cse.c

   1 /*
   2  * CSE - walk the linearized instruction flow, and
   3  * see if we can simplify it and apply CSE on it.
   4  *
   5  * Copyright (C) 2004 Linus Torvalds
   6  */
   7
   8 #include <string.h>
   9 #include <stdarg.h>
  10 #include <stdlib.h>
  11 #include <stdio.h>
  12 #include <stddef.h>
  13 #include <assert.h>
  14
  15 #include "parse.h"
  16 #include "expression.h"
  17 #include "linearize.h"
  18 #include "flow.h"
  19
  20 #define INSN_HASH_SIZE 65536
  21 static struct instruction_list *insn_hash_table[INSN_HASH_SIZE];
  22
  23 int repeat_phase;
  24
  25 static int phi_compare(pseudo_t phi1, pseudo_t phi2)
  26 {
  27         const struct instruction *def1 = phi1->def;
  28         const struct instruction *def2 = phi2->def;
  29
  30         if (def1->src1 != def2->src1)
  31                 return def1->src1 < def2->src1 ? -1 : 1;
  32         if (def1->bb != def2->bb)
  33                 return def1->bb < def2->bb ? -1 : 1;
  34         return 0;
  35 }
  36
  37
  38 static void clean_up_one_instruction(struct basic_block *bb, struct instruction *insn)
  39 {
  40         unsigned long hash;
  41
  42         if (!insn->bb)
  43                 return;
  44         assert(insn->bb == bb);
  45         repeat_phase |= simplify_instruction(insn);
  46         hash = (insn->opcode << 3) + (insn->size >> 3);
  47         switch (insn->opcode) {
  48         case OP_SEL:
  49                 hash += hashval(insn->src3);
  50                 /* Fallthrough */
  51
  52         /* Binary arithmetic */
  53         case OP_ADD: case OP_SUB:
  54         case OP_MUL: case OP_DIV:
  55         case OP_MOD: case OP_SHL:
  56         case OP_SHR:
  57         case OP_AND: case OP_OR:
  58
  59         /* Binary logical */
  60         case OP_XOR: case OP_AND_BOOL:
  61         case OP_OR_BOOL:
  62
  63         /* Binary comparison */
  64         case OP_SET_EQ: case OP_SET_NE:
  65         case OP_SET_LE: case OP_SET_GE:
  66         case OP_SET_LT: case OP_SET_GT:
  67         case OP_SET_B:  case OP_SET_A:
  68         case OP_SET_BE: case OP_SET_AE:
  69                 hash += hashval(insn->src2);
  70                 /* Fallthrough */
  71
  72         /* Unary */
  73         case OP_NOT: case OP_NEG:
  74                 hash += hashval(insn->src1);
  75                 break;
  76
  77         case OP_SETVAL:
  78                 hash += hashval(insn->val);
  79                 hash += hashval(insn->symbol);
  80                 break;
  81
  82         /* Other */
  83         case OP_PHI: {
  84                 pseudo_t phi;
  85                 FOR_EACH_PTR(insn->phi_list, phi) {
  86                         struct instruction *def;
  87                         if (phi == VOID || !phi->def)
  88                                 continue;
  89                         def = phi->def;
  90                         hash += hashval(def->src1);
  91                         hash += hashval(def->bb);
  92                 } END_FOR_EACH_PTR(phi);
  93                 break;
  94         }
  95
  96         default:
  97                 /*
  98                  * Nothing to do, don't even bother hashing them,
  99                  * we're not going to try to CSE them
 100                  */
 101                 return;
 102         }
 103         hash += hash >> 16;
 104         hash &= INSN_HASH_SIZE-1;
 105         add_instruction(insn_hash_table + hash, insn);
 106 }
 107
 108 static void clean_up_insns(struct entrypoint *ep)
 109 {
 110         struct basic_block *bb;
 111
 112         FOR_EACH_PTR(ep->bbs, bb) {
 113                 struct instruction *insn;
 114                 FOR_EACH_PTR(bb->insns, insn) {
 115                         clean_up_one_instruction(bb, insn);
 116                 } END_FOR_EACH_PTR(insn);
 117         } END_FOR_EACH_PTR(bb);
 118 }
 119
 120 /* Compare two (sorted) phi-lists */
 121 static int phi_list_compare(struct pseudo_list *l1, struct pseudo_list *l2)
 122 {
 123         pseudo_t phi1, phi2;
 124
 125         PREPARE_PTR_LIST(l1, phi1);
 126         PREPARE_PTR_LIST(l2, phi2);
 127         for (;;) {
 128                 int cmp;
 129
 130                 while (phi1 && (phi1 == VOID || !phi1->def))
 131                         NEXT_PTR_LIST(phi1);
 132                 while (phi2 && (phi2 == VOID || !phi2->def))
 133                         NEXT_PTR_LIST(phi2);
 134
 135                 if (!phi1)
 136                         return phi2 ? -1 : 0;
 137                 if (!phi2)
 138                         return phi1 ? 1 : 0;
 139                 cmp = phi_compare(phi1, phi2);
 140                 if (cmp)
 141                         return cmp;
 142                 NEXT_PTR_LIST(phi1);
 143                 NEXT_PTR_LIST(phi2);
 144         }
 145         /* Not reached, but we need to make the nesting come out right */
 146         FINISH_PTR_LIST(phi2);
 147         FINISH_PTR_LIST(phi1);
 148 }
 149
 150 static int insn_compare(const void *_i1, const void *_i2)
 151 {
 152         const struct instruction *i1 = _i1;
 153         const struct instruction *i2 = _i2;
 154
 155         if (i1->opcode != i2->opcode)
 156                 return i1->opcode < i2->opcode ? -1 : 1;
 157
 158         switch (i1->opcode) {
 159         case OP_SEL:
 160                 if (i1->src3 != i2->src3)
 161                         return i1->src3 < i2->src3 ? -1 : 1;
 162                 /* Fall-through to binops */
 163
 164         /* Binary arithmetic */
 165         case OP_ADD: case OP_SUB:
 166         case OP_MUL: case OP_DIV:
 167         case OP_MOD: case OP_SHL:
 168         case OP_SHR:
 169         case OP_AND: case OP_OR:
 170
 171         /* Binary logical */
 172         case OP_XOR: case OP_AND_BOOL:
 173         case OP_OR_BOOL:
 174
 175         /* Binary comparison */
 176         case OP_SET_EQ: case OP_SET_NE:
 177         case OP_SET_LE: case OP_SET_GE:
 178         case OP_SET_LT: case OP_SET_GT:
 179         case OP_SET_B:  case OP_SET_A:
 180         case OP_SET_BE: case OP_SET_AE:
 181                 if (i1->src2 != i2->src2)
 182                         return i1->src2 < i2->src2 ? -1 : 1;
 183                 /* Fall-through to unops */
 184
 185         /* Unary */
 186         case OP_NOT: case OP_NEG:
 187                 if (i1->src1 != i2->src1)
 188                         return i1->src1 < i2->src1 ? -1 : 1;
 189                 break;
 190
 191         case OP_SETVAL:
 192                 if (i1->val != i2->val)
 193                         return i1->val < i2->val ? -1 : 1;
 194                 if (i1->symbol != i2->symbol)
 195                         return i1->symbol < i2->symbol ? -1 : 1;
 196                 break;
 197
 198         /* Other */
 199         case OP_PHI:
 200                 return phi_list_compare(i1->phi_list, i2->phi_list);
 201
 202         default:
 203                 warning(i1->bb->pos, "bad instruction on hash chain");
 204         }
 205         if (i1->size != i2->size)
 206                 return i1->size < i2->size ? -1 : 1;
 207         return 0;
 208 }
 209
 210 static void sort_instruction_list(struct instruction_list **list)
 211 {
 212         sort_list((struct ptr_list **)list , insn_compare);
 213 }
 214
 215 static struct instruction * cse_one_instruction(struct instruction *insn, struct instruction *def)
 216 {
 217         convert_instruction_target(insn, def->target);
 218         insn->opcode = OP_NOP;
 219         insn->bb = NULL;
 220         repeat_phase |= REPEAT_CSE;
 221         return def;
 222 }
 223
 224 /*
 225  * Does "bb1" dominate "bb2"?
 226  */
 227 static int bb_dominates(struct entrypoint *ep, struct basic_block *bb1, struct basic_block *bb2, unsigned long generation)
 228 {
 229         struct basic_block *parent;
 230
 231         /* Nothing dominates the entrypoint.. */
 232         if (bb2 == ep->entry->bb)
 233                 return 0;
 234         FOR_EACH_PTR(bb2->parents, parent) {
 235                 if (parent == bb1)
 236                         continue;
 237                 if (parent->generation == generation)
 238                         continue;
 239                 parent->generation = generation;
 240                 if (!bb_dominates(ep, bb1, parent, generation))
 241                         return 0;
 242         } END_FOR_EACH_PTR(parent);
 243         return 1;
 244 }
 245
 246 static struct basic_block *trivial_common_parent(struct basic_block *bb1, struct basic_block *bb2)
 247 {
 248         struct basic_block *parent;
 249
 250         if (bb_list_size(bb1->parents) != 1)
 251                 return 0;
 252         parent = first_basic_block(bb1->parents);
 253         if (bb_list_size(bb2->parents) != 1)
 254                 return 0;
 255         if (first_basic_block(bb2->parents) != parent)
 256                 return 0;
 257         return parent;
 258 }
 259
 260 static inline void remove_instruction(struct instruction_list **list, struct instruction *insn, int count)
 261 {
 262         delete_ptr_list_entry((struct ptr_list **)list, insn, count);
 263 }
 264
 265 static void add_instruction_to_end(struct instruction *insn, struct basic_block *bb)
 266 {
 267         struct instruction *br = delete_last_instruction(&bb->insns);
 268         insn->bb = bb;
 269         add_instruction(&bb->insns, insn);
 270         add_instruction(&bb->insns, br);
 271 }
 272
 273 static struct instruction * try_to_cse(struct entrypoint *ep, struct instruction *i1, struct instruction *i2)
 274 {
 275         struct basic_block *b1, *b2, *common;
 276
 277         /*
 278          * Ok, i1 and i2 are the same instruction, modulo "target".
 279          * We should now see if we can combine them.
 280          */
 281         b1 = i1->bb;
 282         b2 = i2->bb;
 283
 284         /*
 285          * PHI-nodes do not care where they are - the only thing that matters
 286          * are the PHI _sources_.
 287          */
 288         if (i1->opcode == OP_PHI)
 289                 return cse_one_instruction(i1, i2);
 290
 291         /*
 292          * Currently we only handle the uninteresting degenerate case where
 293          * the CSE is inside one basic-block.
 294          */
 295         if (b1 == b2) {
 296                 struct instruction *insn;
 297                 FOR_EACH_PTR(b1->insns, insn) {
 298                         if (insn == i1)
 299                                 return cse_one_instruction(i2, i1);
 300                         if (insn == i2)
 301                                 return cse_one_instruction(i1, i2);
 302                 } END_FOR_EACH_PTR(insn);
 303                 warning(b1->pos, "Whaa? unable to find CSE instructions");
 304                 return i1;
 305         }
 306         if (bb_dominates(ep, b1, b2, ++bb_generation))
 307                 return cse_one_instruction(i2, i1);
 308
 309         if (bb_dominates(ep, b2, b1, ++bb_generation))
 310                 return cse_one_instruction(i1, i2);
 311
 312         /* No direct dominance - but we could try to find a common ancestor.. */
 313         common = trivial_common_parent(b1, b2);
 314         if (common) {
 315                 i1 = cse_one_instruction(i2, i1);
 316                 remove_instruction(&b1->insns, i1, 1);
 317                 add_instruction_to_end(i1, common);
 318         }
 319
 320         return i1;
 321 }
 322
 323 void cleanup_and_cse(struct entrypoint *ep)
 324 {
 325         int i;
 326
 327         simplify_memops(ep);
 328 repeat:
 329         repeat_phase = 0;
 330         clean_up_insns(ep);
 331         for (i = 0; i < INSN_HASH_SIZE; i++) {
 332                 struct instruction_list **list = insn_hash_table + i;
 333                 if (*list) {
 334                         if (instruction_list_size(*list) > 1) {
 335                                 struct instruction *insn, *last;
 336
 337                                 sort_instruction_list(list);
 338
 339                                 last = NULL;
 340                                 FOR_EACH_PTR(*list, insn) {
 341                                         if (!insn->bb)
 342                                                 continue;
 343                                         if (last) {
 344                                                 if (!insn_compare(last, insn))
 345                                                         insn = try_to_cse(ep, last, insn);
 346                                         }
 347                                         last = insn;
 348                                 } END_FOR_EACH_PTR(insn);
 349                         }
 350                         free_ptr_list((struct ptr_list **)list);
 351                 }
 352         }
 353
 354         if (repeat_phase & REPEAT_SYMBOL_CLEANUP)
 355                 simplify_memops(ep);
 356
 357         if (repeat_phase & REPEAT_CSE)
 358                 goto repeat;
 359 }