Optimize moving of several Y registers to X registers

Introduce move_src_window[234] instructions for moving several
consecutively numbered Y registers to discontiguously numbered X
registers. This optimization is effective because the compiler has
sorted the `move` instructions in Y register order.

2 files changed, 60 insertions, 2 deletions

diff --git a/erts/emulator/beam/instrs.tab b/erts/emulator/beam/instrs.tab
index 3cb12dba4a..692408e212 100644
--- a/erts/emulator/beam/instrs.tab
+++ b/erts/emulator/beam/instrs.tab
@@ -515,6 +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;
diff --git a/erts/emulator/beam/ops.tab b/erts/emulator/beam/ops.tab
index 067f81930d..6832e65b1b 100644
--- a/erts/emulator/beam/ops.tab
+++ b/erts/emulator/beam/ops.tab
@@ -270,8 +270,8 @@ move_jump f cxy xy
 move_jump f c r
 
 
-# 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
@@ -299,6 +299,31 @@ 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=xy Tmp=x | move R2=xy R1 | move Tmp R2 => swap_temp R1 R2 Tmp