From c52b2cf226cb3f1bb1b16bee28d47785506adff3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20L=C3=A5ng?= <margnus1@telia.com>
Date: Thu, 16 Mar 2017 15:30:00 +0100
Subject: hipe: Add pseudo_spill_f?move instructions

These pseudo instructions are added to all backends and allow spill slot
to spill slot move coalescing in a clean way.

They have regular move semantics, but contain an additional scratch
register to be used if both source and destination are spilled, and can
not be move coalesced.

Additionally, a register allocator callback
Target:is_spill_move(Instr, Context) is added which allows the spill
slot allocators to check for these instructions and try to coalesce the
spill slots the two temporaries are allocated to.
---
 .../amd64/hipe_amd64_ra_sse2_postconditions.erl    | 17 +++++++-
 lib/hipe/arm/hipe_arm.erl                          |  7 ++++
 lib/hipe/arm/hipe_arm.hrl                          |  1 +
 lib/hipe/arm/hipe_arm_defuse.erl                   |  2 +
 lib/hipe/arm/hipe_arm_frame.erl                    | 22 +++++++++++
 lib/hipe/arm/hipe_arm_ra_finalise.erl              | 25 +++++++++---
 lib/hipe/arm/hipe_arm_ra_postconditions.erl        | 24 +++++++----
 lib/hipe/arm/hipe_arm_subst.erl                    |  2 +
 lib/hipe/ppc/hipe_ppc.erl                          | 14 +++++++
 lib/hipe/ppc/hipe_ppc.hrl                          |  2 +
 lib/hipe/ppc/hipe_ppc_defuse.erl                   |  4 ++
 lib/hipe/ppc/hipe_ppc_frame.erl                    | 36 +++++++++++++++++
 lib/hipe/ppc/hipe_ppc_ra_finalise.erl              | 15 +++++++
 lib/hipe/ppc/hipe_ppc_ra_postconditions.erl        | 24 +++++++----
 lib/hipe/ppc/hipe_ppc_ra_postconditions_fp.erl     | 18 ++++++---
 lib/hipe/ppc/hipe_ppc_subst.erl                    |  6 ++-
 lib/hipe/regalloc/hipe_amd64_specific_sse2.erl     |  4 ++
 lib/hipe/regalloc/hipe_arm_specific.erl            |  4 ++
 lib/hipe/regalloc/hipe_ppc_specific.erl            |  4 ++
 lib/hipe/regalloc/hipe_ppc_specific_fp.erl         |  4 ++
 lib/hipe/regalloc/hipe_sparc_specific.erl          |  4 ++
 lib/hipe/regalloc/hipe_sparc_specific_fp.erl       |  4 ++
 lib/hipe/regalloc/hipe_x86_specific.erl            |  4 ++
 lib/hipe/regalloc/hipe_x86_specific_x87.erl        |  4 ++
 lib/hipe/sparc/hipe_sparc.erl                      | 14 +++++++
 lib/hipe/sparc/hipe_sparc.hrl                      |  2 +
 lib/hipe/sparc/hipe_sparc_defuse.erl               |  4 ++
 lib/hipe/sparc/hipe_sparc_frame.erl                | 36 +++++++++++++++++
 lib/hipe/sparc/hipe_sparc_ra_finalise.erl          | 15 +++++++
 lib/hipe/sparc/hipe_sparc_ra_postconditions.erl    | 24 +++++++----
 lib/hipe/sparc/hipe_sparc_ra_postconditions_fp.erl | 18 ++++++---
 lib/hipe/sparc/hipe_sparc_subst.erl                |  6 ++-
 lib/hipe/x86/hipe_x86.erl                          | 14 +++++++
 lib/hipe/x86/hipe_x86.hrl                          |  2 +
 lib/hipe/x86/hipe_x86_defuse.erl                   |  4 ++
 lib/hipe/x86/hipe_x86_frame.erl                    | 46 ++++++++++++++++++----
 lib/hipe/x86/hipe_x86_ra_finalise.erl              | 10 +++++
 lib/hipe/x86/hipe_x86_ra_postconditions.erl        | 26 ++++++++++--
 lib/hipe/x86/hipe_x86_subst.erl                    |  4 ++
 39 files changed, 422 insertions(+), 54 deletions(-)

diff --git a/lib/hipe/amd64/hipe_amd64_ra_sse2_postconditions.erl b/lib/hipe/amd64/hipe_amd64_ra_sse2_postconditions.erl
index 8a3ea92156..891c874a15 100644
--- a/lib/hipe/amd64/hipe_amd64_ra_sse2_postconditions.erl
+++ b/lib/hipe/amd64/hipe_amd64_ra_sse2_postconditions.erl
@@ -53,6 +53,8 @@ do_insn(I, TempMap, Strategy) ->	% Insn -> {Insn list, DidSpill}
       do_fp_unop(I, TempMap, Strategy);
     #fp_binop{} ->
       do_fp_binop(I, TempMap, Strategy);
+    #pseudo_spill_fmove{} ->
+      do_pseudo_spill_fmove(I, TempMap, Strategy);
     _ ->
       %% All non sse2 ops
       {[I], false}
@@ -95,8 +97,13 @@ do_fmove(I, TempMap, Strategy) ->
   of
     true ->
       Tmp = spill_temp(double, Strategy),
-      {[#fmove{src=Src, dst=Tmp},I#fmove{src=Tmp,dst=Dst}],
-       true};
+      %% pseudo_spill_fmove allows spill slot move coalescing, but must not
+      %% contain memory operands (except for spilled temps)
+      Is = case is_float_temp(Src) andalso is_float_temp(Dst) of
+	     true -> [#pseudo_spill_fmove{src=Src, temp=Tmp, dst=Dst}];
+	     false -> [#fmove{src=Src, dst=Tmp},I#fmove{src=Tmp,dst=Dst}]
+	   end,
+      {Is, true};
     false ->
       {[I], false}
   end.
@@ -104,6 +111,12 @@ do_fmove(I, TempMap, Strategy) ->
 is_float_temp(#x86_temp{type=Type}) -> Type =:= double;
 is_float_temp(#x86_mem{}) -> false.
 
+%%% Fix an pseudo_spill_fmove op.
+do_pseudo_spill_fmove(I = #pseudo_spill_fmove{temp=Temp}, TempMap, _Strategy) ->
+  %% Temp is above the low water mark and must not have been spilled
+  false = is_mem_opnd(Temp, TempMap),
+  {[I], false}. % nothing to do
+
 %%% Check if an operand denotes a memory cell (mem or pseudo).
 
 is_mem_opnd(Opnd, TempMap) ->
diff --git a/lib/hipe/arm/hipe_arm.erl b/lib/hipe/arm/hipe_arm.erl
index e34a00f561..3b090b501a 100644
--- a/lib/hipe/arm/hipe_arm.erl
+++ b/lib/hipe/arm/hipe_arm.erl
@@ -79,6 +79,9 @@
 	 pseudo_move_dst/1,
 	 pseudo_move_src/1,
 
+	 mk_pseudo_spill_move/3,
+	 is_pseudo_spill_move/1,
+
 	 mk_pseudo_switch/3,
 
 	 mk_pseudo_tailcall/4,
@@ -250,6 +253,10 @@ is_pseudo_move(I) -> case I of #pseudo_move{} -> true; _ -> false end.
 pseudo_move_dst(#pseudo_move{dst=Dst}) -> Dst.
 pseudo_move_src(#pseudo_move{src=Src}) -> Src.
 
+mk_pseudo_spill_move(Dst, Temp, Src) ->
+  #pseudo_spill_move{dst=Dst, temp=Temp, src=Src}.
+is_pseudo_spill_move(I) -> is_record(I, pseudo_spill_move).
+
 mk_pseudo_switch(JTab, Index, Labels) ->
   #pseudo_switch{jtab=JTab, index=Index, labels=Labels}.
 
diff --git a/lib/hipe/arm/hipe_arm.hrl b/lib/hipe/arm/hipe_arm.hrl
index 67bc07634e..be06b1ebd7 100644
--- a/lib/hipe/arm/hipe_arm.hrl
+++ b/lib/hipe/arm/hipe_arm.hrl
@@ -101,6 +101,7 @@
 -record(pseudo_call_prepare, {nrstkargs}).
 -record(pseudo_li, {dst, imm, label}).	% pre-generated label for use by the assembler
 -record(pseudo_move, {dst, src}).
+-record(pseudo_spill_move, {dst, temp, src}).
 -record(pseudo_switch, {jtab, index, labels}).
 -record(pseudo_tailcall, {funv, arity, stkargs, linkage}).
 -record(pseudo_tailcall_prepare, {}).
diff --git a/lib/hipe/arm/hipe_arm_defuse.erl b/lib/hipe/arm/hipe_arm_defuse.erl
index 0e62070c6c..652299a514 100644
--- a/lib/hipe/arm/hipe_arm_defuse.erl
+++ b/lib/hipe/arm/hipe_arm_defuse.erl
@@ -40,6 +40,7 @@ insn_def_gpr(I) ->
     #pseudo_call{} -> call_clobbered_gpr();
     #pseudo_li{dst=Dst} -> [Dst];
     #pseudo_move{dst=Dst} -> [Dst];
+    #pseudo_spill_move{dst=Dst, temp=Temp} -> [Dst, Temp];
     #pseudo_tailcall_prepare{} -> tailcall_clobbered_gpr();
     #smull{dstlo=DstLo,dsthi=DstHi,src1=Src1} ->
       %% ARM requires DstLo, DstHi, and Src1 to be distinct.
@@ -83,6 +84,7 @@ insn_use_gpr(I) ->
     #pseudo_call{funv=FunV,sdesc=#arm_sdesc{arity=Arity}} ->
       funv_use(FunV, arity_use_gpr(Arity));
     #pseudo_move{src=Src} -> [Src];
+    #pseudo_spill_move{src=Src} -> [Src];
     #pseudo_switch{jtab=JTabR,index=IndexR} -> addtemp(JTabR, [IndexR]);
     #pseudo_tailcall{funv=FunV,arity=Arity,stkargs=StkArgs} ->
       addargs(StkArgs, addtemps(tailcall_clobbered_gpr(), funv_use(FunV, arity_use_gpr(Arity))));
diff --git a/lib/hipe/arm/hipe_arm_frame.erl b/lib/hipe/arm/hipe_arm_frame.erl
index e323907e31..a1004fb609 100644
--- a/lib/hipe/arm/hipe_arm_frame.erl
+++ b/lib/hipe/arm/hipe_arm_frame.erl
@@ -69,6 +69,8 @@ do_insn(I, LiveOut, Context, FPoff) ->
       do_pseudo_call_prepare(I, FPoff);
     #pseudo_move{} ->
       {do_pseudo_move(I, Context, FPoff), FPoff};
+    #pseudo_spill_move{} ->
+      {do_pseudo_spill_move(I, Context, FPoff), FPoff};
     #pseudo_tailcall{} ->
       {do_pseudo_tailcall(I, Context), context_framesize(Context)};
     _ ->
@@ -99,6 +101,26 @@ do_pseudo_move(I, Context, FPoff) ->
 pseudo_offset(Temp, FPoff, Context) ->
   FPoff + context_offset(Context, Temp).
 
+%%%
+%%% Moves from one spill slot to another
+%%%
+
+do_pseudo_spill_move(I, Context, FPoff) ->
+  #pseudo_spill_move{dst=Dst, temp=Temp, src=Src} = I,
+  case temp_is_pseudo(Src) andalso temp_is_pseudo(Dst) of
+    false -> % Register allocator changed its mind, turn back to move
+      do_pseudo_move(hipe_arm:mk_pseudo_move(Dst, Src), Context, FPoff);
+    true ->
+      SrcOffset = pseudo_offset(Src, FPoff, Context),
+      DstOffset = pseudo_offset(Dst, FPoff, Context),
+      case SrcOffset =:= DstOffset of
+	true -> []; % omit move-to-self
+	false ->
+	  mk_load('ldr', Temp, SrcOffset, mk_sp(),
+		  mk_store('str', Temp, DstOffset, mk_sp(), []))
+      end
+  end.
+
 %%%
 %%% Return - deallocate frame and emit 'ret $N' insn.
 %%%
diff --git a/lib/hipe/arm/hipe_arm_ra_finalise.erl b/lib/hipe/arm/hipe_arm_ra_finalise.erl
index 9bfe0a9a83..80cd470708 100644
--- a/lib/hipe/arm/hipe_arm_ra_finalise.erl
+++ b/lib/hipe/arm/hipe_arm_ra_finalise.erl
@@ -25,11 +25,17 @@ ra_bb(BB, Map) ->
   hipe_bb:code_update(BB, ra_code(hipe_bb:code(BB), Map, [])).
 
 ra_code([I|Insns], Map, Accum) ->
-  ra_code(Insns, Map, [ra_insn(I, Map) | Accum]);
+  ra_code(Insns, Map, ra_insn(I, Map, Accum));
 ra_code([], _Map, Accum) ->
   lists:reverse(Accum).
 
-ra_insn(I, Map) ->
+ra_insn(I, Map, Accum) ->
+  case I of
+    #pseudo_move{} -> ra_pseudo_move(I, Map, Accum);
+    _ -> [ra_insn_1(I, Map) | Accum]
+  end.
+
+ra_insn_1(I, Map) ->
   case I of
     #alu{} -> ra_alu(I, Map);
     #cmp{} -> ra_cmp(I, Map);
@@ -38,7 +44,7 @@ ra_insn(I, Map) ->
     #move{} -> ra_move(I, Map);
     #pseudo_call{} -> ra_pseudo_call(I, Map);
     #pseudo_li{} -> ra_pseudo_li(I, Map);
-    #pseudo_move{} -> ra_pseudo_move(I, Map);
+    #pseudo_spill_move{} -> ra_pseudo_spill_move(I, Map);
     #pseudo_switch{} -> ra_pseudo_switch(I, Map);
     #pseudo_tailcall{} -> ra_pseudo_tailcall(I, Map);
     #smull{} -> ra_smull(I, Map);
@@ -80,10 +86,19 @@ ra_pseudo_li(I=#pseudo_li{dst=Dst}, Map) ->
   NewDst = ra_temp(Dst, Map),
   I#pseudo_li{dst=NewDst}.
 
-ra_pseudo_move(I=#pseudo_move{dst=Dst,src=Src}, Map) ->
+ra_pseudo_move(I=#pseudo_move{dst=Dst,src=Src}, Map, Accum) ->
+  NewDst = ra_temp(Dst, Map),
+  NewSrc = ra_temp(Src, Map),
+  case NewSrc#arm_temp.reg =:= NewDst#arm_temp.reg of
+    true -> Accum;
+    false -> [I#pseudo_move{dst=NewDst,src=NewSrc} | Accum]
+  end.
+
+ra_pseudo_spill_move(I=#pseudo_spill_move{dst=Dst,temp=Temp,src=Src}, Map) ->
   NewDst = ra_temp(Dst, Map),
+  NewTemp = ra_temp(Temp, Map),
   NewSrc = ra_temp(Src, Map),
-  I#pseudo_move{dst=NewDst,src=NewSrc}.
+  I#pseudo_spill_move{dst=NewDst, temp=NewTemp, src=NewSrc}.
 
 ra_pseudo_switch(I=#pseudo_switch{jtab=JTab,index=Index}, Map) ->
   NewJTab = ra_temp(JTab, Map),
diff --git a/lib/hipe/arm/hipe_arm_ra_postconditions.erl b/lib/hipe/arm/hipe_arm_ra_postconditions.erl
index 8d1ee1cb94..23c305511f 100644
--- a/lib/hipe/arm/hipe_arm_ra_postconditions.erl
+++ b/lib/hipe/arm/hipe_arm_ra_postconditions.erl
@@ -56,6 +56,7 @@ do_insn(I, TempMap, Strategy) ->
     #pseudo_call{} -> do_pseudo_call(I, TempMap, Strategy);
     #pseudo_li{} -> do_pseudo_li(I, TempMap, Strategy);
     #pseudo_move{} -> do_pseudo_move(I, TempMap, Strategy);
+    #pseudo_spill_move{} -> do_pseudo_spill_move(I, TempMap, Strategy);
     #pseudo_switch{} -> do_pseudo_switch(I, TempMap, Strategy);
     #pseudo_tailcall{} -> do_pseudo_tailcall(I, TempMap, Strategy);
     #smull{} -> do_smull(I, TempMap, Strategy);
@@ -108,18 +109,25 @@ do_pseudo_li(I=#pseudo_li{dst=Dst}, TempMap, Strategy) ->
 
 do_pseudo_move(I=#pseudo_move{dst=Dst,src=Src}, TempMap, Strategy) ->
   %% Either Dst or Src (but not both) may be a pseudo temp.
-  %% pseudo_move and pseudo_tailcall are special cases: in
-  %% all other instructions, all temps must be non-pseudos
-  %% after register allocation.
-  case temp_is_spilled(Dst, TempMap) of
-    true -> % Src must not be a pseudo
-      {FixSrc,NewSrc,DidSpill} = fix_src1(Src, TempMap, Strategy),
-      NewI = I#pseudo_move{src=NewSrc},
-      {FixSrc ++ [NewI], DidSpill};
+  %% pseudo_move, pseudo_spill_move, and pseudo_tailcall
+  %% are special cases: in all other instructions, all
+  %% temps must be non-pseudos after register allocation.
+  case temp_is_spilled(Dst, TempMap)
+    andalso temp_is_spilled(Dst, TempMap)
+  of
+    true -> % Turn into pseudo_spill_move
+      Temp = clone(Src, temp1(Strategy)),
+      NewI = #pseudo_spill_move{dst=Dst, temp=Temp, src=Src},
+      {[NewI], true};
     _ ->
       {[I], false}
   end.
 
+do_pseudo_spill_move(I = #pseudo_spill_move{temp=Temp}, TempMap, _Strategy) ->
+  %% Temp is above the low water mark and must not have been spilled
+  false = temp_is_spilled(Temp, TempMap),
+  {[I], false}. % nothing to do
+
 do_pseudo_switch(I=#pseudo_switch{jtab=JTab,index=Index}, TempMap, Strategy) ->
   {FixJTab,NewJTab,DidSpill1} = fix_src1(JTab, TempMap, Strategy),
   {FixIndex,NewIndex,DidSpill2} = fix_src2(Index, TempMap, Strategy),
diff --git a/lib/hipe/arm/hipe_arm_subst.erl b/lib/hipe/arm/hipe_arm_subst.erl
index a41a907a4c..4ff245f414 100644
--- a/lib/hipe/arm/hipe_arm_subst.erl
+++ b/lib/hipe/arm/hipe_arm_subst.erl
@@ -59,6 +59,8 @@ insn_temps(T, I) ->
       #pseudo_call{funv=F} -> I#pseudo_call{funv=funv_temps(T, F)};
       #pseudo_call_prepare{} -> I;
       #pseudo_li{dst=D} -> I#pseudo_li{dst=T(D)};
+      #pseudo_spill_move{dst=D,temp=U,src=S} ->
+	  I#pseudo_spill_move{dst=T(D),temp=T(U),src=T(S)};
       #pseudo_switch{jtab=J=#arm_temp{},index=Ix=#arm_temp{}} ->
 	  I#pseudo_switch{jtab=T(J),index=T(Ix)};
       #pseudo_tailcall{funv=F,stkargs=Stk} ->
diff --git a/lib/hipe/ppc/hipe_ppc.erl b/lib/hipe/ppc/hipe_ppc.erl
index df9f193fa3..63ecd0a0b8 100644
--- a/lib/hipe/ppc/hipe_ppc.erl
+++ b/lib/hipe/ppc/hipe_ppc.erl
@@ -98,6 +98,9 @@
 	 pseudo_move_dst/1,
 	 pseudo_move_src/1,
 
+	 mk_pseudo_spill_move/3,
+	 is_pseudo_spill_move/1,
+
 	 mk_pseudo_tailcall/4,
 	 pseudo_tailcall_func/1,
 	 pseudo_tailcall_stkargs/1,
@@ -131,6 +134,9 @@
 	 pseudo_fmove_dst/1,
 	 pseudo_fmove_src/1,
 
+	 mk_pseudo_spill_fmove/3,
+	 is_pseudo_spill_fmove/1,
+
 	 mk_defun/8,
 	 defun_mfa/1,
 	 defun_formals/1,
@@ -412,6 +418,10 @@ is_pseudo_move(I) -> case I of #pseudo_move{} -> true; _ -> false end.
 pseudo_move_dst(#pseudo_move{dst=Dst}) -> Dst.
 pseudo_move_src(#pseudo_move{src=Src}) -> Src.
 
+mk_pseudo_spill_move(Dst, Temp, Src) ->
+  #pseudo_spill_move{dst=Dst, temp=Temp, src=Src}.
+is_pseudo_spill_move(I) -> is_record(I, pseudo_spill_move).
+
 mk_pseudo_tailcall(FunC, Arity, StkArgs, Linkage) ->
   #pseudo_tailcall{func=FunC, arity=Arity, stkargs=StkArgs, linkage=Linkage}.
 pseudo_tailcall_func(#pseudo_tailcall{func=FunC}) -> FunC.
@@ -495,6 +505,10 @@ is_pseudo_fmove(I) -> case I of #pseudo_fmove{} -> true; _ -> false end.
 pseudo_fmove_dst(#pseudo_fmove{dst=Dst}) -> Dst.
 pseudo_fmove_src(#pseudo_fmove{src=Src}) -> Src.
 
+mk_pseudo_spill_fmove(Dst, Temp, Src) ->
+  #pseudo_spill_fmove{dst=Dst, temp=Temp, src=Src}.
+is_pseudo_spill_fmove(I) -> is_record(I, pseudo_spill_fmove).
+
 mk_defun(MFA, Formals, IsClosure, IsLeaf, Code, Data, VarRange, LabelRange) ->
   #defun{mfa=MFA, formals=Formals, code=Code, data=Data,
 	 isclosure=IsClosure, isleaf=IsLeaf,
diff --git a/lib/hipe/ppc/hipe_ppc.hrl b/lib/hipe/ppc/hipe_ppc.hrl
index a96692c52e..3eef8be487 100644
--- a/lib/hipe/ppc/hipe_ppc.hrl
+++ b/lib/hipe/ppc/hipe_ppc.hrl
@@ -87,6 +87,7 @@
 -record(pseudo_call_prepare, {nrstkargs}).
 -record(pseudo_li, {dst, imm}).
 -record(pseudo_move, {dst, src}).
+-record(pseudo_spill_move, {dst, temp, src}).
 -record(pseudo_tailcall, {func, arity, stkargs, linkage}).
 -record(pseudo_tailcall_prepare, {}).
 -record(store, {stop, src, disp, base}).	% non-indexed, non-update form
@@ -99,6 +100,7 @@
 -record(fp_binary, {fp_binop, dst, src1, src2}).
 -record(fp_unary, {fp_unop, dst, src}).
 -record(pseudo_fmove, {dst, src}).
+-record(pseudo_spill_fmove, {dst, temp, src}).
 
 %%% Function definitions.
 
diff --git a/lib/hipe/ppc/hipe_ppc_defuse.erl b/lib/hipe/ppc/hipe_ppc_defuse.erl
index 9a99611493..d8a864f7d5 100644
--- a/lib/hipe/ppc/hipe_ppc_defuse.erl
+++ b/lib/hipe/ppc/hipe_ppc_defuse.erl
@@ -41,6 +41,7 @@ insn_def_gpr(I) ->
     #pseudo_call{} -> call_clobbered_gpr();
     #pseudo_li{dst=Dst} -> [Dst];
     #pseudo_move{dst=Dst} -> [Dst];
+    #pseudo_spill_move{dst=Dst,temp=Temp} -> [Dst, Temp];
     #pseudo_tailcall_prepare{} -> tailcall_clobbered_gpr();
     #unary{dst=Dst} -> [Dst];
     _ -> []
@@ -71,6 +72,7 @@ insn_use_gpr(I) ->
     #mtspr{src=Src} -> [Src];
     #pseudo_call{sdesc=#ppc_sdesc{arity=Arity}} -> arity_use_gpr(Arity);
     #pseudo_move{src=Src} -> [Src];
+    #pseudo_spill_move{src=Src} -> [Src];
     #pseudo_tailcall{arity=Arity,stkargs=StkArgs} ->
       addsrcs(StkArgs, addtemps(tailcall_clobbered_gpr(), arity_use_gpr(Arity)));
     #store{src=Src,base=Base} -> addtemp(Src, [Base]);
@@ -110,6 +112,7 @@ insn_def_fpr(I) ->
     #fp_binary{dst=Dst} -> [Dst];
     #fp_unary{dst=Dst} -> [Dst];
     #pseudo_fmove{dst=Dst} -> [Dst];
+    #pseudo_spill_fmove{dst=Dst,temp=Temp} -> [Dst, Temp];
     _ -> []
   end.
 
@@ -126,6 +129,7 @@ insn_use_fpr(I) ->
     #fp_binary{src1=Src1,src2=Src2} -> addtemp(Src1, [Src2]);
     #fp_unary{src=Src} -> [Src];
     #pseudo_fmove{src=Src} -> [Src];
+    #pseudo_spill_fmove{src=Src} -> [Src];
     _ -> []
   end.
 
diff --git a/lib/hipe/ppc/hipe_ppc_frame.erl b/lib/hipe/ppc/hipe_ppc_frame.erl
index a91cb18cc2..b88b75a5bd 100644
--- a/lib/hipe/ppc/hipe_ppc_frame.erl
+++ b/lib/hipe/ppc/hipe_ppc_frame.erl
@@ -66,10 +66,14 @@ do_insn(I, LiveOut, Context, FPoff) ->
       do_pseudo_call_prepare(I, FPoff);
     #pseudo_move{} ->
       {do_pseudo_move(I, Context, FPoff), FPoff};
+    #pseudo_spill_move{} ->
+      {do_pseudo_spill_move(I, Context, FPoff), FPoff};
     #pseudo_tailcall{} ->
       {do_pseudo_tailcall(I, Context), context_framesize(Context)};
     #pseudo_fmove{} ->
       {do_pseudo_fmove(I, Context, FPoff), FPoff};
+    #pseudo_spill_fmove{} ->
+      {do_pseudo_spill_fmove(I, Context, FPoff), FPoff};
     _ ->
       {[I], FPoff}
   end.
@@ -98,6 +102,22 @@ do_pseudo_move(I, Context, FPoff) ->
       end
   end.
 
+do_pseudo_spill_move(I, Context, FPoff) ->
+  #pseudo_spill_move{dst=Dst,temp=Temp,src=Src} = I,
+  case temp_is_pseudo(Src) andalso temp_is_pseudo(Dst) of
+    false -> % Register allocator changed its mind, turn back to move
+      do_pseudo_move(hipe_ppc:mk_pseudo_move(Dst, Src), Context, FPoff);
+    true ->
+      SrcOffset = pseudo_offset(Src, FPoff, Context),
+      DstOffset = pseudo_offset(Dst, FPoff, Context),
+      case SrcOffset =:= DstOffset of
+	true -> []; % omit move-to-self
+	false ->
+	  mk_load(hipe_ppc:ldop_word(), Temp, SrcOffset, mk_sp(),
+		  mk_store(hipe_ppc:stop_word(), Temp, DstOffset, mk_sp(), []))
+      end
+  end.
+
 do_pseudo_fmove(I, Context, FPoff) ->
   Dst = hipe_ppc:pseudo_fmove_dst(I),
   Src = hipe_ppc:pseudo_fmove_src(I),
@@ -115,6 +135,22 @@ do_pseudo_fmove(I, Context, FPoff) ->
       end
   end.
 
+do_pseudo_spill_fmove(I, Context, FPoff) ->
+  #pseudo_spill_fmove{dst=Dst,temp=Temp,src=Src} = I,
+  case temp_is_pseudo(Src) andalso temp_is_pseudo(Dst) of
+    false -> % Register allocator changed its mind, turn back to move
+      do_pseudo_fmove(hipe_ppc:mk_pseudo_fmove(Dst, Src), Context, FPoff);
+    true ->
+      SrcOffset = pseudo_offset(Src, FPoff, Context),
+      DstOffset = pseudo_offset(Dst, FPoff, Context),
+      case SrcOffset =:= DstOffset of
+	true -> []; % omit move-to-self
+	false ->
+	  hipe_ppc:mk_fload(Temp, SrcOffset, mk_sp(), 0)
+	    ++ hipe_ppc:mk_fstore(Temp, DstOffset, mk_sp(), 0)
+      end
+  end.
+
 pseudo_offset(Temp, FPoff, Context) ->
   FPoff + context_offset(Context, Temp).
 
diff --git a/lib/hipe/ppc/hipe_ppc_ra_finalise.erl b/lib/hipe/ppc/hipe_ppc_ra_finalise.erl
index 74ef7475eb..bca504d754 100644
--- a/lib/hipe/ppc/hipe_ppc_ra_finalise.erl
+++ b/lib/hipe/ppc/hipe_ppc_ra_finalise.erl
@@ -41,6 +41,7 @@ ra_insn(I, Map, FPMap) ->
     #mtspr{} -> ra_mtspr(I, Map);
     #pseudo_li{} -> ra_pseudo_li(I, Map);
     #pseudo_move{} -> ra_pseudo_move(I, Map);
+    #pseudo_spill_move{} -> ra_pseudo_spill_move(I, Map);
     #pseudo_tailcall{} -> ra_pseudo_tailcall(I, Map);
     #store{} -> ra_store(I, Map);
     #storex{} -> ra_storex(I, Map);
@@ -52,6 +53,7 @@ ra_insn(I, Map, FPMap) ->
     #fp_binary{} -> ra_fp_binary(I, FPMap);
     #fp_unary{} -> ra_fp_unary(I, FPMap);
     #pseudo_fmove{} -> ra_pseudo_fmove(I, FPMap);
+    #pseudo_spill_fmove{} -> ra_pseudo_spill_fmove(I, FPMap);
     _ -> I
   end.
 
@@ -98,6 +100,12 @@ ra_pseudo_move(I=#pseudo_move{dst=Dst,src=Src}, Map) ->
   NewSrc = ra_temp(Src, Map),
   I#pseudo_move{dst=NewDst,src=NewSrc}.
 
+ra_pseudo_spill_move(I=#pseudo_spill_move{dst=Dst,temp=Temp,src=Src}, Map) ->
+  NewDst = ra_temp(Dst, Map),
+  NewTemp = ra_temp(Temp, Map),
+  NewSrc = ra_temp(Src, Map),
+  I#pseudo_spill_move{dst=NewDst,temp=NewTemp,src=NewSrc}.
+
 ra_pseudo_tailcall(I=#pseudo_tailcall{stkargs=StkArgs}, Map) ->
   NewStkArgs = ra_args(StkArgs, Map),
   I#pseudo_tailcall{stkargs=NewStkArgs}.
@@ -156,6 +164,13 @@ ra_pseudo_fmove(I=#pseudo_fmove{dst=Dst,src=Src}, FPMap) ->
   NewSrc = ra_temp_fp(Src, FPMap),
   I#pseudo_fmove{dst=NewDst,src=NewSrc}.
 
+ra_pseudo_spill_fmove(I=#pseudo_spill_fmove{dst=Dst,temp=Temp,src=Src},
+		      FPMap) ->
+  NewDst = ra_temp_fp(Dst, FPMap),
+  NewTemp = ra_temp_fp(Temp, FPMap),
+  NewSrc = ra_temp_fp(Src, FPMap),
+  I#pseudo_spill_fmove{dst=NewDst,temp=NewTemp,src=NewSrc}.
+
 ra_args([Arg|Args], Map) ->
   [ra_temp_or_imm(Arg, Map) | ra_args(Args, Map)];
 ra_args([], _) ->
diff --git a/lib/hipe/ppc/hipe_ppc_ra_postconditions.erl b/lib/hipe/ppc/hipe_ppc_ra_postconditions.erl
index 95aa294fe5..0a97129666 100644
--- a/lib/hipe/ppc/hipe_ppc_ra_postconditions.erl
+++ b/lib/hipe/ppc/hipe_ppc_ra_postconditions.erl
@@ -57,6 +57,7 @@ do_insn(I, TempMap, Strategy) ->
     #mtspr{} -> do_mtspr(I, TempMap, Strategy);
     #pseudo_li{} -> do_pseudo_li(I, TempMap, Strategy);
     #pseudo_move{} -> do_pseudo_move(I, TempMap, Strategy);
+    #pseudo_spill_move{} -> do_pseudo_spill_move(I, TempMap, Strategy);
     #store{} -> do_store(I, TempMap, Strategy);
     #storex{} -> do_storex(I, TempMap, Strategy);
     #unary{} -> do_unary(I, TempMap, Strategy);
@@ -117,18 +118,25 @@ do_pseudo_li(I=#pseudo_li{dst=Dst}, TempMap, Strategy) ->
 
 do_pseudo_move(I=#pseudo_move{dst=Dst,src=Src}, TempMap, Strategy) ->
   %% Either Dst or Src (but not both) may be a pseudo temp.
-  %% pseudo_move and pseudo_tailcall are special cases: in
-  %% all other instructions, all temps must be non-pseudos
-  %% after register allocation.
-  case temp_is_spilled(Dst, TempMap) of
-    true -> % Src must not be a pseudo
-      {FixSrc,NewSrc,DidSpill} = fix_src1(Src, TempMap, Strategy),
-      NewI = I#pseudo_move{src=NewSrc},
-      {FixSrc ++ [NewI], DidSpill};
+  %% pseudo_move, pseudo_spill_move, and pseudo_tailcall are
+  %% special cases: in all other instructions, all temps
+  %% must be non-pseudos after register allocation.
+  case temp_is_spilled(Src, TempMap)
+    andalso temp_is_spilled(Dst, TempMap)
+  of
+    true -> % Turn into pseudo_spill_move
+      Temp = clone(Src, temp1(Strategy)),
+      NewI = #pseudo_spill_move{dst=Dst,temp=Temp,src=Src},
+      {[NewI], true};
     _ ->
       {[I], false}
   end.
 
+do_pseudo_spill_move(I=#pseudo_spill_move{temp=Temp}, TempMap, _Strategy) ->
+  %% Temp is above the low water mark and must not have been spilled
+  false = temp_is_spilled(Temp, TempMap),
+  {[I], false}.
+
 do_store(I=#store{src=Src,base=Base}, TempMap, Strategy) ->
   {FixSrc,NewSrc,DidSpill1} = fix_src1(Src, TempMap, Strategy),
   {FixBase,NewBase,DidSpill2} = fix_src2(Base, TempMap, Strategy),
diff --git a/lib/hipe/ppc/hipe_ppc_ra_postconditions_fp.erl b/lib/hipe/ppc/hipe_ppc_ra_postconditions_fp.erl
index 5ec5f29577..7342053620 100644
--- a/lib/hipe/ppc/hipe_ppc_ra_postconditions_fp.erl
+++ b/lib/hipe/ppc/hipe_ppc_ra_postconditions_fp.erl
@@ -42,6 +42,7 @@ do_insn(I, TempMap) ->
     #fp_binary{} -> do_fp_binary(I, TempMap);
     #fp_unary{} -> do_fp_unary(I, TempMap);
     #pseudo_fmove{} -> do_pseudo_fmove(I, TempMap);
+    #pseudo_spill_fmove{} -> do_pseudo_spill_fmove(I, TempMap);
     _ -> {[I], false}
   end.
 
@@ -81,15 +82,22 @@ do_fp_unary(I=#fp_unary{dst=Dst,src=Src}, TempMap) ->
   {FixSrc ++ [NewI | FixDst], DidSpill1 or DidSpill2}.
 
 do_pseudo_fmove(I=#pseudo_fmove{dst=Dst,src=Src}, TempMap) ->
-  case temp_is_spilled(Dst, TempMap) of
-    true ->
-      {FixSrc,NewSrc,DidSpill} = fix_src(Src, TempMap),
-      NewI = I#pseudo_fmove{src=NewSrc},
-      {FixSrc ++ [NewI], DidSpill};
+  case temp_is_spilled(Src, TempMap)
+    andalso temp_is_spilled(Dst, TempMap)
+  of
+    true -> % Turn into pseudo_spill_fmove
+      Temp = clone(Src),
+      NewI = #pseudo_spill_fmove{dst=Dst,temp=Temp,src=Src},
+      {[NewI], true};
     _ ->
       {[I], false}
   end.
 
+do_pseudo_spill_fmove(I=#pseudo_spill_fmove{temp=Temp}, TempMap) ->
+  %% Temp is above the low water mark and must not have been spilled
+  false = temp_is_spilled(Temp, TempMap),
+  {[I], false}.
+
 %%% Fix Dst and Src operands.
 
 fix_src(Src, TempMap) ->
diff --git a/lib/hipe/ppc/hipe_ppc_subst.erl b/lib/hipe/ppc/hipe_ppc_subst.erl
index 1cd18b5c01..e282b22774 100644
--- a/lib/hipe/ppc/hipe_ppc_subst.erl
+++ b/lib/hipe/ppc/hipe_ppc_subst.erl
@@ -48,6 +48,8 @@ insn_temps(T, I) ->
       #pseudo_call_prepare{} -> I;
       #pseudo_li{dst=D} -> I#pseudo_li{dst=T(D)};
       #pseudo_move{dst=D,src=S} -> I#pseudo_move{dst=T(D),src=T(S)};
+      #pseudo_spill_move{dst=D,temp=U,src=S} ->
+	  I#pseudo_spill_move{dst=T(D),temp=T(U),src=T(S)};
       #pseudo_tailcall{func=F,stkargs=Stk} when not is_record(F, ppc_temp) ->
 	  I#pseudo_tailcall{stkargs=lists:map(A,Stk)};
       #pseudo_tailcall_prepare{} -> I;
@@ -62,7 +64,9 @@ insn_temps(T, I) ->
       #fp_binary{dst=D,src1=L,src2=R} ->
 	  I#fp_binary{dst=T(D),src1=T(L),src2=T(R)};
       #fp_unary{dst=D,src=S} -> I#fp_unary{dst=T(D),src=T(S)};
-      #pseudo_fmove{dst=D,src=S} -> I#pseudo_fmove{dst=T(D),src=T(S)}
+      #pseudo_fmove{dst=D,src=S} -> I#pseudo_fmove{dst=T(D),src=T(S)};
+      #pseudo_spill_fmove{dst=D,temp=U,src=S} ->
+	  I#pseudo_spill_fmove{dst=T(D),temp=T(U),src=T(S)}
   end.
 
 -spec oper_temps(subst_fun(), oper()) -> oper().
diff --git a/lib/hipe/regalloc/hipe_amd64_specific_sse2.erl b/lib/hipe/regalloc/hipe_amd64_specific_sse2.erl
index 8914e7cac3..d592ba391c 100644
--- a/lib/hipe/regalloc/hipe_amd64_specific_sse2.erl
+++ b/lib/hipe/regalloc/hipe_amd64_specific_sse2.erl
@@ -30,6 +30,7 @@
 	 def_use/2,
 	 is_arg/2,	%% used by hipe_ls_regalloc
 	 is_move/2,
+	 is_spill_move/2,
 	 is_fixed/2,	%% used by hipe_graph_coloring_regalloc
 	 is_global/2,
 	 is_precoloured/2,
@@ -194,6 +195,9 @@ is_move(Instruction, _) ->
 	andalso hipe_x86:is_temp(Dst) andalso hipe_x86:temp_is_allocatable(Dst);
     false -> false
   end.
+
+is_spill_move(Instruction,_) ->
+  hipe_x86:is_pseudo_spill_fmove(Instruction).
  
 reg_nr(Reg, _) ->
   hipe_x86:temp_reg(Reg).
diff --git a/lib/hipe/regalloc/hipe_arm_specific.erl b/lib/hipe/regalloc/hipe_arm_specific.erl
index 517ad83982..7ebc6aa336 100644
--- a/lib/hipe/regalloc/hipe_arm_specific.erl
+++ b/lib/hipe/regalloc/hipe_arm_specific.erl
@@ -24,6 +24,7 @@
 	 ,reg_nr/2
 	 ,def_use/2
 	 ,is_move/2
+	 ,is_spill_move/2
 	 ,is_precoloured/2
 	 ,var_range/2
 	 ,allocatable/1
@@ -154,6 +155,9 @@ is_move(Instruction, _) ->
     false -> false
   end.
 
+is_spill_move(Instruction, _) ->
+  hipe_arm:is_pseudo_spill_move(Instruction).
+
 reg_nr(Reg, _) ->
   hipe_arm:temp_reg(Reg).
 
diff --git a/lib/hipe/regalloc/hipe_ppc_specific.erl b/lib/hipe/regalloc/hipe_ppc_specific.erl
index 50ab0f04cf..81bb551bd2 100644
--- a/lib/hipe/regalloc/hipe_ppc_specific.erl
+++ b/lib/hipe/regalloc/hipe_ppc_specific.erl
@@ -24,6 +24,7 @@
 	 ,reg_nr/2
 	 ,def_use/2
 	 ,is_move/2
+	 ,is_spill_move/2
 	 ,is_precoloured/2
 	 ,var_range/2
 	 ,allocatable/1
@@ -154,6 +155,9 @@ is_move(Instruction, _) ->
     false -> false
   end.
 
+is_spill_move(Instruction, _) ->
+  hipe_ppc:is_pseudo_spill_move(Instruction).
+
 reg_nr(Reg, _) ->
   hipe_ppc:temp_reg(Reg).
 
diff --git a/lib/hipe/regalloc/hipe_ppc_specific_fp.erl b/lib/hipe/regalloc/hipe_ppc_specific_fp.erl
index 5a4f63fc78..dcfdf6592c 100644
--- a/lib/hipe/regalloc/hipe_ppc_specific_fp.erl
+++ b/lib/hipe/regalloc/hipe_ppc_specific_fp.erl
@@ -24,6 +24,7 @@
 	 ,reg_nr/2
 	 ,def_use/2
 	 ,is_move/2
+	 ,is_spill_move/2
 	 ,is_precoloured/2
 	 ,var_range/2
 	 ,allocatable/1
@@ -135,6 +136,9 @@ defines_all_alloc(I, _) ->
 is_move(I, _) ->
   hipe_ppc:is_pseudo_fmove(I).
  
+is_spill_move(I, _) ->
+  hipe_ppc:is_pseudo_spill_fmove(I).
+
 reg_nr(Reg, _) ->
   hipe_ppc:temp_reg(Reg).
 
diff --git a/lib/hipe/regalloc/hipe_sparc_specific.erl b/lib/hipe/regalloc/hipe_sparc_specific.erl
index 9eb57dafd2..78b6379eba 100644
--- a/lib/hipe/regalloc/hipe_sparc_specific.erl
+++ b/lib/hipe/regalloc/hipe_sparc_specific.erl
@@ -24,6 +24,7 @@
 	 ,reg_nr/2
 	 ,def_use/2
 	 ,is_move/2
+	 ,is_spill_move/2
 	 ,is_precoloured/2
 	 ,var_range/2
 	 ,allocatable/1
@@ -154,6 +155,9 @@ is_move(Instruction, _) ->
     false -> false
   end.
 
+is_spill_move(Instruction, _) ->
+  hipe_sparc:is_pseudo_spill_move(Instruction).
+
 reg_nr(Reg, _) ->
   hipe_sparc:temp_reg(Reg).
 
diff --git a/lib/hipe/regalloc/hipe_sparc_specific_fp.erl b/lib/hipe/regalloc/hipe_sparc_specific_fp.erl
index 9993af81c2..485fdc212a 100644
--- a/lib/hipe/regalloc/hipe_sparc_specific_fp.erl
+++ b/lib/hipe/regalloc/hipe_sparc_specific_fp.erl
@@ -24,6 +24,7 @@
 	 ,reg_nr/2
 	 ,def_use/2
 	 ,is_move/2
+	 ,is_spill_move/2
 	 ,is_precoloured/2
 	 ,var_range/2
 	 ,allocatable/1
@@ -135,6 +136,9 @@ defines_all_alloc(I, _) ->
 is_move(I, _) ->
   hipe_sparc:is_pseudo_fmove(I).
  
+is_spill_move(I, _) ->
+  hipe_sparc:is_pseudo_spill_fmove(I).
+
 reg_nr(Reg, _) ->
   hipe_sparc:temp_reg(Reg).
 
diff --git a/lib/hipe/regalloc/hipe_x86_specific.erl b/lib/hipe/regalloc/hipe_x86_specific.erl
index f21384d377..dacfb71b00 100644
--- a/lib/hipe/regalloc/hipe_x86_specific.erl
+++ b/lib/hipe/regalloc/hipe_x86_specific.erl
@@ -46,6 +46,7 @@
 	 def_use/2,
 	 is_arg/2,	% used by hipe_ls_regalloc
 	 is_move/2,
+	 is_spill_move/2,
 	 is_fixed/2,	% used by hipe_graph_coloring_regalloc
 	 is_global/2,
 	 is_precoloured/2,
@@ -210,6 +211,9 @@ is_move(Instruction,_) ->
     false -> false
   end.
 
+is_spill_move(Instruction,_) ->
+  hipe_x86:is_pseudo_spill_move(Instruction).
+
 reg_nr(Reg,_) ->
   hipe_x86:temp_reg(Reg).
 
diff --git a/lib/hipe/regalloc/hipe_x86_specific_x87.erl b/lib/hipe/regalloc/hipe_x86_specific_x87.erl
index 4b4c83f76d..3fe49e1f00 100644
--- a/lib/hipe/regalloc/hipe_x86_specific_x87.erl
+++ b/lib/hipe/regalloc/hipe_x86_specific_x87.erl
@@ -47,6 +47,7 @@
 	 uses/2,
 	 defines/2,
 	 defines_all_alloc/2,
+	 is_spill_move/2,
 	 is_global/2,
 	 reg_nr/2,
 	 physical_name/2,
@@ -158,6 +159,9 @@ defines(I, _) ->
 
 defines_all_alloc(I, _) -> hipe_amd64_defuse:insn_defs_all(I).
 
+is_spill_move(I, _) ->
+  hipe_x86:is_pseudo_spill_fmove(I).
+
 temp_is_double(Temp) ->
   hipe_x86:temp_type(Temp) =:= 'double'.
 
diff --git a/lib/hipe/sparc/hipe_sparc.erl b/lib/hipe/sparc/hipe_sparc.erl
index 916857b224..22e0761b69 100644
--- a/lib/hipe/sparc/hipe_sparc.erl
+++ b/lib/hipe/sparc/hipe_sparc.erl
@@ -87,6 +87,9 @@
 
 	 mk_pseudo_set/2,
 
+	 mk_pseudo_spill_move/3,
+	 is_pseudo_spill_move/1,
+
 	 mk_pseudo_tailcall/4,
 	 pseudo_tailcall_funv/1,
 	 pseudo_tailcall_linkage/1,
@@ -117,6 +120,9 @@
 	 pseudo_fmove_src/1,
 	 pseudo_fmove_dst/1,
 
+	 mk_pseudo_spill_fmove/3,
+	 is_pseudo_spill_fmove/1,
+
 	 mk_pseudo_fstore/3,
 	 mk_fstore/4,
 
@@ -269,6 +275,10 @@ mk_pseudo_ret() -> #pseudo_ret{}.
 
 mk_pseudo_set(Imm, Dst) -> #pseudo_set{imm=Imm, dst=Dst}.
 
+mk_pseudo_spill_move(Src, Temp, Dst) ->
+  #pseudo_spill_move{src=Src, temp=Temp, dst=Dst}.
+is_pseudo_spill_move(I) -> is_record(I, pseudo_spill_move).
+
 mk_pseudo_tailcall(FunV, Arity, StkArgs, Linkage) ->
   #pseudo_tailcall{funv=FunV, arity=Arity, stkargs=StkArgs, linkage=Linkage}.
 pseudo_tailcall_funv(#pseudo_tailcall{funv=FunV}) -> FunV.
@@ -375,6 +385,10 @@ is_pseudo_fmove(I) -> case I of #pseudo_fmove{} -> true; _ -> false end.
 pseudo_fmove_src(#pseudo_fmove{src=Src}) -> Src.
 pseudo_fmove_dst(#pseudo_fmove{dst=Dst}) -> Dst.
 
+mk_pseudo_spill_fmove(Src, Temp, Dst) ->
+  #pseudo_spill_fmove{src=Src, temp=Temp, dst=Dst}.
+is_pseudo_spill_fmove(I) -> is_record(I, pseudo_spill_fmove).
+
 mk_pseudo_fstore(Src, Base, Disp) ->
   #pseudo_fstore{src=Src, base=Base, disp=Disp}.
 
diff --git a/lib/hipe/sparc/hipe_sparc.hrl b/lib/hipe/sparc/hipe_sparc.hrl
index 4eae6777a9..f60e516e59 100644
--- a/lib/hipe/sparc/hipe_sparc.hrl
+++ b/lib/hipe/sparc/hipe_sparc.hrl
@@ -88,6 +88,8 @@
 -record(pseudo_move, {src, dst}).
 -record(pseudo_ret, {}).
 -record(pseudo_set, {imm, dst}).
+-record(pseudo_spill_fmove, {src, temp, dst}).
+-record(pseudo_spill_move, {src, temp, dst}).
 -record(pseudo_tailcall, {funv, arity, stkargs, linkage}).
 -record(pseudo_tailcall_prepare, {}).
 -record(rdy, {dst}).
diff --git a/lib/hipe/sparc/hipe_sparc_defuse.erl b/lib/hipe/sparc/hipe_sparc_defuse.erl
index cb75f82e2b..4d4b11e301 100644
--- a/lib/hipe/sparc/hipe_sparc_defuse.erl
+++ b/lib/hipe/sparc/hipe_sparc_defuse.erl
@@ -39,6 +39,7 @@ insn_def_gpr(I) ->
     #pseudo_call{} -> call_clobbered_gpr();
     #pseudo_move{dst=Dst} -> [Dst];
     #pseudo_set{dst=Dst} -> [Dst];
+    #pseudo_spill_move{temp=Temp, dst=Dst} -> [Temp, Dst];
     #pseudo_tailcall_prepare{} -> tailcall_clobbered_gpr();
     #rdy{dst=Dst} -> [Dst];
     #sethi{dst=Dst} -> [Dst];
@@ -72,6 +73,7 @@ insn_use_gpr(I) ->
       funv_use(FunV, arity_use_gpr(Arity));
     #pseudo_move{src=Src} -> [Src];
     #pseudo_ret{} -> [hipe_sparc:mk_rv()];
+    #pseudo_spill_move{src=Src} -> [Src];
     #pseudo_tailcall{funv=FunV,arity=Arity,stkargs=StkArgs} ->
       addsrcs(StkArgs, addtemps(tailcall_clobbered_gpr(), funv_use(FunV, arity_use_gpr(Arity))));
     #store{src=Src,base=Base,disp=Disp} ->
@@ -112,6 +114,7 @@ insn_def_fpr(I) ->
     #fp_unary{dst=Dst} -> [Dst];
     #pseudo_fload{dst=Dst} -> [Dst];
     #pseudo_fmove{dst=Dst} -> [Dst];
+    #pseudo_spill_fmove{temp=Temp, dst=Dst} -> [Temp, Dst];
     _ -> []
   end.
 
@@ -130,6 +133,7 @@ insn_use_fpr(I) ->
     #fp_unary{src=Src} -> [Src];
     #pseudo_fmove{src=Src} -> [Src];
     #pseudo_fstore{src=Src} -> [Src];
+    #pseudo_spill_fmove{src=Src} -> [Src];
     _ -> []
   end.
 
diff --git a/lib/hipe/sparc/hipe_sparc_frame.erl b/lib/hipe/sparc/hipe_sparc_frame.erl
index 6f29c3c905..1f2a259ca1 100644
--- a/lib/hipe/sparc/hipe_sparc_frame.erl
+++ b/lib/hipe/sparc/hipe_sparc_frame.erl
@@ -82,6 +82,10 @@ do_insn(I, LiveOut, Context, FPoff) ->
       {do_pseudo_tailcall(I, Context), context_framesize(Context)};
     #pseudo_fmove{} ->
       {do_pseudo_fmove(I, Context, FPoff), FPoff};
+    #pseudo_spill_move{} ->
+      {do_pseudo_spill_move(I, Context, FPoff), FPoff};
+    #pseudo_spill_fmove{} ->
+      {do_pseudo_spill_fmove(I, Context, FPoff), FPoff};
     _ ->
       {[I], FPoff}
   end.
@@ -110,6 +114,22 @@ do_pseudo_move(I, Context, FPoff) ->
       end
   end.
 
+do_pseudo_spill_move(I, Context, FPoff) ->
+  #pseudo_spill_move{src=Src,temp=Temp,dst=Dst} = I,
+  case temp_is_pseudo(Src) andalso temp_is_pseudo(Dst) of
+    false -> % Register allocator changed its mind, turn back to move
+      do_pseudo_move(hipe_sparc:mk_pseudo_move(Src, Dst), Context, FPoff);
+    true ->
+      SrcOffset = pseudo_offset(Src, FPoff, Context),
+      DstOffset = pseudo_offset(Dst, FPoff, Context),
+      case SrcOffset =:= DstOffset of
+	true -> []; % omit move-to-self
+	false ->
+	  mk_load(hipe_sparc:mk_sp(), SrcOffset, Temp,
+		 mk_store(Temp, hipe_sparc:mk_sp(), DstOffset, []))
+      end
+  end.
+
 do_pseudo_fmove(I, Context, FPoff) ->
   Dst = hipe_sparc:pseudo_fmove_dst(I),
   Src = hipe_sparc:pseudo_fmove_src(I),
@@ -127,6 +147,22 @@ do_pseudo_fmove(I, Context, FPoff) ->
       end
   end.
 
+do_pseudo_spill_fmove(I, Context, FPoff) ->
+  #pseudo_spill_fmove{src=Src,temp=Temp,dst=Dst} = I,
+  case temp_is_pseudo(Src) andalso temp_is_pseudo(Dst) of
+    false -> % Register allocator changed its mind, turn back to fmove
+      do_pseudo_fmove(hipe_sparc:mk_pseudo_fmove(Src, Dst), Context, FPoff);
+    true ->
+      SrcOffset = pseudo_offset(Src, FPoff, Context),
+      DstOffset = pseudo_offset(Dst, FPoff, Context),
+      case SrcOffset =:= DstOffset of
+	true -> []; % omit move-to-self
+	false ->
+	  mk_fload(hipe_sparc:mk_sp(), SrcOffset, Temp)
+	    ++ mk_fstore(Temp, hipe_sparc:mk_sp(), DstOffset)
+      end
+  end.
+
 pseudo_offset(Temp, FPoff, Context) ->
   FPoff + context_offset(Context, Temp).
 
diff --git a/lib/hipe/sparc/hipe_sparc_ra_finalise.erl b/lib/hipe/sparc/hipe_sparc_ra_finalise.erl
index 5fdb73e197..a724821992 100644
--- a/lib/hipe/sparc/hipe_sparc_ra_finalise.erl
+++ b/lib/hipe/sparc/hipe_sparc_ra_finalise.erl
@@ -38,6 +38,7 @@ ra_insn(I, Map, FPMap) ->
     #pseudo_call{} -> ra_pseudo_call(I, Map);
     #pseudo_move{} -> ra_pseudo_move(I, Map);
     #pseudo_set{} -> ra_pseudo_set(I, Map);
+    #pseudo_spill_move{} -> ra_pseudo_spill_move(I, Map);
     #pseudo_tailcall{} -> ra_pseudo_tailcall(I, Map);
     #rdy{} -> ra_rdy(I, Map);
     #sethi{} -> ra_sethi(I, Map);
@@ -47,6 +48,7 @@ ra_insn(I, Map, FPMap) ->
     #pseudo_fload{} -> ra_pseudo_fload(I, Map, FPMap);
     #pseudo_fmove{} -> ra_pseudo_fmove(I, FPMap);
     #pseudo_fstore{} -> ra_pseudo_fstore(I, Map, FPMap);
+    #pseudo_spill_fmove{} -> ra_pseudo_spill_fmove(I, FPMap);
     _ -> I
   end.
 
@@ -80,6 +82,12 @@ ra_pseudo_set(I=#pseudo_set{dst=Dst}, Map) ->
   NewDst = ra_temp(Dst, Map),
   I#pseudo_set{dst=NewDst}.
 
+ra_pseudo_spill_move(I=#pseudo_spill_move{src=Src,temp=Temp,dst=Dst}, Map) ->
+  NewSrc = ra_temp(Src, Map),
+  NewTemp = ra_temp(Temp, Map),
+  NewDst = ra_temp(Dst, Map),
+  I#pseudo_spill_move{src=NewSrc,temp=NewTemp,dst=NewDst}.
+
 ra_pseudo_tailcall(I=#pseudo_tailcall{funv=FunV,stkargs=StkArgs}, Map) ->
   NewFunV = ra_funv(FunV, Map),
   NewStkArgs = ra_args(StkArgs, Map),
@@ -120,6 +128,13 @@ ra_pseudo_fmove(I=#pseudo_fmove{src=Src,dst=Dst}, FPMap) ->
   NewDst = ra_temp_fp(Dst, FPMap),
   I#pseudo_fmove{src=NewSrc,dst=NewDst}.
 
+ra_pseudo_spill_fmove(I=#pseudo_spill_fmove{src=Src,temp=Temp,dst=Dst},
+		      FPMap) ->
+  NewSrc = ra_temp_fp(Src, FPMap),
+  NewTemp = ra_temp_fp(Temp, FPMap),
+  NewDst = ra_temp_fp(Dst, FPMap),
+  I#pseudo_spill_fmove{src=NewSrc,temp=NewTemp,dst=NewDst}.
+
 ra_pseudo_fstore(I=#pseudo_fstore{src=Src,base=Base}, Map, FPMap) ->
   NewSrc = ra_temp_fp(Src, FPMap),
   NewBase = ra_temp(Base, Map),
diff --git a/lib/hipe/sparc/hipe_sparc_ra_postconditions.erl b/lib/hipe/sparc/hipe_sparc_ra_postconditions.erl
index 984c97fbd4..d3ecb43ec6 100644
--- a/lib/hipe/sparc/hipe_sparc_ra_postconditions.erl
+++ b/lib/hipe/sparc/hipe_sparc_ra_postconditions.erl
@@ -54,6 +54,7 @@ do_insn(I, TempMap, Strategy) ->
     #pseudo_call{} -> do_pseudo_call(I, TempMap, Strategy);
     #pseudo_move{} -> do_pseudo_move(I, TempMap, Strategy);
     #pseudo_set{} -> do_pseudo_set(I, TempMap, Strategy);
+    #pseudo_spill_move{} -> do_pseudo_spill_move(I, TempMap, Strategy);
     #pseudo_tailcall{} -> do_pseudo_tailcall(I, TempMap, Strategy);
     #rdy{} -> do_rdy(I, TempMap, Strategy);
     #sethi{} -> do_sethi(I, TempMap, Strategy);
@@ -92,14 +93,16 @@ do_pseudo_call(I=#pseudo_call{funv=FunV}, TempMap, Strategy) ->
 
 do_pseudo_move(I=#pseudo_move{src=Src,dst=Dst}, TempMap, Strategy) ->
   %% Either Dst or Src (but not both) may be a pseudo temp.
-  %% pseudo_move is a special case: in [XXX: not pseudo_tailcall]
-  %% all other instructions, all temps must be non-pseudos
-  %% after register allocation.
-  case temp_is_spilled(Dst, TempMap) of
-    true -> % Src must not be a pseudo
-      {FixSrc,NewSrc,DidSpill} = fix_src1(Src, TempMap, Strategy),
-      NewI = I#pseudo_move{src=NewSrc},
-      {FixSrc ++ [NewI], DidSpill};
+  %% pseudo_move and pseudo_spill_move [XXX: not pseudo_tailcall]
+  %% are special cases: in all other instructions, all temps must
+  %% be non-pseudos after register allocation.
+  case temp_is_spilled(Src, TempMap)
+    andalso temp_is_spilled(Dst, TempMap)
+  of
+    true -> % Turn into pseudo_spill_move
+      Temp = clone(Src, temp1(Strategy)),
+      NewI = #pseudo_spill_move{src=Src,temp=Temp,dst=Dst},
+      {[NewI], true};
     _ ->
       {[I], false}
   end.
@@ -109,6 +112,11 @@ do_pseudo_set(I=#pseudo_set{dst=Dst}, TempMap, Strategy) ->
   NewI = I#pseudo_set{dst=NewDst},
   {[NewI | FixDst], DidSpill}.
 
+do_pseudo_spill_move(I=#pseudo_spill_move{temp=Temp}, TempMap, _Strategy) ->
+  %% Temp is above the low water mark and must not have been spilled
+  false = temp_is_spilled(Temp, TempMap),
+  {[I], false}.
+
 do_pseudo_tailcall(I=#pseudo_tailcall{funv=FunV}, TempMap, Strategy) ->
   {FixFunV,NewFunV,DidSpill} = fix_funv(FunV, TempMap, Strategy),
   NewI = I#pseudo_tailcall{funv=NewFunV},
diff --git a/lib/hipe/sparc/hipe_sparc_ra_postconditions_fp.erl b/lib/hipe/sparc/hipe_sparc_ra_postconditions_fp.erl
index 751e91425c..5fa3a5fc59 100644
--- a/lib/hipe/sparc/hipe_sparc_ra_postconditions_fp.erl
+++ b/lib/hipe/sparc/hipe_sparc_ra_postconditions_fp.erl
@@ -43,6 +43,7 @@ do_insn(I, TempMap) ->
     #pseudo_fload{} -> do_pseudo_fload(I, TempMap);
     #pseudo_fmove{} -> do_pseudo_fmove(I, TempMap);
     #pseudo_fstore{} -> do_pseudo_fstore(I, TempMap);
+    #pseudo_spill_fmove{} -> do_pseudo_spill_fmove(I, TempMap);
     _ -> {[I], false}
   end.
 
@@ -67,11 +68,13 @@ do_pseudo_fload(I=#pseudo_fload{dst=Dst}, TempMap) ->
   {[NewI | FixDst], DidSpill}.
 
 do_pseudo_fmove(I=#pseudo_fmove{src=Src,dst=Dst}, TempMap) ->
-  case temp_is_spilled(Dst, TempMap) of
-    true ->
-      {FixSrc,NewSrc,DidSpill} = fix_src(Src, TempMap),
-      NewI = I#pseudo_fmove{src=NewSrc},
-      {FixSrc ++ [NewI], DidSpill};
+  case temp_is_spilled(Src, TempMap)
+    andalso temp_is_spilled(Dst, TempMap)
+  of
+    true -> % Turn into pseudo_spill_fmove
+      Temp = clone(Src),
+      NewI = #pseudo_spill_fmove{src=Src,temp=Temp,dst=Dst},
+      {[NewI], true};
     _ ->
       {[I], false}
   end.
@@ -81,6 +84,11 @@ do_pseudo_fstore(I=#pseudo_fstore{src=Src}, TempMap) ->
   NewI = I#pseudo_fstore{src=NewSrc},
   {FixSrc ++ [NewI], DidSpill}.
 
+do_pseudo_spill_fmove(I=#pseudo_spill_fmove{temp=Temp}, TempMap) ->
+  %% Temp is above the low water mark and must not have been spilled
+  false = temp_is_spilled(Temp, TempMap),
+  {[I], false}.
+
 %%% Fix Dst and Src operands.
 
 fix_src(Src, TempMap) ->
diff --git a/lib/hipe/sparc/hipe_sparc_subst.erl b/lib/hipe/sparc/hipe_sparc_subst.erl
index 1d0671464e..ce3bbb813a 100644
--- a/lib/hipe/sparc/hipe_sparc_subst.erl
+++ b/lib/hipe/sparc/hipe_sparc_subst.erl
@@ -44,6 +44,8 @@ insn_temps(T, I) ->
       #pseudo_move{src=S,dst=D} -> I#pseudo_move{src=T(S),dst=T(D)};
       #pseudo_ret{} -> I;
       #pseudo_set{dst=D}-> I#pseudo_set{dst=T(D)};
+      #pseudo_spill_move{src=S,temp=U,dst=D} ->
+	  I#pseudo_spill_move{src=T(S),temp=T(U),dst=T(D)};
       #pseudo_tailcall{funv=F,stkargs=Stk} ->
 	  I#pseudo_tailcall{funv=funv_temps(T,F),stkargs=lists:map(Arg,Stk)};
       #pseudo_tailcall_prepare{} -> I;
@@ -57,7 +59,9 @@ insn_temps(T, I) ->
 	  I#pseudo_fload{base=T(B),disp=S2(Di),dst=T(Ds)};
       #pseudo_fmove{src=S,dst=D} -> I#pseudo_fmove{src=T(S),dst=T(D)};
       #pseudo_fstore{src=S,base=B,disp=D} ->
-	  I#pseudo_fstore{src=T(S),base=T(B),disp=S2(D)}
+	  I#pseudo_fstore{src=T(S),base=T(B),disp=S2(D)};
+      #pseudo_spill_fmove{src=S,temp=U,dst=D} ->
+	  I#pseudo_spill_fmove{src=T(S),temp=T(U),dst=T(D)}
   end.
 
 -spec src2_temps(subst_fun(), src2()) -> src2().
diff --git a/lib/hipe/x86/hipe_x86.erl b/lib/hipe/x86/hipe_x86.erl
index cc1c75b04d..f514dd1ded 100644
--- a/lib/hipe/x86/hipe_x86.erl
+++ b/lib/hipe/x86/hipe_x86.erl
@@ -167,6 +167,12 @@
 
      mk_pseudo_spill/1,
 
+	 mk_pseudo_spill_fmove/3,
+	 is_pseudo_spill_fmove/1,
+
+	 mk_pseudo_spill_move/3,
+	 is_pseudo_spill_move/1,
+
 	 mk_pseudo_tailcall/4,
 	 %% is_pseudo_tailcall/1,
 	 pseudo_tailcall_fun/1,
@@ -425,6 +431,14 @@ mk_pseudo_jcc_simple(Cc, TrueLabel, FalseLabel, Pred) ->
 mk_pseudo_spill(List) ->
     #pseudo_spill{args=List}.
 
+mk_pseudo_spill_fmove(Src, Temp, Dst) ->
+    #pseudo_spill_fmove{src=Src, temp=Temp, dst=Dst}.
+is_pseudo_spill_fmove(I) -> is_record(I, pseudo_spill_fmove).
+
+mk_pseudo_spill_move(Src, Temp, Dst) ->
+    #pseudo_spill_move{src=Src, temp=Temp, dst=Dst}.
+is_pseudo_spill_move(I) -> is_record(I, pseudo_spill_move).
+
 mk_pseudo_tailcall(Fun, Arity, StkArgs, Linkage) ->
     check_linkage(Linkage),
     #pseudo_tailcall{'fun'=Fun, arity=Arity, stkargs=StkArgs, linkage=Linkage}.
diff --git a/lib/hipe/x86/hipe_x86.hrl b/lib/hipe/x86/hipe_x86.hrl
index 567848bae5..6cd69905b2 100644
--- a/lib/hipe/x86/hipe_x86.hrl
+++ b/lib/hipe/x86/hipe_x86.hrl
@@ -91,6 +91,8 @@
 -record(pseudo_call, {'fun', sdesc, contlab, linkage}).
 -record(pseudo_jcc, {cc, true_label, false_label, pred}).
 -record(pseudo_spill, {args=[]}).
+-record(pseudo_spill_move, {src, temp, dst}).
+-record(pseudo_spill_fmove, {src, temp, dst}).
 -record(pseudo_tailcall, {'fun', arity, stkargs, linkage}).
 -record(pseudo_tailcall_prepare, {}).
 -record(push, {src}).
diff --git a/lib/hipe/x86/hipe_x86_defuse.erl b/lib/hipe/x86/hipe_x86_defuse.erl
index 5d7fadf8e5..2731836dc1 100644
--- a/lib/hipe/x86/hipe_x86_defuse.erl
+++ b/lib/hipe/x86/hipe_x86_defuse.erl
@@ -51,6 +51,8 @@ insn_def(I) ->
     #movzx{dst=Dst} -> dst_def(Dst);
     #pseudo_call{} -> call_clobbered();
     #pseudo_spill{} -> [];
+    #pseudo_spill_fmove{temp=Temp, dst=Dst} -> [Temp, Dst];
+    #pseudo_spill_move{temp=Temp, dst=Dst} -> [Temp, Dst];
     #pseudo_tailcall_prepare{} -> tailcall_clobbered();
     #shift{dst=Dst} -> dst_def(Dst);
     %% call, cmp, comment, jcc, jmp_fun, jmp_label, jmp_switch, label
@@ -108,6 +110,8 @@ insn_use(I) ->
     #pseudo_call{'fun'=Fun,sdesc=#x86_sdesc{arity=Arity}} ->
       addtemp(Fun, arity_use(Arity));
     #pseudo_spill{args=Args} -> Args;
+    #pseudo_spill_fmove{src=Src} -> [Src];
+    #pseudo_spill_move{src=Src} -> [Src];
     #pseudo_tailcall{'fun'=Fun,arity=Arity,stkargs=StkArgs} ->
       addtemp(Fun, addtemps(StkArgs, addtemps(tailcall_clobbered(),
 					      arity_use(Arity))));
diff --git a/lib/hipe/x86/hipe_x86_frame.erl b/lib/hipe/x86/hipe_x86_frame.erl
index 3c2b67967a..558321d0c3 100644
--- a/lib/hipe/x86/hipe_x86_frame.erl
+++ b/lib/hipe/x86/hipe_x86_frame.erl
@@ -95,13 +95,17 @@ do_insn(I, LiveOut, Context, FPoff) ->
     #imul{} ->
       {[do_imul(I, Context, FPoff)], FPoff};
     #move{} ->
-      {[do_move(I, Context, FPoff)], FPoff};
+      {do_move(I, Context, FPoff), FPoff};
     #movsx{} ->
       {[do_movsx(I, Context, FPoff)], FPoff};
     #movzx{} ->
       {[do_movzx(I, Context, FPoff)], FPoff};
     #pseudo_call{} ->
       do_pseudo_call(I, LiveOut, Context, FPoff);
+    #pseudo_spill_fmove{} ->
+      {do_pseudo_spill_fmove(I, Context, FPoff), FPoff};
+    #pseudo_spill_move{} ->
+      {do_pseudo_spill_move(I, Context, FPoff), FPoff};
     #pseudo_tailcall{} ->
       {do_pseudo_tailcall(I, Context), context_framesize(Context)};
     #push{} ->
@@ -144,22 +148,50 @@ do_fp_binop(I, Context, FPoff) ->
   Dst = conv_opnd(Dst0, FPoff, Context),
   [I#fp_binop{src=Src,dst=Dst}].
 
-do_fmove(I, Context, FPoff) ->
-  #fmove{src=Src0,dst=Dst0} = I,
+do_fmove(I0, Context, FPoff) ->
+  #fmove{src=Src0,dst=Dst0} = I0,
   Src = conv_opnd(Src0, FPoff, Context),
   Dst = conv_opnd(Dst0, FPoff, Context),
-  I#fmove{src=Src,dst=Dst}.
+  I = I0#fmove{src=Src,dst=Dst},
+  case Src =:= Dst of
+    true -> []; % omit move-to-self
+    false -> [I]
+  end.
+
+do_pseudo_spill_fmove(I0, Context, FPoff) ->
+  #pseudo_spill_fmove{src=Src0,temp=Temp0,dst=Dst0} = I0,
+  Src = conv_opnd(Src0, FPoff, Context),
+  Temp = conv_opnd(Temp0, FPoff, Context),
+  Dst = conv_opnd(Dst0, FPoff, Context),
+  case Src =:= Dst of
+    true -> []; % omit move-to-self
+    false -> [#fmove{src=Src, dst=Temp}, #fmove{src=Temp, dst=Dst}]
+  end.
 
 do_imul(I, Context, FPoff) ->
   #imul{src=Src0} = I,
   Src = conv_opnd(Src0, FPoff, Context),
   I#imul{src=Src}.
 
-do_move(I, Context, FPoff) ->
-  #move{src=Src0,dst=Dst0} = I,
+do_move(I0, Context, FPoff) ->
+  #move{src=Src0,dst=Dst0} = I0,
   Src = conv_opnd(Src0, FPoff, Context),
   Dst = conv_opnd(Dst0, FPoff, Context),
-  I#move{src=Src,dst=Dst}.
+  I = I0#move{src=Src,dst=Dst},
+  case Src =:= Dst of
+    true -> []; % omit move-to-self
+    false -> [I]
+  end.
+
+do_pseudo_spill_move(I0, Context, FPoff) ->
+  #pseudo_spill_move{src=Src0,temp=Temp0,dst=Dst0} = I0,
+  Src = conv_opnd(Src0, FPoff, Context),
+  Temp = conv_opnd(Temp0, FPoff, Context),
+  Dst = conv_opnd(Dst0, FPoff, Context),
+  case Src =:= Dst of
+    true -> []; % omit move-to-self
+    false -> [#move{src=Src, dst=Temp}, #move{src=Temp, dst=Dst}]
+  end.
 
 do_movsx(I, Context, FPoff) ->
   #movsx{src=Src0,dst=Dst0} = I,
diff --git a/lib/hipe/x86/hipe_x86_ra_finalise.erl b/lib/hipe/x86/hipe_x86_ra_finalise.erl
index 4273e3cee8..e8abe78e00 100644
--- a/lib/hipe/x86/hipe_x86_ra_finalise.erl
+++ b/lib/hipe/x86/hipe_x86_ra_finalise.erl
@@ -140,6 +140,16 @@ ra_insn(I, Map, FpMap) ->
       I#pseudo_call{'fun'=Fun};
     #pseudo_jcc{} ->
       I;
+    #pseudo_spill_fmove{src=Src0, temp=Temp0, dst=Dst0} ->
+      Src = ra_opnd(Src0, Map, FpMap),
+      Temp = ra_opnd(Temp0, Map, FpMap),
+      Dst = ra_opnd(Dst0, Map, FpMap),
+      I#pseudo_spill_fmove{src=Src, temp=Temp, dst=Dst};
+    #pseudo_spill_move{src=Src0, temp=Temp0, dst=Dst0} ->
+      Src = ra_opnd(Src0, Map),
+      Temp = ra_opnd(Temp0, Map),
+      Dst = ra_opnd(Dst0, Map),
+      I#pseudo_spill_move{src=Src, temp=Temp, dst=Dst};
     #pseudo_tailcall{'fun'=Fun0,stkargs=StkArgs0} ->
       Fun = ra_opnd(Fun0, Map),
       StkArgs = ra_args(StkArgs0, Map),
diff --git a/lib/hipe/x86/hipe_x86_ra_postconditions.erl b/lib/hipe/x86/hipe_x86_ra_postconditions.erl
index 28ec9c4277..db6391d5c1 100644
--- a/lib/hipe/x86/hipe_x86_ra_postconditions.erl
+++ b/lib/hipe/x86/hipe_x86_ra_postconditions.erl
@@ -74,6 +74,8 @@ do_insn(I, TempMap, Strategy) ->	% Insn -> {Insn list, DidSpill}
       do_movx(I, TempMap, Strategy);
     #fmove{} ->
       do_fmove(I, TempMap, Strategy);
+    #pseudo_spill_move{} ->
+      do_pseudo_spill_move(I, TempMap, Strategy);
     #shift{} ->
       do_shift(I, TempMap, Strategy);
     #test{} ->
@@ -190,10 +192,19 @@ do_lea(I, TempMap, Strategy) ->
 
 do_move(I, TempMap, Strategy) ->
   #move{src=Src0,dst=Dst0} = I,
-  {FixSrc, Src, FixDst, Dst, DidSpill} =
-    do_check_byte_move(Src0, Dst0, TempMap, Strategy),
-  {FixSrc ++ FixDst ++ [I#move{src=Src,dst=Dst}],
-   DidSpill}.
+  case
+    is_record(Src0, x86_temp) andalso is_record(Dst0, x86_temp)
+    andalso is_spilled(Src0, TempMap) andalso is_spilled(Dst0, TempMap)
+  of
+    true ->
+      Tmp = clone(Src0, Strategy),
+      {[hipe_x86:mk_pseudo_spill_move(Src0, Tmp, Dst0)], true};
+    false ->
+      {FixSrc, Src, FixDst, Dst, DidSpill} =
+	do_check_byte_move(Src0, Dst0, TempMap, Strategy),
+      {FixSrc ++ FixDst ++ [I#move{src=Src,dst=Dst}],
+       DidSpill}
+  end.
 
 -ifdef(HIPE_AMD64).
 
@@ -287,6 +298,13 @@ do_fmove(I, TempMap, Strategy) ->
   {FixSrc ++ FixDst ++ [I#fmove{src=Src,dst=Dst}],
    DidSpill1 or DidSpill2}.
 
+%%% Fix an pseudo_spill_move op.
+
+do_pseudo_spill_move(I = #pseudo_spill_move{temp=Temp}, TempMap, _Strategy) ->
+  %% Temp is above the low water mark and must not have been spilled
+  false = is_spilled(Temp, TempMap),
+  {[I], false}. % nothing to do
+
 %%% Fix a shift operation.
 %%% 1. remove pseudos from any explicit memory operands
 %%% 2. if the source is a register or memory position
diff --git a/lib/hipe/x86/hipe_x86_subst.erl b/lib/hipe/x86/hipe_x86_subst.erl
index adc2e9dc65..7db3b23d92 100644
--- a/lib/hipe/x86/hipe_x86_subst.erl
+++ b/lib/hipe/x86/hipe_x86_subst.erl
@@ -59,6 +59,10 @@ insn_temps(SubstTemp, I) ->
       I#jmp_switch{temp=O(T), jtab=jtab_temps(SubstTemp, J)};
     #pseudo_call{'fun'=F} ->
       I#pseudo_call{'fun'=funv_temps(SubstTemp, F)};
+    #pseudo_spill_fmove{src=S, temp=T, dst=D} ->
+      I#pseudo_spill_fmove{src=O(S), temp=O(T), dst=O(D)};
+    #pseudo_spill_move{src=S, temp=T, dst=D} ->
+      I#pseudo_spill_move{src=O(S), temp=O(T), dst=O(D)};
     #pseudo_tailcall{'fun'=F, stkargs=Stk} ->
       I#pseudo_tailcall{'fun'=funv_temps(SubstTemp, F),
 			stkargs=lists:map(O, Stk)};
-- 
cgit v1.2.3