gcc/hsa-brig.c

   1 /* Producing binary form of HSA BRIG from our internal representation.
   2    Copyright (C) 2013-2018 Free Software Foundation, Inc.
   3    Contributed by Martin Jambor <mjambor@suse.cz> and
   4    Martin Liska <mliska@suse.cz>.
   5
   6 This file is part of GCC.
   7
   8 GCC is free software; you can redistribute it and/or modify
   9 it under the terms of the GNU General Public License as published by
  10 the Free Software Foundation; either version 3, or (at your option)
  11 any later version.
  12
  13 GCC is distributed in the hope that it will be useful,
  14 but WITHOUT ANY WARRANTY; without even the implied warranty of
  15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16 GNU General Public License for more details.
  17
  18 You should have received a copy of the GNU General Public License
  19 along with GCC; see the file COPYING3.  If not see
  20 <http://www.gnu.org/licenses/>.  */
  21
  22 #include "config.h"
  23 #include "system.h"
  24 #include "coretypes.h"
  25 #include "tm.h"
  26 #include "target.h"
  27 #include "memmodel.h"
  28 #include "tm_p.h"
  29 #include "is-a.h"
  30 #include "vec.h"
  31 #include "hash-table.h"
  32 #include "hash-map.h"
  33 #include "tree.h"
  34 #include "tree-iterator.h"
  35 #include "stor-layout.h"
  36 #include "output.h"
  37 #include "basic-block.h"
  38 #include "function.h"
  39 #include "cfg.h"
  40 #include "fold-const.h"
  41 #include "stringpool.h"
  42 #include "gimple-pretty-print.h"
  43 #include "diagnostic-core.h"
  44 #include "cgraph.h"
  45 #include "dumpfile.h"
  46 #include "print-tree.h"
  47 #include "symbol-summary.h"
  48 #include "hsa-common.h"
  49 #include "gomp-constants.h"
  50
  51 /* Convert VAL to little endian form, if necessary.  */
  52
  53 static uint16_t
  54 lendian16 (uint16_t val)
  55 {
  56 #if GCC_VERSION >= 4008
  57 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  58   return val;
  59 #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
  60   return __builtin_bswap16 (val);
  61 #else   /* __ORDER_PDP_ENDIAN__ */
  62   return val;
  63 #endif
  64 #else
  65 // provide a safe slower default, with shifts and masking
  66 #ifndef WORDS_BIGENDIAN
  67   return val;
  68 #else
  69   return (val >> 8) | (val << 8);
  70 #endif
  71 #endif
  72 }
  73
  74 /* Convert VAL to little endian form, if necessary.  */
  75
  76 static uint32_t
  77 lendian32 (uint32_t val)
  78 {
  79 #if GCC_VERSION >= 4006
  80 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  81   return val;
  82 #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
  83   return __builtin_bswap32 (val);
  84 #else  /* __ORDER_PDP_ENDIAN__ */
  85   return (val >> 16) | (val << 16);
  86 #endif
  87 #else
  88 // provide a safe slower default, with shifts and masking
  89 #ifndef WORDS_BIGENDIAN
  90   return val;
  91 #else
  92   val = ((val & 0xff00ff00) >> 8) | ((val & 0xff00ff) << 8);
  93   return (val >> 16) | (val << 16);
  94 #endif
  95 #endif
  96 }
  97
  98 /* Convert VAL to little endian form, if necessary.  */
  99
 100 static uint64_t
 101 lendian64 (uint64_t val)
 102 {
 103 #if GCC_VERSION >= 4006
 104 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
 105   return val;
 106 #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
 107   return __builtin_bswap64 (val);
 108 #else  /* __ORDER_PDP_ENDIAN__ */
 109   return (((val & 0xffffll) << 48)
 110           | ((val & 0xffff0000ll) << 16)
 111           | ((val & 0xffff00000000ll) >> 16)
 112           | ((val & 0xffff000000000000ll) >> 48));
 113 #endif
 114 #else
 115 // provide a safe slower default, with shifts and masking
 116 #ifndef WORDS_BIGENDIAN
 117   return val;
 118 #else
 119   val = (((val & 0xff00ff00ff00ff00ll) >> 8)
 120          | ((val & 0x00ff00ff00ff00ffll) << 8));
 121   val = ((( val & 0xffff0000ffff0000ll) >> 16)
 122          | (( val & 0x0000ffff0000ffffll) << 16));
 123   return (val >> 32) | (val << 32);
 124 #endif
 125 #endif
 126 }
 127
 128 #define BRIG_ELF_SECTION_NAME ".brig"
 129 #define BRIG_LABEL_STRING "hsa_brig"
 130 #define BRIG_SECTION_DATA_NAME    "hsa_data"
 131 #define BRIG_SECTION_CODE_NAME    "hsa_code"
 132 #define BRIG_SECTION_OPERAND_NAME "hsa_operand"
 133
 134 #define BRIG_CHUNK_MAX_SIZE (64 * 1024)
 135
 136 /* Required HSA section alignment.  */
 137
 138 #define HSA_SECTION_ALIGNMENT 16
 139
 140 /* Chunks of BRIG binary data.  */
 141
 142 struct hsa_brig_data_chunk
 143 {
 144   /* Size of the data already stored into a chunk.  */
 145   unsigned size;
 146
 147   /* Pointer to the data.  */
 148   char *data;
 149 };
 150
 151 /* Structure representing a BRIG section, holding and writing its data.  */
 152
 153 class hsa_brig_section
 154 {
 155 public:
 156   /* Section name that will be output to the BRIG.  */
 157   const char *section_name;
 158   /* Size in bytes of all data stored in the section.  */
 159   unsigned total_size;
 160   /* The size of the header of the section including padding.  */
 161   unsigned header_byte_count;
 162   /* The size of the header of the section without any padding.  */
 163   unsigned header_byte_delta;
 164
 165   void init (const char *name);
 166   void release ();
 167   void output ();
 168   unsigned add (const void *data, unsigned len, void **output = NULL);
 169   void round_size_up (int factor);
 170   void *get_ptr_by_offset (unsigned int offset);
 171
 172 private:
 173   void allocate_new_chunk ();
 174
 175   /* Buffers of binary data, each containing BRIG_CHUNK_MAX_SIZE bytes.  */
 176   vec <struct hsa_brig_data_chunk> chunks;
 177
 178   /* More convenient access to the last chunk from the vector above.  */
 179   struct hsa_brig_data_chunk *cur_chunk;
 180 };
 181
 182 static struct hsa_brig_section brig_data, brig_code, brig_operand;
 183 static uint32_t brig_insn_count;
 184 static bool brig_initialized = false;
 185
 186 /* Mapping between emitted HSA functions and their offset in code segment.  */
 187 static hash_map<tree, BrigCodeOffset32_t> *function_offsets;
 188
 189 /* Hash map of emitted function declarations.  */
 190 static hash_map <tree, BrigDirectiveExecutable *> *emitted_declarations;
 191
 192 /* Hash table of emitted internal function declaration offsets.  */
 193 hash_table <hsa_internal_fn_hasher> *hsa_emitted_internal_decls;
 194
 195 /* List of sbr instructions.  */
 196 static vec <hsa_insn_sbr *> *switch_instructions;
 197
 198 struct function_linkage_pair
 199 {
 200   function_linkage_pair (tree decl, unsigned int off)
 201     : function_decl (decl), offset (off) {}
 202
 203   /* Declaration of called function.  */
 204   tree function_decl;
 205
 206   /* Offset in operand section.  */
 207   unsigned int offset;
 208 };
 209
 210 /* Vector of function calls where we need to resolve function offsets.  */
 211 static auto_vec <function_linkage_pair> function_call_linkage;
 212
 213 /* Add a new chunk, allocate data for it and initialize it.  */
 214
 215 void
 216 hsa_brig_section::allocate_new_chunk ()
 217 {
 218   struct hsa_brig_data_chunk new_chunk;
 219
 220   new_chunk.data = XCNEWVEC (char, BRIG_CHUNK_MAX_SIZE);
 221   new_chunk.size = 0;
 222   cur_chunk = chunks.safe_push (new_chunk);
 223 }
 224
 225 /* Initialize the brig section.  */
 226
 227 void
 228 hsa_brig_section::init (const char *name)
 229 {
 230   section_name = name;
 231   /* While the following computation is basically wrong, because the intent
 232      certainly wasn't to have the first character of name and padding, which
 233      are a part of sizeof (BrigSectionHeader), included in the first addend,
 234      this is what the disassembler expects.  */
 235   total_size = sizeof (BrigSectionHeader) + strlen (section_name);
 236   chunks.create (1);
 237   allocate_new_chunk ();
 238   header_byte_delta = total_size;
 239   round_size_up (4);
 240   header_byte_count = total_size;
 241 }
 242
 243 /* Free all data in the section.  */
 244
 245 void
 246 hsa_brig_section::release ()
 247 {
 248   for (unsigned i = 0; i < chunks.length (); i++)
 249     free (chunks[i].data);
 250   chunks.release ();
 251   cur_chunk = NULL;
 252 }
 253
 254 /* Write the section to the output file to a section with the name given at
 255    initialization.  Switches the output section and does not restore it.  */
 256
 257 void
 258 hsa_brig_section::output ()
 259 {
 260   struct BrigSectionHeader section_header;
 261   char padding[8];
 262
 263   section_header.byteCount = lendian64 (total_size);
 264   section_header.headerByteCount = lendian32 (header_byte_count);
 265   section_header.nameLength = lendian32 (strlen (section_name));
 266   assemble_string ((const char *) &section_header, 16);
 267   assemble_string (section_name, (section_header.nameLength));
 268   memset (&padding, 0, sizeof (padding));
 269   /* This is also a consequence of the wrong header size computation described
 270      in a comment in hsa_brig_section::init.  */
 271   assemble_string (padding, 8);
 272   for (unsigned i = 0; i < chunks.length (); i++)
 273     assemble_string (chunks[i].data, chunks[i].size);
 274 }
 275
 276 /* Add to the stream LEN bytes of opaque binary DATA.  Return the offset at
 277    which it was stored.  If OUTPUT is not NULL, store into it the pointer to
 278    the place where DATA was actually stored.  */
 279
 280 unsigned
 281 hsa_brig_section::add (const void *data, unsigned len, void **output)
 282 {
 283   unsigned offset = total_size;
 284
 285   gcc_assert (len <= BRIG_CHUNK_MAX_SIZE);
 286   if (cur_chunk->size > (BRIG_CHUNK_MAX_SIZE - len))
 287     allocate_new_chunk ();
 288
 289   char *dst = cur_chunk->data + cur_chunk->size;
 290   memcpy (dst, data, len);
 291   if (output)
 292     *output = dst;
 293   cur_chunk->size += len;
 294   total_size += len;
 295
 296   return offset;
 297 }
 298
 299 /* Add padding to section so that its size is divisible by FACTOR.  */
 300
 301 void
 302 hsa_brig_section::round_size_up (int factor)
 303 {
 304   unsigned padding, res = total_size % factor;
 305
 306   if (res == 0)
 307     return;
 308
 309   padding = factor - res;
 310   total_size += padding;
 311   if (cur_chunk->size > (BRIG_CHUNK_MAX_SIZE - padding))
 312     {
 313       padding -= BRIG_CHUNK_MAX_SIZE - cur_chunk->size;
 314       cur_chunk->size = BRIG_CHUNK_MAX_SIZE;
 315       allocate_new_chunk ();
 316     }
 317
 318   cur_chunk->size += padding;
 319 }
 320
 321 /* Return pointer to data by global OFFSET in the section.  */
 322
 323 void *
 324 hsa_brig_section::get_ptr_by_offset (unsigned int offset)
 325 {
 326   gcc_assert (offset < total_size);
 327   offset -= header_byte_delta;
 328
 329   unsigned i;
 330   for (i = 0; offset >= chunks[i].size; i++)
 331     offset -= chunks[i].size;
 332
 333   return chunks[i].data + offset;
 334 }
 335
 336 /* BRIG string data hashing.  */
 337
 338 struct brig_string_slot
 339 {
 340   const char *s;
 341   char prefix;
 342   int len;
 343   uint32_t offset;
 344 };
 345
 346 /* Hash table helpers.  */
 347
 348 struct brig_string_slot_hasher : pointer_hash <brig_string_slot>
 349 {
 350   static inline hashval_t hash (const value_type);
 351   static inline bool equal (const value_type, const compare_type);
 352   static inline void remove (value_type);
 353 };
 354
 355 /* Returns a hash code for DS.  Adapted from libiberty's htab_hash_string
 356    to support strings that may not end in '\0'.  */
 357
 358 inline hashval_t
 359 brig_string_slot_hasher::hash (const value_type ds)
 360 {
 361   hashval_t r = ds->len;
 362   int i;
 363
 364   for (i = 0; i < ds->len; i++)
 365      r = r * 67 + (unsigned) ds->s[i] - 113;
 366   r = r * 67 + (unsigned) ds->prefix - 113;
 367   return r;
 368 }
 369
 370 /* Returns nonzero if DS1 and DS2 are equal.  */
 371
 372 inline bool
 373 brig_string_slot_hasher::equal (const value_type ds1, const compare_type ds2)
 374 {
 375   if (ds1->len == ds2->len)
 376     return ds1->prefix == ds2->prefix
 377       && memcmp (ds1->s, ds2->s, ds1->len) == 0;
 378
 379   return 0;
 380 }
 381
 382 /* Deallocate memory for DS upon its removal.  */
 383
 384 inline void
 385 brig_string_slot_hasher::remove (value_type ds)
 386 {
 387   free (const_cast<char *> (ds->s));
 388   free (ds);
 389 }
 390
 391 /* Hash for strings we output in order not to duplicate them needlessly.  */
 392
 393 static hash_table<brig_string_slot_hasher> *brig_string_htab;
 394
 395 /* Emit a null terminated string STR to the data section and return its
 396    offset in it.  If PREFIX is non-zero, output it just before STR too.
 397    Sanitize the string if SANITIZE option is set to true.  */
 398
 399 static unsigned
 400 brig_emit_string (const char *str, char prefix = 0, bool sanitize = true)
 401 {
 402   unsigned slen = strlen (str);
 403   unsigned offset, len = slen + (prefix ? 1 : 0);
 404   uint32_t hdr_len = lendian32 (len);
 405   brig_string_slot s_slot;
 406   brig_string_slot **slot;
 407   char *str2;
 408
 409   str2 = xstrdup (str);
 410
 411   if (sanitize)
 412     hsa_sanitize_name (str2);
 413   s_slot.s = str2;
 414   s_slot.len = slen;
 415   s_slot.prefix = prefix;
 416   s_slot.offset = 0;
 417
 418   slot = brig_string_htab->find_slot (&s_slot, INSERT);
 419   if (*slot == NULL)
 420     {
 421       brig_string_slot *new_slot = XCNEW (brig_string_slot);
 422
 423       /* In theory we should fill in BrigData but that would mean copying
 424          the string to a buffer for no reason, so we just emulate it.  */
 425       offset = brig_data.add (&hdr_len, sizeof (hdr_len));
 426       if (prefix)
 427         brig_data.add (&prefix, 1);
 428
 429       brig_data.add (str2, slen);
 430       brig_data.round_size_up (4);
 431
 432       /* TODO: could use the string we just copied into
 433          brig_string->cur_chunk */
 434       new_slot->s = str2;
 435       new_slot->len = slen;
 436       new_slot->prefix = prefix;
 437       new_slot->offset = offset;
 438       *slot = new_slot;
 439     }
 440   else
 441     {
 442       offset = (*slot)->offset;
 443       free (str2);
 444     }
 445
 446   return offset;
 447 }
 448
 449 /* Linked list of queued operands.  */
 450
 451 static struct operand_queue
 452 {
 453   /* First from the chain of queued operands.  */
 454   hsa_op_base *first_op, *last_op;
 455
 456   /* The offset at which the next operand will be enqueued.  */
 457   unsigned projected_size;
 458
 459 } op_queue;
 460
 461 /* Unless already initialized, initialize infrastructure to produce BRIG.  */
 462
 463 static void
 464 brig_init (void)
 465 {
 466   brig_insn_count = 0;
 467
 468   if (brig_initialized)
 469     return;
 470
 471   brig_string_htab = new hash_table<brig_string_slot_hasher> (37);
 472   brig_data.init (BRIG_SECTION_DATA_NAME);
 473   brig_code.init (BRIG_SECTION_CODE_NAME);
 474   brig_operand.init (BRIG_SECTION_OPERAND_NAME);
 475   brig_initialized = true;
 476
 477   struct BrigDirectiveModule moddir;
 478   memset (&moddir, 0, sizeof (moddir));
 479   moddir.base.byteCount = lendian16 (sizeof (moddir));
 480
 481   char *modname;
 482   if (main_input_filename && *main_input_filename != '\0')
 483     {
 484       const char *part = strrchr (main_input_filename, '/');
 485       if (!part)
 486         part = main_input_filename;
 487       else
 488         part++;
 489       modname = concat ("&__hsa_module_", part, NULL);
 490       char *extension = strchr (modname, '.');
 491       if (extension)
 492         *extension = '\0';
 493
 494       /* As in LTO mode, we have to emit a different module names.  */
 495       if (flag_ltrans)
 496         {
 497           part = strrchr (asm_file_name, '/');
 498           if (!part)
 499             part = asm_file_name;
 500           else
 501             part++;
 502           char *modname2;
 503           modname2 = xasprintf ("%s_%s", modname, part);
 504           free (modname);
 505           modname = modname2;
 506         }
 507
 508       hsa_sanitize_name (modname);
 509       moddir.name = brig_emit_string (modname);
 510       free (modname);
 511     }
 512   else
 513     moddir.name = brig_emit_string ("__hsa_module_unnamed", '&');
 514   moddir.base.kind = lendian16 (BRIG_KIND_DIRECTIVE_MODULE);
 515   moddir.hsailMajor = lendian32 (BRIG_VERSION_HSAIL_MAJOR);
 516   moddir.hsailMinor = lendian32 (BRIG_VERSION_HSAIL_MINOR);
 517   moddir.profile = hsa_full_profile_p () ? BRIG_PROFILE_FULL: BRIG_PROFILE_BASE;
 518   if (hsa_machine_large_p ())
 519     moddir.machineModel = BRIG_MACHINE_LARGE;
 520   else
 521     moddir.machineModel = BRIG_MACHINE_SMALL;
 522   moddir.defaultFloatRound = BRIG_ROUND_FLOAT_DEFAULT;
 523   brig_code.add (&moddir, sizeof (moddir));
 524 }
 525
 526 /* Free all BRIG data.  */
 527
 528 static void
 529 brig_release_data (void)
 530 {
 531   delete brig_string_htab;
 532   brig_data.release ();
 533   brig_code.release ();
 534   brig_operand.release ();
 535
 536   brig_initialized = 0;
 537 }
 538
 539 /* Enqueue operation OP.  Return the offset at which it will be stored.  */
 540
 541 static unsigned int
 542 enqueue_op (hsa_op_base *op)
 543 {
 544   unsigned ret;
 545
 546   if (op->m_brig_op_offset)
 547     return op->m_brig_op_offset;
 548
 549   ret = op_queue.projected_size;
 550   op->m_brig_op_offset = op_queue.projected_size;
 551
 552   if (!op_queue.first_op)
 553     op_queue.first_op = op;
 554   else
 555     op_queue.last_op->m_next = op;
 556   op_queue.last_op = op;
 557
 558   if (is_a <hsa_op_immed *> (op))
 559     op_queue.projected_size += sizeof (struct BrigOperandConstantBytes);
 560   else if (is_a <hsa_op_reg *> (op))
 561     op_queue.projected_size += sizeof (struct BrigOperandRegister);
 562   else if (is_a <hsa_op_address *> (op))
 563     op_queue.projected_size += sizeof (struct BrigOperandAddress);
 564   else if (is_a <hsa_op_code_ref *> (op))
 565     op_queue.projected_size += sizeof (struct BrigOperandCodeRef);
 566   else if (is_a <hsa_op_code_list *> (op))
 567     op_queue.projected_size += sizeof (struct BrigOperandCodeList);
 568   else if (is_a <hsa_op_operand_list *> (op))
 569     op_queue.projected_size += sizeof (struct BrigOperandOperandList);
 570   else
 571     gcc_unreachable ();
 572   return ret;
 573 }
 574
 575 static void emit_immediate_operand (hsa_op_immed *imm);
 576
 577 /* Emit directive describing a symbol if it has not been emitted already.
 578    Return the offset of the directive.  */
 579
 580 static unsigned
 581 emit_directive_variable (struct hsa_symbol *symbol)
 582 {
 583   struct BrigDirectiveVariable dirvar;
 584   unsigned name_offset;
 585   static unsigned res_name_offset;
 586
 587   if (symbol->m_directive_offset)
 588     return symbol->m_directive_offset;
 589
 590   memset (&dirvar, 0, sizeof (dirvar));
 591   dirvar.base.byteCount = lendian16 (sizeof (dirvar));
 592   dirvar.base.kind = lendian16 (BRIG_KIND_DIRECTIVE_VARIABLE);
 593   dirvar.allocation = symbol->m_allocation;
 594
 595   char prefix = symbol->m_global_scope_p ? '&' : '%';
 596
 597   if (symbol->m_decl && TREE_CODE (symbol->m_decl) == RESULT_DECL)
 598     {
 599       if (res_name_offset == 0)
 600         res_name_offset = brig_emit_string (symbol->m_name, '%');
 601       name_offset = res_name_offset;
 602     }
 603   else if (symbol->m_name)
 604     name_offset = brig_emit_string (symbol->m_name, prefix);
 605   else
 606     {
 607       char buf[64];
 608       snprintf (buf, 64, "__%s_%i", hsa_seg_name (symbol->m_segment),
 609                 symbol->m_name_number);
 610       name_offset = brig_emit_string (buf, prefix);
 611     }
 612
 613   dirvar.name = lendian32 (name_offset);
 614
 615   if (symbol->m_decl && TREE_CODE (symbol->m_decl) == CONST_DECL)
 616     {
 617       hsa_op_immed *tmp = new hsa_op_immed (DECL_INITIAL (symbol->m_decl));
 618       dirvar.init = lendian32 (enqueue_op (tmp));
 619     }
 620   else
 621     dirvar.init = 0;
 622   dirvar.type = lendian16 (symbol->m_type);
 623   dirvar.segment = symbol->m_segment;
 624   dirvar.align = symbol->m_align;
 625   dirvar.linkage = symbol->m_linkage;
 626   dirvar.dim.lo = symbol->m_dim;
 627   dirvar.dim.hi = symbol->m_dim >> 32;
 628
 629   /* Global variables are just declared and linked via HSA runtime.  */
 630   if (symbol->m_linkage != BRIG_ALLOCATION_PROGRAM)
 631     dirvar.modifier |= BRIG_VARIABLE_DEFINITION;
 632   dirvar.reserved = 0;
 633
 634   if (symbol->m_cst_value)
 635     {
 636       dirvar.modifier |= BRIG_VARIABLE_CONST;
 637       dirvar.init = lendian32 (enqueue_op (symbol->m_cst_value));
 638     }
 639
 640   symbol->m_directive_offset = brig_code.add (&dirvar, sizeof (dirvar));
 641   return symbol->m_directive_offset;
 642 }
 643
 644 /* Emit directives describing either a function declaration or definition F and
 645    return the produced BrigDirectiveExecutable structure.  The function does
 646    not take into account any instructions when calculating nextModuleEntry
 647    field of the produced BrigDirectiveExecutable structure so when emitting
 648    actual definitions, this field needs to be updated after all of the function
 649    is actually added to the code section.  */
 650
 651 static BrigDirectiveExecutable *
 652 emit_function_directives (hsa_function_representation *f, bool is_declaration)
 653 {
 654   struct BrigDirectiveExecutable fndir;
 655   unsigned name_offset, inarg_off, scoped_off, next_toplev_off;
 656   int count = 0;
 657   void *ptr_to_fndir;
 658   hsa_symbol *sym;
 659
 660   if (!f->m_declaration_p)
 661     for (int i = 0; f->m_global_symbols.iterate (i, &sym); i++)
 662       {
 663         gcc_assert (!sym->m_emitted_to_brig);
 664         sym->m_emitted_to_brig = true;
 665         emit_directive_variable (sym);
 666         brig_insn_count++;
 667       }
 668
 669   name_offset = brig_emit_string (f->m_name, '&');
 670   inarg_off = brig_code.total_size + sizeof (fndir)
 671     + (f->m_output_arg ? sizeof (struct BrigDirectiveVariable) : 0);
 672   scoped_off = inarg_off
 673     + f->m_input_args.length () * sizeof (struct BrigDirectiveVariable);
 674
 675   if (!f->m_declaration_p)
 676     {
 677       count += f->m_spill_symbols.length ();
 678       count += f->m_private_variables.length ();
 679     }
 680
 681   next_toplev_off = scoped_off + count * sizeof (struct BrigDirectiveVariable);
 682
 683   memset (&fndir, 0, sizeof (fndir));
 684   fndir.base.byteCount = lendian16 (sizeof (fndir));
 685   fndir.base.kind = lendian16 (f->m_kern_p ? BRIG_KIND_DIRECTIVE_KERNEL
 686                                : BRIG_KIND_DIRECTIVE_FUNCTION);
 687   fndir.name = lendian32 (name_offset);
 688   fndir.inArgCount = lendian16 (f->m_input_args.length ());
 689   fndir.outArgCount = lendian16 (f->m_output_arg ? 1 : 0);
 690   fndir.firstInArg = lendian32 (inarg_off);
 691   fndir.firstCodeBlockEntry = lendian32 (scoped_off);
 692   fndir.nextModuleEntry = lendian32 (next_toplev_off);
 693   fndir.linkage = f->get_linkage ();
 694   if (!f->m_declaration_p)
 695     fndir.modifier |= BRIG_EXECUTABLE_DEFINITION;
 696   memset (&fndir.reserved, 0, sizeof (fndir.reserved));
 697
 698   /* Once we put a definition of function_offsets, we should not overwrite
 699      it with a declaration of the function.  */
 700   if (f->m_internal_fn == NULL)
 701     {
 702       if (!function_offsets->get (f->m_decl) || !is_declaration)
 703         function_offsets->put (f->m_decl, brig_code.total_size);
 704     }
 705   else
 706     {
 707       /* Internal function.  */
 708       hsa_internal_fn **slot
 709         = hsa_emitted_internal_decls->find_slot (f->m_internal_fn, INSERT);
 710       hsa_internal_fn *int_fn = new hsa_internal_fn (f->m_internal_fn);
 711       int_fn->m_offset = brig_code.total_size;
 712       *slot = int_fn;
 713     }
 714
 715   brig_code.add (&fndir, sizeof (fndir), &ptr_to_fndir);
 716
 717   if (f->m_output_arg)
 718     emit_directive_variable (f->m_output_arg);
 719   for (unsigned i = 0; i < f->m_input_args.length (); i++)
 720     emit_directive_variable (f->m_input_args[i]);
 721
 722   if (!f->m_declaration_p)
 723     {
 724       for (int i = 0; f->m_spill_symbols.iterate (i, &sym); i++)
 725         {
 726           emit_directive_variable (sym);
 727           brig_insn_count++;
 728         }
 729       for (unsigned i = 0; i < f->m_private_variables.length (); i++)
 730         {
 731           emit_directive_variable (f->m_private_variables[i]);
 732           brig_insn_count++;
 733         }
 734     }
 735
 736   return (BrigDirectiveExecutable *) ptr_to_fndir;
 737 }
 738
 739 /* Emit a label directive for the given HBB.  We assume it is about to start on
 740    the current offset in the code section.  */
 741
 742 static void
 743 emit_bb_label_directive (hsa_bb *hbb)
 744 {
 745   struct BrigDirectiveLabel lbldir;
 746
 747   lbldir.base.byteCount = lendian16 (sizeof (lbldir));
 748   lbldir.base.kind = lendian16 (BRIG_KIND_DIRECTIVE_LABEL);
 749   char buf[32];
 750   snprintf (buf, 32, "BB_%u_%i", DECL_UID (current_function_decl),
 751             hbb->m_index);
 752   lbldir.name = lendian32 (brig_emit_string (buf, '@'));
 753
 754   hbb->m_label_ref.m_directive_offset = brig_code.add (&lbldir,
 755                                                        sizeof (lbldir));
 756   brig_insn_count++;
 757 }
 758
 759 /* Map a normal HSAIL type to the type of the equivalent BRIG operand
 760    holding such, for constants and registers.  */
 761
 762 static BrigType16_t
 763 regtype_for_type (BrigType16_t t)
 764 {
 765   switch (t)
 766     {
 767     case BRIG_TYPE_B1:
 768       return BRIG_TYPE_B1;
 769
 770     case BRIG_TYPE_U8:
 771     case BRIG_TYPE_U16:
 772     case BRIG_TYPE_U32:
 773     case BRIG_TYPE_S8:
 774     case BRIG_TYPE_S16:
 775     case BRIG_TYPE_S32:
 776     case BRIG_TYPE_B8:
 777     case BRIG_TYPE_B16:
 778     case BRIG_TYPE_B32:
 779     case BRIG_TYPE_F16:
 780     case BRIG_TYPE_F32:
 781     case BRIG_TYPE_U8X4:
 782     case BRIG_TYPE_U16X2:
 783     case BRIG_TYPE_S8X4:
 784     case BRIG_TYPE_S16X2:
 785     case BRIG_TYPE_F16X2:
 786       return BRIG_TYPE_B32;
 787
 788     case BRIG_TYPE_U64:
 789     case BRIG_TYPE_S64:
 790     case BRIG_TYPE_F64:
 791     case BRIG_TYPE_B64:
 792     case BRIG_TYPE_U8X8:
 793     case BRIG_TYPE_U16X4:
 794     case BRIG_TYPE_U32X2:
 795     case BRIG_TYPE_S8X8:
 796     case BRIG_TYPE_S16X4:
 797     case BRIG_TYPE_S32X2:
 798     case BRIG_TYPE_F16X4:
 799     case BRIG_TYPE_F32X2:
 800       return BRIG_TYPE_B64;
 801
 802     case BRIG_TYPE_B128:
 803     case BRIG_TYPE_U8X16:
 804     case BRIG_TYPE_U16X8:
 805     case BRIG_TYPE_U32X4:
 806     case BRIG_TYPE_U64X2:
 807     case BRIG_TYPE_S8X16:
 808     case BRIG_TYPE_S16X8:
 809     case BRIG_TYPE_S32X4:
 810     case BRIG_TYPE_S64X2:
 811     case BRIG_TYPE_F16X8:
 812     case BRIG_TYPE_F32X4:
 813     case BRIG_TYPE_F64X2:
 814       return BRIG_TYPE_B128;
 815
 816     default:
 817       gcc_unreachable ();
 818     }
 819 }
 820
 821 /* Return the length of the BRIG type TYPE that is going to be streamed out as
 822    an immediate constant (so it must not be B1).  */
 823
 824 unsigned
 825 hsa_get_imm_brig_type_len (BrigType16_t type)
 826 {
 827   BrigType16_t base_type = type & BRIG_TYPE_BASE_MASK;
 828   BrigType16_t pack_type = type & BRIG_TYPE_PACK_MASK;
 829
 830   switch (pack_type)
 831     {
 832     case BRIG_TYPE_PACK_NONE:
 833       break;
 834     case BRIG_TYPE_PACK_32:
 835       return 4;
 836     case BRIG_TYPE_PACK_64:
 837       return 8;
 838     case BRIG_TYPE_PACK_128:
 839       return 16;
 840     default:
 841       gcc_unreachable ();
 842     }
 843
 844   switch (base_type)
 845     {
 846     case BRIG_TYPE_U8:
 847     case BRIG_TYPE_S8:
 848     case BRIG_TYPE_B8:
 849       return 1;
 850     case BRIG_TYPE_U16:
 851     case BRIG_TYPE_S16:
 852     case BRIG_TYPE_F16:
 853     case BRIG_TYPE_B16:
 854       return 2;
 855     case BRIG_TYPE_U32:
 856     case BRIG_TYPE_S32:
 857     case BRIG_TYPE_F32:
 858     case BRIG_TYPE_B32:
 859       return 4;
 860     case BRIG_TYPE_U64:
 861     case BRIG_TYPE_S64:
 862     case BRIG_TYPE_F64:
 863     case BRIG_TYPE_B64:
 864       return 8;
 865     case BRIG_TYPE_B128:
 866       return 16;
 867     default:
 868       gcc_unreachable ();
 869     }
 870 }
 871
 872 /* Emit one scalar VALUE to the buffer DATA intended for BRIG emission.
 873    If NEED_LEN is not equal to zero, shrink or extend the value
 874    to NEED_LEN bytes.  Return how many bytes were written.  */
 875
 876 static int
 877 emit_immediate_scalar_to_buffer (tree value, char *data, unsigned need_len)
 878 {
 879   union hsa_bytes bytes;
 880
 881   memset (&bytes, 0, sizeof (bytes));
 882   tree type = TREE_TYPE (value);
 883   gcc_checking_assert (TREE_CODE (type) != VECTOR_TYPE);
 884
 885   unsigned data_len = tree_to_uhwi (TYPE_SIZE (type)) / BITS_PER_UNIT;
 886   if (INTEGRAL_TYPE_P (type)
 887       || (POINTER_TYPE_P (type) && TREE_CODE (value) == INTEGER_CST))
 888     switch (data_len)
 889       {
 890       case 1:
 891         bytes.b8 = (uint8_t) TREE_INT_CST_LOW (value);
 892         break;
 893       case 2:
 894         bytes.b16 = (uint16_t) TREE_INT_CST_LOW (value);
 895         break;
 896       case 4:
 897         bytes.b32 = (uint32_t) TREE_INT_CST_LOW (value);
 898         break;
 899       case 8:
 900         bytes.b64 = (uint64_t) TREE_INT_CST_LOW (value);
 901         break;
 902       default:
 903         gcc_unreachable ();
 904       }
 905   else if (SCALAR_FLOAT_TYPE_P (type))
 906     {
 907       if (data_len == 2)
 908         {
 909           sorry ("Support for HSA does not implement immediate 16 bit FPU "
 910                  "operands");
 911           return 2;
 912         }
 913       unsigned int_len = GET_MODE_SIZE (SCALAR_FLOAT_TYPE_MODE (type));
 914       /* There are always 32 bits in each long, no matter the size of
 915          the hosts long.  */
 916       long tmp[6];
 917
 918       real_to_target (tmp, TREE_REAL_CST_PTR (value), TYPE_MODE (type));
 919
 920       if (int_len == 4)
 921         bytes.b32 = (uint32_t) tmp[0];
 922       else
 923         {
 924           bytes.b64 = (uint64_t)(uint32_t) tmp[1];
 925           bytes.b64 <<= 32;
 926           bytes.b64 |= (uint32_t) tmp[0];
 927         }
 928     }
 929   else
 930     gcc_unreachable ();
 931
 932   int len;
 933   if (need_len == 0)
 934     len = data_len;
 935   else
 936     len = need_len;
 937
 938   memcpy (data, &bytes, len);
 939   return len;
 940 }
 941
 942 char *
 943 hsa_op_immed::emit_to_buffer (unsigned *brig_repr_size)
 944 {
 945   char *brig_repr;
 946   *brig_repr_size = hsa_get_imm_brig_type_len (m_type);
 947
 948   if (m_tree_value != NULL_TREE)
 949     {
 950       /* Update brig_repr_size for special tree values.  */
 951       if (TREE_CODE (m_tree_value) == STRING_CST)
 952         *brig_repr_size = TREE_STRING_LENGTH (m_tree_value);
 953       else if (TREE_CODE (m_tree_value) == CONSTRUCTOR)
 954         *brig_repr_size
 955           = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (m_tree_value)));
 956
 957       unsigned total_len = *brig_repr_size;
 958
 959       /* As we can have a constructor with fewer elements, fill the memory
 960          with zeros.  */
 961       brig_repr = XCNEWVEC (char, total_len);
 962       char *p = brig_repr;
 963
 964       if (TREE_CODE (m_tree_value) == VECTOR_CST)
 965         {
 966           /* Variable-length vectors aren't supported.  */
 967           int i, num = VECTOR_CST_NELTS (m_tree_value).to_constant ();
 968           for (i = 0; i < num; i++)
 969             {
 970               tree v = VECTOR_CST_ELT (m_tree_value, i);
 971               unsigned actual = emit_immediate_scalar_to_buffer (v, p, 0);
 972               total_len -= actual;
 973               p += actual;
 974             }
 975           /* Vectors should have the exact size.  */
 976           gcc_assert (total_len == 0);
 977         }
 978       else if (TREE_CODE (m_tree_value) == STRING_CST)
 979         memcpy (brig_repr, TREE_STRING_POINTER (m_tree_value),
 980                 TREE_STRING_LENGTH (m_tree_value));
 981       else if (TREE_CODE (m_tree_value) == COMPLEX_CST)
 982         {
 983           gcc_assert (total_len % 2 == 0);
 984           unsigned actual;
 985           actual
 986             = emit_immediate_scalar_to_buffer (TREE_REALPART (m_tree_value), p,
 987                                                total_len / 2);
 988
 989           gcc_assert (actual == total_len / 2);
 990           p += actual;
 991
 992           actual
 993             = emit_immediate_scalar_to_buffer (TREE_IMAGPART (m_tree_value), p,
 994                                                total_len / 2);
 995           gcc_assert (actual == total_len / 2);
 996         }
 997       else if (TREE_CODE (m_tree_value) == CONSTRUCTOR)
 998         {
 999           unsigned len = CONSTRUCTOR_NELTS (m_tree_value);
1000           for (unsigned i = 0; i < len; i++)
1001             {
1002               tree v = CONSTRUCTOR_ELT (m_tree_value, i)->value;
1003               unsigned actual = emit_immediate_scalar_to_buffer (v, p, 0);
1004               total_len -= actual;
1005               p += actual;
1006             }
1007         }
1008       else
1009         emit_immediate_scalar_to_buffer (m_tree_value, p, total_len);
1010     }
1011   else
1012     {
1013       hsa_bytes bytes;
1014
1015       switch (*brig_repr_size)
1016         {
1017         case 1:
1018           bytes.b8 = (uint8_t) m_int_value;
1019           break;
1020         case 2:
1021           bytes.b16 = (uint16_t) m_int_value;
1022           break;
1023         case 4:
1024           bytes.b32 = (uint32_t) m_int_value;
1025           break;
1026         case 8:
1027           bytes.b64 = (uint64_t) m_int_value;
1028           break;
1029         default:
1030           gcc_unreachable ();
1031         }
1032
1033       brig_repr = XNEWVEC (char, *brig_repr_size);
1034       memcpy (brig_repr, &bytes, *brig_repr_size);
1035     }
1036
1037   return brig_repr;
1038 }
1039
1040 /* Emit an immediate BRIG operand IMM.  The BRIG type of the immediate might
1041    have been massaged to comply with various HSA/BRIG type requirements, so the
1042    only important aspect of that is the length (because HSAIL might expect
1043    smaller constants or become bit-data).  The data should be represented
1044    according to what is in the tree representation.  */
1045
1046 static void
1047 emit_immediate_operand (hsa_op_immed *imm)
1048 {
1049   unsigned brig_repr_size;
1050   char *brig_repr = imm->emit_to_buffer (&brig_repr_size);
1051   struct BrigOperandConstantBytes out;
1052
1053   memset (&out, 0, sizeof (out));
1054   out.base.byteCount = lendian16 (sizeof (out));
1055   out.base.kind = lendian16 (BRIG_KIND_OPERAND_CONSTANT_BYTES);
1056   uint32_t byteCount = lendian32 (brig_repr_size);
1057   out.type = lendian16 (imm->m_type);
1058   out.bytes = lendian32 (brig_data.add (&byteCount, sizeof (byteCount)));
1059   brig_operand.add (&out, sizeof (out));
1060   brig_data.add (brig_repr, brig_repr_size);
1061   brig_data.round_size_up (4);
1062
1063   free (brig_repr);
1064 }
1065
1066 /* Emit a register BRIG operand REG.  */
1067
1068 static void
1069 emit_register_operand (hsa_op_reg *reg)
1070 {
1071   struct BrigOperandRegister out;
1072
1073   out.base.byteCount = lendian16 (sizeof (out));
1074   out.base.kind = lendian16 (BRIG_KIND_OPERAND_REGISTER);
1075   out.regNum = lendian32 (reg->m_hard_num);
1076
1077   switch (regtype_for_type (reg->m_type))
1078     {
1079     case BRIG_TYPE_B32:
1080       out.regKind = BRIG_REGISTER_KIND_SINGLE;
1081       break;
1082     case BRIG_TYPE_B64:
1083       out.regKind = BRIG_REGISTER_KIND_DOUBLE;
1084       break;
1085     case BRIG_TYPE_B128:
1086       out.regKind = BRIG_REGISTER_KIND_QUAD;
1087       break;
1088     case BRIG_TYPE_B1:
1089       out.regKind = BRIG_REGISTER_KIND_CONTROL;
1090       break;
1091     default:
1092       gcc_unreachable ();
1093     }
1094
1095   brig_operand.add (&out, sizeof (out));
1096 }
1097
1098 /* Emit an address BRIG operand ADDR.  */
1099
1100 static void
1101 emit_address_operand (hsa_op_address *addr)
1102 {
1103   struct BrigOperandAddress out;
1104
1105   out.base.byteCount = lendian16 (sizeof (out));
1106   out.base.kind = lendian16 (BRIG_KIND_OPERAND_ADDRESS);
1107   out.symbol = addr->m_symbol
1108     ? lendian32 (emit_directive_variable (addr->m_symbol)) : 0;
1109   out.reg = addr->m_reg ? lendian32 (enqueue_op (addr->m_reg)) : 0;
1110
1111   if (sizeof (addr->m_imm_offset) == 8)
1112     {
1113       out.offset.lo = lendian32 (addr->m_imm_offset);
1114       out.offset.hi = lendian32 (addr->m_imm_offset >> 32);
1115     }
1116   else
1117     {
1118       gcc_assert (sizeof (addr->m_imm_offset) == 4);
1119       out.offset.lo = lendian32 (addr->m_imm_offset);
1120       out.offset.hi = 0;
1121     }
1122
1123   brig_operand.add (&out, sizeof (out));
1124 }
1125
1126 /* Emit a code reference operand REF.  */
1127
1128 static void
1129 emit_code_ref_operand (hsa_op_code_ref *ref)
1130 {
1131   struct BrigOperandCodeRef out;
1132
1133   out.base.byteCount = lendian16 (sizeof (out));
1134   out.base.kind = lendian16 (BRIG_KIND_OPERAND_CODE_REF);
1135   out.ref = lendian32 (ref->m_directive_offset);
1136   brig_operand.add (&out, sizeof (out));
1137 }
1138
1139 /* Emit a code list operand CODE_LIST.  */
1140
1141 static void
1142 emit_code_list_operand (hsa_op_code_list *code_list)
1143 {
1144   struct BrigOperandCodeList out;
1145   unsigned args = code_list->m_offsets.length ();
1146
1147   for (unsigned i = 0; i < args; i++)
1148     gcc_assert (code_list->m_offsets[i]);
1149
1150   out.base.byteCount = lendian16 (sizeof (out));
1151   out.base.kind = lendian16 (BRIG_KIND_OPERAND_CODE_LIST);
1152
1153   uint32_t byteCount = lendian32 (4 * args);
1154
1155   out.elements = lendian32 (brig_data.add (&byteCount, sizeof (byteCount)));
1156   brig_data.add (code_list->m_offsets.address (), args * sizeof (uint32_t));
1157   brig_data.round_size_up (4);
1158   brig_operand.add (&out, sizeof (out));
1159 }
1160
1161 /* Emit an operand list operand OPERAND_LIST.  */
1162
1163 static void
1164 emit_operand_list_operand (hsa_op_operand_list *operand_list)
1165 {
1166   struct BrigOperandOperandList out;
1167   unsigned args = operand_list->m_offsets.length ();
1168
1169   for (unsigned i = 0; i < args; i++)
1170     gcc_assert (operand_list->m_offsets[i]);
1171
1172   out.base.byteCount = lendian16 (sizeof (out));
1173   out.base.kind = lendian16 (BRIG_KIND_OPERAND_OPERAND_LIST);
1174
1175   uint32_t byteCount = lendian32 (4 * args);
1176
1177   out.elements = lendian32 (brig_data.add (&byteCount, sizeof (byteCount)));
1178   brig_data.add (operand_list->m_offsets.address (), args * sizeof (uint32_t));
1179   brig_data.round_size_up (4);
1180   brig_operand.add (&out, sizeof (out));
1181 }
1182
1183 /* Emit all operands queued for writing.  */
1184
1185 static void
1186 emit_queued_operands (void)
1187 {
1188   for (hsa_op_base *op = op_queue.first_op; op; op = op->m_next)
1189     {
1190       gcc_assert (op->m_brig_op_offset == brig_operand.total_size);
1191       if (hsa_op_immed *imm = dyn_cast <hsa_op_immed *> (op))
1192         emit_immediate_operand (imm);
1193       else if (hsa_op_reg *reg = dyn_cast <hsa_op_reg *> (op))
1194         emit_register_operand (reg);
1195       else if (hsa_op_address *addr = dyn_cast <hsa_op_address *> (op))
1196         emit_address_operand (addr);
1197       else if (hsa_op_code_ref *ref = dyn_cast <hsa_op_code_ref *> (op))
1198         emit_code_ref_operand (ref);
1199       else if (hsa_op_code_list *code_list = dyn_cast <hsa_op_code_list *> (op))
1200         emit_code_list_operand (code_list);
1201       else if (hsa_op_operand_list *l = dyn_cast <hsa_op_operand_list *> (op))
1202         emit_operand_list_operand (l);
1203       else
1204         gcc_unreachable ();
1205     }
1206 }
1207
1208 /* Emit directives describing the function that is used for
1209    a function declaration.  */
1210
1211 static BrigDirectiveExecutable *
1212 emit_function_declaration (tree decl)
1213 {
1214   hsa_function_representation *f = hsa_generate_function_declaration (decl);
1215
1216   BrigDirectiveExecutable *e = emit_function_directives (f, true);
1217   emit_queued_operands ();
1218
1219   delete f;
1220
1221   return e;
1222 }
1223
1224 /* Emit directives describing the function that is used for
1225    an internal function declaration.  */
1226
1227 static BrigDirectiveExecutable *
1228 emit_internal_fn_decl (hsa_internal_fn *fn)
1229 {
1230   hsa_function_representation *f = hsa_generate_internal_fn_decl (fn);
1231
1232   BrigDirectiveExecutable *e = emit_function_directives (f, true);
1233   emit_queued_operands ();
1234
1235   delete f;
1236
1237   return e;
1238 }
1239
1240 /* Enqueue all operands of INSN and return offset to BRIG data section
1241    to list of operand offsets.  */
1242
1243 static unsigned
1244 emit_insn_operands (hsa_insn_basic *insn)
1245 {
1246   auto_vec<BrigOperandOffset32_t, HSA_BRIG_INT_STORAGE_OPERANDS>
1247     operand_offsets;
1248
1249   unsigned l = insn->operand_count ();
1250
1251   /* We have N operands so use 4 * N for the byte_count.  */
1252   uint32_t byte_count = lendian32 (4 * l);
1253   unsigned offset = brig_data.add (&byte_count, sizeof (byte_count));
1254   if (l > 0)
1255     {
1256       operand_offsets.safe_grow (l);
1257       for (unsigned i = 0; i < l; i++)
1258         operand_offsets[i] = lendian32 (enqueue_op (insn->get_op (i)));
1259
1260       brig_data.add (operand_offsets.address (),
1261                      l * sizeof (BrigOperandOffset32_t));
1262     }
1263   brig_data.round_size_up (4);
1264   return offset;
1265 }
1266
1267 /* Enqueue operand OP0, OP1, OP2 (if different from NULL) and return offset
1268    to BRIG data section to list of operand offsets.  */
1269
1270 static unsigned
1271 emit_operands (hsa_op_base *op0, hsa_op_base *op1 = NULL,
1272                hsa_op_base *op2 = NULL)
1273 {
1274   auto_vec<BrigOperandOffset32_t, HSA_BRIG_INT_STORAGE_OPERANDS>
1275     operand_offsets;
1276
1277   gcc_checking_assert (op0 != NULL);
1278   operand_offsets.safe_push (enqueue_op (op0));
1279
1280   if (op1 != NULL)
1281     {
1282       operand_offsets.safe_push (enqueue_op (op1));
1283       if (op2 != NULL)
1284         operand_offsets.safe_push (enqueue_op (op2));
1285     }
1286
1287   unsigned l = operand_offsets.length ();
1288
1289   /* We have N operands so use 4 * N for the byte_count.  */
1290   uint32_t byte_count = lendian32 (4 * l);
1291
1292   unsigned offset = brig_data.add (&byte_count, sizeof (byte_count));
1293   brig_data.add (operand_offsets.address (),
1294                  l * sizeof (BrigOperandOffset32_t));
1295
1296   brig_data.round_size_up (4);
1297
1298   return offset;
1299 }
1300
1301 /* Emit an HSA memory instruction and all necessary directives, schedule
1302    necessary operands for writing.  */
1303
1304 static void
1305 emit_memory_insn (hsa_insn_mem *mem)
1306 {
1307   struct BrigInstMem repr;
1308   gcc_checking_assert (mem->operand_count () == 2);
1309
1310   hsa_op_address *addr = as_a <hsa_op_address *> (mem->get_op (1));
1311
1312   /* This is necessary because of the erroneous typedef of
1313      BrigMemoryModifier8_t which introduces padding which may then contain
1314      random stuff (which we do not want so that we can test things don't
1315      change).  */
1316   memset (&repr, 0, sizeof (repr));
1317   repr.base.base.byteCount = lendian16 (sizeof (repr));
1318   repr.base.base.kind = lendian16 (BRIG_KIND_INST_MEM);
1319   repr.base.opcode = lendian16 (mem->m_opcode);
1320   repr.base.type = lendian16 (mem->m_type);
1321   repr.base.operands = lendian32 (emit_insn_operands (mem));
1322
1323   if (addr->m_symbol)
1324     repr.segment = addr->m_symbol->m_segment;
1325   else
1326     repr.segment = BRIG_SEGMENT_FLAT;
1327   repr.modifier = 0;
1328   repr.equivClass = mem->m_equiv_class;
1329   repr.align = mem->m_align;
1330   if (mem->m_opcode == BRIG_OPCODE_LD)
1331     repr.width = BRIG_WIDTH_1;
1332   else
1333     repr.width = BRIG_WIDTH_NONE;
1334   memset (&repr.reserved, 0, sizeof (repr.reserved));
1335   brig_code.add (&repr, sizeof (repr));
1336   brig_insn_count++;
1337 }
1338
1339 /* Emit an HSA signal memory instruction and all necessary directives, schedule
1340    necessary operands for writing.  */
1341
1342 static void
1343 emit_signal_insn (hsa_insn_signal *mem)
1344 {
1345   struct BrigInstSignal repr;
1346
1347   memset (&repr, 0, sizeof (repr));
1348   repr.base.base.byteCount = lendian16 (sizeof (repr));
1349   repr.base.base.kind = lendian16 (BRIG_KIND_INST_SIGNAL);
1350   repr.base.opcode = lendian16 (mem->m_opcode);
1351   repr.base.type = lendian16 (mem->m_type);
1352   repr.base.operands = lendian32 (emit_insn_operands (mem));
1353
1354   repr.memoryOrder = mem->m_memory_order;
1355   repr.signalOperation = mem->m_signalop;
1356   repr.signalType = hsa_machine_large_p () ? BRIG_TYPE_SIG64 : BRIG_TYPE_SIG32;
1357
1358   brig_code.add (&repr, sizeof (repr));
1359   brig_insn_count++;
1360 }
1361
1362 /* Emit an HSA atomic memory instruction and all necessary directives, schedule
1363    necessary operands for writing.  */
1364
1365 static void
1366 emit_atomic_insn (hsa_insn_atomic *mem)
1367 {
1368   struct BrigInstAtomic repr;
1369
1370   /* Either operand[0] or operand[1] must be an address operand.  */
1371   hsa_op_address *addr = NULL;
1372   if (is_a <hsa_op_address *> (mem->get_op (0)))
1373     addr = as_a <hsa_op_address *> (mem->get_op (0));
1374   else
1375     addr = as_a <hsa_op_address *> (mem->get_op (1));
1376
1377   memset (&repr, 0, sizeof (repr));
1378   repr.base.base.byteCount = lendian16 (sizeof (repr));
1379   repr.base.base.kind = lendian16 (BRIG_KIND_INST_ATOMIC);
1380   repr.base.opcode = lendian16 (mem->m_opcode);
1381   repr.base.type = lendian16 (mem->m_type);
1382   repr.base.operands = lendian32 (emit_insn_operands (mem));
1383
1384   if (addr->m_symbol)
1385     repr.segment = addr->m_symbol->m_segment;
1386   else
1387     repr.segment = BRIG_SEGMENT_FLAT;
1388   repr.memoryOrder = mem->m_memoryorder;
1389   repr.memoryScope = mem->m_memoryscope;
1390   repr.atomicOperation = mem->m_atomicop;
1391
1392   brig_code.add (&repr, sizeof (repr));
1393   brig_insn_count++;
1394 }
1395
1396 /* Emit an HSA LDA instruction and all necessary directives, schedule
1397    necessary operands for writing.  */
1398
1399 static void
1400 emit_addr_insn (hsa_insn_basic *insn)
1401 {
1402   struct BrigInstAddr repr;
1403
1404   hsa_op_address *addr = as_a <hsa_op_address *> (insn->get_op (1));
1405
1406   repr.base.base.byteCount = lendian16 (sizeof (repr));
1407   repr.base.base.kind = lendian16 (BRIG_KIND_INST_ADDR);
1408   repr.base.opcode = lendian16 (insn->m_opcode);
1409   repr.base.type = lendian16 (insn->m_type);
1410   repr.base.operands = lendian32 (emit_insn_operands (insn));
1411
1412   if (addr->m_symbol)
1413     repr.segment = addr->m_symbol->m_segment;
1414   else
1415     repr.segment = BRIG_SEGMENT_FLAT;
1416   memset (&repr.reserved, 0, sizeof (repr.reserved));
1417
1418   brig_code.add (&repr, sizeof (repr));
1419   brig_insn_count++;
1420 }
1421
1422 /* Emit an HSA segment conversion instruction and all necessary directives,
1423    schedule necessary operands for writing.  */
1424
1425 static void
1426 emit_segment_insn (hsa_insn_seg *seg)
1427 {
1428   struct BrigInstSegCvt repr;
1429
1430   repr.base.base.byteCount = lendian16 (sizeof (repr));
1431   repr.base.base.kind = lendian16 (BRIG_KIND_INST_SEG_CVT);
1432   repr.base.opcode = lendian16 (seg->m_opcode);
1433   repr.base.type = lendian16 (seg->m_type);
1434   repr.base.operands = lendian32 (emit_insn_operands (seg));
1435   repr.sourceType = lendian16 (as_a <hsa_op_reg *> (seg->get_op (1))->m_type);
1436   repr.segment = seg->m_segment;
1437   repr.modifier = 0;
1438
1439   brig_code.add (&repr, sizeof (repr));
1440
1441   brig_insn_count++;
1442 }
1443
1444 /* Emit an HSA alloca instruction and all necessary directives,
1445    schedule necessary operands for writing.  */
1446
1447 static void
1448 emit_alloca_insn (hsa_insn_alloca *alloca)
1449 {
1450   struct BrigInstMem repr;
1451   gcc_checking_assert (alloca->operand_count () == 2);
1452
1453   memset (&repr, 0, sizeof (repr));
1454   repr.base.base.byteCount = lendian16 (sizeof (repr));
1455   repr.base.base.kind = lendian16 (BRIG_KIND_INST_MEM);
1456   repr.base.opcode = lendian16 (alloca->m_opcode);
1457   repr.base.type = lendian16 (alloca->m_type);
1458   repr.base.operands = lendian32 (emit_insn_operands (alloca));
1459   repr.segment = BRIG_SEGMENT_PRIVATE;
1460   repr.modifier = 0;
1461   repr.equivClass = 0;
1462   repr.align = alloca->m_align;
1463   repr.width = BRIG_WIDTH_NONE;
1464   memset (&repr.reserved, 0, sizeof (repr.reserved));
1465   brig_code.add (&repr, sizeof (repr));
1466   brig_insn_count++;
1467 }
1468
1469 /* Emit an HSA comparison instruction and all necessary directives,
1470    schedule necessary operands for writing.  */
1471
1472 static void
1473 emit_cmp_insn (hsa_insn_cmp *cmp)
1474 {
1475   struct BrigInstCmp repr;
1476
1477   memset (&repr, 0, sizeof (repr));
1478   repr.base.base.byteCount = lendian16 (sizeof (repr));
1479   repr.base.base.kind = lendian16 (BRIG_KIND_INST_CMP);
1480   repr.base.opcode = lendian16 (cmp->m_opcode);
1481   repr.base.type = lendian16 (cmp->m_type);
1482   repr.base.operands = lendian32 (emit_insn_operands (cmp));
1483
1484   if (is_a <hsa_op_reg *> (cmp->get_op (1)))
1485     repr.sourceType
1486       = lendian16 (as_a <hsa_op_reg *> (cmp->get_op (1))->m_type);
1487   else
1488     repr.sourceType
1489       = lendian16 (as_a <hsa_op_immed *> (cmp->get_op (1))->m_type);
1490   repr.modifier = 0;
1491   repr.compare = cmp->m_compare;
1492   repr.pack = 0;
1493
1494   brig_code.add (&repr, sizeof (repr));
1495   brig_insn_count++;
1496 }
1497
1498 /* Emit an HSA generic branching/sycnronization instruction.  */
1499
1500 static void
1501 emit_generic_branch_insn (hsa_insn_br *br)
1502 {
1503   struct BrigInstBr repr;
1504   repr.base.base.byteCount = lendian16 (sizeof (repr));
1505   repr.base.base.kind = lendian16 (BRIG_KIND_INST_BR);
1506   repr.base.opcode = lendian16 (br->m_opcode);
1507   repr.width = br->m_width;
1508   repr.base.type = lendian16 (br->m_type);
1509   repr.base.operands = lendian32 (emit_insn_operands (br));
1510   memset (&repr.reserved, 0, sizeof (repr.reserved));
1511
1512   brig_code.add (&repr, sizeof (repr));
1513   brig_insn_count++;
1514 }
1515
1516 /* Emit an HSA conditional branching instruction and all necessary directives,
1517    schedule necessary operands for writing.  */
1518
1519 static void
1520 emit_cond_branch_insn (hsa_insn_cbr *br)
1521 {
1522   struct BrigInstBr repr;
1523
1524   basic_block target = NULL;
1525   edge_iterator ei;
1526   edge e;
1527
1528   /* At the moment we only handle direct conditional jumps.  */
1529   gcc_assert (br->m_opcode == BRIG_OPCODE_CBR);
1530   repr.base.base.byteCount = lendian16 (sizeof (repr));
1531   repr.base.base.kind = lendian16 (BRIG_KIND_INST_BR);
1532   repr.base.opcode = lendian16 (br->m_opcode);
1533   repr.width = br->m_width;
1534   /* For Conditional jumps the type is always B1.  */
1535   repr.base.type = lendian16 (BRIG_TYPE_B1);
1536
1537   FOR_EACH_EDGE (e, ei, br->m_bb->succs)
1538     if (e->flags & EDGE_TRUE_VALUE)
1539       {
1540         target = e->dest;
1541         break;
1542       }
1543   gcc_assert (target);
1544
1545   repr.base.operands
1546     = lendian32 (emit_operands (br->get_op (0),
1547                                 &hsa_bb_for_bb (target)->m_label_ref));
1548   memset (&repr.reserved, 0, sizeof (repr.reserved));
1549
1550   brig_code.add (&repr, sizeof (repr));
1551   brig_insn_count++;
1552 }
1553
1554 /* Emit an HSA unconditional jump branching instruction that points to
1555    a label REFERENCE.  */
1556
1557 static void
1558 emit_unconditional_jump (hsa_op_code_ref *reference)
1559 {
1560   struct BrigInstBr repr;
1561
1562   repr.base.base.byteCount = lendian16 (sizeof (repr));
1563   repr.base.base.kind = lendian16 (BRIG_KIND_INST_BR);
1564   repr.base.opcode = lendian16 (BRIG_OPCODE_BR);
1565   repr.base.type = lendian16 (BRIG_TYPE_NONE);
1566   /* Direct branches to labels must be width(all).  */
1567   repr.width = BRIG_WIDTH_ALL;
1568
1569   repr.base.operands = lendian32 (emit_operands (reference));
1570   memset (&repr.reserved, 0, sizeof (repr.reserved));
1571   brig_code.add (&repr, sizeof (repr));
1572   brig_insn_count++;
1573 }
1574
1575 /* Emit an HSA switch jump instruction that uses a jump table to
1576    jump to a destination label.  */
1577
1578 static void
1579 emit_switch_insn (hsa_insn_sbr *sbr)
1580 {
1581   struct BrigInstBr repr;
1582
1583   gcc_assert (sbr->m_opcode == BRIG_OPCODE_SBR);
1584   repr.base.base.byteCount = lendian16 (sizeof (repr));
1585   repr.base.base.kind = lendian16 (BRIG_KIND_INST_BR);
1586   repr.base.opcode = lendian16 (sbr->m_opcode);
1587   repr.width = BRIG_WIDTH_1;
1588   /* For Conditional jumps the type is always B1.  */
1589   hsa_op_reg *index = as_a <hsa_op_reg *> (sbr->get_op (0));
1590   repr.base.type = lendian16 (index->m_type);
1591   repr.base.operands
1592     = lendian32 (emit_operands (sbr->get_op (0), sbr->m_label_code_list));
1593   memset (&repr.reserved, 0, sizeof (repr.reserved));
1594
1595   brig_code.add (&repr, sizeof (repr));
1596   brig_insn_count++;
1597 }
1598
1599 /* Emit a HSA convert instruction and all necessary directives, schedule
1600    necessary operands for writing.  */
1601
1602 static void
1603 emit_cvt_insn (hsa_insn_cvt *insn)
1604 {
1605   struct BrigInstCvt repr;
1606   BrigType16_t srctype;
1607
1608   repr.base.base.byteCount = lendian16 (sizeof (repr));
1609   repr.base.base.kind = lendian16 (BRIG_KIND_INST_CVT);
1610   repr.base.opcode = lendian16 (insn->m_opcode);
1611   repr.base.type = lendian16 (insn->m_type);
1612   repr.base.operands = lendian32 (emit_insn_operands (insn));
1613
1614   if (is_a <hsa_op_reg *> (insn->get_op (1)))
1615     srctype = as_a <hsa_op_reg *> (insn->get_op (1))->m_type;
1616   else
1617     srctype = as_a <hsa_op_immed *> (insn->get_op (1))->m_type;
1618   repr.sourceType = lendian16 (srctype);
1619   repr.modifier = 0;
1620   /* float to smaller float requires a rounding setting (we default
1621      to 'near'.  */
1622   if (hsa_type_float_p (insn->m_type)
1623       && (!hsa_type_float_p (srctype)
1624           || ((insn->m_type & BRIG_TYPE_BASE_MASK)
1625               < (srctype & BRIG_TYPE_BASE_MASK))))
1626     repr.round = BRIG_ROUND_FLOAT_NEAR_EVEN;
1627   else if (hsa_type_integer_p (insn->m_type) &&
1628            hsa_type_float_p (srctype))
1629     repr.round = BRIG_ROUND_INTEGER_ZERO;
1630   else
1631     repr.round = BRIG_ROUND_NONE;
1632   brig_code.add (&repr, sizeof (repr));
1633   brig_insn_count++;
1634 }
1635
1636 /* Emit call instruction INSN, where this instruction must be closed
1637    within a call block instruction.  */
1638
1639 static void
1640 emit_call_insn (hsa_insn_call *call)
1641 {
1642   struct BrigInstBr repr;
1643
1644   repr.base.base.byteCount = lendian16 (sizeof (repr));
1645   repr.base.base.kind = lendian16 (BRIG_KIND_INST_BR);
1646   repr.base.opcode = lendian16 (BRIG_OPCODE_CALL);
1647   repr.base.type = lendian16 (BRIG_TYPE_NONE);
1648
1649   repr.base.operands
1650     = lendian32 (emit_operands (call->m_result_code_list, &call->m_func,
1651                                 call->m_args_code_list));
1652
1653   /* Internal functions have not set m_called_function.  */
1654   if (call->m_called_function)
1655     {
1656       function_linkage_pair pair (call->m_called_function,
1657                                   call->m_func.m_brig_op_offset);
1658       function_call_linkage.safe_push (pair);
1659     }
1660   else
1661     {
1662       hsa_internal_fn *slot
1663         = hsa_emitted_internal_decls->find (call->m_called_internal_fn);
1664       gcc_assert (slot);
1665       gcc_assert (slot->m_offset > 0);
1666       call->m_func.m_directive_offset = slot->m_offset;
1667     }
1668
1669   repr.width = BRIG_WIDTH_ALL;
1670   memset (&repr.reserved, 0, sizeof (repr.reserved));
1671
1672   brig_code.add (&repr, sizeof (repr));
1673   brig_insn_count++;
1674 }
1675
1676 /* Emit argument block directive.  */
1677
1678 static void
1679 emit_arg_block_insn (hsa_insn_arg_block *insn)
1680 {
1681   switch (insn->m_kind)
1682     {
1683     case BRIG_KIND_DIRECTIVE_ARG_BLOCK_START:
1684       {
1685         struct BrigDirectiveArgBlock repr;
1686         repr.base.byteCount = lendian16 (sizeof (repr));
1687         repr.base.kind = lendian16 (insn->m_kind);
1688         brig_code.add (&repr, sizeof (repr));
1689
1690         for (unsigned i = 0; i < insn->m_call_insn->m_input_args.length (); i++)
1691           {
1692             insn->m_call_insn->m_args_code_list->m_offsets[i]
1693               = lendian32 (emit_directive_variable
1694                            (insn->m_call_insn->m_input_args[i]));
1695             brig_insn_count++;
1696           }
1697
1698         if (insn->m_call_insn->m_output_arg)
1699           {
1700             insn->m_call_insn->m_result_code_list->m_offsets[0]
1701               = lendian32 (emit_directive_variable
1702                            (insn->m_call_insn->m_output_arg));
1703             brig_insn_count++;
1704           }
1705
1706         break;
1707       }
1708     case BRIG_KIND_DIRECTIVE_ARG_BLOCK_END:
1709       {
1710         struct BrigDirectiveArgBlock repr;
1711         repr.base.byteCount = lendian16 (sizeof (repr));
1712         repr.base.kind = lendian16 (insn->m_kind);
1713         brig_code.add (&repr, sizeof (repr));
1714         break;
1715       }
1716     default:
1717       gcc_unreachable ();
1718     }
1719
1720   brig_insn_count++;
1721 }
1722
1723 /* Emit comment directive.  */
1724
1725 static void
1726 emit_comment_insn (hsa_insn_comment *insn)
1727 {
1728   struct BrigDirectiveComment repr;
1729   memset (&repr, 0, sizeof (repr));
1730
1731   repr.base.byteCount = lendian16 (sizeof (repr));
1732   repr.base.kind = lendian16 (insn->m_opcode);
1733   repr.name = brig_emit_string (insn->m_comment, '\0', false);
1734   brig_code.add (&repr, sizeof (repr));
1735 }
1736
1737 /* Emit queue instruction INSN.  */
1738
1739 static void
1740 emit_queue_insn (hsa_insn_queue *insn)
1741 {
1742   BrigInstQueue repr;
1743   memset (&repr, 0, sizeof (repr));
1744
1745   repr.base.base.byteCount = lendian16 (sizeof (repr));
1746   repr.base.base.kind = lendian16 (BRIG_KIND_INST_QUEUE);
1747   repr.base.opcode = lendian16 (insn->m_opcode);
1748   repr.base.type = lendian16 (insn->m_type);
1749   repr.segment = insn->m_segment;
1750   repr.memoryOrder = insn->m_memory_order;
1751   repr.base.operands = lendian32 (emit_insn_operands (insn));
1752   brig_data.round_size_up (4);
1753   brig_code.add (&repr, sizeof (repr));
1754
1755   brig_insn_count++;
1756 }
1757
1758 /* Emit source type instruction INSN.  */
1759
1760 static void
1761 emit_srctype_insn (hsa_insn_srctype *insn)
1762 {
1763   /* We assume that BrigInstMod has a BrigInstBasic prefix.  */
1764   struct BrigInstSourceType repr;
1765   unsigned operand_count = insn->operand_count ();
1766   gcc_checking_assert (operand_count >= 2);
1767
1768   memset (&repr, 0, sizeof (repr));
1769   repr.sourceType = lendian16 (insn->m_source_type);
1770   repr.base.base.byteCount = lendian16 (sizeof (repr));
1771   repr.base.base.kind = lendian16 (BRIG_KIND_INST_SOURCE_TYPE);
1772   repr.base.opcode = lendian16 (insn->m_opcode);
1773   repr.base.type = lendian16 (insn->m_type);
1774
1775   repr.base.operands = lendian32 (emit_insn_operands (insn));
1776   brig_code.add (&repr, sizeof (struct BrigInstSourceType));
1777   brig_insn_count++;
1778 }
1779
1780 /* Emit packed instruction INSN.  */
1781
1782 static void
1783 emit_packed_insn (hsa_insn_packed *insn)
1784 {
1785   /* We assume that BrigInstMod has a BrigInstBasic prefix.  */
1786   struct BrigInstSourceType repr;
1787   unsigned operand_count = insn->operand_count ();
1788   gcc_checking_assert (operand_count >= 2);
1789
1790   memset (&repr, 0, sizeof (repr));
1791   repr.sourceType = lendian16 (insn->m_source_type);
1792   repr.base.base.byteCount = lendian16 (sizeof (repr));
1793   repr.base.base.kind = lendian16 (BRIG_KIND_INST_SOURCE_TYPE);
1794   repr.base.opcode = lendian16 (insn->m_opcode);
1795   repr.base.type = lendian16 (insn->m_type);
1796
1797   if (insn->m_opcode == BRIG_OPCODE_COMBINE)
1798     {
1799       /* Create operand list for packed type.  */
1800       for (unsigned i = 1; i < operand_count; i++)
1801         {
1802           gcc_checking_assert (insn->get_op (i));
1803           insn->m_operand_list->m_offsets[i - 1]
1804             = lendian32 (enqueue_op (insn->get_op (i)));
1805         }
1806
1807       repr.base.operands = lendian32 (emit_operands (insn->get_op (0),
1808                                                      insn->m_operand_list));
1809     }
1810   else if (insn->m_opcode == BRIG_OPCODE_EXPAND)
1811     {
1812       /* Create operand list for packed type.  */
1813       for (unsigned i = 0; i < operand_count - 1; i++)
1814         {
1815           gcc_checking_assert (insn->get_op (i));
1816           insn->m_operand_list->m_offsets[i]
1817             = lendian32 (enqueue_op (insn->get_op (i)));
1818         }
1819
1820       unsigned ops = emit_operands (insn->m_operand_list,
1821                                     insn->get_op (insn->operand_count () - 1));
1822       repr.base.operands = lendian32 (ops);
1823     }
1824
1825
1826   brig_code.add (&repr, sizeof (struct BrigInstSourceType));
1827   brig_insn_count++;
1828 }
1829
1830 /* Emit a basic HSA instruction and all necessary directives, schedule
1831    necessary operands for writing.  */
1832
1833 static void
1834 emit_basic_insn (hsa_insn_basic *insn)
1835 {
1836   /* We assume that BrigInstMod has a BrigInstBasic prefix.  */
1837   struct BrigInstMod repr;
1838   BrigType16_t type;
1839
1840   memset (&repr, 0, sizeof (repr));
1841   repr.base.base.byteCount = lendian16 (sizeof (BrigInstBasic));
1842   repr.base.base.kind = lendian16 (BRIG_KIND_INST_BASIC);
1843   repr.base.opcode = lendian16 (insn->m_opcode);
1844   switch (insn->m_opcode)
1845     {
1846       /* And the bit-logical operations need bit types and whine about
1847          arithmetic types :-/  */
1848       case BRIG_OPCODE_AND:
1849       case BRIG_OPCODE_OR:
1850       case BRIG_OPCODE_XOR:
1851       case BRIG_OPCODE_NOT:
1852         type = regtype_for_type (insn->m_type);
1853         break;
1854       default:
1855         type = insn->m_type;
1856         break;
1857     }
1858   repr.base.type = lendian16 (type);
1859   repr.base.operands = lendian32 (emit_insn_operands (insn));
1860
1861   if (hsa_type_packed_p (type))
1862     {
1863       if (hsa_type_float_p (type)
1864           && !hsa_opcode_floating_bit_insn_p (insn->m_opcode))
1865         repr.round = BRIG_ROUND_FLOAT_NEAR_EVEN;
1866       else
1867         repr.round = 0;
1868       /* We assume that destination and sources agree in packing layout.  */
1869       if (insn->num_used_ops () >= 2)
1870         repr.pack = BRIG_PACK_PP;
1871       else
1872         repr.pack = BRIG_PACK_P;
1873       repr.reserved = 0;
1874       repr.base.base.byteCount = lendian16 (sizeof (BrigInstMod));
1875       repr.base.base.kind = lendian16 (BRIG_KIND_INST_MOD);
1876       brig_code.add (&repr, sizeof (struct BrigInstMod));
1877     }
1878   else
1879     brig_code.add (&repr, sizeof (struct BrigInstBasic));
1880   brig_insn_count++;
1881 }
1882
1883 /* Emit an HSA instruction and all necessary directives, schedule necessary
1884    operands for writing.  */
1885
1886 static void
1887 emit_insn (hsa_insn_basic *insn)
1888 {
1889   gcc_assert (!is_a <hsa_insn_phi *> (insn));
1890
1891   insn->m_brig_offset = brig_code.total_size;
1892
1893   if (hsa_insn_signal *signal = dyn_cast <hsa_insn_signal *> (insn))
1894     emit_signal_insn (signal);
1895   else if (hsa_insn_atomic *atom = dyn_cast <hsa_insn_atomic *> (insn))
1896     emit_atomic_insn (atom);
1897   else if (hsa_insn_mem *mem = dyn_cast <hsa_insn_mem *> (insn))
1898     emit_memory_insn (mem);
1899   else if (insn->m_opcode == BRIG_OPCODE_LDA)
1900     emit_addr_insn (insn);
1901   else if (hsa_insn_seg *seg = dyn_cast <hsa_insn_seg *> (insn))
1902     emit_segment_insn (seg);
1903   else if (hsa_insn_cmp *cmp = dyn_cast <hsa_insn_cmp *> (insn))
1904     emit_cmp_insn (cmp);
1905   else if (hsa_insn_cbr *br = dyn_cast <hsa_insn_cbr *> (insn))
1906     emit_cond_branch_insn (br);
1907   else if (hsa_insn_sbr *sbr = dyn_cast <hsa_insn_sbr *> (insn))
1908     {
1909       if (switch_instructions == NULL)
1910         switch_instructions = new vec <hsa_insn_sbr *> ();
1911
1912       switch_instructions->safe_push (sbr);
1913       emit_switch_insn (sbr);
1914     }
1915   else if (hsa_insn_br *br = dyn_cast <hsa_insn_br *> (insn))
1916     emit_generic_branch_insn (br);
1917   else if (hsa_insn_arg_block *block = dyn_cast <hsa_insn_arg_block *> (insn))
1918     emit_arg_block_insn (block);
1919   else if (hsa_insn_call *call = dyn_cast <hsa_insn_call *> (insn))
1920     emit_call_insn (call);
1921   else if (hsa_insn_comment *comment = dyn_cast <hsa_insn_comment *> (insn))
1922     emit_comment_insn (comment);
1923   else if (hsa_insn_queue *queue = dyn_cast <hsa_insn_queue *> (insn))
1924     emit_queue_insn (queue);
1925   else if (hsa_insn_srctype *srctype = dyn_cast <hsa_insn_srctype *> (insn))
1926     emit_srctype_insn (srctype);
1927   else if (hsa_insn_packed *packed = dyn_cast <hsa_insn_packed *> (insn))
1928     emit_packed_insn (packed);
1929   else if (hsa_insn_cvt *cvt = dyn_cast <hsa_insn_cvt *> (insn))
1930     emit_cvt_insn (cvt);
1931   else if (hsa_insn_alloca *alloca = dyn_cast <hsa_insn_alloca *> (insn))
1932     emit_alloca_insn (alloca);
1933   else
1934     emit_basic_insn (insn);
1935 }
1936
1937 /* We have just finished emitting BB and are about to emit NEXT_BB if non-NULL,
1938    or we are about to finish emitting code, if it is NULL.  If the fall through
1939    edge from BB does not lead to NEXT_BB, emit an unconditional jump.  */
1940
1941 static void
1942 perhaps_emit_branch (basic_block bb, basic_block next_bb)
1943 {
1944   basic_block t_bb = NULL, ff = NULL;
1945
1946   edge_iterator ei;
1947   edge e;
1948
1949   /* If the last instruction of BB is a switch, ignore emission of all
1950      edges.  */
1951   if (hsa_bb_for_bb (bb)->m_last_insn
1952       && is_a <hsa_insn_sbr *> (hsa_bb_for_bb (bb)->m_last_insn))
1953     return;
1954
1955   FOR_EACH_EDGE (e, ei, bb->succs)
1956     if (e->flags & EDGE_TRUE_VALUE)
1957       {
1958         gcc_assert (!t_bb);
1959         t_bb = e->dest;
1960       }
1961     else
1962       {
1963         gcc_assert (!ff);
1964         ff = e->dest;
1965       }
1966
1967   if (!ff || ff == next_bb || ff == EXIT_BLOCK_PTR_FOR_FN (cfun))
1968     return;
1969
1970   emit_unconditional_jump (&hsa_bb_for_bb (ff)->m_label_ref);
1971 }
1972
1973 /* Emit the a function with name NAME to the various brig sections.  */
1974
1975 void
1976 hsa_brig_emit_function (void)
1977 {
1978   basic_block bb, prev_bb;
1979   hsa_insn_basic *insn;
1980   BrigDirectiveExecutable *ptr_to_fndir;
1981
1982   brig_init ();
1983
1984   brig_insn_count = 0;
1985   memset (&op_queue, 0, sizeof (op_queue));
1986   op_queue.projected_size = brig_operand.total_size;
1987
1988   if (!function_offsets)
1989     function_offsets = new hash_map<tree, BrigCodeOffset32_t> ();
1990
1991   if (!emitted_declarations)
1992     emitted_declarations = new hash_map <tree, BrigDirectiveExecutable *> ();
1993
1994   for (unsigned i = 0; i < hsa_cfun->m_called_functions.length (); i++)
1995     {
1996       tree called = hsa_cfun->m_called_functions[i];
1997
1998       /* If the function has no definition, emit a declaration.  */
1999       if (!emitted_declarations->get (called))
2000         {
2001           BrigDirectiveExecutable *e = emit_function_declaration (called);
2002           emitted_declarations->put (called, e);
2003         }
2004     }
2005
2006   for (unsigned i = 0; i < hsa_cfun->m_called_internal_fns.length (); i++)
2007     {
2008       hsa_internal_fn *called = hsa_cfun->m_called_internal_fns[i];
2009       emit_internal_fn_decl (called);
2010     }
2011
2012   ptr_to_fndir = emit_function_directives (hsa_cfun, false);
2013   for (insn = hsa_bb_for_bb (ENTRY_BLOCK_PTR_FOR_FN (cfun))->m_first_insn;
2014        insn;
2015        insn = insn->m_next)
2016     emit_insn (insn);
2017   prev_bb = ENTRY_BLOCK_PTR_FOR_FN (cfun);
2018   FOR_EACH_BB_FN (bb, cfun)
2019     {
2020       perhaps_emit_branch (prev_bb, bb);
2021       emit_bb_label_directive (hsa_bb_for_bb (bb));
2022       for (insn = hsa_bb_for_bb (bb)->m_first_insn; insn; insn = insn->m_next)
2023         emit_insn (insn);
2024       prev_bb = bb;
2025     }
2026   perhaps_emit_branch (prev_bb, NULL);
2027   ptr_to_fndir->nextModuleEntry = lendian32 (brig_code.total_size);
2028
2029   /* Fill up label references for all sbr instructions.  */
2030   if (switch_instructions)
2031     {
2032       for (unsigned i = 0; i < switch_instructions->length (); i++)
2033         {
2034           hsa_insn_sbr *sbr = (*switch_instructions)[i];
2035           for (unsigned j = 0; j < sbr->m_jump_table.length (); j++)
2036             {
2037               hsa_bb *hbb = hsa_bb_for_bb (sbr->m_jump_table[j]);
2038               sbr->m_label_code_list->m_offsets[j]
2039                 = hbb->m_label_ref.m_directive_offset;
2040             }
2041         }
2042
2043       switch_instructions->release ();
2044       delete switch_instructions;
2045       switch_instructions = NULL;
2046     }
2047
2048   if (dump_file)
2049     {
2050       fprintf (dump_file, "------- After BRIG emission: -------\n");
2051       dump_hsa_cfun (dump_file);
2052     }
2053
2054   emit_queued_operands ();
2055 }
2056
2057 /* Emit all OMP symbols related to OMP.  */
2058
2059 void
2060 hsa_brig_emit_omp_symbols (void)
2061 {
2062   brig_init ();
2063   emit_directive_variable (hsa_num_threads);
2064 }
2065
2066 /* Create and return __hsa_global_variables symbol that contains
2067    all informations consumed by libgomp to link global variables
2068    with their string names used by an HSA kernel.  */
2069
2070 static tree
2071 hsa_output_global_variables ()
2072 {
2073   unsigned l = hsa_global_variable_symbols->elements ();
2074
2075   tree variable_info_type = make_node (RECORD_TYPE);
2076   tree id_f1 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2077                            get_identifier ("name"), ptr_type_node);
2078   DECL_CHAIN (id_f1) = NULL_TREE;
2079   tree id_f2 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2080                            get_identifier ("omp_data_size"),
2081                            ptr_type_node);
2082   DECL_CHAIN (id_f2) = id_f1;
2083   finish_builtin_struct (variable_info_type, "__hsa_variable_info", id_f2,
2084                          NULL_TREE);
2085
2086   tree int_num_of_global_vars;
2087   int_num_of_global_vars = build_int_cst (uint32_type_node, l);
2088   tree global_vars_num_index_type = build_index_type (int_num_of_global_vars);
2089   tree global_vars_array_type = build_array_type (variable_info_type,
2090                                                   global_vars_num_index_type);
2091   TYPE_ARTIFICIAL (global_vars_array_type) = 1;
2092
2093   vec<constructor_elt, va_gc> *global_vars_vec = NULL;
2094
2095   for (hash_table <hsa_noop_symbol_hasher>::iterator it
2096        = hsa_global_variable_symbols->begin ();
2097        it != hsa_global_variable_symbols->end (); ++it)
2098     {
2099       unsigned len = strlen ((*it)->m_name);
2100       char *copy = XNEWVEC (char, len + 2);
2101       copy[0] = '&';
2102       memcpy (copy + 1, (*it)->m_name, len);
2103       copy[len + 1] = '\0';
2104       len++;
2105       hsa_sanitize_name (copy);
2106
2107       tree var_name = build_string (len, copy);
2108       TREE_TYPE (var_name)
2109         = build_array_type (char_type_node, build_index_type (size_int (len)));
2110       free (copy);
2111
2112       vec<constructor_elt, va_gc> *variable_info_vec = NULL;
2113       CONSTRUCTOR_APPEND_ELT (variable_info_vec, NULL_TREE,
2114                               build1 (ADDR_EXPR,
2115                                       build_pointer_type (TREE_TYPE (var_name)),
2116                                       var_name));
2117       CONSTRUCTOR_APPEND_ELT (variable_info_vec, NULL_TREE,
2118                               build_fold_addr_expr ((*it)->m_decl));
2119
2120       tree variable_info_ctor = build_constructor (variable_info_type,
2121                                                    variable_info_vec);
2122
2123       CONSTRUCTOR_APPEND_ELT (global_vars_vec, NULL_TREE,
2124                               variable_info_ctor);
2125     }
2126
2127   tree global_vars_ctor = build_constructor (global_vars_array_type,
2128                                              global_vars_vec);
2129
2130   char tmp_name[64];
2131   ASM_GENERATE_INTERNAL_LABEL (tmp_name, "__hsa_global_variables", 1);
2132   tree global_vars_table = build_decl (UNKNOWN_LOCATION, VAR_DECL,
2133                                            get_identifier (tmp_name),
2134                                            global_vars_array_type);
2135   TREE_STATIC (global_vars_table) = 1;
2136   TREE_READONLY (global_vars_table) = 1;
2137   TREE_PUBLIC (global_vars_table) = 0;
2138   DECL_ARTIFICIAL (global_vars_table) = 1;
2139   DECL_IGNORED_P (global_vars_table) = 1;
2140   DECL_EXTERNAL (global_vars_table) = 0;
2141   TREE_CONSTANT (global_vars_table) = 1;
2142   DECL_INITIAL (global_vars_table) = global_vars_ctor;
2143   varpool_node::finalize_decl (global_vars_table);
2144
2145   return global_vars_table;
2146 }
2147
2148 /* Create __hsa_host_functions and __hsa_kernels that contain
2149    all informations consumed by libgomp to register all kernels
2150    in the BRIG binary.  */
2151
2152 static void
2153 hsa_output_kernels (tree *host_func_table, tree *kernels)
2154 {
2155   unsigned map_count = hsa_get_number_decl_kernel_mappings ();
2156
2157   tree int_num_of_kernels;
2158   int_num_of_kernels = build_int_cst (uint32_type_node, map_count);
2159   tree kernel_num_index_type = build_index_type (int_num_of_kernels);
2160   tree host_functions_array_type = build_array_type (ptr_type_node,
2161                                                      kernel_num_index_type);
2162   TYPE_ARTIFICIAL (host_functions_array_type) = 1;
2163
2164   vec<constructor_elt, va_gc> *host_functions_vec = NULL;
2165   for (unsigned i = 0; i < map_count; ++i)
2166     {
2167       tree decl = hsa_get_decl_kernel_mapping_decl (i);
2168       tree host_fn = build_fold_addr_expr (hsa_get_host_function (decl));
2169       CONSTRUCTOR_APPEND_ELT (host_functions_vec, NULL_TREE, host_fn);
2170     }
2171   tree host_functions_ctor = build_constructor (host_functions_array_type,
2172                                                 host_functions_vec);
2173   char tmp_name[64];
2174   ASM_GENERATE_INTERNAL_LABEL (tmp_name, "__hsa_host_functions", 1);
2175   tree hsa_host_func_table = build_decl (UNKNOWN_LOCATION, VAR_DECL,
2176                                          get_identifier (tmp_name),
2177                                          host_functions_array_type);
2178   TREE_STATIC (hsa_host_func_table) = 1;
2179   TREE_READONLY (hsa_host_func_table) = 1;
2180   TREE_PUBLIC (hsa_host_func_table) = 0;
2181   DECL_ARTIFICIAL (hsa_host_func_table) = 1;
2182   DECL_IGNORED_P (hsa_host_func_table) = 1;
2183   DECL_EXTERNAL (hsa_host_func_table) = 0;
2184   TREE_CONSTANT (hsa_host_func_table) = 1;
2185   DECL_INITIAL (hsa_host_func_table) = host_functions_ctor;
2186   varpool_node::finalize_decl (hsa_host_func_table);
2187   *host_func_table = hsa_host_func_table;
2188
2189   /* Following code emits list of kernel_info structures.  */
2190
2191   tree kernel_info_type = make_node (RECORD_TYPE);
2192   tree id_f1 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2193                            get_identifier ("name"), ptr_type_node);
2194   DECL_CHAIN (id_f1) = NULL_TREE;
2195   tree id_f2 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2196                            get_identifier ("omp_data_size"),
2197                            unsigned_type_node);
2198   DECL_CHAIN (id_f2) = id_f1;
2199   tree id_f3 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2200                            get_identifier ("gridified_kernel_p"),
2201                            boolean_type_node);
2202   DECL_CHAIN (id_f3) = id_f2;
2203   tree id_f4 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2204                            get_identifier ("kernel_dependencies_count"),
2205                            unsigned_type_node);
2206   DECL_CHAIN (id_f4) = id_f3;
2207   tree id_f5 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2208                            get_identifier ("kernel_dependencies"),
2209                            build_pointer_type (build_pointer_type
2210                                                (char_type_node)));
2211   DECL_CHAIN (id_f5) = id_f4;
2212   finish_builtin_struct (kernel_info_type, "__hsa_kernel_info", id_f5,
2213                          NULL_TREE);
2214
2215   int_num_of_kernels = build_int_cstu (uint32_type_node, map_count);
2216   tree kernel_info_vector_type
2217     = build_array_type (kernel_info_type,
2218                         build_index_type (int_num_of_kernels));
2219   TYPE_ARTIFICIAL (kernel_info_vector_type) = 1;
2220
2221   vec<constructor_elt, va_gc> *kernel_info_vector_vec = NULL;
2222   tree kernel_dependencies_vector_type = NULL;
2223
2224   for (unsigned i = 0; i < map_count; ++i)
2225     {
2226       tree kernel = hsa_get_decl_kernel_mapping_decl (i);
2227       char *name = hsa_get_decl_kernel_mapping_name (i);
2228       unsigned len = strlen (name);
2229       char *copy = XNEWVEC (char, len + 2);
2230       copy[0] = '&';
2231       memcpy (copy + 1, name, len);
2232       copy[len + 1] = '\0';
2233       len++;
2234
2235       tree kern_name = build_string (len, copy);
2236       TREE_TYPE (kern_name)
2237         = build_array_type (char_type_node, build_index_type (size_int (len)));
2238       free (copy);
2239
2240       unsigned omp_size = hsa_get_decl_kernel_mapping_omp_size (i);
2241       tree omp_data_size = build_int_cstu (unsigned_type_node, omp_size);
2242       bool gridified_kernel_p = hsa_get_decl_kernel_mapping_gridified (i);
2243       tree gridified_kernel_p_tree = build_int_cstu (boolean_type_node,
2244                                                      gridified_kernel_p);
2245       unsigned count = 0;
2246       vec<constructor_elt, va_gc> *kernel_dependencies_vec = NULL;
2247       if (hsa_decl_kernel_dependencies)
2248         {
2249           vec<const char *> **slot;
2250           slot = hsa_decl_kernel_dependencies->get (kernel);
2251           if (slot)
2252             {
2253               vec <const char *> *dependencies = *slot;
2254               count = dependencies->length ();
2255
2256               kernel_dependencies_vector_type
2257                 = build_array_type (build_pointer_type (char_type_node),
2258                                     build_index_type (size_int (count)));
2259               TYPE_ARTIFICIAL (kernel_dependencies_vector_type) = 1;
2260
2261               for (unsigned j = 0; j < count; j++)
2262                 {
2263                   const char *d = (*dependencies)[j];
2264                   len = strlen (d);
2265                   tree dependency_name = build_string (len, d);
2266                   TREE_TYPE (dependency_name)
2267                     = build_array_type (char_type_node,
2268                                         build_index_type (size_int (len)));
2269
2270                   CONSTRUCTOR_APPEND_ELT
2271                     (kernel_dependencies_vec, NULL_TREE,
2272                      build1 (ADDR_EXPR,
2273                              build_pointer_type (TREE_TYPE (dependency_name)),
2274                              dependency_name));
2275                 }
2276             }
2277         }
2278
2279       tree dependencies_count = build_int_cstu (unsigned_type_node, count);
2280
2281       vec<constructor_elt, va_gc> *kernel_info_vec = NULL;
2282       CONSTRUCTOR_APPEND_ELT (kernel_info_vec, NULL_TREE,
2283                               build1 (ADDR_EXPR,
2284                                       build_pointer_type (TREE_TYPE
2285                                                           (kern_name)),
2286                                       kern_name));
2287       CONSTRUCTOR_APPEND_ELT (kernel_info_vec, NULL_TREE, omp_data_size);
2288       CONSTRUCTOR_APPEND_ELT (kernel_info_vec, NULL_TREE,
2289                               gridified_kernel_p_tree);
2290       CONSTRUCTOR_APPEND_ELT (kernel_info_vec, NULL_TREE, dependencies_count);
2291
2292       if (count > 0)
2293         {
2294           ASM_GENERATE_INTERNAL_LABEL (tmp_name, "__hsa_dependencies_list", i);
2295           gcc_checking_assert (kernel_dependencies_vector_type);
2296           tree dependencies_list = build_decl (UNKNOWN_LOCATION, VAR_DECL,
2297                                                get_identifier (tmp_name),
2298                                                kernel_dependencies_vector_type);
2299
2300           TREE_STATIC (dependencies_list) = 1;
2301           TREE_READONLY (dependencies_list) = 1;
2302           TREE_PUBLIC (dependencies_list) = 0;
2303           DECL_ARTIFICIAL (dependencies_list) = 1;
2304           DECL_IGNORED_P (dependencies_list) = 1;
2305           DECL_EXTERNAL (dependencies_list) = 0;
2306           TREE_CONSTANT (dependencies_list) = 1;
2307           DECL_INITIAL (dependencies_list)
2308             = build_constructor (kernel_dependencies_vector_type,
2309                                  kernel_dependencies_vec);
2310           varpool_node::finalize_decl (dependencies_list);
2311
2312           CONSTRUCTOR_APPEND_ELT (kernel_info_vec, NULL_TREE,
2313                                   build1 (ADDR_EXPR,
2314                                           build_pointer_type
2315                                             (TREE_TYPE (dependencies_list)),
2316                                           dependencies_list));
2317         }
2318       else
2319         CONSTRUCTOR_APPEND_ELT (kernel_info_vec, NULL_TREE, null_pointer_node);
2320
2321       tree kernel_info_ctor = build_constructor (kernel_info_type,
2322                                                  kernel_info_vec);
2323
2324       CONSTRUCTOR_APPEND_ELT (kernel_info_vector_vec, NULL_TREE,
2325                               kernel_info_ctor);
2326     }
2327
2328   ASM_GENERATE_INTERNAL_LABEL (tmp_name, "__hsa_kernels", 1);
2329   tree hsa_kernels = build_decl (UNKNOWN_LOCATION, VAR_DECL,
2330                                  get_identifier (tmp_name),
2331                                  kernel_info_vector_type);
2332
2333   TREE_STATIC (hsa_kernels) = 1;
2334   TREE_READONLY (hsa_kernels) = 1;
2335   TREE_PUBLIC (hsa_kernels) = 0;
2336   DECL_ARTIFICIAL (hsa_kernels) = 1;
2337   DECL_IGNORED_P (hsa_kernels) = 1;
2338   DECL_EXTERNAL (hsa_kernels) = 0;
2339   TREE_CONSTANT (hsa_kernels) = 1;
2340   DECL_INITIAL (hsa_kernels) = build_constructor (kernel_info_vector_type,
2341                                                   kernel_info_vector_vec);
2342   varpool_node::finalize_decl (hsa_kernels);
2343   *kernels = hsa_kernels;
2344 }
2345
2346 /* Create a static constructor that will register out brig stuff with
2347    libgomp.  */
2348
2349 static void
2350 hsa_output_libgomp_mapping (tree brig_decl)
2351 {
2352   unsigned kernel_count = hsa_get_number_decl_kernel_mappings ();
2353   unsigned global_variable_count = hsa_global_variable_symbols->elements ();
2354
2355   tree kernels;
2356   tree host_func_table;
2357
2358   hsa_output_kernels (&host_func_table, &kernels);
2359   tree global_vars = hsa_output_global_variables ();
2360
2361   tree hsa_image_desc_type = make_node (RECORD_TYPE);
2362   tree id_f1 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2363                            get_identifier ("brig_module"), ptr_type_node);
2364   DECL_CHAIN (id_f1) = NULL_TREE;
2365   tree id_f2 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2366                            get_identifier ("kernel_count"),
2367                            unsigned_type_node);
2368
2369   DECL_CHAIN (id_f2) = id_f1;
2370   tree id_f3 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2371                            get_identifier ("hsa_kernel_infos"),
2372                            ptr_type_node);
2373   DECL_CHAIN (id_f3) = id_f2;
2374   tree id_f4 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2375                            get_identifier ("global_variable_count"),
2376                            unsigned_type_node);
2377   DECL_CHAIN (id_f4) = id_f3;
2378   tree id_f5 = build_decl (BUILTINS_LOCATION, FIELD_DECL,
2379                            get_identifier ("hsa_global_variable_infos"),
2380                            ptr_type_node);
2381   DECL_CHAIN (id_f5) = id_f4;
2382   finish_builtin_struct (hsa_image_desc_type, "__hsa_image_desc", id_f5,
2383                          NULL_TREE);
2384   TYPE_ARTIFICIAL (hsa_image_desc_type) = 1;
2385
2386   vec<constructor_elt, va_gc> *img_desc_vec = NULL;
2387   CONSTRUCTOR_APPEND_ELT (img_desc_vec, NULL_TREE,
2388                           build_fold_addr_expr (brig_decl));
2389   CONSTRUCTOR_APPEND_ELT (img_desc_vec, NULL_TREE,
2390                           build_int_cstu (unsigned_type_node, kernel_count));
2391   CONSTRUCTOR_APPEND_ELT (img_desc_vec, NULL_TREE,
2392                           build1 (ADDR_EXPR,
2393                                   build_pointer_type (TREE_TYPE (kernels)),
2394                                   kernels));
2395   CONSTRUCTOR_APPEND_ELT (img_desc_vec, NULL_TREE,
2396                           build_int_cstu (unsigned_type_node,
2397                                           global_variable_count));
2398   CONSTRUCTOR_APPEND_ELT (img_desc_vec, NULL_TREE,
2399                           build1 (ADDR_EXPR,
2400                                   build_pointer_type (TREE_TYPE (global_vars)),
2401                                   global_vars));
2402
2403   tree img_desc_ctor = build_constructor (hsa_image_desc_type, img_desc_vec);
2404
2405   char tmp_name[64];
2406   ASM_GENERATE_INTERNAL_LABEL (tmp_name, "__hsa_img_descriptor", 1);
2407   tree hsa_img_descriptor = build_decl (UNKNOWN_LOCATION, VAR_DECL,
2408                                         get_identifier (tmp_name),
2409                                         hsa_image_desc_type);
2410   TREE_STATIC (hsa_img_descriptor) = 1;
2411   TREE_READONLY (hsa_img_descriptor) = 1;
2412   TREE_PUBLIC (hsa_img_descriptor) = 0;
2413   DECL_ARTIFICIAL (hsa_img_descriptor) = 1;
2414   DECL_IGNORED_P (hsa_img_descriptor) = 1;
2415   DECL_EXTERNAL (hsa_img_descriptor) = 0;
2416   TREE_CONSTANT (hsa_img_descriptor) = 1;
2417   DECL_INITIAL (hsa_img_descriptor) = img_desc_ctor;
2418   varpool_node::finalize_decl (hsa_img_descriptor);
2419
2420   /* Construct the "host_table" libgomp expects.  */
2421   tree index_type = build_index_type (build_int_cst (integer_type_node, 4));
2422   tree libgomp_host_table_type = build_array_type (ptr_type_node, index_type);
2423   TYPE_ARTIFICIAL (libgomp_host_table_type) = 1;
2424   vec<constructor_elt, va_gc> *libgomp_host_table_vec = NULL;
2425   tree host_func_table_addr = build_fold_addr_expr (host_func_table);
2426   CONSTRUCTOR_APPEND_ELT (libgomp_host_table_vec, NULL_TREE,
2427                           host_func_table_addr);
2428   offset_int func_table_size
2429     = wi::to_offset (TYPE_SIZE_UNIT (ptr_type_node)) * kernel_count;
2430   CONSTRUCTOR_APPEND_ELT (libgomp_host_table_vec, NULL_TREE,
2431                           fold_build2 (POINTER_PLUS_EXPR,
2432                                        TREE_TYPE (host_func_table_addr),
2433                                        host_func_table_addr,
2434                                        build_int_cst (size_type_node,
2435                                                       func_table_size.to_uhwi
2436                                                       ())));
2437   CONSTRUCTOR_APPEND_ELT (libgomp_host_table_vec, NULL_TREE, null_pointer_node);
2438   CONSTRUCTOR_APPEND_ELT (libgomp_host_table_vec, NULL_TREE, null_pointer_node);
2439   tree libgomp_host_table_ctor = build_constructor (libgomp_host_table_type,
2440                                                     libgomp_host_table_vec);
2441   ASM_GENERATE_INTERNAL_LABEL (tmp_name, "__hsa_libgomp_host_table", 1);
2442   tree hsa_libgomp_host_table = build_decl (UNKNOWN_LOCATION, VAR_DECL,
2443                                             get_identifier (tmp_name),
2444                                             libgomp_host_table_type);
2445
2446   TREE_STATIC (hsa_libgomp_host_table) = 1;
2447   TREE_READONLY (hsa_libgomp_host_table) = 1;
2448   TREE_PUBLIC (hsa_libgomp_host_table) = 0;
2449   DECL_ARTIFICIAL (hsa_libgomp_host_table) = 1;
2450   DECL_IGNORED_P (hsa_libgomp_host_table) = 1;
2451   DECL_EXTERNAL (hsa_libgomp_host_table) = 0;
2452   TREE_CONSTANT (hsa_libgomp_host_table) = 1;
2453   DECL_INITIAL (hsa_libgomp_host_table) = libgomp_host_table_ctor;
2454   varpool_node::finalize_decl (hsa_libgomp_host_table);
2455
2456   /* Generate an initializer with a call to the registration routine.  */
2457
2458   tree offload_register
2459     = builtin_decl_explicit (BUILT_IN_GOMP_OFFLOAD_REGISTER);
2460   gcc_checking_assert (offload_register);
2461
2462   tree *hsa_ctor_stmts = hsa_get_ctor_statements ();
2463   append_to_statement_list
2464     (build_call_expr (offload_register, 4,
2465                       build_int_cstu (unsigned_type_node,
2466                                       GOMP_VERSION_PACK (GOMP_VERSION,
2467                                                          GOMP_VERSION_HSA)),
2468                       build_fold_addr_expr (hsa_libgomp_host_table),
2469                       build_int_cst (integer_type_node, GOMP_DEVICE_HSA),
2470                       build_fold_addr_expr (hsa_img_descriptor)),
2471      hsa_ctor_stmts);
2472
2473   cgraph_build_static_cdtor ('I', *hsa_ctor_stmts, DEFAULT_INIT_PRIORITY);
2474
2475   tree offload_unregister
2476     = builtin_decl_explicit (BUILT_IN_GOMP_OFFLOAD_UNREGISTER);
2477   gcc_checking_assert (offload_unregister);
2478
2479   tree *hsa_dtor_stmts = hsa_get_dtor_statements ();
2480   append_to_statement_list
2481     (build_call_expr (offload_unregister, 4,
2482                       build_int_cstu (unsigned_type_node,
2483                                       GOMP_VERSION_PACK (GOMP_VERSION,
2484                                                          GOMP_VERSION_HSA)),
2485                       build_fold_addr_expr (hsa_libgomp_host_table),
2486                       build_int_cst (integer_type_node, GOMP_DEVICE_HSA),
2487                       build_fold_addr_expr (hsa_img_descriptor)),
2488      hsa_dtor_stmts);
2489   cgraph_build_static_cdtor ('D', *hsa_dtor_stmts, DEFAULT_INIT_PRIORITY);
2490 }
2491
2492 /* Emit the brig module we have compiled to a section in the final assembly and
2493    also create a compile unit static constructor that will register the brig
2494    module with libgomp.  */
2495
2496 void
2497 hsa_output_brig (void)
2498 {
2499   section *saved_section;
2500
2501   if (!brig_initialized)
2502     return;
2503
2504   for (unsigned i = 0; i < function_call_linkage.length (); i++)
2505     {
2506       function_linkage_pair p = function_call_linkage[i];
2507
2508       BrigCodeOffset32_t *func_offset = function_offsets->get (p.function_decl);
2509       gcc_assert (*func_offset);
2510       BrigOperandCodeRef *code_ref
2511         = (BrigOperandCodeRef *) (brig_operand.get_ptr_by_offset (p.offset));
2512       gcc_assert (code_ref->base.kind == BRIG_KIND_OPERAND_CODE_REF);
2513       code_ref->ref = lendian32 (*func_offset);
2514     }
2515
2516   /* Iterate all function declarations and if we meet a function that should
2517      have module linkage and we are unable to emit HSAIL for the function,
2518      then change the linkage to program linkage.  Doing so, we will emit
2519      a valid BRIG image.  */
2520   if (hsa_failed_functions != NULL && emitted_declarations != NULL)
2521     for (hash_map <tree, BrigDirectiveExecutable *>::iterator it
2522          = emitted_declarations->begin ();
2523          it != emitted_declarations->end ();
2524          ++it)
2525       {
2526         if (hsa_failed_functions->contains ((*it).first))
2527           (*it).second->linkage = BRIG_LINKAGE_PROGRAM;
2528       }
2529
2530   saved_section = in_section;
2531
2532   switch_to_section (get_section (BRIG_ELF_SECTION_NAME, SECTION_NOTYPE, NULL));
2533   char tmp_name[64];
2534   ASM_GENERATE_INTERNAL_LABEL (tmp_name, BRIG_LABEL_STRING, 1);
2535   ASM_OUTPUT_LABEL (asm_out_file, tmp_name);
2536   tree brig_id = get_identifier (tmp_name);
2537   tree brig_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, brig_id,
2538                                char_type_node);
2539   SET_DECL_ASSEMBLER_NAME (brig_decl, brig_id);
2540   TREE_ADDRESSABLE (brig_decl) = 1;
2541   TREE_READONLY (brig_decl) = 1;
2542   DECL_ARTIFICIAL (brig_decl) = 1;
2543   DECL_IGNORED_P (brig_decl) = 1;
2544   TREE_STATIC (brig_decl) = 1;
2545   TREE_PUBLIC (brig_decl) = 0;
2546   TREE_USED (brig_decl) = 1;
2547   DECL_INITIAL (brig_decl) = brig_decl;
2548   TREE_ASM_WRITTEN (brig_decl) = 1;
2549
2550   BrigModuleHeader module_header;
2551   memcpy (&module_header.identification, "HSA BRIG",
2552           sizeof (module_header.identification));
2553   module_header.brigMajor = lendian32 (BRIG_VERSION_BRIG_MAJOR);
2554   module_header.brigMinor = lendian32 (BRIG_VERSION_BRIG_MINOR);
2555   uint64_t section_index[3];
2556
2557   int data_padding, code_padding, operand_padding;
2558   data_padding = HSA_SECTION_ALIGNMENT
2559     - brig_data.total_size % HSA_SECTION_ALIGNMENT;
2560   code_padding = HSA_SECTION_ALIGNMENT
2561     - brig_code.total_size % HSA_SECTION_ALIGNMENT;
2562   operand_padding = HSA_SECTION_ALIGNMENT
2563     - brig_operand.total_size % HSA_SECTION_ALIGNMENT;
2564
2565   uint64_t module_size = sizeof (module_header)
2566     + sizeof (section_index)
2567     + brig_data.total_size
2568     + data_padding
2569     + brig_code.total_size
2570     + code_padding
2571     + brig_operand.total_size
2572     + operand_padding;
2573   gcc_assert ((module_size % 16) == 0);
2574   module_header.byteCount = lendian64 (module_size);
2575   memset (&module_header.hash, 0, sizeof (module_header.hash));
2576   module_header.reserved = 0;
2577   module_header.sectionCount = lendian32 (3);
2578   module_header.sectionIndex = lendian64 (sizeof (module_header));
2579   assemble_string ((const char *) &module_header, sizeof (module_header));
2580   uint64_t off = sizeof (module_header) + sizeof (section_index);
2581   section_index[0] = lendian64 (off);
2582   off += brig_data.total_size + data_padding;
2583   section_index[1] = lendian64 (off);
2584   off += brig_code.total_size + code_padding;
2585   section_index[2] = lendian64 (off);
2586   assemble_string ((const char *) &section_index, sizeof (section_index));
2587
2588   char padding[HSA_SECTION_ALIGNMENT];
2589   memset (padding, 0, sizeof (padding));
2590
2591   brig_data.output ();
2592   assemble_string (padding, data_padding);
2593   brig_code.output ();
2594   assemble_string (padding, code_padding);
2595   brig_operand.output ();
2596   assemble_string (padding, operand_padding);
2597
2598   if (saved_section)
2599     switch_to_section (saved_section);
2600
2601   hsa_output_libgomp_mapping (brig_decl);
2602
2603   hsa_free_decl_kernel_mapping ();
2604   brig_release_data ();
2605   hsa_deinit_compilation_unit_data ();
2606
2607   delete emitted_declarations;
2608   emitted_declarations = NULL;
2609   delete function_offsets;
2610   function_offsets = NULL;
2611 }