nsz Git - libfirm/blob - ir/be/sparc/sparc_finish.c

   1 /*
   2  * Copyright (C) 1995-2010 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief    Peephole optimization and legalization of a sparc function
  23  * @author   Matthias Braun
  24  *
  25  * A note on sparc stackpointer (sp) behaviour:
  26  * The ABI expects SPARC_MIN_STACKSIZE bytes to be available at the
  27  * stackpointer. This space will be used to spill register windows,
  28  * and for spilling va_arg arguments (maybe we can optimize this away for
  29  * statically known not-va-arg-functions...)
  30  * This in effect means that we allocate that extra space at the function begin
  31  * which is easy. But this space isn't really fixed at the beginning of the
  32  * stackframe. Instead you should rather imagine the space as always being the
  33  * last-thing on the stack.
  34  * So when addressing anything stack-specific we have to account for this
  35  * area, while our compiler thinks the space is occupied at the beginning
  36  * of the stack frame. The code here among other things adjusts these offsets
  37  * accordingly.
  38  */
  39 #include "config.h"
  40
  41 #include "bearch_sparc_t.h"
  42 #include "gen_sparc_regalloc_if.h"
  43 #include "sparc_new_nodes.h"
  44 #include "sparc_transform.h"
  45 #include "irprog.h"
  46 #include "irgmod.h"
  47 #include "ircons.h"
  48 #include "irgwalk.h"
  49 #include "heights.h"
  50
  51 #include "bepeephole.h"
  52 #include "benode.h"
  53 #include "besched.h"
  54 #include "bespillslots.h"
  55 #include "bestack.h"
  56 #include "beirgmod.h"
  57
  58 static ir_heights_t *heights;
  59
  60 static void kill_unused_stacknodes(ir_node *node)
  61 {
  62         if (get_irn_n_edges(node) > 0)
  63                 return;
  64
  65         if (be_is_IncSP(node)) {
  66                 sched_remove(node);
  67                 kill_node(node);
  68         } else if (is_Phi(node)) {
  69                 int       arity = get_irn_arity(node);
  70                 ir_node **ins   = ALLOCAN(ir_node*, arity);
  71                 int       i;
  72                 sched_remove(node);
  73                 memcpy(ins, get_irn_in(node), arity*sizeof(ins[0]));
  74                 kill_node(node);
  75
  76                 for (i = 0; i < arity; ++i)
  77                         kill_unused_stacknodes(ins[i]);
  78         }
  79 }
  80
  81 static void introduce_epilog(ir_node *ret)
  82 {
  83         const arch_register_t *sp_reg     = &sparc_registers[REG_SP];
  84         ir_graph              *irg        = get_irn_irg(ret);
  85         be_stack_layout_t     *layout     = be_get_irg_stack_layout(irg);
  86         ir_node               *block      = get_nodes_block(ret);
  87         ir_type               *frame_type = get_irg_frame_type(irg);
  88         unsigned               frame_size = get_type_size_bytes(frame_type);
  89         int                    sp_idx     = be_find_return_reg_input(ret, sp_reg);
  90         ir_node               *sp         = get_irn_n(ret, sp_idx);
  91
  92         if (!layout->sp_relative) {
  93                 const arch_register_t *fp_reg = &sparc_registers[REG_FRAME_POINTER];
  94                 const arch_register_t *sp_reg = &sparc_registers[REG_SP];
  95                 ir_node *fp      = be_get_initial_reg_value(irg, fp_reg);
  96                 ir_node *sp      = be_get_initial_reg_value(irg, sp_reg);
  97                 ir_node *restore = new_bd_sparc_RestoreZero(NULL, block, sp, fp);
  98                 sched_add_before(ret, restore);
  99                 arch_set_irn_register(restore, sp_reg);
 100                 set_irn_n(ret, sp_idx, restore);
 101
 102                 kill_unused_stacknodes(sp);
 103         } else {
 104                 ir_node *incsp  = be_new_IncSP(sp_reg, block, sp, -frame_size, 0);
 105                 set_irn_n(ret, sp_idx, incsp);
 106                 sched_add_before(ret, incsp);
 107         }
 108 }
 109
 110 void sparc_introduce_prolog_epilog(ir_graph *irg)
 111 {
 112         const arch_register_t *sp_reg     = &sparc_registers[REG_SP];
 113         ir_node               *start      = get_irg_start(irg);
 114         be_stack_layout_t     *layout     = be_get_irg_stack_layout(irg);
 115         ir_node               *block      = get_nodes_block(start);
 116         ir_node               *initial_sp = be_get_initial_reg_value(irg, sp_reg);
 117         ir_node               *schedpoint = start;
 118         ir_type               *frame_type = get_irg_frame_type(irg);
 119         unsigned               frame_size = get_type_size_bytes(frame_type);
 120
 121         /* introduce epilog for every return node */
 122         {
 123                 ir_node *end_block = get_irg_end_block(irg);
 124                 int      arity     = get_irn_arity(end_block);
 125                 int      i;
 126
 127                 for (i = 0; i < arity; ++i) {
 128                         ir_node *ret = get_irn_n(end_block, i);
 129                         assert(is_sparc_Return(ret));
 130                         introduce_epilog(ret);
 131                 }
 132         }
 133
 134         while (be_is_Keep(sched_next(schedpoint)))
 135                 schedpoint = sched_next(schedpoint);
 136
 137         if (!layout->sp_relative) {
 138                 ir_node *const save = new_bd_sparc_Save_imm(NULL, block, initial_sp, NULL, -(SPARC_MIN_STACKSIZE + frame_size));
 139                 arch_set_irn_register(save, sp_reg);
 140                 sched_add_after(schedpoint, save);
 141                 schedpoint = save;
 142
 143                 edges_reroute_except(initial_sp, save, save);
 144
 145                 /* we still need the Save even if noone is explicitely using the
 146                  * value. (TODO: this isn't 100% correct yet, something at the end of
 147                  * the function should hold the Save, even if we use a restore
 148                  * which just overrides it instead of using the value)
 149                  */
 150                 if (get_irn_n_edges(save) == 0) {
 151                         ir_node *in[] = { save };
 152                         ir_node *keep = be_new_Keep(block, 1, in);
 153                         sched_add_after(schedpoint, keep);
 154                 }
 155         } else {
 156                 ir_node *const incsp = be_new_IncSP(sp_reg, block, initial_sp, frame_size, 0);
 157                 edges_reroute_except(initial_sp, incsp, incsp);
 158                 sched_add_after(schedpoint, incsp);
 159         }
 160 }
 161
 162 /**
 163  * Creates a constant from an immediate value.
 164  */
 165 static ir_node *create_constant_from_immediate(ir_node *node, int offset)
 166 {
 167         dbg_info *dbgi  = get_irn_dbg_info(node);
 168         ir_node  *block = get_nodes_block(node);
 169         ir_node  *high  = new_bd_sparc_SetHi(dbgi, block, NULL, offset);
 170
 171         sched_add_before(node, high);
 172         arch_set_irn_register(high, &sparc_registers[REG_G4]);
 173
 174         if ((offset & 0x3ff) != 0) {
 175                 ir_node *low = new_bd_sparc_Or_imm(dbgi, block, high, NULL, offset & 0x3ff);
 176
 177                 sched_add_before(node, low);
 178                 arch_set_irn_register(low, &sparc_registers[REG_G4]);
 179
 180                 return low;
 181         }
 182
 183         return high;
 184 }
 185
 186 static void finish_sparc_Save(ir_node *node)
 187 {
 188         sparc_attr_t *attr = get_sparc_attr(node);
 189         int offset = attr->immediate_value;
 190
 191         if (! sparc_is_value_imm_encodeable(offset)) {
 192                 ir_node               *base     = get_irn_n(node, n_sparc_Save_stack);
 193                 dbg_info              *dbgi     = get_irn_dbg_info(node);
 194                 ir_node               *block    = get_nodes_block(node);
 195                 ir_node               *constant = create_constant_from_immediate(node, offset);
 196                 ir_node               *new_save = new_bd_sparc_Save_reg(dbgi, block, base, constant);
 197                 const arch_register_t *reg      = arch_get_irn_register(node);
 198
 199                 /* we have a Save with immediate */
 200                 assert(get_irn_arity(node) == 1);
 201
 202                 sched_add_before(node, new_save);
 203                 arch_set_irn_register(new_save, reg);
 204                 be_peephole_exchange(node, new_save);
 205         }
 206 }
 207
 208 /**
 209  * SPARC immediates are limited. Split IncSP with bigger immediates if
 210  * necessary.
 211  */
 212 static void finish_be_IncSP(ir_node *node)
 213 {
 214         int offset = be_get_IncSP_offset(node);
 215
 216         /* we might have to break the IncSP apart if the constant has become too big */
 217         if (! sparc_is_value_imm_encodeable(offset) && ! sparc_is_value_imm_encodeable(-offset)) {
 218                 ir_node               *sp       = be_get_IncSP_pred(node);
 219                 dbg_info              *dbgi     = get_irn_dbg_info(node);
 220                 ir_node               *block    = get_nodes_block(node);
 221                 ir_node               *constant = create_constant_from_immediate(node, offset);
 222                 ir_node               *sub      = new_bd_sparc_Sub_reg(dbgi, block, sp, constant);
 223
 224                 sched_add_before(node, sub);
 225                 arch_set_irn_register(sub, &sparc_registers[REG_SP]);
 226                 be_peephole_exchange(node, sub);
 227         }
 228 }
 229
 230 /**
 231  * Adjust sp-relative offsets.
 232  *
 233  * Split into multiple instructions if offset exceeds SPARC immediate range.
 234  */
 235 static void finish_sparc_FrameAddr(ir_node *node)
 236 {
 237         sparc_attr_t *attr   = get_sparc_attr(node);
 238         int           offset = attr->immediate_value;
 239
 240         if (! sparc_is_value_imm_encodeable(offset)) {
 241                 ir_node               *base          = get_irn_n(node, n_sparc_FrameAddr_base);
 242                 dbg_info              *dbgi          = get_irn_dbg_info(node);
 243                 ir_node               *block         = get_nodes_block(node);
 244                 ir_node               *constant      = create_constant_from_immediate(node, offset);
 245                 ir_node               *new_frameaddr = new_bd_sparc_Add_reg(dbgi, block, base, constant);
 246                 const arch_register_t *reg           = arch_get_irn_register(node);
 247
 248                 sched_add_before(node, new_frameaddr);
 249                 arch_set_irn_register(new_frameaddr, reg);
 250                 be_peephole_exchange(node, new_frameaddr);
 251         }
 252 }
 253
 254 static void finish_sparc_Ld(ir_node *node)
 255 {
 256         sparc_attr_t                  *attr            = get_sparc_attr(node);
 257         int                            offset          = attr->immediate_value;
 258         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 259
 260         if (! load_store_attr->is_frame_entity)
 261                 return;
 262
 263         if (! sparc_is_value_imm_encodeable(offset)) {
 264                 ir_node                 *ptr             = get_irn_n(node, n_sparc_Ld_ptr);
 265                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 266                 ir_node                 *block           = get_nodes_block(node);
 267                 ir_node                 *mem             = get_irn_n(node, n_sparc_Ld_mem);
 268                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 269                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 270                 ir_node                 *new_load        = new_bd_sparc_Ld_reg(dbgi, block, ptr, constant, mem, load_store_mode);
 271                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 272                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 273                 unsigned                 i;
 274
 275                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 276                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 277
 278                 sched_add_before(node, new_load);
 279                 for (i = 0; i < n_outs; i++) {
 280                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 281                 }
 282                 be_peephole_exchange(node, new_load);
 283         }
 284
 285 }
 286
 287 static void split_sparc_ldf(ir_node *node)
 288 {
 289         sparc_load_store_attr_t *attr = get_sparc_load_store_attr(node);
 290         unsigned                 bits = get_mode_size_bits(attr->load_store_mode);
 291         /* split 128bit loads into 2 64bit loads */
 292         if (bits == 128) {
 293                 dbg_info *dbgi  = get_irn_dbg_info(node);
 294                 ir_node  *block = get_nodes_block(node);
 295                 ir_node  *ptr   = get_irn_n(node, n_sparc_Ldf_ptr);
 296                 ir_node  *mem   = get_irn_n(node, n_sparc_Ldf_mem);
 297                 ir_node  *new_load
 298                         = new_bd_sparc_Ldf_d(dbgi, block, ptr, mem, mode_D,
 299                                              attr->base.immediate_value_entity,
 300                                              attr->base.immediate_value + 8,
 301                                              attr->is_frame_entity);
 302                 ir_node  *new_mem = new_r_Proj(new_load, mode_M, pn_sparc_Ldf_M);
 303
 304                 const arch_register_t *reg
 305                         = arch_get_irn_register_out(node, pn_sparc_Ldf_res);
 306                 unsigned reg_index = reg->global_index;
 307
 308                 arch_set_irn_register_out(new_load, pn_sparc_Ldf_res,
 309                                           &sparc_registers[reg_index+2]);
 310
 311                 attr->load_store_mode = mode_D;
 312                 set_irn_n(node, n_sparc_Ldf_mem, new_mem);
 313                 sched_add_before(node, new_load);
 314         }
 315 }
 316
 317 static void finish_sparc_Ldf(ir_node *node)
 318 {
 319         sparc_attr_t                  *attr            = get_sparc_attr(node);
 320         int                            offset          = attr->immediate_value;
 321         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 322
 323         if (! load_store_attr->is_frame_entity)
 324                 return;
 325
 326         if (! sparc_is_value_imm_encodeable(offset)) {
 327                 ir_node                 *ptr             = get_irn_n(node, n_sparc_Ldf_ptr);
 328                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 329                 ir_node                 *block           = get_nodes_block(node);
 330                 ir_node                 *mem             = get_irn_n(node, n_sparc_Ldf_mem);
 331                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 332                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 333                 ir_node                 *new_ptr         = new_bd_sparc_Add_reg(dbgi, block, ptr, constant);
 334                 ir_node                 *new_load        = new_bd_sparc_Ldf_s(dbgi, block, new_ptr, mem, load_store_mode, NULL, 0, true);
 335                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 336                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 337                 unsigned                 i;
 338
 339                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 340                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 341
 342                 sched_add_before(node, new_load);
 343                 for (i = 0; i < n_outs; i++) {
 344                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 345                 }
 346                 be_peephole_exchange(node, new_load);
 347         }
 348
 349 }
 350
 351 static void finish_sparc_St(ir_node *node)
 352 {
 353         sparc_attr_t                  *attr            = get_sparc_attr(node);
 354         int                            offset          = attr->immediate_value;
 355         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 356
 357         if (! load_store_attr->is_frame_entity)
 358                 return;
 359
 360         if (! sparc_is_value_imm_encodeable(offset)) {
 361                 ir_node                 *ptr             = get_irn_n(node, n_sparc_St_ptr);
 362                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 363                 ir_node                 *block           = get_nodes_block(node);
 364                 ir_node                 *mem             = get_irn_n(node, n_sparc_St_mem);
 365                 ir_node                 *value           = get_irn_n(node, n_sparc_St_val);
 366                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 367                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 368                 ir_node                 *new_load        = new_bd_sparc_St_reg(dbgi, block, value, ptr, constant, mem, load_store_mode);
 369                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 370                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 371                 unsigned                 i;
 372
 373                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 374                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 375
 376                 sched_add_before(node, new_load);
 377                 for (i = 0; i < n_outs; i++) {
 378                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 379                 }
 380                 be_peephole_exchange(node, new_load);
 381         }
 382
 383 }
 384
 385 static void finish_sparc_Stf(ir_node *node)
 386 {
 387         sparc_attr_t                  *attr            = get_sparc_attr(node);
 388         int                            offset          = attr->immediate_value;
 389         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 390
 391         if (! load_store_attr->is_frame_entity)
 392                 return;
 393
 394         if (! sparc_is_value_imm_encodeable(offset)) {
 395                 ir_node                 *ptr             = get_irn_n(node, n_sparc_Stf_ptr);
 396                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 397                 ir_node                 *block           = get_nodes_block(node);
 398                 ir_node                 *mem             = get_irn_n(node, n_sparc_Stf_mem);
 399                 ir_node                 *value           = get_irn_n(node, n_sparc_Stf_val);
 400                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 401                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 402                 ir_node                 *new_ptr         = new_bd_sparc_Add_reg(dbgi, block, ptr, constant);
 403                 ir_node                 *new_load        = new_bd_sparc_Stf_s(dbgi, block, value, new_ptr, mem, load_store_mode, NULL, 0, true);
 404                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 405                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 406                 unsigned                 i;
 407
 408                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 409                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 410
 411                 sched_add_before(node, new_load);
 412                 for (i = 0; i < n_outs; i++) {
 413                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 414                 }
 415                 be_peephole_exchange(node, new_load);
 416         }
 417
 418 }
 419
 420 static void peephole_be_IncSP(ir_node *node)
 421 {
 422         ir_node *pred;
 423         node = be_peephole_IncSP_IncSP(node);
 424         if (!be_is_IncSP(node))
 425                 return;
 426
 427         pred = be_get_IncSP_pred(node);
 428         if (is_sparc_Save(pred) && be_has_only_one_user(pred)) {
 429                 int offset = -be_get_IncSP_offset(node);
 430                 sparc_attr_t *attr = get_sparc_attr(pred);
 431                 attr->immediate_value += offset;
 432                 be_peephole_exchange(node, pred);
 433         }
 434 }
 435
 436 static void peephole_sparc_FrameAddr(ir_node *node)
 437 {
 438         /* the peephole code currently doesn't allow this since it changes
 439          * the register. Find out why and how to workaround this... */
 440 #if 0
 441         const sparc_attr_t *attr = get_sparc_attr_const(node);
 442         if (attr->immediate_value == 0) {
 443                 ir_node *base = get_irn_n(node, n_sparc_FrameAddr_base);
 444                 be_peephole_exchange(node, base);
 445         }
 446 #endif
 447         (void) node;
 448 }
 449
 450 /* output must not be local, or out reg. Since the destination of the restore
 451  * is the rotated register-file where only the old in-registers are still
 452  * visible (as out-registers) */
 453 static bool is_restorezeroopt_reg(const arch_register_t *reg)
 454 {
 455         unsigned index = reg->global_index;
 456         return (index >= REG_G0 && index <= REG_G7)
 457             || (index >= REG_I0 && index <= REG_I7);
 458 }
 459
 460 static void replace_with_restore_reg(ir_node *node, ir_node *replaced,
 461                                                                          ir_node *op0, ir_node *op1)
 462 {
 463         dbg_info *dbgi     = get_irn_dbg_info(node);
 464         ir_node  *stack_in = get_irn_n(node, n_sparc_RestoreZero_stack);
 465         ir_node  *fp       = get_irn_n(node, n_sparc_RestoreZero_frame_pointer);
 466         ir_node  *block    = get_nodes_block(node);
 467         ir_mode  *mode     = get_irn_mode(node);
 468         ir_node  *new_node = new_bd_sparc_Restore_reg(dbgi, block, stack_in, fp,
 469                                                       op0, op1);
 470         ir_node  *stack    = new_r_Proj(new_node, mode, pn_sparc_Restore_stack);
 471         ir_node  *res      = new_r_Proj(new_node, mode, pn_sparc_Restore_res);
 472         const arch_register_t *reg = arch_get_irn_register(replaced);
 473         const arch_register_t *sp  = &sparc_registers[REG_SP];
 474         arch_set_irn_register_out(new_node, pn_sparc_Restore_stack, sp);
 475         arch_set_irn_register_out(new_node, pn_sparc_Restore_res, reg);
 476
 477         sched_add_before(node, new_node);
 478         be_peephole_exchange(node, stack);
 479         be_peephole_exchange(replaced, res);
 480 }
 481
 482 static void replace_with_restore_imm(ir_node *node, ir_node *replaced,
 483                                                                          ir_node *op, ir_entity *imm_entity,
 484                                                                          int32_t immediate)
 485 {
 486         dbg_info *dbgi     = get_irn_dbg_info(node);
 487         ir_node  *stack_in = get_irn_n(node, n_sparc_RestoreZero_stack);
 488         ir_node  *fp       = get_irn_n(node, n_sparc_RestoreZero_frame_pointer);
 489         ir_node  *block    = get_nodes_block(node);
 490         ir_mode  *mode     = get_irn_mode(node);
 491         ir_node  *new_node = new_bd_sparc_Restore_imm(dbgi, block, stack_in, fp,
 492                                                       op, imm_entity, immediate);
 493         ir_node  *stack    = new_r_Proj(new_node, mode, pn_sparc_Restore_stack);
 494         ir_node  *res      = new_r_Proj(new_node, mode, pn_sparc_Restore_res);
 495         const arch_register_t *reg = arch_get_irn_register(replaced);
 496         const arch_register_t *sp  = &sparc_registers[REG_SP];
 497         arch_set_irn_register_out(new_node, pn_sparc_Restore_stack, sp);
 498         arch_set_irn_register_out(new_node, pn_sparc_Restore_res, reg);
 499
 500         sched_add_before(node, new_node);
 501         be_peephole_exchange(node, stack);
 502         be_peephole_exchange(replaced, res);
 503 }
 504
 505 static void peephole_sparc_RestoreZero(ir_node *node)
 506 {
 507         /* restore gives us a free "add" instruction, let's try to use that to fold
 508          * an instruction in. We can do the following:
 509          *
 510          * - Copy values                  (g0 + reg)
 511          * - Produce constants            (g0 + immediate)
 512          * - Perform an add               (reg + reg)
 513          * - Perform a sub with immediate (reg + (-immediate))
 514          *
 515          * Note: In an ideal world, this would not be a peephole optimization but
 516          * already performed during code selection. Since about all foldable ops are
 517          * arguments of the return node. However we have a hard time doing this
 518          * since we construct epilogue code only after register allocation
 519          * (and therefore after code selection).
 520          */
 521         int n_tries = 10; /* limit our search */
 522         ir_node *schedpoint = node;
 523
 524         while (sched_has_prev(schedpoint)) {
 525                 const arch_register_t *reg;
 526                 schedpoint = sched_prev(schedpoint);
 527
 528                 if (--n_tries == 0)
 529                         break;
 530
 531                 if (arch_get_irn_n_outs(schedpoint) == 0)
 532                         continue;
 533
 534                 if (!mode_is_data(get_irn_mode(schedpoint)))
 535                         return;
 536
 537                 reg = arch_get_irn_register(schedpoint);
 538                 if (!is_restorezeroopt_reg(reg))
 539                         continue;
 540
 541                 if (be_is_Copy(schedpoint) && be_can_move_down(heights, schedpoint, node)) {
 542                         ir_node *const op = be_get_Copy_op(schedpoint);
 543                         replace_with_restore_imm(node, schedpoint, op, NULL, 0);
 544                 } else if (is_sparc_Or(schedpoint) &&
 545                            arch_get_irn_flags(schedpoint) & ((arch_irn_flags_t)sparc_arch_irn_flag_immediate_form) &&
 546                            arch_get_irn_register_in(schedpoint, 0) == &sparc_registers[REG_G0] &&
 547                            be_can_move_down(heights, schedpoint, node)) {
 548                         /* it's a constant */
 549                         const sparc_attr_t *attr      = get_sparc_attr_const(schedpoint);
 550                         ir_entity          *entity    = attr->immediate_value_entity;
 551                         int32_t             immediate = attr->immediate_value;
 552                         ir_node            *g0        = get_irn_n(schedpoint, 0);
 553                         replace_with_restore_imm(node, schedpoint, g0, entity, immediate);
 554                 } else if (is_sparc_Add(schedpoint) &&
 555                            be_can_move_down(heights, schedpoint, node)) {
 556                         if (arch_get_irn_flags(schedpoint) & ((arch_irn_flags_t)sparc_arch_irn_flag_immediate_form)) {
 557                                 ir_node            *op     = get_irn_n(schedpoint, 0);
 558                                 const sparc_attr_t *attr   = get_sparc_attr_const(schedpoint);
 559                                 ir_entity          *entity = attr->immediate_value_entity;
 560                                 int32_t             imm    = attr->immediate_value;
 561                                 replace_with_restore_imm(node, schedpoint, op, entity, imm);
 562                         } else {
 563                                 ir_node *op0 = get_irn_n(schedpoint, 0);
 564                                 ir_node *op1 = get_irn_n(schedpoint, 1);
 565                                 replace_with_restore_reg(node, schedpoint, op0, op1);
 566                         }
 567                 } else if (is_sparc_Sub(schedpoint) &&
 568                            arch_get_irn_flags(schedpoint) & ((arch_irn_flags_t)sparc_arch_irn_flag_immediate_form) &&
 569                            arch_get_irn_register_in(schedpoint, 0) == &sparc_registers[REG_G0] &&
 570                            be_can_move_down(heights, schedpoint, node)) {
 571                         /* it's a constant */
 572                         const sparc_attr_t *attr   = get_sparc_attr_const(schedpoint);
 573                         ir_entity          *entity = attr->immediate_value_entity;
 574                         int32_t             imm    = attr->immediate_value;
 575                         if (entity == NULL && sparc_is_value_imm_encodeable(-imm)) {
 576                                 ir_node *g0 = get_irn_n(schedpoint, 0);
 577                                 replace_with_restore_imm(node, schedpoint, g0, NULL, -imm);
 578                         } else {
 579                                 continue;
 580                         }
 581                 }
 582                 /* when we're here then we performed a folding and are done */
 583                 return;
 584         }
 585 }
 586
 587 static void finish_sparc_Return(ir_node *node)
 588 {
 589         ir_node *schedpoint = node;
 590         ir_node *restore;
 591         /* see that there is no code between Return and restore, if there is move
 592          * it in front of the restore */
 593         while (true) {
 594                 if (!sched_has_prev(schedpoint))
 595                         return;
 596                 schedpoint = sched_prev(schedpoint);
 597                 if (is_sparc_Restore(schedpoint) || is_sparc_RestoreZero(schedpoint))
 598                         break;
 599         }
 600         restore = schedpoint;
 601         schedpoint = sched_prev(node);
 602         /* move all code between return and restore up */
 603         while (schedpoint != restore) {
 604                 ir_node *next_schedpoint = sched_prev(schedpoint);
 605                 sched_remove(schedpoint);
 606                 sched_add_before(restore, schedpoint);
 607                 schedpoint = next_schedpoint;
 608         }
 609 }
 610
 611 static void register_peephole_optimisation(ir_op *op, peephole_opt_func func)
 612 {
 613         assert(op->ops.generic == NULL);
 614         op->ops.generic = (op_func) func;
 615 }
 616
 617 static void sparc_collect_frame_entity_nodes(ir_node *node, void *data)
 618 {
 619         be_fec_env_t  *env = (be_fec_env_t*)data;
 620         const ir_mode *mode;
 621         int            align;
 622         ir_entity     *entity;
 623         const sparc_load_store_attr_t *attr;
 624
 625         if (be_is_Reload(node) && be_get_frame_entity(node) == NULL) {
 626                 mode  = get_irn_mode(node);
 627                 align = get_mode_size_bytes(mode);
 628                 be_node_needs_frame_entity(env, node, mode, align);
 629                 return;
 630         }
 631
 632         if (!is_sparc_Ld(node) && !is_sparc_Ldf(node))
 633                 return;
 634
 635         attr   = get_sparc_load_store_attr_const(node);
 636         entity = attr->base.immediate_value_entity;
 637         mode   = attr->load_store_mode;
 638         if (entity != NULL)
 639                 return;
 640         if (!attr->is_frame_entity)
 641                 return;
 642         if (arch_get_irn_flags(node) & sparc_arch_irn_flag_needs_64bit_spillslot)
 643                 mode = mode_Lu;
 644         align  = get_mode_size_bytes(mode);
 645         be_node_needs_frame_entity(env, node, mode, align);
 646 }
 647
 648 static void sparc_set_frame_entity(ir_node *node, ir_entity *entity)
 649 {
 650         if (is_be_node(node)) {
 651                 be_node_set_frame_entity(node, entity);
 652         } else {
 653                 /* we only say be_node_needs_frame_entity on nodes with load_store
 654                  * attributes, so this should be fine */
 655                 sparc_load_store_attr_t *attr = get_sparc_load_store_attr(node);
 656                 assert(attr->is_frame_entity);
 657                 assert(attr->base.immediate_value_entity == NULL);
 658                 attr->base.immediate_value_entity = entity;
 659         }
 660 }
 661
 662 void sparc_finish_graph(ir_graph *irg)
 663 {
 664         be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
 665         bool               at_begin     = stack_layout->sp_relative ? true : false;
 666         be_fec_env_t      *fec_env      = be_new_frame_entity_coalescer(irg);
 667
 668         irg_walk_graph(irg, NULL, sparc_collect_frame_entity_nodes, fec_env);
 669         be_assign_entities(fec_env, sparc_set_frame_entity, at_begin);
 670         be_free_frame_entity_coalescer(fec_env);
 671         sparc_adjust_stack_entity_offsets(irg);
 672
 673         sparc_introduce_prolog_epilog(irg);
 674
 675         /* fix stack entity offsets */
 676         be_abi_fix_stack_nodes(irg);
 677         sparc_fix_stack_bias(irg);
 678
 679         heights = heights_new(irg);
 680
 681         /* perform peephole optimizations */
 682         ir_clear_opcodes_generic_func();
 683         register_peephole_optimisation(op_be_IncSP,        peephole_be_IncSP);
 684         register_peephole_optimisation(op_sparc_FrameAddr, peephole_sparc_FrameAddr);
 685         register_peephole_optimisation(op_sparc_RestoreZero,
 686                                        peephole_sparc_RestoreZero);
 687         register_peephole_optimisation(op_sparc_Ldf, split_sparc_ldf);
 688         be_peephole_opt(irg);
 689
 690         /* perform legalizations (mostly fix nodes with too big immediates) */
 691         ir_clear_opcodes_generic_func();
 692         register_peephole_optimisation(op_be_IncSP,        finish_be_IncSP);
 693         register_peephole_optimisation(op_sparc_FrameAddr, finish_sparc_FrameAddr);
 694         register_peephole_optimisation(op_sparc_Ld,        finish_sparc_Ld);
 695         register_peephole_optimisation(op_sparc_Ldf,       finish_sparc_Ldf);
 696         register_peephole_optimisation(op_sparc_Return,    finish_sparc_Return);
 697         register_peephole_optimisation(op_sparc_Save,      finish_sparc_Save);
 698         register_peephole_optimisation(op_sparc_St,        finish_sparc_St);
 699         register_peephole_optimisation(op_sparc_Stf,       finish_sparc_Stf);
 700         be_peephole_opt(irg);
 701
 702         heights_free(heights);
 703
 704         be_remove_dead_nodes_from_schedule(irg);
 705 }