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