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