Replace swap_temp with swap more aggressively

Also support swap of Y registers.

2 files changed, 77 insertions, 2 deletions

diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index 42b03fc031..5961b7d743 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -3146,6 +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.
  */
 
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index 2f6f4d6649..4dfea096b2 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -298,10 +298,23 @@ move_window4 x x x x y
 move_window5 x x x x x y
 
 # 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
@@ -317,9 +330,42 @@ 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 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
+
 swap_temp x x x
 
-swap x x
+swap xy x
+swap y y
 
 # move_dup