Merge branch 'bjorn/erts/beam_load'

* bjorn/erts/beam_load:
  Eliminate unnecessary renaming of bs_put_utf16/3
  Don't let the loader do the compiler's job
  Remove unused variables after code generation
  Avoid rebuilding unchanged instructions
  Introduce a 'rename' instruction
  Simplify window management for the transformation engine
  Eliminate allocation of variables in transform_engine()
  Refactor calls to transform_engine()
  ops.tab: Remove useless transformation

4 files changed, 77 insertions, 110 deletions

diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index a390422040..09a41f2b56 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -4102,7 +4102,7 @@ do {						\
      StoreBifResult(1, result);
  }
 
- OpCase(i_bs_put_utf16_jIs): {
+ OpCase(bs_put_utf16_jIs): {
      Eterm arg;
 
      GetArg1(2, arg);
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index 16cbdbffea..a98900460e 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -2028,42 +2028,47 @@ load_code(LoaderState* stp)
 	ASSERT(arity == last_op->arity);
 
     do_transform:
-	if (stp->genop == NULL) {
-	    last_op_next = NULL;
-	    goto get_next_instr;
-	}
-
+	ASSERT(stp->genop != NULL);
 	if (gen_opc[stp->genop->op].transform != -1) {
-	    int need;
-	    tmp_op = stp->genop;
-
-	    for (need = gen_opc[stp->genop->op].min_window-1; need > 0; need--) {
-		if (tmp_op == NULL) {
-		    goto get_next_instr;
-		}
-		tmp_op = tmp_op->next;
+	    if (stp->genop->next == NULL) {
+		/*
+		 * Simple heuristic: Most transformations requires
+		 * at least two instructions, so make sure that
+		 * there are. That will reduce the number of
+		 * TE_SHORT_WINDOWs.
+		 */
+		goto get_next_instr;
 	    }
 	    switch (transform_engine(stp)) {
 	    case TE_FAIL:
-		last_op_next = NULL;
-		last_op = NULL;
+		/*
+		 * No transformation found. stp->genop != NULL and
+		 * last_op_next is still valid. Go ahead and load
+		 * the instruction.
+		 */
 		break;
 	    case TE_OK:
+		/*
+		 * Some transformation was applied. last_op_next is
+		 * no longer valid and stp->genop may be NULL.
+		 * Try to transform again.
+		 */
+		if (stp->genop == NULL) {
+		    last_op_next = &stp->genop;
+		    goto get_next_instr;
+		}
 		last_op_next = NULL;
-		last_op = NULL;
 		goto do_transform;
 	    case TE_SHORT_WINDOW:
-		last_op_next = NULL;
-		last_op = NULL;
+		/*
+		 * No transformation applied. stp->genop != NULL and
+		 * last_op_next is still valid. Fetch a new instruction
+		 * before trying the transformation again.
+		 */
 		goto get_next_instr;
 	    }
 	}
 
-	if (stp->genop == NULL) {
-	    last_op_next = NULL;
-	    goto get_next_instr;
-	}
-
 	/*
 	 * From the collected generic instruction, find the specific
 	 * instruction.
@@ -2584,7 +2589,10 @@ load_code(LoaderState* stp)
 	{
 	    GenOp* next = stp->genop->next;
 	    FREE_GENOP(stp, stp->genop);
-	    stp->genop = next;
+	    if ((stp->genop = next) == NULL) {
+		last_op_next = &stp->genop;
+		goto get_next_instr;
+	    }
 	    goto do_transform;
 	}
     }
@@ -2728,13 +2736,6 @@ mixed_types(LoaderState* stp, GenOpArg Size, GenOpArg* Rest)
 }
 
 static int
-same_label(LoaderState* stp, GenOpArg Target, GenOpArg Label)
-{
-    return Target.type = TAG_f && Label.type == TAG_u &&
-	Target.val == Label.val;
-}
-
-static int
 is_killed_apply(LoaderState* stp, GenOpArg Reg, GenOpArg Live)
 {
     return Reg.type == TAG_x && Live.type == TAG_u &&
@@ -4805,31 +4806,25 @@ transform_engine(LoaderState* st)
     Uint op;
     int ap;			/* Current argument. */
     Uint* restart;		/* Where to restart if current match fails. */
-    GenOpArg def_vars[TE_MAX_VARS]; /* Default buffer for variables. */
-    GenOpArg* var = def_vars;
-    int num_vars = 0;
+    GenOpArg var[TE_MAX_VARS];	/* Buffer for variables. */
+    GenOpArg* rest_args = NULL;
+    int num_rest_args = 0;
     int i;			/* General index. */
     Uint mask;
     GenOp* instr;
+    GenOp* first = st->genop;
+    GenOp* keep = NULL;
     Uint* pc;
-    int rval;
     static Uint restart_fail[1] = {TOP_fail};
 
-    ASSERT(gen_opc[st->genop->op].transform != -1);
-    pc = op_transform + gen_opc[st->genop->op].transform;
-    restart = pc;
+    ASSERT(gen_opc[first->op].transform != -1);
+    restart = op_transform + gen_opc[first->op].transform;
 
  restart:
-    if (var != def_vars) {
-	erts_free(ERTS_ALC_T_LOADER_TMP, (void *) var);
-	var = def_vars;
-    }
     ASSERT(restart != NULL);
     pc = restart;
     ASSERT(*pc < NUM_TOPS);	/* Valid instruction? */
-    instr = st->genop;
-
-#define RETURN(r) rval = (r); goto do_return;
+    instr = first;
 
 #ifdef DEBUG
     restart = NULL;
@@ -4847,7 +4842,7 @@ transform_engine(LoaderState* st)
 		 * We'll need at least one more instruction to decide whether
 		 * this combination matches or not.
 		 */
-		RETURN(TE_SHORT_WINDOW);
+		return TE_SHORT_WINDOW;
 	    }
 	    if (*pc++ != instr->op)
 		goto restart;
@@ -5009,19 +5004,9 @@ transform_engine(LoaderState* st)
 #if defined(TOP_rest_args)
 	case TOP_rest_args:
 	    {
-		int n = *pc++;
 		int formal_arity = gen_opc[instr->op].arity;
-		int j = formal_arity;
-
-		num_vars = n + (instr->arity - formal_arity);
-		var = erts_alloc(ERTS_ALC_T_LOADER_TMP,
-				 num_vars * sizeof(GenOpArg));
-		for (i = 0; i < n; i++) {
-		    var[i] = def_vars[i];
-		}
-		while (i < num_vars) {
-		    var[i++] = instr->a[j++];
-		}
+		num_rest_args = instr->arity - formal_arity;
+		rest_args = instr->a + formal_arity;
 	    }
 	    break;
 #endif
@@ -5030,21 +5015,22 @@ transform_engine(LoaderState* st)
 	    break;
 	case TOP_commit:
 	    instr = instr->next; /* The next_instr was optimized away. */
-
-	    /*
-	     * The left-hand side of this transformation matched.
-	     * Delete all matched instructions.
-	     */
-	    while (st->genop != instr) {
-		GenOp* next = st->genop->next;
-		FREE_GENOP(st, st->genop);
-		st->genop = next;
-	    }
+	    keep = instr;
+	    st->genop = instr;
 #ifdef DEBUG
 	    instr = 0;
 #endif
 	    break;
-
+#if defined(TOP_keep)
+	case TOP_keep:
+	    /* Keep the current instruction unchanged. */
+	    keep = instr;
+	    st->genop = instr;
+#ifdef DEBUG
+	    instr = 0;
+#endif
+	    break;
+#endif
 #if defined(TOP_call_end)
 	case TOP_call_end:
 	    {
@@ -5069,22 +5055,19 @@ transform_engine(LoaderState* st)
 		    lastp = &((*lastp)->next);
 		}
 		 
-		instr = instr->next; /* The next_instr was optimized away. */
-
-		/*
-		 * The left-hand side of this transformation matched.
-		 * Delete all matched instructions.
-		 */
-		while (st->genop != instr) {
-		    GenOp* next = st->genop->next;
-		    FREE_GENOP(st, st->genop);
-		    st->genop = next;
-		}
-		*lastp = st->genop;
+		keep = instr->next; /* The next_instr was optimized away. */
+		*lastp = keep;
 		st->genop = new_instr;
 	    }
-	    RETURN(TE_OK);
+	    /* FALLTHROUGH */
 #endif
+	case TOP_end:
+	    while (first != keep) {
+		GenOp* next = first->next;
+		FREE_GENOP(st, first);
+		first = next;
+	    }
+	    return TE_OK;
 	case TOP_new_instr:
 	    /*
 	     * Note that the instructions are generated in reverse order.
@@ -5096,6 +5079,12 @@ transform_engine(LoaderState* st)
 	    instr->arity = gen_opc[op].arity;
 	    ap = 0;
 	    break;
+#ifdef TOP_rename
+	case TOP_rename:
+	    instr->op = op = *pc++;
+	    instr->arity = gen_opc[op].arity;
+	    return TE_OK;
+#endif
 	case TOP_store_type:
 	    i = *pc++;
 	    instr->a[ap].type = i;
@@ -5115,14 +5104,10 @@ transform_engine(LoaderState* st)
 #if defined(TOP_store_rest_args)
 	case TOP_store_rest_args:
 	    {
-		int n = *pc++;
-		int num_extra = num_vars - n;
-
-		ASSERT(n <= num_vars);
-		GENOP_ARITY(instr, instr->arity+num_extra);
+		GENOP_ARITY(instr, instr->arity+num_rest_args);
 		memcpy(instr->a, instr->def_args, ap*sizeof(GenOpArg));
-		memcpy(instr->a+ap, var+n, num_extra*sizeof(GenOpArg));
-		ap += num_extra;
+		memcpy(instr->a+ap, rest_args, num_rest_args*sizeof(GenOpArg));
+		ap += num_rest_args;
 	    }
 	    break;
 #endif
@@ -5134,21 +5119,12 @@ transform_engine(LoaderState* st)
 	case TOP_try_me_else_fail:
 	    restart = restart_fail;
 	    break;
-	case TOP_end:
-	    RETURN(TE_OK);
 	case TOP_fail:
-	    RETURN(TE_FAIL);
+	    return TE_FAIL;
 	default:
 	    ASSERT(0);
 	}
     }
-#undef RETURN
-
- do_return:
-    if (var != def_vars) {
-	erts_free(ERTS_ALC_T_LOADER_TMP, (void *) var);
-    }
-    return rval;
 }
 
 static void
diff --git a/erts/emulator/beam/beam_load.h b/erts/emulator/beam/beam_load.h
index 22ab71c868..68f4b96893 100644
--- a/erts/emulator/beam/beam_load.h
+++ b/erts/emulator/beam/beam_load.h
@@ -33,7 +33,6 @@ typedef struct gen_op_entry {
    int specific;
    int num_specific;
    int transform;
-   int min_window;
 } GenOpEntry;
 
 extern GenOpEntry gen_opc[];
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index 772460c177..15f27835a8 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -181,11 +181,6 @@ i_jump_on_val_zero y f I
 i_jump_on_val x f I I
 i_jump_on_val y f I I
 
-jump Target | label Lbl | same_label(Target, Lbl) => label Lbl
-
-is_ne_exact L1 S1 S2 | jump Fail | label L2 | same_label(L1, L2) => \
-  is_eq_exact Fail S1 S2 | label L2
-
 %macro: get_list GetList -pack
 get_list x x x
 get_list x x y
@@ -1355,9 +1350,7 @@ bs_put_utf8 Fail u Src=s => i_bs_put_utf8 Fail Src
 
 i_bs_put_utf8 j s
 
-bs_put_utf16 Fail Flags=u Src=s => i_bs_put_utf16 Fail Flags Src
-
-i_bs_put_utf16 j I s
+bs_put_utf16 j I s
 
 bs_put_utf32 Fail=j Flags=u Src=s => \
    i_bs_validate_unicode Fail Src | bs_put_integer Fail i=32 u=1 Flags Src
@@ -1539,7 +1532,6 @@ gen_minus p Live Reg=d Int=i Dst | negation_is_small(Int) => \
 # GCing arithmetic instructions.
 #
 
-gen_plus Fail Live Y=y X=x Dst => i_plus Fail Live X Y Dst
 gen_plus Fail Live S1 S2 Dst => i_plus Fail Live S1 S2 Dst
 
 gen_minus Fail Live S1 S2 Dst => i_minus Fail Live S1 S2 Dst