Merge pull request #2242 from bjorng/bjorn/add-swap-instruction

Make the swap instruction known to the compiler

1 files changed, 37 insertions, 2 deletions

diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl
index 707974b2c1..a734ca3a10 100644
--- a/lib/compiler/src/beam_block.erl
+++ b/lib/compiler/src/beam_block.erl
@@ -33,8 +33,9 @@ module({Mod,Exp,Attr,Fs0,Lc}, _Opts) ->
 
 function({function,Name,Arity,CLabel,Is0}) ->
     try
-        Is1 = blockify(Is0),
-        Is = embed_lines(Is1),
+        Is1 = swap_opt(Is0),
+        Is2 = blockify(Is1),
+        Is = embed_lines(Is2),
         {function,Name,Arity,CLabel,Is}
     catch
         Class:Error:Stack ->
@@ -42,6 +43,40 @@ function({function,Name,Arity,CLabel,Is0}) ->
 	    erlang:raise(Class, Error, Stack)
     end.
 
+%%%
+%%% Try to use a `swap` instruction instead of a sequence of moves.
+%%%
+%%% Note that beam_ssa_codegen generates `swap` instructions only for
+%%% the moves within a single SSA instruction (such as `call`), not
+%%% for the moves generated by a sequence of SSA instructions.
+%%% Therefore, this optimization is needed.
+%%%
+
+swap_opt([{move,Reg1,{x,X}=Temp}=Move1,
+          {move,Reg2,Reg1}=Move2,
+          {move,Temp,Reg2}=Move3|Is]) when Reg1 =/= Temp ->
+    case is_unused(X, Is) of
+        true ->
+            [{swap,Reg1,Reg2}|swap_opt(Is)];
+        false ->
+            [Move1|swap_opt([Move2,Move3|Is])]
+    end;
+swap_opt([I|Is]) ->
+    [I|swap_opt(Is)];
+swap_opt([]) -> [].
+
+is_unused(X, [{call,A,_}|_]) when A =< X -> true;
+is_unused(X, [{call_ext,A,_}|_]) when A =< X -> true;
+is_unused(X, [{make_fun2,_,_,_,A}|_]) when A =< X -> true;
+is_unused(X, [{move,Src,Dst}|Is]) ->
+    case {Src,Dst} of
+        {{x,X},_} -> false;
+        {_,{x,X}} -> true;
+        {_,_} -> is_unused(X, Is)
+    end;
+is_unused(X, [{line,_}|Is]) -> is_unused(X, Is);
+is_unused(_, _) -> false.
+
 %% blockify(Instructions0) -> Instructions
 %%  Collect sequences of instructions to basic blocks.
 %%  Also do some simple optimations on instructions outside the blocks.