improve lua opcode dispatch performance by using computed goto instead of switch/case - improves performance by about 10% in a simple loop test

git-svn-id: svn://svn.openwrt.org/openwrt/trunk@12378 3c298f89-4303-0410-b956-a3cf2f4a3e73

1 files changed, 363 insertions, 0 deletions

diff --git a/package/lua/patches/300-opcode_performance.patch b/package/lua/patches/300-opcode_performance.patch
new file mode 100644
index 000000000..dbc89a8ed
--- /dev/null
+++ b/package/lua/patches/300-opcode_performance.patch
@@ -0,0 +1,363 @@
+--- a/src/lvm.c
++++ b/src/lvm.c
+@@ -31,6 +31,9 @@
+ /* limit for table tag-method chains (to avoid loops) */
+ #define MAXTAGLOOP	100
+ 
++#ifdef __GNUC__
++#define COMPUTED_GOTO 1
++#endif
+ 
+ /*
+  * If 'obj' is a string, it is tried to be interpreted as a number.
+@@ -562,12 +565,65 @@
+     ARITH_OP1_END
+ #endif
+ 
++#ifdef COMPUTED_GOTO
++#define OPCODE_TARGET(op) DO_OP_##op:
++#define CALL_OPCODE(op) \
++	if ((op < sizeof(opcodes) / sizeof(opcodes[0])) && opcodes[op]) \
++		goto *opcodes[op];
++#define OPCODE_PTR(op) [OP_##op] = &&DO_OP_##op
++#else
++#define OPCODE_TARGET(op) case OP_##op:
++#define CALL_OPCODE(op) switch (op)
++#endif
++
+ 
+ void luaV_execute (lua_State *L, int nexeccalls) {
+   LClosure *cl;
+   StkId base;
+   TValue *k;
+   const Instruction *pc;
++#ifdef COMPUTED_GOTO
++  static const void *opcodes[] = {
++   OPCODE_PTR(MOVE),
++   OPCODE_PTR(LOADK),
++   OPCODE_PTR(LOADBOOL),
++   OPCODE_PTR(LOADNIL),
++   OPCODE_PTR(GETUPVAL),
++   OPCODE_PTR(GETGLOBAL),
++   OPCODE_PTR(GETTABLE),
++   OPCODE_PTR(SETGLOBAL),
++   OPCODE_PTR(SETUPVAL),
++   OPCODE_PTR(SETTABLE),
++   OPCODE_PTR(NEWTABLE),
++   OPCODE_PTR(SELF),
++   OPCODE_PTR(ADD),
++   OPCODE_PTR(SUB),
++   OPCODE_PTR(MUL),
++   OPCODE_PTR(DIV),
++   OPCODE_PTR(MOD),
++   OPCODE_PTR(POW),
++   OPCODE_PTR(UNM),
++   OPCODE_PTR(NOT),
++   OPCODE_PTR(LEN),
++   OPCODE_PTR(CONCAT),
++   OPCODE_PTR(JMP),
++   OPCODE_PTR(EQ),
++   OPCODE_PTR(LT),
++   OPCODE_PTR(LE),
++   OPCODE_PTR(TEST),
++   OPCODE_PTR(TESTSET),
++   OPCODE_PTR(CALL),
++   OPCODE_PTR(TAILCALL),
++   OPCODE_PTR(RETURN),
++   OPCODE_PTR(FORLOOP),
++   OPCODE_PTR(FORPREP),
++   OPCODE_PTR(TFORLOOP),
++   OPCODE_PTR(SETLIST),
++   OPCODE_PTR(CLOSE),
++   OPCODE_PTR(CLOSURE),
++   OPCODE_PTR(VARARG)
++  };
++#endif
+  reentry:  /* entry point */
+   lua_assert(isLua(L->ci));
+   pc = L->savedpc;
+@@ -593,32 +649,32 @@
+     lua_assert(base <= L->top && L->top <= L->stack + L->stacksize);
+     lua_assert(L->top == L->ci->top || luaG_checkopenop(i));
+     switch (GET_OPCODE(i)) {
+-      case OP_MOVE: {
++      OPCODE_TARGET(MOVE) {
+         setobjs2s(L, ra, RB(i));
+         continue;
+       }
+-      case OP_LOADK: {
++      OPCODE_TARGET(LOADK) {
+         setobj2s(L, ra, KBx(i));
+         continue;
+       }
+-      case OP_LOADBOOL: {
++      OPCODE_TARGET(LOADBOOL) {
+         setbvalue(ra, GETARG_B(i));
+         if (GETARG_C(i)) pc++;  /* skip next instruction (if C) */
+         continue;
+       }
+-      case OP_LOADNIL: {
++      OPCODE_TARGET(LOADNIL) {
+         TValue *rb = RB(i);
+         do {
+           setnilvalue(rb--);
+         } while (rb >= ra);
+         continue;
+       }
+-      case OP_GETUPVAL: {
++      OPCODE_TARGET(GETUPVAL) {
+         int b = GETARG_B(i);
+         setobj2s(L, ra, cl->upvals[b]->v);
+         continue;
+       }
+-      case OP_GETGLOBAL: {
++      OPCODE_TARGET(GETGLOBAL) {
+         TValue g;
+         TValue *rb = KBx(i);
+         sethvalue(L, &g, cl->env);
+@@ -626,88 +682,88 @@
+         Protect(luaV_gettable(L, &g, rb, ra));
+         continue;
+       }
+-      case OP_GETTABLE: {
++      OPCODE_TARGET(GETTABLE) {
+         Protect(luaV_gettable(L, RB(i), RKC(i), ra));
+         continue;
+       }
+-      case OP_SETGLOBAL: {
++      OPCODE_TARGET(SETGLOBAL) {
+         TValue g;
+         sethvalue(L, &g, cl->env);
+         lua_assert(ttisstring(KBx(i)));
+         Protect(luaV_settable(L, &g, KBx(i), ra));
+         continue;
+       }
+-      case OP_SETUPVAL: {
++      OPCODE_TARGET(SETUPVAL) {
+         UpVal *uv = cl->upvals[GETARG_B(i)];
+         setobj(L, uv->v, ra);
+         luaC_barrier(L, uv, ra);
+         continue;
+       }
+-      case OP_SETTABLE: {
++      OPCODE_TARGET(SETTABLE) {
+         Protect(luaV_settable(L, ra, RKB(i), RKC(i)));
+         continue;
+       }
+-      case OP_NEWTABLE: {
++      OPCODE_TARGET(NEWTABLE) {
+         int b = GETARG_B(i);
+         int c = GETARG_C(i);
+         sethvalue(L, ra, luaH_new(L, luaO_fb2int(b), luaO_fb2int(c)));
+         Protect(luaC_checkGC(L));
+         continue;
+       }
+-      case OP_SELF: {
++      OPCODE_TARGET(SELF) {
+         StkId rb = RB(i);
+         setobjs2s(L, ra+1, rb);
+         Protect(luaV_gettable(L, rb, RKC(i), ra));
+         continue;
+       }
+-      case OP_ADD: {
++      OPCODE_TARGET(ADD) {
+         TValue *rb = RKB(i), *rc= RKC(i);
+         arith_op_continue( luai_numadd, try_addint, luai_vectadd );
+         Protect(Arith(L, ra, rb, rc, TM_ADD)); \
+         continue;
+       }
+-      case OP_SUB: {
++      OPCODE_TARGET(SUB) {
+         TValue *rb = RKB(i), *rc= RKC(i);
+         arith_op_continue( luai_numsub, try_subint, luai_vectsub );
+         Protect(Arith(L, ra, rb, rc, TM_SUB));
+         continue;
+       }
+-      case OP_MUL: {
++      OPCODE_TARGET(MUL) {
+         TValue *rb = RKB(i), *rc= RKC(i);
+         arith_op_continue(luai_nummul, try_mulint, luai_vectmul);
+         Protect(Arith(L, ra, rb, rc, TM_MUL));
+         continue;
+       }
+-      case OP_DIV: {
++      OPCODE_TARGET(DIV) {
+         TValue *rb = RKB(i), *rc= RKC(i);
+         arith_op_continue(luai_numdiv, try_divint, luai_vectdiv);
+         Protect(Arith(L, ra, rb, rc, TM_DIV));
+         continue;
+       }
+-      case OP_MOD: {
++      OPCODE_TARGET(MOD) {
+         TValue *rb = RKB(i), *rc= RKC(i);
+         arith_op_continue_scalar(luai_nummod, try_modint);  /* scalars only */
+         Protect(Arith(L, ra, rb, rc, TM_MOD));
+         continue;
+       }
+-      case OP_POW: {
++      OPCODE_TARGET(POW) {
+         TValue *rb = RKB(i), *rc= RKC(i);
+         arith_op_continue(luai_numpow, try_powint, luai_vectpow);
+         Protect(Arith(L, ra, rb, rc, TM_POW));
+         continue;
+       }
+-      case OP_UNM: {
++      OPCODE_TARGET(UNM) {
+         TValue *rb = RB(i);
+         arith_op1_continue(luai_numunm, try_unmint, luai_vectunm);
+         Protect(Arith(L, ra, rb, rb, TM_UNM));
+         continue;
+       }
+-      case OP_NOT: {
++      OPCODE_TARGET(NOT) {
+         int res = l_isfalse(RB(i));  /* next assignment may change this value */
+         setbvalue(ra, res);
+         continue;
+       }
+-      case OP_LEN: {
++      OPCODE_TARGET(LEN) {
+         const TValue *rb = RB(i);
+         switch (ttype(rb)) {
+           case LUA_TTABLE: {
+@@ -727,18 +783,18 @@
+         }
+         continue;
+       }
+-      case OP_CONCAT: {
++      OPCODE_TARGET(CONCAT) {
+         int b = GETARG_B(i);
+         int c = GETARG_C(i);
+         Protect(luaV_concat(L, c-b+1, c); luaC_checkGC(L));
+         setobjs2s(L, RA(i), base+b);
+         continue;
+       }
+-      case OP_JMP: {
++      OPCODE_TARGET(JMP) {
+         dojump(L, pc, GETARG_sBx(i));
+         continue;
+       }
+-      case OP_EQ: {
++      OPCODE_TARGET(EQ) {
+         TValue *rb = RKB(i);
+         TValue *rc = RKC(i);
+         Protect(
+@@ -748,7 +804,7 @@
+         pc++;
+         continue;
+       }
+-      case OP_LT: {
++      OPCODE_TARGET(LT) {
+         Protect(
+           if (luaV_lessthan(L, RKB(i), RKC(i)) == GETARG_A(i))
+             dojump(L, pc, GETARG_sBx(*pc));
+@@ -756,7 +812,7 @@
+         pc++;
+         continue;
+       }
+-      case OP_LE: {
++      OPCODE_TARGET(LE) {
+         Protect(
+           if (lessequal(L, RKB(i), RKC(i)) == GETARG_A(i))
+             dojump(L, pc, GETARG_sBx(*pc));
+@@ -764,13 +820,13 @@
+         pc++;
+         continue;
+       }
+-      case OP_TEST: {
++      OPCODE_TARGET(TEST) {
+         if (l_isfalse(ra) != GETARG_C(i))
+           dojump(L, pc, GETARG_sBx(*pc));
+         pc++;
+         continue;
+       }
+-      case OP_TESTSET: {
++      OPCODE_TARGET(TESTSET) {
+         TValue *rb = RB(i);
+         if (l_isfalse(rb) != GETARG_C(i)) {
+           setobjs2s(L, ra, rb);
+@@ -779,7 +835,7 @@
+         pc++;
+         continue;
+       }
+-      case OP_CALL: {
++      OPCODE_TARGET(CALL) {
+         int b = GETARG_B(i);
+         int nresults = GETARG_C(i) - 1;
+         if (b != 0) L->top = ra+b;  /* else previous instruction set top */
+@@ -800,7 +856,7 @@
+           }
+         }
+       }
+-      case OP_TAILCALL: {
++      OPCODE_TARGET(TAILCALL) {
+         int b = GETARG_B(i);
+         if (b != 0) L->top = ra+b;  /* else previous instruction set top */
+         L->savedpc = pc;
+@@ -832,7 +888,7 @@
+           }
+         }
+       }
+-      case OP_RETURN: {
++      OPCODE_TARGET(RETURN) {
+         int b = GETARG_B(i);
+         if (b != 0) L->top = ra+b-1;
+         if (L->openupval) luaF_close(L, base);
+@@ -847,7 +903,7 @@
+           goto reentry;
+         }
+       }
+-      case OP_FORLOOP: {
++      OPCODE_TARGET(FORLOOP) {
+         /* If start,step and limit are all integers, we don't need to check
+          * against overflow in the looping.
+          */
+@@ -875,7 +931,7 @@
+         }
+         continue;
+       }
+-      case OP_FORPREP: {
++      OPCODE_TARGET(FORPREP) {
+         const TValue *init = ra;
+         const TValue *plimit = ra+1;
+         const TValue *pstep = ra+2;
+@@ -898,7 +954,7 @@
+         dojump(L, pc, GETARG_sBx(i));
+         continue;
+       }
+-      case OP_TFORLOOP: {
++      OPCODE_TARGET(TFORLOOP) {
+         StkId cb = ra + 3;  /* call base */
+         setobjs2s(L, cb+2, ra+2);
+         setobjs2s(L, cb+1, ra+1);
+@@ -914,7 +970,7 @@
+         pc++;
+         continue;
+       }
+-      case OP_SETLIST: {
++      OPCODE_TARGET(SETLIST) {
+         int n = GETARG_B(i);
+         int c = GETARG_C(i);
+         int last;
+@@ -936,11 +992,11 @@
+         }
+         continue;
+       }
+-      case OP_CLOSE: {
++      OPCODE_TARGET(CLOSE) {
+         luaF_close(L, ra);
+         continue;
+       }
+-      case OP_CLOSURE: {
++      OPCODE_TARGET(CLOSURE) {
+         Proto *p;
+         Closure *ncl;
+         int nup, j;
+@@ -960,7 +1016,7 @@
+         Protect(luaC_checkGC(L));
+         continue;
+       }
+-      case OP_VARARG: {
++      OPCODE_TARGET(VARARG) {
+         int b = GETARG_B(i) - 1;
+         int j;
+         CallInfo *ci = L->ci;