nsz Git - libfirm/blob - ir/be/beflags.c

   1 /*
   2  * Copyright (C) 1995-2008 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       modifies schedule so flags dependencies are respected.
  23  * @author      Matthias Braun, Christoph Mallon
  24  * @version     $Id: besched.h 14693 2007-06-21 15:35:49Z beck $
  25  *
  26  * Fixup schedule to respect flag constraints by moving and rematerialisation of
  27  * nodes.
  28  *
  29  * Flags are modeled as register classes with ignore registers. However to avoid
  30  * bloating the graph, only flag-consumer -> producer dependencies are
  31  * explicitely modeled in the graph. Nodes that just change the flags are only
  32  * marked with the arch_irn_flags_modify_flags flag.
  33  *
  34  * Flags are usually a limited resource that can't (or at least shouldn't) be
  35  * spilled. So in some situations (for example 2 adc-nodes that use the flags of
  36  * a single add node on x86) operations have to be repeated to work correctly.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include "config.h"
  40 #endif
  41
  42 #include "irgwalk.h"
  43 #include "irnode_t.h"
  44 #include "irtools.h"
  45 #include "ircons.h"
  46 #include "iredges_t.h"
  47 #include "irprintf.h"
  48 #include "error.h"
  49
  50 #include "beflags.h"
  51 #include "bearch_t.h"
  52 #include "beirg_t.h"
  53 #include "besched_t.h"
  54 #include "benode_t.h"
  55
  56 static const arch_env_t            *arch_env   = NULL;
  57 static const arch_register_class_t *flag_class = NULL;
  58 static const arch_register_t       *flags_reg  = NULL;
  59 static func_rematerialize           remat      = NULL;
  60 static int                          changed;
  61
  62 static ir_node *default_remat(ir_node *node, ir_node *after)
  63 {
  64         ir_node *block, *copy;
  65         if(is_Block(after))
  66                 block = after;
  67         else
  68                 block = get_nodes_block(after);
  69
  70         copy = exact_copy(node);
  71         set_nodes_block(copy, block);
  72         sched_add_after(after, copy);
  73
  74         return copy;
  75 }
  76
  77 /**
  78  * tests wether we can legally move node node after node after
  79  * (only works for nodes in same block)
  80  */
  81 static int can_move(ir_node *node, ir_node *after)
  82 {
  83         const ir_edge_t *edge;
  84         assert(get_nodes_block(node) == get_nodes_block(after));
  85
  86         /* TODO respect dep edges */
  87         assert(get_irn_n_edges_kind(node, EDGE_KIND_DEP) == 0);
  88
  89         /** all users have to be after the after node */
  90         foreach_out_edge(node, edge) {
  91                 ir_node *out = get_edge_src_irn(edge);
  92                 if(is_Proj(out)) {
  93                         const ir_edge_t *edge2;
  94                         assert(get_irn_n_edges_kind(out, EDGE_KIND_DEP) == 0);
  95                         foreach_out_edge(out, edge2) {
  96                                 ir_node *out2 = get_edge_src_irn(edge2);
  97                                 /* phi represents a usage at block end */
  98                                 if(is_Phi(out2))
  99                                         continue;
 100                                 if(is_Sync(out2)) {
 101                                         const ir_edge_t *edge3;
 102                                         foreach_out_edge(out2, edge3) {
 103                                                 ir_node *out3 = get_edge_src_irn(edge3);
 104                                                 /* phi represents a usage at block end */
 105                                                 if(is_Phi(out3))
 106                                                         continue;
 107                                                 assert(!is_Sync(out3));
 108                                                 if(sched_get_time_step(out3) <= sched_get_time_step(after)) {
 109                                                         return 0;
 110                                                 }
 111                                         }
 112                                 } else if(sched_get_time_step(out2) <= sched_get_time_step(after)) {
 113                                         return 0;
 114                                 }
 115                         }
 116                 } else {
 117                         /* phi represents a usage at block end */
 118                         if(is_Phi(out))
 119                                 continue;
 120                         if(sched_get_time_step(out) <= sched_get_time_step(after)) {
 121                                 return 0;
 122                         }
 123                 }
 124         }
 125
 126         return 1;
 127 }
 128
 129 static void rematerialize_or_move(ir_node *flags_needed, ir_node *node,
 130                                   ir_node *flag_consumers, int pn, be_lv_t *lv)
 131 {
 132         ir_node *n;
 133         ir_node *copy;
 134         ir_node *value;
 135
 136         if(!is_Block(node) &&
 137                         get_nodes_block(flags_needed) == get_nodes_block(node) &&
 138                         can_move(flags_needed, node)) {
 139                 /* move it */
 140                 sched_remove(flags_needed);
 141                 sched_add_after(node, flags_needed);
 142                 /* No need to update liveness, because the node stays in the same block */
 143                 return;
 144         }
 145
 146         changed = 1;
 147         copy    = remat(flags_needed, node);
 148
 149         if(get_irn_mode(copy) == mode_T) {
 150                 ir_node *block = get_nodes_block(copy);
 151                 ir_mode *mode  = flag_class->mode;
 152                 value = new_rd_Proj(NULL, current_ir_graph, block,
 153                                                         copy, mode, pn);
 154         } else {
 155                 value = copy;
 156         }
 157
 158         n = flag_consumers;
 159         do {
 160                 int i;
 161                 int arity = get_irn_arity(n);
 162                 for(i = 0; i < arity; ++i) {
 163                         ir_node *in = get_irn_n(n, i);
 164                         in = skip_Proj(in);
 165                         if(in == flags_needed) {
 166                                 set_irn_n(n, i, value);
 167                                 break;
 168                         }
 169                 }
 170                 n = get_irn_link(n);
 171         } while(n != NULL);
 172
 173         /* No need to introduce the copy, because it only lives in this block, but
 174          * we have to update the liveness of all operands */
 175         if (is_Block(node) ||
 176                         get_nodes_block(node) != get_nodes_block(flags_needed)) {
 177                 int i;
 178
 179                 for (i = get_irn_arity(copy) - 1; i >= 0; --i) {
 180                         be_liveness_update(lv, get_irn_n(copy, i));
 181                 }
 182         }
 183 }
 184
 185 static int is_modify_flags(ir_node *node) {
 186         int i, arity;
 187
 188         if(arch_irn_is(arch_env, node, modify_flags))
 189                 return 1;
 190         if(!be_is_Keep(node))
 191                 return 0;
 192
 193         arity = get_irn_arity(node);
 194         for(i = 0; i < arity; ++i) {
 195                 ir_node *in = get_irn_n(node, i);
 196                 in = skip_Proj(in);
 197                 if(arch_irn_is(arch_env, in, modify_flags))
 198                         return 1;
 199         }
 200
 201         return 0;
 202 }
 203
 204 /**
 205  * walks up the schedule and makes sure there are no flag-destroying nodes
 206  * between a flag-consumer -> flag-producer chain. Fixes problematic situations
 207  * by moving and/or rematerialisation of the flag-producers.
 208  * (This can be extended in the future to do some register allocation on targets
 209  *  like ppc32 where we conceptually have 8 flag registers)
 210  */
 211 static void fix_flags_walker(ir_node *block, void *env)
 212 {
 213         ir_node *node;
 214         ir_node *flags_needed   = NULL;
 215         ir_node *flag_consumers = NULL;
 216         int      pn = -1;
 217
 218         sched_foreach_reverse(block, node) {
 219                 int i, arity;
 220                 ir_node *new_flags_needed = NULL;
 221
 222                 if(is_Phi(node))
 223                         break;
 224
 225                 if(node == flags_needed) {
 226                         /* all ok */
 227                         flags_needed   = NULL;
 228                         flag_consumers = NULL;
 229                 }
 230
 231                 /* test whether node destroys the flags */
 232                 if(flags_needed != NULL && is_modify_flags(node)) {
 233                         /* rematerialize */
 234                         rematerialize_or_move(flags_needed, node, flag_consumers, pn, env);
 235                         flags_needed   = NULL;
 236                         flag_consumers = NULL;
 237                 }
 238
 239                 /* test whether the current node needs flags */
 240                 arity = get_irn_arity(node);
 241                 for(i = 0; i < arity; ++i) {
 242                         const arch_register_class_t *cls
 243                                 = arch_get_irn_reg_class(arch_env, node, i);
 244                         if(cls == flag_class) {
 245                                 assert(new_flags_needed == NULL);
 246                                 new_flags_needed = get_irn_n(node, i);
 247                         }
 248                 }
 249
 250                 if(new_flags_needed == NULL)
 251                         continue;
 252
 253                 /* spiller can't (correctly) remat flag consumers at the moment */
 254                 assert(!arch_irn_is(arch_env, node, rematerializable));
 255
 256                 if(skip_Proj(new_flags_needed) != flags_needed) {
 257                         if(flags_needed != NULL) {
 258                                 /* rematerialize node */
 259                                 rematerialize_or_move(flags_needed, node, flag_consumers, pn, env);
 260                                 flags_needed   = NULL;
 261                                 flag_consumers = NULL;
 262                         }
 263
 264                         flags_needed = new_flags_needed;
 265                         arch_set_irn_register(arch_env, flags_needed, flags_reg);
 266                         if(is_Proj(flags_needed)) {
 267                                 pn           = get_Proj_proj(flags_needed);
 268                                 flags_needed = get_Proj_pred(flags_needed);
 269                         }
 270                         flag_consumers = node;
 271                         set_irn_link(flag_consumers, NULL);
 272                         assert(arch_irn_is(arch_env, flags_needed, rematerializable));
 273                 } else {
 274                         /* link all consumers in a list */
 275                         set_irn_link(node, flag_consumers);
 276                         flag_consumers = node;
 277                 }
 278         }
 279
 280         if(flags_needed != NULL) {
 281                 assert(get_nodes_block(flags_needed) != block);
 282                 rematerialize_or_move(flags_needed, node, flag_consumers, pn, env);
 283                 flags_needed   = NULL;
 284                 flag_consumers = NULL;
 285         }
 286
 287         assert(flags_needed   == NULL);
 288         assert(flag_consumers == NULL);
 289 }
 290
 291 void be_sched_fix_flags(be_irg_t *birg, const arch_register_class_t *flag_cls,
 292                         func_rematerialize remat_func)
 293 {
 294         ir_graph *irg = be_get_birg_irg(birg);
 295
 296         arch_env   = be_get_birg_arch_env(birg);
 297         flag_class = flag_cls;
 298         flags_reg  = & flag_class->regs[0];
 299         remat      = remat_func;
 300         changed    = 0;
 301         if(remat == NULL)
 302                 remat = &default_remat;
 303
 304         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
 305         irg_block_walk_graph(irg, fix_flags_walker, NULL, birg->lv);
 306         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
 307
 308         if(changed) {
 309                 be_remove_dead_nodes_from_schedule(birg);
 310         }
 311 }