Optimize creation of tuples

Combine the put_tuple/2 and all following put/1 instructions
to one i_put_tuple/2 instruction. In general, that will reduce
the number of instruction words by 50 percent.

Measurements seem to indicate that the speed is about the same.

4 files changed, 149 insertions, 34 deletions

diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index 4fa9d60bca..6c16c24d0d 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -330,6 +330,7 @@ print_op(int to, void *to_arg, int op, int size, BeamInstr* addr)
     BeamInstr packed = 0;		/* Accumulator for packed operations. */
     BeamInstr args[8];		/* Arguments for this instruction. */
     BeamInstr* ap;			/* Pointer to arguments. */
+    BeamInstr* unpacked;		/* Unpacked arguments */
 
     start_prog = opc[op].pack;
 
@@ -551,6 +552,7 @@ print_op(int to, void *to_arg, int op, int size, BeamInstr* addr)
      * Print more information about certain instructions.
      */
 
+    unpacked = ap;
     ap = addr + size;
     switch (op) {
     case op_i_select_val_sfI:
@@ -598,6 +600,32 @@ print_op(int to, void *to_arg, int op, int size, BeamInstr* addr)
 	    }
 	}
 	break;
+    case op_i_put_tuple_rI:
+    case op_i_put_tuple_xI:
+    case op_i_put_tuple_yI:
+	{
+	    int n = unpacked[-1];
+
+	    while (n > 0) {
+		if (!is_header(ap[0])) {
+		    erts_print(to, to_arg, " %T", (Eterm) ap[0]);
+		} else {
+		    switch ((ap[0] >> 2) & 0x03) {
+		    case R_REG_DEF:
+			erts_print(to, to_arg, " x(0)");
+			break;
+		    case X_REG_DEF:
+			erts_print(to, to_arg, " x(%d)", ap[0] >> 4);
+			break;
+		    case Y_REG_DEF:
+			erts_print(to, to_arg, " y(%d)", ap[0] >> 4);
+			break;
+		    }
+		}
+		ap++, size++, n--;
+	    }
+	}
+	break;
     }
     erts_print(to, to_arg, "\n");
 
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 82925fda2c..e36ac1f1e6 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -525,6 +525,11 @@ extern int count_instructions;
      SWAPIN;							\
   } while (0)
 
+#define PutTuple(Dst, Arity)			\
+ do {						\
+   Dst = make_tuple(HTOP);			\
+   pt_arity = (Arity);				\
+ } while (0)
 
 /*
  * Check that we haven't used the reductions and jump to function pointed to by
@@ -732,15 +737,6 @@ extern int count_instructions;
      (Dest) = (* (Eterm *) EXPAND_POINTER(tmp_arg1));		\
   } while (0)
 
-#define PutTuple(Arity, Src, Dest)		\
-     ASSERT(is_arity_value(Arity));		\
-     Dest = make_tuple(HTOP);			\
-     HTOP[0] = (Arity);				\
-     HTOP[1] = (Src);				\
-     HTOP += 2
-
-#define Put(Word) *HTOP++ = (Word)
-
 #define EqualImmed(X, Y, Action) if (X != Y) { Action; }
 
 #define IsFloat(Src, Fail) if (is_not_float(Src)) { Fail; }
@@ -1155,6 +1151,8 @@ void process_main(void)
 
     Uint temp_bits; /* Temporary used by BsSkipBits2 & BsGetInteger2 */
 
+    Eterm pt_arity;		/* Used by do_put_tuple */
+
     ERL_BITS_DECLARE_STATEP; /* Has to be last declaration */
 
 
@@ -1924,6 +1922,32 @@ void process_main(void)
      Goto(*I);
  }
 
+ do_put_tuple: {
+     Eterm* hp = HTOP;
+
+     *hp++ = make_arityval(pt_arity);
+
+     do {
+	 Eterm term = *I++;
+	 switch (term & _TAG_IMMED1_MASK) {
+	 case (R_REG_DEF << _TAG_PRIMARY_SIZE) | TAG_PRIMARY_HEADER:
+	     *hp++ = r(0);
+	     break;
+	 case (X_REG_DEF << _TAG_PRIMARY_SIZE) | TAG_PRIMARY_HEADER:
+	     *hp++ = x(term >> _TAG_IMMED1_SIZE);
+	     break;
+	 case (Y_REG_DEF << _TAG_PRIMARY_SIZE) | TAG_PRIMARY_HEADER:
+	     *hp++ = y(term >> _TAG_IMMED1_SIZE);
+	     break;
+	 default:
+	     *hp++ = term;
+	     break;
+	 }
+     } while (--pt_arity != 0);
+     HTOP = hp;
+     Goto(*I);
+ }
+
     /*
      * All guards with zero arguments have special instructions:
      * 	self/0
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index bc6186751e..a476e439ca 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -2015,6 +2015,30 @@ load_code(LoaderState* stp)
 		stp->labels[tmp_op->a[arg].val].patches = ci;
 		ci++;
 		break;
+	    case TAG_r:
+		CodeNeed(1);
+		code[ci++] = (R_REG_DEF << _TAG_PRIMARY_SIZE) |
+		    TAG_PRIMARY_HEADER;
+		break;
+	    case TAG_x:
+		CodeNeed(1);
+		code[ci++] = (tmp_op->a[arg].val << _TAG_IMMED1_SIZE) |
+		    (X_REG_DEF << _TAG_PRIMARY_SIZE) | TAG_PRIMARY_HEADER;
+		break;
+	    case TAG_y:
+		CodeNeed(1);
+		code[ci++] = (tmp_op->a[arg].val << _TAG_IMMED1_SIZE) |
+		    (Y_REG_DEF << _TAG_PRIMARY_SIZE) | TAG_PRIMARY_HEADER;
+		break;
+	    case TAG_n:
+		CodeNeed(1);
+		code[ci++] = NIL;
+		break;
+	    case TAG_q:
+		CodeNeed(1);
+		new_literal_patch(stp, ci);
+		code[ci++] = tmp_op->a[arg].val;
+		break;
 	    default:
 		LoadError1(stp, "unsupported primitive type '%c'",
 			   tag_to_letter[tmp_op->a[arg].type]);
@@ -3440,6 +3464,56 @@ gen_guard_bif3(LoaderState* stp, GenOpArg Fail, GenOpArg Live, GenOpArg Bif,
     return op;
 }
 
+static GenOp*
+tuple_append_put5(LoaderState* stp, GenOpArg Arity, GenOpArg Dst,
+		  GenOpArg* Puts, GenOpArg S1, GenOpArg S2, GenOpArg S3,
+		  GenOpArg S4, GenOpArg S5)
+{
+    GenOp* op;
+    int arity = Arity.val;	/* Arity of tuple, not the instruction */
+    int i;
+
+    NEW_GENOP(stp, op);
+    op->next = NULL;
+    GENOP_ARITY(op, arity+2+5);
+    op->op = genop_i_put_tuple_2;
+    op->a[0] = Dst;
+    op->a[1].type = TAG_u;
+    op->a[1].val = arity + 5;
+    for (i = 0; i < arity; i++) {
+	op->a[i+2] = Puts[i];
+    }
+    op->a[arity+2] = S1;
+    op->a[arity+3] = S2;
+    op->a[arity+4] = S3;
+    op->a[arity+5] = S4;
+    op->a[arity+6] = S5;
+    return op;
+}
+
+static GenOp*
+tuple_append_put(LoaderState* stp, GenOpArg Arity, GenOpArg Dst,
+		 GenOpArg* Puts, GenOpArg S)
+{
+    GenOp* op;
+    int arity = Arity.val;	/* Arity of tuple, not the instruction */
+    int i;
+
+    NEW_GENOP(stp, op);
+    op->next = NULL;
+    GENOP_ARITY(op, arity+2+1);
+    op->op = genop_i_put_tuple_2;
+    op->a[0] = Dst;
+    op->a[1].type = TAG_u;
+    op->a[1].val = arity + 1;
+    for (i = 0; i < arity; i++) {
+	op->a[i+2] = Puts[i];
+    }
+    op->a[arity+2] = S;
+    return op;
+}
+
+
 
 /*
  * Freeze the code in memory, move the string table into place,
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index 4546f056a9..f629cc80b6 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -345,25 +345,21 @@ i_is_eq_immed f y c
 # Putting things.
 #
 
-put_tuple Arity Dst | put V => i_put_tuple Arity V Dst
-
-%macro: i_put_tuple PutTuple -pack
-i_put_tuple A x x
-i_put_tuple A y x
-i_put_tuple A r x
-i_put_tuple A n x
-i_put_tuple A c x
-i_put_tuple A x y
-i_put_tuple A x r
-i_put_tuple A y r
-i_put_tuple A n r
-i_put_tuple A c r
+put_tuple Arity Dst => i_put_tuple Dst u
 
-%cold
-i_put_tuple A r y
-i_put_tuple A y y
-i_put_tuple A c y
-%hot
+i_put_tuple Dst Arity Puts=* | put S1 | put S2 | \
+  put S3 | put S4 | put S5 => \
+	    tuple_append_put5(Arity, Dst, Puts, S1, S2, S3, S4, S5)
+
+i_put_tuple Dst Arity Puts=* | put S => \
+	    tuple_append_put(Arity, Dst, Puts, S)
+
+i_put_tuple/2
+
+%macro:i_put_tuple PutTuple -pack -goto:do_put_tuple
+i_put_tuple r I
+i_put_tuple x I
+i_put_tuple y I
 
 %macro:put_list PutList -pack -gen_dest
 
@@ -416,13 +412,6 @@ put_list x r r
 put_list s s d
 %hot
 
-%macro: put Put
-put x
-put r
-put y
-put c
-put n
-
 %macro: i_fetch FetchArgs -pack
 i_fetch c c
 i_fetch c r