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