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 256
  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                 /* Fall through */
  51
  52         /* Binary arithmetic */
  53         case OP_ADD: case OP_SUB:
  54         case OP_MULU: case OP_MULS:
  55         case OP_DIVU: case OP_DIVS:
  56         case OP_MODU: case OP_MODS:
  57         case OP_SHL:
  58         case OP_LSR: case OP_ASR:
  59         case OP_AND: case OP_OR:
  60
  61         /* Binary logical */
  62         case OP_XOR: case OP_AND_BOOL:
  63         case OP_OR_BOOL:
  64
  65         /* Binary comparison */
  66         case OP_SET_EQ: case OP_SET_NE:
  67         case OP_SET_LE: case OP_SET_GE:
  68         case OP_SET_LT: case OP_SET_GT:
  69         case OP_SET_B:  case OP_SET_A:
  70         case OP_SET_BE: case OP_SET_AE:
  71                 hash += hashval(insn->src2);
  72                 /* Fall through */
  73
  74         /* Unary */
  75         case OP_NOT: case OP_NEG:
  76                 hash += hashval(insn->src1);
  77                 break;
  78
  79         case OP_SETVAL:
  80                 hash += hashval(insn->val);
  81                 break;
  82
  83         case OP_SYMADDR:
  84                 hash += hashval(insn->symbol);
  85                 break;
  86
  87         case OP_CAST:
  88         case OP_SCAST:
  89         case OP_PTRCAST:
  90                 /*
  91                  * This is crap! Many "orig_types" are the
  92                  * same as far as casts go, we should generate
  93                  * some kind of "type hash" that is identical
  94                  * for identical casts
  95                  */
  96                 hash += hashval(insn->orig_type);
  97                 hash += hashval(insn->src);
  98                 break;
  99
 100         /* Other */
 101         case OP_PHI: {
 102                 pseudo_t phi;
 103                 FOR_EACH_PTR(insn->phi_list, phi) {
 104                         struct instruction *def;
 105                         if (phi == VOID || !phi->def)
 106                                 continue;
 107                         def = phi->def;
 108                         hash += hashval(def->src1);
 109                         hash += hashval(def->bb);
 110                 } END_FOR_EACH_PTR(phi);
 111                 break;
 112         }
 113
 114         default:
 115                 /*
 116                  * Nothing to do, don't even bother hashing them,
 117                  * we're not going to try to CSE them
 118                  */
 119                 return;
 120         }
 121         hash += hash >> 16;
 122         hash &= INSN_HASH_SIZE-1;
 123         add_instruction(insn_hash_table + hash, insn);
 124 }
 125
 126 static void clean_up_insns(struct entrypoint *ep)
 127 {
 128         struct basic_block *bb;
 129
 130         FOR_EACH_PTR(ep->bbs, bb) {
 131                 struct instruction *insn;
 132                 FOR_EACH_PTR(bb->insns, insn) {
 133                         clean_up_one_instruction(bb, insn);
 134                 } END_FOR_EACH_PTR(insn);
 135         } END_FOR_EACH_PTR(bb);
 136 }
 137
 138 /* Compare two (sorted) phi-lists */
 139 static int phi_list_compare(struct pseudo_list *l1, struct pseudo_list *l2)
 140 {
 141         pseudo_t phi1, phi2;
 142
 143         PREPARE_PTR_LIST(l1, phi1);
 144         PREPARE_PTR_LIST(l2, phi2);
 145         for (;;) {
 146                 int cmp;
 147
 148                 while (phi1 && (phi1 == VOID || !phi1->def))
 149                         NEXT_PTR_LIST(phi1);
 150                 while (phi2 && (phi2 == VOID || !phi2->def))
 151                         NEXT_PTR_LIST(phi2);
 152
 153                 if (!phi1)
 154                         return phi2 ? -1 : 0;
 155                 if (!phi2)
 156                         return phi1 ? 1 : 0;
 157                 cmp = phi_compare(phi1, phi2);
 158                 if (cmp)
 159                         return cmp;
 160                 NEXT_PTR_LIST(phi1);
 161                 NEXT_PTR_LIST(phi2);
 162         }
 163         /* Not reached, but we need to make the nesting come out right */
 164         FINISH_PTR_LIST(phi2);
 165         FINISH_PTR_LIST(phi1);
 166 }
 167
 168 static int insn_compare(const void *_i1, const void *_i2)
 169 {
 170         const struct instruction *i1 = _i1;
 171         const struct instruction *i2 = _i2;
 172
 173         if (i1->opcode != i2->opcode)
 174                 return i1->opcode < i2->opcode ? -1 : 1;
 175
 176         switch (i1->opcode) {
 177
 178         /* commutative binop */
 179         case OP_ADD:
 180         case OP_MULU: case OP_MULS:
 181         case OP_AND_BOOL: case OP_OR_BOOL:
 182         case OP_AND: case OP_OR:
 183         case OP_XOR:
 184         case OP_SET_EQ: case OP_SET_NE:
 185                 if (i1->src1 == i2->src2 && i1->src2 == i2->src1)
 186                         return 0;
 187                 goto case_binops;
 188
 189         case OP_SEL:
 190                 if (i1->src3 != i2->src3)
 191                         return i1->src3 < i2->src3 ? -1 : 1;
 192                 /* Fall-through to binops */
 193
 194         /* Binary arithmetic */
 195         case OP_SUB:
 196         case OP_DIVU: case OP_DIVS:
 197         case OP_MODU: case OP_MODS:
 198         case OP_SHL:
 199         case OP_LSR: case OP_ASR:
 200
 201         /* Binary comparison */
 202         case OP_SET_LE: case OP_SET_GE:
 203         case OP_SET_LT: case OP_SET_GT:
 204         case OP_SET_B:  case OP_SET_A:
 205         case OP_SET_BE: case OP_SET_AE:
 206         case_binops:
 207                 if (i1->src2 != i2->src2)
 208                         return i1->src2 < i2->src2 ? -1 : 1;
 209                 /* Fall through to unops */
 210
 211         /* Unary */
 212         case OP_NOT: case OP_NEG:
 213                 if (i1->src1 != i2->src1)
 214                         return i1->src1 < i2->src1 ? -1 : 1;
 215                 break;
 216
 217         case OP_SYMADDR:
 218                 if (i1->symbol != i2->symbol)
 219                         return i1->symbol < i2->symbol ? -1 : 1;
 220                 break;
 221
 222         case OP_SETVAL:
 223                 if (i1->val != i2->val)
 224                         return i1->val < i2->val ? -1 : 1;
 225                 break;
 226
 227         /* Other */
 228         case OP_PHI:
 229                 return phi_list_compare(i1->phi_list, i2->phi_list);
 230
 231         case OP_CAST:
 232         case OP_SCAST:
 233         case OP_PTRCAST:
 234                 /*
 235                  * This is crap! See the comments on hashing.
 236                  */
 237                 if (i1->orig_type != i2->orig_type)
 238                         return i1->orig_type < i2->orig_type ? -1 : 1;
 239                 if (i1->src != i2->src)
 240                         return i1->src < i2->src ? -1 : 1;
 241                 break;
 242
 243         default:
 244                 warning(i1->pos, "bad instruction on hash chain");
 245         }
 246         if (i1->size != i2->size)
 247                 return i1->size < i2->size ? -1 : 1;
 248         return 0;
 249 }
 250
 251 static void sort_instruction_list(struct instruction_list **list)
 252 {
 253         sort_list((struct ptr_list **)list , insn_compare);
 254 }
 255
 256 static struct instruction * cse_one_instruction(struct instruction *insn, struct instruction *def)
 257 {
 258         convert_instruction_target(insn, def->target);
 259
 260         kill_instruction(insn);
 261         repeat_phase |= REPEAT_CSE;
 262         return def;
 263 }
 264
 265 /*
 266  * Does "bb1" dominate "bb2"?
 267  */
 268 static int bb_dominates(struct entrypoint *ep, struct basic_block *bb1, struct basic_block *bb2, unsigned long generation)
 269 {
 270         struct basic_block *parent;
 271
 272         /* Nothing dominates the entrypoint.. */
 273         if (bb2 == ep->entry->bb)
 274                 return 0;
 275         FOR_EACH_PTR(bb2->parents, parent) {
 276                 if (parent == bb1)
 277                         continue;
 278                 if (parent->generation == generation)
 279                         continue;
 280                 parent->generation = generation;
 281                 if (!bb_dominates(ep, bb1, parent, generation))
 282                         return 0;
 283         } END_FOR_EACH_PTR(parent);
 284         return 1;
 285 }
 286
 287 static struct basic_block *trivial_common_parent(struct basic_block *bb1, struct basic_block *bb2)
 288 {
 289         struct basic_block *parent;
 290
 291         if (bb_list_size(bb1->parents) != 1)
 292                 return NULL;
 293         parent = first_basic_block(bb1->parents);
 294         if (bb_list_size(bb2->parents) != 1)
 295                 return NULL;
 296         if (first_basic_block(bb2->parents) != parent)
 297                 return NULL;
 298         return parent;
 299 }
 300
 301 static inline void remove_instruction(struct instruction_list **list, struct instruction *insn, int count)
 302 {
 303         delete_ptr_list_entry((struct ptr_list **)list, insn, count);
 304 }
 305
 306 static void add_instruction_to_end(struct instruction *insn, struct basic_block *bb)
 307 {
 308         struct instruction *br = delete_last_instruction(&bb->insns);
 309         insn->bb = bb;
 310         add_instruction(&bb->insns, insn);
 311         add_instruction(&bb->insns, br);
 312 }
 313
 314 static struct instruction * try_to_cse(struct entrypoint *ep, struct instruction *i1, struct instruction *i2)
 315 {
 316         struct basic_block *b1, *b2, *common;
 317
 318         /*
 319          * OK, i1 and i2 are the same instruction, modulo "target".
 320          * We should now see if we can combine them.
 321          */
 322         b1 = i1->bb;
 323         b2 = i2->bb;
 324
 325         /*
 326          * Currently we only handle the uninteresting degenerate case where
 327          * the CSE is inside one basic-block.
 328          */
 329         if (b1 == b2) {
 330                 struct instruction *insn;
 331                 FOR_EACH_PTR(b1->insns, insn) {
 332                         if (insn == i1)
 333                                 return cse_one_instruction(i2, i1);
 334                         if (insn == i2)
 335                                 return cse_one_instruction(i1, i2);
 336                 } END_FOR_EACH_PTR(insn);
 337                 warning(b1->pos, "Whaa? unable to find CSE instructions");
 338                 return i1;
 339         }
 340         if (bb_dominates(ep, b1, b2, ++bb_generation))
 341                 return cse_one_instruction(i2, i1);
 342
 343         if (bb_dominates(ep, b2, b1, ++bb_generation))
 344                 return cse_one_instruction(i1, i2);
 345
 346         /* No direct dominance - but we could try to find a common ancestor.. */
 347         common = trivial_common_parent(b1, b2);
 348         if (common) {
 349                 i1 = cse_one_instruction(i2, i1);
 350                 remove_instruction(&b1->insns, i1, 1);
 351                 add_instruction_to_end(i1, common);
 352         }
 353
 354         return i1;
 355 }
 356
 357 void cleanup_and_cse(struct entrypoint *ep)
 358 {
 359         int i;
 360
 361         simplify_memops(ep);
 362 repeat:
 363         repeat_phase = 0;
 364         clean_up_insns(ep);
 365         for (i = 0; i < INSN_HASH_SIZE; i++) {
 366                 struct instruction_list **list = insn_hash_table + i;
 367                 if (*list) {
 368                         if (instruction_list_size(*list) > 1) {
 369                                 struct instruction *insn, *last;
 370
 371                                 sort_instruction_list(list);
 372
 373                                 last = NULL;
 374                                 FOR_EACH_PTR(*list, insn) {
 375                                         if (!insn->bb)
 376                                                 continue;
 377                                         if (last) {
 378                                                 if (!insn_compare(last, insn))
 379                                                         insn = try_to_cse(ep, last, insn);
 380                                         }
 381                                         last = insn;
 382                                 } END_FOR_EACH_PTR(insn);
 383                         }
 384                         free_ptr_list((struct ptr_list **)list);
 385                 }
 386         }
 387
 388         if (repeat_phase & REPEAT_SYMBOL_CLEANUP)
 389                 simplify_memops(ep);
 390
 391         if (repeat_phase & REPEAT_CSE)
 392                 goto repeat;
 393 }