Merge branch 'bjorn/compiler/eprof'

* bjorn/compiler/eprof:
  v3_life: Optimize updating of the variable data base
  beam_jump: Replace use of lists:dropwhile/2 with a custom function
  beam_asm: Eliminate unnecessary use of iolist_to_binary/1
  beam_bsm: Optimize btb_index()
  beam_type: Eliminate redundant calls to checkerror_1/2
  erl_expand_records: Simplify handling of call_ext instructions
  beam_utils: Optimize index_labels_1/2
  beam_block: Optimize matching of binary literals
  Move rewriting of bs_match from beam_clean to beam_z
  v3_codegen: Reduce cost for fixing up bs_match_string instructions
  v3_codegen: Optimize "turning" of y registers
  v3_kernel: Optimize subst_vsub/3
  orddict: Eliminate unnecessary consing in store/3 and others
  compile: Add the {eprof,Pass} option for easy eprof running
  compile: Eliminate unnecessary wrappers for compiler passes
  Add z_SUITE to validate loaded code
  test suite: Always place .core files in data directories
  test suites: Unload modules compiled from .core or .S
  compilation_SUITE: Unload tested modules using the code server

14 files changed, 217 insertions, 164 deletions

diff --git a/lib/compiler/src/beam_a.erl b/lib/compiler/src/beam_a.erl
index dd7e03dd28..410f598665 100644
--- a/lib/compiler/src/beam_a.erl
+++ b/lib/compiler/src/beam_a.erl
@@ -91,6 +91,10 @@ rename_instr({bs_private_append=I,F,Sz,U,Src,Flags,Dst}) ->
     {bs_init,F,{I,U,Flags},none,[Sz,Src],Dst};
 rename_instr(bs_init_writable=I) ->
     {bs_init,{f,0},I,1,[{x,0}],{x,0}};
+rename_instr({test,Op,F,[Ctx,Bits,{string,Str}]}) ->
+    %% When compiling from a .S file.
+    <<Bs:Bits/bits,_/bits>> = list_to_binary(Str),
+    {test,Op,F,[Ctx,Bs]};
 rename_instr({put_map_assoc,Fail,S,D,R,L}) ->
     {put_map,Fail,assoc,S,D,R,L};
 rename_instr({put_map_exact,Fail,S,D,R,L}) ->
diff --git a/lib/compiler/src/beam_asm.erl b/lib/compiler/src/beam_asm.erl
index f8cf178d2e..084686def7 100644
--- a/lib/compiler/src/beam_asm.erl
+++ b/lib/compiler/src/beam_asm.erl
@@ -132,10 +132,10 @@ build_file(Code, Attr, Dict, NumLabels, NumFuncs, Abst, SourceFile, Opts) ->
     LiteralChunk = case beam_dict:literal_table(Dict) of
 		       {0,[]} -> [];
 		       {NumLiterals,LitTab0} ->
-			   LitTab1 = iolist_to_binary(LitTab0),
-			   LitTab2 = <<NumLiterals:32,LitTab1/binary>>,
-			   LitTab = iolist_to_binary(zlib:compress(LitTab2)),
-			   chunk(<<"LitT">>, <<(byte_size(LitTab2)):32>>, LitTab)
+			   LitTab1 = [<<NumLiterals:32>>,LitTab0],
+			   LitTab = zlib:compress(LitTab1),
+			   chunk(<<"LitT">>, <<(iolist_size(LitTab1)):32>>,
+				 LitTab)
 		   end,
 
     %% Create the line chunk.
diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl
index 5216f39296..e2639e9cac 100644
--- a/lib/compiler/src/beam_block.erl
+++ b/lib/compiler/src/beam_block.erl
@@ -61,15 +61,6 @@ blockify(Is) ->
 blockify([{loop_rec,{f,Fail},{x,0}},{loop_rec_end,_Lbl},{label,Fail}|Is], Acc) ->
     %% Useless instruction sequence.
     blockify(Is, Acc);
-
-%% New bit syntax matching.
-blockify([{bs_save2,R,Point}=I,{bs_restore2,R,Point}|Is], Acc) ->
-    blockify([I|Is], Acc);
-blockify([{bs_save2,R,Point}=I,{test,is_eq_exact,_,_}=Test,
-	  {bs_restore2,R,Point}|Is], Acc) ->
-    blockify([I,Test|Is], Acc);
-
-%% Do other peep-hole optimizations.
 blockify([{test,is_atom,{f,Fail},[Reg]}=I|
 	  [{select,select_val,Reg,{f,Fail},
 	    [{atom,false},{f,_}=BrFalse,
diff --git a/lib/compiler/src/beam_bsm.erl b/lib/compiler/src/beam_bsm.erl
index d54c2a9fde..427b7071ac 100644
--- a/lib/compiler/src/beam_bsm.erl
+++ b/lib/compiler/src/beam_bsm.erl
@@ -20,7 +20,7 @@
 -module(beam_bsm).
 -export([module/2,format_error/1]).
 
--import(lists, [member/2,foldl/3,reverse/1,sort/1,all/2,dropwhile/2]).
+-import(lists, [member/2,foldl/3,reverse/1,sort/1,all/2]).
 
 %%%
 %%% We optimize bit syntax matching where the tail end of a binary is
@@ -542,16 +542,13 @@ btb_context_regs_1(Regs, N, Tag, Acc) ->
 %%  a binary. MustSave is true if the function may pass the match
 %%  context to the bs_context_to_binary instruction (in which case
 %%  the current position in the binary must have saved into the
-%%  start position using "bs_save_2 Ctx start".
+%%  start position using "bs_save_2 Ctx start").
 
 btb_index(Fs) ->
     btb_index_1(Fs, []).
 
 btb_index_1([{function,_,_,Entry,Is0}|Fs], Acc0) ->
-    [{label,Entry}|Is] =
-	dropwhile(fun({label,L}) when L =:= Entry -> false;
-		     (_) -> true
-		  end, Is0),
+    Is = drop_to_label(Is0, Entry),
     Acc = btb_index_2(Is, Entry, false, Acc0),
     btb_index_1(Fs, Acc);
 btb_index_1([], Acc) -> gb_trees:from_orddict(sort(Acc)).
@@ -566,6 +563,9 @@ btb_index_2(Is0, Entry, _, Acc) ->
 	throw:none -> Acc
     end.
 
+drop_to_label([{label,L}|Is], L) -> Is;
+drop_to_label([_|Is], L) -> drop_to_label(Is, L).
+
 btb_index_find_start_match([{test,_,{f,F},_},{bs_context_to_binary,_}|Is]) ->
     btb_index_find_label(Is, F);
 btb_index_find_start_match(_) ->
diff --git a/lib/compiler/src/beam_clean.erl b/lib/compiler/src/beam_clean.erl
index b68b8702e0..1d26993103 100644
--- a/lib/compiler/src/beam_clean.erl
+++ b/lib/compiler/src/beam_clean.erl
@@ -184,14 +184,6 @@ function_replace([{function,Name,Arity,Entry,Asm0}|Fs], Dict, Acc) ->
     function_replace(Fs, Dict, [{function,Name,Arity,Entry,Asm}|Acc]);
 function_replace([], _, Acc) -> Acc.
 
-replace([{test,bs_match_string=Op,{f,Lbl},[Ctx,Bin0]}|Is], Acc, D) ->
-    Bits = bit_size(Bin0),
-    Bin = case Bits rem 8 of
-	      0 -> Bin0;
-	      Rem -> <<Bin0/bitstring,0:(8-Rem)>>
-	  end,
-    I = {test,Op,{f,label(Lbl, D)},[Ctx,Bits,{string,binary_to_list(Bin)}]},
-    replace(Is, [I|Acc], D);
 replace([{test,Test,{f,Lbl},Ops}|Is], Acc, D) ->
     replace(Is, [{test,Test,{f,label(Lbl, D)},Ops}|Acc], D);
 replace([{test,Test,{f,Lbl},Live,Ops,Dst}|Is], Acc, D) ->
diff --git a/lib/compiler/src/beam_dead.erl b/lib/compiler/src/beam_dead.erl
index f4515ba2a7..5932d8ce1d 100644
--- a/lib/compiler/src/beam_dead.erl
+++ b/lib/compiler/src/beam_dead.erl
@@ -255,6 +255,16 @@ backward([{jump,{f,To}}=J|[{bif,Op,_,Ops,Reg}|Is]=Is0], D, Acc) ->
     catch
 	throw:not_possible -> backward(Is0, D, [J|Acc])
     end;
+backward([{test,bs_start_match2,F,_,[R,_],Ctxt}=I|Is], D,
+	 [{test,bs_match_string,F,[Ctxt,Bs]},
+	  {test,bs_test_tail2,F,[Ctxt,0]}|Acc0]=Acc) ->
+    case beam_utils:is_killed(Ctxt, Acc0, D) of
+	true ->
+	    Eq = {test,is_eq_exact,F,[R,{literal,Bs}]},
+	    backward(Is, D, [Eq|Acc0]);
+	false ->
+	    backward(Is, D, [I|Acc])
+    end;
 backward([{test,bs_start_match2,{f,To0},Live,[Src|_]=Info,Dst}|Is], D, Acc) ->
     To = shortcut_bs_start_match(To0, Src, D),
     I = {test,bs_start_match2,{f,To},Live,Info,Dst},
@@ -459,8 +469,8 @@ count_bits_matched([{test,_,_,_,[_,Sz,U,{field_flags,_}],_}|Is], SavePoint, Bits
 	{integer,N} -> count_bits_matched(Is, SavePoint, Bits+N*U);
 	_ -> count_bits_matched(Is, SavePoint, Bits)
     end;
-count_bits_matched([{test,bs_match_string,_,[_,Bits,_]}|Is], SavePoint, Bits0) ->
-    count_bits_matched(Is, SavePoint, Bits0+Bits);
+count_bits_matched([{test,bs_match_string,_,[_,Bs]}|Is], SavePoint, Bits) ->
+    count_bits_matched(Is, SavePoint, Bits+bit_size(Bs));
 count_bits_matched([{test,_,_,_}|Is], SavePoint, Bits) ->
     count_bits_matched(Is, SavePoint, Bits);
 count_bits_matched([{bs_save2,Reg,SavePoint}|_], {Reg,SavePoint}, Bits) ->
diff --git a/lib/compiler/src/beam_jump.erl b/lib/compiler/src/beam_jump.erl
index ba71d4efae..52b6464c7f 100644
--- a/lib/compiler/src/beam_jump.erl
+++ b/lib/compiler/src/beam_jump.erl
@@ -127,7 +127,7 @@
 %%% on the program state.
 %%% 
 
--import(lists, [reverse/1,reverse/2,foldl/3,dropwhile/2]).
+-import(lists, [reverse/1,reverse/2,foldl/3]).
 
 module({Mod,Exp,Attr,Fs0,Lc}, _Opt) ->
     Fs = [function(F) || F <- Fs0],
@@ -509,10 +509,7 @@ rem_unused([{label,Lbl}=I|Is0], Used, [Prev|_]=Acc) ->
     case gb_sets:is_member(Lbl, Used) of
 	false ->
 	    Is = case is_unreachable_after(Prev) of
-		     true ->
-			 dropwhile(fun({label,_}) -> false;
-				      (_) -> true
-				   end, Is0);
+		     true -> drop_upto_label(Is0);
 		     false -> Is0
 		 end,
 	    rem_unused(Is, Used, Acc);
@@ -533,6 +530,10 @@ initial_labels([{label,Lbl}|Is], Acc) ->
 initial_labels([{func_info,_,_,_},{label,Lbl}|_], Acc) ->
     gb_sets:from_list([Lbl|Acc]).
 
+drop_upto_label([{label,_}|_]=Is) -> Is;
+drop_upto_label([_|Is]) -> drop_upto_label(Is);
+drop_upto_label([]) -> [].
+
 %% ulbl(Instruction, UsedGbSet) -> UsedGbSet'
 %%  Update the gb_set UsedGbSet with any function-local labels
 %%  (i.e. not with labels in call instructions) referenced by
diff --git a/lib/compiler/src/beam_type.erl b/lib/compiler/src/beam_type.erl
index 26c933481a..4731b5e78e 100644
--- a/lib/compiler/src/beam_type.erl
+++ b/lib/compiler/src/beam_type.erl
@@ -149,9 +149,10 @@ simplify_basic_1([], Ts, Acc) ->
 %%
 simplify_float(Is0, Ts0) ->
     {Is1,Ts} = simplify_float_1(Is0, Ts0, [], []),
-    Is2 = flt_need_heap(Is1),
+    Is2 = opt_fmoves(Is1, []),
+    Is3 = flt_need_heap(Is2),
     try
-	{flt_liveness(Is2),Ts}
+	{flt_liveness(Is3),Ts}
     catch
 	throw:not_possible -> not_possible
     end.
@@ -202,14 +203,15 @@ simplify_float_1([{set,_,_,{'catch',_}}=I|Is]=Is0, _Ts, Rs0, Acc0) ->
     simplify_float_1(Is, tdb_new(), Rs0, [I|Acc]);
 simplify_float_1([{set,_,_,{line,_}}=I|Is], Ts, Rs, Acc) ->
     simplify_float_1(Is, Ts, Rs, [I|Acc]);
+simplify_float_1([I|Is], Ts0, [], Acc) ->
+    Ts = update(I, Ts0),
+    simplify_float_1(Is, Ts, [], [I|Acc]);
 simplify_float_1([I|Is]=Is0, Ts0, Rs0, Acc0) ->
     Ts = update(I, Ts0),
     {Rs,Acc} = flush(Rs0, Is0, Acc0),
     simplify_float_1(Is, Ts, Rs, [I|checkerror(Acc)]);
-simplify_float_1([], Ts, Rs, Acc0) ->
-    Acc = checkerror(Acc0),
-    Is0 = reverse(flush_all(Rs, [], Acc)),
-    Is = opt_fmoves(Is0, []),
+simplify_float_1([], Ts, [], Acc) ->
+    Is = reverse(Acc),
     {Is,Ts}.
 
 coerce_to_float({integer,I}=Int) ->
diff --git a/lib/compiler/src/beam_utils.erl b/lib/compiler/src/beam_utils.erl
index 7704690f86..fd666be41e 100644
--- a/lib/compiler/src/beam_utils.erl
+++ b/lib/compiler/src/beam_utils.erl
@@ -612,13 +612,15 @@ is_reg_used_at_1(R, Lbl, St0) ->
     end.
 
 index_labels_1([{label,Lbl}|Is0], Acc) ->
-    Is = lists:dropwhile(fun({label,_}) -> true;
-			    (_) -> false end, Is0),
+    Is = drop_labels(Is0),
     index_labels_1(Is0, [{Lbl,Is}|Acc]);
 index_labels_1([_|Is], Acc) ->
     index_labels_1(Is, Acc);
 index_labels_1([], Acc) -> gb_trees:from_orddict(sort(Acc)).
 
+drop_labels([{label,_}|Is]) -> drop_labels(Is);
+drop_labels(Is) -> Is.
+
 %% Help functions for combine_heap_needs.
 
 combine_alloc_lists(Al1, Al2) ->
diff --git a/lib/compiler/src/beam_z.erl b/lib/compiler/src/beam_z.erl
index 0c7bef9183..47e786034d 100644
--- a/lib/compiler/src/beam_z.erl
+++ b/lib/compiler/src/beam_z.erl
@@ -74,6 +74,13 @@ undo_rename({bs_init,F,{I,Extra,U,Flags},Live,[Sz,Src],Dst}) ->
     {I,F,Sz,Extra,Live,U,Src,Flags,Dst};
 undo_rename({bs_init,_,bs_init_writable=I,_,_,_}) ->
     I;
+undo_rename({test,bs_match_string=Op,F,[Ctx,Bin0]}) ->
+    Bits = bit_size(Bin0),
+    Bin = case Bits rem 8 of
+	      0 -> Bin0;
+	      Rem -> <<Bin0/bitstring,0:(8-Rem)>>
+	  end,
+    {test,Op,F,[Ctx,Bits,{string,binary_to_list(Bin)}]};
 undo_rename({put_map,Fail,assoc,S,D,R,L}) ->
     {put_map_assoc,Fail,S,D,R,L};
 undo_rename({put_map,Fail,exact,S,D,R,L}) ->
diff --git a/lib/compiler/src/compile.erl b/lib/compiler/src/compile.erl
index c45c9a1a29..0b021073db 100644
--- a/lib/compiler/src/compile.erl
+++ b/lib/compiler/src/compile.erl
@@ -285,11 +285,20 @@ internal_comp(Passes, File, Suffix, St0) ->
     St1 = St0#compile{filename=File, dir=Dir, base=Base,
 		      ifile=erlfile(Dir, Base, Suffix),
 		      ofile=objfile(Base, St0)},
-    Run = case member(time, St1#compile.options) of
-	      true  ->
-		  io:format("Compiling ~tp\n", [File]),
-		  fun run_tc/2;
-	      false -> fun({_Name,Fun}, St) -> catch Fun(St) end
+    Opts = St1#compile.options,
+    Run0 = case member(time, Opts) of
+	       true  ->
+		   io:format("Compiling ~tp\n", [File]),
+		   fun run_tc/2;
+	       false -> fun({_Name,Fun}, St) -> catch Fun(St) end
+	   end,
+    Run = case keyfind(eprof, 1, Opts) of
+	      {eprof,EprofPass} ->
+		  fun(P, St) ->
+			  run_eprof(P, EprofPass, St)
+		  end;
+	      false ->
+		  Run0
 	  end,
     case fold_comp(Passes, Run, St1) of
 	{ok,St2} -> comp_ret_ok(St2);
@@ -331,6 +340,16 @@ run_tc({Name,Fun}, St) ->
 	      [Name,(After_c-Before_c) / 1000,Mem]),
     Val.
 
+run_eprof({Name,Fun}, Name, St) ->
+    io:format("~p: Running eprof\n", [Name]),
+    eprof:start_profiling([self()]),
+    Val = (catch Fun(St)),
+    eprof:stop_profiling(),
+    eprof:analyze(),
+    Val;
+run_eprof({_,Fun}, _, St) ->
+    catch Fun(St).
+
 comp_ret_ok(#compile{code=Code,warnings=Warn0,module=Mod,options=Opts}=St) ->
     case werror(St) of
         true ->
@@ -606,7 +625,7 @@ standard_passes() ->
      {iff,'to_exp',{done,"E"}},
 
      %% Conversion to Core Erlang.
-     ?pass(core_module),
+     {pass,v3_core},
      {iff,'dcore',{listing,"core"}},
      {iff,'to_core0',{done,"core"}}
      | core_passes()].
@@ -618,7 +637,7 @@ core_passes() ->
       [{unless,no_copt,
        [{core_old_inliner,fun test_old_inliner/1,fun core_old_inliner/1},
 	{iff,doldinline,{listing,"oldinline"}},
-	?pass(core_fold_module),
+	{pass,sys_core_fold},
 	{iff,dcorefold,{listing,"corefold"}},
 	{core_inline_module,fun test_core_inliner/1,fun core_inline_module/1},
 	{iff,dinline,{listing,"inline"}},
@@ -631,14 +650,14 @@ core_passes() ->
 
 kernel_passes() ->
     %% Destructive setelement/3 optimization and core lint.
-    [?pass(core_dsetel_module),
+    [{pass,sys_core_dsetel},
      {iff,dsetel,{listing,"dsetel"}},
 
      {iff,clint,?pass(core_lint_module)},
      {iff,core,?pass(save_core_code)},
 
      %% Kernel Erlang and code generation.
-     ?pass(kernel_module),
+     {pass,v3_kernel},
      {iff,dkern,{listing,"kernel"}},
      {iff,'to_kernel',{done,"kernel"}},
      {pass,v3_life},
@@ -1176,14 +1195,6 @@ expand_module(#compile{code=Code,options=Opts0}=St0) ->
     Opts = expand_opts(Opts1),
     {ok,St0#compile{module=Mod,options=Opts,code={Mod,Exp,Forms}}}.
 
-core_module(#compile{code=Code0,options=Opts}=St) ->
-    {ok,Code,Ws} = v3_core:module(Code0, Opts),
-    {ok,St#compile{code=Code,warnings=St#compile.warnings ++ Ws}}.
-
-core_fold_module(#compile{code=Code0,options=Opts,warnings=Warns}=St) ->
-    {ok,Code,Ws} = sys_core_fold:module(Code0, Opts),
-    {ok,St#compile{code=Code,warnings=Warns ++ Ws}}.
-
 core_fold_module_after_inlining(#compile{code=Code0,options=Opts}=St) ->
     %% Inlining may produce code that generates spurious warnings.
     %% Ignore all warnings.
@@ -1219,14 +1230,6 @@ core_inline_module(#compile{code=Code0,options=Opts}=St) ->
     Code = cerl_inline:core_transform(Code0, Opts),
     {ok,St#compile{code=Code}}.
 
-core_dsetel_module(#compile{code=Code0,options=Opts}=St) ->
-    {ok,Code} = sys_core_dsetel:module(Code0, Opts),
-    {ok,St#compile{code=Code}}.
-
-kernel_module(#compile{code=Code0,options=Opts}=St) ->
-    {ok,Code,Ws} = v3_kernel:module(Code0, Opts),
-    {ok,St#compile{code=Code,warnings=St#compile.warnings ++ Ws}}.
-
 save_abstract_code(#compile{ifile=File}=St) ->
     case abstract_code(St) of
 	{ok,Code} ->
diff --git a/lib/compiler/src/v3_codegen.erl b/lib/compiler/src/v3_codegen.erl
index 40235d6767..15a54a5886 100644
--- a/lib/compiler/src/v3_codegen.erl
+++ b/lib/compiler/src/v3_codegen.erl
@@ -121,24 +121,15 @@ cg_fun(Les, Hvs, Vdb, AtomMod, NameArity, Anno, St0) ->
 					   put_reg(V, Reg)
 				   end, [], Hvs),
 			 stk=[]}, 0, Vdb),
-    {B0,_Aft,St} = cg_list(Les, 0, Vdb, Bef,
+    {B,_Aft,St} = cg_list(Les, 0, Vdb, Bef,
 			  St3#cg{bfail=0,
 				 ultimate_failure=UltimateMatchFail,
 				 is_top_block=true}),
-    B = fix_bs_match_strings(B0),
     {Name,Arity} = NameArity,
     Asm = [{label,Fi},line(Anno),{func_info,AtomMod,{atom,Name},Arity},
 	   {label,Fl}|B++[{label,UltimateMatchFail},if_end]],
     {Asm,Fl,St}.
 
-fix_bs_match_strings([{test,bs_match_string,F,[Ctx,BinList]}|Is])
-  when is_list(BinList) ->
-    I = {test,bs_match_string,F,[Ctx,list_to_bitstring(BinList)]},
-    [I|fix_bs_match_strings(Is)];
-fix_bs_match_strings([I|Is]) ->
-    [I|fix_bs_match_strings(Is)];
-fix_bs_match_strings([]) -> [].
-
 %% cg(Lkexpr, Vdb, StackReg, State) -> {[Ainstr],StackReg,State}.
 %%  Generate code for a kexpr.
 %%  Split function into two steps for clarity, not efficiency.
@@ -584,7 +575,7 @@ top_level_block(Keis, Bef, MaxRegs, _St) ->
 			(return) ->
 			    [{deallocate,FrameSz},return];
 			(Tuple) when is_tuple(Tuple) ->
-			    [turn_yregs(tuple_size(Tuple), Tuple, MaxY)];
+			    [turn_yregs(Tuple, MaxY)];
 			(Other) ->
 			    [Other]
 		    end, Keis),
@@ -596,14 +587,49 @@ top_level_block(Keis, Bef, MaxRegs, _St) ->
 %%   catches work.  The code generation algorithm gives a lower register
 %%   number to the outer catch, which is wrong.
 
-turn_yregs(0, Tp, _) -> Tp;
-turn_yregs(El, Tp, MaxY) ->
-    turn_yregs(El-1,setelement(El,Tp,turn_yreg(element(El,Tp),MaxY)),MaxY).
-
-turn_yreg({yy,YY},MaxY) -> {y,MaxY-YY};
-turn_yreg({list,Ls},MaxY) -> {list, turn_yreg(Ls,MaxY)};
-turn_yreg(Ts,MaxY) when is_list(Ts) -> [turn_yreg(T,MaxY)||T<-Ts];
-turn_yreg(Other,_MaxY) -> Other.
+turn_yregs({call,_,_}=I, _MaxY) -> I;
+turn_yregs({call_ext,_,_}=I, _MaxY) -> I;
+turn_yregs({jump,_}=I, _MaxY) -> I;
+turn_yregs({label,_}=I, _MaxY) -> I;
+turn_yregs({line,_}=I, _MaxY) -> I;
+turn_yregs({test_heap,_,_}=I, _MaxY) -> I;
+turn_yregs({bif,Op,F,A,B}, MaxY) ->
+    {bif,Op,F,turn_yreg(A, MaxY),turn_yreg(B, MaxY)};
+turn_yregs({gc_bif,Op,F,Live,A,B}, MaxY) when is_integer(Live) ->
+    {gc_bif,Op,F,Live,turn_yreg(A, MaxY),turn_yreg(B, MaxY)};
+turn_yregs({get_tuple_element,S,N,D}, MaxY) ->
+    {get_tuple_element,turn_yreg(S, MaxY),N,turn_yreg(D, MaxY)};
+turn_yregs({put_tuple,Arity,D}, MaxY) ->
+    {put_tuple,Arity,turn_yreg(D, MaxY)};
+turn_yregs({select_val,R,F,L}, MaxY) ->
+    {select_val,turn_yreg(R, MaxY),F,L};
+turn_yregs({test,Op,F,L}, MaxY) ->
+    {test,Op,F,turn_yreg(L, MaxY)};
+turn_yregs({test,Op,F,Live,A,B}, MaxY) when is_integer(Live) ->
+    {test,Op,F,Live,turn_yreg(A, MaxY),turn_yreg(B, MaxY)};
+turn_yregs({Op,A}, MaxY) ->
+    {Op,turn_yreg(A, MaxY)};
+turn_yregs({Op,A,B}, MaxY) ->
+    {Op,turn_yreg(A, MaxY),turn_yreg(B, MaxY)};
+turn_yregs({Op,A,B,C}, MaxY) ->
+    {Op,turn_yreg(A, MaxY),turn_yreg(B, MaxY),turn_yreg(C, MaxY)};
+turn_yregs(Tuple, MaxY) ->
+    turn_yregs(tuple_size(Tuple), Tuple, MaxY).
+
+turn_yregs(1, Tp, _) ->
+    Tp;
+turn_yregs(N, Tp, MaxY) ->
+    E = turn_yreg(element(N, Tp), MaxY),
+    turn_yregs(N-1, setelement(N, Tp, E), MaxY).
+
+turn_yreg({yy,YY}, MaxY) ->
+    {y,MaxY-YY};
+turn_yreg({list,Ls},MaxY) ->
+    {list,turn_yreg(Ls, MaxY)};
+turn_yreg([_|_]=Ts, MaxY) ->
+    [turn_yreg(T, MaxY) || T <- Ts];
+turn_yreg(Other, _MaxY) ->
+    Other.
 
 %% select_cg(Sclause, V, TypeFail, ValueFail, StackReg, State) ->
 %%      {Is,StackReg,State}.
@@ -682,22 +708,37 @@ select_nil(#l{ke={val_clause,nil,B}}, V, Tf, Vf, Bef, St0) ->
 select_binary(#l{ke={val_clause,{binary,{var,V}},B},i=I,vdb=Vdb},
 	      V, Tf, Vf, Bef, St0) ->
     Int0 = clear_dead(Bef#sr{reg=Bef#sr.reg}, I, Vdb),
-    {Bis,Aft,St1} = match_cg(B, Vf, Int0, St0),
+    {Bis0,Aft,St1} = match_cg(B, Vf, Int0, St0),
     CtxReg = fetch_var(V, Int0),
     Live = max_reg(Bef#sr.reg),
-    {[{test,bs_start_match2,{f,Tf},Live,[CtxReg,V],CtxReg},
-      {bs_save2,CtxReg,{V,V}}|Bis],
-     Aft,St1};
+    Bis1 = [{test,bs_start_match2,{f,Tf},Live,[CtxReg,V],CtxReg},
+	    {bs_save2,CtxReg,{V,V}}|Bis0],
+    Bis = finish_select_binary(Bis1),
+    {Bis,Aft,St1};
 select_binary(#l{ke={val_clause,{binary,{var,Ivar}},B},i=I,vdb=Vdb},
 	      V, Tf, Vf, Bef, St0) ->
     Regs = put_reg(Ivar, Bef#sr.reg),
     Int0 = clear_dead(Bef#sr{reg=Regs}, I, Vdb),
-    {Bis,Aft,St1} = match_cg(B, Vf, Int0, St0),
+    {Bis0,Aft,St1} = match_cg(B, Vf, Int0, St0),
     CtxReg = fetch_var(Ivar, Int0),
     Live = max_reg(Bef#sr.reg),
-    {[{test,bs_start_match2,{f,Tf},Live,[fetch_var(V, Bef),Ivar],CtxReg},
-      {bs_save2,CtxReg,{Ivar,Ivar}}|Bis],
-     Aft,St1}.
+    Bis1 = [{test,bs_start_match2,{f,Tf},Live,[fetch_var(V, Bef),Ivar],CtxReg},
+	    {bs_save2,CtxReg,{Ivar,Ivar}}|Bis0],
+    Bis = finish_select_binary(Bis1),
+    {Bis,Aft,St1}.
+
+finish_select_binary([{bs_save2,R,Point}=I,{bs_restore2,R,Point}|Is]) ->
+    [I|finish_select_binary(Is)];
+finish_select_binary([{bs_save2,R,Point}=I,{test,is_eq_exact,_,_}=Test,
+		      {bs_restore2,R,Point}|Is]) ->
+    [I,Test|finish_select_binary(Is)];
+finish_select_binary([{test,bs_match_string,F,[Ctx,BinList]}|Is])
+  when is_list(BinList) ->
+    I = {test,bs_match_string,F,[Ctx,list_to_bitstring(BinList)]},
+    [I|finish_select_binary(Is)];
+finish_select_binary([I|Is]) ->
+    [I|finish_select_binary(Is)];
+finish_select_binary([]) -> [].
 
 %% New instructions for selection of binary segments.
 
diff --git a/lib/compiler/src/v3_kernel.erl b/lib/compiler/src/v3_kernel.erl
index 0ac1aaf158..7dff58582e 100644
--- a/lib/compiler/src/v3_kernel.erl
+++ b/lib/compiler/src/v3_kernel.erl
@@ -836,12 +836,23 @@ get_vsub(V, Vsub) ->
 set_vsub(V, S, Vsub) ->
     orddict:store(V, S, Vsub).
 
-subst_vsub(V, S, Vsub0) ->
-    %% Fold chained substitutions.
-    Vsub1 = orddict:map(fun (_, V1) when V1 =:= V -> S;
-			    (_, V1) -> V1
-			end, Vsub0),
-    orddict:store(V, S, Vsub1).
+subst_vsub(Key, New, [{K,Key}|Dict]) ->
+    %% Fold chained substitution.
+    [{K,New}|subst_vsub(Key, New, Dict)];
+subst_vsub(Key, New, [{K,_}|_]=Dict) when Key < K ->
+    %% Insert the new substitution here, and continue
+    %% look for chained substitutions.
+    [{Key,New}|subst_vsub_1(Key, New, Dict)];
+subst_vsub(Key, New, [{K,_}=E|Dict]) when Key > K ->
+    [E|subst_vsub(Key, New, Dict)];
+subst_vsub(Key, New, []) -> [{Key,New}].
+
+subst_vsub_1(V, S, [{K,V}|Dict]) ->
+    %% Fold chained substitution.
+    [{K,S}|subst_vsub_1(V, S, Dict)];
+subst_vsub_1(V, S, [E|Dict]) ->
+    [E|subst_vsub_1(V, S, Dict)];
+subst_vsub_1(_, _, []) -> [].
 
 get_fsub(F, A, Fsub) ->
     case orddict:find({F,A}, Fsub) of
diff --git a/lib/compiler/src/v3_life.erl b/lib/compiler/src/v3_life.erl
index 75bd188479..4b1f1c3f71 100644
--- a/lib/compiler/src/v3_life.erl
+++ b/lib/compiler/src/v3_life.erl
@@ -45,7 +45,7 @@
 
 -export([vdb_find/2]).
 
--import(lists, [member/2,map/2,foldl/3,reverse/1,sort/1]).
+-import(lists, [member/2,map/2,reverse/1,sort/1]).
 -import(ordsets, [add_element/2,intersection/2,union/2]).
 
 -include("v3_kernel.hrl").
@@ -68,7 +68,7 @@ functions([], Acc) -> reverse(Acc).
 function(#k_fdef{anno=#k{a=Anno},func=F,arity=Ar,vars=Vs,body=Kb}) ->
     try
 	As = var_list(Vs),
-	Vdb0 = foldl(fun ({var,N}, Vdb) -> new_var(N, 0, Vdb) end, [], As),
+	Vdb0 = init_vars(As),
 	%% Force a top-level match!
 	B0 = case Kb of
 		 #k_match{} -> Kb;
@@ -94,14 +94,14 @@ function(#k_fdef{anno=#k{a=Anno},func=F,arity=Ar,vars=Vs,body=Kb}) ->
 body(#k_seq{arg=Ke,body=Kb}, I, Vdb0) ->
     %%ok = io:fwrite("life ~w:~p~n", [?LINE,{Ke,I,Vdb0}]),
     A = get_kanno(Ke),
-    Vdb1 = use_vars(A#k.us, I, new_vars(A#k.ns, I, Vdb0)),
+    Vdb1 = use_vars(union(A#k.us, A#k.ns), I, Vdb0),
     {Es,MaxI,Vdb2} = body(Kb, I+1, Vdb1),
     E = expr(Ke, I, Vdb2),
     {[E|Es],MaxI,Vdb2};
 body(Ke, I, Vdb0) ->
     %%ok = io:fwrite("life ~w:~p~n", [?LINE,{Ke,I,Vdb0}]),
     A = get_kanno(Ke),
-    Vdb1 = use_vars(A#k.us, I, new_vars(A#k.ns, I, Vdb0)),
+    Vdb1 = use_vars(union(A#k.us, A#k.ns), I, Vdb0),
     E = expr(Ke, I, Vdb1),
     {[E],I,Vdb1}.
 
@@ -150,12 +150,12 @@ expr(#k_try_enter{anno=A,arg=Ka,vars=Vs,body=Kb,evars=Evs,handler=Kh}, I, Vdb) -
     %% the body and handler. Add try tag 'variable'.
     Ab = get_kanno(Kb),
     Ah = get_kanno(Kh),
-    Tdb1 = use_vars(Ab#k.us, I+3, use_vars(Ah#k.us, I+3, Tdb0)),
+    Tdb1 = use_vars(union(Ab#k.us, Ah#k.us), I+3, Tdb0),
     Tdb2 = vdb_sub(I, I+2, Tdb1),
     Vnames = fun (Kvar) -> Kvar#k_var.name end,	%Get the variable names
     {Aes,_,Adb} = body(Ka, I+2, add_var({catch_tag,I+1}, I+1, 1000000, Tdb2)),
-    {Bes,_,Bdb} = body(Kb, I+4, new_vars(map(Vnames, Vs), I+3, Tdb2)),
-    {Hes,_,Hdb} = body(Kh, I+4, new_vars(map(Vnames, Evs), I+3, Tdb2)),
+    {Bes,_,Bdb} = body(Kb, I+4, new_vars(sort(map(Vnames, Vs)), I+3, Tdb2)),
+    {Hes,_,Hdb} = body(Kh, I+4, new_vars(sort(map(Vnames, Evs)), I+3, Tdb2)),
     #l{ke={try_enter,#l{ke={block,Aes},i=I+1,vdb=Adb,a=[]},
 	   var_list(Vs),#l{ke={block,Bes},i=I+3,vdb=Bdb,a=[]},
 	   var_list(Evs),#l{ke={block,Hes},i=I+3,vdb=Hdb,a=[]}},
@@ -171,7 +171,7 @@ expr(#k_receive{anno=A,var=V,body=Kb,timeout=T,action=Ka,ret=Rs}, I, Vdb) ->
     %% Work out imported variables which need to be locked.
     Rdb = vdb_sub(I, I+1, Vdb),
     M = match(Kb, add_element(V#k_var.name, A#k.us), I+1, [],
- 	      new_var(V#k_var.name, I, Rdb)),
+	      new_vars([V#k_var.name], I, Rdb)),
     {Tes,_,Adb} = body(Ka, I+1, Rdb),
     #l{ke={receive_loop,atomic(T),variable(V),M,
 	   #l{ke=Tes,i=I+1,vdb=Adb,a=[]},var_list(Rs)},
@@ -199,12 +199,12 @@ body_try(#k_try{anno=A,arg=Ka,vars=Vs,body=Kb,evars=Evs,handler=Kh,ret=Rs},
     %% the body and handler. Add try tag 'variable'.
     Ab = get_kanno(Kb),
     Ah = get_kanno(Kh),
-    Tdb1 = use_vars(Ab#k.us, I+3, use_vars(Ah#k.us, I+3, Tdb0)),
+    Tdb1 = use_vars(union(Ab#k.us, Ah#k.us), I+3, Tdb0),
     Tdb2 = vdb_sub(I, I+2, Tdb1),
     Vnames = fun (Kvar) -> Kvar#k_var.name end,	%Get the variable names
     {Aes,_,Adb} = body(Ka, I+2, add_var({catch_tag,I+1}, I+1, locked, Tdb2)),
-    {Bes,_,Bdb} = body(Kb, I+4, new_vars(map(Vnames, Vs), I+3, Tdb2)),
-    {Hes,_,Hdb} = body(Kh, I+4, new_vars(map(Vnames, Evs), I+3, Tdb2)),
+    {Bes,_,Bdb} = body(Kb, I+4, new_vars(sort(map(Vnames, Vs)), I+3, Tdb2)),
+    {Hes,_,Hdb} = body(Kh, I+4, new_vars(sort(map(Vnames, Evs)), I+3, Tdb2)),
     #l{ke={'try',#l{ke={block,Aes},i=I+1,vdb=Adb,a=[]},
 	   var_list(Vs),#l{ke={block,Bes},i=I+3,vdb=Bdb,a=[]},
 	   var_list(Evs),#l{ke={block,Hes},i=I+3,vdb=Hdb,a=[]},
@@ -400,79 +400,68 @@ is_gc_bif(Bif, Arity) ->
 	 erl_internal:new_type_test(Bif, Arity) orelse
 	 erl_internal:comp_op(Bif, Arity)).
 
-%% new_var(VarName, I, Vdb) -> Vdb.
+%% Keep track of life time for variables.
+%%
+%% init_vars([{var,VarName}]) -> Vdb.
 %% new_vars([VarName], I, Vdb) -> Vdb.
-%% use_var(VarName, I, Vdb) -> Vdb.
 %% use_vars([VarName], I, Vdb) -> Vdb.
 %% add_var(VarName, F, L, Vdb) -> Vdb.
+%%
+%% The list of variable names for new_vars/3 and use_vars/3
+%% must be sorted.
 
-new_var(V, I, Vdb) ->
-    vdb_store_new(V, I, I, Vdb).
+init_vars(Vs) ->
+    sort([{V,0,0} || {var,V} <- Vs]).
 
-new_vars(Vs, I, Vdb0) ->
-    foldl(fun (V, Vdb) -> new_var(V, I, Vdb) end, Vdb0, Vs).
+new_vars([], _, Vdb) -> Vdb;
+new_vars([V], I, Vdb) -> vdb_store_new(V, {V,I,I}, Vdb);
+new_vars(Vs, I, Vdb) -> vdb_update_vars(Vs, Vdb, I).
 
-use_var(V, I, Vdb) ->
+use_vars([], _, Vdb) ->
+    Vdb;
+use_vars([V], I, Vdb) ->
     case vdb_find(V, Vdb) of
-	{V,F,L} when I > L -> vdb_update(V, F, I, Vdb);
+	{V,F,L} when I > L -> vdb_update(V, {V,F,I}, Vdb);
 	{V,_,_} -> Vdb;
-	error -> vdb_store_new(V, I, I, Vdb)
-    end.
-
-use_vars([], _, Vdb) -> Vdb;
-use_vars([V], I, Vdb) -> use_var(V, I, Vdb);
-use_vars(Vs, I, Vdb) ->
-    Res = use_vars_1(sort(Vs), Vdb, I),
-    %% The following line can be used as an assertion.
-    %%   Res = foldl(fun (V, Vdb) -> use_var(V, I, Vdb) end, Vdb, Vs),
-    Res.
-
-%% Measurements show that it is worthwhile having this special
-%% function that updates/inserts several variables at once.
-
-use_vars_1([V|_]=Vs, [{V1,_,_}=Vd|Vdb], I) when V > V1 ->
-    [Vd|use_vars_1(Vs, Vdb, I)];
-use_vars_1([V|Vs], [{V1,_,_}|_]=Vdb, I) when V < V1 ->
-    %% New variable.
-    [{V,I,I}|use_vars_1(Vs, Vdb, I)];
-use_vars_1([V|Vs], [{_,F,L}=Vd|Vdb], I) ->
-    %% Existing variable.
-    if
-	I > L ->[{V,F,I}|use_vars_1(Vs, Vdb, I)];
-	true -> [Vd|use_vars_1(Vs, Vdb, I)]
+	error -> vdb_store_new(V, {V,I,I}, Vdb)
     end;
-use_vars_1([V|Vs], [], I) ->
-    %% New variable.
-    [{V,I,I}|use_vars_1(Vs, [], I)];
-use_vars_1([], Vdb, _) -> Vdb.
+use_vars(Vs, I, Vdb) -> vdb_update_vars(Vs, Vdb, I).
 
 add_var(V, F, L, Vdb) ->
-    vdb_store_new(V, F, L, Vdb).
+    vdb_store_new(V, {V,F,L}, Vdb).
 
 vdb_find(V, Vdb) ->
-    %% Performance note: Profiling shows that this function accounts for
-    %% a lot of the execution time when huge constant terms are built.
-    %% Using the BIF lists:keyfind/3 is a lot faster than the
-    %% original Erlang version.
     case lists:keyfind(V, 1, Vdb) of
 	false -> error;
 	Vd -> Vd
     end.
 
-%vdb_find(V, [{V1,F,L}=Vd|Vdb]) when V < V1 -> error;
-%vdb_find(V, [{V1,F,L}=Vd|Vdb]) when V == V1 -> Vd;
-%vdb_find(V, [{V1,F,L}=Vd|Vdb]) when V > V1 -> vdb_find(V, Vdb);
-%vdb_find(V, []) -> error.
+vdb_update(V, Update, [{V,_,_}|Vdb]) ->
+    [Update|Vdb];
+vdb_update(V, Update, [Vd|Vdb]) ->
+    [Vd|vdb_update(V, Update, Vdb)].
 
-vdb_update(V, F, L, [{V1,_,_}=Vd|Vdb]) when V > V1 ->
-    [Vd|vdb_update(V, F, L, Vdb)];
-vdb_update(V, F, L, [{V1,_,_}|Vdb]) when V == V1 ->
-    [{V,F,L}|Vdb].
+vdb_store_new(V, New, [{V1,_,_}=Vd|Vdb]) when V > V1 ->
+    [Vd|vdb_store_new(V, New, Vdb)];
+vdb_store_new(V, New, [{V1,_,_}|_]=Vdb) when V < V1 ->
+    [New|Vdb];
+vdb_store_new(_, New, []) -> [New].
 
-vdb_store_new(V, F, L, [{V1,_,_}=Vd|Vdb]) when V > V1 ->
-    [Vd|vdb_store_new(V, F, L, Vdb)];
-vdb_store_new(V, F, L, [{V1,_,_}|_]=Vdb) when V < V1 -> [{V,F,L}|Vdb];
-vdb_store_new(V, F, L, []) -> [{V,F,L}].
+vdb_update_vars([V|_]=Vs, [{V1,_,_}=Vd|Vdb], I) when V > V1 ->
+    [Vd|vdb_update_vars(Vs, Vdb, I)];
+vdb_update_vars([V|Vs], [{V1,_,_}|_]=Vdb, I) when V < V1 ->
+    %% New variable.
+    [{V,I,I}|vdb_update_vars(Vs, Vdb, I)];
+vdb_update_vars([V|Vs], [{_,F,L}=Vd|Vdb], I) ->
+    %% Existing variable.
+    if
+	I > L -> [{V,F,I}|vdb_update_vars(Vs, Vdb, I)];
+	true ->  [Vd|vdb_update_vars(Vs, Vdb, I)]
+    end;
+vdb_update_vars([V|Vs], [], I) ->
+    %% New variable.
+    [{V,I,I}|vdb_update_vars(Vs, [], I)];
+vdb_update_vars([], Vdb, _) -> Vdb.
 
 %% vdb_sub(Min, Max, Vdb) -> Vdb.
 %%  Extract variables which are used before and after Min.  Lock