Merge pull request #2188 from bjorng/bjorn/tune-beam-3

Tune BEAM instructions for the new compiler (part 3)

6 files changed, 379 insertions, 206 deletions

diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index f71efd708f..762c5da9be 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -332,6 +332,13 @@ erts_debug_disassemble_1(BIF_ALIST_1)
 		   "unknown " HEXF "\n", instr);
 	code_ptr++;
     }
+    if (i == op_call_nif) {
+        /*
+         * The rest of the code will not be executed. Don't disassemble any
+         * more code in this function.
+         */
+        code_ptr = 0;
+    }
     bin = new_binary(p, (byte *) dsbufp->str, dsbufp->str_len);
     erts_destroy_tmp_dsbuf(dsbufp);
     hsz = 4+4;
@@ -569,6 +576,7 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
                 }
                 break;
 	    case op_i_bs_match_string_xfWW:
+	    case op_i_bs_match_string_yfWW:
                 if (ap - first_arg < 3) {
                     erts_print(to, to_arg, "%d", *ap);
                 } else {
@@ -772,8 +780,11 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
     case op_put_tuple2_xI:
     case op_put_tuple2_yI:
     case op_new_map_dtI:
-    case op_update_map_assoc_sdtI:
-    case op_update_map_exact_jsdtI:
+    case op_update_map_assoc_xdtI:
+    case op_update_map_assoc_ydtI:
+    case op_update_map_assoc_cdtI:
+    case op_update_map_exact_xjdtI:
+    case op_update_map_exact_yjdtI:
 	{
 	    int n = unpacked[-1];
 
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index 0ad5329b2f..21740caa2c 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -492,6 +492,11 @@ typedef struct LoaderState {
 			   (Genop)->arity * sizeof(GenOpArg)); \
   } while (0)
 
+#define GENOP_NAME_ARITY(Genop, Name, Arity)    \
+  do {                                          \
+    (Genop)->op = genop_##Name##_##Arity;       \
+    (Genop)->arity = Arity;                     \
+  } while (0)
 
 static void free_loader_state(Binary* magic);
 static ErlHeapFragment* new_literal_fragment(Uint size);
@@ -3141,6 +3146,35 @@ is_killed(LoaderState* stp, GenOpArg Reg, GenOpArg Live)
 }
 
 /*
+ * Test whether register Reg is killed by make_fun instruction that
+ * creates the fun given by index idx.
+ */
+
+static int
+is_killed_by_make_fun(LoaderState* stp, GenOpArg Reg, GenOpArg idx)
+{
+    Uint num_free;
+
+    if (idx.val >= stp->num_lambdas) {
+        /* Invalid index. Ignore the error for now. */
+        return 0;
+    } else {
+        num_free = stp->lambdas[idx.val].num_free;
+        return Reg.type == TAG_x && num_free <= Reg.val;
+    }
+}
+
+/*
+ * Test whether register Reg is killed by the send instruction that follows.
+ */
+
+static int
+is_killed_by_send(LoaderState* stp, GenOpArg Reg)
+{
+    return Reg.type == TAG_x && 2 <= Reg.val;
+}
+
+/*
  * Generate an instruction for element/2.
  */
 
@@ -3151,20 +3185,19 @@ gen_element(LoaderState* stp, GenOpArg Fail, GenOpArg Index,
     GenOp* op;
 
     NEW_GENOP(stp, op);
-    op->arity = 4;
     op->next = NULL;
 
     if (Index.type == TAG_i && Index.val > 0 &&
         Index.val <= ERTS_MAX_TUPLE_SIZE &&
 	(Tuple.type == TAG_x || Tuple.type == TAG_y)) {
-	op->op = genop_i_fast_element_4;
+	GENOP_NAME_ARITY(op, i_fast_element, 4);
 	op->a[0] = Tuple;
 	op->a[1] = Fail;
 	op->a[2].type = TAG_u;
 	op->a[2].val = Index.val;
 	op->a[3] = Dst;
     } else {
-	op->op = genop_i_element_4;
+        GENOP_NAME_ARITY(op, i_element, 4);
 	op->a[0] = Tuple;
 	op->a[1] = Fail;
 	op->a[2] = Index;
@@ -3180,8 +3213,7 @@ gen_bs_save(LoaderState* stp, GenOpArg Reg, GenOpArg Index)
     GenOp* op;
 
     NEW_GENOP(stp, op);
-    op->op = genop_i_bs_save2_2;
-    op->arity = 2;
+    GENOP_NAME_ARITY(op, i_bs_save2, 2);
     op->a[0] = Reg;
     op->a[1] = Index;
     if (Index.type == TAG_u) {
@@ -3200,8 +3232,7 @@ gen_bs_restore(LoaderState* stp, GenOpArg Reg, GenOpArg Index)
     GenOp* op;
 
     NEW_GENOP(stp, op);
-    op->op = genop_i_bs_restore2_2;
-    op->arity = 2;
+    GENOP_NAME_ARITY(op, i_bs_restore2, 2);
     op->a[0] = Reg;
     op->a[1] = Index;
     if (Index.type == TAG_u) {
@@ -3235,21 +3266,18 @@ gen_get_integer2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 	} else if ((Flags.val & BSF_SIGNED) != 0) {
 	    goto generic;
 	} else if (bits == 8) {
-	    op->op = genop_i_bs_get_integer_8_3;
-	    op->arity = 3;
-	    op->a[0] = Ms;
+	    GENOP_NAME_ARITY(op, i_bs_get_integer_8, 3);
+            op->a[0] = Ms;
 	    op->a[1] = Fail;
 	    op->a[2] = Dst;
 	} else if (bits == 16 && (Flags.val & BSF_LITTLE) == 0) {
-	    op->op = genop_i_bs_get_integer_16_3;
-	    op->arity = 3;
+	    GENOP_NAME_ARITY(op, i_bs_get_integer_16, 3);
 	    op->a[0] = Ms;
 	    op->a[1] = Fail;
 	    op->a[2] = Dst;
 #ifdef ARCH_64
 	} else if (bits == 32 && (Flags.val & BSF_LITTLE) == 0) {
-	    op->op = genop_i_bs_get_integer_32_3;
-	    op->arity = 3;
+	    GENOP_NAME_ARITY(op, i_bs_get_integer_32, 3);
 	    op->a[0] = Ms;
 	    op->a[1] = Fail;
 	    op->a[2] = Dst;
@@ -3257,8 +3285,7 @@ gen_get_integer2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 	} else {
 	generic:
 	    if (bits < SMALL_BITS) {
-		op->op = genop_i_bs_get_integer_small_imm_5;
-		op->arity = 5;
+		GENOP_NAME_ARITY(op, i_bs_get_integer_small_imm, 5);
 		op->a[0] = Ms;
 		op->a[1].type = TAG_u;
 		op->a[1].val = bits;
@@ -3266,8 +3293,7 @@ gen_get_integer2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 		op->a[3] = Flags;
 		op->a[4] = Dst;
 	    } else {
-		op->op = genop_i_bs_get_integer_imm_6;
-		op->arity = 6;
+		GENOP_NAME_ARITY(op, i_bs_get_integer_imm, 6);
 		op->a[0] = Ms;
 		op->a[1].type = TAG_u;
 		op->a[1].val = bits;
@@ -3283,8 +3309,7 @@ gen_get_integer2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 
 	if (!term_to_Uint(big, &bigval)) {
 	error:
-	    op->op = genop_jump_1;
-	    op->arity = 1;
+	    GENOP_NAME_ARITY(op, jump, 1);
 	    op->a[0] = Fail;
 	} else {
 	    if (!safe_mul(bigval, Unit.val, &bits)) {
@@ -3293,8 +3318,7 @@ gen_get_integer2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 	    goto generic;
 	}
     } else {
-	op->op = genop_i_bs_get_integer_6;
-	op->arity = 6;
+	GENOP_NAME_ARITY(op, i_bs_get_integer, 6);
 	op->a[0] = Ms;
 	op->a[1] = Fail;
 	op->a[2] = Live;
@@ -3324,23 +3348,20 @@ gen_get_binary2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
     NATIVE_ENDIAN(Flags);
     if (Size.type == TAG_a && Size.val == am_all) {
 	if (Ms.type == Dst.type && Ms.val == Dst.val) {
-	    op->op = genop_i_bs_get_binary_all_reuse_3;
-	    op->arity = 3;
+	    GENOP_NAME_ARITY(op, i_bs_get_binary_all_reuse, 3);
 	    op->a[0] = Ms;
 	    op->a[1] = Fail;
 	    op->a[2] = Unit;
 	} else {
-	    op->op = genop_i_bs_get_binary_all2_5;
-	    op->arity = 5;
+	    GENOP_NAME_ARITY(op, i_bs_get_binary_all2, 5);
 	    op->a[0] = Ms;
 	    op->a[1] = Fail;
-	    op->a[2] = Live;	
+	    op->a[2] = Live;
 	    op->a[3] = Unit;
 	    op->a[4] = Dst;
 	}
     } else if (Size.type == TAG_i) {
-	op->op = genop_i_bs_get_binary_imm2_6;
-	op->arity = 6;
+	GENOP_NAME_ARITY(op, i_bs_get_binary_imm2, 6);
 	op->a[0] = Ms;
 	op->a[1] = Fail;
 	op->a[2] = Live;
@@ -3356,12 +3377,10 @@ gen_get_binary2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 
 	if (!term_to_Uint(big, &bigval)) {
 	error:
-	    op->op = genop_jump_1;
-	    op->arity = 1;
+	    GENOP_NAME_ARITY(op, jump, 1);
 	    op->a[0] = Fail;
 	} else {
-	    op->op = genop_i_bs_get_binary_imm2_6;
-	    op->arity = 6;
+	    GENOP_NAME_ARITY(op, i_bs_get_binary_imm2, 6);
 	    op->a[0] = Ms;
 	    op->a[1] = Fail;
 	    op->a[2] = Live;
@@ -3373,8 +3392,7 @@ gen_get_binary2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
 	    op->a[5] = Dst;
 	}
     } else {
-	op->op = genop_i_bs_get_binary2_6;
-	op->arity = 6;
+	GENOP_NAME_ARITY(op, i_bs_get_binary2, 6);
 	op->a[0] = Ms;
 	op->a[1] = Fail;
 	op->a[2] = Live;
@@ -3407,22 +3425,19 @@ gen_put_binary(LoaderState* stp, GenOpArg Fail,GenOpArg Size,
 
     NATIVE_ENDIAN(Flags);
     if (Size.type == TAG_a && Size.val == am_all) {
-	op->op = genop_i_new_bs_put_binary_all_3;
-	op->arity = 3;
+	GENOP_NAME_ARITY(op, i_new_bs_put_binary_all, 3);
 	op->a[0] = Src;
 	op->a[1] = Fail;
 	op->a[2] = Unit;
     } else if (Size.type == TAG_i) {
-	op->op = genop_i_new_bs_put_binary_imm_3;
-	op->arity = 3;
+	GENOP_NAME_ARITY(op, i_new_bs_put_binary_imm, 3);
 	op->a[0] = Fail;
 	op->a[1].type = TAG_u;
 	if (safe_mul(Size.val, Unit.val, &op->a[1].val)) {
 	    op->a[2] = Src;
 	} else {
         error:
-	    op->op = genop_badarg_1;
-	    op->arity = 1;
+            GENOP_NAME_ARITY(op, badarg, 1);
 	    op->a[0] = Fail;
 	}
     } else if (Size.type == TAG_q) {
@@ -3440,16 +3455,14 @@ gen_put_binary(LoaderState* stp, GenOpArg Fail,GenOpArg Size,
             !safe_mul(bigval, Unit.val, &size)) {
 	    goto error;
 	}
-        op->op = genop_i_new_bs_put_binary_imm_3;
-        op->arity = 3;
+        GENOP_NAME_ARITY(op, i_new_bs_put_binary_imm, 3);
         op->a[0] = Fail;
         op->a[1].type = TAG_u;
         op->a[1].val = size;
         op->a[2] = Src;
 #endif
     } else {
-	op->op = genop_i_new_bs_put_binary_4;
-	op->arity = 4;
+	GENOP_NAME_ARITY(op, i_new_bs_put_binary, 4);
 	op->a[0] = Fail;
 	op->a[1] = Size;
 	op->a[2].type = TAG_u;
@@ -3474,14 +3487,12 @@ gen_put_integer(LoaderState* stp, GenOpArg Fail, GenOpArg Size,
         Uint size;
 	if (!safe_mul(Size.val, Unit.val, &size)) {
         error:
-            op->op = genop_badarg_1;
-            op->arity = 1;
+            GENOP_NAME_ARITY(op, badarg, 1);
             op->a[0] = Fail;
             op->next = NULL;
             return op;
 	}
-	op->op = genop_i_new_bs_put_integer_imm_4;
-	op->arity = 4;
+	GENOP_NAME_ARITY(op, i_new_bs_put_integer_imm, 4);
 	op->a[0] = Src;
 	op->a[1] = Fail;
 	op->a[2].type = TAG_u;
@@ -3497,8 +3508,7 @@ gen_put_integer(LoaderState* stp, GenOpArg Fail, GenOpArg Size,
             !safe_mul(bigval, Unit.val, &size)) {
 	    goto error;
 	}
-	op->op = genop_i_new_bs_put_integer_imm_4;
-	op->arity = 4;
+	GENOP_NAME_ARITY(op, i_new_bs_put_integer_imm, 4);
 	op->a[0] = Src;
 	op->a[1] = Fail;
 	op->a[2].type = TAG_u;
@@ -3506,8 +3516,7 @@ gen_put_integer(LoaderState* stp, GenOpArg Fail, GenOpArg Size,
 	op->a[3].type = Flags.type;
 	op->a[3].val = (Flags.val & 7);
     } else {
-	op->op = genop_i_new_bs_put_integer_4;
-	op->arity = 4;
+	GENOP_NAME_ARITY(op, i_new_bs_put_integer, 4);
 	op->a[0] = Fail;
 	op->a[1] = Size;
 	op->a[2].type = TAG_u;
@@ -3527,21 +3536,18 @@ gen_put_float(LoaderState* stp, GenOpArg Fail, GenOpArg Size,
 
     NATIVE_ENDIAN(Flags);
     if (Size.type == TAG_i) {
-	op->op = genop_i_new_bs_put_float_imm_4;
-	op->arity = 4;
+	GENOP_NAME_ARITY(op, i_new_bs_put_float_imm, 4);
 	op->a[0] = Fail;
 	op->a[1].type = TAG_u;
 	if (!safe_mul(Size.val, Unit.val, &op->a[1].val)) {
-	    op->op = genop_badarg_1;
-	    op->arity = 1;
+	    GENOP_NAME_ARITY(op, badarg, 1);
 	    op->a[0] = Fail;
 	} else {
 	    op->a[2] = Flags;
 	    op->a[3] = Src;
 	}
     } else {
-	op->op = genop_i_new_bs_put_float_4;
-	op->arity = 4;
+	GENOP_NAME_ARITY(op, i_new_bs_put_float, 4);
 	op->a[0] = Fail;
 	op->a[1] = Size;
 	op->a[2].type = TAG_u;
@@ -3564,8 +3570,7 @@ gen_get_float2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
     NEW_GENOP(stp, op);
 
     NATIVE_ENDIAN(Flags);
-    op->op = genop_i_bs_get_float2_6;
-    op->arity = 6;
+    GENOP_NAME_ARITY(op, i_bs_get_float2, 6);
     op->a[0] = Ms;
     op->a[1] = Fail;
     op->a[2] = Live;
@@ -3582,7 +3587,7 @@ gen_get_float2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, GenOpArg Live,
  */
 
 static GenOp*
-gen_skip_bits2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms, 
+gen_skip_bits2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms,
 	       GenOpArg Size, GenOpArg Unit, GenOpArg Flags)
 {
     GenOp* op;
@@ -3602,16 +3607,14 @@ gen_skip_bits2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms,
          * a bs_restore2 instruction which will overwrite the position
          * by one of the stored positions.
          */
-	op->op = genop_bs_test_unit_3;
-	op->arity = 3;
+	GENOP_NAME_ARITY(op, bs_test_unit, 3);
 	op->a[0] = Fail;
 	op->a[1] = Ms;
 	op->a[2] = Unit;
     } else if (Size.type == TAG_i) {
-	op->op = genop_i_bs_skip_bits_imm2_3;
-	op->arity = 3;
+	GENOP_NAME_ARITY(op, i_bs_skip_bits_imm2, 3);
 	op->a[0] = Fail;
-	op->a[1] = Ms; 
+	op->a[1] = Ms;
 	op->a[2].type = TAG_u;
 	if (!safe_mul(Size.val, Unit.val, &op->a[2].val)) {
 	    goto error;
@@ -3622,22 +3625,19 @@ gen_skip_bits2(LoaderState* stp, GenOpArg Fail, GenOpArg Ms,
 
 	if (!term_to_Uint(big, &bigval)) {
 	error:
-	    op->op = genop_jump_1;
-	    op->arity = 1;
+	    GENOP_NAME_ARITY(op, jump, 1);
 	    op->a[0] = Fail;
 	} else {
-	    op->op = genop_i_bs_skip_bits_imm2_3;
-	    op->arity = 3;
+	    GENOP_NAME_ARITY(op, i_bs_skip_bits_imm2, 3);
 	    op->a[0] = Fail;
-	    op->a[1] = Ms; 
+	    op->a[1] = Ms;
 	    op->a[2].type = TAG_u;
 	    if (!safe_mul(bigval, Unit.val, &op->a[2].val)) {
 		goto error;
 	    }
 	}
     } else {
-	op->op = genop_i_bs_skip_bits2_4;
-	op->arity = 4;
+	GENOP_NAME_ARITY(op, i_bs_skip_bits2, 4);
 	op->a[0] = Ms;
 	op->a[1] = Size;
 	op->a[2] = Fail;
@@ -3654,8 +3654,7 @@ gen_increment(LoaderState* stp, GenOpArg Reg,
     GenOp* op;
 
     NEW_GENOP(stp, op);
-    op->op = genop_i_increment_3;
-    op->arity = 3;
+    GENOP_NAME_ARITY(op, i_increment, 3);
     op->next = NULL;
     op->a[0] = Reg;
     op->a[1].type = TAG_u;
@@ -3671,8 +3670,7 @@ gen_increment_from_minus(LoaderState* stp, GenOpArg Reg,
     GenOp* op;
 
     NEW_GENOP(stp, op);
-    op->op = genop_i_increment_3;
-    op->arity = 3;
+    GENOP_NAME_ARITY(op, i_increment, 3);
     op->next = NULL;
     op->a[0] = Reg;
     op->a[1].type = TAG_u;
@@ -3681,6 +3679,24 @@ gen_increment_from_minus(LoaderState* stp, GenOpArg Reg,
     return op;
 }
 
+static GenOp*
+gen_plus_from_minus(LoaderState* stp, GenOpArg Fail, GenOpArg Live,
+                    GenOpArg Src, GenOpArg Integer, GenOpArg Dst)
+{
+    GenOp* op;
+
+    NEW_GENOP(stp, op);
+    GENOP_NAME_ARITY(op, gen_plus, 5);
+    op->next = NULL;
+    op->a[0] = Fail;
+    op->a[1] = Live;
+    op->a[2] = Src;
+    op->a[3].type = TAG_i;
+    op->a[3].val = -Integer.val;
+    op->a[4] = Dst;
+    return op;
+}
+
 /*
  * Test whether the negation of the given number is small.
  */
@@ -3727,12 +3743,11 @@ gen_literal_timeout(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
     Sint timeout;
 
     NEW_GENOP(stp, op);
-    op->op = genop_wait_timeout_unlocked_int_2;
+    GENOP_NAME_ARITY(op, wait_timeout_unlocked_int, 2);
     op->next = NULL;
-    op->arity = 2;
     op->a[0].type = TAG_u;
     op->a[1] = Fail;
-    
+
     if (Time.type == TAG_i && (timeout = Time.val) >= 0 &&
 #if defined(ARCH_64)
 	(timeout >> 32) == 0
@@ -3761,8 +3776,7 @@ gen_literal_timeout(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
 #if !defined(ARCH_64)
     error:
 #endif
-	op->op = genop_i_wait_error_0;
-	op->arity = 0;
+	GENOP_NAME_ARITY(op, i_wait_error, 0);
     }
     return op;
 }
@@ -3774,9 +3788,8 @@ gen_literal_timeout_locked(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
     Sint timeout;
 
     NEW_GENOP(stp, op);
-    op->op = genop_wait_timeout_locked_int_2;
+    GENOP_NAME_ARITY(op, wait_timeout_locked_int, 2);
     op->next = NULL;
-    op->arity = 2;
     op->a[0].type = TAG_u;
     op->a[1] = Fail;
 
@@ -3808,8 +3821,7 @@ gen_literal_timeout_locked(LoaderState* stp, GenOpArg Fail, GenOpArg Time)
 #if !defined(ARCH_64)
     error:
 #endif
-	op->op = genop_i_wait_error_locked_0;
-	op->arity = 0;
+        GENOP_NAME_ARITY(op, i_wait_error_locked, 0);
     }
     return op;
 }
@@ -3846,9 +3858,9 @@ gen_select_tuple_arity(LoaderState* stp, GenOpArg S, GenOpArg Fail,
      */
     if (size == 2) {
 	NEW_GENOP(stp, op);
-	op->next = NULL;
-	op->op = genop_i_select_tuple_arity2_4;
+	GENOP_NAME_ARITY(op, i_select_tuple_arity2, 4);
 	GENOP_ARITY(op, arity - 1);
+	op->next = NULL;
 	op->a[0] = S;
 	op->a[1] = Fail;
 	op->a[2].type = TAG_u;
@@ -3874,9 +3886,9 @@ gen_select_tuple_arity(LoaderState* stp, GenOpArg S, GenOpArg Fail,
     size  += align;
 
     NEW_GENOP(stp, op);
-    op->next = NULL;
-    op->op = genop_i_select_tuple_arity_3;
+    GENOP_NAME_ARITY(op, i_select_tuple_arity, 3);
     GENOP_ARITY(op, arity);
+    op->next = NULL;
     op->a[0] = S;
     op->a[1] = Fail;
     op->a[2].type = TAG_u;
@@ -3932,21 +3944,18 @@ gen_split_values(LoaderState* stp, GenOpArg S, GenOpArg TypeFail,
     ASSERT(Size.val >= 2 && Size.val % 2 == 0);
 
     NEW_GENOP(stp, is_integer);
-    is_integer->op = genop_is_integer_2;
-    is_integer->arity = 2;
+    GENOP_NAME_ARITY(is_integer, is_integer, 2);
     is_integer->a[0] = TypeFail;
     is_integer->a[1] = S;
 
     NEW_GENOP(stp, label);
-    label->op = genop_label_1;
-    label->arity = 1;
+    GENOP_NAME_ARITY(label, label, 1);
     label->a[0].type = TAG_u;
     label->a[0].val = new_label(stp);
 
     NEW_GENOP(stp, op1);
-    op1->op = genop_select_val_3;
+    GENOP_NAME_ARITY(op1, select_val, 3);
     GENOP_ARITY(op1, 3 + Size.val);
-    op1->arity = 3;
     op1->a[0] = S;
     op1->a[1].type = TAG_f;
     op1->a[1].val = label->a[0].val;
@@ -3954,9 +3963,8 @@ gen_split_values(LoaderState* stp, GenOpArg S, GenOpArg TypeFail,
     op1->a[2].val = 0;
 
     NEW_GENOP(stp, op2);
-    op2->op = genop_select_val_3;
+    GENOP_NAME_ARITY(op2, select_val, 3);
     GENOP_ARITY(op2, 3 + Size.val);
-    op2->arity = 3;
     op2->a[0] = S;
     op2->a[1] = Fail;
     op2->a[2].type = TAG_u;
@@ -4034,19 +4042,17 @@ gen_jump_tab(LoaderState* stp, GenOpArg S, GenOpArg Fail, GenOpArg Size, GenOpAr
 	GenOp* jump;
 
 	NEW_GENOP(stp, op);
-	op->arity = 3;
-	op->op = genop_is_ne_exact_3;
+	GENOP_NAME_ARITY(op, is_ne_exact, 3);
 	op->a[0] = Rest[1];
 	op->a[1] = S;
 	op->a[2] = Rest[0];
 
 	NEW_GENOP(stp, jump);
-	jump->next = NULL;
-	jump->arity = 1;
-	jump->op = genop_jump_1;
+        GENOP_NAME_ARITY(jump, jump, 1);
 	jump->a[0] = Fail;
 
 	op->next = jump;
+	jump->next = NULL;
 	return op;
     }
 
@@ -4073,12 +4079,11 @@ gen_jump_tab(LoaderState* stp, GenOpArg S, GenOpArg Fail, GenOpArg Size, GenOpAr
     NEW_GENOP(stp, op);
     op->next = NULL;
     if (min == 0) {
-	op->op = genop_i_jump_on_val_zero_3;
-	fixed_args = 3;
+	GENOP_NAME_ARITY(op, i_jump_on_val_zero, 3);
     } else {
-	op->op = genop_i_jump_on_val_4;
-	fixed_args = 4;
+	GENOP_NAME_ARITY(op, i_jump_on_val, 4);
     }
+    fixed_args = op->arity;
     arity = fixed_args + size;
     GENOP_ARITY(op, arity);
     op->a[0] = S;
@@ -4143,7 +4148,7 @@ gen_select_val(LoaderState* stp, GenOpArg S, GenOpArg Fail,
 
 	NEW_GENOP(stp, op);
 	op->next = NULL;
-	op->op = genop_i_select_val2_4;
+        GENOP_NAME_ARITY(op, i_select_val2, 4);
 	GENOP_ARITY(op, arity - 1);
 	op->a[0] = S;
 	op->a[1] = Fail;
@@ -4165,7 +4170,11 @@ gen_select_val(LoaderState* stp, GenOpArg S, GenOpArg Fail,
 
     NEW_GENOP(stp, op);
     op->next = NULL;
-    op->op = (align == 0) ? genop_i_select_val_bins_3 : genop_i_select_val_lins_3;
+    if (align == 0) {
+        GENOP_NAME_ARITY(op, i_select_val_bins, 3);
+    } else {
+        GENOP_NAME_ARITY(op, i_select_val_lins, 3);
+    }
     GENOP_ARITY(op, arity);
     op->a[0] = S;
     op->a[1] = Fail;
@@ -4229,8 +4238,7 @@ gen_select_literals(LoaderState* stp, GenOpArg S, GenOpArg Fail,
 	ASSERT(Rest[i].type == TAG_q);
 
 	NEW_GENOP(stp, op);
-	op->op = genop_is_ne_exact_3;
-	op->arity = 3;
+	GENOP_NAME_ARITY(op, is_ne_exact, 3);
 	op->a[0] = Rest[i+1];
 	op->a[1] = S;
 	op->a[2] = Rest[i];
@@ -4239,9 +4247,8 @@ gen_select_literals(LoaderState* stp, GenOpArg S, GenOpArg Fail,
     }
 
     NEW_GENOP(stp, jump);
+    GENOP_NAME_ARITY(jump, jump, 1);
     jump->next = NULL;
-    jump->op = genop_jump_1;
-    jump->arity = 1;
     jump->a[0] = Fail;
     *prev_next = jump;
     return op;
@@ -4263,9 +4270,8 @@ const_select_val(LoaderState* stp, GenOpArg S, GenOpArg Fail,
     ASSERT(Size.type == TAG_u);
 
     NEW_GENOP(stp, op);
+    GENOP_NAME_ARITY(op, jump, 1);
     op->next = NULL;
-    op->op = genop_jump_1;
-    op->arity = 1;
 
     /*
      * Search for a literal matching the controlling expression.
@@ -4346,15 +4352,17 @@ gen_make_fun2(LoaderState* stp, GenOpArg idx)
         funp->creator = erts_init_process_id;
         funp->arity = arity;
 
-        op->op = genop_move_2;
-        op->arity = 2;
+        /*
+         * Use a move_fun/2 instruction to load the fun to enable
+         * further optimizations.
+         */
+        GENOP_NAME_ARITY(op, move_fun, 2);
         op->a[0].type = TAG_q;
         op->a[0].val = lit;
         op->a[1].type = TAG_x;
         op->a[1].val = 0;
     } else {
-        op->op = genop_i_make_fun_2;
-        op->arity = 2;
+        GENOP_NAME_ARITY(op, i_make_fun, 2);
         op->a[0].type = TAG_u;
         op->a[0].val = (BeamInstr) fe;
         op->a[1].type = TAG_u;
@@ -4382,13 +4390,11 @@ gen_is_function2(LoaderState* stp, GenOpArg Fail, GenOpArg Fun, GenOpArg Arity)
          */
         if (Arity.val > MAX_ARG) {
             /* Arity is negative or too big. */
-            op->op = genop_jump_1;
-            op->arity = 1;
+            GENOP_NAME_ARITY(op, jump, 1);
             op->a[0] = Fail;
             return op;
         } else {
-            op->op = genop_hot_is_function2_3;
-            op->arity = 3;
+            GENOP_NAME_ARITY(op, hot_is_function2, 3);
             op->a[0] = Fail;
             op->a[1] = Fun;
             op->a[2].type = TAG_u;
@@ -4408,20 +4414,17 @@ gen_is_function2(LoaderState* stp, GenOpArg Fail, GenOpArg Fun, GenOpArg Arity)
         move_fun->next = move_arity;
         move_arity->next = op;
 
-        move_fun->arity = 2;
-        move_fun->op = genop_move_2;
+        GENOP_NAME_ARITY(move_fun, move, 2);
         move_fun->a[0] = Fun;
         move_fun->a[1].type = TAG_x;
         move_fun->a[1].val = 1022;
 
-        move_arity->arity = 2;
-        move_arity->op = genop_move_2;
+        GENOP_NAME_ARITY(move_arity, move, 2);
         move_arity->a[0] = Arity;
         move_arity->a[1].type = TAG_x;
         move_arity->a[1].val = 1023;
 
-        op->op = genop_cold_is_function2_3;
-        op->arity = 3;
+        GENOP_NAME_ARITY(op, cold_is_function2, 3);
         op->a[0] = Fail;
         op->a[1].type = TAG_x;
         op->a[1].val = 1022;
@@ -4442,8 +4445,8 @@ tuple_append_put5(LoaderState* stp, GenOpArg Arity, GenOpArg Dst,
 
     NEW_GENOP(stp, op);
     op->next = NULL;
+    GENOP_NAME_ARITY(op, i_put_tuple, 2);
     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;
@@ -4468,8 +4471,8 @@ tuple_append_put(LoaderState* stp, GenOpArg Arity, GenOpArg Dst,
 
     NEW_GENOP(stp, op);
     op->next = NULL;
+    GENOP_NAME_ARITY(op, i_put_tuple, 2);
     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;
@@ -4537,9 +4540,9 @@ gen_new_small_map_lit(LoaderState* stp, GenOpArg Dst, GenOpArg Live,
     Eterm keys;
 
     NEW_GENOP(stp, op);
+    GENOP_NAME_ARITY(op, i_new_small_map_lit, 3);
     GENOP_ARITY(op, 3 + size/2);
     op->next = NULL;
-    op->op = genop_i_new_small_map_lit_3;
 
     tmp = thp = erts_alloc(ERTS_ALC_T_LOADER_TMP, (1 + size/2) * sizeof(*tmp));
     keys = make_tuple(thp);
@@ -4720,14 +4723,12 @@ gen_get_map_element(LoaderState* stp, GenOpArg Fail, GenOpArg Src,
 
     Key = Rest[0];
     if (hash_genop_arg(stp, Key, &hx)) {
-	op->arity = 5;
-	op->op = genop_i_get_map_element_hash_5;
+	GENOP_NAME_ARITY(op, i_get_map_element_hash, 5);
 	op->a[3].type = TAG_u;
 	op->a[3].val = (BeamInstr) hx;
 	op->a[4] = Rest[1];
     } else {
-	op->arity = 4;
-	op->op = genop_i_get_map_element_4;
+	GENOP_NAME_ARITY(op, i_get_map_element, 4);
 	op->a[3] = Rest[1];
     }
     return op;
@@ -4767,15 +4768,13 @@ gen_get(LoaderState* stp, GenOpArg Src, GenOpArg Dst)
     NEW_GENOP(stp, op);
     op->next = NULL;
     if (hash_internal_genop_arg(stp, Src, &hx)) {
-	op->arity = 3;
-	op->op = genop_i_get_hash_3;
+	GENOP_NAME_ARITY(op, i_get_hash, 3);
         op->a[0] = Src;
 	op->a[1].type = TAG_u;
 	op->a[1].val = (BeamInstr) hx;
 	op->a[2] = Dst;
     } else {
-	op->arity = 2;
-	op->op = genop_i_get_2;
+	GENOP_NAME_ARITY(op, i_get, 2);
         op->a[0] = Src;
 	op->a[1] = Dst;
     }
@@ -4799,7 +4798,7 @@ gen_get_map_elements(LoaderState* stp, GenOpArg Fail, GenOpArg Src,
     ASSERT(Size.type == TAG_u);
 
     NEW_GENOP(stp, op);
-    op->op = genop_i_get_map_elements_3;
+    GENOP_NAME_ARITY(op, i_get_map_elements, 3);
     GENOP_ARITY(op, 3 + 3*(Size.val/2));
     op->next = NULL;
     op->a[0] = Fail;
@@ -4837,9 +4836,9 @@ gen_has_map_fields(LoaderState* stp, GenOpArg Fail, GenOpArg Src,
     n = Size.val;
 
     NEW_GENOP(stp, op);
+    GENOP_NAME_ARITY(op, get_map_elements, 3);
     GENOP_ARITY(op, 3 + 2*n);
     op->next = NULL;
-    op->op = genop_get_map_elements_3;
 
     op->a[0] = Fail;
     op->a[1] = Src;
diff --git a/erts/emulator/beam/instrs.tab b/erts/emulator/beam/instrs.tab
index 1eb83b61f2..692408e212 100644
--- a/erts/emulator/beam/instrs.tab
+++ b/erts/emulator/beam/instrs.tab
@@ -477,6 +477,13 @@ i_make_fun(FunP, NumFree) {
     HEAVY_SWAPIN;
 }
 
+move_trim(Src, Dst, Words) {
+    Uint cp = E[0];
+    $Dst = $Src;
+    E += $Words;
+    E[0] = cp;
+}
+
 i_trim(Words) {
     Uint cp = E[0];
     E += $Words;
@@ -493,10 +500,6 @@ move3(S1, D1, S2, D2, S3, D3) {
     $D3 = $S3;
 }
 
-move_dup(Src, D1, D2) {
-    $D1 = $D2 = $Src;
-}
-
 move2_par(S1, D1, S2, D2) {
     Eterm V1, V2;
     V1 = $S1;
@@ -512,6 +515,39 @@ move_shift(Src, SD, D) {
     $SD = V;
 }
 
+move_src_window2(Src, D1, D2) {
+    Eterm* src = &$Src;
+    Eterm s1, s2;
+    s1 = src[0];
+    s2 = src[1];
+    $D1 = s1;
+    $D2 = s2;
+}
+
+move_src_window3(Src, D1, D2, D3) {
+    Eterm* src = &$Src;
+    Eterm s1, s2, s3;
+    s1 = src[0];
+    s2 = src[1];
+    s3 = src[2];
+    $D1 = s1;
+    $D2 = s2;
+    $D3 = s3;
+}
+
+move_src_window4(Src, D1, D2, D3, D4) {
+    Eterm* src = &$Src;
+    Eterm s1, s2, s3, s4;
+    s1 = src[0];
+    s2 = src[1];
+    s3 = src[2];
+    s4 = src[3];
+    $D1 = s1;
+    $D2 = s2;
+    $D3 = s3;
+    $D4 = s4;
+}
+
 move_window2(S1, S2, D) {
     Eterm xt0, xt1;
     Eterm* y = &$D;
@@ -703,9 +739,11 @@ jump(Fail) {
     $JUMP($Fail);
 }
 
-move_jump(Fail, Src) {
-    x(0) = $Src;
-    $jump($Fail);
+move_jump(Lbl, Src, Dst) {
+    Eterm lbl = $Lbl;
+    Eterm src = $Src;
+    $Dst = src;
+    $JUMP(lbl);
 }
 
 //
@@ -817,6 +855,16 @@ is_tagged_tuple(Fail, Src, Arityval, Tag) {
     }
 }
 
+is_tagged_tuple_ff(NotTupleFail, WrongRecordFail, Src, Arityval, Tag) {
+    Eterm term = $Src;
+    if (is_not_tuple(term)) {
+        $FAIL($NotTupleFail);
+    } else if (ERTS_UNLIKELY((tuple_val(term))[0] != $Arityval ||
+                             (tuple_val(term))[1] != $Tag)) {
+        $FAIL($WrongRecordFail);
+    }
+}
+
 is_tuple(Fail, Src) {
     if (is_not_tuple($Src)) {
         $FAIL($Fail);
diff --git a/erts/emulator/beam/map_instrs.tab b/erts/emulator/beam/map_instrs.tab
index c594a87298..5620d5a2d1 100644
--- a/erts/emulator/beam/map_instrs.tab
+++ b/erts/emulator/beam/map_instrs.tab
@@ -127,11 +127,21 @@ i_get_map_elements(Fail, Src, N) {
     }
 }
 
-update_map_assoc(Src, Dst, Live, N) {
+update_map_assoc := update_map_assoc.fetch.execute;
+
+update_map_assoc.head() {
+    Eterm map;
+}
+
+update_map_assoc.fetch(Src) {
+    map = $Src;
+}
+
+update_map_assoc.execute(Dst, Live, N) {
     Eterm res;
     Uint live = $Live;
 
-    reg[live] = $Src;
+    reg[live] = map;
     HEAVY_SWAPOUT;
     res = erts_gc_update_map_assoc(c_p, reg, live, $N, $NEXT_INSTRUCTION);
     HEAVY_SWAPIN;
@@ -141,11 +151,21 @@ update_map_assoc(Src, Dst, Live, N) {
     $NEXT($NEXT_INSTRUCTION+$N);
 }
 
-update_map_exact(Fail, Src, Dst, Live, N) {
+update_map_exact := update_map_exact.fetch.execute;
+
+update_map_exact.head() {
+    Eterm map;
+}
+
+update_map_exact.fetch(Src) {
+    map = $Src;
+}
+
+update_map_exact.execute(Fail, Dst, Live, N) {
     Eterm res;
     Uint live = $Live;
 
-    reg[live] = $Src;
+    reg[live] = map;
     HEAVY_SWAPOUT;
     res = erts_gc_update_map_exact(c_p, reg, live, $N, $NEXT_INSTRUCTION);
     HEAVY_SWAPIN;
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index da5364183c..6832e65b1b 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -106,7 +106,10 @@ init y
 allocate_zero t t?
 allocate_heap_zero t I t?
 
+move Src=y Dst=x | trim N Remaining => move_trim Src Dst N
 trim N Remaining => i_trim N
+
+move_trim y x t
 i_trim t
 
 test_heap I t?
@@ -261,12 +264,14 @@ system_limit j
 # Move instructions.
 #
 
-move C=cxy x==0 | jump Lbl => move_jump Lbl C
+move Src=cxy Dst=xy | jump Lbl => move_jump Lbl Src Dst
+
+move_jump f cxy xy
+move_jump f c r
 
-move_jump f ncxy
 
-# Movement to and from the stack is common
-# Try to pack as much as we can into one instruction
+# Movement to and from the stack is common.
+# Try to pack as much as we can into one instruction.
 
 # Window move
 move_window/5
@@ -294,11 +299,49 @@ move_window3 x x x y
 move_window4 x x x x y
 move_window5 x x x x x y
 
+# y -> x
+
+move_src_window/4
+move_src_window/5
+
+move Y1=y X1=x | move Y2=y X2=x | succ(Y1, Y2) => \
+    move_src_window Y1 Y2 X1 X2
+
+move_src_window Y1 Y2 X1 X2 | move Y3=y X3=x | succ(Y2, Y3) => \
+    move_src_window Y1 Y3 X1 X2 X3
+move_src_window Y1 Y2 X1 X2 | move Y3=y X3=x | move Y4=y X4=x | succ(Y3, Y4) => \
+    move_src_window2 Y1 X1 X2 | move_src_window Y3 Y4 X3 X4
+move_src_window Y1 Y2 X1 X2 | move Y3=y X3=x => \
+    move3 Y1 X1 Y2 X2 Y3 X3
+
+move_src_window Y1 Y3 X1 X2 X3 | move Y4=y X4=x | succ(Y3, Y4) => \
+    move_src_window4 Y1 X1 X2 X3 X4
+
+move_src_window Y1 y X1 X2    => move_src_window2 Y1 X1 X2
+move_src_window Y1 y X1 X2 X3 => move_src_window3 Y1 X1 X2 X3
+
+move_src_window2 y x x
+move_src_window3 y x x x
+move_src_window4 y x x x x
+
 # Swap registers.
-move R1=x Tmp=x | move R2=x R1 | move Tmp R2 => swap_temp R1 R2 Tmp
+move R1=xy Tmp=x | move R2=xy R1 | move Tmp R2 => swap_temp R1 R2 Tmp
+
+# The compiler uses x(1022) when swapping registers. It will definitely
+# not be used again.
+swap_temp R1 R2 Tmp=x==1022 => swap R1 R2
+
+swap_temp R1 R2 Tmp | move Src Tmp => swap R1 R2 | move Src Tmp
 
 swap_temp R1 R2 Tmp | line Loc | apply Live | is_killed_apply(Tmp, Live) => \
   swap R1 R2 | line Loc | apply Live
+swap_temp R1 R2 Tmp | line Loc | apply_last Live D | is_killed_apply(Tmp, Live) => \
+  swap R1 R2 | line Loc | apply_last Live D
+
+swap_temp R1 R2 Tmp | line Loc | call_fun Live | is_killed(Tmp, Live) => \
+  swap R1 R2 | line Loc | call_fun Live
+swap_temp R1 R2 Tmp | make_fun2 OldIndex=u | is_killed_by_make_fun(Tmp, OldIndex) => \
+  swap R1 R2 | make_fun2 OldIndex
 
 swap_temp R1 R2 Tmp | line Loc | call Live Addr | is_killed(Tmp, Live) => \
   swap R1 R2 | line Loc | call Live Addr
@@ -314,27 +357,55 @@ swap_temp R1 R2 Tmp | line Loc | call_ext_only Live Addr | \
 swap_temp R1 R2 Tmp | line Loc | call_ext_last Live Addr D | \
   is_killed(Tmp, Live) => swap R1 R2 | line Loc | call_ext_last Live Addr D
 
-swap_temp x x x
-
-swap x x
+swap_temp R1 R2 Tmp | call_ext Live Addr | is_killed(Tmp, Live) => \
+  swap R1 R2 | call_ext Live Addr
+swap_temp R1 R2 Tmp | call_ext_only Live Addr | is_killed(Tmp, Live) => \
+  swap R1 R2 | call_ext_only Live Addr
+swap_temp R1 R2 Tmp | call_ext_last Live Addr D | is_killed(Tmp, Live) => \
+  swap R1 R2 | call_ext_last Live Addr D
+
+swap_temp R1 R2 Tmp | move Src Any | line Loc | call Live Addr | \
+  is_killed(Tmp, Live) | distinct(Tmp, Src) => \
+     swap R1 R2 | move Src Any | line Loc | call Live Addr
+swap_temp R1 R2 Tmp | move Src Any | line Loc | call_ext Live Addr | \
+  is_killed(Tmp, Live) | distinct(Tmp, Src) => \
+     swap R1 R2 | move Src Any | line Loc | call_ext Live Addr
+swap_temp R1 R2 Tmp | move Src Any | call_only Live Addr | \
+  is_killed(Tmp, Live) | distinct(Tmp, Src) => \
+    swap R1 R2 | move Src Any | call_only Live Addr
+swap_temp R1 R2 Tmp | move Src Any | line Loc | call_ext_only Live Addr | \
+  is_killed(Tmp, Live) | distinct(Tmp, Src) => \
+    swap R1 R2 | move Src Any | line Loc | call_ext_only Live Addr
+swap_temp R1 R2 Tmp | move Src Any | line Loc | call_fun Live | \
+  is_killed(Tmp, Live) | distinct(Tmp, Src) => \
+    swap R1 R2 | move Src Any | line Loc | call_fun Live
+
+swap_temp R1 R2 Tmp | line Loc | send | is_killed_by_send(Tmp) => \
+  swap R1 R2 | line Loc | send
+
+# swap_temp/3 with Y register operands are rare.
+swap_temp R1 R2=y Tmp => swap R1 R2 | move R2 Tmp
+swap_temp R1=y R2 Tmp => swap R1 R2 | move R2 Tmp
+
+swap R1=x R2=y => swap R2 R1
 
-# move_dup
-
-move Src=x D1=x | move Src=x D2=x => move_dup Src D1 D2
-move Src=x SD=x | move SD=x  D=x  => move_dup Src SD D
+swap_temp x x x
 
-move_dup x x x
+swap xy x
+swap y y
 
 # move_shift
 
-move SD=x D=x | move Src=xy SD=x | distinct(D, Src) => move_shift Src SD D
-move SD=y D=x | move Src=x  SD=y | distinct(D, Src) => move_shift Src SD D
-move SD=x D=y | move Src=x  SD=x | distinct(D, Src) => move_shift Src SD D
+move SD=x    D=x | move Src=cxy SD=x  | distinct(D, Src) => move_shift Src SD D
+move SD=y    D=x | move Src=x  SD=y   | distinct(D, Src) => move_shift Src SD D
+move SD=y    D=x | init SD            |                  => move_shift n   SD D
+move SD=x    D=y | move Src=x  SD=x   | distinct(D, Src) => move_shift Src SD D
+move SD=x==0 D=y | move Src=y SD=x==0 | distinct(D, Src) => move_shift Src SD D
 
-move_shift x x x
-move_shift y x x
-move_shift x y x
+move_shift cxy x x
+move_shift nx y x
 move_shift x x y
+move_shift y r y
 
 # move2_par x x x x
 
@@ -530,32 +601,26 @@ put_list Src Dst=x Dst => update_list Src Dst
 
 update_list xyc x
 
-put_list x n x
-put_list y n x
-put_list x x x
-put_list y x x
+# put_list SrcReg1 SrcReg2 => Dst
 
-put_list y y x
-put_list x y x
+put_list xy xy x
 
-# put_list SrcReg Constant Dst
+# put_list SrcReg [] => Dst
 
-put_list x c x
-put_list x c y
+put_list xy n xy
 
-put_list y c x
+# put_list SrcReg Constant => x
 
-# put_list Constant SrcReg Dst
+put_list xy c x
 
-put_list c x x
-put_list c y x
+# put_list Constant SrcReg => Dst
+
+put_list c xy x
 
 # The following put_list instructions using x(0) are frequently used.
 
-put_list r n r
-put_list r n x
-put_list r x x
-put_list r x r
+put_list r n rx
+put_list r x rx
 put_list x x r
 
 %cold
@@ -629,6 +694,11 @@ test_arity f? xy A
 
 test_arity_get_tuple_element f? x A P x
 
+is_tuple NotTupleFail Tuple=x | is_tagged_tuple WrongRecordFail Tuple Arity Atom => \
+   is_tagged_tuple_ff NotTupleFail WrongRecordFail Tuple Arity Atom
+
+is_tagged_tuple_ff f? f? rx A a
+
 get_tuple_element Reg=x P1 D1=x | get_tuple_element Reg=x P2 D2=x | \
    get_tuple_element Reg=x P3 D3=x | \
    succ(P1, P2) | succ(P2, P3) | \
@@ -1103,8 +1173,25 @@ call_fun Arity => i_call_fun Arity
 i_call_fun t
 i_call_fun_last t Q
 
+
+#
+# A fun with an empty environment can be converted to a literal.
+# As a further optimization, the we try to move the fun to its
+# final destination directly.
+
 make_fun2 OldIndex=u => gen_make_fun2(OldIndex)
 
+move_fun/2
+move_fun Fun X0 | move X0 Dst | move Src X0 => move Fun Dst | move Src X0
+move_fun Fun X0 | move A B | move X0 Dst | move Src X0 | \
+  independent_moves(Fun, X0, A, B) | distinct(Dst, A) => \
+    move Fun Dst | move A B | move Src X0
+move_fun Fun X0 | move X0 Dst | make_fun2 OldIndex | \
+  is_killed_by_make_fun(X0, OldIndex)=> \
+    move Fun Dst | make_fun2 OldIndex
+
+move_fun Fun Dst => move Fun Dst
+
 %cold
 i_make_fun W t
 %hot
@@ -1484,23 +1571,22 @@ put_map_exact F Map Dst Live Size Rest=* | map_key_sort(Size, Rest) => \
 
 sorted_put_map_assoc Map Dst Live Size Rest=* | is_empty_map(Map) => \
    new_map Dst Live Size Rest
-sorted_put_map_assoc Src=s Dst Live Size Rest=* => \
-	       update_map_assoc Src Dst Live Size Rest
-sorted_put_map_assoc Src Dst Live Size Rest=* => \
-	       move Src x | update_map_assoc x Dst Live Size Rest
+sorted_put_map_assoc Src=xyc Dst Live Size Rest=* => \
+   update_map_assoc Src Dst Live Size Rest
 
-sorted_put_map_exact F Src=s Dst Live Size Rest=* => \
-	      update_map_exact F Src Dst Live Size Rest
-sorted_put_map_exact F Src Dst Live Size Rest=* => \
-	      move Src x | update_map_exact F x Dst Live Size Rest
+sorted_put_map_exact Fail Src=xy Dst Live Size Rest=* => \
+   update_map_exact Src Fail Dst Live Size Rest
+# Literal map arguments for an exact update operation are extremely rare.
+sorted_put_map_exact Fail Src Dst Live Size Rest=* => \
+   move Src x | update_map_exact x Fail Dst Live Size Rest
 
 new_map Dst Live Size Rest=* | is_small_map_literal_keys(Size, Rest) => \
    gen_new_small_map_lit(Dst, Live, Size, Rest)
 
 new_map d t I
 i_new_small_map_lit d t q
-update_map_assoc s d t I
-update_map_exact j? s d t I
+update_map_assoc xyc d t I
+update_map_exact xy j? d t I
 
 is_map Fail Lit=q | literal_is_map(Lit) =>
 is_map Fail cq => jump Fail
@@ -1562,6 +1648,11 @@ gen_minus p Live Reg=d Int=i Dst | negation_is_small(Int) => \
 # Arithmetic instructions.
 #
 
+# It is OK to swap arguments for '+' in a guard. It is also
+# OK to turn minus into plus in a guard.
+gen_plus Fail=f Live S1=c S2 Dst => i_plus S2 S1 Fail Dst
+gen_minus Fail=f Live S1 S2=i Dst => gen_plus_from_minus(Fail, Live, S1, S2, Dst)
+
 gen_plus Fail Live S1 S2 Dst => i_plus S1 S2 Fail Dst
 
 gen_minus Fail Live S1 S2 Dst => i_minus S1 S2 Fail Dst
@@ -1595,10 +1686,13 @@ gc_bif1 Fail Live u$bif:erlang:bnot/1 Src Dst=d => i_int_bnot Fail Src Dst
 
 i_increment rxy W d
 
-i_plus x xy j? d
-i_plus s s  j? d
+# Handle unoptimized code.
+i_plus S1=c S2=c Fail Dst => move S1 x | i_plus x S2 Fail Dst
+
+i_plus xy xyc j? d
 
 i_minus x x j? d
+i_minus c x j? d
 i_minus s s j? d
 
 i_times j? s s d
@@ -1639,8 +1733,9 @@ bif1 Fail u$bif:erlang:trunc/1 s d => too_old_compiler
 gc_bif1 Fail=j Live u$bif:erlang:length/1 Src Dst => \
    i_length_setup Live Src | i_length Fail Live Dst
 
-i_length_setup t xyc
+i_length_setup Live Src=c => move Src x | i_length_setup Live x
 
+i_length_setup t xy
 i_length j? t d
 
 #
diff --git a/erts/emulator/utils/beam_makeops b/erts/emulator/utils/beam_makeops
index 1625b2cc65..605a402f2a 100755
--- a/erts/emulator/utils/beam_makeops
+++ b/erts/emulator/utils/beam_makeops
@@ -2737,7 +2737,7 @@ sub tr_maybe_keep {
 		return;
 	    }
 	} elsif ($op eq 'store_var_next_arg') {
-	    return unless shift(@last_instr) eq $args[0];
+	    return unless @last_instr and shift(@last_instr) eq $args[0];
 	} elsif (defined $pos) {
 	    return;
 	}