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

   1 /*
   2  * Copyright (C) 1995-2007 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
  61 static ir_node *default_remat(ir_node *node, ir_node *after)
  62 {
  63         ir_node *block, *copy;
  64         if(is_Block(after))
  65                 block = after;
  66         else
  67                 block = get_nodes_block(after);
  68
  69         copy = exact_copy(node);
  70         set_nodes_block(copy, block);
  71         sched_add_after(after, copy);
  72
  73         return copy;
  74 }
  75
  76 /**
  77  * tests wether we can legally move node node after node after
  78  * (only works for nodes in same block)
  79  */
  80 static int can_move(ir_node *node, ir_node *after)
  81 {
  82         const ir_edge_t *edge;
  83         assert(get_nodes_block(node) == get_nodes_block(after));
  84
  85         /* TODO respect dep edges */
  86         assert(get_irn_n_edges_kind(node, EDGE_KIND_DEP) == 0);
  87
  88         /** all users have to be after the after node */
  89         foreach_out_edge(node, edge) {
  90                 ir_node *out = get_edge_src_irn(edge);
  91                 if(is_Proj(out)) {
  92                         const ir_edge_t *edge2;
  93                         assert(get_irn_n_edges_kind(out, EDGE_KIND_DEP) == 0);
  94                         foreach_out_edge(out, edge2) {
  95                                 ir_node *out2 = get_edge_src_irn(edge2);
  96                                 /* phi represents a usage at block end */
  97                                 if(is_Phi(out2))
  98                                         continue;
  99                                 if(is_Sync(out2)) {
 100                                         const ir_edge_t *edge3;
 101                                         foreach_out_edge(out2, edge3) {
 102                                                 ir_node *out3 = get_edge_src_irn(edge3);
 103                                                 /* phi represents a usage at block end */
 104                                                 if(is_Phi(out3))
 105                                                         continue;
 106                                                 assert(!is_Sync(out3));
 107                                                 if(sched_get_time_step(out3) <= sched_get_time_step(after)) {
 108                                                         return 0;
 109                                                 }
 110                                         }
 111                                 } else if(sched_get_time_step(out2) <= sched_get_time_step(after)) {
 112                                         return 0;
 113                                 }
 114                         }
 115                 } else {
 116                         /* phi represents a usage at block end */
 117                         if(is_Phi(out))
 118                                 continue;
 119                         if(sched_get_time_step(out) <= sched_get_time_step(after)) {
 120                                 return 0;
 121                         }
 122                 }
 123         }
 124
 125         return 1;
 126 }
 127
 128 static void rematerialize_or_move(ir_node *flags_needed, ir_node *node,
 129                                   ir_node *flag_consumers, int pn)
 130 {
 131         ir_node *n;
 132         ir_node *copy;
 133         ir_node *value;
 134
 135         if(!is_Block(node) &&
 136                         get_nodes_block(flags_needed) == get_nodes_block(node) &&
 137                         can_move(flags_needed, node)) {
 138                 /* move it */
 139                 sched_remove(flags_needed);
 140                 sched_add_after(node, flags_needed);
 141                 return;
 142         }
 143
 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
 171 static int is_modify_flags(ir_node *node) {
 172         int i, arity;
 173
 174         if(arch_irn_is(arch_env, node, modify_flags))
 175                 return 1;
 176         if(!be_is_Keep(node))
 177                 return 0;
 178
 179         arity = get_irn_arity(node);
 180         for(i = 0; i < arity; ++i) {
 181                 ir_node *in = get_irn_n(node, i);
 182                 in = skip_Proj(in);
 183                 if(arch_irn_is(arch_env, in, modify_flags))
 184                         return 1;
 185         }
 186
 187         return 0;
 188 }
 189
 190 /**
 191  * walks up the schedule and makes sure there are no flag-destroying nodes
 192  * between a flag-consumer -> flag-producer chain. Fixes problematic situations
 193  * by moving and/or rematerialisation of the flag-producers.
 194  * (This can be extended in the future to do some register allocation on targets
 195  *  like ppc32 where we conceptually have 8 flag registers)
 196  */
 197 static void fix_flags_walker(ir_node *block, void *env)
 198 {
 199         ir_node *node;
 200         ir_node *flags_needed   = NULL;
 201         ir_node *flag_consumers = NULL;
 202         int      pn;
 203         (void) env;
 204
 205         sched_foreach_reverse(block, node) {
 206                 int i, arity;
 207                 ir_node *new_flags_needed = NULL;
 208
 209                 if(is_Phi(node))
 210                         break;
 211
 212                 if(node == flags_needed) {
 213                         /* all ok */
 214                         flags_needed   = NULL;
 215                         flag_consumers = NULL;
 216                 }
 217
 218                 /* test wether node destroys the flags */
 219                 if(flags_needed != NULL && is_modify_flags(node)) {
 220                         /* rematerialize */
 221                         rematerialize_or_move(flags_needed, node, flag_consumers, pn);
 222                         flags_needed   = NULL;
 223                         flag_consumers = NULL;
 224                 }
 225
 226                 /* test wether the current node needs flags */
 227                 arity = get_irn_arity(node);
 228                 for(i = 0; i < arity; ++i) {
 229                         //ir_node *in = get_irn_n(node, i);
 230                         const arch_register_class_t *cls
 231                                 = arch_get_irn_reg_class(arch_env, node, i);
 232                         if(cls == flag_class) {
 233                                 assert(new_flags_needed == NULL);
 234                                 new_flags_needed = get_irn_n(node, i);
 235                         }
 236                 }
 237
 238                 if(new_flags_needed == NULL)
 239                         continue;
 240
 241                 /* spiller can't (correctly) remat flag consumers at the moment */
 242                 assert(!arch_irn_is(arch_env, node, rematerializable));
 243                 if(new_flags_needed != flags_needed) {
 244                         if(flags_needed != NULL) {
 245                                 /* rematerialize node */
 246                                 rematerialize_or_move(flags_needed, node, flag_consumers, pn);
 247                                 flags_needed   = NULL;
 248                                 flag_consumers = NULL;
 249                         }
 250
 251                         flags_needed = new_flags_needed;
 252                                         arch_set_irn_register(arch_env, flags_needed, flags_reg);
 253                         if(is_Proj(flags_needed)) {
 254                                 pn           = get_Proj_proj(flags_needed);
 255                                 flags_needed = get_Proj_pred(flags_needed);
 256                         }
 257                         flag_consumers = node;
 258                         set_irn_link(flag_consumers, NULL);
 259                         assert(arch_irn_is(arch_env, flags_needed, rematerializable));
 260                 } else {
 261                         /* link all consumers in a list */
 262                         set_irn_link(node, flag_consumers);
 263                         flag_consumers = node;
 264                 }
 265         }
 266
 267         if(flags_needed != NULL) {
 268                 assert(get_nodes_block(flags_needed) != block);
 269                 rematerialize_or_move(flags_needed, node, flag_consumers, pn);
 270                 flags_needed   = NULL;
 271                 flag_consumers = NULL;
 272         }
 273
 274         assert(flags_needed   == NULL);
 275         assert(flag_consumers == NULL);
 276 }
 277
 278 void be_sched_fix_flags(be_irg_t *birg, const arch_register_class_t *flag_cls,
 279                         func_rematerialize remat_func)
 280 {
 281         ir_graph *irg = be_get_birg_irg(birg);
 282
 283         arch_env   = be_get_birg_arch_env(birg);
 284         flag_class = flag_cls;
 285         flags_reg  = & flag_class->regs[0];
 286         remat      = remat_func;
 287         if(remat == NULL)
 288                 remat = &default_remat;
 289
 290         set_using_irn_link(irg);
 291         irg_block_walk_graph(irg, fix_flags_walker, NULL, NULL);
 292         clear_using_irn_link(irg);
 293 }