Specialize metatables for method calls

src/lj_bc.h

@@ -144,6 +144,7 @@
   _(GSET,	var,	___,	str,	newindex) \
   _(TGETV,	dst,	var,	var,	index) \
   _(TGETS,	dst,	var,	str,	index) \
+  _(TGETSS,	dst,	var,	str,	index) \
   _(TGETB,	dst,	var,	lit,	index) \
   _(TGETR,	dst,	var,	var,	index) \
   _(TSETV,	var,	var,	var,	newindex) \
@@ -259,6 +260,8 @@ static LJ_AINLINE int bc_isret(BCOp op)
   return (op == BC_RETM || op == BC_RET || op == BC_RET0 || op == BC_RET1);
 }
 
+#define bc_isalias(op) ((op) == BC_TGETSS)
+
 LJ_DATA const uint16_t lj_bc_mode[];
 LJ_DATA const uint16_t lj_bc_ofs[];
 

src/lj_debug.c

@@ -271,13 +271,10 @@ restart:
 	return "global";
       case BC_TGETS:
 	*name = strdata(gco2str(proto_kgc(pt, ~(ptrdiff_t)bc_c(ins))));
-	if (ip > proto_bc(pt)) {
-	  BCIns insp = ip[-1];
-	  if (bc_op(insp) == BC_MOV && bc_a(insp) == ra+1+LJ_FR2 &&
-	      bc_d(insp) == bc_b(ins))
-	    return "method";
-	}
 	return "field";
+      case BC_TGETSS:
+	*name = strdata(gco2str(proto_kgc(pt, ~(ptrdiff_t)bc_c(ins))));
+	return "method";
       case BC_UGET:
 	*name = lj_debug_uvname(pt, bc_d(ins));
 	return "upvalue";

src/lj_ffrecord.c

@@ -225,6 +225,7 @@ static void LJ_FASTCALL recff_getmetatable(jit_State *J, RecordFFData *rd)
     RecordIndex ix;
     ix.tab = tr;
     copyTV(J->L, &ix.tabv, &rd->argv[0]);
+    ix.mtspec = 0;
     if (lj_record_mm_lookup(J, &ix, MM_metatable))
       J->base[0] = ix.mobj;
     else
@@ -241,6 +242,7 @@ static void LJ_FASTCALL recff_setmetatable(jit_State *J, RecordFFData *rd)
     RecordIndex ix;
     ix.tab = tr;
     copyTV(J->L, &ix.tabv, &rd->argv[0]);
+    ix.mtspec = 0;
     lj_record_mm_lookup(J, &ix, MM_metatable); /* Guard for no __metatable. */
     fref = emitir(IRT(IR_FREF, IRT_PGC), tr, IRFL_TAB_META);
     mtref = tref_isnil(mt) ? lj_ir_knull(J, IRT_TAB) : mt;
@@ -389,6 +391,7 @@ static int recff_metacall(jit_State *J, RecordFFData *rd, MMS mm)
   RecordIndex ix;
   ix.tab = J->base[0];
   copyTV(J->L, &ix.tabv, &rd->argv[0]);
+  ix.mtspec = 1;
   if (lj_record_mm_lookup(J, &ix, mm)) {  /* Has metamethod? */
     int errcode;
     TValue argv0;
@@ -548,6 +551,7 @@ static void LJ_FASTCALL recff_next(jit_State *J, RecordFFData *rd)
     ix.idxchain = (J->framedepth && frame_islua(J->L->base-1) &&
 		   bc_b(frame_pc(J->L->base-1)[-1])-1 < 2);
     ix.mobj = 0;  /* We don't need the next index. */
+    ix.mtspec = 0;
     rd->nres = lj_record_next(J, &ix);
     J->base[0] = ix.key;
     J->base[1] = ix.val;

src/lj_parse.c

@@ -675,7 +675,7 @@ static void bcemit_method(FuncState *fs, ExpDesc *e, ExpDesc *key)
   idx = const_str(fs, key);
   if (idx <= BCMAX_C) {
     bcreg_reserve(fs, 2+LJ_FR2);
-    bcemit_ABC(fs, BC_TGETS, func, obj, idx);
+    bcemit_ABC(fs, BC_TGETSS, func, obj, idx);
   } else {
     bcreg_reserve(fs, 3+LJ_FR2);
     bcemit_AD(fs, BC_KSTR, func+2+LJ_FR2, idx);

src/lj_record.c

@@ -686,6 +686,7 @@ static LoopEvent rec_itern(jit_State *J, BCReg ra, BCReg rb)
   copyTV(J->L, &ix.keyv, &J->L->base[ra-1]);
   ix.idxchain = (rb < 3);  /* Omit value type check, if unused. */
   ix.mobj = 1;  /* We need the next index, too. */
+  ix.mtspec = 0;
   J->maxslot = ra + lj_record_next(J, &ix);
   J->needsnap = 1;
   if (!tref_isnil(ix.key)) {  /* Looping back? */
@@ -818,6 +819,7 @@ static void rec_call_setup(jit_State *J, BCReg func, ptrdiff_t nargs)
   if (!tref_isfunc(fbase[0])) {  /* Resolve __call metamethod. */
     ix.tab = fbase[0];
     copyTV(J->L, &ix.tabv, functv);
+    ix.mtspec = 1;
     if (!lj_record_mm_lookup(J, &ix, MM_call) || !tref_isfunc(ix.mobj))
       lj_trace_err(J, LJ_TRERR_NOMM);
     for (i = ++nargs; i > LJ_FR2; i--)  /* Shift arguments up. */
@@ -1103,7 +1105,14 @@ int lj_record_mm_lookup(jit_State *J, RecordIndex *ix, MMS mm)
     goto nocheck;
   }
   ix->mt = mt ? mix.tab : TREF_NIL;
-  emitir(IRTG(mt ? IR_NE : IR_EQ, IRT_TAB), mix.tab, lj_ir_knull(J, IRT_TAB));
+  if (ix->mtspec && mt) {
+    TRef kmt = lj_ir_ktab(J, mt);
+    emitir(IRTG(IR_EQ, IRT_TAB), mix.tab, kmt);
+    mix.tab = kmt;
+    ix->mt = kmt;
+  } else {
+    emitir(IRTG(mt ? IR_NE : IR_EQ, IRT_TAB), mix.tab, lj_ir_knull(J, IRT_TAB));
+  }
 nocheck:
   if (mt) {
     GCstr *mmstr = mmname_str(J2G(J), mm);
@@ -1157,6 +1166,7 @@ static TRef rec_mm_len(jit_State *J, TRef tr, TValue *tv)
   RecordIndex ix;
   ix.tab = tr;
   copyTV(J->L, &ix.tabv, tv);
+  ix.mtspec = 1;
   if (lj_record_mm_lookup(J, &ix, MM_len)) {
     BCReg func = rec_mm_prep(J, lj_cont_ra);
     TRef *base = J->base + func;
@@ -2086,6 +2096,7 @@ static TRef rec_cat(jit_State *J, BCReg baseslot, BCReg topslot)
   ix.tab = top[-1];
   ix.key = top[0];
   memcpy(savetv, &J->L->base[topslot-1], sizeof(savetv));  /* Save slots. */
+  ix.mtspec = 1;
   rec_mm_arith(J, &ix, MM_concat);  /* Call __concat metamethod. */
   memcpy(&J->L->base[topslot-1], savetv, sizeof(savetv));  /* Restore slots. */
   return 0;  /* No result yet. */
@@ -2290,6 +2301,7 @@ void lj_record_ins(jit_State *J)
 	rc = lj_ir_kint(J, 0);
 	ta = IRT_INT;
       } else {
+	ix.mtspec = 1;
 	rec_mm_comp(J, &ix, (int)op);
 	break;
       }
@@ -2313,10 +2325,12 @@ void lj_record_ins(jit_State *J)
       int diff;
       rec_comp_prep(J);
       diff = lj_record_objcmp(J, ra, rc, rav, rcv);
-      if (diff == 2 || !(tref_istab(ra) || tref_isudata(ra)))
+      if (diff == 2 || !(tref_istab(ra) || tref_isudata(ra))) {
 	rec_comp_fixup(J, J->pc, ((int)op & 1) == !diff);
-      else if (diff == 1)  /* Only check __eq if different, but same type. */
+      } else if (diff == 1) { /* Only check __eq if different, but same type. */
+	ix.mtspec = 1;
 	rec_mm_equal(J, &ix, (int)op);
+      }
     }
     break;
 
@@ -2367,6 +2381,7 @@ void lj_record_ins(jit_State *J)
     } else {
       ix.tab = rc;
       copyTV(J->L, &ix.tabv, rcv);
+      ix.mtspec = 1;
       rc = rec_mm_arith(J, &ix, MM_unm);
     }
     break;
@@ -2383,27 +2398,33 @@ void lj_record_ins(jit_State *J)
   case BC_ADDVN: case BC_SUBVN: case BC_MULVN: case BC_DIVVN:
   case BC_ADDVV: case BC_SUBVV: case BC_MULVV: case BC_DIVVV: {
     MMS mm = bcmode_mm(op);
-    if (tref_isnumber_str(rb) && tref_isnumber_str(rc))
+    if (tref_isnumber_str(rb) && tref_isnumber_str(rc)) {
       rc = lj_opt_narrow_arith(J, rb, rc, rbv, rcv,
 			       (int)mm - (int)MM_add + (int)IR_ADD);
-    else
+    } else {
+      ix.mtspec = 1;
       rc = rec_mm_arith(J, &ix, mm);
+    }
     break;
     }
 
   case BC_MODVN: case BC_MODVV:
   recmod:
-    if (tref_isnumber_str(rb) && tref_isnumber_str(rc))
+    if (tref_isnumber_str(rb) && tref_isnumber_str(rc)) {
       rc = lj_opt_narrow_mod(J, rb, rc, rbv, rcv);
-    else
+    } else {
+      ix.mtspec = 1;
       rc = rec_mm_arith(J, &ix, MM_mod);
+    }
     break;
 
   case BC_POW:
-    if (tref_isnumber_str(rb) && tref_isnumber_str(rc))
+    if (tref_isnumber_str(rb) && tref_isnumber_str(rc)) {
       rc = lj_opt_narrow_arith(J, rb, rc, rbv, rcv, IR_POW);
-    else
+    } else {
+      ix.mtspec = 1;
       rc = rec_mm_arith(J, &ix, MM_pow);
+    }
     break;
 
   /* -- Miscellaneous ops ------------------------------------------------- */
@@ -2457,6 +2478,7 @@ void lj_record_ins(jit_State *J)
     settabV(J->L, &ix.tabv, tabref(J->fn->l.env));
     ix.tab = emitir(IRT(IR_FLOAD, IRT_TAB), getcurrf(J), IRFL_FUNC_ENV);
     ix.idxchain = LJ_MAX_IDXCHAIN;
+    ix.mtspec = 0;
     rc = lj_record_idx(J, &ix);
     break;
 
@@ -2466,6 +2488,12 @@ void lj_record_ins(jit_State *J)
     /* fallthrough */
   case BC_TGETV: case BC_TGETS: case BC_TSETV: case BC_TSETS:
     ix.idxchain = LJ_MAX_IDXCHAIN;
+    ix.mtspec = 0;
+    rc = lj_record_idx(J, &ix);
+    break;
+  case BC_TGETSS:
+    ix.idxchain = LJ_MAX_IDXCHAIN;
+    ix.mtspec = 1;
     rc = lj_record_idx(J, &ix);
     break;
   case BC_TGETR: case BC_TSETR:

src/lj_record.h

@@ -23,7 +23,8 @@ typedef struct RecordIndex {
   TRef val;		/* Value reference for a store or 0 for a load. */
   TRef mt;		/* Metatable reference. */
   TRef mobj;		/* Metamethod object reference. */
-  int idxchain;		/* Index indirections left or 0 for raw lookup. */
+  short idxchain;	/* Index indirections left or 0 for raw lookup. */
+  short mtspec;	/* Specialize to metatable. */
 } RecordIndex;
 
 LJ_FUNC int lj_record_objcmp(jit_State *J, TRef a, TRef b,

src/vm_arm.dasc

@@ -2592,6 +2592,10 @@ static void build_subroutines(BuildCtx *ctx)
 static void build_ins(BuildCtx *ctx, BCOp op, int defop)
 {
   int vk = 0;
+
+  if (bc_isalias(op))
+    return;
+
   |=>defop:
 
   switch (op) {
@@ -3542,6 +3546,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
     |  b ->vmeta_tgetv
     break;
   case BC_TGETS:
+    |=>BC_TGETSS:
     |  decode_RB8 RB, INS
     |   and RC, RC, #255
     |  // RA = dst*8, RB = table*8, RC = str_const (~)

src/vm_arm64.dasc

@@ -2226,6 +2226,10 @@ static void build_subroutines(BuildCtx *ctx)
 static void build_ins(BuildCtx *ctx, BCOp op, int defop)
 {
   int vk = 0;
+
+  if (bc_isalias(op))
+    return;
+
   |=>defop:
 
   switch (op) {
@@ -2992,6 +2996,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
     |  b ->BC_TGETS_Z
     break;
   case BC_TGETS:
+    |=>BC_TGETSS:
     |  decode_RB RB, INS
     |   and RC, RC, #255
     |  // RA = dst, RB = table, RC = str_const (~)

src/vm_mips.dasc

@@ -2981,6 +2981,10 @@ static void build_subroutines(BuildCtx *ctx)
 static void build_ins(BuildCtx *ctx, BCOp op, int defop)
 {
   int vk = 0;
+
+  if (bc_isalias(op))
+    return;
+
   |=>defop:
 
   switch (op) {
@@ -4076,6 +4080,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
     |.  nop
     break;
   case BC_TGETS:
+    |=>BC_TGETSS:
     |  // RA = dst*8, RB = table*8, RC = str_const*4 (~)
     |  decode_RB8a RB, INS
     |  decode_RB8b RB

src/vm_mips64.dasc

@@ -3104,6 +3104,10 @@ static void build_subroutines(BuildCtx *ctx)
 static void build_ins(BuildCtx *ctx, BCOp op, int defop)
 {
   int vk = 0;
+
+  if (bc_isalias(op))
+    return;
+
   |=>defop:
 
   switch (op) {
@@ -4290,6 +4294,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
     |.  nop
     break;
   case BC_TGETS:
+    |=>BC_TGETSS:
     |  // RA = dst*8, RB = table*8, RC = str_const*8 (~)
     |  decode_RB8a RB, INS
     |  decode_RB8b RB

src/vm_ppc.dasc

@@ -3310,6 +3310,10 @@ static void build_subroutines(BuildCtx *ctx)
 static void build_ins(BuildCtx *ctx, BCOp op, int defop)
 {
   int vk = 0;
+
+  if (bc_isalias(op))
+    return;
+
   |=>defop:
 
   switch (op) {
@@ -4563,6 +4567,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
     |  b ->BC_TGETS_Z			// String key?
     break;
   case BC_TGETS:
+    |=>BC_TGETSS:
     |  // RA = dst*8, RB = table*8, RC = str_const*8 (~)
     |  lwzux CARG1, RB, BASE
     |   srwi TMP1, RC, 1

src/vm_x64.dasc

@@ -2796,6 +2796,10 @@ static void build_subroutines(BuildCtx *ctx)
 static void build_ins(BuildCtx *ctx, BCOp op, int defop)
 {
   int vk = 0;
+
+  if (bc_isalias(op))
+    return;
+
   |// Note: aligning all instructions does not pay off.
   |=>defop:
 
@@ -3678,6 +3682,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
     |  jmp ->BC_TGETS_Z
     break;
   case BC_TGETS:
+    |=>BC_TGETSS:
     |  ins_ABC	// RA = dst, RB = table, RC = str const (~)
     |  mov TAB:RB, [BASE+RB*8]
     |  not RC

src/vm_x86.dasc

@@ -3381,6 +3381,10 @@ static void build_subroutines(BuildCtx *ctx)
 static void build_ins(BuildCtx *ctx, BCOp op, int defop)
 {
   int vk = 0;
+
+  if (bc_isalias(op))
+    return;
+
   |// Note: aligning all instructions does not pay off.
   |=>defop:
 
@@ -4309,6 +4313,7 @@ static void build_ins(BuildCtx *ctx, BCOp op, int defop)
     |  jmp ->BC_TGETS_Z
     break;
   case BC_TGETS:
+    |=>BC_TGETSS:
     |  ins_ABC	// RA = dst, RB = table, RC = str const (~)
     |  not RCa
     |  mov STR:RC, [KBASE+RC*4]