hipe_rtl: unify branch and alub

branch and alub overlap in their use cases, but the backends rely on
knowing that the result is unused in their lowering of branch. By
extending alub so that the destination is optional, it can fully replace
branch.

This simplifies rtl by reducing code duplication and the number of
instructions.

Also, in the x86 and arm backends, we can now use 'test' and
{'tst','mvn','teq'} to lower some alubs without destinations. This is
particularly good for x86, as sequences such as 'is_boxed' type tests
now get shorter (both from not needing a mov to copy the variable, but
also from the fact that 'testb' encodes shorter than 'andq').

14 files changed, 269 insertions, 432 deletions

diff --git a/lib/hipe/arm/hipe_rtl_to_arm.erl b/lib/hipe/arm/hipe_rtl_to_arm.erl
index 2f9181d517..c964c222aa 100644
--- a/lib/hipe/arm/hipe_rtl_to_arm.erl
+++ b/lib/hipe/arm/hipe_rtl_to_arm.erl
@@ -62,7 +62,6 @@ conv_insn(I, Map, Data) ->
   case I of
     #alu{} -> conv_alu(I, Map, Data);
     #alub{} -> conv_alub(I, Map, Data);
-    #branch{} -> conv_branch(I, Map, Data);
     #call{} -> conv_call(I, Map, Data);
     #comment{} -> conv_comment(I, Map, Data);
     #enter{} -> conv_enter(I, Map, Data);
@@ -111,6 +110,17 @@ commute_arithop(ArithOp) ->
     _ -> ArithOp
   end.
 
+conv_cmpop('add') -> 'cmn';
+conv_cmpop('sub') -> 'cmp';
+conv_cmpop('and') -> 'tst';
+conv_cmpop('xor') -> 'teq';
+conv_cmpop(_) -> none.
+
+cmpop_commutes('cmp') -> false;
+cmpop_commutes('cmn') -> true;
+cmpop_commutes('tst') -> true;
+cmpop_commutes('teq') -> true.
+
 mk_alu(S, Dst, Src1, RtlAluOp, Src2) ->
   case hipe_rtl:is_shift_op(RtlAluOp) of
     true ->
@@ -223,71 +233,77 @@ fix_aluop_imm(AluOp, Imm) -> % {FixAm1,NewAluOp,Am1}
 
 conv_alub(I, Map, Data) ->
   %% dst = src1 aluop src2; if COND goto label
-  {Dst, Map0} = conv_dst(hipe_rtl:alub_dst(I), Map),
-  {Src1, Map1} = conv_src(hipe_rtl:alub_src1(I), Map0),
-  {Src2, Map2} = conv_src(hipe_rtl:alub_src2(I), Map1),
+  {Src1, Map0} = conv_src(hipe_rtl:alub_src1(I), Map),
+  {Src2, Map1} = conv_src(hipe_rtl:alub_src2(I), Map0),
   RtlAluOp = hipe_rtl:alub_op(I),
-  Cond0 = conv_alub_cond(RtlAluOp, hipe_rtl:alub_cond(I)),
-  Cond =
-    case {RtlAluOp,Cond0} of
-      {'mul','vs'} -> 'ne';	% overflow becomes not-equal
-      {'mul','vc'} -> 'eq';	% no-overflow becomes equal
-      {'mul',_} -> exit({?MODULE,I});
-      {_,_} -> Cond0
-    end,
-  I2 = mk_pseudo_bc(
-	  Cond,
-	  hipe_rtl:alub_true_label(I),
-	  hipe_rtl:alub_false_label(I),
-	  hipe_rtl:alub_pred(I)),
-  S = true,
-  I1 = mk_alu(S, Dst, Src1, RtlAluOp, Src2),
-  {I1 ++ I2, Map2, Data}.
-
-conv_branch(I, Map, Data) ->
-  %% <unused> = src1 - src2; if COND goto label
-  {Src1, Map0} = conv_src(hipe_rtl:branch_src1(I), Map),
-  {Src2, Map1} = conv_src(hipe_rtl:branch_src2(I), Map0),
-  Cond = conv_branch_cond(hipe_rtl:branch_cond(I)),
-  I2 = mk_branch(Src1, Cond, Src2,
-		 hipe_rtl:branch_true_label(I),
-		 hipe_rtl:branch_false_label(I),
-		 hipe_rtl:branch_pred(I)),
-  {I2, Map1, Data}.
+  RtlCond = hipe_rtl:alub_cond(I),
+  HasDst = hipe_rtl:alub_has_dst(I),
+  CmpOp = conv_cmpop(RtlAluOp),
+  Cond0 = conv_alub_cond(RtlAluOp, RtlCond),
+  case (not HasDst) andalso CmpOp =/= none of
+    true ->
+      I1 = mk_branch(Src1, CmpOp, Src2, Cond0,
+		     hipe_rtl:alub_true_label(I),
+		     hipe_rtl:alub_false_label(I),
+		     hipe_rtl:alub_pred(I)),
+      {I1, Map1, Data};
+    false ->
+      {Dst, Map2} =
+	case HasDst of
+	  false -> {new_untagged_temp(), Map1};
+	  true -> conv_dst(hipe_rtl:alub_dst(I), Map1)
+	end,
+      Cond =
+	case {RtlAluOp,Cond0} of
+	  {'mul','vs'} -> 'ne';	% overflow becomes not-equal
+	  {'mul','vc'} -> 'eq';	% no-overflow becomes equal
+	  {'mul',_} -> exit({?MODULE,I});
+	  {_,_} -> Cond0
+	end,
+      I2 = mk_pseudo_bc(
+	     Cond,
+	     hipe_rtl:alub_true_label(I),
+	     hipe_rtl:alub_false_label(I),
+	     hipe_rtl:alub_pred(I)),
+      S = true,
+      I1 = mk_alu(S, Dst, Src1, RtlAluOp, Src2),
+      {I1 ++ I2, Map2, Data}
+  end.
 
-mk_branch(Src1, Cond, Src2, TrueLab, FalseLab, Pred) ->
+mk_branch(Src1, CmpOp, Src2, Cond, TrueLab, FalseLab, Pred) ->
   case hipe_arm:is_temp(Src1) of
     true ->
       case hipe_arm:is_temp(Src2) of
 	true ->
-	  mk_branch_rr(Src1, Src2, Cond, TrueLab, FalseLab, Pred);
+	  mk_branch_rr(Src1, CmpOp, Src2, Cond, TrueLab, FalseLab, Pred);
 	_ ->
-	  mk_branch_ri(Src1, Cond, Src2, TrueLab, FalseLab, Pred)
+	  mk_branch_ri(Src1, CmpOp, Src2, Cond, TrueLab, FalseLab, Pred)
       end;
     _ ->
       case hipe_arm:is_temp(Src2) of
 	true ->
-	  NewCond = commute_cond(Cond),
-	  mk_branch_ri(Src2, NewCond, Src1, TrueLab, FalseLab, Pred);
+	  NewCond =
+	    case cmpop_commutes(CmpOp) of
+	      true -> Cond;
+	      false ->  commute_cond(Cond)
+	    end,
+	  mk_branch_ri(Src2, CmpOp, Src1, NewCond, TrueLab, FalseLab, Pred);
 	_ ->
-	  mk_branch_ii(Src1, Cond, Src2, TrueLab, FalseLab, Pred)
+	  mk_branch_ii(Src1, CmpOp, Src2, Cond, TrueLab, FalseLab, Pred)
       end
   end.
 
-mk_branch_ii(Imm1, Cond, Imm2, TrueLab, FalseLab, Pred) ->
+mk_branch_ii(Imm1, CmpOp, Imm2, Cond, TrueLab, FalseLab, Pred) ->
   Tmp = new_untagged_temp(),
   mk_li(Tmp, Imm1,
-	mk_branch_ri(Tmp, Cond, Imm2,
+	mk_branch_ri(Tmp, CmpOp, Imm2, Cond,
 		     TrueLab, FalseLab, Pred)).
 
-mk_branch_ri(Src, Cond, Imm, TrueLab, FalseLab, Pred) ->
-  {FixAm1,NewCmpOp,Am1} = fix_aluop_imm('cmp', Imm),
-  FixAm1 ++ mk_cmp_bc(NewCmpOp, Src, Am1, Cond, TrueLab, FalseLab, Pred).
-
-mk_branch_rr(Src1, Src2, Cond, TrueLab, FalseLab, Pred) ->
-  mk_cmp_bc('cmp', Src1, Src2, Cond, TrueLab, FalseLab, Pred).
+mk_branch_ri(Src, CmpOp, Imm, Cond, TrueLab, FalseLab, Pred) ->
+  {FixAm1,NewCmpOp,Am1} = fix_aluop_imm(CmpOp, Imm),
+  FixAm1 ++ mk_branch_rr(Src, NewCmpOp, Am1, Cond, TrueLab, FalseLab, Pred).
 
-mk_cmp_bc(CmpOp, Src, Am1, Cond, TrueLab, FalseLab, Pred) ->
+mk_branch_rr(Src, CmpOp, Am1, Cond, TrueLab, FalseLab, Pred) ->
   [hipe_arm:mk_cmp(CmpOp, Src, Am1) |
    mk_pseudo_bc(Cond, TrueLab, FalseLab, Pred)].
 
@@ -637,6 +653,7 @@ conv_alub_cond(RtlAluOp, Cond) ->	% may be unsigned, depends on aluop
   case {RtlAluOp, Cond} of	% handle allowed alub unsigned conditions
     {'add', 'ltu'} -> 'hs';	% add+ltu == unsigned overflow == carry set == hs
     %% add more cases when needed
+    {'sub', _} -> conv_branch_cond(Cond);
     _ -> conv_cond(Cond)
   end.
 
diff --git a/lib/hipe/llvm/hipe_rtl_to_llvm.erl b/lib/hipe/llvm/hipe_rtl_to_llvm.erl
index 20813f8bd7..f8911c1909 100644
--- a/lib/hipe/llvm/hipe_rtl_to_llvm.erl
+++ b/lib/hipe/llvm/hipe_rtl_to_llvm.erl
@@ -156,9 +156,6 @@ translate_instr(I, Relocs, Data) ->
     #alub{} ->
       {I2, Relocs2} = trans_alub(I, Relocs),
       {I2, Relocs2, Data};
-    #branch{} ->
-      {I2, Relocs2} = trans_branch(I, Relocs),
-      {I2, Relocs2, Data};
     #call{} ->
       {I2, Relocs2} =
         case hipe_rtl:call_fun(I) of
@@ -255,7 +252,6 @@ trans_alub(I, Relocs) ->
 trans_alub_overflow(I, Sign, Relocs) ->
   {Src1, I1} = trans_src(hipe_rtl:alub_src1(I)),
   {Src2, I2} = trans_src(hipe_rtl:alub_src2(I)),
-  RtlDst = hipe_rtl:alub_dst(I),
   TmpDst = mk_temp(),
   Name = trans_alub_op(I, Sign),
   NewRelocs = relocs_store(Name, {call, remote, {llvm, Name, 2}}, Relocs),
@@ -266,7 +262,10 @@ trans_alub_overflow(I, Sign, Relocs) ->
 			                   [{WordTy, Src1}, {WordTy, Src2}], []),
   %% T1{0}: result of the operation
   I4 = hipe_llvm:mk_extractvalue(TmpDst, ReturnType, T1 , "0", []),
-  I5 = store_stack_dst(TmpDst, RtlDst),
+  I5 = case hipe_rtl:alub_has_dst(I) of
+	 false -> [];
+	 true -> store_stack_dst(TmpDst, hipe_rtl:alub_dst(I))
+       end,
   T2 = mk_temp(),
   %% T1{1}: Boolean variable indicating overflow
   I6 = hipe_llvm:mk_extractvalue(T2, ReturnType, T1, "1", []),
@@ -311,42 +310,35 @@ trans_alub_op(I, Sign) ->
   Name ++ Type.
 
 trans_alub_no_overflow(I, Relocs) ->
+  {Src1, I1} = trans_src(hipe_rtl:alub_src1(I)),
+  {Src2, I2} = trans_src(hipe_rtl:alub_src2(I)),
+  WordTy = hipe_llvm:mk_int(?BITS_IN_WORD),
   %% alu
-  T = hipe_rtl:mk_alu(hipe_rtl:alub_dst(I), hipe_rtl:alub_src1(I),
-                      hipe_rtl:alub_op(I), hipe_rtl:alub_src2(I)),
-  %% A trans_alu instruction cannot change relocations
-  {I1, _} = trans_alu(T, Relocs),
+  {CmpLhs, CmpRhs, I5, Cond} =
+    case {hipe_rtl:alub_has_dst(I), hipe_rtl:alub_op(I)} of
+      {false, 'sub'} ->
+	Cond0 = trans_branch_rel_op(hipe_rtl:alub_cond(I)),
+	{Src1, Src2, [], Cond0};
+      {HasDst, AlubOp} ->
+	TmpDst = mk_temp(),
+	Op = trans_op(AlubOp),
+	I3 = hipe_llvm:mk_operation(TmpDst, Op, WordTy, Src1, Src2, []),
+	I4 = case HasDst of
+	       false -> [];
+	       true -> store_stack_dst(TmpDst, hipe_rtl:alub_dst(I))
+	     end,
+	Cond0 = trans_alub_rel_op(hipe_rtl:alub_cond(I)),
+	{TmpDst, "0", [I4, I3], Cond0}
+    end,
   %% icmp
-  %% Translate destination as src, to match with the semantics of instruction
-  {Dst, I2} = trans_src(hipe_rtl:alub_dst(I)),
-  Cond = trans_rel_op(hipe_rtl:alub_cond(I)),
   T3 = mk_temp(),
-  WordTy = hipe_llvm:mk_int(?BITS_IN_WORD),
-  I5 = hipe_llvm:mk_icmp(T3, Cond, WordTy, Dst, "0"),
+  I6 = hipe_llvm:mk_icmp(T3, Cond, WordTy, CmpLhs, CmpRhs),
   %% br
   Metadata = branch_metadata(hipe_rtl:alub_pred(I)),
   True_label = mk_jump_label(hipe_rtl:alub_true_label(I)),
   False_label = mk_jump_label(hipe_rtl:alub_false_label(I)),
-  I6 = hipe_llvm:mk_br_cond(T3, True_label, False_label, Metadata),
-  {[I6, I5, I2, I1], Relocs}.
-
-%%
-%% branch
-%%
-trans_branch(I, Relocs) ->
-  {Src1, I1} = trans_src(hipe_rtl:branch_src1(I)),
-  {Src2, I2} = trans_src(hipe_rtl:branch_src2(I)),
-  Cond = trans_rel_op(hipe_rtl:branch_cond(I)),
-  %% icmp
-  T1 = mk_temp(),
-  WordTy = hipe_llvm:mk_int(?BITS_IN_WORD),
-  I3 = hipe_llvm:mk_icmp(T1, Cond, WordTy, Src1, Src2),
-  %% br
-  True_label = mk_jump_label(hipe_rtl:branch_true_label(I)),
-  False_label = mk_jump_label(hipe_rtl:branch_false_label(I)),
-  Metadata = branch_metadata(hipe_rtl:branch_pred(I)),
-  I4 = hipe_llvm:mk_br_cond(T1, True_label, False_label, Metadata),
-  {[I4, I3, I2, I1], Relocs}.
+  I7 = hipe_llvm:mk_br_cond(T3, True_label, False_label, Metadata),
+  {[I7, I6, I5, I2, I1], Relocs}.
 
 branch_metadata(X) when X =:= 0.5 -> [];
 branch_metadata(X) when X > 0.5 -> ?BRANCH_META_TAKEN;
@@ -1162,7 +1154,7 @@ trans_dst(A) ->
 		       true ->
 			 "%DL" ++ integer_to_list(hipe_rtl:const_label_label(A)) ++ "_var";
 		       false ->
-			 exit({?MODULE, trans_dst, {"Bad RTL argument",A}})
+			 error(badarg, [A])
 		     end
 		 end
 	     end,
@@ -1260,14 +1252,19 @@ trans_op(Op) ->
     Other -> exit({?MODULE, trans_op, {"Unknown RTL operator", Other}})
   end.
 
-trans_rel_op(Op) ->
+trans_branch_rel_op(Op) ->
   case Op of
-    eq -> eq;
-    ne -> ne;
     gtu -> ugt;
     geu -> uge;
     ltu -> ult;
     leu -> ule;
+    _ -> trans_alub_rel_op(Op)
+  end.
+
+trans_alub_rel_op(Op) ->
+  case Op of
+    eq -> eq;
+    ne -> ne;
     gt -> sgt;
     ge -> sge;
     lt -> slt;
@@ -1300,7 +1297,10 @@ insn_dst(I) ->
     #alu{} ->
       [hipe_rtl:alu_dst(I)];
     #alub{} ->
-      [hipe_rtl:alub_dst(I)];
+      case hipe_rtl:alub_has_dst(I) of
+	true -> [hipe_rtl:alub_dst(I)];
+	false -> []
+      end;
     #call{} ->
       case hipe_rtl:call_dstlist(I) of
         [] -> [];
diff --git a/lib/hipe/ppc/hipe_rtl_to_ppc.erl b/lib/hipe/ppc/hipe_rtl_to_ppc.erl
index a01e67a789..09f1ce5a49 100644
--- a/lib/hipe/ppc/hipe_rtl_to_ppc.erl
+++ b/lib/hipe/ppc/hipe_rtl_to_ppc.erl
@@ -80,7 +80,6 @@ conv_insn(I, Map, Data) ->
   case I of
     #alu{} -> conv_alu(I, Map, Data);
     #alub{} -> conv_alub(I, Map, Data);
-    #branch{} -> conv_branch(I, Map, Data);
     #call{} -> conv_call(I, Map, Data);
     #comment{} -> conv_comment(I, Map, Data);
     #enter{} -> conv_enter(I, Map, Data);
@@ -441,36 +440,53 @@ mk_alu_rr(Dst, Src1, RtlAluOp, Src2) ->
 
 conv_alub(I, Map, Data) ->
   %% dst = src1 aluop src2; if COND goto label
-  {Dst, Map0} = conv_dst(hipe_rtl:alub_dst(I), Map),
-  {Src1, Map1} = conv_src(hipe_rtl:alub_src1(I), Map0),
-  {Src2, Map2} = conv_src(hipe_rtl:alub_src2(I), Map1),
-  {AluOp, BCond} =
-    case {hipe_rtl:alub_op(I), hipe_rtl:alub_cond(I)} of
-      {'add', 'ltu'} ->
-	{'addc', 'eq'};
-      {RtlAlubOp, RtlAlubCond} ->
-	{conv_alub_op(RtlAlubOp), conv_alub_cond(RtlAlubCond)}
-    end,
-  BC = mk_pseudo_bc(BCond,
-		    hipe_rtl:alub_true_label(I),
-		    hipe_rtl:alub_false_label(I),
-		    hipe_rtl:alub_pred(I)),
-  I2 =
-    case {AluOp, BCond} of
-      {'addc', 'eq'} ->	% copy XER[CA] to CR0[EQ] before the BC
-	TmpR = new_untagged_temp(),
-	[hipe_ppc:mk_mfspr(TmpR, 'xer'),
-	 hipe_ppc:mk_mtcr(TmpR) |
-	 BC];
-      _ -> BC
-    end,
-  {NewSrc1, NewSrc2} =
-    case AluOp of
-      'subf' -> {Src2, Src1};
-      _ -> {Src1, Src2}
-    end,
-  I1 = mk_alub(Dst, NewSrc1, AluOp, NewSrc2, BCond),
-  {I1 ++ I2, Map2, Data}.
+  HasDst = hipe_rtl:alub_has_dst(I),
+  {Src1, Map0} = conv_src(hipe_rtl:alub_src1(I), Map),
+  {Src2, Map1} = conv_src(hipe_rtl:alub_src2(I), Map0),
+  RtlAlubOp = hipe_rtl:alub_op(I),
+  RtlAlubCond = hipe_rtl:alub_cond(I),
+  case {HasDst, RtlAlubOp} of
+    {false, sub} ->
+      {BCond,Sign} = conv_branch_cond(RtlAlubCond),
+      I2 = mk_branch(Src1, BCond, Sign, Src2,
+		     hipe_rtl:alub_true_label(I),
+		     hipe_rtl:alub_false_label(I),
+		     hipe_rtl:alub_pred(I)),
+      {I2, Map1, Data};
+    _ ->
+      {Dst, Map2} =
+	case HasDst of
+	  false -> {new_untagged_temp(), Map1};
+	  true -> conv_dst(hipe_rtl:alub_dst(I), Map1)
+	end,
+      {AluOp, BCond} =
+	case {RtlAlubOp, RtlAlubCond} of
+	  {'add', 'ltu'} ->
+	    {'addc', 'eq'};
+	  {_, _} ->
+	    {conv_alub_op(RtlAlubOp), conv_alub_cond(RtlAlubCond)}
+	end,
+      BC = mk_pseudo_bc(BCond,
+			hipe_rtl:alub_true_label(I),
+			hipe_rtl:alub_false_label(I),
+			hipe_rtl:alub_pred(I)),
+      I2 =
+	case {AluOp, BCond} of
+	  {'addc', 'eq'} ->	% copy XER[CA] to CR0[EQ] before the BC
+	    TmpR = new_untagged_temp(),
+	    [hipe_ppc:mk_mfspr(TmpR, 'xer'),
+	     hipe_ppc:mk_mtcr(TmpR) |
+	     BC];
+	  _ -> BC
+	end,
+      {NewSrc1, NewSrc2} =
+	case AluOp of
+	  'subf' -> {Src2, Src1};
+	  _ -> {Src1, Src2}
+	end,
+      I1 = mk_alub(Dst, NewSrc1, AluOp, NewSrc2, BCond),
+      {I1 ++ I2, Map2, Data}
+  end.
 
 conv_alub_op(RtlAluOp) ->
   case {get(hipe_target_arch), RtlAluOp} of
@@ -689,17 +705,6 @@ mk_alub_rr_Rc(Dst, Src1, AluOp, Src2) ->
     end,
   [hipe_ppc:mk_alu(AluOpDot, Dst, Src1, Src2)].
 
-conv_branch(I, Map, Data) ->
-  %% <unused> = src1 - src2; if COND goto label
-  {Src1, Map0} = conv_src(hipe_rtl:branch_src1(I), Map),
-  {Src2, Map1} = conv_src(hipe_rtl:branch_src2(I), Map0),
-  {BCond,Sign} = conv_branch_cond(hipe_rtl:branch_cond(I)),
-  I2 = mk_branch(Src1, BCond, Sign, Src2,
-		 hipe_rtl:branch_true_label(I),
-		 hipe_rtl:branch_false_label(I),
-		 hipe_rtl:branch_pred(I)),
-  {I2, Map1, Data}.
-
 conv_branch_cond(Cond) -> % may be unsigned
   case Cond of
     gtu -> {'gt', 'unsigned'};
diff --git a/lib/hipe/rtl/hipe_rtl.erl b/lib/hipe/rtl/hipe_rtl.erl
index 0726827299..e4e857ec43 100644
--- a/lib/hipe/rtl/hipe_rtl.erl
+++ b/lib/hipe/rtl/hipe_rtl.erl
@@ -187,18 +187,14 @@
 
 	 mk_branch/5,
 	 mk_branch/6,
-	 branch_src1/1,
-	 branch_src2/1,
-	 branch_cond/1,
-	 branch_true_label/1,
-	 branch_false_label/1,
-	 branch_pred/1,
+	 mk_branch/7,
 	 %% is_branch/1,
 	 %% branch_true_label_update/2,
 	 %% branch_false_label_update/2,
 
 	 mk_alub/7,
 	 mk_alub/8,
+	 alub_has_dst/1,
 	 alub_dst/1,
 	 alub_src1/1,
 	 alub_op/1,
@@ -588,37 +584,25 @@ is_label(#label{}) -> true;
 is_label(_) -> false.
 
 %%
-%% branch
-%%
-
-mk_branch(Src1, Op, Src2, True, False) ->
-  mk_branch(Src1, Op, Src2, True, False, 0.5).
-mk_branch(Src1, Op, Src2, True, False, P) ->
-  #branch{src1=Src1, 'cond'=Op, src2=Src2, true_label=True,
-	  false_label=False, p=P}.
-branch_src1(#branch{src1=Src1}) -> Src1.
-branch_src1_update(Br, NewSrc) -> Br#branch{src1=NewSrc}.
-branch_src2(#branch{src2=Src2}) -> Src2.
-branch_src2_update(Br, NewSrc) -> Br#branch{src2=NewSrc}.
-branch_cond(#branch{'cond'=Cond}) -> Cond.
-branch_true_label(#branch{true_label=TrueLbl}) -> TrueLbl.
-branch_true_label_update(Br, NewTrue) -> Br#branch{true_label=NewTrue}.
-branch_false_label(#branch{false_label=FalseLbl}) -> FalseLbl.
-branch_false_label_update(Br, NewFalse) -> Br#branch{false_label=NewFalse}.
-branch_pred(#branch{p=P}) -> P.
-
-%%
 %% alub
 %%
 
 -type alub_cond() :: 'eq' | 'ne' | 'ge' | 'geu' | 'gt' | 'gtu' | 'le'
                    | 'leu' | 'lt' | 'ltu' | 'overflow' | 'not_overflow'.
 
+mk_branch(Src1, Cond, Src2, True, False) ->
+  mk_branch(Src1, Cond, Src2, True, False, 0.5).
+mk_branch(Src1, Cond, Src2, True, False, P) ->
+  mk_branch(Src1, 'sub', Src2, Cond, True, False, P).
+mk_branch(Src1, Op, Src2, Cond, True, False, P) ->
+  mk_alub([], Src1, Op, Src2, Cond, True, False, P).
+
 mk_alub(Dst, Src1, Op, Src2, Cond, True, False) ->
   mk_alub(Dst, Src1, Op, Src2, Cond, True, False, 0.5).
 mk_alub(Dst, Src1, Op, Src2, Cond, True, False, P) ->
   #alub{dst=Dst, src1=Src1, op=Op, src2=Src2, 'cond'=Cond,
 	true_label=True, false_label=False, p=P}.
+alub_has_dst(#alub{dst=Dst}) -> Dst =/= [].
 alub_dst(#alub{dst=Dst}) -> Dst.
 alub_dst_update(A, NewDst) -> A#alub{dst=NewDst}.
 alub_src1(#alub{src1=Src1}) -> Src1.
@@ -943,8 +927,7 @@ args(I) ->
   case I of
     #alu{} -> [alu_src1(I), alu_src2(I)];
     #alub{} -> [alub_src1(I), alub_src2(I)];
-    #branch{} -> [branch_src1(I), branch_src2(I)];
-    #call{} -> 
+    #call{} ->
       Args = call_arglist(I) ++ hipe_rtl_arch:call_used(),
       case call_is_known(I) of
 	false -> [call_fun(I) | Args];
@@ -987,8 +970,8 @@ args(I) ->
 defines(Instr) ->
   Defs = case Instr of
 	   #alu{} -> [alu_dst(Instr)];
+	   #alub{dst=[]} -> [];
 	   #alub{} -> [alub_dst(Instr)];
-	   #branch{} -> [];
 	   #call{} -> call_dstlist(Instr) ++ hipe_rtl_arch:call_defined();
 	   #comment{} -> [];
 	   #enter{} -> [];
@@ -1042,9 +1025,6 @@ subst_uses(Subst, I) ->
     #alub{} ->
       I0 = alub_src1_update(I, subst1(Subst, alub_src1(I))),
       alub_src2_update(I0, subst1(Subst, alub_src2(I)));
-    #branch{} ->
-      I0 = branch_src1_update(I, subst1(Subst, branch_src1(I))),
-      branch_src2_update(I0, subst1(Subst, branch_src2(I)));
     #call{} ->
       case call_is_known(I) of
 	false ->
@@ -1126,11 +1106,6 @@ subst_uses_llvm(Subst, I) ->
       {NewSrc1, _ } = subst1_llvm(Subst1, alub_src1(I)),
       I0 =  alub_src1_update(I, NewSrc1),
       alub_src2_update(I0, NewSrc2);
-    #branch{} ->
-      {NewSrc2, Subst1} = subst1_llvm(Subst, branch_src2(I)),
-      {NewSrc1, _ } = subst1_llvm(Subst1, branch_src1(I)),
-      I0 = branch_src1_update(I, NewSrc1),
-      branch_src2_update(I0, NewSrc2);
     #call{} ->
       case call_is_known(I) of
         false ->
@@ -1243,10 +1218,10 @@ subst_defines(Subst, I)->
   case I of
     #alu{} ->
       alu_dst_update(I, subst1(Subst, alu_dst(I)));
+    #alub{dst=[]} ->
+      I;
     #alub{} ->
       alub_dst_update(I, subst1(Subst, alub_dst(I)));
-    #branch{} ->
-      I;
     #call{} ->
       call_dstlist_update(I, subst_list(Subst, call_dstlist(I)));
     #comment{} ->
@@ -1313,7 +1288,6 @@ is_safe(Instr) ->
   case Instr of
     #alu{} -> true;
     #alub{} -> false;
-    #branch{} -> false;
     #call{} -> false;
     #comment{} -> false;
     #enter{} -> false;
@@ -1386,17 +1360,6 @@ redirect_jmp(Jmp, ToOld, ToNew) ->
   %% OBS: In a jmp instruction more than one labels may be identical
   %%      and thus need redirection!
   case Jmp of
-    #branch{} ->
-      TmpJmp = case branch_true_label(Jmp) of
-		 ToOld -> branch_true_label_update(Jmp, ToNew);
-		 _ -> Jmp
-	       end,
-      case branch_false_label(TmpJmp) of
-	ToOld ->
-	  branch_false_label_update(TmpJmp, ToNew);
-	_ ->
-	  TmpJmp
-      end;
     #switch{} ->
       NewLbls = [case Lbl =:= ToOld of
 		   true -> ToNew;
@@ -1591,13 +1554,6 @@ pp_instr(Dev, I) ->
       io:format(Dev, "~n", []);
     #label{} ->
       io:format(Dev, "L~w:~n", [label_name(I)]);
-    #branch{} ->
-      io:format(Dev, "    if (", []),
-      pp_arg(Dev, branch_src1(I)),
-      io:format(Dev, " ~w ", [branch_cond(I)]),
-      pp_arg(Dev, branch_src2(I)),
-      io:format(Dev, ") then L~w (~.2f) else L~w~n", 
-		[branch_true_label(I), branch_pred(I), branch_false_label(I)]);
     #switch{} ->
       io:format(Dev, "    switch (", []),
       pp_arg(Dev, switch_src(I)),
@@ -1606,7 +1562,10 @@ pp_instr(Dev, I) ->
       io:format(Dev, ">\n", []);
     #alub{} ->
       io:format(Dev, "    ", []),
-      pp_arg(Dev, alub_dst(I)),
+      case alub_has_dst(I) of
+	true -> pp_arg(Dev, alub_dst(I));
+	false -> io:format(Dev, "_", [])
+      end,
       io:format(Dev, " <- ", []),
       pp_arg(Dev, alub_src1(I)),
       io:format(Dev, " ~w ", [alub_op(I)]),
diff --git a/lib/hipe/rtl/hipe_rtl.hrl b/lib/hipe/rtl/hipe_rtl.hrl
index cc76e7e5c4..74020c6045 100644
--- a/lib/hipe/rtl/hipe_rtl.hrl
+++ b/lib/hipe/rtl/hipe_rtl.hrl
@@ -28,7 +28,6 @@
 
 -record(alu, {dst, src1, op, src2}).
 -record(alub, {dst, src1, op, src2, 'cond', true_label, false_label, p}).
--record(branch, {src1, src2, 'cond', true_label, false_label, p}).
 -record(call, {dstlist, 'fun', arglist, type, continuation,
     failcontinuation, normalcontinuation = []}).
 -record(comment, {text}).
diff --git a/lib/hipe/rtl/hipe_rtl_binary_construct.erl b/lib/hipe/rtl/hipe_rtl_binary_construct.erl
index 367d76b24d..2922972085 100644
--- a/lib/hipe/rtl/hipe_rtl_binary_construct.erl
+++ b/lib/hipe/rtl/hipe_rtl_binary_construct.erl
@@ -757,9 +757,9 @@ test_alignment(SrcOffset, NumBits, Offset, AlignedCode, CCode) ->
   [AlignedLbl, CLbl] = create_lbls(2),
    [hipe_rtl:mk_alu(Tmp, SrcOffset, 'or', NumBits),
    hipe_rtl:mk_alu(Tmp, Tmp, 'or', Offset),
-   hipe_rtl:mk_alub(Tmp, Tmp, 'and', ?LOW_BITS, 'eq',
-		    hipe_rtl:label_name(AlignedLbl),
-		    hipe_rtl:label_name(CLbl)),
+   hipe_rtl:mk_branch(Tmp, 'and', ?LOW_BITS, 'eq',
+		      hipe_rtl:label_name(AlignedLbl),
+		      hipe_rtl:label_name(CLbl), 0.5),
    AlignedLbl,
    AlignedCode,
    CLbl,
@@ -1284,8 +1284,7 @@ is_divisible(Dividend, Divisor, SuccLbl, FailLbl) ->
     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)];
+      [hipe_rtl:mk_branch(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)],
diff --git a/lib/hipe/rtl/hipe_rtl_binary_match.erl b/lib/hipe/rtl/hipe_rtl_binary_match.erl
index d999cd2743..520b055ba7 100644
--- a/lib/hipe/rtl/hipe_rtl_binary_match.erl
+++ b/lib/hipe/rtl/hipe_rtl_binary_match.erl
@@ -659,9 +659,8 @@ test_alignment_code(Size, Unit, SLblName, FalseLblName) ->
   end.
 
 get_fast_test_code(Size, AndTest, SLblName, FalseLblName) ->
-  [Tmp] = create_gcsafe_regs(1),
-  [hipe_rtl:mk_alub(Tmp, Size, 'and', hipe_rtl:mk_imm(AndTest),
-		    'eq', SLblName, FalseLblName)].
+  [hipe_rtl:mk_branch(Size, 'and', hipe_rtl:mk_imm(AndTest), 'eq',
+		      SLblName, FalseLblName, 0.5)].
 
 %% This is really slow
 get_slow_test_code(Size, Unit, SLblName, FalseLblName) ->
diff --git a/lib/hipe/rtl/hipe_rtl_cfg.erl b/lib/hipe/rtl/hipe_rtl_cfg.erl
index f49e8f815f..e802b320c2 100644
--- a/lib/hipe/rtl/hipe_rtl_cfg.erl
+++ b/lib/hipe/rtl/hipe_rtl_cfg.erl
@@ -83,9 +83,7 @@ mk_goto(Name) ->
 
 branch_successors(Instr) ->
   case Instr of
-    #branch{} -> [hipe_rtl:branch_true_label(Instr), 
-		  hipe_rtl:branch_false_label(Instr)];
-    #alub{} -> [hipe_rtl:alub_true_label(Instr), 
+    #alub{} -> [hipe_rtl:alub_true_label(Instr),
 	        hipe_rtl:alub_false_label(Instr)];
     #switch{} -> hipe_rtl:switch_labels(Instr);
     #call{} -> 
@@ -106,7 +104,6 @@ fails_to(Instr) ->
 
 is_branch(Instr) ->
    case Instr of
-     #branch{} -> true;
      #alub{} -> true;
      #switch{} -> true;
      #goto{} -> true;
@@ -127,7 +124,7 @@ is_branch(Instr) ->
 
 is_pure_branch(Instr) ->
   case Instr of
-    #branch{} -> true;
+    #alub{} -> not hipe_rtl:alub_has_dst(Instr);
     #switch{} -> true;
     #goto{} -> true;
     _ -> false
diff --git a/lib/hipe/rtl/hipe_rtl_lcm.erl b/lib/hipe/rtl/hipe_rtl_lcm.erl
index 71bd06c0df..67ddd0f649 100644
--- a/lib/hipe/rtl/hipe_rtl_lcm.erl
+++ b/lib/hipe/rtl/hipe_rtl_lcm.erl
@@ -378,7 +378,6 @@ is_expr(I) ->
 %% 	  end;
 	       
         #alub{} -> false; %% TODO: Split instruction to consider alu expression?
-        #branch{} -> false;
         #call{} -> false; %% We cannot prove that a call has no side-effects
         #comment{} -> false;
         #enter{} -> false;
diff --git a/lib/hipe/rtl/hipe_rtl_ssa_const_prop.erl b/lib/hipe/rtl/hipe_rtl_ssa_const_prop.erl
index 7158383010..f887eeab66 100644
--- a/lib/hipe/rtl/hipe_rtl_ssa_const_prop.erl
+++ b/lib/hipe/rtl/hipe_rtl_ssa_const_prop.erl
@@ -110,8 +110,6 @@ visit_expression(Instruction, Environment) ->
       visit_alu(Instruction, Environment);
     #alub{} ->
       visit_alub(Instruction, Environment);
-    #branch{} ->
-      visit_branch(Instruction, Environment);
     #call{} ->
       visit_call(Instruction, Environment);
 %%    #comment{} ->
@@ -184,42 +182,6 @@ set_to(Dst, Val, Env) ->
   {[], SSAWork, Env1}.
 
 %%-----------------------------------------------------------------------------
-%% Procedure : visit_branch/2
-%% Purpose   : do symbolic exection of branch instructions.
-%% Arguments : Inst - The instruction
-%%             Env  - The environment
-%% Returns   : { FlowWorkList, SSAWorkList, NewEnvironment}
-%%-----------------------------------------------------------------------------
-
-visit_branch(Inst, Env) -> %% Titta också på exekverbarflagga
-  Val1 = lookup_lattice_value(hipe_rtl:branch_src1(Inst), Env),
-  Val2 = lookup_lattice_value(hipe_rtl:branch_src2(Inst), Env),
-  CFGWL = case evaluate_relop(Val1, hipe_rtl:branch_cond(Inst), Val2) of
-            true   -> [hipe_rtl:branch_true_label(Inst)];
-            false  -> [hipe_rtl:branch_false_label(Inst)];
-            bottom -> [hipe_rtl:branch_true_label(Inst), 
-	               hipe_rtl:branch_false_label(Inst)];
-            top    -> []
-          end,
-  {CFGWL, [], Env}.
-
-%%-----------------------------------------------------------------------------
-%% Procedure : evaluate_relop/3
-%% Purpose   : evaluate the given relop. While taking care to handle top & 
-%%             bottom in some sane way.
-%% Arguments : Val1, Val2 - The operands Integers or top or bottom
-%%             RelOp  - some relop atom from rtl. 
-%% Returns   : bottom, top, true or false
-%%-----------------------------------------------------------------------------
-
-evaluate_relop(Val1, RelOp, Val2) ->
-  if 
-    (Val1==bottom) or (Val2==bottom) -> bottom ;
-    (Val1==top) or (Val2==top)       ->  top;
-    true ->  hipe_rtl_arch:eval_cond(RelOp, Val1, Val2)
-  end.
-
-%%-----------------------------------------------------------------------------
 %% Procedure : evaluate_fixnumop/2 
 %% Purpose   : try to evaluate a fixnumop.
 %% Arguments : Val1 - operand (an integer, 'top' or 'bottom')
@@ -408,6 +370,7 @@ partial_eval_branch(Cond, N0, Z0, V0, C0) ->
        Cond =:= 'ne'           -> {true, Z0,   true, true};
        Cond =:= 'gt';
        Cond =:= 'le'           -> {N0,   Z0,   V0,   true};
+       Cond =:= 'leu';
        Cond =:= 'gtu'          -> {true, Z0,   true, C0  };
        Cond =:= 'lt';
        Cond =:= 'ge'           -> {N0,   true, V0,   true};
@@ -450,7 +413,11 @@ visit_alub(Inst, Env) ->
           false  -> [hipe_rtl:alub_false_label(Inst)]
         end
      end,
-  {[], NewSSA, NewEnv} = set_to(hipe_rtl:alub_dst(Inst), NewVal,  Env),
+  {[], NewSSA, NewEnv} =
+    case hipe_rtl:alub_has_dst(Inst) of
+      false -> {[], [], Env};
+      true -> set_to(hipe_rtl:alub_dst(Inst), NewVal, Env)
+    end,
   {Labels, NewSSA, NewEnv}.
       
 %%-----------------------------------------------------------------------------
@@ -688,8 +655,6 @@ update_instruction(Inst, Env) ->
       update_alu(Inst, Env);
     #alub{} ->
       update_alub(Inst, Env);
-    #branch{} ->
-      update_branch(Inst, Env);
     #call{} ->
       subst_all_uses(Inst, Env);
 %%    #comment{} ->
@@ -902,33 +867,6 @@ update_alu(Inst, Env) ->
       {Val,_,_,_,_} = evaluate_alu(Val1, hipe_rtl:alu_op(Inst), Val2),
       [hipe_rtl:mk_move(hipe_rtl:alu_dst(Inst), hipe_rtl:mk_imm(Val))]
   end.
- 
-%%-----------------------------------------------------------------------------
-%% Procedure : update_branch/2
-%% Purpose   : update an branch-instruction
-%% Arguments : Inst - the instruction.
-%%             Env - in which everything happens.
-%% Returns   : list of new instruction
-%%-----------------------------------------------------------------------------
-
-update_branch(Inst, Env) ->
-  Src1 = hipe_rtl:branch_src1(Inst),
-  Src2 = hipe_rtl:branch_src2(Inst),
-  Val1 = lookup_lattice_value(Src1, Env),
-  Val2 = lookup_lattice_value(Src2, Env),
-  if
-    (Val1 =:= bottom) and (Val2 =:= bottom) ->
-      [Inst];
-    Val1 =:= bottom ->
-      [hipe_rtl:subst_uses([{Src2, hipe_rtl:mk_imm(Val2)}], Inst)];
-    Val2 =:= bottom -> 
-      [hipe_rtl:subst_uses([{Src1, hipe_rtl:mk_imm(Val1)}], Inst)];
-    true ->
-      case hipe_rtl_arch:eval_cond(hipe_rtl:branch_cond(Inst), Val1, Val2) of
-        true  -> [hipe_rtl:mk_goto(hipe_rtl:branch_true_label(Inst))];
-        false -> [hipe_rtl:mk_goto(hipe_rtl:branch_false_label(Inst))]
-      end
-  end.
 
 %%-----------------------------------------------------------------------------
 %% Procedure : update_alub/2
@@ -943,8 +881,12 @@ update_branch(Inst, Env) ->
 
 %% some small helpers.
 alub_to_move(Inst, Res, Lab) ->
-  [hipe_rtl:mk_move(hipe_rtl:alub_dst(Inst), Res),
-   hipe_rtl:mk_goto(Lab)].
+  Goto = [hipe_rtl:mk_goto(Lab)],
+  case hipe_rtl:alub_has_dst(Inst) of
+    false -> Goto;
+    true ->
+      [hipe_rtl:mk_move(hipe_rtl:alub_dst(Inst), Res) | Goto]
+  end.
 
 make_alub_subst_list(bottom, _, Tail) ->  Tail;
 make_alub_subst_list(top, Src, _) ->
diff --git a/lib/hipe/rtl/hipe_tagscheme.erl b/lib/hipe/rtl/hipe_tagscheme.erl
index 8cf45772b5..8d9514ae82 100644
--- a/lib/hipe/rtl/hipe_tagscheme.erl
+++ b/lib/hipe/rtl/hipe_tagscheme.erl
@@ -171,14 +171,12 @@ test_nil(X, TrueLab, FalseLab, Pred) ->
   hipe_rtl:mk_branch(X, eq, hipe_rtl:mk_imm(?NIL), TrueLab, FalseLab, Pred).
 
 test_cons(X, TrueLab, FalseLab, Pred) ->
-  Tmp = hipe_rtl:mk_new_reg_gcsafe(),
   Mask = hipe_rtl:mk_imm(?TAG_PRIMARY_MASK - ?TAG_PRIMARY_LIST),
-  hipe_rtl:mk_alub(Tmp, X, 'and', Mask, 'eq', TrueLab, FalseLab, Pred).
+  hipe_rtl:mk_branch(X, 'and', Mask, 'eq', TrueLab, FalseLab, Pred).
 
 test_is_boxed(X, TrueLab, FalseLab, Pred) ->
-  Tmp = hipe_rtl:mk_new_reg_gcsafe(),
   Mask = hipe_rtl:mk_imm(?TAG_PRIMARY_MASK - ?TAG_PRIMARY_BOXED),
-  hipe_rtl:mk_alub(Tmp, X, 'and', Mask, 'eq', TrueLab, FalseLab, Pred).
+  hipe_rtl:mk_branch(X, 'and', Mask, 'eq', TrueLab, FalseLab, Pred).
 
 get_header(Res, X) ->
   hipe_rtl:mk_load(Res, X, hipe_rtl:mk_imm(-(?TAG_PRIMARY_BOXED))).
@@ -238,13 +236,12 @@ test_atom(X, TrueLab, FalseLab, Pred) ->
 
 test_tuple(X, TrueLab, FalseLab, Pred) ->
   Tmp = hipe_rtl:mk_new_reg_gcsafe(),
-  Tmp2 = hipe_rtl:mk_new_reg_gcsafe(),
   HalfTrueLab = hipe_rtl:mk_new_label(),
   [test_is_boxed(X, hipe_rtl:label_name(HalfTrueLab), FalseLab, Pred),
    HalfTrueLab,
    get_header(Tmp, X),
-   hipe_rtl:mk_alub(Tmp2, Tmp, 'and', hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
-		    TrueLab, FalseLab, Pred)].
+   hipe_rtl:mk_branch(Tmp, 'and', hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
+		      TrueLab, FalseLab, Pred)].
 
 test_tuple_N(X, N, TrueLab, FalseLab, Pred) ->
   Tmp = hipe_rtl:mk_new_reg_gcsafe(),
@@ -687,7 +684,6 @@ element(Dst, Index, Tuple, FailLabName, unknown, IndexInfo) ->
   IndexOkLab = hipe_rtl:mk_new_label(),
   Ptr = hipe_rtl:mk_new_reg(), % offset from Tuple
   Header = hipe_rtl:mk_new_reg_gcsafe(),
-  Tmp = hipe_rtl:mk_new_reg_gcsafe(),
   UIndex = hipe_rtl:mk_new_reg_gcsafe(),
   Arity = hipe_rtl:mk_new_reg_gcsafe(),
   InvIndex = hipe_rtl:mk_new_reg_gcsafe(),
@@ -700,9 +696,9 @@ element(Dst, Index, Tuple, FailLabName, unknown, IndexInfo) ->
        BoxedOkLab,
        hipe_rtl:mk_alu(Ptr, Tuple, 'sub', hipe_rtl:mk_imm(?TAG_PRIMARY_BOXED)),
        hipe_rtl:mk_load(Header, Ptr, hipe_rtl:mk_imm(0)),
-       hipe_rtl:mk_alub(Tmp, Header, 'and', 
-			hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
-			hipe_rtl:label_name(TupleOkLab), FailLabName, 0.99),
+       hipe_rtl:mk_branch(Header, 'and',
+			  hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
+			  hipe_rtl:label_name(TupleOkLab), FailLabName, 0.99),
        TupleOkLab,
        untag_fixnum(UIndex, Index),
        hipe_rtl:mk_alu(Arity, Header, 'srl',
@@ -716,9 +712,9 @@ element(Dst, Index, Tuple, FailLabName, unknown, IndexInfo) ->
        BoxedOkLab,
        hipe_rtl:mk_alu(Ptr, Tuple, 'sub', hipe_rtl:mk_imm(?TAG_PRIMARY_BOXED)),
        hipe_rtl:mk_load(Header, Ptr, hipe_rtl:mk_imm(0)),
-       hipe_rtl:mk_alub(Tmp, Header, 'and', 
-			hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
-			hipe_rtl:label_name(TupleOkLab), FailLabName, 0.99),
+       hipe_rtl:mk_branch(Header, 'and',
+			  hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
+			  hipe_rtl:label_name(TupleOkLab), FailLabName, 0.99),
        TupleOkLab,
        hipe_rtl:mk_alu(Arity, Header, 'srl', 
 		       hipe_rtl:mk_imm(?HEADER_ARITY_OFFS))|
@@ -734,9 +730,9 @@ element(Dst, Index, Tuple, FailLabName, unknown, IndexInfo) ->
        BoxedOkLab,
        hipe_rtl:mk_alu(Ptr, Tuple, 'sub', hipe_rtl:mk_imm(?TAG_PRIMARY_BOXED)),
        hipe_rtl:mk_load(Header, Ptr, hipe_rtl:mk_imm(0)),
-       hipe_rtl:mk_alub(Tmp, Header, 'and', 
-			hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
-			hipe_rtl:label_name(TupleOkLab), FailLabName, 0.99),
+       hipe_rtl:mk_branch(Header, 'and',
+			  hipe_rtl:mk_imm(?TAG_HEADER_MASK), 'eq',
+			  hipe_rtl:label_name(TupleOkLab), FailLabName, 0.99),
        TupleOkLab,
        untag_fixnum(UIndex, Index),
        hipe_rtl:mk_alu(Arity, Header, 'srl',
diff --git a/lib/hipe/sparc/hipe_rtl_to_sparc.erl b/lib/hipe/sparc/hipe_rtl_to_sparc.erl
index e170fec3d6..7fab0d95c7 100644
--- a/lib/hipe/sparc/hipe_rtl_to_sparc.erl
+++ b/lib/hipe/sparc/hipe_rtl_to_sparc.erl
@@ -63,7 +63,6 @@ conv_insn(I, Map, Data) ->
   case I of
     #alu{} -> conv_alu(I, Map, Data);
     #alub{} -> conv_alub(I, Map, Data);
-    #branch{} -> conv_branch(I, Map, Data);
     #call{} -> conv_call(I, Map, Data);
     #comment{} -> conv_comment(I, Map, Data);
     #enter{} -> conv_enter(I, Map, Data);
@@ -281,7 +280,12 @@ mk_alu_rs(XAluOp, Src1, Src2, Dst) ->
 
 conv_alub(I, Map, Data) ->
   %% dst = src1 aluop src2; if COND goto label
-  {Dst, Map0} = conv_dst(hipe_rtl:alub_dst(I), Map),
+  HasDst = hipe_rtl:alub_has_dst(I),
+  {Dst, Map0} =
+    case HasDst of
+      false -> {hipe_sparc:mk_g0(), Map};
+      true -> conv_dst(hipe_rtl:alub_dst(I), Map)
+    end,
   {Src1, Map1} = conv_src(hipe_rtl:alub_src1(I), Map0),
   {Src2, Map2} = conv_src(hipe_rtl:alub_src2(I), Map1),
   Cond = conv_cond(hipe_rtl:alub_cond(I)),
@@ -307,67 +311,33 @@ conv_alub(I, Map, Data) ->
 	I1 ++
 	[hipe_sparc:mk_rdy(TmpHi),
 	 hipe_sparc:mk_alu('sra', Dst, hipe_sparc:mk_uimm5(31), TmpSign) |
-	 conv_alub2(G0, TmpSign, 'sub', NewCond, TmpHi, I)];
+	 conv_alub2(G0, TmpSign, 'cmpcc', NewCond, TmpHi, I)];
       _ ->
-	conv_alub2(Dst, Src1, RtlAlubOp, Cond, Src2, I)
+	XAluOp =
+	  case (not HasDst) andalso RtlAlubOp =:= 'sub' of
+	    true -> 'cmpcc'; % == a subcc that commutes
+	    false -> conv_alubop_cc(RtlAlubOp)
+	  end,
+	conv_alub2(Dst, Src1, XAluOp, Cond, Src2, I)
     end,
   {I2, Map2, Data}.
 
--ifdef(notdef).	% XXX: only for sparc64, alas
-conv_alub2(Dst, Src1, RtlAlubOp, Cond, Src2, I) ->
-  case conv_cond_rcond(Cond) of
-    [] ->
-      conv_alub_bp(Dst, Src1, RtlAlubOp, Cond, Src2, I);
-    RCond ->
-      conv_alub_br(Dst, Src1, RtlAlubOp, RCond, Src2, I)
-  end.
+conv_alub2(Dst, Src1, XAluOp, Cond, Src2, I) ->
+  conv_alub_bp(Dst, Src1, XAluOp, Cond, Src2, I).
 
-conv_alub_br(Dst, Src1, RtlAlubOp, RCond, Src2, I) ->
-  TrueLab = hipe_rtl:alub_true_label(I),
-  FalseLab = hipe_rtl:alub_false_label(I),
-  Pred = hipe_rtl:alub_pred(I),
-  %% "Dst = Src1 AluOp Src2; if COND" becomes
-  %% "Dst = Src1 AluOp Src2; if-COND(Dst)"
-  {I2, _DidCommute} = mk_alu(conv_alubop_nocc(RtlAlubOp), Src1, Src2, Dst),
-  I2 ++ mk_pseudo_br(RCond, Dst, TrueLab, FalseLab, Pred).
-
-conv_cond_rcond(Cond) ->
-  case Cond of
-    'e'  -> 'z';
-    'ne' -> 'nz';
-    'g'  -> 'gz';
-    'ge' -> 'gez';
-    'l'  -> 'lz';
-    'le' -> 'lez';
-    _	 -> []	% vs, vc, gu, geu, lu, leu
-  end.
-
-conv_alubop_nocc(RtlAlubOp) ->
-  case RtlAlubOp of
-    'add' -> 'add';
-    'sub' -> 'sub';
-    %% mul: handled elsewhere
-    'or' -> 'or';
-    'and' -> 'and';
-    'xor' -> 'xor'
-    %% no shift ops
-  end.
-
-mk_pseudo_br(RCond, Dst, TrueLab, FalseLab, Pred) ->
-  [hipe_sparc:mk_pseudo_br(RCond, Dst, TrueLab, FalseLab, Pred)].
--else.
-conv_alub2(Dst, Src1, RtlAlubOp, Cond, Src2, I) ->
-  conv_alub_bp(Dst, Src1, RtlAlubOp, Cond, Src2, I).
--endif.
-
-conv_alub_bp(Dst, Src1, RtlAlubOp, Cond, Src2, I) ->
+conv_alub_bp(Dst, Src1, XAluOp, Cond, Src2, I) ->
   TrueLab = hipe_rtl:alub_true_label(I),
   FalseLab = hipe_rtl:alub_false_label(I),
   Pred = hipe_rtl:alub_pred(I),
   %% "Dst = Src1 AluOp Src2; if COND" becomes
   %% "Dst = Src1 AluOpCC Src22; if-COND(CC)"
-  {I2, _DidCommute} = mk_alu(conv_alubop_cc(RtlAlubOp), Src1, Src2, Dst),
-  I2 ++ mk_pseudo_bp(Cond, TrueLab, FalseLab, Pred).
+  {I2, DidCommute} = mk_alu(XAluOp, Src1, Src2, Dst),
+  NewCond =
+    case DidCommute andalso XAluOp =:= 'cmpcc' of
+      true -> commute_cond(Cond); % subcc does not commute; its conditions do
+      false -> Cond
+    end,
+  I2 ++ mk_pseudo_bp(NewCond, TrueLab, FalseLab, Pred).
 
 conv_alubop_cc(RtlAlubOp) ->
   case RtlAlubOp of
@@ -380,69 +350,6 @@ conv_alubop_cc(RtlAlubOp) ->
     %% no shift ops
   end.
 
-conv_branch(I, Map, Data) ->
-  %% <unused> = src1 - src2; if COND goto label
-  {Src1, Map0} = conv_src(hipe_rtl:branch_src1(I), Map),
-  {Src2, Map1} = conv_src(hipe_rtl:branch_src2(I), Map0),
-  Cond = conv_cond(hipe_rtl:branch_cond(I)),
-  I2 = conv_branch2(Src1, Cond, Src2, I),
-  {I2, Map1, Data}.
-
--ifdef(notdef).	% XXX: only for sparc64, alas
-conv_branch2(Src1, Cond, Src2, I) ->
-  case conv_cond_rcond(Cond) of
-    [] ->
-      conv_branch_bp(Src1, Cond, Src2, I);
-    RCond ->
-      conv_branch_br(Src1, RCond, Src2, I)
-  end.
-
-conv_branch_br(Src1, RCond, Src2, I) ->
-  TrueLab = hipe_rtl:branch_true_label(I),
-  FalseLab = hipe_rtl:branch_false_label(I),
-  Pred = hipe_rtl:branch_pred(I),
-  %% "if src1-COND-src2" becomes
-  %% "sub src1,src2,tmp; if-COND(tmp)"
-  Dst = hipe_sparc:mk_new_temp('untagged'),
-  XAluOp = 'cmp',	% == a sub that commutes
-  {I1, DidCommute} = mk_alu(XAluOp, Src1, Src2, Dst),
-  NewRCond =
-    case DidCommute of
-      true -> commute_rcond(RCond);
-      false -> RCond
-    end,
-  I1 ++ mk_pseudo_br(NewRCond, Dst, TrueLab, FalseLab, Pred).
-
-commute_rcond(RCond) ->	% if x RCond y, then y commute_rcond(RCond) x
-  case RCond of
-    'z'   -> 'z';	% ==, ==
-    'nz'  -> 'nz';	% !=, !=
-    'gz'  -> 'lz';	% >, <
-    'gez' -> 'lez';	% >=, <=
-    'lz'  -> 'gz';	% <, >
-    'lez' -> 'gez'	% <=, >=
-  end.
--else.
-conv_branch2(Src1, Cond, Src2, I) ->
-  conv_branch_bp(Src1, Cond, Src2, I).
--endif.
-
-conv_branch_bp(Src1, Cond, Src2, I) ->
-  TrueLab = hipe_rtl:branch_true_label(I),
-  FalseLab = hipe_rtl:branch_false_label(I),
-  Pred = hipe_rtl:branch_pred(I),
-  %% "if src1-COND-src2" becomes
-  %% "subcc src1,src2,%g0; if-COND(CC)"
-  Dst = hipe_sparc:mk_g0(),
-  XAluOp = 'cmpcc',	% == a subcc that commutes
-  {I1, DidCommute} = mk_alu(XAluOp, Src1, Src2, Dst),
-  NewCond =
-    case DidCommute of
-      true -> commute_cond(Cond);
-      false -> Cond
-    end,
-  I1 ++ mk_pseudo_bp(NewCond, TrueLab, FalseLab, Pred).
-
 conv_call(I, Map, Data) ->
   {Args, Map0} = conv_src_list(hipe_rtl:call_arglist(I), Map),
   {Dsts, Map1} = conv_dst_list(hipe_rtl:call_dstlist(I), Map0),
diff --git a/lib/hipe/x86/hipe_rtl_to_x86.erl b/lib/hipe/x86/hipe_rtl_to_x86.erl
index 4c8c98551c..ccb9b7632b 100644
--- a/lib/hipe/x86/hipe_rtl_to_x86.erl
+++ b/lib/hipe/x86/hipe_rtl_to_x86.erl
@@ -91,26 +91,31 @@ conv_insn(I, Map, Data) ->
     #alub{} ->
       %% dst = src1 op src2; if COND goto label
       BinOp = conv_binop(hipe_rtl:alub_op(I)),
-      {Dst, Map0} = conv_dst(hipe_rtl:alub_dst(I), Map),
-      {FixSrc1, Src1, Map1} = conv_src(hipe_rtl:alub_src1(I), Map0),
-      {FixSrc2, Src2, Map2} = conv_src(hipe_rtl:alub_src2(I), Map1),
+      {FixSrc1, Src1, Map0} = conv_src(hipe_rtl:alub_src1(I), Map),
+      {FixSrc2, Src2, Map1} = conv_src(hipe_rtl:alub_src2(I), Map0),
       Cc = conv_cond(hipe_rtl:alub_cond(I)),
-      I1 = [hipe_x86:mk_pseudo_jcc(Cc,
-				   hipe_rtl:alub_true_label(I),
-				   hipe_rtl:alub_false_label(I),
-				   hipe_rtl:alub_pred(I))],
-      I2 = conv_alu(Dst, Src1, BinOp, Src2, I1),
-      {FixSrc1++FixSrc2++I2, Map2, Data};
-    #branch{} ->
-      %% <unused> = src1 - src2; if COND goto label
-      {FixSrc1, Src1, Map0} = conv_src(hipe_rtl:branch_src1(I), Map),
-      {FixSrc2, Src2, Map1} = conv_src(hipe_rtl:branch_src2(I), Map0),
-      Cc = conv_cond(hipe_rtl:branch_cond(I)),
-      I2 = conv_branch(Src1, Cc, Src2,
-		       hipe_rtl:branch_true_label(I),
-		       hipe_rtl:branch_false_label(I),
-		       hipe_rtl:branch_pred(I)),
-      {FixSrc1++FixSrc2++I2, Map1, Data};
+      BranchOp = conv_branchop(BinOp),
+      HasDst = hipe_rtl:alub_has_dst(I),
+      {I2, Map3} =
+	case (not HasDst) andalso BranchOp =/= none of
+	  true ->
+	    {conv_branch(Src1, BranchOp, Src2, Cc,
+			 hipe_rtl:alub_true_label(I),
+			 hipe_rtl:alub_false_label(I),
+			 hipe_rtl:alub_pred(I)), Map1};
+	  false ->
+	    {Dst, Map2} =
+	      case HasDst of
+		false -> {new_untagged_temp(), Map1};
+		true -> conv_dst(hipe_rtl:alub_dst(I), Map1)
+	      end,
+	    I1 = [hipe_x86:mk_pseudo_jcc(Cc,
+					 hipe_rtl:alub_true_label(I),
+					 hipe_rtl:alub_false_label(I),
+					 hipe_rtl:alub_pred(I))],
+	    {conv_alu(Dst, Src1, BinOp, Src2, I1), Map2}
+	end,
+      {FixSrc1++FixSrc2++I2, Map3, Data};
     #call{} ->
       %%	push <arg1>
       %%	...
@@ -360,28 +365,41 @@ conv_shift(Dst, Src1, BinOp, Src2) ->
 %%% Finalise the conversion of a conditional branch operation, taking
 %%% care to not introduce more temps and moves than necessary.
 
-conv_branch(Src1, Cc, Src2, TrueLab, FalseLab, Pred) ->
+conv_branchop('sub') -> 'cmp';
+conv_branchop('and') ->  'test';
+conv_branchop(_) -> none.
+
+branchop_commutes('cmp') -> false;
+branchop_commutes('test') -> true.
+
+conv_branch(Src1, Op, Src2, Cc, TrueLab, FalseLab, Pred) ->
   case hipe_x86:is_imm(Src1) of
     false ->
-      mk_branch(Src1, Cc, Src2, TrueLab, FalseLab, Pred);
+      mk_branch(Src1, Op, Src2, Cc, TrueLab, FalseLab, Pred);
     true ->
       case hipe_x86:is_imm(Src2) of
 	false ->
-	  NewCc = commute_cc(Cc),
-	  mk_branch(Src2, NewCc, Src1, TrueLab, FalseLab, Pred);
+	  NewCc = case branchop_commutes(Op) of
+		    true -> Cc;
+		    false -> commute_cc(Cc)
+		  end,
+	  mk_branch(Src2, Op, Src1, NewCc, TrueLab, FalseLab, Pred);
 	true ->
 	  %% two immediates, let the optimiser clean it up
 	  Tmp = new_untagged_temp(),
 	  [hipe_x86:mk_move(Src1, Tmp) |
-	   mk_branch(Tmp, Cc, Src2, TrueLab, FalseLab, Pred)]
+	   mk_branch(Tmp, Op, Src2, Cc, TrueLab, FalseLab, Pred)]
       end
   end.
 
-mk_branch(Src1, Cc, Src2, TrueLab, FalseLab, Pred) ->
+mk_branch(Src1, Op, Src2, Cc, TrueLab, FalseLab, Pred) ->
   %% PRE: not(is_imm(Src1))
-  [hipe_x86:mk_cmp(Src2, Src1),
+  [mk_branchtest(Src1, Op, Src2),
    hipe_x86:mk_pseudo_jcc(Cc, TrueLab, FalseLab, Pred)].
 
+mk_branchtest(Src1, cmp, Src2) -> hipe_x86:mk_cmp(Src2, Src1);
+mk_branchtest(Src1, test, Src2) -> hipe_x86:mk_test(Src2, Src1).
+
 %%% Convert an RTL ALU or ALUB binary operator.
 
 conv_binop(BinOp) ->
diff --git a/lib/hipe/x86/hipe_x86.erl b/lib/hipe/x86/hipe_x86.erl
index 33d7f77cf1..e7c4497cda 100644
--- a/lib/hipe/x86/hipe_x86.erl
+++ b/lib/hipe/x86/hipe_x86.erl
@@ -201,7 +201,7 @@
 	 shift_src/1,
 	 shift_dst/1,
 
-	 %% mk_test/2,
+	 mk_test/2,
 	 test_src/1,
 	 test_dst/1,
 
@@ -305,7 +305,7 @@ mk_cmp(Src, Dst) -> #cmp{src=Src, dst=Dst}.
 cmp_src(#cmp{src=Src}) -> Src.
 cmp_dst(#cmp{dst=Dst}) -> Dst.
 
-%% mk_test(Src, Dst) -> #test{src=Src, dst=Dst}.
+mk_test(Src, Dst) -> #test{src=Src, dst=Dst}.
 test_src(#test{src=Src}) -> Src.
 test_dst(#test{dst=Dst}) -> Dst.