hipe_rtl_to_x86: Use LEA only for immediate adds

It seems that most 3-address adds of temps can be move coalesced.
Therefore, we limit the behaviour added by 1567585dda8 to only affect
immediate adds.

Also, add conversion of immediate mov+sub to lea.

2 files changed, 15 insertions, 3 deletions

diff --git a/lib/hipe/x86/hipe_rtl_to_x86.erl b/lib/hipe/x86/hipe_rtl_to_x86.erl
index ccb9b7632b..851b7da2dd 100644
--- a/lib/hipe/x86/hipe_rtl_to_x86.erl
+++ b/lib/hipe/x86/hipe_rtl_to_x86.erl
@@ -257,7 +257,9 @@ conv_insn(I, Map, Data) ->
 
 conv_alu_nocc(Dst, Src1, 'add', Src2, Tail) ->
   case (not same_opnd(Dst, Src1)) andalso (not same_opnd(Dst, Src2))
-    andalso (hipe_x86:is_temp(Src1) orelse hipe_x86:is_temp(Src2))
+    %% We could use orelse instead of xor here to generate lea T1(T2), T3, but
+    %% they seem to move coalesce so well that move+add is better for them.
+    andalso (hipe_x86:is_temp(Src1) xor hipe_x86:is_temp(Src2))
   of
     false -> conv_alu(Dst, Src1, 'add', Src2, Tail);
     true -> % Use LEA
@@ -268,6 +270,16 @@ conv_alu_nocc(Dst, Src1, 'add', Src2, Tail) ->
 	    end,
       [hipe_x86:mk_lea(Mem, Dst) | Tail]
   end;
+conv_alu_nocc(Dst, Src1, 'sub', Src2, Tail) ->
+  case (not same_opnd(Dst, Src1)) andalso hipe_x86:is_temp(Src1)
+    andalso (not hipe_x86:is_temp(Src2))
+  of
+    false -> conv_alu(Dst, Src1, 'sub', Src2, Tail);
+    true -> % Use LEA
+      Imm = hipe_x86:mk_imm(-hipe_x86:imm_value(Src2)),
+      Mem = hipe_x86:mk_mem(Src1, Imm, typeof_dst(Dst)),
+      [hipe_x86:mk_lea(Mem, Dst) | Tail]
+  end;
 conv_alu_nocc(Dst, Src1, BinOp, Src2, Tail) ->
   conv_alu(Dst, Src1, BinOp, Src2, Tail).
 
diff --git a/lib/hipe/x86/hipe_x86.erl b/lib/hipe/x86/hipe_x86.erl
index e7c4497cda..95af3f9c67 100644
--- a/lib/hipe/x86/hipe_x86.erl
+++ b/lib/hipe/x86/hipe_x86.erl
@@ -37,7 +37,7 @@
 	 mk_imm_from_addr/2,
 	 mk_imm_from_atom/1,
 	 is_imm/1,
-	 %% imm_value/1,
+	 imm_value/1,
 
 	 mk_mem/3,
 	 %% is_mem/1,
@@ -241,7 +241,7 @@ mk_imm_from_addr(Addr, Type) ->
 mk_imm_from_atom(Atom) ->
     mk_imm(Atom).
 is_imm(X) -> case X of #x86_imm{} -> true; _ -> false end.
-%% imm_value(#x86_imm{value=Value}) -> Value.
+imm_value(#x86_imm{value=Value}) -> Value.
 
 mk_mem(Base, Off, Type) -> #x86_mem{base=Base, off=Off, type=Type}.
 %% is_mem(X) -> case X of #x86_mem{} -> true; _ -> false end.