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               *sp         = initial_sp;
 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 *save = new_bd_sparc_Save_imm(NULL, block, sp, NULL,
 140                                                       -SPARC_MIN_STACKSIZE-frame_size);
 141                 arch_set_irn_register(save, sp_reg);
 142                 sched_add_after(schedpoint, save);
 143                 schedpoint = save;
 144
 145                 edges_reroute(initial_sp, save);
 146                 set_irn_n(save, n_sparc_Save_stack, initial_sp);
 147
 148                 /* we still need the Save even if noone is explicitely using the
 149                  * value. (TODO: this isn't 100% correct yet, something at the end of
 150                  * the function should hold the Save, even if we use a restore
 151                  * which just overrides it instead of using the value)
 152                  */
 153                 if (get_irn_n_edges(save) == 0) {
 154                         ir_node *in[] = { save };
 155                         ir_node *keep = be_new_Keep(block, 1, in);
 156                         sched_add_after(schedpoint, keep);
 157                 }
 158         } else {
 159                 ir_node *incsp = be_new_IncSP(sp_reg, block, sp, frame_size, 0);
 160                 edges_reroute(initial_sp, incsp);
 161                 be_set_IncSP_pred(incsp, sp);
 162                 sched_add_after(schedpoint, incsp);
 163         }
 164 }
 165
 166 /**
 167  * Creates a constant from an immediate value.
 168  */
 169 static ir_node *create_constant_from_immediate(ir_node *node, int offset)
 170 {
 171         dbg_info *dbgi  = get_irn_dbg_info(node);
 172         ir_node  *block = get_nodes_block(node);
 173         ir_node  *high  = new_bd_sparc_SetHi(dbgi, block, NULL, offset);
 174
 175         sched_add_before(node, high);
 176         arch_set_irn_register(high, &sparc_registers[REG_G4]);
 177
 178         if ((offset & 0x3ff) != 0) {
 179                 ir_node *low = new_bd_sparc_Or_imm(dbgi, block, high, NULL, offset & 0x3ff);
 180
 181                 sched_add_before(node, low);
 182                 arch_set_irn_register(low, &sparc_registers[REG_G4]);
 183
 184                 return low;
 185         }
 186
 187         return high;
 188 }
 189
 190 static void finish_sparc_Save(ir_node *node)
 191 {
 192         sparc_attr_t *attr = get_sparc_attr(node);
 193         int offset = attr->immediate_value;
 194
 195         if (! sparc_is_value_imm_encodeable(offset)) {
 196                 ir_node               *base     = get_irn_n(node, n_sparc_Save_stack);
 197                 dbg_info              *dbgi     = get_irn_dbg_info(node);
 198                 ir_node               *block    = get_nodes_block(node);
 199                 ir_node               *constant = create_constant_from_immediate(node, offset);
 200                 ir_node               *new_save = new_bd_sparc_Save_reg(dbgi, block, base, constant);
 201                 const arch_register_t *reg      = arch_get_irn_register(node);
 202
 203                 /* we have a Save with immediate */
 204                 assert(get_irn_arity(node) == 1);
 205
 206                 sched_add_before(node, new_save);
 207                 arch_set_irn_register(new_save, reg);
 208                 be_peephole_exchange(node, new_save);
 209         }
 210 }
 211
 212 /**
 213  * SPARC immediates are limited. Split IncSP with bigger immediates if
 214  * necessary.
 215  */
 216 static void finish_be_IncSP(ir_node *node)
 217 {
 218         int offset = be_get_IncSP_offset(node);
 219
 220         /* we might have to break the IncSP apart if the constant has become too big */
 221         if (! sparc_is_value_imm_encodeable(offset) && ! sparc_is_value_imm_encodeable(-offset)) {
 222                 ir_node               *sp       = be_get_IncSP_pred(node);
 223                 dbg_info              *dbgi     = get_irn_dbg_info(node);
 224                 ir_node               *block    = get_nodes_block(node);
 225                 ir_node               *constant = create_constant_from_immediate(node, offset);
 226                 ir_node               *sub      = new_bd_sparc_Sub_reg(dbgi, block, sp, constant);
 227
 228                 sched_add_before(node, sub);
 229                 arch_set_irn_register(sub, &sparc_registers[REG_SP]);
 230                 be_peephole_exchange(node, sub);
 231         }
 232 }
 233
 234 /**
 235  * Adjust sp-relative offsets.
 236  *
 237  * Split into multiple instructions if offset exceeds SPARC immediate range.
 238  */
 239 static void finish_sparc_FrameAddr(ir_node *node)
 240 {
 241         sparc_attr_t *attr   = get_sparc_attr(node);
 242         int           offset = attr->immediate_value;
 243
 244         if (! sparc_is_value_imm_encodeable(offset)) {
 245                 ir_node               *base          = get_irn_n(node, n_sparc_FrameAddr_base);
 246                 dbg_info              *dbgi          = get_irn_dbg_info(node);
 247                 ir_node               *block         = get_nodes_block(node);
 248                 ir_node               *constant      = create_constant_from_immediate(node, offset);
 249                 ir_node               *new_frameaddr = new_bd_sparc_Add_reg(dbgi, block, base, constant);
 250                 const arch_register_t *reg           = arch_get_irn_register(node);
 251
 252                 sched_add_before(node, new_frameaddr);
 253                 arch_set_irn_register(new_frameaddr, reg);
 254                 be_peephole_exchange(node, new_frameaddr);
 255         }
 256 }
 257
 258 static void finish_sparc_Ld(ir_node *node)
 259 {
 260         sparc_attr_t                  *attr            = get_sparc_attr(node);
 261         int                            offset          = attr->immediate_value;
 262         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 263
 264         if (! load_store_attr->is_frame_entity)
 265                 return;
 266
 267         if (! sparc_is_value_imm_encodeable(offset)) {
 268                 ir_node                 *ptr             = get_irn_n(node, n_sparc_Ld_ptr);
 269                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 270                 ir_node                 *block           = get_nodes_block(node);
 271                 ir_node                 *mem             = get_irn_n(node, n_sparc_Ld_mem);
 272                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 273                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 274                 ir_node                 *new_load        = new_bd_sparc_Ld_reg(dbgi, block, ptr, constant, mem, load_store_mode);
 275                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 276                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 277                 unsigned                 i;
 278
 279                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 280                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 281
 282                 sched_add_before(node, new_load);
 283                 for (i = 0; i < n_outs; i++) {
 284                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 285                 }
 286                 be_peephole_exchange(node, new_load);
 287         }
 288
 289 }
 290
 291 static void split_sparc_ldf(ir_node *node)
 292 {
 293         sparc_load_store_attr_t *attr = get_sparc_load_store_attr(node);
 294         unsigned                 bits = get_mode_size_bits(attr->load_store_mode);
 295         /* split 128bit loads into 2 64bit loads */
 296         if (bits == 128) {
 297                 dbg_info *dbgi  = get_irn_dbg_info(node);
 298                 ir_node  *block = get_nodes_block(node);
 299                 ir_node  *ptr   = get_irn_n(node, n_sparc_Ldf_ptr);
 300                 ir_node  *mem   = get_irn_n(node, n_sparc_Ldf_mem);
 301                 ir_node  *new_load
 302                         = new_bd_sparc_Ldf_d(dbgi, block, ptr, mem, mode_D,
 303                                              attr->base.immediate_value_entity,
 304                                              attr->base.immediate_value + 8,
 305                                              attr->is_frame_entity);
 306                 ir_node  *new_mem = new_r_Proj(new_load, mode_M, pn_sparc_Ldf_M);
 307
 308                 const arch_register_t *reg
 309                         = arch_get_irn_register_out(node, pn_sparc_Ldf_res);
 310                 unsigned reg_index = reg->global_index;
 311
 312                 arch_set_irn_register_out(new_load, pn_sparc_Ldf_res,
 313                                           &sparc_registers[reg_index+2]);
 314
 315                 attr->load_store_mode = mode_D;
 316                 set_irn_n(node, n_sparc_Ldf_mem, new_mem);
 317                 sched_add_before(node, new_load);
 318         }
 319 }
 320
 321 static void finish_sparc_Ldf(ir_node *node)
 322 {
 323         sparc_attr_t                  *attr            = get_sparc_attr(node);
 324         int                            offset          = attr->immediate_value;
 325         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 326
 327         if (! load_store_attr->is_frame_entity)
 328                 return;
 329
 330         if (! sparc_is_value_imm_encodeable(offset)) {
 331                 ir_node                 *ptr             = get_irn_n(node, n_sparc_Ldf_ptr);
 332                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 333                 ir_node                 *block           = get_nodes_block(node);
 334                 ir_node                 *mem             = get_irn_n(node, n_sparc_Ldf_mem);
 335                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 336                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 337                 ir_node                 *new_ptr         = new_bd_sparc_Add_reg(dbgi, block, ptr, constant);
 338                 ir_node                 *new_load        = new_bd_sparc_Ldf_s(dbgi, block, new_ptr, mem, load_store_mode, NULL, 0, true);
 339                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 340                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 341                 unsigned                 i;
 342
 343                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 344                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 345
 346                 sched_add_before(node, new_load);
 347                 for (i = 0; i < n_outs; i++) {
 348                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 349                 }
 350                 be_peephole_exchange(node, new_load);
 351         }
 352
 353 }
 354
 355 static void finish_sparc_St(ir_node *node)
 356 {
 357         sparc_attr_t                  *attr            = get_sparc_attr(node);
 358         int                            offset          = attr->immediate_value;
 359         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 360
 361         if (! load_store_attr->is_frame_entity)
 362                 return;
 363
 364         if (! sparc_is_value_imm_encodeable(offset)) {
 365                 ir_node                 *ptr             = get_irn_n(node, n_sparc_St_ptr);
 366                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 367                 ir_node                 *block           = get_nodes_block(node);
 368                 ir_node                 *mem             = get_irn_n(node, n_sparc_St_mem);
 369                 ir_node                 *value           = get_irn_n(node, n_sparc_St_val);
 370                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 371                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 372                 ir_node                 *new_load        = new_bd_sparc_St_reg(dbgi, block, value, ptr, constant, mem, load_store_mode);
 373                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 374                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 375                 unsigned                 i;
 376
 377                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 378                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 379
 380                 sched_add_before(node, new_load);
 381                 for (i = 0; i < n_outs; i++) {
 382                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 383                 }
 384                 be_peephole_exchange(node, new_load);
 385         }
 386
 387 }
 388
 389 static void finish_sparc_Stf(ir_node *node)
 390 {
 391         sparc_attr_t                  *attr            = get_sparc_attr(node);
 392         int                            offset          = attr->immediate_value;
 393         const sparc_load_store_attr_t *load_store_attr = get_sparc_load_store_attr_const(node);
 394
 395         if (! load_store_attr->is_frame_entity)
 396                 return;
 397
 398         if (! sparc_is_value_imm_encodeable(offset)) {
 399                 ir_node                 *ptr             = get_irn_n(node, n_sparc_Stf_ptr);
 400                 dbg_info                *dbgi            = get_irn_dbg_info(node);
 401                 ir_node                 *block           = get_nodes_block(node);
 402                 ir_node                 *mem             = get_irn_n(node, n_sparc_Stf_mem);
 403                 ir_node                 *value           = get_irn_n(node, n_sparc_Stf_val);
 404                 ir_mode                 *load_store_mode = load_store_attr->load_store_mode;
 405                 ir_node                 *constant        = create_constant_from_immediate(node, offset);
 406                 ir_node                 *new_ptr         = new_bd_sparc_Add_reg(dbgi, block, ptr, constant);
 407                 ir_node                 *new_load        = new_bd_sparc_Stf_s(dbgi, block, value, new_ptr, mem, load_store_mode, NULL, 0, true);
 408                 sparc_load_store_attr_t *new_load_attr   = get_sparc_load_store_attr(new_load);
 409                 unsigned                 n_outs          = arch_get_irn_n_outs(node);
 410                 unsigned                 i;
 411
 412                 new_load_attr->is_frame_entity = load_store_attr->is_frame_entity;
 413                 new_load_attr->is_reg_reg      = load_store_attr->is_reg_reg;
 414
 415                 sched_add_before(node, new_load);
 416                 for (i = 0; i < n_outs; i++) {
 417                         arch_set_irn_register_out(new_load, i, arch_get_irn_register_out(node, i));
 418                 }
 419                 be_peephole_exchange(node, new_load);
 420         }
 421
 422 }
 423
 424 static void peephole_be_IncSP(ir_node *node)
 425 {
 426         ir_node *pred;
 427         node = be_peephole_IncSP_IncSP(node);
 428         if (!be_is_IncSP(node))
 429                 return;
 430
 431         pred = be_get_IncSP_pred(node);
 432         if (is_sparc_Save(pred) && be_has_only_one_user(pred)) {
 433                 int offset = -be_get_IncSP_offset(node);
 434                 sparc_attr_t *attr = get_sparc_attr(pred);
 435                 attr->immediate_value += offset;
 436                 be_peephole_exchange(node, pred);
 437         }
 438 }
 439
 440 static void peephole_sparc_FrameAddr(ir_node *node)
 441 {
 442         /* the peephole code currently doesn't allow this since it changes
 443          * the register. Find out why and how to workaround this... */
 444 #if 0
 445         const sparc_attr_t *attr = get_sparc_attr_const(node);
 446         if (attr->immediate_value == 0) {
 447                 ir_node *base = get_irn_n(node, n_sparc_FrameAddr_base);
 448                 be_peephole_exchange(node, base);
 449         }
 450 #endif
 451         (void) node;
 452 }
 453
 454 /* output must not be local, or out reg. Since the destination of the restore
 455  * is the rotated register-file where only the old in-registers are still
 456  * visible (as out-registers) */
 457 static bool is_restorezeroopt_reg(const arch_register_t *reg)
 458 {
 459         unsigned index = reg->global_index;
 460         return (index >= REG_G0 && index <= REG_G7)
 461             || (index >= REG_I0 && index <= REG_I7);
 462 }
 463
 464 static void replace_with_restore_reg(ir_node *node, ir_node *replaced,
 465                                                                          ir_node *op0, ir_node *op1)
 466 {
 467         dbg_info *dbgi     = get_irn_dbg_info(node);
 468         ir_node  *stack_in = get_irn_n(node, n_sparc_RestoreZero_stack);
 469         ir_node  *fp       = get_irn_n(node, n_sparc_RestoreZero_frame_pointer);
 470         ir_node  *block    = get_nodes_block(node);
 471         ir_mode  *mode     = get_irn_mode(node);
 472         ir_node  *new_node = new_bd_sparc_Restore_reg(dbgi, block, stack_in, fp,
 473                                                       op0, op1);
 474         ir_node  *stack    = new_r_Proj(new_node, mode, pn_sparc_Restore_stack);
 475         ir_node  *res      = new_r_Proj(new_node, mode, pn_sparc_Restore_res);
 476         const arch_register_t *reg = arch_get_irn_register(replaced);
 477         const arch_register_t *sp  = &sparc_registers[REG_SP];
 478         arch_set_irn_register_out(new_node, pn_sparc_Restore_stack, sp);
 479         arch_set_irn_register_out(new_node, pn_sparc_Restore_res, reg);
 480
 481         sched_add_before(node, new_node);
 482         be_peephole_exchange(node, stack);
 483         be_peephole_exchange(replaced, res);
 484 }
 485
 486 static void replace_with_restore_imm(ir_node *node, ir_node *replaced,
 487                                                                          ir_node *op, ir_entity *imm_entity,
 488                                                                          int32_t immediate)
 489 {
 490         dbg_info *dbgi     = get_irn_dbg_info(node);
 491         ir_node  *stack_in = get_irn_n(node, n_sparc_RestoreZero_stack);
 492         ir_node  *fp       = get_irn_n(node, n_sparc_RestoreZero_frame_pointer);
 493         ir_node  *block    = get_nodes_block(node);
 494         ir_mode  *mode     = get_irn_mode(node);
 495         ir_node  *new_node = new_bd_sparc_Restore_imm(dbgi, block, stack_in, fp,
 496                                                       op, imm_entity, immediate);
 497         ir_node  *stack    = new_r_Proj(new_node, mode, pn_sparc_Restore_stack);
 498         ir_node  *res      = new_r_Proj(new_node, mode, pn_sparc_Restore_res);
 499         const arch_register_t *reg = arch_get_irn_register(replaced);
 500         const arch_register_t *sp  = &sparc_registers[REG_SP];
 501         arch_set_irn_register_out(new_node, pn_sparc_Restore_stack, sp);
 502         arch_set_irn_register_out(new_node, pn_sparc_Restore_res, reg);
 503
 504         sched_add_before(node, new_node);
 505         be_peephole_exchange(node, stack);
 506         be_peephole_exchange(replaced, res);
 507 }
 508
 509 static void peephole_sparc_RestoreZero(ir_node *node)
 510 {
 511         /* restore gives us a free "add" instruction, let's try to use that to fold
 512          * an instruction in. We can do the following:
 513          *
 514          * - Copy values                  (g0 + reg)
 515          * - Produce constants            (g0 + immediate)
 516          * - Perform an add               (reg + reg)
 517          * - Perform a sub with immediate (reg + (-immediate))
 518          *
 519          * Note: In an ideal world, this would not be a peephole optimization but
 520          * already performed during code selection. Since about all foldable ops are
 521          * arguments of the return node. However we have a hard time doing this
 522          * since we construct epilogue code only after register allocation
 523          * (and therefore after code selection).
 524          */
 525         int n_tries = 10; /* limit our search */
 526         ir_node *schedpoint = node;
 527
 528         while (sched_has_prev(schedpoint)) {
 529                 const arch_register_t *reg;
 530                 schedpoint = sched_prev(schedpoint);
 531
 532                 if (--n_tries == 0)
 533                         break;
 534
 535                 if (arch_get_irn_n_outs(schedpoint) == 0)
 536                         continue;
 537
 538                 if (!mode_is_data(get_irn_mode(schedpoint)))
 539                         return;
 540
 541                 reg = arch_get_irn_register(schedpoint);
 542                 if (!is_restorezeroopt_reg(reg))
 543                         continue;
 544
 545                 if (be_is_Copy(schedpoint) && be_can_move_down(heights, schedpoint, node)) {
 546                         ir_node *op = get_irn_n(schedpoint, n_be_Copy_op);
 547                         replace_with_restore_imm(node, schedpoint, op, NULL, 0);
 548                 } else if (is_sparc_Or(schedpoint) &&
 549                            arch_get_irn_flags(schedpoint) & ((arch_irn_flags_t)sparc_arch_irn_flag_immediate_form) &&
 550                            arch_get_irn_register_in(schedpoint, 0) == &sparc_registers[REG_G0] &&
 551                            be_can_move_down(heights, schedpoint, node)) {
 552                         /* it's a constant */
 553                         const sparc_attr_t *attr      = get_sparc_attr_const(schedpoint);
 554                         ir_entity          *entity    = attr->immediate_value_entity;
 555                         int32_t             immediate = attr->immediate_value;
 556                         ir_node            *g0        = get_irn_n(schedpoint, 0);
 557                         replace_with_restore_imm(node, schedpoint, g0, entity, immediate);
 558                 } else if (is_sparc_Add(schedpoint) &&
 559                            be_can_move_down(heights, schedpoint, node)) {
 560                         if (arch_get_irn_flags(schedpoint) & ((arch_irn_flags_t)sparc_arch_irn_flag_immediate_form)) {
 561                                 ir_node            *op     = get_irn_n(schedpoint, 0);
 562                                 const sparc_attr_t *attr   = get_sparc_attr_const(schedpoint);
 563                                 ir_entity          *entity = attr->immediate_value_entity;
 564                                 int32_t             imm    = attr->immediate_value;
 565                                 replace_with_restore_imm(node, schedpoint, op, entity, imm);
 566                         } else {
 567                                 ir_node *op0 = get_irn_n(schedpoint, 0);
 568                                 ir_node *op1 = get_irn_n(schedpoint, 1);
 569                                 replace_with_restore_reg(node, schedpoint, op0, op1);
 570                         }
 571                 } else if (is_sparc_Sub(schedpoint) &&
 572                            arch_get_irn_flags(schedpoint) & ((arch_irn_flags_t)sparc_arch_irn_flag_immediate_form) &&
 573                            arch_get_irn_register_in(schedpoint, 0) == &sparc_registers[REG_G0] &&
 574                            be_can_move_down(heights, schedpoint, node)) {
 575                         /* it's a constant */
 576                         const sparc_attr_t *attr   = get_sparc_attr_const(schedpoint);
 577                         ir_entity          *entity = attr->immediate_value_entity;
 578                         int32_t             imm    = attr->immediate_value;
 579                         if (entity == NULL && sparc_is_value_imm_encodeable(-imm)) {
 580                                 ir_node *g0 = get_irn_n(schedpoint, 0);
 581                                 replace_with_restore_imm(node, schedpoint, g0, NULL, -imm);
 582                         } else {
 583                                 continue;
 584                         }
 585                 }
 586                 /* when we're here then we performed a folding and are done */
 587                 return;
 588         }
 589 }
 590
 591 static void finish_sparc_Return(ir_node *node)
 592 {
 593         ir_node *schedpoint = node;
 594         ir_node *restore;
 595         /* see that there is no code between Return and restore, if there is move
 596          * it in front of the restore */
 597         while (true) {
 598                 if (!sched_has_prev(schedpoint))
 599                         return;
 600                 schedpoint = sched_prev(schedpoint);
 601                 if (is_sparc_Restore(schedpoint) || is_sparc_RestoreZero(schedpoint))
 602                         break;
 603         }
 604         restore = schedpoint;
 605         schedpoint = sched_prev(node);
 606         /* move all code between return and restore up */
 607         while (schedpoint != restore) {
 608                 ir_node *next_schedpoint = sched_prev(schedpoint);
 609                 sched_remove(schedpoint);
 610                 sched_add_before(restore, schedpoint);
 611                 schedpoint = next_schedpoint;
 612         }
 613 }
 614
 615 static void register_peephole_optimisation(ir_op *op, peephole_opt_func func)
 616 {
 617         assert(op->ops.generic == NULL);
 618         op->ops.generic = (op_func) func;
 619 }
 620
 621 static void sparc_collect_frame_entity_nodes(ir_node *node, void *data)
 622 {
 623         be_fec_env_t  *env = (be_fec_env_t*)data;
 624         const ir_mode *mode;
 625         int            align;
 626         ir_entity     *entity;
 627         const sparc_load_store_attr_t *attr;
 628
 629         if (be_is_Reload(node) && be_get_frame_entity(node) == NULL) {
 630                 mode  = get_irn_mode(node);
 631                 align = get_mode_size_bytes(mode);
 632                 be_node_needs_frame_entity(env, node, mode, align);
 633                 return;
 634         }
 635
 636         if (!is_sparc_Ld(node) && !is_sparc_Ldf(node))
 637                 return;
 638
 639         attr   = get_sparc_load_store_attr_const(node);
 640         entity = attr->base.immediate_value_entity;
 641         mode   = attr->load_store_mode;
 642         if (entity != NULL)
 643                 return;
 644         if (!attr->is_frame_entity)
 645                 return;
 646         if (arch_get_irn_flags(node) & sparc_arch_irn_flag_needs_64bit_spillslot)
 647                 mode = mode_Lu;
 648         align  = get_mode_size_bytes(mode);
 649         be_node_needs_frame_entity(env, node, mode, align);
 650 }
 651
 652 static void sparc_set_frame_entity(ir_node *node, ir_entity *entity)
 653 {
 654         if (is_be_node(node)) {
 655                 be_node_set_frame_entity(node, entity);
 656         } else {
 657                 /* we only say be_node_needs_frame_entity on nodes with load_store
 658                  * attributes, so this should be fine */
 659                 sparc_load_store_attr_t *attr = get_sparc_load_store_attr(node);
 660                 assert(attr->is_frame_entity);
 661                 assert(attr->base.immediate_value_entity == NULL);
 662                 attr->base.immediate_value_entity = entity;
 663         }
 664 }
 665
 666 void sparc_finish_graph(ir_graph *irg)
 667 {
 668         be_stack_layout_t *stack_layout = be_get_irg_stack_layout(irg);
 669         bool               at_begin     = stack_layout->sp_relative ? true : false;
 670         be_fec_env_t      *fec_env      = be_new_frame_entity_coalescer(irg);
 671
 672         irg_walk_graph(irg, NULL, sparc_collect_frame_entity_nodes, fec_env);
 673         be_assign_entities(fec_env, sparc_set_frame_entity, at_begin);
 674         be_free_frame_entity_coalescer(fec_env);
 675         sparc_adjust_stack_entity_offsets(irg);
 676
 677         sparc_introduce_prolog_epilog(irg);
 678
 679         /* fix stack entity offsets */
 680         be_abi_fix_stack_nodes(irg);
 681         sparc_fix_stack_bias(irg);
 682
 683         heights = heights_new(irg);
 684
 685         /* perform peephole optimizations */
 686         ir_clear_opcodes_generic_func();
 687         register_peephole_optimisation(op_be_IncSP,        peephole_be_IncSP);
 688         register_peephole_optimisation(op_sparc_FrameAddr, peephole_sparc_FrameAddr);
 689         register_peephole_optimisation(op_sparc_RestoreZero,
 690                                        peephole_sparc_RestoreZero);
 691         register_peephole_optimisation(op_sparc_Ldf, split_sparc_ldf);
 692         be_peephole_opt(irg);
 693
 694         /* perform legalizations (mostly fix nodes with too big immediates) */
 695         ir_clear_opcodes_generic_func();
 696         register_peephole_optimisation(op_be_IncSP,        finish_be_IncSP);
 697         register_peephole_optimisation(op_sparc_FrameAddr, finish_sparc_FrameAddr);
 698         register_peephole_optimisation(op_sparc_Ld,        finish_sparc_Ld);
 699         register_peephole_optimisation(op_sparc_Ldf,       finish_sparc_Ldf);
 700         register_peephole_optimisation(op_sparc_Return,    finish_sparc_Return);
 701         register_peephole_optimisation(op_sparc_Save,      finish_sparc_Save);
 702         register_peephole_optimisation(op_sparc_St,        finish_sparc_St);
 703         register_peephole_optimisation(op_sparc_Stf,       finish_sparc_Stf);
 704         be_peephole_opt(irg);
 705
 706         heights_free(heights);
 707
 708         be_remove_dead_nodes_from_schedule(irg);
 709 }