Improve optimization of switches

Part of the switch optimization done by `ssa_opt_sw` can be better
done in `beam_ssa_type`.

2 files changed, 79 insertions, 93 deletions

diff --git a/lib/compiler/src/beam_ssa_opt.erl b/lib/compiler/src/beam_ssa_opt.erl
index 0f03d0e736..f8e19d0aa7 100644
--- a/lib/compiler/src/beam_ssa_opt.erl
+++ b/lib/compiler/src/beam_ssa_opt.erl
@@ -164,12 +164,14 @@ repeated_passes(Opts) ->
 epilogue_passes(Opts) ->
     Ps = [?PASS(ssa_opt_type_finish),
           ?PASS(ssa_opt_float),
-          ?PASS(ssa_opt_live),                  %One last time to clean up the
-                                                %mess left by the float pass.
+          ?PASS(ssa_opt_sw),
+
+          %% Run live one more time to clean up after the float and sw
+          %% passes.
+          ?PASS(ssa_opt_live),
           ?PASS(ssa_opt_bsm),
           ?PASS(ssa_opt_bsm_units),
           ?PASS(ssa_opt_bsm_shortcut),
-          ?PASS(ssa_opt_sw),
           ?PASS(ssa_opt_blockify),
           ?PASS(ssa_opt_sink),
           ?PASS(ssa_opt_merge_blocks),
@@ -1830,12 +1832,16 @@ opt_tup_size_is([], _, _, _Acc) -> none.
 %%%
 
 ssa_opt_sw({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) ->
-    {Linear,Count} = opt_sw(Linear0, #{}, Count0, []),
+    {Linear,Count} = opt_sw(Linear0, Count0, []),
     {St#st{ssa=Linear,cnt=Count}, FuncDb}.
 
-opt_sw([{L,#b_blk{is=Is,last=#b_switch{}=Last0}=Blk0}|Bs], Phis0, Count0, Acc) ->
-    Phis = opt_sw_phis(Is, Phis0),
-    case opt_sw_last(Last0, Phis) of
+opt_sw([{L,#b_blk{is=Is,last=#b_switch{}=Sw0}=Blk0}|Bs], Count0, Acc) ->
+    %% Ensure that no label in the switch list is the same
+    %% as the failure label.
+    #b_switch{fail=Fail,list=List0} = Sw0,
+    List = [{Val,Lbl} || {Val,Lbl} <- List0, Lbl =/= Fail],
+    Sw1 = beam_ssa:normalize(Sw0#b_switch{list=List}),
+    case Sw1 of
         #b_switch{arg=Arg,fail=Fail,list=[{Lit,Lbl}]} ->
             %% Rewrite a single value switch to a br.
             Bool = #b_var{name={'@ssa_bool',Count0}},
@@ -1843,7 +1849,7 @@ opt_sw([{L,#b_blk{is=Is,last=#b_switch{}=Last0}=Blk0}|Bs], Phis0, Count0, Acc) -
             IsEq = #b_set{op={bif,'=:='},dst=Bool,args=[Arg,Lit]},
             Br = #b_br{bool=Bool,succ=Lbl,fail=Fail},
             Blk = Blk0#b_blk{is=Is++[IsEq],last=Br},
-            opt_sw(Bs, Phis, Count, [{L,Blk}|Acc]);
+            opt_sw(Bs, Count, [{L,Blk}|Acc]);
         #b_switch{arg=Arg,fail=Fail,
                   list=[{#b_literal{val=B1},Lbl},{#b_literal{val=B2},Lbl}]}
           when B1 =:= not B2 ->
@@ -1853,71 +1859,18 @@ opt_sw([{L,#b_blk{is=Is,last=#b_switch{}=Last0}=Blk0}|Bs], Phis0, Count0, Acc) -
             IsBool = #b_set{op={bif,is_boolean},dst=Bool,args=[Arg]},
             Br = #b_br{bool=Bool,succ=Lbl,fail=Fail},
             Blk = Blk0#b_blk{is=Is++[IsBool],last=Br},
-            opt_sw(Bs, Phis, Count, [{L,Blk}|Acc]);
-        Last0 ->
-            opt_sw(Bs, Phis, Count0, [{L,Blk0}|Acc]);
-        Last ->
-            Blk = Blk0#b_blk{last=Last},
-            opt_sw(Bs, Phis, Count0, [{L,Blk}|Acc])
+            opt_sw(Bs, Count, [{L,Blk}|Acc]);
+        Sw0 ->
+            opt_sw(Bs, Count0, [{L,Blk0}|Acc]);
+        Sw ->
+            Blk = Blk0#b_blk{last=Sw},
+            opt_sw(Bs, Count0, [{L,Blk}|Acc])
     end;
-opt_sw([{L,#b_blk{is=Is}=Blk}|Bs], Phis0, Count, Acc) ->
-    Phis = opt_sw_phis(Is, Phis0),
-    opt_sw(Bs, Phis, Count, [{L,Blk}|Acc]);
-opt_sw([], _Phis, Count, Acc) ->
+opt_sw([{L,#b_blk{}=Blk}|Bs], Count, Acc) ->
+    opt_sw(Bs, Count, [{L,Blk}|Acc]);
+opt_sw([], Count, Acc) ->
     {reverse(Acc),Count}.
 
-opt_sw_phis([#b_set{op=phi,dst=Dst,args=Args}|Is], Phis) ->
-    case opt_sw_literals(Args, []) of
-        error ->
-            opt_sw_phis(Is, Phis);
-        Literals ->
-            opt_sw_phis(Is, Phis#{Dst=>Literals})
-    end;
-opt_sw_phis(_, Phis) -> Phis.
-
-opt_sw_last(#b_switch{arg=Arg,fail=Fail,list=List0}=Sw0, Phis) ->
-    case Phis of
-        #{Arg:=Values0} ->
-            Values = gb_sets:from_list(Values0),
-
-            %% Prune the switch list to only contain the possible values.
-            List1 = [P || {Lit,_}=P <- List0, gb_sets:is_member(Lit, Values)],
-
-            %% Now test whether the failure label can ever be reached.
-            Sw = case gb_sets:size(Values) =:= length(List1) of
-                     true ->
-                         %% The switch list has the same number of values as the phi node.
-                         %% The values must be the same, because the values that were not
-                         %% possible were pruned from the switch list. Therefore, the
-                         %% failure label can't possibly be reached, and we can choose a
-                         %% a new failure label by picking a value from the list.
-                         case List1 of
-                             [{#b_literal{},Lbl}|List] ->
-                                 Sw0#b_switch{fail=Lbl,list=List};
-                             [] ->
-                                 Sw0#b_switch{list=List1}
-                         end;
-                     false ->
-                         %% There are some values in the phi node that are not in the
-                         %% switch list; thus, the failure label can still be reached.
-                         Sw0
-                 end,
-            beam_ssa:normalize(Sw);
-        #{} ->
-            %% Ensure that no label in the switch list is the same
-            %% as the failure label.
-            List = [{Val,Lbl} || {Val,Lbl} <- List0, Lbl =/= Fail],
-            Sw = Sw0#b_switch{list=List},
-            beam_ssa:normalize(Sw)
-    end.
-
-opt_sw_literals([{#b_literal{}=Lit,_}|T], Acc) ->
-    opt_sw_literals(T, [Lit|Acc]);
-opt_sw_literals([_|_], _Acc) ->
-    error;
-opt_sw_literals([], Acc) -> Acc.
-
-
 %%%
 %%% Merge blocks.
 %%%
diff --git a/lib/compiler/src/beam_ssa_type.erl b/lib/compiler/src/beam_ssa_type.erl
index 4b53848a74..83f8a3cc56 100644
--- a/lib/compiler/src/beam_ssa_type.erl
+++ b/lib/compiler/src/beam_ssa_type.erl
@@ -23,7 +23,8 @@
 
 -include("beam_ssa_opt.hrl").
 -import(lists, [all/2,any/2,droplast/1,foldl/3,last/1,member/2,
-                partition/2,reverse/1,reverse/2,seq/2,sort/1]).
+                keyfind/3,partition/2,reverse/1,reverse/2,
+                seq/2,sort/1]).
 
 -define(UNICODE_INT, #t_integer{elements={0,16#10FFFF}}).
 
@@ -664,25 +665,45 @@ opt_terminator(#b_br{bool=#b_var{}}=Br, Ts, Ds) ->
     simplify_not(Br, Ts, Ds);
 opt_terminator(#b_switch{arg=#b_literal{}}=Sw, _Ts, _Ds) ->
     beam_ssa:normalize(Sw);
-opt_terminator(#b_switch{arg=#b_var{}=V}=Sw0, Ts, Ds) ->
-    Type = get_type(V, Ts),
+opt_terminator(#b_switch{arg=#b_var{}=V}=Sw, Ts, Ds) ->
+    case get_type(V, Ts) of
+        any ->
+            beam_ssa:normalize(Sw);
+        Type ->
+            beam_ssa:normalize(opt_switch(Sw, Type, Ts, Ds))
+    end;
+opt_terminator(#b_ret{}=Ret, _Ts, _Ds) -> Ret.
+
+
+opt_switch(#b_switch{fail=Fail,list=List0}=Sw0, Type, Ts, Ds) ->
+    List = prune_switch_list(List0, Fail, Type, Ts),
+    Sw1 = Sw0#b_switch{list=List},
     case Type of
         #t_integer{elements={_,_}=Range} ->
-            simplify_switch_int(Sw0, Range);
-        _ ->
+            simplify_switch_int(Sw1, Range);
+        #t_atom{elements=[_|_]} ->
             case t_is_boolean(Type) of
                 true ->
-                    case simplify_switch_bool(Sw0, Ts, Ds) of
-                        #b_br{}=Br ->
-                            opt_terminator(Br, Ts, Ds);
-                        Sw ->
-                            beam_ssa:normalize(Sw)
-                    end;
+                    #b_br{} = Br = simplify_switch_bool(Sw1, Ts, Ds),
+                    opt_terminator(Br, Ts, Ds);
                 false ->
-                    beam_ssa:normalize(Sw0)
-            end
+                    simplify_switch_atom(Type, Sw1)
+            end;
+        _ ->
+            Sw1
+    end.
+
+prune_switch_list([{_,Fail}|T], Fail, Type, Ts) ->
+    prune_switch_list(T, Fail, Type, Ts);
+prune_switch_list([{Arg,_}=Pair|T], Fail, Type, Ts) ->
+    case meet(get_type(Arg, Ts), Type) of
+        none ->
+            %% Different types. This value can never match.
+            prune_switch_list(T, Fail, Type, Ts);
+        _ ->
+            [Pair|prune_switch_list(T, Fail, Type, Ts)]
     end;
-opt_terminator(#b_ret{}=Ret, _Ts, _Ds) -> Ret.
+prune_switch_list([], _, _, _) -> [].
 
 update_successors(#b_br{bool=#b_literal{val=true},succ=S}, Ts, D) ->
     update_successor(S, Ts, D);
@@ -1079,6 +1100,17 @@ bs_match_type(utf16, _) ->
 bs_match_type(utf32, _) ->
     ?UNICODE_INT.
 
+simplify_switch_atom(#t_atom{elements=Atoms}, #b_switch{list=List0}=Sw) ->
+    case sort([A || {#b_literal{val=A},_} <- List0]) of
+        Atoms ->
+            %% All possible atoms are included in the list. The
+            %% failure label will never be used.
+            [{_,Fail}|List] = List0,
+            Sw#b_switch{fail=Fail,list=List};
+        _ ->
+            Sw
+    end.
+
 simplify_switch_int(#b_switch{list=List0}=Sw, {Min,Max}) ->
     List1 = sort(List0),
     Vs = [V || {#b_literal{val=V},_} <- List1],
@@ -1123,14 +1155,14 @@ simplify_is_record(I, any, _Size, _Tag, _Ts) ->
 simplify_is_record(_I, _Type, _Size, _Tag, _Ts) ->
     #b_literal{val=false}.
 
-simplify_switch_bool(#b_switch{arg=B,list=List0}=Sw, Ts, Ds) ->
-    List = sort(List0),
-    case List of
-        [{#b_literal{val=false},Fail},{#b_literal{val=true},Succ}] ->
-            simplify_not(#b_br{bool=B,succ=Succ,fail=Fail}, Ts, Ds);
-        [_|_] ->
-            Sw
-    end.
+simplify_switch_bool(#b_switch{arg=B,fail=Fail,list=List0}, Ts, Ds) ->
+    FalseVal = #b_literal{val=false},
+    TrueVal = #b_literal{val=true},
+    List1 = List0 ++ [{FalseVal,Fail},{TrueVal,Fail}],
+    {_,FalseLbl} = keyfind(FalseVal, 1, List1),
+    {_,TrueLbl} = keyfind(TrueVal, 1, List1),
+    Br = beam_ssa:normalize(#b_br{bool=B,succ=TrueLbl,fail=FalseLbl}),
+    simplify_not(Br, Ts, Ds).
 
 simplify_not(#b_br{bool=#b_var{}=V,succ=Succ,fail=Fail}=Br0, Ts, Ds) ->
     case Ds of
@@ -1144,7 +1176,8 @@ simplify_not(#b_br{bool=#b_var{}=V,succ=Succ,fail=Fail}=Br0, Ts, Ds) ->
             end;
         #{} ->
             Br0
-    end.
+    end;
+simplify_not(#b_br{bool=#b_literal{}}=Br, _Ts, _Ds) -> Br.
 
 %%%
 %%% Calculate the set of variables that are only used once in the