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
  45 typedef struct parallelize_info
  46 {
  47         ir_node      *origin_block;
  48         ir_node      *origin_ptr;
  49         ir_mode      *origin_mode;
  50         ir_nodeset_t  this_mem;
  51         ir_nodeset_t  user_mem;
  52 } parallelize_info;
  53
  54 static void parallelize_load(parallelize_info *pi, ir_node *irn)
  55 {
  56         /* There is no point in investigating the same subgraph twice */
  57         if (ir_nodeset_contains(&pi->user_mem, irn))
  58                 return;
  59
  60         if (get_nodes_block(irn) == pi->origin_block) {
  61                 if (is_Proj(irn)) {
  62                         ir_node *pred = get_Proj_pred(irn);
  63                         if (is_Load(pred) &&
  64                                         get_Load_volatility(pred) == volatility_non_volatile) {
  65                                 ir_node *mem = get_Load_mem(pred);
  66                                 //ir_nodeset_insert(&pi->this_mem, mem);
  67                                 ir_nodeset_insert(&pi->user_mem, irn);
  68                                 parallelize_load(pi, mem);
  69                                 return;
  70                         } else if (is_Store(pred) &&
  71                                         get_Store_volatility(pred) == volatility_non_volatile) {
  72                                 ir_mode *org_mode   = pi->origin_mode;
  73                                 ir_node *org_ptr    = pi->origin_ptr;
  74                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
  75                                 ir_node *store_ptr  = get_Store_ptr(pred);
  76                                 if (get_alias_relation(org_ptr, org_mode, store_ptr, store_mode) == ir_no_alias) {
  77                                         ir_node *mem = get_Store_mem(pred);
  78                                         ir_nodeset_insert(&pi->user_mem, irn);
  79                                         parallelize_load(pi, mem);
  80                                         return;
  81                                 }
  82                         }
  83                 } else if (is_Sync(irn)) {
  84                         int n = get_Sync_n_preds(irn);
  85                         int i;
  86
  87                         for (i = 0; i < n; ++i) {
  88                                 ir_node *sync_pred = get_Sync_pred(irn, i);
  89                                 parallelize_load(pi, sync_pred);
  90                         }
  91                         return;
  92                 }
  93         }
  94         ir_nodeset_insert(&pi->this_mem, irn);
  95 }
  96
  97 static void parallelize_store(parallelize_info *pi, ir_node *irn)
  98 {
  99         /* There is no point in investigating the same subgraph twice */
 100         if (ir_nodeset_contains(&pi->user_mem, irn))
 101                 return;
 102
 103         //ir_fprintf(stderr, "considering %+F\n", irn);
 104         if (get_nodes_block(irn) == pi->origin_block) {
 105                 if (is_Proj(irn)) {
 106                         ir_node *pred = get_Proj_pred(irn);
 107                         if (is_Load(pred) &&
 108                                         get_Load_volatility(pred) == volatility_non_volatile) {
 109                                 ir_mode *org_mode  = pi->origin_mode;
 110                                 ir_node *org_ptr   = pi->origin_ptr;
 111                                 ir_mode *load_mode = get_Load_mode(pred);
 112                                 ir_node *load_ptr  = get_Load_ptr(pred);
 113                                 if (get_alias_relation(org_ptr, org_mode, load_ptr, load_mode) == ir_no_alias) {
 114                                         ir_node *mem = get_Load_mem(pred);
 115                                         ir_nodeset_insert(&pi->user_mem, irn);
 116                                         parallelize_store(pi, mem);
 117                                         return;
 118                                 }
 119                         } else if (is_Store(pred) &&
 120                                         get_Store_volatility(pred) == volatility_non_volatile) {
 121                                 ir_mode *org_mode   = pi->origin_mode;
 122                                 ir_node *org_ptr    = pi->origin_ptr;
 123                                 ir_mode *store_mode = get_irn_mode(get_Store_value(pred));
 124                                 ir_node *store_ptr  = get_Store_ptr(pred);
 125                                 if (get_alias_relation(org_ptr, org_mode, store_ptr, store_mode) == ir_no_alias) {
 126                                         ir_node *mem;
 127
 128                                         ir_nodeset_insert(&pi->user_mem, irn);
 129                                         mem = get_Store_mem(pred);
 130                                         parallelize_store(pi, mem);
 131                                         return;
 132                                 }
 133                         }
 134                 } else if (is_Sync(irn)) {
 135                         int n = get_Sync_n_preds(irn);
 136                         int i;
 137
 138                         for (i = 0; i < n; ++i) {
 139                                 ir_node *sync_pred = get_Sync_pred(irn, i);
 140                                 parallelize_store(pi, sync_pred);
 141                         }
 142                         return;
 143                 }
 144         }
 145         ir_nodeset_insert(&pi->this_mem, irn);
 146 }
 147
 148 static void walker(ir_node *proj, void *env)
 149 {
 150         ir_node          *mem_op;
 151         ir_node          *pred;
 152         ir_node          *block;
 153         size_t            n;
 154         parallelize_info  pi;
 155
 156         (void)env;
 157
 158         if (!is_Proj(proj)) return;
 159         if (get_irn_mode(proj) != mode_M) return;
 160
 161         mem_op = get_Proj_pred(proj);
 162         if (is_Load(mem_op)) {
 163                 if (get_Load_volatility(mem_op) != volatility_non_volatile) return;
 164
 165                 block = get_nodes_block(mem_op);
 166                 pred  = get_Load_mem(mem_op);
 167
 168                 pi.origin_block = block,
 169                 pi.origin_ptr   = get_Load_ptr(mem_op);
 170                 pi.origin_mode  = get_Load_mode(mem_op);
 171                 ir_nodeset_init(&pi.this_mem);
 172                 ir_nodeset_init(&pi.user_mem);
 173
 174                 parallelize_load(&pi, pred);
 175         } else if (is_Store(mem_op)) {
 176                 if (get_Store_volatility(mem_op) != volatility_non_volatile) return;
 177
 178                 block = get_nodes_block(mem_op);
 179                 pred  = get_Store_mem(mem_op);
 180
 181                 pi.origin_block = block,
 182                 pi.origin_ptr   = get_Store_ptr(mem_op);
 183                 pi.origin_mode  = get_irn_mode(get_Store_value(mem_op));
 184                 ir_nodeset_init(&pi.this_mem);
 185                 ir_nodeset_init(&pi.user_mem);
 186
 187                 parallelize_store(&pi, pred);
 188         } else {
 189                 return;
 190         }
 191
 192         n = ir_nodeset_size(&pi.user_mem);
 193         if (n != 0) { /* nothing happened otherwise */
 194                 ir_graph               *irg  = get_irn_irg(block);
 195                 ir_node                *sync;
 196                 ir_node               **in;
 197                 ir_nodeset_iterator_t   iter;
 198                 size_t                  i;
 199
 200                 ++n;
 201                 NEW_ARR_A(ir_node*, in, n);
 202                 i = 0;
 203                 in[i++] = new_r_Unknown(irg, mode_M);
 204                 ir_nodeset_iterator_init(&iter, &pi.user_mem);
 205                 for (;;) {
 206                         ir_node* p = ir_nodeset_iterator_next(&iter);
 207                         if (p == NULL) break;
 208                         in[i++] = p;
 209                 }
 210                 assert(i == n);
 211                 sync = new_r_Sync(block, n, in);
 212                 exchange(proj, sync);
 213
 214                 assert((long)pn_Load_M == (long)pn_Store_M);
 215                 proj = new_r_Proj(mem_op, mode_M, pn_Load_M);
 216                 set_Sync_pred(sync, 0, proj);
 217
 218                 n = ir_nodeset_size(&pi.this_mem);
 219                 ir_nodeset_iterator_init(&iter, &pi.this_mem);
 220                 if (n == 1) {
 221                         sync = ir_nodeset_iterator_next(&iter);
 222                 } else {
 223                         NEW_ARR_A(ir_node*, in, n);
 224                         i = 0;
 225                         for (;;) {
 226                                 ir_node* p = ir_nodeset_iterator_next(&iter);
 227                                 if (p == NULL) break;
 228                                 in[i++] = p;
 229                         }
 230                         assert(i == n);
 231                         sync = new_r_Sync(block, n, in);
 232                 }
 233                 set_memop_mem(mem_op, sync);
 234         }
 235
 236         ir_nodeset_destroy(&pi.this_mem);
 237         ir_nodeset_destroy(&pi.user_mem);
 238 }
 239
 240 void opt_parallelize_mem(ir_graph *irg)
 241 {
 242         irg_walk_graph(irg, NULL, walker, NULL);
 243         confirm_irg_properties(irg, IR_GRAPH_PROPERTIES_CONTROL_FLOW);
 244 }
 245
 246 ir_graph_pass_t *opt_parallelize_mem_pass(const char *name)
 247 {
 248         return def_graph_pass(name ? name : "parallelize-mem", opt_parallelize_mem);
 249 }