                                         ia64_stfs_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 4, 0);
1956                                         break;
1957                                 case AggregateDoubleHFA:
1958                                         ia64_stfd_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 8, 0);
1959                                         break;
1960                                 default:
1961                                         g_assert_not_reached ();
1962                                 }
1963                         }
1964                 }
1965                 break;
1966         }
1967         default:
1968                 g_assert_not_reached ();
1969         }
1970
1971         return code;
1972 }
1973
1974 #define add_patch_info(cfg,code,patch_type,data) do { \
1975         mono_add_patch_info (cfg, code.buf + code.nins - cfg->native_code, patch_type, data); \
1976 } while (0)
1977
1978 #define emit_cond_system_exception(cfg,code,exc_name,predicate) do { \
1979         MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
1980     if (tins == NULL) \
1981         add_patch_info (cfg, code, MONO_PATCH_INFO_EXC, exc_name); \
1982     else \
1983                 add_patch_info (cfg, code, MONO_PATCH_INFO_BB, tins->inst_true_bb); \
1984         ia64_br_cond_pred (code, (predicate), 0); \
1985 } while (0)
1986
1987 static Ia64CodegenState
1988 emit_call (MonoCompile *cfg, Ia64CodegenState code, guint32 patch_type, gconstpointer data)
1989 {
1990         add_patch_info (cfg, code, patch_type, data);
1991
1992         if ((patch_type == MONO_PATCH_INFO_ABS) || (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD)) {
1993                 /* Indirect call */
1994                 /* mono_arch_patch_callsite will patch this */
1995                 ia64_movl (code, GP_SCRATCH_REG, 0);
1996                 ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
1997                 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
1998                 ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
1999                 ia64_br_call_reg (code, IA64_B0, IA64_B6);
2000         }
2001         else {
2002                 /* Can't use a direct call since the displacement might be too small */
2003                 /* mono_arch_patch_callsite will patch this */
2004                 ia64_movl (code, GP_SCRATCH_REG, 0);
2005                 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
2006                 ia64_br_call_reg (code, IA64_B0, IA64_B6);
2007         }
2008
2009         return code;
2010 }
2011
2012 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
2013
2014 void
2015 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
2016 {
2017         MonoInst *ins;
2018         MonoCallInst *call;
2019         guint offset;
2020         Ia64CodegenState code;
2021         guint8 *code_start = cfg->native_code + cfg->code_len;
2022         MonoInst *last_ins = NULL;
2023         guint last_offset = 0;
2024         int max_len, cpos;
2025
2026         if (cfg->opt & MONO_OPT_LOOP) {
2027                 /* FIXME: */
2028         }
2029
2030         if (cfg->verbose_level > 2)
2031                 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
2032
2033         cpos = bb->max_offset;
2034
2035         if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
2036                 NOT_IMPLEMENTED;
2037         }
2038
2039         offset = code_start - cfg->native_code;
2040
2041         ia64_codegen_init (code, code_start);
2042
2043 #if 0
2044         if (strstr (cfg->method->name, "conv_ovf_i1") && (bb->block_num == 2))
2045                 break_count ();
2046 #endif
2047
2048         MONO_BB_FOR_EACH_INS (bb, ins) {
2049                 offset = code.buf - cfg->native_code;
2050
2051                 max_len = ((int)(((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN])) + 128;
2052
2053                 while (offset + max_len + 16 > cfg->code_size) {
2054                         ia64_codegen_close (code);
2055
2056                         offset = code.buf - cfg->native_code;
2057
2058                         cfg->code_size *= 2;
2059                         cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
2060                         code_start = cfg->native_code + offset;
2061                         cfg->stat_code_reallocs++;
2062
2063                         ia64_codegen_init (code, code_start);
2064                 }
2065
2066                 mono_debug_record_line_number (cfg, ins, offset);
2067
2068                 switch (ins->opcode) {
2069                 case OP_ICONST:
2070                 case OP_I8CONST:
2071                         if (ia64_is_imm14 (ins->inst_c0))
2072                                 ia64_adds_imm (code, ins->dreg, ins->inst_c0, IA64_R0);
2073                         else
2074                                 ia64_movl (code, ins->dreg, ins->inst_c0);
2075                         break;
2076                 case OP_JUMP_TABLE:
2077                         add_patch_info (cfg, code, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
2078                         ia64_movl (code, ins->dreg, 0);
2079                         break;
2080                 case OP_MOVE:
2081                         ia64_mov (code, ins->dreg, ins->sreg1);
2082                         break;
2083                 case OP_BR:
2084                 case OP_IA64_BR_COND: {
2085                         int pred = 0;
2086                         if (ins->opcode == OP_IA64_BR_COND)
2087                                 pred = 6;
2088                         if (ins->inst_target_bb->native_offset) {
2089                                 guint8 *pos = code.buf + code.nins;
2090
2091                                 ia64_br_cond_pred (code, pred, 0);
2092                                 ia64_begin_bundle (code);
2093                                 ia64_patch (pos, cfg->native_code + ins->inst_target_bb->native_offset);
2094                         } else {
2095                                 add_patch_info (cfg, code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
2096                                 ia64_br_cond_pred (code, pred, 0);
2097                         }
2098                         break;
2099                 }
2100                 case OP_LABEL:
2101                         ia64_begin_bundle (code);
2102                         ins->inst_c0 = code.buf - cfg->native_code;
2103                         break;
2104                 case OP_NOP:
2105                 case OP_RELAXED_NOP:
2106                 case OP_DUMMY_USE:
2107                 case OP_DUMMY_STORE:
2108                 case OP_NOT_REACHED:
2109                 case OP_NOT_NULL:
2110                         break;
2111                 case OP_BR_REG:
2112                         ia64_mov_to_br (code, IA64_B6, ins->sreg1);
2113                         ia64_br_cond_reg (code, IA64_B6);
2114                         break;
2115                 case OP_IADD:
2116                 case OP_LADD:
2117                         ia64_add (code, ins->dreg, ins->sreg1, ins->sreg2);
2118                         break;
2119                 case OP_ISUB:
2120                 case OP_LSUB:
2121                         ia64_sub (code, ins->dreg, ins->sreg1, ins->sreg2);
2122                         break;
2123                 case OP_IAND:
2124                 case OP_LAND:
2125                         ia64_and (code, ins->dreg, ins->sreg1, ins->sreg2);
2126                         break;
2127                 case OP_IOR:
2128                 case OP_LOR:
2129                         ia64_or (code, ins->dreg, ins->sreg1, ins->sreg2);
2130                         break;
2131                 case OP_IXOR:
2132                 case OP_LXOR:
2133                         ia64_xor (code, ins->dreg, ins->sreg1, ins->sreg2);
2134                         break;
2135                 case OP_INEG:
2136                 case OP_LNEG:
2137                         ia64_sub (code, ins->dreg, IA64_R0, ins->sreg1);
2138                         break;
2139                 case OP_INOT:
2140                 case OP_LNOT:
2141                         ia64_andcm_imm (code, ins->dreg, -1, ins->sreg1);
2142                         break;
2143                 case OP_ISHL:
2144                 case OP_LSHL:
2145                         ia64_shl (code, ins->dreg, ins->sreg1, ins->sreg2);
2146                         break;
2147                 case OP_ISHR:
2148                         ia64_sxt4 (code, GP_SCRATCH_REG, ins->sreg1);
2149                         ia64_shr (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
2150                         break;
2151                 case OP_LSHR:
2152                         ia64_shr (code, ins->dreg, ins->sreg1, ins->sreg2);
2153                         break;
2154                 case OP_ISHR_UN:
2155                         ia64_zxt4 (code, GP_SCRATCH_REG, ins->sreg1);
2156                         ia64_shr_u (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
2157                         break;
2158                 case OP_LSHR_UN:
2159                         ia64_shr_u (code, ins->dreg, ins->sreg1, ins->sreg2);
2160                         break;
2161                 case OP_IADDCC:
2162                         /* p6 and p7 is set if there is signed/unsigned overflow */
2163
2164                         /* Set p8-p9 == (sreg2 > 0) */
2165                         ia64_cmp4_lt (code, 8, 9, IA64_R0, ins->sreg2);
2166
2167                         ia64_add (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2168
2169                         /* (sreg2 > 0) && (res < ins->sreg1) => signed overflow */
2170                         ia64_cmp4_lt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2171                         /* (sreg2 <= 0) && (res > ins->sreg1) => signed overflow */
2172                         ia64_cmp4_lt_pred (code, 9, 6, 10, ins->sreg1, GP_SCRATCH_REG);
2173
2174                         /* res <u sreg1 => unsigned overflow */
2175                         ia64_cmp4_ltu (code, 7, 10, GP_SCRATCH_REG, ins->sreg1);
2176
2177                         /* FIXME: Predicate this since this is a side effect */
2178                         ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2179                         break;
2180                 case OP_ISUBCC:
2181                         /* p6 and p7 is set if there is signed/unsigned overflow */
2182
2183                         /* Set p8-p9 == (sreg2 > 0) */
2184                         ia64_cmp4_lt (code, 8, 9, IA64_R0, ins->sreg2);
2185
2186                         ia64_sub (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2187
2188                         /* (sreg2 > 0) && (res > ins->sreg1) => signed overflow */
2189                         ia64_cmp4_gt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2190                         /* (sreg2 <= 0) && (res < ins->sreg1) => signed overflow */
2191                         ia64_cmp4_lt_pred (code, 9, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2192
2193                         /* sreg1 <u sreg2 => unsigned overflow */
2194                         ia64_cmp4_ltu (code, 7, 10, ins->sreg1, ins->sreg2);
2195
2196                         /* FIXME: Predicate this since this is a side effect */
2197                         ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2198                         break;
2199                 case OP_ADDCC:
2200                         /* Same as OP_IADDCC */
2201                         ia64_cmp_lt (code, 8, 9, IA64_R0, ins->sreg2);
2202
2203                         ia64_add (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2204
2205                         ia64_cmp_lt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2206                         ia64_cmp_lt_pred (code, 9, 6, 10, ins->sreg1, GP_SCRATCH_REG);
2207
2208                         ia64_cmp_ltu (code, 7, 10, GP_SCRATCH_REG, ins->sreg1);
2209
2210                         ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2211                         break;
2212                 case OP_SUBCC:
2213                         /* Same as OP_ISUBCC */
2214
2215                         ia64_cmp_lt (code, 8, 9, IA64_R0, ins->sreg2);
2216
2217                         ia64_sub (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2218
2219                         ia64_cmp_gt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2220                         ia64_cmp_lt_pred (code, 9, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2221
2222                         ia64_cmp_ltu (code, 7, 10, ins->sreg1, ins->sreg2);
2223
2224                         ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2225                         break;
2226                 case OP_ADD_IMM:
2227                 case OP_IADD_IMM:
2228                 case OP_LADD_IMM:
2229                         ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2230                         break;
2231                 case OP_IAND_IMM:
2232                 case OP_AND_IMM:
2233                 case OP_LAND_IMM:
2234                         ia64_and_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2235                         break;
2236                 case OP_IOR_IMM:
2237                 case OP_LOR_IMM:
2238                         ia64_or_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2239                         break;
2240                 case OP_IXOR_IMM:
2241                 case OP_LXOR_IMM:
2242                         ia64_xor_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2243                         break;
2244                 case OP_SHL_IMM:
2245                 case OP_ISHL_IMM:
2246                 case OP_LSHL_IMM:
2247                         ia64_shl_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2248                         break;
2249                 case OP_SHR_IMM:
2250                 case OP_LSHR_IMM:
2251                         ia64_shr_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2252                         break;
2253                 case OP_ISHR_IMM:
2254                         g_assert (ins->inst_imm <= 64);
2255                         ia64_extr (code, ins->dreg, ins->sreg1, ins->inst_imm, 32 - ins->inst_imm);
2256                         break;
2257                 case OP_ISHR_UN_IMM:
2258                         ia64_zxt4 (code, GP_SCRATCH_REG, ins->sreg1);
2259                         ia64_shr_u_imm (code, ins->dreg, GP_SCRATCH_REG, ins->inst_imm);
2260                         break;
2261                 case OP_LSHR_UN_IMM:
2262                         ia64_shr_u_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2263                         break;
2264                 case OP_LMUL:
2265                         /* Based on gcc code */
2266                         ia64_setf_sig (code, FP_SCRATCH_REG, ins->sreg1);
2267                         ia64_setf_sig (code, FP_SCRATCH_REG2, ins->sreg2);
2268                         ia64_xmpy_l (code, FP_SCRATCH_REG, FP_SCRATCH_REG, FP_SCRATCH_REG2);
2269                         ia64_getf_sig (code, ins->dreg, FP_SCRATCH_REG);
2270                         break;
2271
2272                 case OP_STOREI1_MEMBASE_REG:
2273                         ia64_st1_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2274                         break;
2275                 case OP_STOREI2_MEMBASE_REG:
2276                         ia64_st2_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2277                         break;
2278                 case OP_STOREI4_MEMBASE_REG:
2279                         ia64_st4_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2280                         break;
2281                 case OP_STOREI8_MEMBASE_REG:
2282                 case OP_STORE_MEMBASE_REG:
2283                         if (ins->inst_offset != 0) {
2284                                 /* This is generated by local regalloc */
2285                                 if (ia64_is_imm14 (ins->inst_offset)) {
2286                                         ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_destbasereg);
2287                                 } else {
2288                                         ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2289                                         ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_destbasereg);
2290                                 }
2291                                 ins->inst_destbasereg = GP_SCRATCH_REG;
2292                         }
2293                         ia64_st8_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2294                         break;
2295
2296                 case OP_IA64_STOREI1_MEMBASE_INC_REG:
2297                         ia64_st1_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 1, 0);
2298                         break;
2299                 case OP_IA64_STOREI2_MEMBASE_INC_REG:
2300                         ia64_st2_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 2, 0);
2301                         break;
2302                 case OP_IA64_STOREI4_MEMBASE_INC_REG:
2303                         ia64_st4_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 4, 0);
2304                         break;
2305                 case OP_IA64_STOREI8_MEMBASE_INC_REG:
2306                         ia64_st8_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 8, 0);
2307                         break;
2308
2309                 case OP_LOADU1_MEMBASE:
2310                         ia64_ld1 (code, ins->dreg, ins->inst_basereg);
2311                         break;
2312                 case OP_LOADU2_MEMBASE:
2313                         ia64_ld2 (code, ins->dreg, ins->inst_basereg);
2314                         break;
2315                 case OP_LOADU4_MEMBASE:
2316                         ia64_ld4 (code, ins->dreg, ins->inst_basereg);
2317                         break;
2318                 case OP_LOADI1_MEMBASE:
2319                         ia64_ld1 (code, ins->dreg, ins->inst_basereg);
2320                         ia64_sxt1 (code, ins->dreg, ins->dreg);
2321                         break;
2322                 case OP_LOADI2_MEMBASE:
2323                         ia64_ld2 (code, ins->dreg, ins->inst_basereg);
2324                         ia64_sxt2 (code, ins->dreg, ins->dreg);
2325                         break;
2326                 case OP_LOADI4_MEMBASE:
2327                         ia64_ld4 (code, ins->dreg, ins->inst_basereg);
2328                         ia64_sxt4 (code, ins->dreg, ins->dreg);
2329                         break;
2330                 case OP_LOAD_MEMBASE:
2331                 case OP_LOADI8_MEMBASE:
2332                         if (ins->inst_offset != 0) {
2333                                 /* This is generated by local regalloc */
2334                                 if (ia64_is_imm14 (ins->inst_offset)) {
2335                                         ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_basereg);
2336                                 } else {
2337                                         ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2338                                         ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_basereg);
2339                                 }
2340                                 ins->inst_basereg = GP_SCRATCH_REG;
2341                         }
2342                         ia64_ld8 (code, ins->dreg, ins->inst_basereg);
2343                         break;
2344
2345                 case OP_IA64_LOADU1_MEMBASE_INC:
2346                         ia64_ld1_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 1, 0);
2347                         break;
2348                 case OP_IA64_LOADU2_MEMBASE_INC:
2349                         ia64_ld2_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 2, 0);
2350                         break;
2351                 case OP_IA64_LOADU4_MEMBASE_INC:
2352                         ia64_ld4_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 4, 0);
2353                         break;
2354                 case OP_IA64_LOADI8_MEMBASE_INC:
2355                         ia64_ld8_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 8, 0);
2356                         break;
2357
2358                 case OP_SEXT_I1:
2359                         ia64_sxt1 (code, ins->dreg, ins->sreg1);
2360                         break;
2361                 case OP_SEXT_I2:
2362                         ia64_sxt2 (code, ins->dreg, ins->sreg1);
2363                         break;
2364                 case OP_SEXT_I4:
2365                         ia64_sxt4 (code, ins->dreg, ins->sreg1);
2366                         break;
2367                 case OP_ZEXT_I1:
2368                         ia64_zxt1 (code, ins->dreg, ins->sreg1);
2369                         break;
2370                 case OP_ZEXT_I2:
2371                         ia64_zxt2 (code, ins->dreg, ins->sreg1);
2372                         break;
2373                 case OP_ZEXT_I4:
2374                         ia64_zxt4 (code, ins->dreg, ins->sreg1);
2375                         break;
2376
2377                         /* Compare opcodes */
2378                 case OP_IA64_CMP4_EQ:
2379                         ia64_cmp4_eq (code, 6, 7, ins->sreg1, ins->sreg2);
2380                         break;
2381                 case OP_IA64_CMP4_NE:
2382                         ia64_cmp4_ne (code, 6, 7, ins->sreg1, ins->sreg2);
2383                         break;
2384                 case OP_IA64_CMP4_LE:
2385                         ia64_cmp4_le (code, 6, 7, ins->sreg1, ins->sreg2);
2386                         break;
2387                 case OP_IA64_CMP4_LT:
2388                         ia64_cmp4_lt (code, 6, 7, ins->sreg1, ins->sreg2);
2389                         break;
2390                 case OP_IA64_CMP4_GE:
2391                         ia64_cmp4_ge (code, 6, 7, ins->sreg1, ins->sreg2);
2392                         break;
2393                 case OP_IA64_CMP4_GT:
2394                         ia64_cmp4_gt (code, 6, 7, ins->sreg1, ins->sreg2);
2395                         break;
2396                 case OP_IA64_CMP4_LT_UN:
2397                         ia64_cmp4_ltu (code, 6, 7, ins->sreg1, ins->sreg2);
2398                         break;
2399                 case OP_IA64_CMP4_LE_UN:
2400                         ia64_cmp4_leu (code, 6, 7, ins->sreg1, ins->sreg2);
2401                         break;
2402                 case OP_IA64_CMP4_GT_UN:
2403                         ia64_cmp4_gtu (code, 6, 7, ins->sreg1, ins->sreg2);
2404                         break;
2405                 case OP_IA64_CMP4_GE_UN:
2406                         ia64_cmp4_geu (code, 6, 7, ins->sreg1, ins->sreg2);
2407                         break;
2408                 case OP_IA64_CMP_EQ:
2409                         ia64_cmp_eq (code, 6, 7, ins->sreg1, ins->sreg2);
2410                         break;
2411                 case OP_IA64_CMP_NE:
2412                         ia64_cmp_ne (code, 6, 7, ins->sreg1, ins->sreg2);
2413                         break;
2414                 case OP_IA64_CMP_LE:
2415                         ia64_cmp_le (code, 6, 7, ins->sreg1, ins->sreg2);
2416                         break;
2417                 case OP_IA64_CMP_LT:
2418                         ia64_cmp_lt (code, 6, 7, ins->sreg1, ins->sreg2);
2419                         break;
2420                 case OP_IA64_CMP_GE:
2421                         ia64_cmp_ge (code, 6, 7, ins->sreg1, ins->sreg2);
2422                         break;
2423                 case OP_IA64_CMP_GT:
2424                         ia64_cmp_gt (code, 6, 7, ins->sreg1, ins->sreg2);
2425                         break;
2426                 case OP_IA64_CMP_GT_UN:
2427                         ia64_cmp_gtu (code, 6, 7, ins->sreg1, ins->sreg2);
2428                         break;
2429                 case OP_IA64_CMP_LT_UN:
2430                         ia64_cmp_ltu (code, 6, 7, ins->sreg1, ins->sreg2);
2431                         break;
2432                 case OP_IA64_CMP_GE_UN:
2433                         ia64_cmp_geu (code, 6, 7, ins->sreg1, ins->sreg2);
2434                         break;
2435                 case OP_IA64_CMP_LE_UN:
2436                         ia64_cmp_leu (code, 6, 7, ins->sreg1, ins->sreg2);
2437                         break;
2438                 case OP_IA64_CMP4_EQ_IMM:
2439                         ia64_cmp4_eq_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2440                         break;
2441                 case OP_IA64_CMP4_NE_IMM:
2442                         ia64_cmp4_ne_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2443                         break;
2444                 case OP_IA64_CMP4_LE_IMM:
2445                         ia64_cmp4_le_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2446                         break;
2447                 case OP_IA64_CMP4_LT_IMM:
2448                         ia64_cmp4_lt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2449                         break;
2450                 case OP_IA64_CMP4_GE_IMM:
2451                         ia64_cmp4_ge_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2452                         break;
2453                 case OP_IA64_CMP4_GT_IMM:
2454                         ia64_cmp4_gt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2455                         break;
2456                 case OP_IA64_CMP4_LT_UN_IMM:
2457                         ia64_cmp4_ltu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2458                         break;
2459                 case OP_IA64_CMP4_LE_UN_IMM:
2460                         ia64_cmp4_leu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2461                         break;
2462                 case OP_IA64_CMP4_GT_UN_IMM:
2463                         ia64_cmp4_gtu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2464                         break;
2465                 case OP_IA64_CMP4_GE_UN_IMM:
2466                         ia64_cmp4_geu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2467                         break;
2468                 case OP_IA64_CMP_EQ_IMM:
2469                         ia64_cmp_eq_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2470                         break;
2471                 case OP_IA64_CMP_NE_IMM:
2472                         ia64_cmp_ne_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2473                         break;
2474                 case OP_IA64_CMP_LE_IMM:
2475                         ia64_cmp_le_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2476                         break;
2477                 case OP_IA64_CMP_LT_IMM:
2478                         ia64_cmp_lt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2479                         break;
2480                 case OP_IA64_CMP_GE_IMM:
2481                         ia64_cmp_ge_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2482                         break;
2483                 case OP_IA64_CMP_GT_IMM:
2484                         ia64_cmp_gt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2485                         break;
2486                 case OP_IA64_CMP_GT_UN_IMM:
2487                         ia64_cmp_gtu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2488                         break;
2489                 case OP_IA64_CMP_LT_UN_IMM:
2490                         ia64_cmp_ltu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2491                         break;
2492                 case OP_IA64_CMP_GE_UN_IMM:
2493                         ia64_cmp_geu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2494                         break;
2495                 case OP_IA64_CMP_LE_UN_IMM:
2496                         ia64_cmp_leu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2497                         break;
2498                 case OP_IA64_FCMP_EQ:
2499                         ia64_fcmp_eq_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2500                         break;
2501                 case OP_IA64_FCMP_NE:
2502                         ia64_fcmp_ne_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2503                         break;
2504                 case OP_IA64_FCMP_LT:
2505                         ia64_fcmp_lt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2506                         break;
2507                 case OP_IA64_FCMP_GT:
2508                         ia64_fcmp_gt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2509                         break;
2510                 case OP_IA64_FCMP_LE:
2511                         ia64_fcmp_le_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2512                         break;
2513                 case OP_IA64_FCMP_GE:
2514                         ia64_fcmp_ge_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2515                         break;
2516                 case OP_IA64_FCMP_GT_UN:
2517                         ia64_fcmp_gt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2518                         ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2519                         break;
2520                 case OP_IA64_FCMP_LT_UN:
2521                         ia64_fcmp_lt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2522                         ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2523                         break;
2524                 case OP_IA64_FCMP_GE_UN:
2525                         ia64_fcmp_ge_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2526                         ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2527                         break;
2528                 case OP_IA64_FCMP_LE_UN:
2529                         ia64_fcmp_le_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2530                         ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2531                         break;
2532
2533                 case OP_COND_EXC_IOV:
2534                 case OP_COND_EXC_OV:
2535                         emit_cond_system_exception (cfg, code, "OverflowException", 6);
2536                         break;
2537                 case OP_COND_EXC_IC:
2538                 case OP_COND_EXC_C:
2539                         emit_cond_system_exception (cfg, code, "OverflowException", 7);
2540                         break;
2541                 case OP_IA64_COND_EXC:
2542                         emit_cond_system_exception (cfg, code, ins->inst_p1, 6);
2543                         break;
2544                 case OP_IA64_CSET:
2545                         ia64_mov_pred (code, 7, ins->dreg, IA64_R0);
2546                         ia64_no_stop (code);
2547                         ia64_add1_pred (code, 6, ins->dreg, IA64_R0, IA64_R0);
2548                         break;
2549                 case OP_ICONV_TO_I1:
2550                 case OP_LCONV_TO_I1:
2551                         /* FIXME: Is this needed ? */
2552                         ia64_sxt1 (code, ins->dreg, ins->sreg1);
2553                         break;
2554                 case OP_ICONV_TO_I2:
2555                 case OP_LCONV_TO_I2:
2556                         /* FIXME: Is this needed ? */
2557                         ia64_sxt2 (code, ins->dreg, ins->sreg1);
2558                         break;
2559                 case OP_LCONV_TO_I4:
2560                         /* FIXME: Is this needed ? */
2561                         ia64_sxt4 (code, ins->dreg, ins->sreg1);
2562                         break;
2563                 case OP_ICONV_TO_U1:
2564                 case OP_LCONV_TO_U1:
2565                         /* FIXME: Is this needed */
2566                         ia64_zxt1 (code, ins->dreg, ins->sreg1);
2567                         break;
2568                 case OP_ICONV_TO_U2:
2569                 case OP_LCONV_TO_U2:
2570                         /* FIXME: Is this needed */
2571                         ia64_zxt2 (code, ins->dreg, ins->sreg1);
2572                         break;
2573                 case OP_LCONV_TO_U4:
2574                         /* FIXME: Is this needed */
2575                         ia64_zxt4 (code, ins->dreg, ins->sreg1);
2576                         break;
2577                 case OP_ICONV_TO_I8:
2578                 case OP_ICONV_TO_I:
2579                 case OP_LCONV_TO_I8:
2580                 case OP_LCONV_TO_I:
2581                         ia64_sxt4 (code, ins->dreg, ins->sreg1);
2582                         break;
2583                 case OP_LCONV_TO_U8:
2584                 case OP_LCONV_TO_U:
2585                         ia64_zxt4 (code, ins->dreg, ins->sreg1);
2586                         break;
2587
2588                         /*
2589                          * FLOAT OPCODES
2590                          */
2591                 case OP_R8CONST: {
2592                         double d = *(double *)ins->inst_p0;
2593
2594                         if ((d == 0.0) && (mono_signbit (d) == 0))
2595                                 ia64_fmov (code, ins->dreg, 0);
2596                         else if (d == 1.0)
2597                                 ia64_fmov (code, ins->dreg, 1);
2598                         else {
2599                                 add_patch_info (cfg, code, MONO_PATCH_INFO_R8, ins->inst_p0);
2600                                 ia64_movl (code, GP_SCRATCH_REG, 0);
2601                                 ia64_ldfd (code, ins->dreg, GP_SCRATCH_REG);
2602                         }
2603                         break;
2604                 }
2605                 case OP_R4CONST: {
2606                         float f = *(float *)ins->inst_p0;
2607
2608                         if ((f == 0.0) && (mono_signbit (f) == 0))
2609                                 ia64_fmov (code, ins->dreg, 0);
2610                         else if (f == 1.0)
2611                                 ia64_fmov (code, ins->dreg, 1);
2612                         else {
2613                                 add_patch_info (cfg, code, MONO_PATCH_INFO_R4, ins->inst_p0);
2614                                 ia64_movl (code, GP_SCRATCH_REG, 0);
2615                                 ia64_ldfs (code, ins->dreg, GP_SCRATCH_REG);
2616                         }
2617                         break;
2618                 }
2619                 case OP_FMOVE:
2620                         ia64_fmov (code, ins->dreg, ins->sreg1);
2621                         break;
2622                 case OP_STORER8_MEMBASE_REG:
2623                         if (ins->inst_offset != 0) {
2624                                 /* This is generated by local regalloc */
2625                                 if (ia64_is_imm14 (ins->inst_offset)) {
2626                                         ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_destbasereg);
2627                                 } else {
2628                                         ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2629                                         ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_destbasereg);
2630                                 }
2631                                 ins->inst_destbasereg = GP_SCRATCH_REG;
2632                         }
2633                         ia64_stfd_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2634                         break;
2635                 case OP_STORER4_MEMBASE_REG:
2636                         ia64_fnorm_s_sf (code, FP_SCRATCH_REG, ins->sreg1, 0);
2637                         ia64_stfs_hint (code, ins->inst_destbasereg, FP_SCRATCH_REG, 0);
2638                         break;
2639                 case OP_LOADR8_MEMBASE:
2640                         if (ins->inst_offset != 0) {
2641                                 /* This is generated by local regalloc */
2642                                 if (ia64_is_imm14 (ins->inst_offset)) {
2643                                         ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_basereg);
2644                                 } else {
2645                                         ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2646                                         ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_basereg);
2647                                 }
2648                                 ins->inst_basereg = GP_SCRATCH_REG;
2649                         }
2650                         ia64_ldfd (code, ins->dreg, ins->inst_basereg);
2651                         break;
2652                 case OP_LOADR4_MEMBASE:
2653                         ia64_ldfs (code, ins->dreg, ins->inst_basereg);
2654                         ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
2655                         break;
2656                 case OP_ICONV_TO_R4:
2657                 case OP_LCONV_TO_R4:
2658                         ia64_setf_sig (code, ins->dreg, ins->sreg1);
2659                         ia64_fcvt_xf (code, ins->dreg, ins->dreg);
2660                         ia64_fnorm_s_sf (code, ins->dreg, ins->dreg, 0);
2661                         break;
2662                 case OP_ICONV_TO_R8:
2663                 case OP_LCONV_TO_R8:
2664                         ia64_setf_sig (code, ins->dreg, ins->sreg1);
2665                         ia64_fcvt_xf (code, ins->dreg, ins->dreg);
2666                         ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
2667                         break;
2668                 case OP_FCONV_TO_R4:
2669                         ia64_fnorm_s_sf (code, ins->dreg, ins->sreg1, 0);
2670                         break;
2671                 case OP_FCONV_TO_I8:
2672                 case OP_FCONV_TO_I:
2673                         ia64_fcvt_fx_trunc_sf (code, FP_SCRATCH_REG, ins->sreg1, 0);
2674                         ia64_getf_sig (code, ins->dreg, FP_SCRATCH_REG);
2675                         break;
2676                 case OP_FADD:
2677                         ia64_fma_d_sf (code, ins->dreg, ins->sreg1, 1, ins->sreg2, 0);
2678                         break;
2679                 case OP_FSUB:
2680                         ia64_fms_d_sf (code, ins->dreg, ins->sreg1, 1, ins->sreg2, 0);
2681                         break;
2682                 case OP_FMUL:
2683                         ia64_fma_d_sf (code, ins->dreg, ins->sreg1, ins->sreg2, 0, 0);
2684                         break;
2685                 case OP_FNEG:
2686                         ia64_fmerge_ns (code, ins->dreg, ins->sreg1, ins->sreg1);
2687                         break;
2688                 case OP_CKFINITE:
2689                         /* Quiet NaN */
2690                         ia64_fclass_m (code, 6, 7, ins->sreg1, 0x080);
2691                         emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2692                         /* Signaling NaN */
2693                         ia64_fclass_m (code, 6, 7, ins->sreg1, 0x040);
2694                         emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2695                         /* Positive infinity */
2696                         ia64_fclass_m (code, 6, 7, ins->sreg1, 0x021);
2697                         emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2698                         /* Negative infinity */
2699                         ia64_fclass_m (code, 6, 7, ins->sreg1, 0x022);
2700                         emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2701                         break;
2702
2703                 /* Calls */
2704                 case OP_CHECK_THIS:
2705                         /* ensure ins->sreg1 is not NULL */
2706                         /* Can't use ld8 as this could be a vtype address */
2707                         ia64_ld1 (code, GP_SCRATCH_REG, ins->sreg1);
2708                         break;
2709                 case OP_ARGLIST:
2710                         ia64_adds_imm (code, GP_SCRATCH_REG, cfg->sig_cookie, cfg->frame_reg);
2711                         ia64_st8 (code, ins->sreg1, GP_SCRATCH_REG);
2712                         break;
2713                 case OP_FCALL:
2714                 case OP_LCALL:
2715                 case OP_VCALL:
2716                 case OP_VCALL2:
2717                 case OP_VOIDCALL:
2718                 case OP_CALL:
2719                         call = (MonoCallInst*)ins;
2720
2721                         if (ins->flags & MONO_INST_HAS_METHOD)
2722                                 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
2723                         else
2724                                 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr);
2725
2726                         code = emit_move_return_value (cfg, ins, code);
2727                         break;
2728
2729                 case OP_CALL_REG:
2730                 case OP_FCALL_REG:
2731                 case OP_LCALL_REG:
2732                 case OP_VCALL_REG:
2733                 case OP_VCALL2_REG:
2734                 case OP_VOIDCALL_REG: {
2735                         MonoCallInst *call = (MonoCallInst*)ins;
2736                         CallInfo *cinfo;
2737                         int out_reg;
2738
2739                         /*
2740                          * mono_arch_get_this_arg_from_call () needs to find the this argument in a global
2741                          * register.
2742                          */
2743                         cinfo = get_call_info (cfg, cfg->mempool, call->signature, FALSE);
2744                         out_reg = cfg->arch.reg_out0;
2745                         ia64_mov (code, IA64_R10, out_reg);
2746
2747                         /* Indirect call */
2748                         ia64_mov (code, IA64_R8, ins->sreg1);
2749                         ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, IA64_R8, 8);
2750                         ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
2751                         ia64_ld8 (code, IA64_GP, IA64_R8);
2752                         ia64_br_call_reg (code, IA64_B0, IA64_B6);
2753
2754                         code = emit_move_return_value (cfg, ins, code);
2755                         break;
2756                 }
2757                 case OP_FCALL_MEMBASE:
2758                 case OP_LCALL_MEMBASE:
2759                 case OP_VCALL_MEMBASE:
2760                 case OP_VCALL2_MEMBASE:
2761                 case OP_VOIDCALL_MEMBASE:
2762                 case OP_CALL_MEMBASE: {
2763                         MonoCallInst *call = (MonoCallInst*)ins;
2764                         CallInfo *cinfo;
2765                         int out_reg;
2766
2767                         ia64_mov (code, IA64_R11, ins->sreg1);
2768                         if (ia64_is_imm14 (ins->inst_offset))
2769                                 ia64_adds_imm (code, IA64_R8, ins->inst_offset, ins->sreg1);
2770                         else {
2771                                 ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2772                                 ia64_add (code, IA64_R8, GP_SCRATCH_REG, ins->sreg1);
2773                         }
2774
2775                         if (call->method && ins->inst_offset < 0) {
2776                                 /*
2777                                  * This is a possible IMT call so save the IMT method in a global
2778                                  * register where mono_arch_find_imt_method () and its friends can
2779                                  * access it.
2780                                  */
2781                                 ia64_movl (code, IA64_R9, call->method);
2782                         }
2783
2784                         /*
2785                          * mono_arch_find_this_arg () needs to find the this argument in a global
2786                          * register.
2787                          */
2788                         cinfo = get_call_info (cfg, cfg->mempool, call->signature, FALSE);
2789                         out_reg = cfg->arch.reg_out0;
2790                         ia64_mov (code, IA64_R10, out_reg);
2791
2792                         ia64_ld8 (code, GP_SCRATCH_REG, IA64_R8);
2793
2794                         ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
2795
2796                         ia64_br_call_reg (code, IA64_B0, IA64_B6);
2797
2798                         code = emit_move_return_value (cfg, ins, code);
2799                         break;
2800                 }
2801                 case OP_JMP: {
2802                         /*
2803                          * Keep in sync with the code in emit_epilog.
2804                          */
2805
2806                         if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
2807                                 NOT_IMPLEMENTED;
2808
2809                         g_assert (!cfg->method->save_lmf);
2810
2811                         /* Load arguments into their original registers */
2812                         code = emit_load_volatile_arguments (cfg, code);
2813
2814                         if (cfg->arch.stack_alloc_size) {
2815                                 if (cfg->arch.omit_fp) {
2816                                         if (ia64_is_imm14 (cfg->arch.stack_alloc_size))
2817                                                 ia64_adds_imm (code, IA64_SP, (cfg->arch.stack_alloc_size), IA64_SP);
2818                                         else {
2819                                                 ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
2820                                                 ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
2821                                         }
2822                                 }
2823                                 else
2824                                         ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
2825                         }
2826                         ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
2827                         ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
2828
2829                         add_patch_info (cfg, code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
2830                         ia64_movl (code, GP_SCRATCH_REG, 0);
2831                         ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
2832                         ia64_br_cond_reg (code, IA64_B6);
2833
2834                         break;
2835                 }
2836                 case OP_BREAK:
2837                         code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, mono_break);
2838                         break;
2839
2840                 case OP_LOCALLOC: {
2841                         gint32 abi_offset;
2842
2843                         /* FIXME: Sigaltstack support */
2844
2845                         /* keep alignment */
2846                         ia64_adds_imm (code, GP_SCRATCH_REG, MONO_ARCH_LOCALLOC_ALIGNMENT - 1, ins->sreg1);
2847                         ia64_movl (code, GP_SCRATCH_REG2, ~(MONO_ARCH_LOCALLOC_ALIGNMENT - 1));
2848                         ia64_and (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
2849
2850                         ia64_sub (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
2851
2852                         ia64_mov (code, ins->dreg, IA64_SP);
2853
2854                         /* An area at sp is reserved by the ABI for parameter passing */
2855                         abi_offset = - ALIGN_TO (cfg->param_area + 16, MONO_ARCH_LOCALLOC_ALIGNMENT);
2856                         if (ia64_is_adds_imm (abi_offset))
2857                                 ia64_adds_imm (code, IA64_SP, abi_offset, IA64_SP);
2858                         else {
2859                                 ia64_movl (code, GP_SCRATCH_REG2, abi_offset);
2860                                 ia64_add (code, IA64_SP, IA64_SP, GP_SCRATCH_REG2);
2861                         }
2862
2863                         if (ins->flags & MONO_INST_INIT) {
2864                                 /* Upper limit */
2865                                 ia64_add (code, GP_SCRATCH_REG2, ins->dreg, GP_SCRATCH_REG);
2866
2867                                 ia64_codegen_set_one_ins_per_bundle (code, TRUE);
2868
2869                                 /* Init loop */
2870                                 ia64_st8_inc_imm_hint (code, ins->dreg, IA64_R0, 8, 0);
2871                                 ia64_cmp_lt (code, 8, 9, ins->dreg, GP_SCRATCH_REG2);
2872                                 ia64_br_cond_pred (code, 8, -2);
2873
2874                                 ia64_codegen_set_one_ins_per_bundle (code, FALSE);
2875
2876                                 ia64_sub (code, ins->dreg, GP_SCRATCH_REG2, GP_SCRATCH_REG);
2877                         }
2878
2879                         break;
2880                 }
2881                 case OP_LOCALLOC_IMM: {
2882                         gint32 abi_offset;
2883
2884                         /* FIXME: Sigaltstack support */
2885
2886                         gssize size = ins->inst_imm;
2887                         size = (size + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
2888
2889                         if (ia64_is_adds_imm (size))
2890                                 ia64_adds_imm (code, GP_SCRATCH_REG, size, IA64_R0);
2891                         else
2892                                 ia64_movl (code, GP_SCRATCH_REG, size);
2893
2894                         ia64_sub (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
2895                         ia64_mov (code, ins->dreg, IA64_SP);
2896
2897                         /* An area at sp is reserved by the ABI for parameter passing */
2898                         abi_offset = - ALIGN_TO (cfg->param_area + 16, MONO_ARCH_FRAME_ALIGNMENT);
2899                         if (ia64_is_adds_imm (abi_offset))
2900                                 ia64_adds_imm (code, IA64_SP, abi_offset, IA64_SP);
2901                         else {
2902                                 ia64_movl (code, GP_SCRATCH_REG2, abi_offset);
2903                                 ia64_add (code, IA64_SP, IA64_SP, GP_SCRATCH_REG2);
2904                         }
2905
2906                         if (ins->flags & MONO_INST_INIT) {
2907                                 /* Upper limit */
2908                                 ia64_add (code, GP_SCRATCH_REG2, ins->dreg, GP_SCRATCH_REG);
2909
2910                                 ia64_codegen_set_one_ins_per_bundle (code, TRUE);
2911
2912                                 /* Init loop */
2913                                 ia64_st8_inc_imm_hint (code, ins->dreg, IA64_R0, 8, 0);
2914                                 ia64_cmp_lt (code, 8, 9, ins->dreg, GP_SCRATCH_REG2);
2915                                 ia64_br_cond_pred (code, 8, -2);
2916
2917                                 ia64_codegen_set_one_ins_per_bundle (code, FALSE);
2918
2919                                 ia64_sub (code, ins->dreg, GP_SCRATCH_REG2, GP_SCRATCH_REG);
2920                         }
2921
2922                         break;
2923                 }
2924                 case OP_TLS_GET:
2925                         ia64_adds_imm (code, ins->dreg, ins->inst_offset, IA64_TP);
2926                         ia64_ld8 (code, ins->dreg, ins->dreg);
2927                         break;
2928
2929                         /* Synchronization */
2930                 case OP_MEMORY_BARRIER:
2931                         ia64_mf (code);
2932                         break;
2933                 case OP_ATOMIC_ADD_IMM_I4:
2934                         g_assert (ins->inst_offset == 0);
2935                         ia64_fetchadd4_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
2936                         ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
2937                         break;
2938                 case OP_ATOMIC_ADD_IMM_I8:
2939                         g_assert (ins->inst_offset == 0);
2940                         ia64_fetchadd8_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
2941                         ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
2942                         break;
2943                 case OP_ATOMIC_EXCHANGE_I4:
2944                         ia64_xchg4_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
2945                         ia64_sxt4 (code, ins->dreg, ins->dreg);
2946                         break;
2947                 case OP_ATOMIC_EXCHANGE_I8:
2948                         ia64_xchg8_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
2949                         break;
2950                 case OP_ATOMIC_ADD_I4: {
2951                         guint8 *label, *buf;
2952
2953                         /* From libatomic_ops */
2954                         ia64_mf (code);
2955
2956                         ia64_begin_bundle (code);
2957                         label = code.buf + code.nins;
2958                         ia64_ld4_acq (code, GP_SCRATCH_REG, ins->sreg1);
2959                         ia64_add (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, ins->sreg2);
2960                         ia64_mov_to_ar_m (code, IA64_CCV, GP_SCRATCH_REG);
2961                         ia64_cmpxchg4_acq_hint (code, GP_SCRATCH_REG2, ins->sreg1, GP_SCRATCH_REG2, 0);
2962                         ia64_cmp4_eq (code, 6, 7, GP_SCRATCH_REG, GP_SCRATCH_REG2);
2963                         buf = code.buf + code.nins;
2964                         ia64_br_cond_pred (code, 7, 0);
2965                         ia64_begin_bundle (code);
2966                         ia64_patch (buf, label);
2967                         ia64_add (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
2968                         break;
2969                 }
2970                 case OP_ATOMIC_ADD_I8: {
2971                         guint8 *label, *buf;
2972
2973                         /* From libatomic_ops */
2974                         ia64_mf (code);
2975
2976                         ia64_begin_bundle (code);
2977                         label = code.buf + code.nins;
2978                         ia64_ld8_acq (code, GP_SCRATCH_REG, ins->sreg1);
2979                         ia64_add (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, ins->sreg2);
2980                         ia64_mov_to_ar_m (code, IA64_CCV, GP_SCRATCH_REG);
2981                         ia64_cmpxchg8_acq_hint (code, GP_SCRATCH_REG2, ins->sreg1, GP_SCRATCH_REG2, 0);
2982                         ia64_cmp_eq (code, 6, 7, GP_SCRATCH_REG, GP_SCRATCH_REG2);
2983                         buf = code.buf + code.nins;
2984                         ia64_br_cond_pred (code, 7, 0);
2985                         ia64_begin_bundle (code);
2986                         ia64_patch (buf, label);
2987                         ia64_add (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
2988                         break;
2989                 }
2990
2991                         /* Exception handling */
2992                 case OP_CALL_HANDLER:
2993                         /*
2994                          * Using a call instruction would mess up the register stack, so
2995                          * save the return address to a register and use a
2996                          * branch.
2997                          */
2998                         ia64_codegen_set_one_ins_per_bundle (code, TRUE);
2999                         ia64_mov (code, IA64_R15, IA64_R0);
3000                         ia64_mov_from_ip (code, GP_SCRATCH_REG);
3001                         /* Add the length of OP_CALL_HANDLER */
3002                         ia64_adds_imm (code, GP_SCRATCH_REG, 5 * 16, GP_SCRATCH_REG);
3003                         add_patch_info (cfg, code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
3004                         ia64_movl (code, GP_SCRATCH_REG2, 0);
3005                         ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
3006                         ia64_br_cond_reg (code, IA64_B6);
3007                         // FIXME:
3008                         //mono_cfg_add_try_hole (cfg, ins->inst_eh_block, code, bb);
3009                         ia64_codegen_set_one_ins_per_bundle (code, FALSE);
3010                         break;
3011                 case OP_START_HANDLER: {
3012                         /*
3013                          * We receive the return address in GP_SCRATCH_REG.
3014                          */
3015                         MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
3016
3017                         /*
3018                          * R15 determines our caller. It is used since it is writable using
3019                          * libunwind.
3020                          * R15 == 0 means we are called by OP_CALL_HANDLER or via resume_context ()
3021                          * R15 != 0 means we are called by call_filter ().
3022                          */
3023                         ia64_codegen_set_one_ins_per_bundle (code, TRUE);
3024                         ia64_cmp_eq (code, 6, 7, IA64_R15, IA64_R0);
3025
3026                         ia64_br_cond_pred (code, 6, 6);
3027
3028                         /*
3029                          * Called by call_filter:
3030                          * Allocate a new stack frame, and set the fp register from the
3031                          * value passed in by the caller.
3032                          * We allocate a similar frame as is done by the prolog, so
3033                          * if an exception is thrown while executing the filter, the
3034                          * unwinder can unwind through the filter frame using the unwind
3035                          * info for the prolog.
3036                          */
3037                         ia64_alloc (code, cfg->arch.reg_saved_ar_pfs, cfg->arch.reg_local0 - cfg->arch.reg_in0, cfg->arch.reg_out0 - cfg->arch.reg_local0, cfg->arch.n_out_regs, 0);
3038                         ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
3039                         ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
3040                         ia64_mov (code, cfg->frame_reg, IA64_R15);
3041                         /* Signal to endfilter that we are called by call_filter */
3042                         ia64_mov (code, GP_SCRATCH_REG, IA64_R0);
3043
3044                         /* Branch target: */
3045                         if (ia64_is_imm14 (spvar->inst_offset))
3046                                 ia64_adds_imm (code, GP_SCRATCH_REG2, spvar->inst_offset, cfg->frame_reg);
3047                         else {
3048                                 ia64_movl (code, GP_SCRATCH_REG2, spvar->inst_offset);
3049                                 ia64_add (code, GP_SCRATCH_REG2, cfg->frame_reg, GP_SCRATCH_REG2);
3050                         }
3051
3052                         /* Save the return address */
3053                         ia64_st8_hint (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 0);
3054                         ia64_codegen_set_one_ins_per_bundle (code, FALSE);
3055
3056                         break;
3057                 }
3058                 case OP_ENDFINALLY:
3059                 case OP_ENDFILTER: {
3060                         /* FIXME: Return the value in ENDFILTER */
3061                         MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
3062
3063                         /* Load the return address */
3064                         if (ia64_is_imm14 (spvar->inst_offset)) {
3065                                 ia64_adds_imm (code, GP_SCRATCH_REG, spvar->inst_offset, cfg->frame_reg);
3066                         } else {
3067                                 ia64_movl (code, GP_SCRATCH_REG, spvar->inst_offset);
3068                                 ia64_add (code, GP_SCRATCH_REG, cfg->frame_reg, GP_SCRATCH_REG);
3069                         }
3070                         ia64_ld8_hint (code, GP_SCRATCH_REG, GP_SCRATCH_REG, 0);
3071
3072                         /* Test caller */
3073                         ia64_cmp_eq (code, 6, 7, GP_SCRATCH_REG, IA64_R0);
3074                         ia64_br_cond_pred (code, 7, 4);
3075
3076                         /* Called by call_filter */
3077                         /* Pop frame */
3078                         ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
3079                         ia64_mov_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
3080                         ia64_br_ret_reg (code, IA64_B0);
3081
3082                         /* Called by CALL_HANDLER */
3083                         ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
3084                         ia64_br_cond_reg (code, IA64_B6);
3085                         break;
3086                 }
3087                 case OP_THROW:
3088                         ia64_mov (code, cfg->arch.reg_out0, ins->sreg1);
3089                         code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3090                                                           (gpointer)"mono_arch_throw_exception");
3091
3092                         /*
3093                          * This might be the last instruction in the method, so add a dummy
3094                          * instruction so the unwinder will work.
3095                          */
3096                         ia64_break_i (code, 0);
3097                         break;
3098                 case OP_RETHROW:
3099                         ia64_mov (code, cfg->arch.reg_out0, ins->sreg1);
3100                         code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3101                                                           (gpointer)"mono_arch_rethrow_exception");
3102
3103                         ia64_break_i (code, 0);
3104                         break;
3105                 case OP_GC_SAFE_POINT:
3106                         break;
3107
3108                 default:
3109                         g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
3110                         g_assert_not_reached ();
3111                 }
3112
3113                 if ((code.buf - cfg->native_code - offset) > max_len) {
3114                         g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %ld)",
3115                                    mono_inst_name (ins->opcode), max_len, code.buf - cfg->native_code - offset);
3116                         g_assert_not_reached ();
3117                 }
3118
3119                 cpos += max_len;
3120
3121                 last_ins = ins;
3122                 last_offset = offset;
3123         }
3124
3125         ia64_codegen_close (code);
3126
3127         cfg->code_len = code.buf - cfg->native_code;
3128 }
3129
3130 void
3131 mono_arch_register_lowlevel_calls (void)
3132 {
3133 }
3134
3135 static Ia64InsType ins_types_in_template [32][3] = {
3136         {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3137         {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3138         {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3139         {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3140         {IA64_INS_TYPE_M, IA64_INS_TYPE_LX, IA64_INS_TYPE_LX},
3141         {IA64_INS_TYPE_M, IA64_INS_TYPE_LX, IA64_INS_TYPE_LX},
3142         {0, 0, 0},
3143         {0, 0, 0},
3144         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3145         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3146         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3147         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3148         {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_I},
3149         {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_I},
3150         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_F},
3151         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_F},
3152         {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_B},
3153         {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_B},
3154         {IA64_INS_TYPE_M, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3155         {IA64_INS_TYPE_M, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3156         {0, 0, 0},
3157         {0, 0, 0},
3158         {IA64_INS_TYPE_B, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3159         {IA64_INS_TYPE_B, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3160         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_B},
3161         {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_B},
3162         {0, 0, 0},
3163         {0, 0, 0},
3164         {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_B},
3165         {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_B},
3166         {0, 0, 0},
3167         {0, 0, 0}
3168 };
3169
3170 static gboolean stops_in_template [32][3] = {
3171         { FALSE, FALSE, FALSE },
3172         { FALSE, FALSE, TRUE },
3173         { FALSE, TRUE, FALSE },
3174         { FALSE, TRUE, TRUE },
3175         { FALSE, FALSE, FALSE },
3176         { FALSE, FALSE, TRUE },
3177         { FALSE, FALSE, FALSE },
3178         { FALSE, FALSE, FALSE },
3179
3180         { FALSE, FALSE, FALSE },
3181         { FALSE, FALSE, TRUE },
3182         { TRUE, FALSE, FALSE },
3183         { TRUE, FALSE, TRUE },
3184         { FALSE, FALSE, FALSE },
3185         { FALSE, FALSE, TRUE },
3186         { FALSE, FALSE, FALSE },
3187         { FALSE, FALSE, TRUE },
3188
3189         { FALSE, FALSE, FALSE },
3190         { FALSE, FALSE, TRUE },
3191         { FALSE, FALSE, FALSE },
3192         { FALSE, FALSE, TRUE },
3193         { FALSE, FALSE, FALSE },
3194         { FALSE, FALSE, FALSE },
3195         { FALSE, FALSE, FALSE },
3196         { FALSE, FALSE, TRUE },
3197
3198         { FALSE, FALSE, FALSE },
3199         { FALSE, FALSE, TRUE },
3200         { FALSE, FALSE, FALSE },
3201         { FALSE, FALSE, FALSE },
3202         { FALSE, FALSE, FALSE },
3203         { FALSE, FALSE, TRUE },
3204         { FALSE, FALSE, FALSE },
3205         { FALSE, FALSE, FALSE }
3206 };
3207
3208 static int last_stop_in_template [32] = {
3209         -1, 2, 1, 2, -1, 2, -1, -1,
3210         -1, 2, 0, 2, -1, 2, -1, 2,
3211         -1, 2, -1, 2, -1, -1, -1, 2,
3212         -1, 2, -1, -1, -1, 2, -1, -1
3213 };
3214
3215 static guint64 nops_for_ins_types [6] = {
3216         IA64_NOP_I,
3217         IA64_NOP_I,
3218         IA64_NOP_M,
3219         IA64_NOP_F,
3220         IA64_NOP_B,
3221         IA64_NOP_X
3222 };
3223
3224 #define ITYPE_MATCH(itype1, itype2) (((itype1) == (itype2)) || (((itype2) == IA64_INS_TYPE_A) && (((itype1) == IA64_INS_TYPE_I) || ((itype1) == IA64_INS_TYPE_M))))
3225
3226 /*
3227  * Debugging support
3228  */
3229
3230 #if 0
3231 #define DEBUG_INS_SCHED(a) do { a; } while (0)
3232 #else
3233 #define DEBUG_INS_SCHED(a)
3234 #endif
3235
3236 static void
3237 ia64_analyze_deps (Ia64CodegenState *code, int *deps_start, int *stops)
3238 {
3239         int i, pos, ins_index, current_deps_start, current_ins_start, reg;
3240         guint8 *deps = code->dep_info;
3241         gboolean need_stop, no_stop;
3242
3243         for (i = 0; i < code->nins; ++i)
3244                 stops [i] = FALSE;
3245
3246         ins_index = 0;
3247         current_deps_start = 0;
3248         current_ins_start = 0;
3249         deps_start [ins_index] = current_ins_start;
3250         pos = 0;
3251         no_stop = FALSE;
3252         DEBUG_INS_SCHED (printf ("BEGIN.\n"));
3253         while (pos < code->dep_info_pos) {
3254                 need_stop = FALSE;
3255                 switch (deps [pos]) {
3256                 case IA64_END_OF_INS:
3257                         ins_index ++;
3258                         current_ins_start = pos + 2;
3259                         deps_start [ins_index] = current_ins_start;
3260                         no_stop = FALSE;
3261                         DEBUG_INS_SCHED (printf ("(%d) END INS.\n", ins_index - 1));
3262                         break;
3263                 case IA64_NONE:
3264                         break;
3265                 case IA64_READ_GR:
3266                         reg = deps [pos + 1];
3267
3268                         DEBUG_INS_SCHED (printf ("READ GR: %d\n", reg));
3269                         for (i = current_deps_start; i < current_ins_start; i += 2)
3270                                 if (deps [i] == IA64_WRITE_GR && deps [i + 1] == reg)
3271                                         need_stop = TRUE;
3272                         break;
3273                 case IA64_WRITE_GR:
3274                         reg = code->dep_info [pos + 1];
3275
3276                         DEBUG_INS_SCHED (printf ("WRITE GR: %d\n", reg));
3277                         for (i = current_deps_start; i < current_ins_start; i += 2)
3278                                 if (deps [i] == IA64_WRITE_GR && deps [i + 1] == reg)
3279                                         need_stop = TRUE;
3280                         break;
3281                 case IA64_READ_PR:
3282                         reg = deps [pos + 1];
3283
3284                         DEBUG_INS_SCHED (printf ("READ PR: %d\n", reg));
3285                         for (i = current_deps_start; i < current_ins_start; i += 2)
3286                                 if (((deps [i] == IA64_WRITE_PR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
3287                                         need_stop = TRUE;
3288                         break;
3289                 case IA64_READ_PR_BRANCH:
3290                         reg = deps [pos + 1];
3291
3292                         /* Writes to prs by non-float instructions are visible to branches */
3293                         DEBUG_INS_SCHED (printf ("READ PR BRANCH: %d\n", reg));
3294                         for (i = current_deps_start; i < current_ins_start; i += 2)
3295                                 if (deps [i] == IA64_WRITE_PR_FLOAT && deps [i + 1] == reg)
3296                                         need_stop = TRUE;
3297                         break;
3298                 case IA64_WRITE_PR:
3299                         reg = code->dep_info [pos + 1];
3300
3301                         DEBUG_INS_SCHED (printf ("WRITE PR: %d\n", reg));
3302                         for (i = current_deps_start; i < current_ins_start; i += 2)
3303                                 if (((deps [i] == IA64_WRITE_PR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
3304                                         need_stop = TRUE;
3305                         break;
3306                 case IA64_WRITE_PR_FLOAT:
3307                         reg = code->dep_info [pos + 1];
3308
3309                         DEBUG_INS_SCHED (printf ("WRITE PR FP: %d\n", reg));
3310                         for (i = current_deps_start; i < current_ins_start; i += 2)
3311                                 if (((deps [i] == IA64_WRITE_GR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
3312                                         need_stop = TRUE;
3313                         break;
3314                 case IA64_READ_BR:
3315                         reg = deps [pos + 1];
3316
3317                         DEBUG_INS_SCHED (printf ("READ BR: %d\n", reg));
3318                         for (i = current_deps_start; i < current_ins_start; i += 2)
3319                                 if (deps [i] == IA64_WRITE_BR && deps [i + 1] == reg)
3320                                         need_stop = TRUE;
3321                         break;
3322                 case IA64_WRITE_BR:
3323                         reg = code->dep_info [pos + 1];
3324
3325                         DEBUG_INS_SCHED (printf ("WRITE BR: %d\n", reg));
3326                         for (i = current_deps_start; i < current_ins_start; i += 2)
3327                                 if (deps [i] == IA64_WRITE_BR && deps [i + 1] == reg)
3328                                         need_stop = TRUE;
3329                         break;
3330                 case IA64_READ_BR_BRANCH:
3331                         reg = deps [pos + 1];
3332
3333                         /* Writes to brs are visible to branches */
3334                         DEBUG_INS_SCHED (printf ("READ BR BRACH: %d\n", reg));
3335                         break;
3336                 case IA64_READ_FR:
3337                         reg = deps [pos + 1];
3338
3339                         DEBUG_INS_SCHED (printf ("READ BR: %d\n", reg));
3340                         for (i = current_deps_start; i < current_ins_start; i += 2)
3341                                 if (deps [i] == IA64_WRITE_FR && deps [i + 1] == reg)
3342                                         need_stop = TRUE;
3343                         break;
3344                 case IA64_WRITE_FR:
3345                         reg = code->dep_info [pos + 1];
3346
3347                         DEBUG_INS_SCHED (printf ("WRITE BR: %d\n", reg));
3348                         for (i = current_deps_start; i < current_ins_start; i += 2)
3349                                 if (deps [i] == IA64_WRITE_FR && deps [i + 1] == reg)
3350                                         need_stop = TRUE;
3351                         break;
3352                 case IA64_READ_AR:
3353                         reg = deps [pos + 1];
3354
3355                         DEBUG_INS_SCHED (printf ("READ AR: %d\n", reg));
3356                         for (i = current_deps_start; i < current_ins_start; i += 2)
3357                                 if (deps [i] == IA64_WRITE_AR && deps [i + 1] == reg)
3358                                         need_stop = TRUE;
3359                         break;
3360                 case IA64_WRITE_AR:
3361                         reg = code->dep_info [pos + 1];
3362
3363                         DEBUG_INS_SCHED (printf ("WRITE AR: %d\n", reg));
3364                         for (i = current_deps_start; i < current_ins_start; i += 2)
3365                                 if (deps [i] == IA64_WRITE_AR && deps [i + 1] == reg)
3366                                         need_stop = TRUE;
3367                         break;
3368                 case IA64_NO_STOP:
3369                         /*
3370                          * Explicitly indicate that a stop is not required. Useful for
3371                          * example when two predicated instructions with negated predicates
3372                          * write the same registers.
3373                          */
3374                         no_stop = TRUE;
3375                         break;
3376                 default:
3377                         g_assert_not_reached ();
3378                 }
3379                 pos += 2;
3380
3381                 if (need_stop && !no_stop) {
3382                         g_assert (ins_index > 0);
3383                         stops [ins_index - 1] = 1;
3384
3385                         DEBUG_INS_SCHED (printf ("STOP\n"));
3386                         current_deps_start = current_ins_start;
3387
3388                         /* Skip remaining deps for this instruction */
3389                         while (deps [pos] != IA64_END_OF_INS)
3390                                 pos += 2;
3391                 }
3392         }
3393
3394         if (code->nins > 0) {
3395                 /* No dependency info for the last instruction */
3396                 stops [code->nins - 1] = 1;
3397         }
3398
3399         deps_start [code->nins] = code->dep_info_pos;
3400 }
3401
3402 static void
3403 ia64_real_emit_bundle (Ia64CodegenState *code, int *deps_start, int *stops, int n, guint64 template, guint64 ins1, guint64 ins2, guint64 ins3, guint8 nops)
3404 {
3405         int stop_pos, i, deps_to_shift, dep_shift;
3406
3407         g_assert (n <= code->nins);
3408
3409         // if (n > 1) printf ("FOUND: %ld.\n", template);
3410
3411         ia64_emit_bundle_template (code, template, ins1, ins2, ins3);
3412
3413         stop_pos = last_stop_in_template [template] + 1;
3414         if (stop_pos > n)
3415                 stop_pos = n;
3416
3417         /* Compute the number of 'real' instructions before the stop */
3418         deps_to_shift = stop_pos;
3419         if (stop_pos >= 3 && (nops & (1 << 2)))
3420                 deps_to_shift --;
3421         if (stop_pos >= 2 && (nops & (1 << 1)))
3422                 deps_to_shift --;
3423         if (stop_pos >= 1 && (nops & (1 << 0)))
3424                 deps_to_shift --;
3425
3426         /*
3427          * We have to keep some dependencies whose instructions have been shifted
3428          * out of the buffer. So nullify the end_of_ins markers in the dependency
3429          * array.
3430          */
3431         for (i = deps_start [deps_to_shift]; i < deps_start [n]; i += 2)
3432                 if (code->dep_info [i] == IA64_END_OF_INS)
3433                         code->dep_info [i] = IA64_NONE;
3434
3435         g_assert (deps_start [deps_to_shift] <= code->dep_info_pos);
3436         memcpy (code->dep_info, &code->dep_info [deps_start [deps_to_shift]], code->dep_info_pos - deps_start [deps_to_shift]);
3437         code->dep_info_pos = code->dep_info_pos - deps_start [deps_to_shift];
3438
3439         dep_shift = deps_start [deps_to_shift];
3440         for (i = 0; i < code->nins + 1 - n; ++i)
3441                 deps_start [i] = deps_start [n + i] - dep_shift;
3442
3443         /* Determine the exact positions of instructions with unwind ops */
3444         if (code->unw_op_count) {
3445                 int ins_pos [16];
3446                 int curr_ins, curr_ins_pos;
3447
3448                 curr_ins = 0;
3449                 curr_ins_pos = ((code->buf - code->region_start - 16) / 16) * 3;
3450                 for (i = 0; i < 3; ++i) {
3451                         if (! (nops & (1 << i))) {
3452                                 ins_pos [curr_ins] = curr_ins_pos + i;
3453                                 curr_ins ++;
3454                         }
3455                 }
3456
3457                 for (i = code->unw_op_pos; i < code->unw_op_count; ++i) {
3458                         if (code->unw_ops_pos [i] < n) {
3459                                 code->unw_ops [i].when = ins_pos [code->unw_ops_pos [i]];
3460                                 //printf ("UNW-OP: %d -> %d\n", code->unw_ops_pos [i], code->unw_ops [i].when);
3461                         }
3462                 }
3463                 if (code->unw_op_pos < code->unw_op_count)
3464                         code->unw_op_pos += n;
3465         }
3466
3467         if (n == code->nins) {
3468                 code->template = 0;
3469                 code->nins = 0;
3470         }
3471         else {
3472                 memcpy (&code->instructions [0], &code->instructions [n], (code->nins - n) * sizeof (guint64));
3473                 memcpy (&code->itypes [0], &code->itypes [n], (code->nins - n) * sizeof (int));
3474                 memcpy (&stops [0], &stops [n], (code->nins - n) * sizeof (int));
3475                 code->nins -= n;
3476         }
3477 }
3478
3479 void
3480 ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
3481 {
3482         int i, ins_type, template, nins_to_emit;
3483         int deps_start [16];
3484         int stops [16];
3485         gboolean found;
3486
3487         /*
3488          * We implement a simple scheduler which tries to put three instructions
3489          * per bundle, then two, then one.
3490          */
3491         ia64_analyze_deps (code, deps_start, stops);
3492
3493         if ((code->nins >= 3) && !code->one_ins_per_bundle) {
3494                 /* Find a suitable template */
3495                 for (template = 0; template < 32; ++template) {
3496                         if (stops_in_template [template][0] != stops [0] ||
3497                                 stops_in_template [template][1] != stops [1] ||
3498                                 stops_in_template [template][2] != stops [2])
3499                                 continue;
3500
3501                         found = TRUE;
3502                         for (i = 0; i < 3; ++i) {
3503                                 ins_type = ins_types_in_template [template][i];
3504                                 switch (code->itypes [i]) {
3505                                 case IA64_INS_TYPE_A:
3506                                         found &= (ins_type == IA64_INS_TYPE_I) || (ins_type == IA64_INS_TYPE_M);
3507                                         break;
3508                                 default:
3509                                         found &= (ins_type == code->itypes [i]);
3510                                         break;
3511                                 }
3512                         }
3513
3514                         if (found)
3515                                 found = debug_ins_sched ();
3516
3517                         if (found) {
3518                                 ia64_real_emit_bundle (code, deps_start, stops, 3, template, code->instructions [0], code->instructions [1], code->instructions [2], 0);
3519                                 break;
3520                         }
3521                 }
3522         }
3523
3524         if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
3525                 /* Wait for more instructions */
3526                 return;
3527
3528         /* If it didn't work out, try putting two instructions into one bundle */
3529         if ((code->nins >= 2) && !code->one_ins_per_bundle) {
3530                 /* Try a nop at the end */
3531                 for (template = 0; template < 32; ++template) {
3532                         if (stops_in_template [template][0] != stops [0] ||
3533                                 ((stops_in_template [template][1] != stops [1]) &&
3534                                  (stops_in_template [template][2] != stops [1])))
3535
3536                                 continue;
3537
3538                         if (!ITYPE_MATCH (ins_types_in_template [template][0], code->itypes [0]) ||
3539                                 !ITYPE_MATCH (ins_types_in_template [template][1], code->itypes [1]))
3540                                 continue;
3541
3542                         if (!debug_ins_sched ())
3543                                 continue;
3544
3545                         ia64_real_emit_bundle (code, deps_start, stops, 2, template, code->instructions [0], code->instructions [1], nops_for_ins_types [ins_types_in_template [template][2]], 1 << 2);
3546                         break;
3547                 }
3548         }
3549
3550         if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
3551                 /* Wait for more instructions */
3552                 return;
3553
3554         if ((code->nins >= 2) && !code->one_ins_per_bundle) {
3555                 /* Try a nop in the middle */
3556                 for (template = 0; template < 32; ++template) {
3557                         if (((stops_in_template [template][0] != stops [0]) &&
3558                                  (stops_in_template [template][1] != stops [0])) ||
3559                                 stops_in_template [template][2] != stops [1])
3560                                 continue;
3561
3562                         if (!ITYPE_MATCH (ins_types_in_template [template][0], code->itypes [0]) ||
3563                                 !ITYPE_MATCH (ins_types_in_template [template][2], code->itypes [1]))
3564                                 continue;
3565
3566                         if (!debug_ins_sched ())
3567                                 continue;
3568
3569                         ia64_real_emit_bundle (code, deps_start, stops, 2, template, code->instructions [0], nops_for_ins_types [ins_types_in_template [template][1]], code->instructions [1], 1 << 1);
3570                         break;
3571                 }
3572         }
3573
3574         if ((code->nins >= 2) && flush && !code->one_ins_per_bundle) {
3575                 /* Try a nop at the beginning */
3576                 for (template = 0; template < 32; ++template) {
3577                         if ((stops_in_template [template][1] != stops [0]) ||
3578                                 (stops_in_template [template][2] != stops [1]))
3579                                 continue;
3580
3581                         if (!ITYPE_MATCH (ins_types_in_template [template][1], code->itypes [0]) ||
3582                                 !ITYPE_MATCH (ins_types_in_template [template][2], code->itypes [1]))
3583                                 continue;
3584
3585                         if (!debug_ins_sched ())
3586                                 continue;
3587
3588                         ia64_real_emit_bundle (code, deps_start, stops, 2, template, nops_for_ins_types [ins_types_in_template [template][0]], code->instructions [0], code->instructions [1], 1 << 0);
3589                         break;
3590                 }
3591         }
3592
3593         if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
3594                 /* Wait for more instructions */
3595                 return;
3596
3597         if (flush)
3598                 nins_to_emit = code->nins;
3599         else
3600                 nins_to_emit = 1;
3601
3602         while (nins_to_emit > 0) {
3603                 if (!debug_ins_sched ())
3604                         stops [0] = 1;
3605                 switch (code->itypes [0]) {
3606                 case IA64_INS_TYPE_A:
3607                         if (stops [0])
3608                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3609                         else
3610                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3611                         break;
3612                 case IA64_INS_TYPE_I:
3613                         if (stops [0])
3614                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3615                         else
3616                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3617                         break;
3618                 case IA64_INS_TYPE_M:
3619                         if (stops [0])
3620                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3621                         else
3622                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3623                         break;
3624                 case IA64_INS_TYPE_B:
3625                         if (stops [0])
3626                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIBS, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
3627                         else
3628                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIB, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
3629                         break;
3630                 case IA64_INS_TYPE_F:
3631                         if (stops [0])
3632                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3633                         else
3634                                 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFI, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3635                         break;
3636                 case IA64_INS_TYPE_LX:
3637                         if (stops [0] || stops [1])
3638                                 ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLXS, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
3639                         else
3640                                 ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLX, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
3641                         nins_to_emit --;
3642                         break;
3643                 default:
3644                         g_assert_not_reached ();
3645                 }
3646                 nins_to_emit --;
3647         }
3648 }
3649
3650 unw_dyn_region_info_t*
3651 mono_ia64_create_unwind_region (Ia64CodegenState *code)
3652 {
3653         unw_dyn_region_info_t *r;
3654
3655         g_assert (code->nins == 0);
3656         r = g_malloc0 (_U_dyn_region_info_size (code->unw_op_count));
3657         memcpy (&r->op, &code->unw_ops, sizeof (unw_dyn_op_t) * code->unw_op_count);
3658         r->op_count = code->unw_op_count;
3659         r->insn_count = ((code->buf - code->region_start) >> 4) * 3;
3660         code->unw_op_count = 0;
3661         code->unw_op_pos = 0;
3662         code->region_start = code->buf;
3663
3664         return r;
3665 }
3666
3667 static void
3668 ia64_patch (unsigned char* code, gpointer target)
3669 {
3670         int template, i;
3671         guint64 instructions [3];
3672         guint8 gen_buf [16];
3673         Ia64CodegenState gen;
3674         int ins_to_skip;
3675         gboolean found;
3676
3677         /*
3678          * code encodes both the position inside the buffer and code.nins when
3679          * the instruction was emitted.
3680          */
3681         ins_to_skip = (guint64)code % 16;
3682         code = (unsigned char*)((guint64)code & ~15);
3683
3684         /*
3685          * Search for the first instruction which is 'patchable', skipping
3686          * ins_to_skip instructions.
3687          */
3688
3689         while (TRUE) {
3690
3691         template = ia64_bundle_template (code);
3692         instructions [0] = ia64_bundle_ins1 (code);
3693         instructions [1] = ia64_bundle_ins2 (code);
3694         instructions [2] = ia64_bundle_ins3 (code);
3695
3696         ia64_codegen_init (gen, gen_buf);
3697
3698         found = FALSE;
3699         for (i = 0; i < 3; ++i) {
3700                 guint64 ins = instructions [i];
3701                 int opcode = ia64_ins_opcode (ins);
3702
3703                 if (ins == nops_for_ins_types [ins_types_in_template [template][i]])
3704                         continue;
3705
3706                 if (ins_to_skip) {
3707                         ins_to_skip --;
3708                         continue;
3709                 }
3710
3711                 switch (ins_types_in_template [template][i]) {
3712                 case IA64_INS_TYPE_A:
3713                 case IA64_INS_TYPE_M:
3714                         if ((opcode == 8) && (ia64_ins_x2a (ins) == 2) && (ia64_ins_ve (ins) == 0)) {
3715                                 /* adds */
3716                                 ia64_adds_imm_pred (gen, ia64_ins_qp (ins), ia64_ins_r1 (ins), (guint64)target, ia64_ins_r3 (ins));
3717                                 instructions [i] = gen.instructions [0];
3718                                 found = TRUE;
3719                         }
3720                         else
3721                                 NOT_IMPLEMENTED;
3722                         break;
3723                 case IA64_INS_TYPE_B:
3724                         if ((opcode == 4) && (ia64_ins_btype (ins) == 0)) {
3725                                 /* br.cond */
3726                                 gint64 disp = ((guint8*)target - code) >> 4;
3727
3728                                 /* FIXME: hints */
3729                                 ia64_br_cond_hint_pred (gen, ia64_ins_qp (ins), disp, 0, 0, 0);
3730
3731                                 instructions [i] = gen.instructions [0];
3732                                 found = TRUE;
3733                         }
3734                         else if (opcode == 5) {
3735                                 /* br.call */
3736                                 gint64 disp = ((guint8*)target - code) >> 4;
3737
3738                                 /* FIXME: hints */
3739                                 ia64_br_call_hint_pred (gen, ia64_ins_qp (ins), ia64_ins_b1 (ins), disp, 0, 0, 0);
3740                                 instructions [i] = gen.instructions [0];
3741                                 found = TRUE;
3742                         }
3743                         else
3744                                 NOT_IMPLEMENTED;
3745                         break;
3746                 case IA64_INS_TYPE_LX:
3747                         if (i == 1)
3748                                 break;
3749
3750                         if ((opcode == 6) && (ia64_ins_vc (ins) == 0)) {
3751                                 /* movl */
3752                                 ia64_movl_pred (gen, ia64_ins_qp (ins), ia64_ins_r1 (ins), target);
3753                                 instructions [1] = gen.instructions [0];
3754                                 instructions [2] = gen.instructions [1];
3755                                 found = TRUE;
3756                         }
3757                         else
3758                                 NOT_IMPLEMENTED;
3759
3760                         break;
3761                 default:
3762                         NOT_IMPLEMENTED;
3763                 }
3764
3765                 if (found) {
3766                         /* Rewrite code */
3767                         ia64_codegen_init (gen, code);
3768                         ia64_emit_bundle_template (&gen, template, instructions [0], instructions [1], instructions [2]);
3769                         return;
3770                 }
3771         }
3772
3773         code += 16;
3774         }
3775 }
3776
3777 void
3778 mono_arch_patch_code (MonoCompile *cfg, MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gboolean run_cctors)
3779 {
3780         MonoJumpInfo *patch_info;
3781
3782         for (patch_info = ji; patch_info; patch_info = patch_info->next) {
3783                 unsigned char *ip = patch_info->ip.i + code;
3784                 const unsigned char *target;
3785
3786                 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
3787
3788                 if (patch_info->type == MONO_PATCH_INFO_NONE)
3789                         continue;
3790                 if (mono_compile_aot) {
3791                         NOT_IMPLEMENTED;
3792                 }
3793
3794                 ia64_patch (ip, (gpointer)target);
3795         }
3796 }
3797
3798 guint8 *
3799 mono_arch_emit_prolog (MonoCompile *cfg)
3800 {
3801         MonoMethod *method = cfg->method;
3802         MonoMethodSignature *sig;
3803         MonoInst *inst;
3804         int alloc_size, pos, i;
3805         Ia64CodegenState code;
3806         CallInfo *cinfo;
3807
3808         sig = mono_method_signature (method);
3809         pos = 0;
3810
3811         cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
3812
3813         cfg->code_size =  MAX (cfg->header->code_size * 4, 512);
3814
3815         if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
3816                 cfg->code_size += 1024;
3817         if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
3818                 cfg->code_size += 1024;
3819
3820         cfg->native_code = g_malloc (cfg->code_size);
3821
3822         ia64_codegen_init (code, cfg->native_code);
3823
3824         alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
3825         if (cfg->param_area)
3826                 alloc_size += cfg->param_area;
3827         if (alloc_size)
3828                 /* scratch area */
3829                 alloc_size += 16;
3830         alloc_size = ALIGN_TO (alloc_size, MONO_ARCH_FRAME_ALIGNMENT);
3831
3832         if (cfg->flags & MONO_CFG_HAS_ALLOCA)
3833                 /* Force sp to be saved/restored */
3834                 alloc_size += MONO_ARCH_FRAME_ALIGNMENT;
3835
3836         cfg->arch.stack_alloc_size = alloc_size;
3837
3838         pos = 0;
3839
3840         if (method->save_lmf) {
3841                 /* No LMF on IA64 */
3842         }
3843
3844         alloc_size -= pos;
3845
3846         ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + cfg->arch.reg_saved_ar_pfs);
3847         ia64_alloc (code, cfg->arch.reg_saved_ar_pfs, cfg->arch.reg_local0 - cfg->arch.reg_in0, cfg->arch.reg_out0 - cfg->arch.reg_local0, cfg->arch.n_out_regs, 0);
3848         ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + cfg->arch.reg_saved_b0);
3849         ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
3850
3851         if ((alloc_size || cinfo->stack_usage) && !cfg->arch.omit_fp) {
3852                 ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + cfg->arch.reg_saved_sp);
3853                 ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
3854                 if (cfg->frame_reg != cfg->arch.reg_saved_sp)
3855                         ia64_mov (code, cfg->frame_reg, IA64_SP);
3856         }
3857
3858         if (alloc_size) {
3859 #if defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3860                 int pagesize = getpagesize ();
3861
3862                 if (alloc_size >= pagesize) {
3863                         gint32 remaining_size = alloc_size;
3864
3865                         /* Generate stack touching code */
3866                         ia64_mov (code, GP_SCRATCH_REG, IA64_SP);
3867                         while (remaining_size >= pagesize) {
3868                                 ia64_movl (code, GP_SCRATCH_REG2, pagesize);
3869                                 ia64_sub (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
3870                                 ia64_ld8 (code, GP_SCRATCH_REG2, GP_SCRATCH_REG);
3871                                 remaining_size -= pagesize;
3872                         }
3873                 }
3874 #endif
3875                 if (ia64_is_imm14 (-alloc_size)) {
3876                         if (cfg->arch.omit_fp)
3877                                 ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
3878                         ia64_adds_imm (code, IA64_SP, (-alloc_size), IA64_SP);
3879                 }
3880                 else {
3881                         ia64_movl (code, GP_SCRATCH_REG, -alloc_size);
3882                         if (cfg->arch.omit_fp)
3883                                 ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
3884                         ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
3885                 }
3886         }
3887
3888         ia64_begin_bundle (code);
3889
3890         /* Initialize unwind info */
3891         cfg->arch.r_pro = mono_ia64_create_unwind_region (&code);
3892
3893         if (sig->ret->type != MONO_TYPE_VOID) {
3894                 if ((cinfo->ret.storage == ArgInIReg) && (cfg->ret->opcode != OP_REGVAR)) {
3895                         /* Save volatile arguments to the stack */
3896                         NOT_IMPLEMENTED;
3897                 }
3898         }
3899
3900         /* Keep this in sync with emit_load_volatile_arguments */
3901         for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
3902                 ArgInfo *ainfo = cinfo->args + i;
3903                 gint32 stack_offset;
3904                 MonoType *arg_type;
3905
3906                 inst = cfg->args [i];
3907
3908                 if (sig->hasthis && (i == 0))
3909                         arg_type = &mono_defaults.object_class->byval_arg;
3910                 else
3911                         arg_type = sig->params [i - sig->hasthis];
3912
3913                 arg_type = mini_get_underlying_type (arg_type);
3914
3915                 stack_offset = ainfo->offset + ARGS_OFFSET;
3916
3917                 /*
3918                  * FIXME: Native code might pass non register sized integers
3919                  * without initializing the upper bits.
3920                  */
3921                 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED && !arg_type->byref && ainfo->storage == ArgInIReg) {
3922                         int reg = cfg->arch.reg_in0 + ainfo->reg;
3923
3924                         switch (mono_type_to_load_membase (cfg, arg_type)) {
3925                         case OP_LOADI1_MEMBASE:
3926                                 ia64_sxt1 (code, reg, reg);
3927                                 break;
3928                         case OP_LOADU1_MEMBASE:
3929                                 ia64_zxt1 (code, reg, reg);
3930                                 break;
3931                         case OP_LOADI2_MEMBASE:
3932                                 ia64_sxt2 (code, reg, reg);
3933                                 break;
3934                         case OP_LOADU2_MEMBASE:
3935                                 ia64_zxt2 (code, reg, reg);
3936                                 break;
3937                         default:
3938                                 break;
3939                         }
3940                 }
3941
3942                 /* Save volatile arguments to the stack */
3943                 if (inst->opcode != OP_REGVAR) {
3944                         switch (ainfo->storage) {
3945                         case ArgInIReg:
3946                         case ArgInFloatReg:
3947                         case ArgInFloatRegR4:
3948                                 g_assert (inst->opcode == OP_REGOFFSET);
3949                                 if (ia64_is_adds_imm (inst->inst_offset))
3950                                         ia64_adds_imm (code, GP_SCRATCH_REG, inst->inst_offset, inst->inst_basereg);
3951                                 else {
3952                                         ia64_movl (code, GP_SCRATCH_REG2, inst->inst_offset);
3953                                         ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
3954                                 }
3955                                 if (arg_type->byref)
3956                                         ia64_st8_hint (code, GP_SCRATCH_REG, cfg->arch.reg_in0 + ainfo->reg, 0);
3957                                 else {
3958                                         switch (arg_type->type) {
3959                                         case MONO_TYPE_R4:
3960                                                 ia64_stfs_hint (code, GP_SCRATCH_REG, ainfo->reg, 0);
3961                                                 break;
3962                                         case MONO_TYPE_R8:
3963                                                 ia64_stfd_hint (code, GP_SCRATCH_REG, ainfo->reg, 0);
3964                                                 break;
3965                                         default:
3966                                                 ia64_st8_hint (code, GP_SCRATCH_REG, cfg->arch.reg_in0 + ainfo->reg, 0);
3967                                                 break;
3968                                         }
3969                                 }
3970                                 break;
3971                         case ArgOnStack:
3972                                 break;
3973                         case ArgAggregate:
3974                                 if (ainfo->nslots != ainfo->nregs)
3975                                         NOT_IMPLEMENTED;
3976
3977                                 g_assert (inst->opcode == OP_REGOFFSET);
3978                                 ia64_adds_imm (code, GP_SCRATCH_REG, inst->inst_offset, inst->inst_basereg);
3979                                 for (i = 0; i < ainfo->nregs; ++i) {
3980                                         switch (ainfo->atype) {
3981                                         case AggregateNormal:
3982                                                 ia64_st8_inc_imm_hint (code, GP_SCRATCH_REG, cfg->arch.reg_in0 + ainfo->reg + i, sizeof (gpointer), 0);
3983                                                 break;
3984                                         case AggregateSingleHFA:
3985                                                 ia64_stfs_inc_imm_hint (code, GP_SCRATCH_REG, ainfo->reg + i, 4, 0);
3986                                                 break;
3987                                         case AggregateDoubleHFA:
3988                                                 ia64_stfd_inc_imm_hint (code, GP_SCRATCH_REG, ainfo->reg + i, sizeof (gpointer), 0);
3989                                                 break;
3990                                         default:
3991                                                 NOT_IMPLEMENTED;
3992                                         }
3993                                 }
3994                                 break;
3995                         default:
3996                                 g_assert_not_reached ();
3997                         }
3998                 }
3999
4000                 if (inst->opcode == OP_REGVAR) {
4001                         /* Argument allocated to (non-volatile) register */
4002                         switch (ainfo->storage) {
4003                         case ArgInIReg:
4004                                 if (inst->dreg != cfg->arch.reg_in0 + ainfo->reg)
4005                                         ia64_mov (code, inst->dreg, cfg->arch.reg_in0 + ainfo->reg);
4006                                 break;
4007                         case ArgOnStack:
4008                                 ia64_adds_imm (code, GP_SCRATCH_REG, 16 + ainfo->offset, cfg->frame_reg);
4009                                 ia64_ld8 (code, inst->dreg, GP_SCRATCH_REG);
4010                                 break;
4011                         default:
4012                                 NOT_IMPLEMENTED;
4013                         }
4014                 }
4015         }
4016
4017         if (method->save_lmf) {
4018                 /* No LMF on IA64 */
4019         }
4020
4021         ia64_codegen_close (code);
4022
4023         if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4024                 code.buf = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code.buf, TRUE);
4025
4026         cfg->code_len = code.buf - cfg->native_code;
4027
4028         g_assert (cfg->code_len < cfg->code_size);
4029
4030         cfg->arch.prolog_end_offset = cfg->code_len;
4031
4032         return code.buf;
4033 }
4034
4035 void
4036 mono_arch_emit_epilog (MonoCompile *cfg)
4037 {
4038         MonoMethod *method = cfg->method;
4039         int i, pos;
4040         int max_epilog_size = 16 * 4;
4041         Ia64CodegenState code;
4042         guint8 *buf;
4043         CallInfo *cinfo;
4044         ArgInfo *ainfo;
4045
4046         if (mono_jit_trace_calls != NULL)
4047                 max_epilog_size += 1024;
4048
4049         cfg->arch.epilog_begin_offset = cfg->code_len;
4050
4051         while (cfg->code_len + max_epilog_size > cfg->code_size) {
4052                 cfg->code_size *= 2;
4053                 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4054                 cfg->stat_code_reallocs++;
4055         }
4056
4057         /* FIXME: Emit unwind info */
4058
4059         buf = cfg->native_code + cfg->code_len;
4060
4061         if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4062                 buf = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, buf, TRUE);
4063
4064         ia64_codegen_init (code, buf);
4065
4066         /* the code restoring the registers must be kept in sync with OP_JMP */
4067         pos = 0;
4068
4069         if (method->save_lmf) {
4070                 /* No LMF on IA64 */
4071         }
4072
4073         /* Load returned vtypes into registers if needed */
4074         cinfo = get_call_info (cfg, cfg->mempool, mono_method_signature (method), FALSE);
4075         ainfo = &cinfo->ret;
4076         switch (ainfo->storage) {
4077         case ArgAggregate:
4078                 if (ainfo->nslots != ainfo->nregs)
4079                         NOT_IMPLEMENTED;
4080
4081                 g_assert (cfg->ret->opcode == OP_REGOFFSET);
4082                 ia64_adds_imm (code, GP_SCRATCH_REG, cfg->ret->inst_offset, cfg->ret->inst_basereg);
4083                 for (i = 0; i < ainfo->nregs; ++i) {
4084                         switch (ainfo->atype) {
4085                         case AggregateNormal:
4086                                 ia64_ld8_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, sizeof (gpointer), 0);
4087                                 break;
4088                         case AggregateSingleHFA:
4089                                 ia64_ldfs_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, 4, 0);
4090                                 break;
4091                         case AggregateDoubleHFA:
4092                                 ia64_ldfd_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, sizeof (gpointer), 0);
4093                                 break;
4094                         default:
4095                                 g_assert_not_reached ();
4096                         }
4097                 }
4098                 break;
4099         default:
4100                 break;
4101         }
4102
4103         ia64_begin_bundle (code);
4104
4105         code.region_start = cfg->native_code;
4106
4107         /* Label the unwind state at the start of the exception throwing region */
4108         //ia64_unw_label_state (code, 1234);
4109
4110         if (cfg->arch.stack_alloc_size) {
4111                 if (cfg->arch.omit_fp) {
4112                         if (ia64_is_imm14 (cfg->arch.stack_alloc_size)) {
4113                                 ia64_unw_pop_frames (code, 1);
4114                                 ia64_adds_imm (code, IA64_SP, (cfg->arch.stack_alloc_size), IA64_SP);
4115                         } else {
4116                                 ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
4117                                 ia64_unw_pop_frames (code, 1);
4118                                 ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
4119                         }
4120                 }
4121                 else {
4122                         ia64_unw_pop_frames (code, 1);
4123                         ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
4124                 }
4125         }
4126         ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
4127         ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
4128         ia64_br_ret_reg (code, IA64_B0);
4129
4130         ia64_codegen_close (code);
4131
4132         cfg->arch.r_epilog = mono_ia64_create_unwind_region (&code);
4133         cfg->arch.r_pro->next = cfg->arch.r_epilog;
4134
4135         cfg->code_len = code.buf - cfg->native_code;
4136
4137         g_assert (cfg->code_len < cfg->code_size);
4138 }
4139
4140 void
4141 mono_arch_emit_exceptions (MonoCompile *cfg)
4142 {
4143         MonoJumpInfo *patch_info;
4144         int i, nthrows;
4145         Ia64CodegenState code;
4146         gboolean empty = TRUE;
4147         //unw_dyn_region_info_t *r_exceptions;
4148         MonoClass *exc_classes [16];
4149         guint8 *exc_throw_start [16], *exc_throw_end [16];
4150         guint32 code_size = 0;
4151
4152         /* Compute needed space */
4153         for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4154                 if (patch_info->type == MONO_PATCH_INFO_EXC)
4155                         code_size += 256;
4156                 if (patch_info->type == MONO_PATCH_INFO_R8)
4157                         code_size += 8 + 7; /* sizeof (double) + alignment */
4158                 if (patch_info->type == MONO_PATCH_INFO_R4)
4159                         code_size += 4 + 7; /* sizeof (float) + alignment */
4160         }
4161
4162         if (code_size == 0)
4163                 return;
4164
4165         while (cfg->code_len + code_size > (cfg->code_size - 16)) {
4166                 cfg->code_size *= 2;
4167                 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4168                 cfg->stat_code_reallocs++;
4169         }
4170
4171         ia64_codegen_init (code, cfg->native_code + cfg->code_len);
4172
4173         /* The unwind state here is the same as before the epilog */
4174         //ia64_unw_copy_state (code, 1234);
4175
4176         /* add code to raise exceptions */
4177         /* FIXME: Optimize this */
4178         nthrows = 0;
4179         for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4180                 switch (patch_info->type) {
4181                 case MONO_PATCH_INFO_EXC: {
4182                         MonoClass *exc_class;
4183                         guint8* throw_ip;
4184                         guint8* buf;
4185                         guint64 exc_token_index;
4186
4187                         exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
4188                         g_assert (exc_class);
4189                         exc_token_index = mono_metadata_token_index (exc_class->type_token);
4190                         throw_ip = cfg->native_code + patch_info->ip.i;
4191
4192                         ia64_begin_bundle (code);
4193
4194                         ia64_patch (cfg->native_code + patch_info->ip.i, code.buf);
4195
4196                         /* Find a throw sequence for the same exception class */
4197                         for (i = 0; i < nthrows; ++i)
4198                                 if (exc_classes [i] == exc_class)
4199                                         break;
4200
4201                         if (i < nthrows) {
4202                                 gint64 offset = exc_throw_end [i] - 16 - throw_ip;
4203
4204                                 if (ia64_is_adds_imm (offset))
4205                                         ia64_adds_imm (code, cfg->arch.reg_out0 + 1, offset, IA64_R0);
4206                                 else
4207                                         ia64_movl (code, cfg->arch.reg_out0 + 1, offset);
4208
4209                                 buf = code.buf + code.nins;
4210                                 ia64_br_cond_pred (code, 0, 0);
4211                                 ia64_begin_bundle (code);
4212                                 ia64_patch (buf, exc_throw_start [i]);
4213
4214                                 patch_info->type = MONO_PATCH_INFO_NONE;
4215                         }
4216                         else {
4217                                 /* Arg1 */
4218                                 buf = code.buf;
4219                                 ia64_movl (code, cfg->arch.reg_out0 + 1, 0);
4220
4221                                 ia64_begin_bundle (code);
4222
4223                                 if (nthrows < 16) {
4224                                         exc_classes [nthrows] = exc_class;
4225                                         exc_throw_start [nthrows] = code.buf;
4226                                 }
4227
4228                                 /* Arg2 */
4229                                 if (ia64_is_adds_imm (exc_token_index))
4230                                         ia64_adds_imm (code, cfg->arch.reg_out0 + 0, exc_token_index, IA64_R0);
4231                                 else
4232                                         ia64_movl (code, cfg->arch.reg_out0 + 0, exc_token_index);
4233
4234                                 patch_info->data.name = "mono_arch_throw_corlib_exception";
4235                                 patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
4236                                 patch_info->ip.i = code.buf + code.nins - cfg->native_code;
4237
4238                                 /* Indirect call */
4239                                 ia64_movl (code, GP_SCRATCH_REG, 0);
4240                                 ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
4241                                 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
4242                                 ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
4243
4244                                 ia64_br_call_reg (code, IA64_B0, IA64_B6);
4245
4246                                 /* Patch up the throw offset */
4247                                 ia64_begin_bundle (code);
4248
4249                                 ia64_patch (buf, (gpointer)(code.buf - 16 - throw_ip));
4250
4251                                 if (nthrows < 16) {
4252                                         exc_throw_end [nthrows] = code.buf;
4253                                         nthrows ++;
4254                                 }
4255                         }
4256
4257                         empty = FALSE;
4258                         break;
4259                 }
4260                 default:
4261                         break;
4262                 }
4263         }
4264
4265         if (!empty)
4266                 /* The unwinder needs this to work */
4267                 ia64_break_i (code, 0);
4268
4269         ia64_codegen_close (code);
4270
4271         /* FIXME: */
4272         //r_exceptions = mono_ia64_create_unwind_region (&code);
4273         //cfg->arch.r_epilog = r_exceptions;
4274
4275         cfg->code_len = code.buf - cfg->native_code;
4276
4277         g_assert (cfg->code_len < cfg->code_size);
4278 }
4279
4280 void*
4281 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
4282 {
4283         Ia64CodegenState code;
4284         CallInfo *cinfo = NULL;
4285         MonoMethodSignature *sig;
4286         MonoInst *ins;
4287         int i, n, stack_area = 0;
4288
4289         ia64_codegen_init (code, p);
4290
4291         /* Keep this in sync with mono_arch_get_argument_info */
4292
4293         if (enable_arguments) {
4294                 /* Allocate a new area on the stack and save arguments there */
4295                 sig = mono_method_signature (cfg->method);
4296
4297                 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
4298
4299                 n = sig->param_count + sig->hasthis;
4300
4301                 stack_area = ALIGN_TO (n * 8, 16);
4302
4303                 if (n) {
4304                         ia64_movl (code, GP_SCRATCH_REG, stack_area);
4305
4306                         ia64_sub (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
4307
4308                         /* FIXME: Allocate out registers */
4309
4310                         ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_SP);
4311
4312                         /* Required by the ABI */
4313                         ia64_adds_imm (code, IA64_SP, -16, IA64_SP);
4314
4315                         add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, cfg->method);
4316                         ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
4317
4318                         /* Save arguments to the stack */
4319                         for (i = 0; i < n; ++i) {
4320                                 ins = cfg->args [i];
4321
4322                                 if (ins->opcode == OP_REGVAR) {
4323                                         ia64_movl (code, GP_SCRATCH_REG, (i * 8));
4324                                         ia64_add (code, GP_SCRATCH_REG, cfg->arch.reg_out0 + 1, GP_SCRATCH_REG);
4325                                         ia64_st8 (code, GP_SCRATCH_REG, ins->dreg);
4326                                 }
4327                                 else {
4328                                         ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
4329                                         ia64_add (code, GP_SCRATCH_REG, ins->inst_basereg, GP_SCRATCH_REG);
4330                                         ia64_ld8 (code, GP_SCRATCH_REG2, GP_SCRATCH_REG);
4331                                         ia64_movl (code, GP_SCRATCH_REG, (i * 8));
4332                                         ia64_add (code, GP_SCRATCH_REG, cfg->arch.reg_out0 + 1, GP_SCRATCH_REG);
4333                                         ia64_st8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG2);
4334                                 }
4335                         }
4336                 }
4337                 else
4338                         ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_R0);
4339         }
4340         else
4341                 ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_R0);
4342
4343         add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, cfg->method);
4344         ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
4345
4346         code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
4347
4348         if (enable_arguments && stack_area) {
4349                 ia64_movl (code, GP_SCRATCH_REG, stack_area);
4350
4351                 ia64_add (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
4352
4353                 ia64_adds_imm (code, IA64_SP, 16, IA64_SP);
4354         }
4355
4356         ia64_codegen_close (code);
4357
4358         return code.buf;
4359 }
4360
4361 void*
4362 mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
4363 {
4364         Ia64CodegenState code;
4365         CallInfo *cinfo = NULL;
4366         MonoMethod *method = cfg->method;
4367         MonoMethodSignature *sig = mono_method_signature (cfg->method);
4368
4369         ia64_codegen_init (code, p);
4370
4371         cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
4372
4373         /* Save return value + pass it to func */
4374         switch (cinfo->ret.storage) {
4375         case ArgNone:
4376                 break;
4377         case ArgInIReg:
4378                 ia64_mov (code, cfg->arch.reg_saved_return_val, cinfo->ret.reg);
4379                 ia64_mov (code, cfg->arch.reg_out0 + 1, cinfo->ret.reg);
4380                 break;
4381         case ArgInFloatReg:
4382                 ia64_adds_imm (code, IA64_SP, -16, IA64_SP);
4383                 ia64_adds_imm (code, GP_SCRATCH_REG, 16, IA64_SP);
4384                 ia64_stfd_hint (code, GP_SCRATCH_REG, cinfo->ret.reg, 0);
4385                 ia64_fmov (code, 8 + 1, cinfo->ret.reg);
4386                 break;
4387         case ArgValuetypeAddrInIReg:
4388                 ia64_mov (code, cfg->arch.reg_out0 + 1, cfg->arch.reg_in0 + cinfo->ret.reg);
4389                 break;
4390         case ArgAggregate:
4391                 NOT_IMPLEMENTED;
4392                 break;
4393         default:
4394                 break;
4395         }
4396
4397         add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, method);
4398         ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
4399         code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
4400
4401         /* Restore return value */
4402         switch (cinfo->ret.storage) {
4403         case ArgNone:
4404                 break;
4405         case ArgInIReg:
4406                 ia64_mov (code, cinfo->ret.reg, cfg->arch.reg_saved_return_val);
4407                 break;
4408         case ArgInFloatReg:
4409                 ia64_adds_imm (code, GP_SCRATCH_REG, 16, IA64_SP);
4410                 ia64_ldfd (code, cinfo->ret.reg, GP_SCRATCH_REG);
4411                 break;
4412         case ArgValuetypeAddrInIReg:
4413                 break;
4414         case ArgAggregate:
4415                 break;
4416         default:
4417                 break;
4418         }
4419
4420         ia64_codegen_close (code);
4421
4422         return code.buf;
4423 }
4424
4425 void
4426 mono_arch_save_unwind_info (MonoCompile *cfg)
4427 {
4428         unw_dyn_info_t *di;
4429
4430         /* FIXME: Unregister this for dynamic methods */
4431
4432         di = g_malloc0 (sizeof (unw_dyn_info_t));
4433         di->start_ip = (unw_word_t) cfg->native_code;
4434         di->end_ip = (unw_word_t) cfg->native_code + cfg->code_len;
4435         di->gp = 0;
4436         di->format = UNW_INFO_FORMAT_DYNAMIC;
4437         di->u.pi.name_ptr = (unw_word_t)mono_method_full_name (cfg->method, TRUE);
4438         di->u.pi.regions = cfg->arch.r_pro;
4439
4440         _U_dyn_register (di);
4441
4442         /*
4443         {
4444                 unw_dyn_region_info_t *region = di->u.pi.regions;
4445
4446                 printf ("Unwind info for method %s:\n", mono_method_full_name (cfg->method, TRUE));
4447                 while (region) {
4448                         printf ("    [Region: %d]\n", region->insn_count);
4449                         region = region->next;
4450                 }
4451         }
4452         */
4453 }
4454
4455 void
4456 mono_arch_flush_icache (guint8 *code, gint size)
4457 {
4458         guint8* p = (guint8*)((guint64)code & ~(0x3f));
4459         guint8* end = (guint8*)((guint64)code + size);
4460
4461 #ifdef __INTEL_COMPILER
4462         /* icc doesn't define an fc.i instrinsic, but fc==fc.i on itanium 2 */
4463         while (p < end) {
4464                 __fc ((guint64)p);
4465                 p += 32;
4466         }
4467 #else
4468         while (p < end) {
4469                 __asm__ __volatile__ ("fc.i %0"::"r"(p));
4470                 /* FIXME: This could be increased to 128 on some cpus */
4471                 p += 32;
4472         }
4473 #endif
4474 }
4475
4476 void
4477 mono_arch_flush_register_windows (void)
4478 {
4479         /* Not needed because of libunwind */
4480 }
4481
4482 gboolean
4483 mono_arch_is_inst_imm (gint64 imm)
4484 {
4485         /* The lowering pass will take care of it */
4486
4487         return TRUE;
4488 }
4489
4490 /*
4491  * Determine whenever the trap whose info is in SIGINFO is caused by
4492  * integer overflow.
4493  */
4494 gboolean
4495 mono_arch_is_int_overflow (void *sigctx, void *info)
4496 {
4497         /* Division is emulated with explicit overflow checks */
4498         return FALSE;
4499 }
4500
4501 guint32
4502 mono_arch_get_patch_offset (guint8 *code)
4503 {
4504         NOT_IMPLEMENTED;
4505
4506         return 0;
4507 }
4508
4509 gpointer*
4510 mono_arch_get_delegate_method_ptr_addr (guint8* code, mgreg_t *regs)
4511 {
4512         NOT_IMPLEMENTED;
4513
4514         return NULL;
4515 }
4516
4517 void
4518 mono_arch_finish_init (void)
4519 {
4520 }
4521
4522 void
4523 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
4524 {
4525 }
4526
4527 /*
4528  * LOCKING: called with the domain lock held
4529  */
4530 gpointer
4531 mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count,
4532         gpointer fail_tramp)
4533 {
4534         int i;
4535         int size = 0;
4536         guint8 *start, *buf;
4537         Ia64CodegenState code;
4538
4539         size = count * 256;
4540         buf = g_malloc0 (size);
4541         ia64_codegen_init (code, buf);
4542
4543         /* IA64_R9 contains the IMT method */
4544
4545         for (i = 0; i < count; ++i) {
4546                 MonoIMTCheckItem *item = imt_entries [i];
4547                 ia64_begin_bundle (code);
4548                 item->code_target = (guint8*)code.buf + code.nins;
4549                 if (item->is_equals) {
4550                         gboolean fail_case = !item->check_target_idx && fail_tramp;
4551
4552                         if (item->check_target_idx || fail_case) {
4553                                 if (!item->compare_done || fail_case) {
4554                                         ia64_movl (code, GP_SCRATCH_REG, item->key);
4555                                         ia64_cmp_eq (code, 6, 7, IA64_R9, GP_SCRATCH_REG);
4556                                 }
4557                                 item->jmp_code = (guint8*)code.buf + code.nins;
4558                                 ia64_br_cond_pred (code, 7, 0);
4559
4560                                 if (item->has_target_code) {
4561                                         ia64_movl (code, GP_SCRATCH_REG, item->value.target_code);
4562                                 } else {
4563                                         ia64_movl (code, GP_SCRATCH_REG, &(vtable->vtable [item->value.vtable_slot]));
4564                                         ia64_ld8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG);
4565                                 }
4566                                 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
4567                                 ia64_br_cond_reg (code, IA64_B6);
4568
4569                                 if (fail_case) {
4570                                         ia64_begin_bundle (code);
4571                                         ia64_patch (item->jmp_code, (guint8*)code.buf + code.nins);
4572                                         ia64_movl (code, GP_SCRATCH_REG, fail_tramp);
4573                                         ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
4574                                         ia64_br_cond_reg (code, IA64_B6);
4575                                         item->jmp_code = NULL;
4576                                 }
4577                         } else {
4578                                 /* enable the commented code to assert on wrong method */
4579 #if ENABLE_WRONG_METHOD_CHECK
4580                                 g_assert_not_reached ();
4581 #endif
4582                                 ia64_movl (code, GP_SCRATCH_REG, &(vtable->vtable [item->value.vtable_slot]));
4583                                 ia64_ld8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG);
4584                                 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
4585                                 ia64_br_cond_reg (code, IA64_B6);
4586 #if ENABLE_WRONG_METHOD_CHECK
4587                                 g_assert_not_reached ();
4588 #endif
4589                         }
4590                 } else {
4591                         ia64_movl (code, GP_SCRATCH_REG, item->key);
4592                         ia64_cmp_geu (code, 6, 7, IA64_R9, GP_SCRATCH_REG);
4593                         item->jmp_code = (guint8*)code.buf + code.nins;
4594                         ia64_br_cond_pred (code, 6, 0);
4595                 }
4596         }
4597         /* patch the branches to get to the target items */
4598         for (i = 0; i < count; ++i) {
4599                 MonoIMTCheckItem *item = imt_entries [i];
4600                 if (item->jmp_code) {
4601                         if (item->check_target_idx) {
4602                                 ia64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
4603                         }
4604                 }
4605         }
4606
4607         ia64_codegen_close (code);
4608         g_assert (code.buf - buf <= size);
4609
4610         size = code.buf - buf;
4611         if (fail_tramp) {
4612                 start = mono_method_alloc_generic_virtual_thunk (domain, size + 16);
4613                 start = (gpointer)ALIGN_TO (start, 16);
4614         } else {
4615                 start = mono_domain_code_reserve (domain, size);
4616         }
4617         memcpy (start, buf, size);
4618
4619         mono_arch_flush_icache (start, size);
4620
4621         mono_stats.imt_thunks_size += size;
4622
4623         mono_tramp_info_register (mono_tramp_info_create (NULL, start, size, NULL, NULL), domain);
4624
4625         return start;
4626 }
4627
4628 MonoMethod*
4629 mono_arch_find_imt_method (mgreg_t *regs, guint8 *code)
4630 {
4631         return (MonoMethod*)regs [IA64_R9];
4632 }
4633
4634 gpointer
4635 mono_arch_get_this_arg_from_call (mgreg_t *regs, guint8 *code)
4636 {
4637         return (gpointer)regs [IA64_R10];
4638 }
4639
4640 gpointer
4641 mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
4642 {
4643         return NULL;
4644 }
4645
4646 gpointer
4647 mono_arch_get_delegate_virtual_invoke_impl (MonoMethodSignature *sig, MonoMethod *method, int offset, gboolean load_imt_reg)
4648 {
4649         return NULL;
4650 }
4651
4652 MonoInst*
4653 mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
4654 {
4655         MonoInst *ins = NULL;
4656
4657         if (cmethod->klass->image == mono_defaults.corlib &&
4658                 (strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
4659                 (strcmp (cmethod->klass->name, "Interlocked") == 0)) {
4660
4661                 /*
4662                  * We don't use the generic version in mini_emit_inst_for_method () since we
4663                  * ia64 has atomic_add_imm opcodes.
4664                  */
4665                 if (strcmp (cmethod->name, "Increment") == 0) {
4666                         guint32 opcode;
4667
4668                         if (fsig->params [0]->type == MONO_TYPE_I4)
4669                                 opcode = OP_ATOMIC_ADD_IMM_I4;
4670                         else if (fsig->params [0]->type == MONO_TYPE_I8)
4671                                 opcode = OP_ATOMIC_ADD_IMM_I8;
4672                         else
4673                                 g_assert_not_reached ();
4674                         MONO_INST_NEW (cfg, ins, opcode);
4675                         ins->dreg = mono_alloc_preg (cfg);
4676                         ins->inst_imm = 1;
4677                         ins->inst_basereg = args [0]->dreg;
4678                         ins->inst_offset = 0;
4679                         MONO_ADD_INS (cfg->cbb, ins);
4680                 } else if (strcmp (cmethod->name, "Decrement") == 0) {
4681                         guint32 opcode;
4682
4683                         if (fsig->params [0]->type == MONO_TYPE_I4)
4684                                 opcode = OP_ATOMIC_ADD_IMM_I4;
4685                         else if (fsig->params [0]->type == MONO_TYPE_I8)
4686                                 opcode = OP_ATOMIC_ADD_IMM_I8;
4687                         else
4688                                 g_assert_not_reached ();
4689                         MONO_INST_NEW (cfg, ins, opcode);
4690                         ins->dreg = mono_alloc_preg (cfg);
4691                         ins->inst_imm = -1;
4692                         ins->inst_basereg = args [0]->dreg;
4693                         ins->inst_offset = 0;
4694                         MONO_ADD_INS (cfg->cbb, ins);
4695                 } else if (strcmp (cmethod->name, "Add") == 0) {
4696                         guint32 opcode;
4697                         gboolean is_imm = FALSE;
4698                         gint64 imm = 0;
4699
4700                         if ((args [1]->opcode == OP_ICONST) || (args [1]->opcode == OP_I8CONST)) {
4701                                 imm = (args [1]->opcode == OP_ICONST) ? args [1]->inst_c0 : args [1]->inst_l;
4702
4703                                 is_imm = (imm == 1 || imm == 4 || imm == 8 || imm == 16 || imm == -1 || imm == -4 || imm == -8 || imm == -16);
4704                         }
4705
4706                         if (is_imm) {
4707                                 if (fsig->params [0]->type == MONO_TYPE_I4)
4708                                         opcode = OP_ATOMIC_ADD_IMM_I4;
4709                                 else if (fsig->params [0]->type == MONO_TYPE_I8)
4710                                         opcode = OP_ATOMIC_ADD_IMM_I8;
4711                                 else
4712                                         g_assert_not_reached ();
4713
4714                                 MONO_INST_NEW (cfg, ins, opcode);
4715                                 ins->dreg = mono_alloc_ireg (cfg);
4716                                 ins->inst_basereg = args [0]->dreg;
4717                                 ins->inst_offset = 0;
4718                                 ins->inst_imm = imm;
4719                                 ins->type = (opcode == OP_ATOMIC_ADD_IMM_I4) ? STACK_I4 : STACK_I8;
4720                         } else {
4721                                 if (fsig->params [0]->type == MONO_TYPE_I4)
4722                                         opcode = OP_ATOMIC_ADD_I4;
4723                                 else if (fsig->params [0]->type == MONO_TYPE_I8)
4724                                         opcode = OP_ATOMIC_ADD_I8;
4725                                 else
4726                                         g_assert_not_reached ();
4727
4728                                 MONO_INST_NEW (cfg, ins, opcode);
4729                                 ins->dreg = mono_alloc_ireg (cfg);
4730                                 ins->inst_basereg = args [0]->dreg;
4731                                 ins->inst_offset = 0;
4732                                 ins->sreg2 = args [1]->dreg;
4733                                 ins->type = (opcode == OP_ATOMIC_ADD_I4) ? STACK_I4 : STACK_I8;
4734                         }
4735                         MONO_ADD_INS (cfg->cbb, ins);
4736                 }
4737         }
4738
4739         return ins;
4740 }
4741
4742 gboolean
4743 mono_arch_print_tree (MonoInst *tree, int arity)
4744 {
4745         return 0;
4746 }
4747
4748 mgreg_t
4749 mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
4750 {
4751         /* FIXME: implement */
4752         g_assert_not_reached ();
4753 }
4754
4755 gboolean
4756 mono_arch_opcode_supported (int opcode)
4757 {
4758         switch (opcode) {
4759         case OP_ATOMIC_ADD_I4:
4760         case OP_ATOMIC_ADD_I8:
4761         case OP_ATOMIC_EXCHANGE_I4:
4762         case OP_ATOMIC_EXCHANGE_I8:
4763                 return TRUE;
4764         default:
4765                 return FALSE;
4766         }
4767 }