tools/lua/patches/300-opcode_performance.patch

   1 --- a/src/lvm.c
   2 +++ b/src/lvm.c
   3 @@ -31,6 +31,9 @@
   4  /* limit for table tag-method chains (to avoid loops) */
   5  #define MAXTAGLOOP     100
   6
   7 +#ifdef __GNUC__
   8 +#define COMPUTED_GOTO 1
   9 +#endif
  10
  11  /*
  12   * If 'obj' is a string, it is tried to be interpreted as a number.
  13 @@ -562,12 +565,63 @@
  14      ARITH_OP1_END
  15  #endif
  16
  17 +#ifdef COMPUTED_GOTO
  18 +#define OPCODE_TARGET(op) DO_OP_##op:
  19 +#define CALL_OPCODE(op) goto *opcodes[op];
  20 +#define OPCODE_PTR(op) [OP_##op] = &&DO_OP_##op
  21 +#else
  22 +#define OPCODE_TARGET(op) case OP_##op:
  23 +#define CALL_OPCODE(op) switch (op)
  24 +#endif
  25 +
  26
  27  void luaV_execute (lua_State *L, int nexeccalls) {
  28    LClosure *cl;
  29    StkId base;
  30    TValue *k;
  31    const Instruction *pc;
  32 +#ifdef COMPUTED_GOTO
  33 +  static const void *opcodes[] = {
  34 +   OPCODE_PTR(MOVE),
  35 +   OPCODE_PTR(LOADK),
  36 +   OPCODE_PTR(LOADBOOL),
  37 +   OPCODE_PTR(LOADNIL),
  38 +   OPCODE_PTR(GETUPVAL),
  39 +   OPCODE_PTR(GETGLOBAL),
  40 +   OPCODE_PTR(GETTABLE),
  41 +   OPCODE_PTR(SETGLOBAL),
  42 +   OPCODE_PTR(SETUPVAL),
  43 +   OPCODE_PTR(SETTABLE),
  44 +   OPCODE_PTR(NEWTABLE),
  45 +   OPCODE_PTR(SELF),
  46 +   OPCODE_PTR(ADD),
  47 +   OPCODE_PTR(SUB),
  48 +   OPCODE_PTR(MUL),
  49 +   OPCODE_PTR(DIV),
  50 +   OPCODE_PTR(MOD),
  51 +   OPCODE_PTR(POW),
  52 +   OPCODE_PTR(UNM),
  53 +   OPCODE_PTR(NOT),
  54 +   OPCODE_PTR(LEN),
  55 +   OPCODE_PTR(CONCAT),
  56 +   OPCODE_PTR(JMP),
  57 +   OPCODE_PTR(EQ),
  58 +   OPCODE_PTR(LT),
  59 +   OPCODE_PTR(LE),
  60 +   OPCODE_PTR(TEST),
  61 +   OPCODE_PTR(TESTSET),
  62 +   OPCODE_PTR(CALL),
  63 +   OPCODE_PTR(TAILCALL),
  64 +   OPCODE_PTR(RETURN),
  65 +   OPCODE_PTR(FORLOOP),
  66 +   OPCODE_PTR(FORPREP),
  67 +   OPCODE_PTR(TFORLOOP),
  68 +   OPCODE_PTR(SETLIST),
  69 +   OPCODE_PTR(CLOSE),
  70 +   OPCODE_PTR(CLOSURE),
  71 +   OPCODE_PTR(VARARG)
  72 +  };
  73 +#endif
  74   reentry:  /* entry point */
  75    lua_assert(isLua(L->ci));
  76    pc = L->savedpc;
  77 @@ -592,33 +646,33 @@
  78      lua_assert(base == L->base && L->base == L->ci->base);
  79      lua_assert(base <= L->top && L->top <= L->stack + L->stacksize);
  80      lua_assert(L->top == L->ci->top || luaG_checkopenop(i));
  81 -    switch (GET_OPCODE(i)) {
  82 -      case OP_MOVE: {
  83 +    CALL_OPCODE(GET_OPCODE(i)) {
  84 +      OPCODE_TARGET(MOVE) {
  85          setobjs2s(L, ra, RB(i));
  86          continue;
  87        }
  88 -      case OP_LOADK: {
  89 +      OPCODE_TARGET(LOADK) {
  90          setobj2s(L, ra, KBx(i));
  91          continue;
  92        }
  93 -      case OP_LOADBOOL: {
  94 +      OPCODE_TARGET(LOADBOOL) {
  95          setbvalue(ra, GETARG_B(i));
  96          if (GETARG_C(i)) pc++;  /* skip next instruction (if C) */
  97          continue;
  98        }
  99 -      case OP_LOADNIL: {
 100 +      OPCODE_TARGET(LOADNIL) {
 101          TValue *rb = RB(i);
 102          do {
 103            setnilvalue(rb--);
 104          } while (rb >= ra);
 105          continue;
 106        }
 107 -      case OP_GETUPVAL: {
 108 +      OPCODE_TARGET(GETUPVAL) {
 109          int b = GETARG_B(i);
 110          setobj2s(L, ra, cl->upvals[b]->v);
 111          continue;
 112        }
 113 -      case OP_GETGLOBAL: {
 114 +      OPCODE_TARGET(GETGLOBAL) {
 115          TValue g;
 116          TValue *rb = KBx(i);
 117          sethvalue(L, &g, cl->env);
 118 @@ -626,88 +680,88 @@
 119          Protect(luaV_gettable(L, &g, rb, ra));
 120          continue;
 121        }
 122 -      case OP_GETTABLE: {
 123 +      OPCODE_TARGET(GETTABLE) {
 124          Protect(luaV_gettable(L, RB(i), RKC(i), ra));
 125          continue;
 126        }
 127 -      case OP_SETGLOBAL: {
 128 +      OPCODE_TARGET(SETGLOBAL) {
 129          TValue g;
 130          sethvalue(L, &g, cl->env);
 131          lua_assert(ttisstring(KBx(i)));
 132          Protect(luaV_settable(L, &g, KBx(i), ra));
 133          continue;
 134        }
 135 -      case OP_SETUPVAL: {
 136 +      OPCODE_TARGET(SETUPVAL) {
 137          UpVal *uv = cl->upvals[GETARG_B(i)];
 138          setobj(L, uv->v, ra);
 139          luaC_barrier(L, uv, ra);
 140          continue;
 141        }
 142 -      case OP_SETTABLE: {
 143 +      OPCODE_TARGET(SETTABLE) {
 144          Protect(luaV_settable(L, ra, RKB(i), RKC(i)));
 145          continue;
 146        }
 147 -      case OP_NEWTABLE: {
 148 +      OPCODE_TARGET(NEWTABLE) {
 149          int b = GETARG_B(i);
 150          int c = GETARG_C(i);
 151          sethvalue(L, ra, luaH_new(L, luaO_fb2int(b), luaO_fb2int(c)));
 152          Protect(luaC_checkGC(L));
 153          continue;
 154        }
 155 -      case OP_SELF: {
 156 +      OPCODE_TARGET(SELF) {
 157          StkId rb = RB(i);
 158          setobjs2s(L, ra+1, rb);
 159          Protect(luaV_gettable(L, rb, RKC(i), ra));
 160          continue;
 161        }
 162 -      case OP_ADD: {
 163 +      OPCODE_TARGET(ADD) {
 164          TValue *rb = RKB(i), *rc= RKC(i);
 165          arith_op_continue( luai_numadd, try_addint, luai_vectadd );
 166          Protect(Arith(L, ra, rb, rc, TM_ADD)); \
 167          continue;
 168        }
 169 -      case OP_SUB: {
 170 +      OPCODE_TARGET(SUB) {
 171          TValue *rb = RKB(i), *rc= RKC(i);
 172          arith_op_continue( luai_numsub, try_subint, luai_vectsub );
 173          Protect(Arith(L, ra, rb, rc, TM_SUB));
 174          continue;
 175        }
 176 -      case OP_MUL: {
 177 +      OPCODE_TARGET(MUL) {
 178          TValue *rb = RKB(i), *rc= RKC(i);
 179          arith_op_continue(luai_nummul, try_mulint, luai_vectmul);
 180          Protect(Arith(L, ra, rb, rc, TM_MUL));
 181          continue;
 182        }
 183 -      case OP_DIV: {
 184 +      OPCODE_TARGET(DIV) {
 185          TValue *rb = RKB(i), *rc= RKC(i);
 186          arith_op_continue(luai_numdiv, try_divint, luai_vectdiv);
 187          Protect(Arith(L, ra, rb, rc, TM_DIV));
 188          continue;
 189        }
 190 -      case OP_MOD: {
 191 +      OPCODE_TARGET(MOD) {
 192          TValue *rb = RKB(i), *rc= RKC(i);
 193          arith_op_continue_scalar(luai_nummod, try_modint);  /* scalars only */
 194          Protect(Arith(L, ra, rb, rc, TM_MOD));
 195          continue;
 196        }
 197 -      case OP_POW: {
 198 +      OPCODE_TARGET(POW) {
 199          TValue *rb = RKB(i), *rc= RKC(i);
 200          arith_op_continue(luai_numpow, try_powint, luai_vectpow);
 201          Protect(Arith(L, ra, rb, rc, TM_POW));
 202          continue;
 203        }
 204 -      case OP_UNM: {
 205 +      OPCODE_TARGET(UNM) {
 206          TValue *rb = RB(i);
 207          arith_op1_continue(luai_numunm, try_unmint, luai_vectunm);
 208          Protect(Arith(L, ra, rb, rb, TM_UNM));
 209          continue;
 210        }
 211 -      case OP_NOT: {
 212 +      OPCODE_TARGET(NOT) {
 213          int res = l_isfalse(RB(i));  /* next assignment may change this value */
 214          setbvalue(ra, res);
 215          continue;
 216        }
 217 -      case OP_LEN: {
 218 +      OPCODE_TARGET(LEN) {
 219          const TValue *rb = RB(i);
 220          switch (ttype(rb)) {
 221            case LUA_TTABLE: {
 222 @@ -727,18 +781,18 @@
 223          }
 224          continue;
 225        }
 226 -      case OP_CONCAT: {
 227 +      OPCODE_TARGET(CONCAT) {
 228          int b = GETARG_B(i);
 229          int c = GETARG_C(i);
 230          Protect(luaV_concat(L, c-b+1, c); luaC_checkGC(L));
 231          setobjs2s(L, RA(i), base+b);
 232          continue;
 233        }
 234 -      case OP_JMP: {
 235 +      OPCODE_TARGET(JMP) {
 236          dojump(L, pc, GETARG_sBx(i));
 237          continue;
 238        }
 239 -      case OP_EQ: {
 240 +      OPCODE_TARGET(EQ) {
 241          TValue *rb = RKB(i);
 242          TValue *rc = RKC(i);
 243          Protect(
 244 @@ -748,7 +802,7 @@
 245          pc++;
 246          continue;
 247        }
 248 -      case OP_LT: {
 249 +      OPCODE_TARGET(LT) {
 250          Protect(
 251            if (luaV_lessthan(L, RKB(i), RKC(i)) == GETARG_A(i))
 252              dojump(L, pc, GETARG_sBx(*pc));
 253 @@ -756,7 +810,7 @@
 254          pc++;
 255          continue;
 256        }
 257 -      case OP_LE: {
 258 +      OPCODE_TARGET(LE) {
 259          Protect(
 260            if (lessequal(L, RKB(i), RKC(i)) == GETARG_A(i))
 261              dojump(L, pc, GETARG_sBx(*pc));
 262 @@ -764,13 +818,13 @@
 263          pc++;
 264          continue;
 265        }
 266 -      case OP_TEST: {
 267 +      OPCODE_TARGET(TEST) {
 268          if (l_isfalse(ra) != GETARG_C(i))
 269            dojump(L, pc, GETARG_sBx(*pc));
 270          pc++;
 271          continue;
 272        }
 273 -      case OP_TESTSET: {
 274 +      OPCODE_TARGET(TESTSET) {
 275          TValue *rb = RB(i);
 276          if (l_isfalse(rb) != GETARG_C(i)) {
 277            setobjs2s(L, ra, rb);
 278 @@ -779,7 +833,7 @@
 279          pc++;
 280          continue;
 281        }
 282 -      case OP_CALL: {
 283 +      OPCODE_TARGET(CALL) {
 284          int b = GETARG_B(i);
 285          int nresults = GETARG_C(i) - 1;
 286          if (b != 0) L->top = ra+b;  /* else previous instruction set top */
 287 @@ -800,7 +854,7 @@
 288            }
 289          }
 290        }
 291 -      case OP_TAILCALL: {
 292 +      OPCODE_TARGET(TAILCALL) {
 293          int b = GETARG_B(i);
 294          if (b != 0) L->top = ra+b;  /* else previous instruction set top */
 295          L->savedpc = pc;
 296 @@ -832,7 +886,7 @@
 297            }
 298          }
 299        }
 300 -      case OP_RETURN: {
 301 +      OPCODE_TARGET(RETURN) {
 302          int b = GETARG_B(i);
 303          if (b != 0) L->top = ra+b-1;
 304          if (L->openupval) luaF_close(L, base);
 305 @@ -847,7 +901,7 @@
 306            goto reentry;
 307          }
 308        }
 309 -      case OP_FORLOOP: {
 310 +      OPCODE_TARGET(FORLOOP) {
 311          /* If start,step and limit are all integers, we don't need to check
 312           * against overflow in the looping.
 313           */
 314 @@ -875,7 +929,7 @@
 315          }
 316          continue;
 317        }
 318 -      case OP_FORPREP: {
 319 +      OPCODE_TARGET(FORPREP) {
 320          const TValue *init = ra;
 321          const TValue *plimit = ra+1;
 322          const TValue *pstep = ra+2;
 323 @@ -898,7 +952,7 @@
 324          dojump(L, pc, GETARG_sBx(i));
 325          continue;
 326        }
 327 -      case OP_TFORLOOP: {
 328 +      OPCODE_TARGET(TFORLOOP) {
 329          StkId cb = ra + 3;  /* call base */
 330          setobjs2s(L, cb+2, ra+2);
 331          setobjs2s(L, cb+1, ra+1);
 332 @@ -914,7 +968,7 @@
 333          pc++;
 334          continue;
 335        }
 336 -      case OP_SETLIST: {
 337 +      OPCODE_TARGET(SETLIST) {
 338          int n = GETARG_B(i);
 339          int c = GETARG_C(i);
 340          int last;
 341 @@ -936,11 +990,11 @@
 342          }
 343          continue;
 344        }
 345 -      case OP_CLOSE: {
 346 +      OPCODE_TARGET(CLOSE) {
 347          luaF_close(L, ra);
 348          continue;
 349        }
 350 -      case OP_CLOSURE: {
 351 +      OPCODE_TARGET(CLOSURE) {
 352          Proto *p;
 353          Closure *ncl;
 354          int nup, j;
 355 @@ -960,7 +1014,7 @@
 356          Protect(luaC_checkGC(L));
 357          continue;
 358        }
 359 -      case OP_VARARG: {
 360 +      OPCODE_TARGET(VARARG) {
 361          int b = GETARG_B(i) - 1;
 362          int j;
 363          CallInfo *ci = L->ci;