aboutsummaryrefslogtreecommitdiffstats
path: root/lib/hipe/x86/hipe_rtl_to_x86.erl
diff options
context:
space:
mode:
authorSverker Eriksson <[email protected]>2016-09-02 14:51:25 +0200
committerSverker Eriksson <[email protected]>2016-09-02 14:51:25 +0200
commitc2f8b61ca3682281752fa0984699214dfcbf7ccd (patch)
tree3f44fa5c58d0d0d845045c3c5535aefad333b6dd /lib/hipe/x86/hipe_rtl_to_x86.erl
parent87643cf92c061d7518299fdebb326e315c32e528 (diff)
parenta19e3f0e1e82b793d58f9ef0db907ba637793fb6 (diff)
downloadotp-c2f8b61ca3682281752fa0984699214dfcbf7ccd.tar.gz
otp-c2f8b61ca3682281752fa0984699214dfcbf7ccd.tar.bz2
otp-c2f8b61ca3682281752fa0984699214dfcbf7ccd.zip
Merge branch 'sverker/hipe-performance-o1/PR-1154/OTP-13862'
* sverker/hipe-performance-o1/PR-1154: hipe_sparc: Minimise CFG<->linear conversions hipe_ppc: Minimise CFG<->linear conversions hipe_arm: Minimise CFG<->linear conversions hipe_x86: Use lea instead of move+add hipe_arm: Improve peephole optimiser hipe_arm: Be resilient to crappy RTL hipe_ppc: Be resilient to crappy RTL hipe_sparc: Be resilient to crappy RTL hipe: Reuse liveness info for spillmin hipe_x86: Minimise CFG<->linear conversions hipe: Fix o0 and o1 hipe: Add o0 and o1 to tests hipe_rtl_binary:get_word_integer/4: Handle imms hipe_x86: Be resilient to crappy RTL hipe_x86: LSRA for SSE2
Diffstat (limited to 'lib/hipe/x86/hipe_rtl_to_x86.erl')
-rw-r--r--lib/hipe/x86/hipe_rtl_to_x86.erl37
1 files changed, 32 insertions, 5 deletions
diff --git a/lib/hipe/x86/hipe_rtl_to_x86.erl b/lib/hipe/x86/hipe_rtl_to_x86.erl
index d13f63b1d9..4c8c98551c 100644
--- a/lib/hipe/x86/hipe_rtl_to_x86.erl
+++ b/lib/hipe/x86/hipe_rtl_to_x86.erl
@@ -85,7 +85,7 @@ conv_insn(I, Map, Data) ->
true ->
conv_shift(Dst, Src1, BinOp, Src2);
false ->
- conv_alu(Dst, Src1, BinOp, Src2, [])
+ conv_alu_nocc(Dst, Src1, BinOp, Src2, [])
end,
{FixSrc1++FixSrc2++I2, Map2, Data};
#alub{} ->
@@ -144,7 +144,7 @@ conv_insn(I, Map, Data) ->
{I2, Map, Data};
#load{} ->
{Dst, Map0} = conv_dst(hipe_rtl:load_dst(I), Map),
- {FixSrc, Src, Map1} = conv_src(hipe_rtl:load_src(I), Map0),
+ {FixSrc, Src, Map1} = conv_src_noimm(hipe_rtl:load_src(I), Map0),
{FixOff, Off, Map2} = conv_src(hipe_rtl:load_offset(I), Map1),
I2 = case {hipe_rtl:load_size(I), hipe_rtl:load_sign(I)} of
{byte, signed} ->
@@ -171,6 +171,7 @@ conv_insn(I, Map, Data) ->
Src = hipe_x86:mk_imm_from_atom(hipe_rtl:load_atom_atom(I)),
I2 = [hipe_x86:mk_move(Src, Dst)],
{I2, Map0, Data};
+ #move{src=Dst, dst=Dst} -> {[], Map, Data};
#move{} ->
{Dst, Map0} = conv_dst(hipe_rtl:move_dst(I), Map),
{FixSrc, Src, Map1} = conv_src(hipe_rtl:move_src(I), Map0),
@@ -182,11 +183,11 @@ conv_insn(I, Map, Data) ->
I2 = move_retvals(Args, [hipe_x86:mk_ret(-1)]),
{FixArgs++I2, Map0, Data};
#store{} ->
- {Ptr, Map0} = conv_dst(hipe_rtl:store_base(I), Map),
+ {FixPtr, Ptr, Map0} = conv_src_noimm(hipe_rtl:store_base(I), Map),
{FixSrc, Src, Map1} = conv_src(hipe_rtl:store_src(I), Map0),
{FixOff, Off, Map2} = conv_src(hipe_rtl:store_offset(I), Map1),
I2 = mk_store(hipe_rtl:store_size(I), Src, Ptr, Off),
- {FixSrc++FixOff++I2, Map2, Data};
+ {FixPtr++FixSrc++FixOff++I2, Map2, Data};
#switch{} -> % this one also updates Data :-(
%% from hipe_rtl2sparc, but we use a hairy addressing mode
%% instead of doing the arithmetic manually
@@ -206,7 +207,7 @@ conv_insn(I, Map, Data) ->
{I2, Map1, NewData};
#fload{} ->
{Dst, Map0} = conv_dst(hipe_rtl:fload_dst(I), Map),
- {[], Src, Map1} = conv_src(hipe_rtl:fload_src(I), Map0),
+ {[], Src, Map1} = conv_src_noimm(hipe_rtl:fload_src(I), Map0),
{[], Off, Map2} = conv_src(hipe_rtl:fload_offset(I), Map1),
I2 = [hipe_x86:mk_fmove(hipe_x86:mk_mem(Src, Off, 'double'),Dst)],
{I2, Map2, Data};
@@ -249,6 +250,22 @@ conv_insn(I, Map, Data) ->
%%% Finalise the conversion of a 3-address ALU operation, taking
%%% care to not introduce more temps and moves than necessary.
+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))
+ of
+ false -> conv_alu(Dst, Src1, 'add', Src2, Tail);
+ true -> % Use LEA
+ Type = typeof_dst(Dst),
+ Mem = case hipe_x86:is_temp(Src1) of
+ true -> hipe_x86:mk_mem(Src1, Src2, Type);
+ false -> hipe_x86:mk_mem(Src2, Src1, Type)
+ end,
+ [hipe_x86:mk_lea(Mem, Dst) | Tail]
+ end;
+conv_alu_nocc(Dst, Src1, BinOp, Src2, Tail) ->
+ conv_alu(Dst, Src1, BinOp, Src2, Tail).
+
conv_alu(Dst, Src1, 'imul', Src2, Tail) ->
mk_imul(Src1, Src2, Dst, Tail);
conv_alu(Dst, Src1, BinOp, Src2, Tail) ->
@@ -572,6 +589,16 @@ conv_fun(Fun, Map) ->
end
end.
+conv_src_noimm(Opnd, Map) ->
+ R={FixSrc0, Src, NewMap} = conv_src(Opnd, Map),
+ case hipe_x86:is_imm(Src) of
+ false -> R;
+ true ->
+ Tmp = new_untagged_temp(),
+ {FixSrc0 ++ [hipe_x86:mk_move(Src, Tmp)],
+ Tmp, NewMap}
+ end.
+
%%% Convert an RTL source operand (imm/var/reg).
conv_src(Opnd, Map) ->