diff options
Diffstat (limited to 'lib/compiler/src/beam_ssa_codegen.erl')
| -rw-r--r-- | lib/compiler/src/beam_ssa_codegen.erl | 334 |
1 files changed, 292 insertions, 42 deletions
diff --git a/lib/compiler/src/beam_ssa_codegen.erl b/lib/compiler/src/beam_ssa_codegen.erl index 357352269c..c2d5035b19 100644 --- a/lib/compiler/src/beam_ssa_codegen.erl +++ b/lib/compiler/src/beam_ssa_codegen.erl @@ -108,7 +108,8 @@ module(#b_module{name=Mod,exports=Es,attributes=Attrs,body=Fs}, _Opts) -> -type ssa_register() :: xreg() | yreg() | {'fr',reg_num()} | {'z',reg_num()}. functions(Forms, AtomMod) -> - mapfoldl(fun (F, St) -> function(F, AtomMod, St) end, #cg{lcount=1}, Forms). + mapfoldl(fun (F, St) -> function(F, AtomMod, St) end, + #cg{lcount=1}, Forms). function(#b_function{anno=Anno,bs=Blocks}, AtomMod, St0) -> #{func_info:={_,Name,Arity}} = Anno, @@ -125,8 +126,9 @@ function(#b_function{anno=Anno,bs=Blocks}, AtomMod, St0) -> ultimate_fail=Ult}, {Body,St} = cg_fun(Blocks, St5), Asm = [{label,Fi},line(Anno), - {func_info,AtomMod,{atom,Name},Arity}] ++ Body ++ - [{label,Ult},if_end], + {func_info,AtomMod,{atom,Name},Arity}] ++ + add_parameter_annos(Body, Anno) ++ + [{label,Ult},if_end], Func = {function,Name,Arity,Entry,Asm}, {Func,St} catch @@ -150,6 +152,17 @@ assert_badarg_block(Blocks) -> ok end. +add_parameter_annos([{label, _}=Entry | Body], Anno) -> + ParamInfo = maps:get(parameter_type_info, Anno, #{}), + Annos = maps:fold( + fun(K, V, Acc) when is_map_key(K, ParamInfo) -> + TypeInfo = maps:get(K, ParamInfo), + [{'%', {type_info, V, TypeInfo}} | Acc]; + (_K, _V, Acc) -> + Acc + end, [], maps:get(registers, Anno)), + [Entry | sort(Annos)] ++ Body. + cg_fun(Blocks, St0) -> Linear0 = linearize(Blocks), St = collect_catch_labels(Linear0, St0), @@ -218,7 +231,7 @@ need_heap_never(_) -> false. need_heap_blks([{L,#cg_blk{is=Is0}=Blk0}|Bs], H0, Acc) -> {Is1,H1} = need_heap_is(reverse(Is0), H0, []), - {Ns,H} = need_heap_terminator(Bs, H1), + {Ns,H} = need_heap_terminator(Bs, L, H1), Is = Ns ++ Is1, Blk = Blk0#cg_blk{is=Is}, need_heap_blks(Bs, H, [{L,Blk}|Acc]); @@ -228,6 +241,13 @@ need_heap_blks([], H, Acc) -> need_heap_is([#cg_alloc{words=Words}=Alloc0|Is], N, Acc) -> Alloc = Alloc0#cg_alloc{words=add_heap_words(N, Words)}, need_heap_is(Is, #need{}, [Alloc|Acc]); +need_heap_is([#cg_set{anno=Anno,op=bs_init}=I0|Is], N, Acc) -> + Alloc = case need_heap_need(N) of + [#cg_alloc{words=Need}] -> alloc(Need); + [] -> 0 + end, + I = I0#cg_set{anno=Anno#{alloc=>Alloc}}, + need_heap_is(Is, #need{}, [I|Acc]); need_heap_is([#cg_set{op=Op,args=Args}=I|Is], N, Acc) -> case classify_heap_need(Op, Args) of {put,Words} -> @@ -243,11 +263,31 @@ need_heap_is([#cg_set{op=Op,args=Args}=I|Is], N, Acc) -> need_heap_is([], N, Acc) -> {Acc,N}. -need_heap_terminator([{_,#cg_blk{last=#cg_br{succ=Same,fail=Same}}}|_], N) -> +need_heap_terminator([{_,#cg_blk{last=#cg_br{succ=L,fail=L}}}|_], L, N) -> + %% Fallthrough. {[],N}; -need_heap_terminator([{_,#cg_blk{}}|_], N) -> +need_heap_terminator([{_,#cg_blk{is=Is,last=#cg_br{succ=L}}}|_], L, N) -> + case need_heap_need(N) of + [] -> + {[],#need{}}; + [_|_]=Alloc -> + %% If the preceding instructions are a binary construction, + %% hoist the allocation and incorporate into the bs_init + %% instruction. + case reverse(Is) of + [#cg_set{op=succeeded},#cg_set{op=bs_init}|_] -> + {[],N}; + [#cg_set{op=bs_put}|_] -> + {[],N}; + _ -> + %% Not binary construction. Must emit an allocation + %% instruction in this block. + {Alloc,#need{}} + end + end; +need_heap_terminator([{_,#cg_blk{}}|_], _, N) -> {need_heap_need(N),#need{}}; -need_heap_terminator([], H) -> +need_heap_terminator([], _, H) -> {need_heap_need(H),#need{}}. need_heap_need(#need{h=0,f=0}) -> []; @@ -286,6 +326,8 @@ classify_heap_need(put_list, _) -> {put,2}; classify_heap_need(put_tuple_arity, [#b_literal{val=Words}]) -> {put,Words+1}; +classify_heap_need(put_tuple, Elements) -> + {put,length(Elements)+1}; classify_heap_need({bif,Name}, Args) -> case is_gc_bif(Name, Args) of false -> neutral; @@ -313,12 +355,15 @@ classify_heap_need(Name, _Args) -> classify_heap_need(bs_add) -> gc; classify_heap_need(bs_get) -> gc; +classify_heap_need(bs_get_tail) -> gc; classify_heap_need(bs_init) -> gc; classify_heap_need(bs_init_writable) -> gc; classify_heap_need(bs_match_string) -> gc; classify_heap_need(bs_put) -> neutral; classify_heap_need(bs_restore) -> neutral; classify_heap_need(bs_save) -> neutral; +classify_heap_need(bs_get_position) -> gc; +classify_heap_need(bs_set_position) -> neutral; classify_heap_need(bs_skip) -> gc; classify_heap_need(bs_start_match) -> neutral; classify_heap_need(bs_test_tail) -> neutral; @@ -327,7 +372,6 @@ classify_heap_need(bs_utf8_size) -> neutral; classify_heap_need(build_stacktrace) -> gc; classify_heap_need(call) -> gc; classify_heap_need(catch_end) -> gc; -classify_heap_need(context_to_binary) -> gc; classify_heap_need(copy) -> neutral; classify_heap_need(extract) -> gc; classify_heap_need(get_hd) -> neutral; @@ -590,8 +634,7 @@ liveness_successors(Terminator) -> liveness_is([#cg_alloc{}=I0|Is], Regs, Live, Acc) -> I = I0#cg_alloc{live=num_live(Live, Regs)}, liveness_is(Is, Regs, Live, [I|Acc]); -liveness_is([#cg_set{dst=Dst0,args=Args}=I0|Is], Regs, Live0, Acc) -> - #b_var{name=Dst} = Dst0, +liveness_is([#cg_set{dst=Dst,args=Args}=I0|Is], Regs, Live0, Acc) -> Live1 = liveness_clobber(I0, Live0, Regs), I1 = liveness_yregs_anno(I0, Live1, Regs), Live2 = liveness_args(Args, Live1), @@ -608,7 +651,7 @@ liveness_terminator(#cg_switch{arg=Arg}, Live) -> liveness_terminator(#cg_ret{arg=Arg}, Live) -> liveness_terminator_1(Arg, Live). -liveness_terminator_1(#b_var{name=V}, Live) -> +liveness_terminator_1(#b_var{}=V, Live) -> ordsets:add_element(V, Live); liveness_terminator_1(#b_literal{}, Live) -> Live; @@ -616,7 +659,7 @@ liveness_terminator_1(Reg, Live) -> _ = verify_beam_register(Reg), ordsets:add_element(Reg, Live). -liveness_args([#b_var{name=V}|As], Live) -> +liveness_args([#b_var{}=V|As], Live) -> liveness_args(As, ordsets:add_element(V, Live)); liveness_args([#b_remote{mod=Mod,name=Name}|As], Live) -> liveness_args([Mod,Name|As], Live); @@ -639,7 +682,7 @@ liveness_anno(#cg_set{op=Op}=I, Live, Regs) -> I end. -liveness_yregs_anno(#cg_set{op=Op,dst=#b_var{name=Dst}}=I, Live0, Regs) -> +liveness_yregs_anno(#cg_set{op=Op,dst=Dst}=I, Live0, Regs) -> case need_live_anno(Op) of true -> Live = ordsets:del_element(Dst, Live0), @@ -694,6 +737,8 @@ need_live_anno(Op) -> {bif,_} -> true; bs_get -> true; bs_init -> true; + bs_get_position -> true; + bs_get_tail -> true; bs_start_match -> true; bs_skip -> true; call -> true; @@ -702,7 +747,16 @@ need_live_anno(Op) -> end. %%% -%%% Add annotations for defined Y registers. +%%% Add the following annotations for Y registers: +%%% +%%% def_yregs An ordset with variables that refer to live Y registers. +%%% That is, Y registers that that have been killed +%%% are not included. This annotation is added to all +%%% instructions that require Y registers to be initialized. +%%% +%%% kill_yregs This annotation is added to call instructions. It is +%%% an ordset containing variables referring to Y registers +%%% that will no longer be used after the call instruction. %%% defined(Linear, #cg{regs=Regs}) -> @@ -726,13 +780,13 @@ def_get(L, DefMap) -> def_is([#cg_alloc{anno=Anno0}=I0|Is], Regs, Def, Acc) -> I = I0#cg_alloc{anno=Anno0#{def_yregs=>Def}}, def_is(Is, Regs, Def, [I|Acc]); -def_is([#cg_set{op=kill_try_tag,args=[#b_var{name=Tag}]}=I|Is], Regs, Def0, Acc) -> +def_is([#cg_set{op=kill_try_tag,args=[#b_var{}=Tag]}=I|Is], Regs, Def0, Acc) -> Def = ordsets:del_element(Tag, Def0), def_is(Is, Regs, Def, [I|Acc]); -def_is([#cg_set{op=catch_end,args=[#b_var{name=Tag}|_]}=I|Is], Regs, Def0, Acc) -> +def_is([#cg_set{op=catch_end,args=[#b_var{}=Tag|_]}=I|Is], Regs, Def0, Acc) -> Def = ordsets:del_element(Tag, Def0), def_is(Is, Regs, Def, [I|Acc]); -def_is([#cg_set{anno=Anno0,op=call,dst=#b_var{name=Dst}}=I0|Is], +def_is([#cg_set{anno=Anno0,op=call,dst=Dst}=I0|Is], Regs, Def0, Acc) -> #{live_yregs:=LiveYregVars} = Anno0, LiveRegs = gb_sets:from_list([maps:get(V, Regs) || V <- LiveYregVars]), @@ -747,7 +801,7 @@ def_is([#cg_set{anno=Anno0,op=call,dst=#b_var{name=Dst}}=I0|Is], Def1 = ordsets:subtract(Def0, Kill), Def = def_add_yreg(Dst, Def1, Regs), def_is(Is, Regs, Def, [I|Acc]); -def_is([#cg_set{anno=Anno0,op={bif,Bif},dst=#b_var{name=Dst},args=Args}=I0|Is], +def_is([#cg_set{anno=Anno0,op={bif,Bif},dst=Dst,args=Args}=I0|Is], Regs, Def0, Acc) -> Arity = length(Args), I = case is_gc_bif(Bif, Args) orelse not erl_bifs:is_safe(erlang, Bif, Arity) of @@ -758,7 +812,7 @@ def_is([#cg_set{anno=Anno0,op={bif,Bif},dst=#b_var{name=Dst},args=Args}=I0|Is], end, Def = def_add_yreg(Dst, Def0, Regs), def_is(Is, Regs, Def, [I|Acc]); -def_is([#cg_set{anno=Anno0,dst=#b_var{name=Dst}}=I0|Is], Regs, Def0, Acc) -> +def_is([#cg_set{anno=Anno0,dst=Dst}=I0|Is], Regs, Def0, Acc) -> I = case need_y_init(I0) of true -> I0#cg_set{anno=Anno0#{def_yregs=>Def0}}; @@ -793,6 +847,8 @@ def_successors([], _, DefMap) -> DefMap. need_y_init(#cg_set{anno=#{clobbers:=Clobbers}}) -> Clobbers; need_y_init(#cg_set{op=bs_get}) -> true; +need_y_init(#cg_set{op=bs_get_position}) -> true; +need_y_init(#cg_set{op=bs_get_tail}) -> true; need_y_init(#cg_set{op=bs_init}) -> true; need_y_init(#cg_set{op=bs_skip,args=[#b_literal{val=Type}|_]}) -> case Type of @@ -816,13 +872,35 @@ opt_allocate(Linear, #cg{regs=Regs}) -> opt_allocate_1([{L,#cg_blk{is=[#cg_alloc{stack=Stk}=I0|Is]}=Blk0}|Bs]=Bs0, Regs) when is_integer(Stk) -> - Yregs = opt_alloc_def(Bs0, gb_sets:singleton(L), []), - I = I0#cg_alloc{def_yregs=Yregs}, - [{L,Blk0#cg_blk{is=[I|Is]}}|opt_allocate_1(Bs, Regs)]; + %% Collect the variables that are initialized by copy + %% instruction in this block. + case ordsets:from_list(opt_allocate_defs(Is, Regs)) of + Yregs when length(Yregs) =:= Stk -> + %% Those copy instructions are sufficient to fully + %% initialize the stack frame. + I = I0#cg_alloc{def_yregs=Yregs}, + [{L,Blk0#cg_blk{is=[I|Is]}}|opt_allocate_1(Bs, Regs)]; + Yregs0 -> + %% Determine a conservative approximation of the Y + %% registers that are guaranteed to be initialized by all + %% successors of this block, and to it add the variables + %% initialized by copy instructions in this block. + Yregs1 = opt_alloc_def(Bs0, gb_sets:singleton(L), []), + Yregs = ordsets:union(Yregs0, Yregs1), + I = I0#cg_alloc{def_yregs=Yregs}, + [{L,Blk0#cg_blk{is=[I|Is]}}|opt_allocate_1(Bs, Regs)] + end; opt_allocate_1([B|Bs], Regs) -> [B|opt_allocate_1(Bs, Regs)]; opt_allocate_1([], _) -> []. +opt_allocate_defs([#cg_set{op=copy,dst=Dst}|Is], Regs) -> + case is_yreg(Dst, Regs) of + true -> [Dst|opt_allocate_defs(Is, Regs)]; + false -> [] + end; +opt_allocate_defs(_, _Regs) -> []. + opt_alloc_def([{L,#cg_blk{is=Is,last=Last}}|Bs], Ws0, Def0) -> case gb_sets:is_member(L, Ws0) of false -> @@ -993,8 +1071,8 @@ cg_block([#cg_set{op={bif,Name},dst=Dst0,args=Args0}]=Is0, {Dst0,Fail}, St0) -> {z,_} -> %% The result of the BIF call will only be used once. Convert to %% a test instruction. - Test = bif_to_test(Name, Args, ensure_label(Fail, St0)), - {Test,St0}; + {Test,St1} = bif_to_test(Name, Args, ensure_label(Fail, St0), St0), + {Test,St1}; _ -> %% Must explicitly call the BIF since the result will be used %% more than once. @@ -1021,12 +1099,13 @@ cg_block([#cg_set{op=bs_init,dst=Dst0,args=Args0,anno=Anno}=I, #cg_set{op=succeeded,dst=Bool}], {Bool,Fail0}, St) -> Fail = bif_fail(Fail0), Line = line(Anno), + Alloc = map_get(alloc, Anno), [#b_literal{val=Kind}|Args1] = Args0, case Kind of new -> [Dst,Size,{integer,Unit}] = beam_args([Dst0|Args1], St), Live = get_live(I), - {[Line|cg_bs_init(Dst, Size, Unit, Live, Fail)],St}; + {[Line|cg_bs_init(Dst, Size, Alloc, Unit, Live, Fail)],St}; private_append -> [Dst,Src,Bits,{integer,Unit}] = beam_args([Dst0|Args1], St), Flags = {field_flags,[]}, @@ -1036,17 +1115,23 @@ cg_block([#cg_set{op=bs_init,dst=Dst0,args=Args0,anno=Anno}=I, [Dst,Src,Bits,{integer,Unit}] = beam_args([Dst0|Args1], St), Flags = {field_flags,[]}, Live = get_live(I), - Is = [Line,{bs_append,Fail,Bits,0,Live,Unit,Src,Flags,Dst}], + Is = [Line,{bs_append,Fail,Bits,Alloc,Live,Unit,Src,Flags,Dst}], {Is,St} end; cg_block([#cg_set{anno=Anno,op=bs_start_match,dst=Ctx0,args=[Bin0]}=I, #cg_set{op=succeeded,dst=Bool}], {Bool,Fail}, St) -> - #{num_slots:=Slots} = Anno, [Dst,Bin1] = beam_args([Ctx0,Bin0], St), {Bin,Pre} = force_reg(Bin1, Dst), Live = get_live(I), - Is = Pre ++ [{test,bs_start_match2,Fail,Live,[Bin,Slots],Dst}], - {Is,St}; + %% num_slots is only set when using the old instructions. + case maps:find(num_slots, Anno) of + {ok, Slots} -> + Is = Pre ++ [{test,bs_start_match2,Fail,Live,[Bin,Slots],Dst}], + {Is,St}; + error -> + Is = Pre ++ [{test,bs_start_match3,Fail,Live,[Bin],Dst}], + {Is,St} + end; cg_block([#cg_set{op=bs_get}=Set, #cg_set{op=succeeded,dst=Bool}], {Bool,Fail}, St) -> {cg_bs_get(Fail, Set, St),St}; @@ -1178,18 +1263,119 @@ cg_copy_1([#cg_set{dst=Dst0,args=Args}|T], St) -> end; cg_copy_1([], _St) -> []. +-define(IS_LITERAL(Val), (Val =:= nil orelse + element(1, Val) =:= integer orelse + element(1, Val) =:= float orelse + element(1, Val) =:= atom orelse + element(1, Val) =:= literal)). + +bif_to_test('or', [V1,V2], {f,Lbl}=Fail, St0) when Lbl =/= 0 -> + {SuccLabel,St} = new_label(St0), + {[{test,is_eq_exact,{f,SuccLabel},[V1,{atom,false}]}, + {test,is_eq_exact,Fail,[V2,{atom,true}]}, + {label,SuccLabel}],St}; +bif_to_test(Op, Args, Fail, St) -> + {bif_to_test(Op, Args, Fail),St}. + +bif_to_test('and', [V1,V2], Fail) -> + [{test,is_eq_exact,Fail,[V1,{atom,true}]}, + {test,is_eq_exact,Fail,[V2,{atom,true}]}]; bif_to_test('not', [Var], Fail) -> [{test,is_eq_exact,Fail,[Var,{atom,false}]}]; bif_to_test(Name, Args, Fail) -> - [beam_utils:bif_to_test(Name, Args, Fail)]. + [bif_to_test_1(Name, Args, Fail)]. + +bif_to_test_1(is_atom, [_]=Ops, Fail) -> + {test,is_atom,Fail,Ops}; +bif_to_test_1(is_boolean, [_]=Ops, Fail) -> + {test,is_boolean,Fail,Ops}; +bif_to_test_1(is_binary, [_]=Ops, Fail) -> + {test,is_binary,Fail,Ops}; +bif_to_test_1(is_bitstring,[_]=Ops, Fail) -> + {test,is_bitstr,Fail,Ops}; +bif_to_test_1(is_float, [_]=Ops, Fail) -> + {test,is_float,Fail,Ops}; +bif_to_test_1(is_function, [_]=Ops, Fail) -> + {test,is_function,Fail,Ops}; +bif_to_test_1(is_function, [_,_]=Ops, Fail) -> + {test,is_function2,Fail,Ops}; +bif_to_test_1(is_integer, [_]=Ops, Fail) -> + {test,is_integer,Fail,Ops}; +bif_to_test_1(is_list, [_]=Ops, Fail) -> + {test,is_list,Fail,Ops}; +bif_to_test_1(is_map, [_]=Ops, Fail) -> + {test,is_map,Fail,Ops}; +bif_to_test_1(is_number, [_]=Ops, Fail) -> + {test,is_number,Fail,Ops}; +bif_to_test_1(is_pid, [_]=Ops, Fail) -> + {test,is_pid,Fail,Ops}; +bif_to_test_1(is_port, [_]=Ops, Fail) -> + {test,is_port,Fail,Ops}; +bif_to_test_1(is_reference, [_]=Ops, Fail) -> + {test,is_reference,Fail,Ops}; +bif_to_test_1(is_tuple, [_]=Ops, Fail) -> + {test,is_tuple,Fail,Ops}; +bif_to_test_1('=<', [A,B], Fail) -> + {test,is_ge,Fail,[B,A]}; +bif_to_test_1('>', [A,B], Fail) -> + {test,is_lt,Fail,[B,A]}; +bif_to_test_1('<', [_,_]=Ops, Fail) -> + {test,is_lt,Fail,Ops}; +bif_to_test_1('>=', [_,_]=Ops, Fail) -> + {test,is_ge,Fail,Ops}; +bif_to_test_1('==', [C,A], Fail) when ?IS_LITERAL(C) -> + {test,is_eq,Fail,[A,C]}; +bif_to_test_1('==', [_,_]=Ops, Fail) -> + {test,is_eq,Fail,Ops}; +bif_to_test_1('/=', [C,A], Fail) when ?IS_LITERAL(C) -> + {test,is_ne,Fail,[A,C]}; +bif_to_test_1('/=', [_,_]=Ops, Fail) -> + {test,is_ne,Fail,Ops}; +bif_to_test_1('=:=', [C,A], Fail) when ?IS_LITERAL(C) -> + {test,is_eq_exact,Fail,[A,C]}; +bif_to_test_1('=:=', [_,_]=Ops, Fail) -> + {test,is_eq_exact,Fail,Ops}; +bif_to_test_1('=/=', [C,A], Fail) when ?IS_LITERAL(C) -> + {test,is_ne_exact,Fail,[A,C]}; +bif_to_test_1('=/=', [_,_]=Ops, Fail) -> + {test,is_ne_exact,Fail,Ops}. opt_call_moves(Is0, Arity) -> {Moves0,Is} = splitwith(fun({move,_,_}) -> true; + ({kill,_}) -> true; (_) -> false end, Is0), Moves = opt_call_moves_1(Moves0, Arity), Moves ++ Is. +opt_call_moves_1([{move,Src,{x,_}=Tmp}=M1|[{kill,_}|_]=Is], Arity) -> + %% There could be a {move,Tmp,{x,0}} instruction after the + %% kill/1 instructions (moved to there by opt_move_to_x0/1). + case splitwith(fun({kill,_}) -> true; + (_) -> false + end, Is) of + {Kills,[{move,{x,_}=Tmp,{x,0}}=M2]} -> + %% The two move/2 instructions (M1 and M2) can be combined + %% to one. The question is, though, is it safe to place + %% them after the kill/1 instructions? + case is_killed(Src, Kills, Arity) of + true -> + %% Src (a Y register) is killed by one of the + %% kill/1 instructions. Thus M1 and M2 + %% must be placed before the kill/1 instructions + %% (essentially undoing what opt_move_to_x0/1 + %% did, which turned out to be a pessimization + %% in this case). + opt_call_moves_1([M1,M2|Kills], Arity); + false -> + %% Src is not killed by any of the kill/1 + %% instructions. Thus it is safe to place + %% M1 and M2 after the kill/1 instructions. + opt_call_moves_1(Kills++[M1,M2], Arity) + end; + {_,_} -> + [M1|Is] + end; opt_call_moves_1([{move,Src,{x,_}=Tmp}=M1,{move,Tmp,Dst}=M2|Is], Arity) -> case is_killed(Tmp, Is, Arity) of true -> @@ -1203,6 +1389,10 @@ opt_call_moves_1([M|Ms], Arity) -> [M|opt_call_moves_1(Ms, Arity)]; opt_call_moves_1([], _Arity) -> []. +is_killed(Y, [{kill,Y}|_], _) -> + true; +is_killed(R, [{kill,_}|Is], Arity) -> + is_killed(R, Is, Arity); is_killed(R, [{move,R,_}|_], _) -> false; is_killed(R, [{move,_,R}|_], _) -> @@ -1210,7 +1400,9 @@ is_killed(R, [{move,_,R}|_], _) -> is_killed(R, [{move,_,_}|Is], Arity) -> is_killed(R, Is, Arity); is_killed({x,X}, [], Arity) -> - X >= Arity. + X >= Arity; +is_killed({y,_}, [], _) -> + false. cg_alloc(#cg_alloc{stack=none,words=#need{h=0,f=0}}, _St) -> []; @@ -1257,22 +1449,35 @@ cg_call(#cg_set{anno=Anno,op=call,dst=Dst0,args=[#b_local{}=Func0|Args0]}, Line = call_line(Where, local, Anno), Call = build_call(call, Arity, {f,FuncLbl}, Context, Dst), Is = setup_args(Args, Anno, Context, St) ++ Line ++ Call, - {Is,St}; -cg_call(#cg_set{anno=Anno,op=call,dst=Dst0,args=[#b_remote{}=Func0|Args0]}, + case Anno of + #{ result_type := Info } -> + {Is ++ [{'%', {type_info, Dst, Info}}], St}; + #{} -> + {Is, St} + end; +cg_call(#cg_set{anno=Anno0,op=call,dst=Dst0,args=[#b_remote{}=Func0|Args0]}, Where, Context, St) -> [Dst|Args] = beam_args([Dst0|Args0], St), #b_remote{mod=Mod0,name=Name0,arity=Arity} = Func0, case {beam_arg(Mod0, St),beam_arg(Name0, St)} of {{atom,Mod},{atom,Name}} -> Func = {extfunc,Mod,Name,Arity}, - Line = call_line(Where, Func, Anno), + Line = call_line(Where, Func, Anno0), Call = build_call(call_ext, Arity, Func, Context, Dst), + Anno = case erl_bifs:is_exit_bif(Mod, Name, Arity) of + true -> + %% There is no need to kill Y registers + %% before calling an exit BIF. + maps:remove(kill_yregs, Anno0); + false -> + Anno0 + end, Is = setup_args(Args, Anno, Context, St) ++ Line ++ Call, {Is,St}; {Mod,Name} -> Apply = build_apply(Arity, Context, Dst), - Is = setup_args(Args++[Mod,Name], Anno, Context, St) ++ - [line(Anno)] ++ Apply, + Is = setup_args(Args++[Mod,Name], Anno0, Context, St) ++ + [line(Anno0)] ++ Apply, {Is,St} end; cg_call(#cg_set{anno=Anno,op=call,dst=Dst0,args=Args0}, @@ -1325,6 +1530,12 @@ build_apply(Arity, none, Dst) -> cg_instr(put_map, [{atom,assoc},SrcMap|Ss], Dst, Set) -> Live = get_live(Set), [{put_map_assoc,{f,0},SrcMap,Dst,Live,{list,Ss}}]; +cg_instr(bs_get_tail, [Src], Dst, Set) -> + Live = get_live(Set), + [{bs_get_tail,Src,Dst,Live}]; +cg_instr(bs_get_position, [Ctx], Dst, Set) -> + Live = get_live(Set), + [{bs_get_position,Ctx,Dst,Live}]; cg_instr(Op, Args, Dst, _Set) -> cg_instr(Op, Args, Dst). @@ -1340,10 +1551,10 @@ cg_instr(bs_restore, [Ctx,Slot], _Dst) -> cg_instr(bs_save, [Ctx,Slot], _Dst) -> {integer,N} = Slot, [{bs_save2,Ctx,N}]; +cg_instr(bs_set_position, [Ctx,Pos], _Dst) -> + [{bs_set_position,Ctx,Pos}]; cg_instr(build_stacktrace, Args, Dst) -> setup_args(Args) ++ [build_stacktrace|copy({x,0}, Dst)]; -cg_instr(context_to_binary, [Src], _Dst) -> - [{bs_context_to_binary,Src}]; cg_instr(set_tuple_element=Op, [New,Tuple,{integer,Index}], _Dst) -> [{Op,New,Tuple,Index}]; cg_instr({float,clearerror}, [], _Dst) -> @@ -1360,6 +1571,8 @@ cg_instr(get_tuple_element=Op, [Src,{integer,N}], Dst) -> [{Op,Src,N,Dst}]; cg_instr(put_list=Op, [Hd,Tl], Dst) -> [{Op,Hd,Tl,Dst}]; +cg_instr(put_tuple, Elements, Dst) -> + [{put_tuple2,Dst,{list,Elements}}]; cg_instr(put_tuple_arity, [{integer,Arity}], Dst) -> [{put_tuple,Arity,Dst}]; cg_instr(put_tuple_elements, Elements, _Dst) -> @@ -1475,13 +1688,13 @@ cg_bs_put(Fail, [{atom,Type},{literal,Flags}|Args]) -> [{Op,Fail,{field_flags,Flags},Src}] end. -cg_bs_init(Dst, Size0, Unit, Live, Fail) -> +cg_bs_init(Dst, Size0, Alloc, Unit, Live, Fail) -> Op = case Unit of 1 -> bs_init_bits; 8 -> bs_init2 end, Size = cg_bs_init_size(Size0), - [{Op,Fail,Size,0,Live,{field_flags,[]},Dst}]. + [{Op,Fail,Size,Alloc,Live,{field_flags,[]},Dst}]. cg_bs_init_size({x,_}=R) -> R; cg_bs_init_size({y,_}=R) -> R; @@ -1517,6 +1730,14 @@ copy(Src, Dst) -> [{move,Src,Dst}]. force_reg({literal,_}=Lit, Reg) -> {Reg,[{move,Lit,Reg}]}; +force_reg({integer,_}=Lit, Reg) -> + {Reg,[{move,Lit,Reg}]}; +force_reg({atom,_}=Lit, Reg) -> + {Reg,[{move,Lit,Reg}]}; +force_reg({float,_}=Lit, Reg) -> + {Reg,[{move,Lit,Reg}]}; +force_reg(nil=Lit, Reg) -> + {Reg,[{move,Lit,Reg}]}; force_reg({Kind,_}=R, _) when Kind =:= x; Kind =:= y -> {R,[]}. @@ -1600,12 +1821,41 @@ phi_copies([#b_set{dst=Dst,args=PhiArgs}|Sets], L) -> [#cg_set{op=copy,dst=Dst,args=CopyArgs}|phi_copies(Sets, L)]; phi_copies([], _) -> []. +%% opt_move_to_x0([Instruction]) -> [Instruction]. +%% Simple peep-hole optimization to move a {move,Any,{x,0}} past +%% any kill up to the next call instruction. (To give the loader +%% an opportunity to combine the 'move' and the 'call' instructions.) + +opt_move_to_x0(Moves) -> + opt_move_to_x0(Moves, []). + +opt_move_to_x0([{move,_,{x,0}}=I|Is0], Acc0) -> + case move_past_kill(Is0, I, Acc0) of + impossible -> opt_move_to_x0(Is0, [I|Acc0]); + {Is,Acc} -> opt_move_to_x0(Is, Acc) + end; +opt_move_to_x0([I|Is], Acc) -> + opt_move_to_x0(Is, [I|Acc]); +opt_move_to_x0([], Acc) -> reverse(Acc). + +move_past_kill([{kill,Src}|_], {move,Src,_}, _) -> + impossible; +move_past_kill([{kill,_}=I|Is], Move, Acc) -> + move_past_kill(Is, Move, [I|Acc]); +move_past_kill(Is, Move, Acc) -> + {Is,[Move|Acc]}. + %% setup_args(Args, Anno, Context) -> [Instruction]. %% setup_args(Args) -> [Instruction]. %% Set up X registers for a call. setup_args(Args, Anno, none, St) -> - setup_args(Args) ++ kill_yregs(Anno, St); + case {setup_args(Args),kill_yregs(Anno, St)} of + {Moves,[]} -> + Moves; + {Moves,Kills} -> + opt_move_to_x0(Moves ++ Kills) + end; setup_args(Args, _, _, _) -> setup_args(Args). @@ -1694,7 +1944,7 @@ get_register(V, Regs) -> beam_args(As, St) -> [beam_arg(A, St) || A <- As]. -beam_arg(#b_var{name=Name}, #cg{regs=Regs}) -> +beam_arg(#b_var{}=Name, #cg{regs=Regs}) -> maps:get(Name, Regs); beam_arg(#b_literal{val=Val}, _) -> if |