Merge branch 'maint'

Conflicts:
	erts/emulator/beam/beam_emu.c

12 files changed, 468 insertions, 307 deletions

diff --git a/lib/hipe/cerl/cerl_to_icode.erl b/lib/hipe/cerl/cerl_to_icode.erl
index 97b95f2e7c..ab131c2d01 100644
--- a/lib/hipe/cerl/cerl_to_icode.erl
+++ b/lib/hipe/cerl/cerl_to_icode.erl
@@ -794,9 +794,9 @@ bitstr_gen_op([V], #ctxt{fail=FL, class=guard}, SizeInfo, ConstInfo,
 	      Type, Flags, Base, Offset) ->
     SL = new_label(),
     case SizeInfo of
-	{all,_NewUnit, NewAlign, S1} ->
+	{all, NewUnit, NewAlign, S1} ->
 	    Type = binary,
-	    Name = {bs_put_binary_all, Flags},
+	    Name = {bs_put_binary_all, NewUnit, Flags},
 	    Primop = {hipe_bs_primop, Name},
 	    {add_code([icode_guardop([Offset], Primop,
 				     [V, Base, Offset], SL, FL),
@@ -819,9 +819,9 @@ bitstr_gen_op([V], #ctxt{fail=FL, class=guard}, SizeInfo, ConstInfo,
 bitstr_gen_op([V], _Ctxt, SizeInfo, ConstInfo, Type, Flags, Base,
 	      Offset) ->
     case SizeInfo of
-	{all, _NewUnit, NewAlign, S} ->
+	{all, NewUnit, NewAlign, S} ->
 	    Type = binary,
-	    Name = {bs_put_binary_all, Flags},
+	    Name = {bs_put_binary_all, NewUnit, Flags},
 	    Primop = {hipe_bs_primop, Name},
 	    {add_code([icode_call_primop([Offset], Primop, 
 					 [V, Base, Offset])], S), 
diff --git a/lib/hipe/icode/hipe_beam_to_icode.erl b/lib/hipe/icode/hipe_beam_to_icode.erl
index 0c7bdb788a..37c6ba9c60 100644
--- a/lib/hipe/icode/hipe_beam_to_icode.erl
+++ b/lib/hipe/icode/hipe_beam_to_icode.erl
@@ -881,6 +881,15 @@ trans_fun([{bs_init_bits,{f,Lbl},Size,_Words,_LiveRegs,{field_flags,Flags0},X}|
 trans_fun([{bs_add, {f,Lbl}, [Old,New,Unit], Res}|Instructions], Env) ->
   Dst = mk_var(Res),
   Temp = mk_var(new),
+  {FailLblName, FailCode} =
+    if Lbl =:= 0 ->
+	FailLbl = mk_label(new),
+	{hipe_icode:label_name(FailLbl),
+	 [FailLbl,
+	  hipe_icode:mk_fail([hipe_icode:mk_const(badarg)], error)]};
+       true ->
+	{map_label(Lbl), []}
+    end,
   MultIs =
     case {New,Unit} of
       {{integer, NewInt}, _} ->
@@ -890,40 +899,26 @@ trans_fun([{bs_add, {f,Lbl}, [Old,New,Unit], Res}|Instructions], Env) ->
 	[hipe_icode:mk_move(Temp, NewVar)];
       _ ->
 	NewVar = mk_var(New),
-	if Lbl =:= 0 ->
-	    [hipe_icode:mk_primop([Temp], '*', 
-				  [NewVar, hipe_icode:mk_const(Unit)])];
-	   true ->
-	    Succ = mk_label(new),
-	    [hipe_icode:mk_primop([Temp], '*', 
-				  [NewVar, hipe_icode:mk_const(Unit)],
-				  hipe_icode:label_name(Succ), map_label(Lbl)),
-	     Succ]
-	end
+	Succ = mk_label(new),
+	[hipe_icode:mk_primop([Temp], '*',
+			      [NewVar, hipe_icode:mk_const(Unit)],
+			      hipe_icode:label_name(Succ), FailLblName),
+	 Succ]
     end,
   Succ2 = mk_label(new),
-  {FailLblName, FailCode} = 
-    if Lbl =:= 0 ->
-	FailLbl = mk_label(new),
-	{hipe_icode:label_name(FailLbl),
-	 [FailLbl,
-	  hipe_icode:mk_fail([hipe_icode:mk_const(badarg)], error)]};
-       true ->
-	{map_label(Lbl), []}
-    end,
   IsPos = 
     [hipe_icode:mk_if('>=', [Temp, hipe_icode:mk_const(0)], 
 		      hipe_icode:label_name(Succ2), FailLblName)] ++
-    FailCode ++ [Succ2], 
-  AddI =
+    FailCode ++ [Succ2],
+  AddRhs =
     case Old of
-      {integer,OldInt} ->
-	hipe_icode:mk_primop([Dst], '+', [Temp, hipe_icode:mk_const(OldInt)]);
-      _ ->
-	OldVar = mk_var(Old),
-	hipe_icode:mk_primop([Dst], '+', [Temp, OldVar])
+      {integer,OldInt} -> hipe_icode:mk_const(OldInt);
+      _ -> mk_var(Old)
     end,
-  MultIs ++ IsPos ++ [AddI|trans_fun(Instructions, Env)];
+  Succ3 = mk_label(new),
+  AddI = hipe_icode:mk_primop([Dst], '+', [Temp, AddRhs],
+			      hipe_icode:label_name(Succ3), FailLblName),
+  MultIs ++ IsPos ++ [AddI,Succ3|trans_fun(Instructions, Env)];
 %%--------------------------------------------------------------------
 %% Bit syntax instructions added in R12B-5 (Fall 2008)
 %%--------------------------------------------------------------------
@@ -1306,7 +1301,7 @@ trans_bin([{bs_put_binary,{f,Lbl},Size,Unit,{field_flags,Flags},Source}|
   {Name, Args, Env2} =
     case Size of
       {atom,all} -> %% put all bits
-	{{bs_put_binary_all, Flags}, [Src,Base,Offset], Env};
+	{{bs_put_binary_all, Unit, Flags}, [Src,Base,Offset], Env};
       {integer,NoBits} when is_integer(NoBits), NoBits >= 0 ->
 	%% Create a N*Unit bits subbinary
 	{{bs_put_binary, NoBits*Unit, Flags}, [Src,Base,Offset], Env};
diff --git a/lib/hipe/icode/hipe_icode_primops.erl b/lib/hipe/icode/hipe_icode_primops.erl
index ffb51b2ea4..84aae30291 100644
--- a/lib/hipe/icode/hipe_icode_primops.erl
+++ b/lib/hipe/icode/hipe_icode_primops.erl
@@ -116,7 +116,7 @@ is_safe({hipe_bs_primop, {bs_init, _, _}}) -> false;
 is_safe({hipe_bs_primop, {bs_init_bits, _}}) -> false;
 is_safe({hipe_bs_primop, {bs_init_bits, _, _}}) -> false;
 is_safe({hipe_bs_primop, {bs_put_binary, _, _}}) -> false;
-is_safe({hipe_bs_primop, {bs_put_binary_all, _}}) -> false;  
+is_safe({hipe_bs_primop, {bs_put_binary_all, _, _}}) -> false;
 is_safe({hipe_bs_primop, {bs_put_float, _, _, _}}) -> false;
 is_safe({hipe_bs_primop, {bs_put_integer, _, _, _}}) -> false;
 is_safe({hipe_bs_primop, {bs_put_string, _, _}}) -> false;  
@@ -219,7 +219,7 @@ fails({hipe_bs_primop, {bs_init, _, _}}) -> true;
 fails({hipe_bs_primop, {bs_init_bits, _}}) -> true;
 fails({hipe_bs_primop, {bs_init_bits, _, _}}) -> true;
 fails({hipe_bs_primop, {bs_put_binary, _, _}}) -> true;
-fails({hipe_bs_primop, {bs_put_binary_all, _}}) -> true;  
+fails({hipe_bs_primop, {bs_put_binary_all, _, _}}) -> true;
 fails({hipe_bs_primop, {bs_put_float, _, _, _}}) -> true;
 fails({hipe_bs_primop, {bs_put_integer, _, _, _}}) -> true;
 fails({hipe_bs_primop, {bs_put_string, _, _}}) -> true;  
@@ -265,8 +265,8 @@ pp(Dev, Op) ->
       io:format(Dev, "gc_test<~w>", [N]);
     {hipe_bs_primop, BsOp}  ->
       case BsOp of
-	{bs_put_binary_all, Flags} -> 
-	  io:format(Dev, "bs_put_binary_all<~w>", [Flags]);
+	{bs_put_binary_all, Unit, Flags} ->
+	  io:format(Dev, "bs_put_binary_all<~w, ~w>", [Unit,Flags]);
 	{bs_put_binary, Size} ->
 	  io:format(Dev, "bs_put_binary<~w>", [Size]);
 	{bs_put_binary, Flags, Size} ->
@@ -505,11 +505,13 @@ type(Primop, Args) ->
 	  NewMatchState = 
 	    erl_types:t_matchstate_update_present(NewBinType, MatchState),
 	  if Signed =:= 0 ->
-	      erl_types:t_product([erl_types:t_from_range(0, 1 bsl Size - 1), 
+	      UpperBound = inf_add(safe_bsl(1, Size), -1),
+	      erl_types:t_product([erl_types:t_from_range(0, UpperBound),
 				   NewMatchState]);
 	     Signed =:= 4 ->
-	      erl_types:t_product([erl_types:t_from_range(- (1 bsl (Size-1)), 
-							  (1 bsl (Size-1)) - 1),
+	      erl_types:t_product([erl_types:t_from_range(
+				     inf_inv(safe_bsl(1, Size-1)),
+				     inf_add(safe_bsl(1, Size-1), -1)),
 				   NewMatchState])
 	  end;
 	[_Arg] ->
@@ -628,8 +630,9 @@ type(Primop, Args) ->
 	[_SrcType, _BitsType, _Base, Type] ->
 	  erl_types:t_bitstr_concat(Type, erl_types:t_bitstr(Size, 0))
       end;
-    {hipe_bs_primop, {bs_put_binary_all, _Flags}} ->
-      [SrcType, _Base, Type] = Args,
+    {hipe_bs_primop, {bs_put_binary_all, Unit, _Flags}} ->
+      [SrcType0, _Base, Type] = Args,
+      SrcType = erl_types:t_inf(erl_types:t_bitstr(Unit, 0), SrcType0),
       erl_types:t_bitstr_concat(SrcType,Type);
     {hipe_bs_primop, {bs_put_string, _, Size}} ->
       [_Base, Type] = Args,
@@ -965,3 +968,19 @@ check_fun_args(_, _) ->
 
 match_bin(Pattern, Match) ->
   erl_types:t_bitstr_match(Pattern, Match).
+
+safe_bsl(0, _) -> 0;
+safe_bsl(Base, Shift) when Shift =< 128 -> Base bsl Shift;
+safe_bsl(Base, _Shift) when Base > 0 -> pos_inf;
+safe_bsl(Base, _Shift) when Base < 0 -> neg_inf.
+
+inf_inv(pos_inf) -> neg_inf;
+inf_inv(neg_inf) -> pos_inf;
+inf_inv(Number) -> -Number.
+
+inf_add(pos_inf, _Number) -> pos_inf;
+inf_add(neg_inf, _Number) -> neg_inf;
+inf_add(_Number, pos_inf) -> pos_inf;
+inf_add(_Number, neg_inf) -> neg_inf;
+inf_add(Number1, Number2) when is_integer(Number1), is_integer(Number2) ->
+  Number1 + Number2.
diff --git a/lib/hipe/icode/hipe_icode_range.erl b/lib/hipe/icode/hipe_icode_range.erl
index ba5fa1bfd7..9a20527a83 100644
--- a/lib/hipe/icode/hipe_icode_range.erl
+++ b/lib/hipe/icode/hipe_icode_range.erl
@@ -1202,11 +1202,11 @@ basic_type(#unsafe_update_element{}) -> not_analysed.
 analyse_bs_get_integer(Size, Flags, true) ->
   Signed = Flags band 4,
   if Signed =:= 0 ->
-      Max = 1 bsl Size - 1,
+      Max = inf_add(inf_bsl(1, Size), -1),
       Min = 0;
      true ->
-      Max = 1 bsl (Size-1) - 1,
-      Min = -(1 bsl (Size-1))
+      Max = inf_add(inf_bsl(1, Size-1), -1),
+      Min = inf_inv(inf_bsl(1, Size-1))
   end,
   {Min, Max};
 analyse_bs_get_integer(Size, Flags, false) when is_integer(Size),
diff --git a/lib/hipe/rtl/hipe_rtl_arith.inc b/lib/hipe/rtl/hipe_rtl_arith.inc
index 1f455f47ed..1dff56b074 100644
--- a/lib/hipe/rtl/hipe_rtl_arith.inc
+++ b/lib/hipe/rtl/hipe_rtl_arith.inc
@@ -30,13 +30,13 @@
 %% Returns a tuple
 %%  {Res, Sign, Zero, Overflow, Carry}
 %%  Res will be a number in the range 
-%%   MAX_SIGNED_INT >= Res >= MIN_SIGNED_INT
+%%   MAX_UNSIGNED_INT >= Res >= 0
 %% The other four values are flags that are either true or false
 %% 
 eval_alu(Op, Arg1, Arg2) 
-  when Arg1 =< ?MAX_SIGNED_INT, 
+  when Arg1 =< ?MAX_UNSIGNED_INT,
        Arg1 >= ?MIN_SIGNED_INT,
-       Arg2 =< ?MAX_SIGNED_INT, 
+       Arg2 =< ?MAX_UNSIGNED_INT,
        Arg2 >= ?MIN_SIGNED_INT ->
 
   Sign1 = sign_bit(Arg1),
@@ -111,7 +111,7 @@ eval_alu(Op, Arg1, Arg2)
       Res = N = Z = V = C = 0,
       ?EXIT({"unknown alu op", Op})
   end,
-  {two_comp_to_erl(Res), N, Z, V, C};
+  {Res, N, Z, V, C};
 eval_alu(Op, Arg1, Arg2) ->
   ?EXIT({argument_overflow,Op,Arg1,Arg2}).
 
@@ -162,16 +162,9 @@ eval_cond(Cond, Arg1, Arg2) ->
 sign_bit(Val) ->
   ((Val bsr ?SIGN_BIT) band 1) =:= 1.
 
-two_comp_to_erl(V) ->
-  if V > ?MAX_SIGNED_INT ->
-      - ((?MAX_UNSIGNED_INT + 1) - V);
-     true -> V
-  end.
-
 shiftmask(Arg) ->
   Setbits = ?BITS - Arg,
   (1 bsl Setbits) - 1.
 
 zero(Val) ->
   Val =:= 0.
-
diff --git a/lib/hipe/rtl/hipe_rtl_arith_32.erl b/lib/hipe/rtl/hipe_rtl_arith_32.erl
index 572556be1c..d790a8b981 100644
--- a/lib/hipe/rtl/hipe_rtl_arith_32.erl
+++ b/lib/hipe/rtl/hipe_rtl_arith_32.erl
@@ -24,7 +24,8 @@
 %%  Filename : 	hipe_rtl_arith_32.erl
 %%  Module   :	hipe_rtl_arith_32
 %%  Purpose  :  To implement 32-bit RTL-arithmetic 
-%%  Notes    :  The arithmetic works on 32-bit signed integers. 
+%%  Notes    :  The arithmetic works on 32-bit signed and unsigned
+%%              integers.
 %%              The implementation is taken from the implementation
 %%              of arithmetic on SPARC.
 %%              XXX: This code is seldom used, and hence also
diff --git a/lib/hipe/rtl/hipe_rtl_binary.erl b/lib/hipe/rtl/hipe_rtl_binary.erl
index b549073050..9cbab08ee2 100644
--- a/lib/hipe/rtl/hipe_rtl_binary.erl
+++ b/lib/hipe/rtl/hipe_rtl_binary.erl
@@ -1,3 +1,4 @@
+%% -*- erlang-indent-level: 2 -*-
 %%%
 %%% %CopyrightBegin%
 %%% 
@@ -28,11 +29,20 @@
 
 -export([gen_rtl/7]).
 
+-export([floorlog2/1, get_word_integer/4, make_size/3, make_size/4]).
+
+%%--------------------------------------------------------------------
+
+-define(BYTE_SHIFT, 3). %% Turn bits into bytes or vice versa
+-define(BYTE_SIZE, 8).
+
+%%--------------------------------------------------------------------
+
 gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SysLimName, ConstTab) ->
   case type_of_operation(BsOP) of
     match ->
       {hipe_rtl_binary_match:gen_rtl(
-	BsOP, Dst, Args, TrueLblName, FalseLblName),ConstTab};
+	 BsOP, Dst, Args, TrueLblName, FalseLblName),ConstTab};
     construct ->
       hipe_rtl_binary_construct:gen_rtl(
 	BsOP, Dst, Args, TrueLblName, FalseLblName, SysLimName, ConstTab)
@@ -62,7 +72,7 @@ type_of_operation({bs_init,_,_}) -> construct;
 type_of_operation({bs_init_bits,_}) -> construct;
 type_of_operation({bs_init_bits,_,_}) -> construct;
 type_of_operation({bs_put_binary,_,_}) -> construct;
-type_of_operation({bs_put_binary_all,_}) -> construct;  
+type_of_operation({bs_put_binary_all,_,_}) -> construct;
 type_of_operation({bs_put_float,_,_,_}) -> construct;
 type_of_operation({bs_put_integer,_,_,_}) -> construct;
 type_of_operation({bs_put_string,_,_}) -> construct;  
@@ -79,3 +89,133 @@ type_of_operation(bs_final) -> construct;
 type_of_operation({bs_append,_,_,_,_}) -> construct;
 type_of_operation({bs_private_append,_,_}) -> construct;
 type_of_operation(bs_init_writable) -> construct.
+
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+%%
+%% Small utility functions:
+%%
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+
+create_lbls(X) when X > 0 ->
+  [hipe_rtl:mk_new_label()|create_lbls(X-1)];
+create_lbls(0) ->
+  [].
+
+%%------------------------------------------------------------------------------
+%% Utilities used by both hipe_rtl_binary_construct and hipe_rtl_binary_match
+%%------------------------------------------------------------------------------
+
+get_word_integer(Var, Register, SystemLimitLblName, FalseLblName) ->
+  [EndLbl] = create_lbls(1),
+  EndName = hipe_rtl:label_name(EndLbl),
+  get_word_integer(Var, Register,SystemLimitLblName,  FalseLblName, EndName, EndName,
+		   [EndLbl]).
+
+get_word_integer(Var, Register, SystemLimitLblName, FalseLblName, TrueLblName,
+		 BigLblName, Tail) ->
+  [FixnumLbl, NotFixnumLbl, BignumLbl, SuccessLbl] = create_lbls(4),
+  [hipe_tagscheme:test_fixnum(Var, hipe_rtl:label_name(FixnumLbl),
+			      hipe_rtl:label_name(NotFixnumLbl), 0.99),
+   FixnumLbl,
+   hipe_tagscheme:fixnum_ge(Var, hipe_rtl:mk_imm(hipe_tagscheme:mk_fixnum(0)),
+			    hipe_rtl:label_name(SuccessLbl), FalseLblName,
+			    0.99),
+   SuccessLbl,
+   hipe_tagscheme:untag_fixnum(Register, Var),
+   hipe_rtl:mk_goto(TrueLblName),
+   NotFixnumLbl,
+   hipe_tagscheme:test_pos_bignum_arity(Var, 1, hipe_rtl:label_name(BignumLbl),
+					FalseLblName, SystemLimitLblName, 0.99),
+   BignumLbl,
+   hipe_tagscheme:unsafe_get_one_word_pos_bignum(Register, Var),
+   hipe_rtl:mk_goto(BigLblName) | Tail].
+
+make_size(UnitImm, BitsVar, FailLblName) ->
+  make_size(UnitImm, BitsVar, FailLblName, FailLblName).
+
+make_size(1, BitsVar, OverflowLblName, FalseLblName) ->
+  DstReg = hipe_rtl:mk_new_reg_gcsafe(),
+  {get_word_integer(BitsVar, DstReg, OverflowLblName, FalseLblName), DstReg};
+make_size(?BYTE_SIZE, BitsVar, OverflowLblName, FalseLblName) ->
+  DstReg = hipe_rtl:mk_new_reg_gcsafe(),
+  [FixnumLbl, BignumLbl] = create_lbls(2),
+  WordBits = hipe_rtl_arch:word_size() * ?BYTE_SIZE,
+  FixnumLblName = hipe_rtl:label_name(FixnumLbl),
+  Tail = [BignumLbl,
+	  hipe_rtl:mk_branch(DstReg, 'ltu',
+			     hipe_rtl:mk_imm(1 bsl (WordBits - ?BYTE_SHIFT)),
+			     FixnumLblName, OverflowLblName, 0.99),
+	  FixnumLbl,
+	  hipe_rtl:mk_alu(DstReg, DstReg, sll, hipe_rtl:mk_imm(?BYTE_SHIFT))],
+  Code = get_word_integer(BitsVar, DstReg, OverflowLblName, FalseLblName,
+			  FixnumLblName, hipe_rtl:label_name(BignumLbl), Tail),
+  {Code, DstReg};
+make_size(UnitImm, BitsVar, OverflowLblName, FalseLblName) ->
+  DstReg = hipe_rtl:mk_new_reg_gcsafe(),
+  UnitList = number2list(UnitImm),
+  Code = multiply_code(UnitList, BitsVar, DstReg, OverflowLblName, FalseLblName),
+  {Code, DstReg}.
+
+multiply_code(List=[Head|_Tail], Variable, Result, OverflowLblName,
+	      FalseLblName) ->
+  Test = set_high(Head),
+  Tmp1 = hipe_rtl:mk_new_reg(),
+  SuccessLbl = hipe_rtl:mk_new_label(),
+  Register = hipe_rtl:mk_new_reg(),
+  Code = [hipe_rtl:mk_move(Result, hipe_rtl:mk_imm(0))|
+	  get_word_integer(Variable, Register, OverflowLblName, FalseLblName)]
+    ++
+    [hipe_rtl:mk_alub(Tmp1, Register, 'and', hipe_rtl:mk_imm(Test),
+		      eq, hipe_rtl:label_name(SuccessLbl),
+		      OverflowLblName, 0.99),
+     SuccessLbl],
+  multiply_code(List, Register, Result, OverflowLblName, Tmp1, Code).
+
+multiply_code([ShiftSize|Rest], Register, Result, OverflowLblName, Tmp1,
+	      OldCode) ->
+  SuccessLbl = hipe_rtl:mk_new_label(),
+  Code =
+    OldCode ++
+    [hipe_rtl:mk_alu(Tmp1, Register, sll, hipe_rtl:mk_imm(ShiftSize)),
+     hipe_rtl:mk_alub(Result, Tmp1, 'add', Result, not_overflow,
+		      hipe_rtl:label_name(SuccessLbl), OverflowLblName, 0.99),
+     SuccessLbl],
+  multiply_code(Rest, Register, Result, OverflowLblName, Tmp1, Code);
+multiply_code([], _Register, _Result, _OverflowLblName, _Tmp1, Code) ->
+  Code.
+
+set_high(X) ->
+  WordBits = hipe_rtl_arch:word_size() * ?BYTE_SIZE,
+  set_high(min(X, WordBits), WordBits, 0).
+
+set_high(0, _, Y) ->
+  Y;
+set_high(X, WordBits, Y) ->
+  set_high(X-1, WordBits, Y+(1 bsl (WordBits-X))).
+
+
+number2list(X) when is_integer(X), X >= 0 ->
+  number2list(X, []).
+
+number2list(1, Acc) ->
+  lists:reverse([0|Acc]);
+number2list(0, Acc) ->
+  lists:reverse(Acc);
+number2list(X, Acc) ->
+  F = floorlog2(X),
+  number2list(X-(1 bsl F), [F|Acc]).
+
+floorlog2(X) ->
+  %% Double-precision floats do not have enough precision to make floorlog2
+  %% exact for integers larger than 2^47.
+  Approx = round(math:log(X)/math:log(2)-0.5),
+  floorlog2_refine(X, Approx).
+
+floorlog2_refine(X, Approx) ->
+  if (1 bsl Approx) > X ->
+      floorlog2_refine(X, Approx - 1);
+     (1 bsl (Approx+1)) > X ->
+      Approx;
+     true ->
+      floorlog2_refine(X, Approx + 1)
+  end.
diff --git a/lib/hipe/rtl/hipe_rtl_binary_construct.erl b/lib/hipe/rtl/hipe_rtl_binary_construct.erl
index 692bad7d96..4403aa552f 100644
--- a/lib/hipe/rtl/hipe_rtl_binary_construct.erl
+++ b/lib/hipe/rtl/hipe_rtl_binary_construct.erl
@@ -34,6 +34,10 @@
 			 get_field_from_term/3,
 			 set_field_from_pointer/3,
 			 get_field_from_pointer/3]).
+
+-import(hipe_rtl_binary, [floorlog2/1,
+			  get_word_integer/4,
+			  make_size/4]).
 %%-------------------------------------------------------------------------
 
 -include("../main/hipe.hrl").
@@ -94,10 +98,11 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 	    var_init_bits(Size, Dst0, Base, Offset, TrueLblName, 
 			  SystemLimitLblName, FalseLblName);
 	  
-	  {bs_put_binary_all, _Flags} ->
+	  {bs_put_binary_all, Unit, _Flags} ->
 	    [Src, Base, Offset] = Args,
 	    [NewOffset] = get_real(Dst),
-	    put_binary_all(NewOffset, Src, Base, Offset, TrueLblName, FalseLblName);
+	    put_binary_all(NewOffset, Src, Base, Offset, Unit,
+			   TrueLblName, FalseLblName);
 	   
 	  {bs_put_binary, Size, _Flags} ->
 	    case is_illegal_const(Size) of
@@ -110,7 +115,9 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 		    put_static_binary(NewOffset, Src, Size, Base, Offset, 
 				      TrueLblName, FalseLblName);
 		  [Src, Bits, Base, Offset]  ->
-		    {SizeCode, SizeReg} = make_size(Size, Bits, FalseLblName),
+		    {SizeCode, SizeReg} = make_size(Size, Bits,
+						    SystemLimitLblName,
+						    FalseLblName),
 		    InCode = put_dynamic_binary(NewOffset, Src, SizeReg, Base, 
 						Offset, TrueLblName, FalseLblName),
 		    SizeCode ++ InCode
@@ -132,7 +139,9 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 		   put_float(NewOffset, Src, Base, Offset, Size, CCode, Aligned, 
 			     LittleEndian, ConstInfo, TrueLblName);
 		 [Src, Bits, Base, Offset] ->
-		   {SizeCode, SizeReg} = make_size(Size, Bits, FalseLblName),
+		   {SizeCode, SizeReg} = make_size(Size, Bits,
+						   SystemLimitLblName,
+						   FalseLblName),
 		   InCode = float_c_code(NewOffset, Src, Base, Offset, SizeReg, 
 					 Flags, TrueLblName, FalseLblName),
 		   SizeCode ++ InCode
@@ -161,6 +170,7 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 				       CCode, Aligned, LittleEndian, TrueLblName); 
 		      [Src, Bits, Base, Offset] -> 
 		    {SizeCode, SizeReg} = make_size(Size, Bits,
+						    SystemLimitLblName,
 						    FalseLblName), 
 			CCode = int_c_code(NewOffset, Src, Base,
 					   Offset, SizeReg, Flags, 
@@ -209,6 +219,7 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 					       TrueLblName); 
 		       [Src, Bits, Base, Offset] -> 
 			 {SizeCode, SizeReg} = make_size(Size, Bits,
+							 SystemLimitLblName,
 							 FalseLblName), 
 			 CCode = int_c_code(NewOffset, Src, Base,
 					    Offset, SizeReg, Flags, 
@@ -293,7 +304,7 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 	    [SizeReg] = create_regs(1),
 	    [Base] = create_unsafe_regs(1),
 	    [hipe_rtl:mk_gctest(?PROC_BIN_WORDSIZE + ?SUB_BIN_WORDSIZE),
-	     check_and_untag_fixnum(Size, SizeReg, FalseLblName),
+	     get_word_integer(Size, SizeReg, SystemLimitLblName, FalseLblName),
 	     allocate_writable(DstVar, Base, SizeReg, Zero, Zero),
 	     hipe_rtl:mk_goto(TrueLblName)]; 
 	  
@@ -305,7 +316,7 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 	    SubBinSize = {sub_binary, binsize},
 	    [get_field_from_term({sub_binary, orig}, Bin, ProcBin),
 	     get_field_from_term(SubBinSize, Bin, SubSize),
-	     check_and_untag_fixnum(Size, SizeReg, FalseLblName),
+	     get_word_integer(Size, SizeReg, SystemLimitLblName, FalseLblName),
 	     realloc_binary(SizeReg, ProcBin, Base),
 	     calculate_sizes(Bin, SizeReg, Offset, EndSubSize, EndSubBitSize),
 	     set_field_from_term(SubBinSize, Bin, EndSubSize),
@@ -313,20 +324,21 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 	     hipe_rtl:mk_move(DstVar, Bin),
 	     hipe_rtl:mk_goto(TrueLblName)]; 
 
-	  {bs_append, _U, _F, _B, _Bla} -> 
+	  {bs_append, _U, _F, Unit, _Bla} ->
 	    [Size, Bin] = Args, 
 	    [DstVar, Base, Offset] = Dst, 
 	    [ProcBin] = create_vars(1),
 	    [Flags, SizeReg, IsWritable, EndSubSize, EndSubBitSize] = 
 	      create_regs(5),
-	    [ContLbl,ContLbl2,ContLbl3,WritableLbl,NotWritableLbl] = Lbls =
-	      create_lbls(5),
-	    [ContLblName, ContLbl2Name, ContLbl3Name, Writable, NotWritable] =
+	    [ContLbl,ContLbl2,ContLbl3,ContLbl4,WritableLbl,NotWritableLbl] =
+	      Lbls = create_lbls(6),
+	    [ContLblName, ContLbl2Name, ContLbl3Name, ContLbl4Name,
+	     Writable, NotWritable] =
 	      [hipe_rtl:label_name(Lbl) || Lbl <- Lbls],
 	    Zero = hipe_rtl:mk_imm(0),
 	    SubIsWritable = {sub_binary, is_writable},
 	    [hipe_rtl:mk_gctest(?SUB_BIN_WORDSIZE + ?PROC_BIN_WORDSIZE),
-	     check_and_untag_fixnum(Size, SizeReg, FalseLblName),
+	     get_word_integer(Size, SizeReg, SystemLimitLblName, FalseLblName),
 	     hipe_tagscheme:test_bitstr(Bin, ContLblName, FalseLblName, 0.99), 
 	     ContLbl, 
 	     hipe_tagscheme:test_subbinary(Bin,ContLbl2Name, NotWritable), 
@@ -339,17 +351,19 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 	     get_field_from_term({proc_bin, flags}, ProcBin, Flags),
 	     hipe_rtl:mk_alub(Flags, Flags, 'and',
 			      hipe_rtl:mk_imm(?PB_IS_WRITABLE), 
-			      eq, NotWritable, Writable, 0.01),
+			      eq, NotWritable, ContLbl4Name, 0.01),
+	     ContLbl4,
+	     calculate_sizes(Bin, SizeReg, Offset, EndSubSize, EndSubBitSize),
+	     is_divisible(Offset, Unit, Writable, FalseLblName),
 	     WritableLbl,
 	     set_field_from_term(SubIsWritable, Bin, Zero),
 	     realloc_binary(SizeReg, ProcBin, Base),
-	     calculate_sizes(Bin, SizeReg, Offset, EndSubSize, EndSubBitSize),
 	     hipe_tagscheme:mk_sub_binary(DstVar, EndSubSize, Zero,
 					  EndSubBitSize, Zero,
 					  hipe_rtl:mk_imm(1), ProcBin),
 	     hipe_rtl:mk_goto(TrueLblName),
 	     NotWritableLbl,
-	     not_writable_code(Bin, SizeReg, DstVar, Base, Offset, 
+	     not_writable_code(Bin, SizeReg, DstVar, Base, Offset, Unit,
 			       TrueLblName, FalseLblName)]
 	end,
       {Code, ConstTab}
@@ -361,7 +375,7 @@ gen_rtl(BsOP, Dst, Args, TrueLblName, FalseLblName, SystemLimitLblName, ConstTab
 %%
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 
-not_writable_code(Bin, SizeReg, Dst, Base, Offset, 
+not_writable_code(Bin, SizeReg, Dst, Base, Offset, Unit,
 		  TrueLblName, FalseLblName) ->
   [SrcBase] = create_unsafe_regs(1),
   [SrcOffset, SrcSize, TotSize, TotBytes, UsedBytes] = create_regs(5),
@@ -372,13 +386,13 @@ not_writable_code(Bin, SizeReg, Dst, Base, Offset,
    hipe_rtl:mk_alu(TotBytes, TotSize, add, ?LOW_BITS),
    hipe_rtl:mk_alu(TotBytes, TotBytes, srl, ?BYTE_SHIFT),
    hipe_rtl:mk_alu(UsedBytes, TotBytes, sll, hipe_rtl:mk_imm(1)),
-   hipe_rtl:mk_branch(UsedBytes, ge, hipe_rtl:mk_imm(256), 
+   hipe_rtl:mk_branch(UsedBytes, geu, hipe_rtl:mk_imm(256),
 		      AllLblName, IncLblName),
    IncLbl,
    hipe_rtl:mk_move(UsedBytes, hipe_rtl:mk_imm(256)),
    AllLbl,
    allocate_writable(Dst, Base, UsedBytes, TotBytes, TotSize),
-   put_binary_all(Offset, Bin, Base, hipe_rtl:mk_imm(0), 
+   put_binary_all(Offset, Bin, Base, hipe_rtl:mk_imm(0), Unit,
 		  TrueLblName, FalseLblName)].
   
 allocate_writable(Dst, Base, UsedBytes, TotBytes, TotSize) ->
@@ -397,16 +411,6 @@ allocate_writable(Dst, Base, UsedBytes, TotBytes, TotSize) ->
    hipe_tagscheme:mk_sub_binary(Dst, EndSubSize, Zero, EndSubBitSize,
 				Zero, hipe_rtl:mk_imm(1), ProcBin)].
 
-check_and_untag_fixnum(Size, SizeReg, FalseLblName) ->
-  [ContLbl,NextLbl] = Lbls = create_lbls(2),
-  [ContLblName,NextLblName] = get_label_names(Lbls),
-  [hipe_tagscheme:test_fixnum(Size, ContLblName, FalseLblName, 0.99),
-   ContLbl,
-   hipe_tagscheme:untag_fixnum(SizeReg,Size),
-   hipe_rtl:mk_branch(SizeReg, ge, hipe_rtl:mk_imm(0), NextLblName, 
-		      FalseLblName),
-   NextLbl].
-
 realloc_binary(SizeReg, ProcBin, Base) -> 
   [NoReallocLbl, ReallocLbl, NextLbl, ContLbl] = Lbls = create_lbls(4),
   [NoReallocLblName, ReallocLblName, NextLblName, ContLblName] =
@@ -428,7 +432,7 @@ realloc_binary(SizeReg, ProcBin, Base) ->
    set_field_from_term(ProcBinFlagsTag, ProcBin, Flags),
    get_field_from_term(ProcBinValTag, ProcBin, BinPointer),
    get_field_from_pointer(BinOrigSizeTag, BinPointer, OrigSize),
-   hipe_rtl:mk_branch(OrigSize, 'lt', ResultingSize, 
+   hipe_rtl:mk_branch(OrigSize, 'ltu', ResultingSize,
 		      ReallocLblName, NoReallocLblName),
    NoReallocLbl,
    get_field_from_term(ProcBinBytesTag, ProcBin, Base),
@@ -669,14 +673,14 @@ var_init2(Size, Dst, Base, Offset, TrueLblName, SystemLimitLblName, FalseLblName
   WordSize = hipe_rtl_arch:word_size(),
   [ContLbl,HeapLbl,REFCLbl,NextLbl] = create_lbls(4),
   [USize,Tmp] = create_unsafe_regs(2),
-  [get_32_bit_value(Size, USize, SystemLimitLblName, FalseLblName),
-   hipe_rtl:mk_branch(USize, le, hipe_rtl:mk_imm(?MAX_BINSIZE), 
-		      hipe_rtl:label_name(ContLbl), 
-		      SystemLimitLblName),
+  [get_word_integer(Size, USize, SystemLimitLblName, FalseLblName),
+   hipe_rtl:mk_branch(USize, leu, hipe_rtl:mk_imm(?MAX_BINSIZE),
+                     hipe_rtl:label_name(ContLbl),
+                     SystemLimitLblName),
    ContLbl,
    hipe_rtl:mk_move(Offset, hipe_rtl:mk_imm(0)),
-   hipe_rtl:mk_branch(USize, le, hipe_rtl:mk_imm(?MAX_HEAP_BIN_SIZE), 
-		      hipe_rtl:label_name(HeapLbl), 
+   hipe_rtl:mk_branch(USize, leu, hipe_rtl:mk_imm(?MAX_HEAP_BIN_SIZE),
+		      hipe_rtl:label_name(HeapLbl),
 		      hipe_rtl:label_name(REFCLbl)),
    HeapLbl,
    hipe_rtl:mk_alu(Tmp, USize, add, hipe_rtl:mk_imm(3*WordSize-1)),
@@ -694,8 +698,8 @@ var_init2(Size, Dst, Base, Offset, TrueLblName, SystemLimitLblName, FalseLblName
    hipe_rtl:mk_goto(TrueLblName)].
 
 var_init_bits(Size, Dst, Base, Offset, TrueLblName, SystemLimitLblName, FalseLblName) ->
-  [HeapLbl,REFCLbl,NextLbl,NoSubLbl,SubLbl,ContLbl,
-   NoCreateSubBin, CreateSubBin, JoinLbl, JoinLbl2] = create_lbls(10),
+  [HeapLbl,REFCLbl,NextLbl,NoSubLbl,SubLbl,
+   NoCreateSubBin, CreateSubBin, JoinLbl, JoinLbl2] = create_lbls(9),
   [USize,ByteSize,TotByteSize,OffsetBits] = create_regs(4),
   [TmpDst] = create_unsafe_regs(1),
   Log2WordSize = hipe_rtl_arch:log2_word_size(),
@@ -705,7 +709,7 @@ var_init_bits(Size, Dst, Base, Offset, TrueLblName, SystemLimitLblName, FalseLbl
 	       ?PROC_BIN_WORDSIZE) + ?SUB_BIN_WORDSIZE,
   Zero = hipe_rtl:mk_imm(0),
   [hipe_rtl:mk_gctest(MaximumWords),
-   get_32_bit_value(Size, USize, SystemLimitLblName, FalseLblName),
+   get_word_integer(Size, USize, SystemLimitLblName, FalseLblName),
    hipe_rtl:mk_alu(ByteSize, USize, srl, ?BYTE_SHIFT),
    hipe_rtl:mk_alub(OffsetBits, USize, 'and', ?LOW_BITS, eq, 
 		    hipe_rtl:label_name(NoSubLbl),
@@ -716,11 +720,7 @@ var_init_bits(Size, Dst, Base, Offset, TrueLblName, SystemLimitLblName, FalseLbl
    SubLbl,
    hipe_rtl:mk_alu(TotByteSize, ByteSize, 'add', hipe_rtl:mk_imm(1)),
    JoinLbl,
-   hipe_rtl:mk_branch(USize, le, hipe_rtl:mk_imm(?MAX_BINSIZE), 
-		      hipe_rtl:label_name(ContLbl), 
-		      SystemLimitLblName),
-   ContLbl,
-   hipe_rtl:mk_branch(TotByteSize, 'le', hipe_rtl:mk_imm(?MAX_HEAP_BIN_SIZE),
+   hipe_rtl:mk_branch(TotByteSize, 'leu', hipe_rtl:mk_imm(?MAX_HEAP_BIN_SIZE),
 		      hipe_rtl:label_name(HeapLbl), 
 		      hipe_rtl:label_name(REFCLbl)),
    HeapLbl,
@@ -743,13 +743,16 @@ var_init_bits(Size, Dst, Base, Offset, TrueLblName, SystemLimitLblName, FalseLbl
    hipe_rtl:mk_move(Dst, TmpDst),
    hipe_rtl:mk_goto(TrueLblName)].
 
-put_binary_all(NewOffset, Src, Base, Offset, TLName, FLName) -> 
+put_binary_all(NewOffset, Src, Base, Offset, Unit, TLName, FLName) ->
   [SrcBase,SrcOffset,NumBits] = create_regs(3),
+  [ContLbl] = create_lbls(1),
   CCode = binary_c_code(NewOffset, Src, Base, Offset, NumBits, TLName),
   AlignedCode = copy_aligned_bytes(SrcBase, SrcOffset, NumBits, Base, Offset, 
 				   NewOffset, TLName),
-  get_base_offset_size(Src, SrcBase, SrcOffset, NumBits,FLName) ++
-    test_alignment(SrcOffset, NumBits, Offset, AlignedCode, CCode).
+  [get_base_offset_size(Src, SrcBase, SrcOffset, NumBits,FLName),
+   is_divisible(NumBits, Unit, hipe_rtl:label_name(ContLbl), FLName),
+   ContLbl
+   |test_alignment(SrcOffset, NumBits, Offset, AlignedCode, CCode)].
 
 test_alignment(SrcOffset, NumBits, Offset, AlignedCode, CCode) ->
   [Tmp] = create_regs(1),  
@@ -976,7 +979,7 @@ copy_string(StringBase, StringSize, BinBase, BinOffset, NewOffset, TrueLblName)
 
 small_check(SizeVar, CopySize, FalseLblName) ->
   SuccessLbl = hipe_rtl:mk_new_label(),
-  [hipe_rtl:mk_branch(SizeVar, le, CopySize,
+  [hipe_rtl:mk_branch(SizeVar, leu, CopySize,
 		      hipe_rtl:label_name(SuccessLbl), FalseLblName),
    SuccessLbl].
 
@@ -1279,89 +1282,18 @@ copy_float_big(Base, Offset, NewOffset, Src, FalseLblName, TrueLblName, var) ->
   hipe_tagscheme:test_flonum(Src, hipe_rtl:label_name(SuccessLbl), FalseLblName, 0.99) ++
     [SuccessLbl|copy_float_big(Base, Offset, NewOffset, Src, FalseLblName, TrueLblName, pass)].
 
-make_size(1, BitsVar, FalseLblName) ->
-  [DstReg] = create_regs(1),
-  {first_part(BitsVar, DstReg, FalseLblName), DstReg};
-make_size(?BYTE_SIZE, BitsVar, FalseLblName) ->
-  [DstReg] = create_regs(1),
-  Code = 
-    first_part(BitsVar, DstReg, FalseLblName) ++
-    [hipe_rtl:mk_alu(DstReg, DstReg, 'sll', ?BYTE_SHIFT)],
-  {Code, DstReg};
-make_size(UnitImm, BitsVar, FalseLblName) ->
-  [DstReg] = create_regs(1),
-  UnitList = number2list(UnitImm),
-  Code = multiply_code(UnitList, BitsVar, DstReg, FalseLblName),
-  {Code, DstReg}.
-
-multiply_code(List=[Head|_Tail], Variable, Result, FalseLblName) ->
-  Test = set_high(Head),
-  Tmp1 = hipe_rtl:mk_new_reg(),
-  SuccessLbl = hipe_rtl:mk_new_label(),
-  Register = hipe_rtl:mk_new_reg(),
-  Code = [hipe_rtl:mk_move(Result, hipe_rtl:mk_imm(0))|
-	  first_part(Variable, Register, FalseLblName)]
-    ++
-    [hipe_rtl:mk_alub(Tmp1, Register, 'and', hipe_rtl:mk_imm(Test), 
-		      'eq', hipe_rtl:label_name(SuccessLbl), 
-		      FalseLblName, 0.99),
-     SuccessLbl],
-  multiply_code(List, Register, Result, FalseLblName, Tmp1, Code).
-
-multiply_code([ShiftSize|Rest], Register, Result, FalseLblName, Tmp1, OldCode) ->
-  SuccessLbl = hipe_rtl:mk_new_label(),
-  Code = OldCode ++ [hipe_rtl:mk_alu(Tmp1, Register, 'sll',
-				     hipe_rtl:mk_imm(ShiftSize)),
-		     hipe_rtl:mk_alub(Result, Tmp1, 'add', Result, not_overflow, hipe_rtl:label_name(SuccessLbl), FalseLblName, 0.99),
-		     SuccessLbl],
-  multiply_code(Rest, Register, Result, FalseLblName, Tmp1, Code);
-multiply_code([], _Register, _Result, _FalseLblName, _Tmp1, Code) ->
-  Code.
-
-number2list(X) when is_integer(X), X >= 0 ->
-  number2list(X, []).
-
-number2list(1, Acc) ->
-  lists:reverse([0|Acc]);
-number2list(0, Acc) ->
-  lists:reverse(Acc);
-number2list(X, Acc) ->
-  F = floorlog2(X),
-  number2list(X-(1 bsl F), [F|Acc]).
-
-floorlog2(X) ->
-  round(math:log(X)/math:log(2)-0.5). 
-
-set_high(X) ->
-  set_high(X, 0).
-
-set_high(0, Y) ->
-  Y;
-set_high(X, Y) ->
-  set_high(X-1, Y+(1 bsl (27-X))).
-
-get_32_bit_value(Size, USize, SystemLimitLblName, NegLblName) ->
-  Lbls = [FixLbl, BigLbl, OkLbl, PosBigLbl] = create_lbls(4),
-  [FixLblName, BigLblName, OkLblName, PosBigLblName] = [hipe_rtl:label_name(Lbl) || Lbl <- Lbls],
-  [hipe_tagscheme:test_fixnum(Size, FixLblName, BigLblName, 0.99),
-   FixLbl,
-   hipe_tagscheme:untag_fixnum(USize, Size),
-   hipe_rtl:mk_branch(USize, ge, hipe_rtl:mk_imm(0), OkLblName, NegLblName), 
-   BigLbl,
-   hipe_tagscheme:test_pos_bignum(Size, PosBigLblName, NegLblName, 0.99),
-   PosBigLbl,
-   hipe_tagscheme:get_one_word_pos_bignum(USize, Size, SystemLimitLblName),
-   OkLbl].  
-   
-
-first_part(Var, Register, FalseLblName) ->
-  [SuccessLbl1, SuccessLbl2] = create_lbls(2),
-  [hipe_tagscheme:test_fixnum(Var, hipe_rtl:label_name(SuccessLbl1),
-			      FalseLblName, 0.99),
-  SuccessLbl1,
-  hipe_tagscheme:fixnum_ge(Var, hipe_rtl:mk_imm(hipe_tagscheme:mk_fixnum(0)), 
-			   hipe_rtl:label_name(SuccessLbl2), FalseLblName, 0.99),
-  SuccessLbl2,
-  hipe_tagscheme:untag_fixnum(Register, Var)].
-
-
+is_divisible(_Dividend, 1, SuccLbl, _FailLbl) ->
+  [hipe_rtl:mk_goto(SuccLbl)];
+is_divisible(Dividend, Divisor, SuccLbl, FailLbl) ->
+  Log2 = floorlog2(Divisor),
+  case Divisor =:= 1 bsl Log2 of
+    true -> %% Divisor is a power of 2
+      %% Test that the Log2-1 lowest bits are clear
+      Mask = hipe_rtl:mk_imm(Divisor - 1),
+      [Tmp] = create_regs(1),
+      [hipe_rtl:mk_alub(Tmp, Dividend, 'and', Mask, eq, SuccLbl, FailLbl, 0.99)];
+    false ->
+      %% We need division, fall back to a primop
+      [hipe_rtl:mk_call([], is_divisible, [Dividend, hipe_rtl:mk_imm(Divisor)],
+			SuccLbl, FailLbl, not_remote)]
+  end.
diff --git a/lib/hipe/rtl/hipe_rtl_binary_match.erl b/lib/hipe/rtl/hipe_rtl_binary_match.erl
index cf1b8e453f..be4c35dae0 100644
--- a/lib/hipe/rtl/hipe_rtl_binary_match.erl
+++ b/lib/hipe/rtl/hipe_rtl_binary_match.erl
@@ -31,11 +31,12 @@
 
 -import(hipe_tagscheme, [set_field_from_term/3, get_field_from_term/3]).
 
+-import(hipe_rtl_binary, [make_size/3]).
+
 -include("hipe_literals.hrl").
 
 %%--------------------------------------------------------------------
 
--define(MAX_BINSIZE, trunc(?MAX_HEAP_BIN_SIZE / hipe_rtl_arch:word_size()) + 2).
 -define(BYTE_SHIFT, 3). %% Turn bits into bytes or vice versa
 -define(LOW_BITS, 7). %% Three lowest bits set 
 -define(BYTE_SIZE, 8).
@@ -835,10 +836,10 @@ create_lbls(0) ->
 create_lbls(X) when X > 0 ->
   [hipe_rtl:mk_new_label()|create_lbls(X-1)].
 
-make_dyn_prep(SizeReg, CCode) ->   
+make_dyn_prep(SizeReg, CCode) ->
   [CLbl, SuccessLbl] = create_lbls(2),
-  Init = [hipe_rtl:mk_branch(SizeReg, le, hipe_rtl:mk_imm(?MAX_SMALL_BITS),  
-			     hipe_rtl:label_name(SuccessLbl), 
+  Init = [hipe_rtl:mk_branch(SizeReg, leu, hipe_rtl:mk_imm(?MAX_SMALL_BITS),
+			     hipe_rtl:label_name(SuccessLbl),
 			     hipe_rtl:label_name(CLbl)),
 	  SuccessLbl],
   End = [CLbl|CCode],
@@ -883,8 +884,8 @@ get_unaligned_int_to_reg(Reg, Size, Base, Offset, LowBits, Shiftr, Type) ->
 			 hipe_rtl:mk_imm((WordSize-MinLoad)*?BYTE_SIZE))];
       {_, WordSize} ->
 	UnsignedBig = {unsigned, big},
-	[hipe_rtl:mk_branch(TotBits, le, hipe_rtl:mk_imm(MinLoad*?BYTE_SIZE), 
-			    hipe_rtl:label_name(LessLbl), 
+	[hipe_rtl:mk_branch(TotBits, leu, hipe_rtl:mk_imm(MinLoad*?BYTE_SIZE),
+			    hipe_rtl:label_name(LessLbl),
 			    hipe_rtl:label_name(MoreLbl)),
 	 LessLbl,
 	 load_bytes(LoadDst, Base, ByteOffset, Type, MinLoad),
@@ -944,7 +945,7 @@ get_big_unknown_int(Dst1, Base, Offset, NewOffset,
    hipe_rtl:mk_alu(ByteOffset, Offset, srl, hipe_rtl:mk_imm(?BYTE_SHIFT)),
    load_bytes(LoadDst, Base, ByteOffset, Type, 1),
    BackLbl,
-   hipe_rtl:mk_branch(ByteOffset, le, Limit, hipe_rtl:label_name(LoopLbl), 
+   hipe_rtl:mk_branch(ByteOffset, leu, Limit, hipe_rtl:label_name(LoopLbl),
 		      hipe_rtl:label_name(EndLbl)),
    LoopLbl,
    load_bytes(Tmp, Base, ByteOffset, {unsigned, big}, 1),
@@ -973,8 +974,8 @@ get_little_unknown_int(Dst1, Base, Offset, NewOffset,
    hipe_rtl:mk_alu(Limit, Tmp, srl, hipe_rtl:mk_imm(?BYTE_SHIFT)),
    hipe_rtl:mk_move(ShiftReg, hipe_rtl:mk_imm(0)),
    BackLbl,
-   hipe_rtl:mk_branch(ByteOffset, lt, Limit, 
-		      hipe_rtl:label_name(LoopLbl), 
+   hipe_rtl:mk_branch(ByteOffset, ltu, Limit,
+		      hipe_rtl:label_name(LoopLbl),
 		      hipe_rtl:label_name(DoneLbl)),
    LoopLbl,
    load_bytes(Tmp, Base, ByteOffset, {unsigned, big}, 1),
@@ -1090,7 +1091,7 @@ load_bytes(Dst, Base, Offset, {Signedness, Endianness}, X) when X > 1 ->
        hipe_rtl:mk_alu(Offset, Offset, add, hipe_rtl:mk_imm(1)),
        hipe_rtl:mk_alu(Dst, Dst, sll, hipe_rtl:mk_imm(8)),
        hipe_rtl:mk_alu(Dst, Dst, 'or', Tmp1),
-       hipe_rtl:mk_branch(Offset, lt, Limit, hipe_rtl:label_name(LoopLbl),
+       hipe_rtl:mk_branch(Offset, ltu, Limit, hipe_rtl:label_name(LoopLbl),
 			  hipe_rtl:label_name(EndLbl)),
        EndLbl];
     little ->
@@ -1102,7 +1103,7 @@ load_bytes(Dst, Base, Offset, {Signedness, Endianness}, X) when X > 1 ->
        hipe_rtl:mk_load(Tmp1, Base, TmpOffset, byte, Signedness),
        hipe_rtl:mk_alu(Dst, Dst, sll, hipe_rtl:mk_imm(8)),
        hipe_rtl:mk_alu(Dst, Dst, 'or', Tmp1),
-       hipe_rtl:mk_branch(Offset, lt, TmpOffset, hipe_rtl:label_name(LoopLbl),
+       hipe_rtl:mk_branch(Offset, ltu, TmpOffset, hipe_rtl:label_name(LoopLbl),
 			  hipe_rtl:label_name(EndLbl)),
        EndLbl,
        hipe_rtl:mk_move(Offset, Limit)]
@@ -1118,103 +1119,5 @@ create_gcsafe_regs(X) when X > 0 ->
 create_gcsafe_regs(0) ->
   [].
 
-first_part(Var, Register, FalseLblName) ->
-  [EndLbl] = create_lbls(1),
-  EndName = hipe_rtl:label_name(EndLbl),
-  first_part(Var, Register, FalseLblName, EndName, EndName, [EndLbl]).
-
-first_part(Var, Register, FalseLblName, TrueLblName, BigLblName, Tail) ->
-  [FixnumLbl, NotFixnumLbl, BignumLbl, SuccessLbl] = create_lbls(4),
-  [hipe_tagscheme:test_fixnum(Var, hipe_rtl:label_name(FixnumLbl),
-			      hipe_rtl:label_name(NotFixnumLbl), 0.99),
-   FixnumLbl,
-   hipe_tagscheme:fixnum_ge(Var, hipe_rtl:mk_imm(hipe_tagscheme:mk_fixnum(0)),
-			    hipe_rtl:label_name(SuccessLbl), FalseLblName,
-			    0.99),
-   SuccessLbl,
-   hipe_tagscheme:untag_fixnum(Register, Var),
-   hipe_rtl:mk_goto(TrueLblName),
-   NotFixnumLbl,
-   %% Since binaries are not allowed to be larger than 2^wordsize bits
-   %% and since bignum digits are words, we know that a bignum with an
-   %% arity larger than one can't match.
-   hipe_tagscheme:test_pos_bignum_arity(Var, 1, hipe_rtl:label_name(BignumLbl),
-					FalseLblName, 0.99),
-   BignumLbl,
-   hipe_tagscheme:unsafe_get_one_word_pos_bignum(Register, Var),
-   hipe_rtl:mk_goto(BigLblName) | Tail].
-
-make_size(1, BitsVar, FalseLblName) ->
-  [DstReg] = create_regs(1),
-  {first_part(BitsVar, DstReg, FalseLblName), DstReg};
-make_size(?BYTE_SIZE, BitsVar, FalseLblName) ->
-  [DstReg] = create_regs(1),
-  [FixnumLbl, BignumLbl] = create_lbls(2),
-  WordBits = hipe_rtl_arch:word_size() * ?BYTE_SIZE,
-  FixnumLblName = hipe_rtl:label_name(FixnumLbl),
-  Tail = [BignumLbl,
-	  hipe_rtl:mk_branch(DstReg, 'ltu',
-			     hipe_rtl:mk_imm(1 bsl (WordBits - ?BYTE_SHIFT)),
-			     FixnumLblName, FalseLblName, 0.99),
-	  FixnumLbl,
-	  hipe_rtl:mk_alu(DstReg, DstReg, sll, hipe_rtl:mk_imm(?BYTE_SHIFT))],
-  Code = first_part(BitsVar, DstReg, FalseLblName, FixnumLblName,
-		    hipe_rtl:label_name(BignumLbl), Tail),
-  {Code, DstReg};
-make_size(UnitImm, BitsVar, FalseLblName) ->
-  [DstReg] = create_regs(1),
-  UnitList = number2list(UnitImm),
-  Code = multiply_code(UnitList, BitsVar, DstReg, FalseLblName),
-  {Code, DstReg}.
-
-multiply_code(List=[Head|_Tail], Variable, Result, FalseLblName) ->
-  Test = set_high(Head),
-  Tmp1 = hipe_rtl:mk_new_reg(),
-  SuccessLbl = hipe_rtl:mk_new_label(),
-  Register = hipe_rtl:mk_new_reg(),
-  Code = [hipe_rtl:mk_move(Result, hipe_rtl:mk_imm(0))|
-	  first_part(Variable, Register, FalseLblName)]
-	++
-	 [hipe_rtl:mk_alub(Tmp1, Register, 'and', hipe_rtl:mk_imm(Test), 
-			   eq, hipe_rtl:label_name(SuccessLbl), 
-			   FalseLblName, 0.99),
-	  SuccessLbl],
-  multiply_code(List, Register, Result, FalseLblName, Tmp1, Code).
-
-multiply_code([ShiftSize|Rest], Register, Result, FalseLblName, Tmp1, OldCode) ->
-  SuccessLbl = hipe_rtl:mk_new_label(),
-  Code =
-    OldCode ++
-    [hipe_rtl:mk_alu(Tmp1, Register, sll, hipe_rtl:mk_imm(ShiftSize)),
-     hipe_rtl:mk_alub(Result, Tmp1, 'add', Result, not_overflow,
-		      hipe_rtl:label_name(SuccessLbl), FalseLblName, 0.99),
-     SuccessLbl],
-  multiply_code(Rest, Register, Result, FalseLblName, Tmp1, Code);
-multiply_code([], _Register, _Result, _FalseLblName, _Tmp1, Code) ->
-  Code.
-
-number2list(X) when is_integer(X), X >= 0 ->
-  number2list(X, []).
-
-number2list(1, Acc) ->
-  lists:reverse([0|Acc]);
-number2list(0, Acc) ->
-  lists:reverse(Acc);
-number2list(X, Acc) ->
-  F = floorlog2(X),
-  number2list(X-(1 bsl F), [F|Acc]).
-
-floorlog2(X) ->
-  round(math:log(X)/math:log(2)-0.5). 
-
-set_high(X) ->
-  WordBits = hipe_rtl_arch:word_size() * ?BYTE_SIZE,
-  set_high(min(X, WordBits), WordBits, 0).
-
-set_high(0, _, Y) ->
-  Y;
-set_high(X, WordBits, Y) ->
-  set_high(X-1, WordBits, Y+(1 bsl (WordBits-X))).
-
 is_illegal_const(Const) ->
   Const >=  1 bsl (hipe_rtl_arch:word_size() * ?BYTE_SIZE) orelse Const < 0.
diff --git a/lib/hipe/rtl/hipe_tagscheme.erl b/lib/hipe/rtl/hipe_tagscheme.erl
index d77078acb6..8825a3ade3 100644
--- a/lib/hipe/rtl/hipe_tagscheme.erl
+++ b/lib/hipe/rtl/hipe_tagscheme.erl
@@ -42,7 +42,7 @@
 	 test_ref/4, test_fun/4, test_fun2/5, test_matchstate/4,
 	 test_binary/4, test_bitstr/4, test_list/4, test_map/4,
 	 test_integer/4, test_number/4, test_tuple_N/5,
-	 test_pos_bignum_arity/5]).
+	 test_pos_bignum_arity/6]).
 -export([realtag_fixnum/2, tag_fixnum/2, realuntag_fixnum/2, untag_fixnum/2]).
 -export([test_two_fixnums/3, test_fixnums/4, unsafe_fixnum_add/3,
 	 unsafe_fixnum_sub/3,
@@ -351,14 +351,23 @@ test_pos_bignum(X, TrueLab, FalseLab, Pred) ->
    mask_and_compare(Tmp, BigMask, ?TAG_HEADER_POS_BIG,
 		    TrueLab, FalseLab, Pred)].
 
-test_pos_bignum_arity(X, Arity, TrueLab, FalseLab, Pred) ->
+test_pos_bignum_arity(X, Arity, TrueLab, NotPosBignumLab, FalseLab, Pred) ->
   Tmp = hipe_rtl:mk_new_reg_gcsafe(),
-  HalfTrueLab = hipe_rtl:mk_new_label(),
+  BoxedLab = hipe_rtl:mk_new_label(),
   HeaderImm = hipe_rtl:mk_imm(mk_header(Arity, ?TAG_HEADER_POS_BIG)),
-  [test_is_boxed(X, hipe_rtl:label_name(HalfTrueLab), FalseLab, Pred),
-   HalfTrueLab,
-   get_header(Tmp, X),
-   hipe_rtl:mk_branch(Tmp, 'eq', HeaderImm, TrueLab, FalseLab, Pred)].
+  [test_is_boxed(X, hipe_rtl:label_name(BoxedLab), NotPosBignumLab, Pred),
+   BoxedLab,
+   get_header(Tmp, X)] ++
+    case NotPosBignumLab =:= FalseLab of
+      true -> [];
+      false ->
+	BignumLab = hipe_rtl:mk_new_label(),
+	BigMask = ?TAG_HEADER_MASK,
+	[mask_and_compare(Tmp, BigMask, ?TAG_HEADER_POS_BIG,
+			  hipe_rtl:label_name(BignumLab), NotPosBignumLab, Pred),
+	 BignumLab]
+    end ++
+    [hipe_rtl:mk_branch(Tmp, 'eq', HeaderImm, TrueLab, FalseLab, Pred)].
 
 test_matchstate(X, TrueLab, FalseLab, Pred) ->
   Tmp = hipe_rtl:mk_new_reg_gcsafe(),
diff --git a/lib/hipe/test/bs_SUITE_data/bs_add.erl b/lib/hipe/test/bs_SUITE_data/bs_add.erl
index af5a3b2f23..4b92e6b413 100644
--- a/lib/hipe/test/bs_SUITE_data/bs_add.erl
+++ b/lib/hipe/test/bs_SUITE_data/bs_add.erl
@@ -2,7 +2,7 @@
 %%-------------------------------------------------------------------------
 %% The guard in f/3 revealed a problem in the translation of the 'bs_add'
 %% BEAM instruction to Icode. The fail label was not properly translated.
-%% Fixed 3/2/2011.
+%% Fixed 3/2/2011. Then in 2015 we found another issue: g/2.  Also fixed.
 %%-------------------------------------------------------------------------
 -module(bs_add).
 
@@ -10,9 +10,17 @@
 
 test() ->
   42 = f(<<12345:16>>, 4711, <<42>>),
+  true = g(<<1:13>>, 3),  %% was handled OK, but
+  false = g(<<>>, gurka), %% this one leaked badarith
   ok.
 
 f(Bin, A, B) when <<A:9, B:7/binary>> == Bin ->
   gazonk;
 f(Bin, _, _) when is_binary(Bin) ->
   42.
+
+%% Complex way of testing (bit_size(Bin) + Len) rem 8 =:= 0
+g(Bin, Len) when is_binary(<<Bin/binary-unit:1, 123:Len>>) ->
+    true;
+g(_, _) ->
+    false.
diff --git a/lib/hipe/test/bs_SUITE_data/bs_construct.erl b/lib/hipe/test/bs_SUITE_data/bs_construct.erl
index 37a54c1981..b9e7d93570 100644
--- a/lib/hipe/test/bs_SUITE_data/bs_construct.erl
+++ b/lib/hipe/test/bs_SUITE_data/bs_construct.erl
@@ -14,6 +14,11 @@ test() ->
   16#10000008 = bit_size(large_bin(1, 2, 3, 4)),
   ok = bad_ones(),
   ok = zero_width(),
+  ok = not_used(),
+  ok = bad_append(),
+  ok = system_limit(),
+  ok = bad_floats(),
+  ok = huge_binaries(),
   ok.
 
 %%--------------------------------------------------------------------
@@ -142,3 +147,159 @@ zero_width() ->
   ok.
 
 id(X) -> X.
+
+%%--------------------------------------------------------------------
+%% Taken from bs_construct_SUITE. The test checks that constructed
+%% binaries that are not used would still give a `badarg' exception.
+%% Problem was that in native code one of them gave `badarith'.
+
+not_used() ->
+    ok = not_used1(3, <<"dum">>),
+    {'EXIT',{badarg,_}} = (catch not_used1(42, "dum_string")),
+    {'EXIT',{badarg,_}} = (catch not_used2(666, -2)),
+    {'EXIT',{badarg,_}} = (catch not_used2(666, "bad_size")), % this one
+    {'EXIT',{badarg,_}} = (catch not_used3(666)),
+    ok.
+
+not_used1(I, BinString) ->
+    <<I:32,BinString/binary>>,
+    ok.
+
+not_used2(I, Sz) ->
+    <<I:Sz>>,
+    ok.
+
+not_used3(I) ->
+    <<I:(-8)>>,
+    ok.
+
+%%--------------------------------------------------------------------
+%% Taken from bs_construct_SUITE.
+
+bad_append() ->
+    do_bad_append(<<127:1>>, fun append_unit_3/1),
+    do_bad_append(<<127:2>>, fun append_unit_3/1),
+    do_bad_append(<<127:17>>, fun append_unit_3/1),
+
+    do_bad_append(<<127:3>>, fun append_unit_4/1),
+    do_bad_append(<<127:5>>, fun append_unit_4/1),
+    do_bad_append(<<127:7>>, fun append_unit_4/1),
+    do_bad_append(<<127:199>>, fun append_unit_4/1),
+
+    do_bad_append(<<127:7>>, fun append_unit_8/1),
+    do_bad_append(<<127:9>>, fun append_unit_8/1),
+
+    do_bad_append(<<0:8>>, fun append_unit_16/1),
+    do_bad_append(<<0:15>>, fun append_unit_16/1),
+    do_bad_append(<<0:17>>, fun append_unit_16/1),
+    ok.
+
+do_bad_append(Bin0, Appender) ->
+    {'EXIT',{badarg,_}} = (catch Appender(Bin0)),
+
+    Bin1 = id(<<0:3,Bin0/bitstring>>),
+    <<_:3,Bin2/bitstring>> = Bin1,
+    {'EXIT',{badarg,_}} = (catch Appender(Bin2)),
+
+    %% Create a writable binary.
+    Empty = id(<<>>),
+    Bin3 = <<Empty/bitstring,Bin0/bitstring>>,
+    {'EXIT',{badarg,_}} = (catch Appender(Bin3)),
+    ok.
+
+append_unit_3(Bin) ->
+    <<Bin/binary-unit:3,0:1>>.
+
+append_unit_4(Bin) ->
+    <<Bin/binary-unit:4,0:1>>.
+
+append_unit_8(Bin) ->
+    <<Bin/binary,0:1>>.
+
+append_unit_16(Bin) ->
+    <<Bin/binary-unit:16,0:1>>.
+
+%%--------------------------------------------------------------------
+%% Taken from bs_construct_SUITE.
+
+system_limit() ->
+    WordSize = erlang:system_info(wordsize),
+    BitsPerWord = WordSize * 8,
+    {'EXIT',{system_limit,_}} =
+	(catch <<0:(id(0)),42:(id(1 bsl BitsPerWord))>>),
+    {'EXIT',{system_limit,_}} =
+	(catch <<42:(id(1 bsl BitsPerWord)),0:(id(0))>>),
+    {'EXIT',{system_limit,_}} =
+	(catch <<(id(<<>>))/binary,0:(id(1 bsl 100))>>),
+
+    %% Would fail to load.
+    {'EXIT',{system_limit,_}} = (catch <<0:(1 bsl 67)>>),
+    {'EXIT',{system_limit,_}} = (catch <<0:((1 bsl 64)+1)>>),
+    case WordSize of
+	4 ->
+	    system_limit_32();
+	8 ->
+	    ok
+    end.
+
+system_limit_32() ->
+    {'EXIT',{badarg,_}} = (catch <<42:(-1)>>),
+    {'EXIT',{badarg,_}} = (catch <<42:(id(-1))>>),
+    {'EXIT',{badarg,_}} = (catch <<42:(id(-389739873536870912))/unit:8>>),
+    {'EXIT',{system_limit,_}} = (catch <<42:536870912/unit:8>>),
+    {'EXIT',{system_limit,_}} = (catch <<42:(id(536870912))/unit:8>>),
+    {'EXIT',{system_limit,_}} = (catch <<0:(id(8)),42:536870912/unit:8>>),
+    {'EXIT',{system_limit,_}} =	(catch <<0:(id(8)),42:(id(536870912))/unit:8>>),
+
+    %% The size would be silently truncated, resulting in a crash.
+    {'EXIT',{system_limit,_}} = (catch <<0:(1 bsl 35)>>),
+    {'EXIT',{system_limit,_}} = (catch <<0:((1 bsl 32)+1)>>),
+
+    %% Would fail to load.
+    {'EXIT',{system_limit,_}} = (catch <<0:(1 bsl 43)>>),
+    {'EXIT',{system_limit,_}} = (catch <<0:((1 bsl 40)+1)>>),
+    ok.
+
+%%--------------------------------------------------------------------
+
+bad_floats() ->
+  WordSize = erlang:system_info(wordsize),
+  BitsPerWord = WordSize * 8,
+  {'EXIT',{badarg,_}} = (catch <<3.14:(id(33))/float>>),
+  {'EXIT',{badarg,_}} = (catch <<3.14:(id(64 bor 32))/float>>),
+  {'EXIT',{badarg,_}} = (catch <<3.14:(id((1 bsl 28) bor 32))/float>>),
+  {'EXIT',{system_limit,_}} = (catch <<3.14:(id(1 bsl BitsPerWord))/float>>),
+  ok.
+
+%%--------------------------------------------------------------------
+%% A bug in the implementation of binaries compared sizes in bits with sizes in
+%% bytes, causing <<0:(id((1 bsl 31)-1))>> to fail to construct with
+%% 'system_limit'.
+%% <<0:(id((1 bsl 32)-1))>> was succeeding because the comparison was
+%% (incorrectly) signed.
+
+huge_binaries() ->
+  AlmostIllegal = id(<<0:(id((1 bsl 32)-8))>>),
+  case erlang:system_info(wordsize) of
+    4 -> huge_binaries_32(AlmostIllegal);
+    8 -> ok
+  end,
+  garbage_collect(),
+  id(<<0:(id((1 bsl 31)-1))>>),
+  id(<<0:(id((1 bsl 30)-1))>>),
+  garbage_collect(),
+  ok.
+
+huge_binaries_32(AlmostIllegal) ->
+  %% Attempt construction of too large binary using bs_init/1 (which takes the
+  %% number of bytes as an argument, which should be compared to the maximum
+  %% size in bytes).
+  {'EXIT',{system_limit,_}} = (catch <<0:32,AlmostIllegal/binary>>),
+  %% Attempt construction of too large binary using bs_init/1 with a size in
+  %% bytes that has the msb set (and would be negative if it was signed).
+  {'EXIT',{system_limit,_}} =
+    (catch <<0:8, AlmostIllegal/binary, AlmostIllegal/binary,
+	     AlmostIllegal/binary, AlmostIllegal/binary,
+	     AlmostIllegal/binary, AlmostIllegal/binary,
+	     AlmostIllegal/binary, AlmostIllegal/binary>>),
+  ok.