Pack sequences of trailing 'f' operands

Pack sequences of trailing 'f' operands for instructions
such at jump_on_val or i_select_val_lins.

3 files changed, 100 insertions, 18 deletions

diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index 39e96126bf..9ba65e2464 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -54,6 +54,7 @@ void dbg_where(BeamInstr* addr, Eterm x0, Eterm* reg);
 static int print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr);
 static void print_bif_name(fmtfn_t to, void* to_arg, BifFunction bif);
 static BeamInstr* f_to_addr(BeamInstr* base, int op, BeamInstr* ap);
+static BeamInstr* f_to_addr_packed(BeamInstr* base, int op, Sint32* ap);
 
 BIF_RETTYPE
 erts_debug_same_2(BIF_ALIST_2)
@@ -634,6 +635,7 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
 	{
 	    int n = ap[-1];
 	    int ix = n;
+            Sint32* jump_tab = (Sint32 *)(ap + n);
 
 	    while (ix--) {
 		erts_print(to, to_arg, "%T ", (Eterm) ap[0]);
@@ -642,11 +644,11 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
 	    }
 	    ix = n;
 	    while (ix--) {
-                BeamInstr* target = f_to_addr(addr, op, ap);
+                BeamInstr* target = f_to_addr_packed(addr, op, jump_tab);
 		erts_print(to, to_arg, "f(" HEXF ") ", target);
-		ap++;
-		size++;
+                jump_tab++;
 	    }
+            size += (n+1) / 2;
 	}
 	break;
     case op_i_select_val_bins_xfI:
@@ -668,6 +670,7 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
         {
             int n = ap[-1];
             int ix = n - 1; /* without sentinel */
+            Sint32* jump_tab = (Sint32 *)(ap + n);
 
             while (ix--) {
                 Uint arity = arityval(ap[0]);
@@ -681,34 +684,38 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
             size++;
             ix = n;
             while (ix--) {
-                BeamInstr* target = f_to_addr(addr, op, ap);
+                BeamInstr* target = f_to_addr_packed(addr, op, jump_tab);
                 erts_print(to, to_arg, "f(" HEXF ") ", target);
-                ap++;
-                size++;
+                jump_tab++;
             }
+            size += (n+1) / 2;
         }
         break;
     case op_i_jump_on_val_xfIW:
     case op_i_jump_on_val_yfIW:
 	{
-	    int n;
-	    for (n = ap[-2]; n > 0; n--) {
-                BeamInstr* target = f_to_addr(addr, op, ap);
+	    int n = ap[-2];
+            Sint32* jump_tab = (Sint32 *) ap;
+
+            size += (n+1) / 2;
+            while (n-- > 0) {
+                BeamInstr* target = f_to_addr_packed(addr, op, jump_tab);
 		erts_print(to, to_arg, "f(" HEXF ") ", target);
-		ap++;
-		size++;
+                jump_tab++;
 	    }
 	}
 	break;
     case op_i_jump_on_val_zero_xfI:
     case op_i_jump_on_val_zero_yfI:
 	{
-	    int n;
-	    for (n = ap[-1]; n > 0; n--) {
-                BeamInstr* target = f_to_addr(addr, op, ap);
+	    int n = ap[-1];
+            Sint32* jump_tab = (Sint32 *) ap;
+
+            size += (n+1) / 2;
+            while (n-- > 0) {
+                BeamInstr* target = f_to_addr_packed(addr, op, jump_tab);
 		erts_print(to, to_arg, "f(" HEXF ") ", target);
-		ap++;
-		size++;
+                jump_tab++;
 	    }
 	}
 	break;
@@ -810,6 +817,11 @@ static BeamInstr* f_to_addr(BeamInstr* base, int op, BeamInstr* ap)
     return base - 1 + opc[op].adjust + (Sint32) *ap;
 }
 
+static BeamInstr* f_to_addr_packed(BeamInstr* base, int op, Sint32* ap)
+{
+    return base - 1 + opc[op].adjust + *ap;
+}
+
 
 /*
  * Dirty BIF testing.
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index 2acf3a2c1c..9bb5866cb3 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -1976,6 +1976,9 @@ load_code(LoaderState* stp)
     GenOp** last_op_next = NULL;
     int arity;
     int retval = 1;
+#if defined(BEAM_WIDE_SHIFT)
+    int num_trailing_f;     /* Number of extra 'f' arguments in a list */
+#endif
 
     /*
      * The size of the loaded func_info instruction is needed
@@ -2661,7 +2664,17 @@ load_code(LoaderState* stp)
 	 * Load any list arguments using the primitive tags.
 	 */
 
+#if defined(BEAM_WIDE_SHIFT)
+        num_trailing_f = 0;
+#endif
 	for ( ; arg < tmp_op->arity; arg++) {
+#if defined(BEAM_WIDE_SHIFT)
+	    if (tmp_op->a[arg].type == TAG_f) {
+                num_trailing_f++;
+            } else {
+                num_trailing_f = 0;
+            }
+#endif
 	    switch (tmp_op->a[arg].type) {
 	    case TAG_i:
 		CodeNeed(1);
@@ -2701,6 +2714,61 @@ load_code(LoaderState* stp)
 	    }
 	}
 
+        /*
+         * If all the extra arguments were 'f' operands,
+         * and the wordsize is 64 bits, pack two 'f' operands
+         * into each word.
+         */
+
+#if defined(BEAM_WIDE_SHIFT)
+        if (num_trailing_f >= 1) {
+            Uint src_index = ci - num_trailing_f;
+            Uint src_limit = ci;
+            Uint dst_limit = src_index + (num_trailing_f+1)/2;
+
+            ci = src_index;
+            while (ci < dst_limit) {
+                Uint w[2];
+                BeamInstr packed = 0;
+                int wi;
+
+                w[0] = code[src_index];
+                if (src_index+1 < src_limit) {
+                    w[1] = code[src_index+1];
+                } else {
+                    w[1] = 0;
+                }
+                for (wi = 0; wi < 2; wi++) {
+                    Uint lbl = w[wi];
+                    LabelPatch* lp = stp->labels[lbl].patches;
+                    int num_patches = stp->labels[lbl].num_patches;
+
+#if defined(WORDS_BIGENDIAN)
+                    packed <<= BEAM_WIDE_SHIFT;
+                    packed |= lbl & BEAM_WIDE_MASK;
+#else
+                    packed >>= BEAM_WIDE_SHIFT;
+                    packed |= lbl << BEAM_WIDE_SHIFT;
+#endif
+                    while (num_patches-- > 0) {
+                        if (lp->pos == src_index + wi) {
+                            lp->pos = ci;
+#if defined(WORDS_BIGENDIAN)
+                            lp->packed = 2 - wi;
+#else
+                            lp->packed = wi + 1;
+#endif
+                            break;
+                        }
+                        lp++;
+                    }
+                }
+                code[ci++] = packed;
+                src_index += 2;
+            }
+        }
+#endif
+
 	/*
 	 * Handle a few special cases.
 	 */
diff --git a/erts/emulator/beam/select_instrs.tab b/erts/emulator/beam/select_instrs.tab
index da6b7dbe62..6b59f02925 100644
--- a/erts/emulator/beam/select_instrs.tab
+++ b/erts/emulator/beam/select_instrs.tab
@@ -150,7 +150,8 @@ select_val_lin.execute(N) {
     }
 
     if (vs[ix] == select_val) {
-        Eterm offset = *($NEXT_INSTRUCTION + $N + ix);
+        Sint32* jump_tab = (Sint32 *) ($NEXT_INSTRUCTION + $N);
+        Eterm offset = jump_tab[ix];
         $JUMP(offset);
     } else {
         $JUMP(*select_fail);
@@ -161,7 +162,8 @@ JUMP_ON_VAL(Fail, Index, N, Base) {
     if (is_small($Index)) {
         $Index = (Uint) (signed_val($Index) - $Base);
         if ($Index < $N) {
-            $JUMP((($NEXT_INSTRUCTION)[$Index]));
+            Sint32* jump_tab = (Sint32 *) ($NEXT_INSTRUCTION);
+            $JUMP(jump_tab[$Index]);
         }
     }
     $FAIL($Fail);