nsz Git - libfirm/blob - ir/opt/parallelize_mem.c

   1 /*
   2  * Copyright (C) 1995-2011 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   parallelizing Load/Store optimisation
  23  * @author  Christoph Mallon
  24  */
  25 #include "config.h"
  26
  27 #include "iroptimize.h"
  28
  29 #include "array_t.h"
  30 #include "debug.h"
  31 #include "ircons.h"
  32 #include "irgraph.h"
  33 #include "irgmod.h"
  34 #include "irgopt.h"
  35 #include "irgwalk.h"
  36 #include "irmemory.h"
  37 #include "irnode.h"
  38 #include "irnodeset.h"
  39 #include "obst.h"
  40 #include "irdump.h"
  41 #include "irflag_t.h"
  42 #include "irprintf.h"
  43 #include "irpass.h"
  44 #include "iredges.h"
  45
  46 typedef struct parallelize_info
  47 {
  48         ir_node      *origin_block;
  49         ir_node      *origin_ptr;
  50         ir_mode      *origin_mode;
  51         ir_nodeset_t  this_mem;
  52         ir_nodeset_t  user_mem;
  53 } parallelize_info;
  54
  55 static void parallelize_load(parallelize_info *pi, ir_node *irn)
  56 {
  57         /* There is no point in investigating the same subgraph twice */
  58         if (ir_nodeset_contains(&pi->user_mem, irn))
  59                 return;
  60
  61         if (get_nodes_block(irn) == pi->origin_block) {
  62                 if (is_Proj(irn)) {
  63                         ir_node *pred = get_Proj_pred(irn);
  64                         if (is_Load(pred) &&
  65                                         get_Load_volatility(pred) == volatility_non_volatile) {
  66                                 ir_node *mem = get_Load_mem(pred);
  67                                 //ir_nodeset_insert(&pi->this_mem, mem);
  68                                 ir_nodeset_insert(&pi->user_mem, irn);
  69                                 parallelize_load(pi, mem);
  70                                 return;
  71                         } else if (is_Store(pred) &&
  72                                         get_Store_volatility(pred) == volatility_non_volatile) {
  73                                 ir_mode *org_mode   = pi->origin_mode;
  74                                 ir_node *org_ptr    = pi->origin_ptr;
  75                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
  76                                 ir_node *store_ptr  = get_Store_ptr(pred);
  77                                 if (get_alias_relation(org_ptr, org_mode, store_ptr, store_mode) == ir_no_alias) {
  78                                         ir_node *mem = get_Store_mem(pred);
  79                                         ir_nodeset_insert(&pi->user_mem, irn);
  80                                         parallelize_load(pi, mem);
  81                                         return;
  82                                 }
  83                         }
  84                 } else if (is_Sync(irn)) {
  85                         int n = get_Sync_n_preds(irn);
  86                         int i;
  87
  88                         for (i = 0; i < n; ++i) {
  89                                 ir_node *sync_pred = get_Sync_pred(irn, i);
  90                                 parallelize_load(pi, sync_pred);
  91                         }
  92                         return;
  93                 }
  94         }
  95         ir_nodeset_insert(&pi->this_mem, irn);
  96 }
  97
  98 static void parallelize_store(parallelize_info *pi, ir_node *irn)
  99 {
 100         /* There is no point in investigating the same subgraph twice */
 101         if (ir_nodeset_contains(&pi->user_mem, irn))
 102                 return;
 103
 104         //ir_fprintf(stderr, "considering %+F\n", irn);
 105         if (get_nodes_block(irn) == pi->origin_block) {
 106                 if (is_Proj(irn)) {
 107                         ir_node *pred = get_Proj_pred(irn);
 108                         if (is_Load(pred) &&
 109                                         get_Load_volatility(pred) == volatility_non_volatile) {
 110                                 ir_mode *org_mode  = pi->origin_mode;
 111                                 ir_node *org_ptr   = pi->origin_ptr;
 112                                 ir_mode *load_mode = get_Load_mode(pred);
 113                                 ir_node *load_ptr  = get_Load_ptr(pred);
 114                                 if (get_alias_relation(org_ptr, org_mode, load_ptr, load_mode) == ir_no_alias) {
 115                                         ir_node *mem = get_Load_mem(pred);
 116                                         ir_nodeset_insert(&pi->user_mem, irn);
 117                                         parallelize_store(pi, mem);
 118                                         return;
 119                                 }
 120                         } else if (is_Store(pred) &&
 121                                         get_Store_volatility(pred) == volatility_non_volatile) {
 122                                 ir_mode *org_mode   = pi->origin_mode;
 123                                 ir_node *org_ptr    = pi->origin_ptr;
 124                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
 125                                 ir_node *store_ptr  = get_Store_ptr(pred);
 126                                 if (get_alias_relation(org_ptr, org_mode, store_ptr, store_mode) == ir_no_alias) {
 127                                         ir_node *mem;
 128
 129                                         ir_nodeset_insert(&pi->user_mem, irn);
 130                                         mem = get_Store_mem(pred);
 131                                         parallelize_store(pi, mem);
 132                                         return;
 133                                 }
 134                         }
 135                 } else if (is_Sync(irn)) {
 136                         int n = get_Sync_n_preds(irn);
 137                         int i;
 138
 139                         for (i = 0; i < n; ++i) {
 140                                 ir_node *sync_pred = get_Sync_pred(irn, i);
 141                                 parallelize_store(pi, sync_pred);
 142                         }
 143                         return;
 144                 }
 145         }
 146         ir_nodeset_insert(&pi->this_mem, irn);
 147 }
 148
 149 static void walker(ir_node *proj, void *env)
 150 {
 151         ir_node          *mem_op;
 152         ir_node          *pred;
 153         ir_node          *block;
 154         size_t            n;
 155         parallelize_info  pi;
 156
 157         (void)env;
 158
 159         if (!is_Proj(proj)) return;
 160         if (get_irn_mode(proj) != mode_M) return;
 161
 162         mem_op = get_Proj_pred(proj);
 163         if (is_Load(mem_op)) {
 164                 if (get_Load_volatility(mem_op) != volatility_non_volatile) return;
 165
 166                 block = get_nodes_block(mem_op);
 167                 pred  = get_Load_mem(mem_op);
 168
 169                 pi.origin_block = block,
 170                 pi.origin_ptr   = get_Load_ptr(mem_op);
 171                 pi.origin_mode  = get_Load_mode(mem_op);
 172                 ir_nodeset_init(&pi.this_mem);
 173                 ir_nodeset_init(&pi.user_mem);
 174
 175                 parallelize_load(&pi, pred);
 176         } else if (is_Store(mem_op)) {
 177                 if (get_Store_volatility(mem_op) != volatility_non_volatile) return;
 178
 179                 block = get_nodes_block(mem_op);
 180                 pred  = get_Store_mem(mem_op);
 181
 182                 pi.origin_block = block,
 183                 pi.origin_ptr   = get_Store_ptr(mem_op);
 184                 pi.origin_mode  = get_irn_mode(get_Store_value(mem_op));
 185                 ir_nodeset_init(&pi.this_mem);
 186                 ir_nodeset_init(&pi.user_mem);
 187
 188                 parallelize_store(&pi, pred);
 189         } else {
 190                 return;
 191         }
 192
 193         n = ir_nodeset_size(&pi.user_mem);
 194         if (n > 0) { /* nothing happened otherwise */
 195                 ir_node                *sync;
 196                 ir_node               **in   = XMALLOCN(ir_node*, n+1);
 197                 ir_node                *node;
 198                 ir_nodeset_iterator_t   iter;
 199                 size_t                  i;
 200
 201                 i = 0;
 202                 in[i++] = proj;
 203                 foreach_ir_nodeset(&pi.user_mem, node, iter) {
 204                         in[i++] = node;
 205                 }
 206                 assert(i == n+1);
 207                 sync = new_r_Sync(block, i, in);
 208                 xfree(in);
 209                 edges_reroute_except(proj, sync, sync);
 210
 211                 n = ir_nodeset_size(&pi.this_mem);
 212                 if (n == 1) {
 213                         ir_nodeset_iterator_init(&iter, &pi.this_mem);
 214                         sync = ir_nodeset_iterator_next(&iter);
 215                 } else {
 216                         in = XMALLOCN(ir_node*, n);
 217                         i = 0;
 218                         foreach_ir_nodeset(&pi.this_mem, node, iter) {
 219                                 in[i++] = node;
 220                         }
 221                         assert(i == n);
 222                         sync = new_r_Sync(block, i, in);
 223                 }
 224                 set_memop_mem(mem_op, sync);
 225         }
 226
 227         ir_nodeset_destroy(&pi.this_mem);
 228         ir_nodeset_destroy(&pi.user_mem);
 229 }
 230
 231 void opt_parallelize_mem(ir_graph *irg)
 232 {
 233         assure_irg_properties(irg, IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES);
 234         irg_walk_graph(irg, NULL, walker, NULL);
 235         confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_CONTROL_FLOW);
 236 }
 237
 238 ir_graph_pass_t *opt_parallelize_mem_pass(const char *name)
 239 {
 240         return def_graph_pass(name ? name : "parallelize-mem", opt_parallelize_mem);
 241 }