diff options
Diffstat (limited to 'lib/compiler/src')
43 files changed, 7342 insertions, 4484 deletions
diff --git a/lib/compiler/src/Makefile b/lib/compiler/src/Makefile index bd35f20442..97c73d0e07 100644 --- a/lib/compiler/src/Makefile +++ b/lib/compiler/src/Makefile @@ -49,10 +49,7 @@ MODULES = \ beam_a \ beam_asm \ beam_block \ - beam_bs \ - beam_bsm \ beam_clean \ - beam_dead \ beam_dict \ beam_disasm \ beam_except \ @@ -61,14 +58,17 @@ MODULES = \ beam_listing \ beam_opcodes \ beam_peep \ - beam_split \ beam_ssa \ + beam_ssa_bsm \ beam_ssa_codegen \ + beam_ssa_dead \ + beam_ssa_funs \ beam_ssa_lint \ beam_ssa_opt \ beam_ssa_pp \ beam_ssa_pre_codegen \ beam_ssa_recv \ + beam_ssa_share \ beam_ssa_type \ beam_kernel_to_ssa \ beam_trim \ @@ -103,6 +103,7 @@ BEAM_H = $(wildcard ../priv/beam_h/*.h) HRL_FILES= \ beam_disasm.hrl \ + beam_ssa_opt.hrl \ beam_ssa.hrl \ core_parse.hrl \ v3_kernel.hrl @@ -194,6 +195,7 @@ $(EBIN)/beam_listing.beam: core_parse.hrl v3_kernel.hrl beam_ssa.hrl $(EBIN)/beam_kernel_to_ssa.beam: v3_kernel.hrl beam_ssa.hrl $(EBIN)/beam_ssa.beam: beam_ssa.hrl $(EBIN)/beam_ssa_codegen.beam: beam_ssa.hrl +$(EBIN)/beam_ssa_dead.beam: beam_ssa.hrl $(EBIN)/beam_ssa_lint.beam: beam_ssa.hrl $(EBIN)/beam_ssa_opt.beam: beam_ssa.hrl $(EBIN)/beam_ssa_pp.beam: beam_ssa.hrl diff --git a/lib/compiler/src/beam_a.erl b/lib/compiler/src/beam_a.erl index 266e8f46c8..1ac892a8f1 100644 --- a/lib/compiler/src/beam_a.erl +++ b/lib/compiler/src/beam_a.erl @@ -52,6 +52,16 @@ function({function,Name,Arity,CLabel,Is0}) -> erlang:raise(Class, Error, Stack) end. +rename_instrs([{test,is_eq_exact,_,[Dst,Src]}=Test, + {move,Src,Dst}|Is]) -> + %% The move instruction is not needed. + rename_instrs([Test|Is]); +rename_instrs([{test,is_eq_exact,_,[Same,Same]}|Is]) -> + %% Same literal or same register. Will always succeed. + rename_instrs(Is); +rename_instrs([{loop_rec,{f,Fail},{x,0}},{loop_rec_end,_},{label,Fail}|Is]) -> + %% This instruction sequence does nothing. + rename_instrs(Is); rename_instrs([{apply_last,A,N}|Is]) -> [{apply,A},{deallocate,N},return|rename_instrs(Is)]; rename_instrs([{call_last,A,F,N}|Is]) -> @@ -90,8 +100,12 @@ rename_instr({bs_put_utf16=I,F,Fl,Src}) -> {bs_put,F,{I,Fl},[Src]}; rename_instr({bs_put_utf32=I,F,Fl,Src}) -> {bs_put,F,{I,Fl},[Src]}; -rename_instr({bs_put_string,_,_}=I) -> - {bs_put,{f,0},I,[]}; +rename_instr({bs_put_string,_,{string,String}}) -> + %% Only happens when compiling from .S files. In old + %% .S files, String is a list. In .S in OTP 22 and later, + %% String is a binary. + {bs_put,{f,0},{bs_put_binary,8,{field_flags,[unsigned,big]}}, + [{atom,all},{literal,iolist_to_binary([String])}]}; rename_instr({bs_add=I,F,[Src1,Src2,U],Dst}) when is_integer(U) -> {bif,I,F,[Src1,Src2,{integer,U}],Dst}; rename_instr({bs_utf8_size=I,F,Src,Dst}) -> @@ -108,8 +122,8 @@ rename_instr({bs_private_append=I,F,Sz,U,Src,Flags,Dst}) -> {bs_init,F,{I,U,Flags},none,[Sz,Src],Dst}; rename_instr(bs_init_writable=I) -> {bs_init,{f,0},I,1,[{x,0}],{x,0}}; -rename_instr({test,Op,F,[Ctx,Bits,{string,Str}]}) -> - %% When compiling from a .S file. +rename_instr({test,bs_match_string=Op,F,[Ctx,Bits,{string,Str}]}) when is_list(Str) -> + %% When compiling from an old .S file. Starting from OTP 22, Str is a binary. <<Bs:Bits/bits,_/bits>> = list_to_binary(Str), {test,Op,F,[Ctx,Bs]}; rename_instr({put_map_assoc,Fail,S,D,R,L}) -> diff --git a/lib/compiler/src/beam_asm.erl b/lib/compiler/src/beam_asm.erl index df0321e85a..bc1290f6fd 100644 --- a/lib/compiler/src/beam_asm.erl +++ b/lib/compiler/src/beam_asm.erl @@ -424,8 +424,8 @@ encode_arg({f, W}, Dict) -> {encode(?tag_f, W), Dict}; %% encode_arg({'char', C}, Dict) -> %% {encode(?tag_h, C), Dict}; -encode_arg({string, String}, Dict0) -> - {Offset, Dict} = beam_dict:string(String, Dict0), +encode_arg({string, BinString}, Dict0) when is_binary(BinString) -> + {Offset, Dict} = beam_dict:string(BinString, Dict0), {encode(?tag_u, Offset), Dict}; encode_arg({extfunc, M, F, A}, Dict0) -> {Index, Dict} = beam_dict:import(M, F, A, Dict0), diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl index 0ed2a03a81..707974b2c1 100644 --- a/lib/compiler/src/beam_block.erl +++ b/lib/compiler/src/beam_block.erl @@ -22,7 +22,7 @@ -module(beam_block). -export([module/2]). --import(lists, [reverse/1,splitwith/2]). +-import(lists, [keysort/2,reverse/1,splitwith/2]). -spec module(beam_utils:module_code(), [compile:option()]) -> {'ok',beam_utils:module_code()}. @@ -49,14 +49,12 @@ function({function,Name,Arity,CLabel,Is0}) -> blockify(Is) -> blockify(Is, []). -blockify([{loop_rec,{f,Fail},{x,0}},{loop_rec_end,_Lbl},{label,Fail}|Is], Acc) -> - %% Useless instruction sequence. - blockify(Is, Acc); blockify([I|Is0]=IsAll, Acc) -> case collect(I) of error -> blockify(Is0, [I|Acc]); Instr when is_tuple(Instr) -> - {Block,Is} = collect_block(IsAll), + {Block0,Is} = collect_block(IsAll), + Block = sort_moves(Block0), blockify(Is, [{block,Block}|Acc]) end; blockify([], Acc) -> reverse(Acc). @@ -83,8 +81,8 @@ collect({allocate_heap,Ns,Nh,R}) -> {set,[],[],{alloc,R,{nozero,Ns,Nh,[]}}}; collect({allocate_heap_zero,Ns,Nh,R}) -> {set,[],[],{alloc,R,{zero,Ns,Nh,[]}}}; collect({init,D}) -> {set,[D],[],init}; collect({test_heap,N,R}) -> {set,[],[],{alloc,R,{nozero,nostack,N,[]}}}; -collect({bif,N,F,As,D}) -> {set,[D],As,{bif,N,F}}; -collect({gc_bif,N,F,R,As,D}) -> {set,[D],As,{alloc,R,{gc_bif,N,F}}}; +collect({bif,N,{f,0},As,D}) -> {set,[D],As,{bif,N,{f,0}}}; +collect({gc_bif,N,{f,0},R,As,D}) -> {set,[D],As,{alloc,R,{gc_bif,N,{f,0}}}}; collect({move,S,D}) -> {set,[D],[S],move}; collect({put_list,S1,S2,D}) -> {set,[D],[S1,S2],put_list}; collect({put_tuple,A,D}) -> {set,[D],[],{put_tuple,A}}; @@ -95,12 +93,8 @@ collect({set_tuple_element,S,D,I}) -> {set,[],[S,D],{set_tuple_element,I}}; collect({get_hd,S,D}) -> {set,[D],[S],get_hd}; collect({get_tl,S,D}) -> {set,[D],[S],get_tl}; collect(remove_message) -> {set,[],[],remove_message}; -collect({put_map,F,Op,S,D,R,{list,Puts}}) -> - {set,[D],[S|Puts],{alloc,R,{put_map,Op,F}}}; -collect({'catch'=Op,R,L}) -> - {set,[R],[],{try_catch,Op,L}}; -collect({'try'=Op,R,L}) -> - {set,[R],[],{try_catch,Op,L}}; +collect({put_map,{f,0},Op,S,D,R,{list,Puts}}) -> + {set,[D],[S|Puts],{alloc,R,{put_map,Op,{f,0}}}}; collect(fclearerror) -> {set,[],[],fclearerror}; collect({fcheckerror,{f,0}}) -> {set,[],[],fcheckerror}; collect({fmove,S,D}) -> {set,[D],[S],fmove}; @@ -124,3 +118,36 @@ embed_lines([{block,B1},{line,_}=Line|T], Acc) -> embed_lines([I|Is], Acc) -> embed_lines(Is, [I|Acc]); embed_lines([], Acc) -> Acc. + +%% sort_moves([Instruction]) -> [Instruction]. +%% Sort move instructions on the Y register to give the loader +%% more opportunities for combining instructions. + +sort_moves([{set,[{x,_}],[{y,_}],move}=I|Is0]) -> + {Moves,Is} = sort_moves_1(Is0, x, y, [I]), + Moves ++ sort_moves(Is); +sort_moves([{set,[{y,_}],[{x,_}],move}=I|Is0]) -> + {Moves,Is} = sort_moves_1(Is0, y, x, [I]), + Moves ++ sort_moves(Is); +sort_moves([I|Is]) -> + [I|sort_moves(Is)]; +sort_moves([]) -> []. + +sort_moves_1([{set,[{x,0}],[_],move}=I|Is], _DTag, _STag, Acc) -> + %% The loader sometimes combines a move to x0 with the + %% instruction that follows, producing, for example, a move_call + %% instruction. Therefore, we don't want include this move + %% instruction in the sorting. + {sort_on_yreg(Acc)++[I],Is}; +sort_moves_1([{set,[{DTag,_}],[{STag,_}],move}=I|Is], DTag, STag, Acc) -> + sort_moves_1(Is, DTag, STag, [I|Acc]); +sort_moves_1(Is, _DTag, _STag, Acc) -> + {sort_on_yreg(Acc),Is}. + +sort_on_yreg([{set,[Dst],[Src],move}|_]=Moves) -> + case {Dst,Src} of + {{y,_},{x,_}} -> + keysort(2, Moves); + {{x,_},{y,_}} -> + keysort(3, Moves) + end. diff --git a/lib/compiler/src/beam_bs.erl b/lib/compiler/src/beam_bs.erl deleted file mode 100644 index 15d8d687fc..0000000000 --- a/lib/compiler/src/beam_bs.erl +++ /dev/null @@ -1,183 +0,0 @@ -%% -%% %CopyrightBegin% -%% -%% Copyright Ericsson AB 1999-2018. All Rights Reserved. -%% -%% Licensed under the Apache License, Version 2.0 (the "License"); -%% you may not use this file except in compliance with the License. -%% You may obtain a copy of the License at -%% -%% http://www.apache.org/licenses/LICENSE-2.0 -%% -%% Unless required by applicable law or agreed to in writing, software -%% distributed under the License is distributed on an "AS IS" BASIS, -%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -%% See the License for the specific language governing permissions and -%% limitations under the License. -%% -%% %CopyrightEnd% -%% -%% Purpose: Peephole optimization of binary syntax instructions. - --module(beam_bs). - --export([module/2]). --import(lists, [reverse/1]). - --spec module(beam_utils:module_code(), [compile:option()]) -> - {'ok',beam_utils:module_code()}. - -module({Mod,Exp,Attr,Fs0,Lc}, _Opt) -> - Fs = [function(F) || F <- Fs0], - {ok,{Mod,Exp,Attr,Fs,Lc}}. - -function({function,Name,Arity,CLabel,Is0}) -> - try - Is = bs_opt(Is0), - {function,Name,Arity,CLabel,Is} - catch - Class:Error:Stack -> - io:fwrite("Function: ~w/~w\n", [Name,Arity]), - erlang:raise(Class, Error, Stack) - end. - -%%% -%%% Evaluate construction of constant bit fields. -%%% Combine bs_skip_bits2 and bs_test_tail2 instructions. -%%% - -bs_opt([{bs_put,_,_,_}=I|Is0]) -> - {BsPuts0,Is} = collect_bs_puts(Is0, [I]), - BsPuts = opt_bs_puts(BsPuts0), - BsPuts ++ bs_opt(Is); -bs_opt([{test,bs_skip_bits2,F,[Ctx,{integer,I},Unit,_Flags]}, - {test,bs_test_tail2,F,[Ctx,Bits]}|Is]) -> - [{test,bs_test_tail2,F,[Ctx,Bits+I*Unit]}|bs_opt(Is)]; -bs_opt([{test,bs_skip_bits2,F,[Ctx,{integer,I1},Unit1,Flags]}, - {test,bs_skip_bits2,F,[Ctx,{integer,I2},Unit2,_]}|Is]) -> - I = {test,bs_skip_bits2,F, - [Ctx,{integer,I1*Unit1+I2*Unit2},1,Flags]}, - bs_opt([I|Is]); -bs_opt([I|Is]) -> - [I|bs_opt(Is)]; -bs_opt([]) -> []. - -collect_bs_puts([{bs_put,_,_,_}=I|Is], Acc) -> - collect_bs_puts(Is, [I|Acc]); -collect_bs_puts([_|_]=Is, Acc) -> - {reverse(Acc),Is}. - -opt_bs_puts(Is) -> - opt_bs_1(Is, []). - -opt_bs_1([{bs_put,Fail, - {bs_put_float,1,Flags0},[{integer,Sz},Src]}=I0|Is], Acc) -> - try eval_put_float(Src, Sz, Flags0) of - <<Int:Sz>> -> - Flags = force_big(Flags0), - I = {bs_put,Fail,{bs_put_integer,1,Flags}, - [{integer,Sz},{integer,Int}]}, - opt_bs_1([I|Is], Acc) - catch - error:_ -> - opt_bs_1(Is, [I0|Acc]) - end; -opt_bs_1([{bs_put,_,{bs_put_integer,1,_},[{integer,8},{integer,_}]}|_]=IsAll, - Acc0) -> - {Is,Acc} = bs_collect_string(IsAll, Acc0), - opt_bs_1(Is, Acc); -opt_bs_1([{bs_put,Fail,{bs_put_integer,1,F},[{integer,Sz},{integer,N}]}=I|Is0], - Acc) when Sz > 8 -> - case field_endian(F) of - big -> - %% We can do this optimization for any field size without - %% risk for code explosion. - case bs_split_int(N, Sz, Fail, Is0) of - no_split -> opt_bs_1(Is0, [I|Acc]); - Is -> opt_bs_1(Is, Acc) - end; - little when Sz < 128 -> - %% We only try to optimize relatively small fields, to - %% avoid an explosion in code size. - <<Int:Sz>> = <<N:Sz/little>>, - Flags = force_big(F), - Is = [{bs_put,Fail,{bs_put_integer,1,Flags}, - [{integer,Sz},{integer,Int}]}|Is0], - opt_bs_1(Is, Acc); - _ -> %native or too wide little field - opt_bs_1(Is0, [I|Acc]) - end; -opt_bs_1([{bs_put,Fail,{Op,U,F},[{integer,Sz},Src]}|Is], Acc) when U > 1 -> - opt_bs_1([{bs_put,Fail,{Op,1,F},[{integer,U*Sz},Src]}|Is], Acc); -opt_bs_1([I|Is], Acc) -> - opt_bs_1(Is, [I|Acc]); -opt_bs_1([], Acc) -> reverse(Acc). - -eval_put_float(Src, Sz, Flags) when Sz =< 256 -> - %%Only evaluate if Sz is reasonable. - Val = value(Src), - case field_endian(Flags) of - little -> <<Val:Sz/little-float-unit:1>>; - big -> <<Val:Sz/big-float-unit:1>> - %% native intentionally not handled here - we can't optimize - %% it. - end. - -value({integer,I}) -> I; -value({float,F}) -> F. - -bs_collect_string(Is, [{bs_put,_,{bs_put_string,Len,{string,Str}},[]}|Acc]) -> - bs_coll_str_1(Is, Len, reverse(Str), Acc); -bs_collect_string(Is, Acc) -> - bs_coll_str_1(Is, 0, [], Acc). - -bs_coll_str_1([{bs_put,_,{bs_put_integer,U,_},[{integer,Sz},{integer,V}]}|Is], - Len, StrAcc, IsAcc) when U*Sz =:= 8 -> - Byte = V band 16#FF, - bs_coll_str_1(Is, Len+1, [Byte|StrAcc], IsAcc); -bs_coll_str_1(Is, Len, StrAcc, IsAcc) -> - {Is,[{bs_put,{f,0},{bs_put_string,Len,{string,reverse(StrAcc)}},[]}|IsAcc]}. - -field_endian({field_flags,F}) -> field_endian_1(F). - -field_endian_1([big=E|_]) -> E; -field_endian_1([little=E|_]) -> E; -field_endian_1([native=E|_]) -> E; -field_endian_1([_|Fs]) -> field_endian_1(Fs). - -force_big({field_flags,F}) -> - {field_flags,force_big_1(F)}. - -force_big_1([big|_]=Fs) -> Fs; -force_big_1([little|Fs]) -> [big|Fs]; -force_big_1([F|Fs]) -> [F|force_big_1(Fs)]. - -bs_split_int(0, Sz, _, _) when Sz > 64 -> - %% We don't want to split in this case because the - %% string will consist of only zeroes. - no_split; -bs_split_int(-1, Sz, _, _) when Sz > 64 -> - %% We don't want to split in this case because the - %% string will consist of only 255 bytes. - no_split; -bs_split_int(N, Sz, Fail, Acc) -> - FirstByteSz = case Sz rem 8 of - 0 -> 8; - Rem -> Rem - end, - bs_split_int_1(N, FirstByteSz, Sz, Fail, Acc). - -bs_split_int_1(-1, _, Sz, Fail, Acc) when Sz > 64 -> - I = {bs_put,Fail,{bs_put_integer,1,{field_flags,[big]}}, - [{integer,Sz},{integer,-1}]}, - [I|Acc]; -bs_split_int_1(0, _, Sz, Fail, Acc) when Sz > 64 -> - I = {bs_put,Fail,{bs_put_integer,1,{field_flags,[big]}}, - [{integer,Sz},{integer,0}]}, - [I|Acc]; -bs_split_int_1(N, ByteSz, Sz, Fail, Acc) when Sz > 0 -> - Mask = (1 bsl ByteSz) - 1, - I = {bs_put,Fail,{bs_put_integer,1,{field_flags,[big]}}, - [{integer,ByteSz},{integer,N band Mask}]}, - bs_split_int_1(N bsr ByteSz, 8, Sz-ByteSz, Fail, [I|Acc]); -bs_split_int_1(_, _, _, _, Acc) -> Acc. diff --git a/lib/compiler/src/beam_bsm.erl b/lib/compiler/src/beam_bsm.erl deleted file mode 100644 index abc6e96c85..0000000000 --- a/lib/compiler/src/beam_bsm.erl +++ /dev/null @@ -1,719 +0,0 @@ -%% -%% %CopyrightBegin% -%% -%% Copyright Ericsson AB 2007-2018. All Rights Reserved. -%% -%% Licensed under the Apache License, Version 2.0 (the "License"); -%% you may not use this file except in compliance with the License. -%% You may obtain a copy of the License at -%% -%% http://www.apache.org/licenses/LICENSE-2.0 -%% -%% Unless required by applicable law or agreed to in writing, software -%% distributed under the License is distributed on an "AS IS" BASIS, -%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -%% See the License for the specific language governing permissions and -%% limitations under the License. -%% -%% %CopyrightEnd% -%% - --module(beam_bsm). --export([module/2,format_error/1]). - --import(lists, [member/2,foldl/3,reverse/1,sort/1,all/2]). - -%%% -%%% We optimize bit syntax matching where the tail end of a binary is -%%% matched out and immediately passed on to a bs_start_match2 instruction, -%%% such as in this code sequence: -%%% -%%% func_info ... -%%% L1 test bs_start_match2 {f,...} {x,0} Live SavePositions {x,0} -%%% . . . -%%% test bs_get_binary2 {f,...} {x,0} all 1 Flags {x,0} -%%% . . . -%%% call_only 2 L1 -%%% -%%% The sequence can be optimized simply by removing the bs_get_binary2 -%%% instruction. Another example: -%%% -%%% func_info ... -%%% L1 test bs_start_match2 {f,...} {x,0} Live SavePositions {x,0} -%%% . . . -%%% test bs_get_binary2 {f,...} {x,0} all 8 Flags {x,1} -%%% . . . -%%% move {x,1} {x,0} -%%% call_only 2 L1 -%%% -%%% In this case, the bs_get_binary2 instruction must be replaced by -%%% -%%% test bs_unit {x,1} 8 -%%% -%%% to ensure that the match fail if the length of the binary in bits -%%% is not evenly divisible by 8. -%%% -%%% Note that the bs_start_match2 instruction doesn't need to be in the same -%%% function as the caller. It can be in the beginning of any function, or -%%% follow the bs_get_binary2 instruction in the same function. The important -%%% thing is that the match context register is not copied or built into -%%% data structures or passed to BIFs. -%%% - --type label() :: beam_asm:label(). --type func_info() :: {beam_asm:reg(),boolean()}. - --record(btb, - {f :: gb_trees:tree(label(), func_info()), - index :: beam_utils:code_index(), %{Label,Code} index (for liveness). - ok_br=gb_sets:empty() :: gb_sets:set(label()), %Labels that are OK. - must_not_save=false :: boolean(), %Must not save position when - % optimizing (reaches - % bs_context_to_binary). - must_save=false :: boolean() %Must save position when optimizing. - }). - - --spec module(beam_utils:module_code(), [compile:option()]) -> - {'ok',beam_utils:module_code()}. - -module({Mod,Exp,Attr,Fs0,Lc}, Opts) -> - FIndex = btb_index(Fs0), - Fs = [function(F, FIndex) || F <- Fs0], - Code = {Mod,Exp,Attr,Fs,Lc}, - case proplists:get_bool(bin_opt_info, Opts) of - true -> - {ok,Code,collect_warnings(Fs)}; - false -> - {ok,Code} - end. - --spec format_error('bin_opt' | {'no_bin_opt', term()}) -> nonempty_string(). - -format_error(bin_opt) -> - "OPTIMIZED: creation of sub binary delayed"; -format_error({no_bin_opt,Reason}) -> - lists:flatten(["NOT OPTIMIZED: "|format_error_1(Reason)]). - -%%% -%%% Local functions. -%%% - -function({function,Name,Arity,Entry,Is}, FIndex) -> - try - Index = beam_utils:index_labels(Is), - D = #btb{f=FIndex,index=Index}, - {function,Name,Arity,Entry,btb_opt_1(Is, D, [])} - catch - Class:Error:Stack -> - io:fwrite("Function: ~w/~w\n", [Name,Arity]), - erlang:raise(Class, Error, Stack) - end. - -btb_opt_1([{test,bs_get_binary2,F,_,[Reg,{atom,all},U,Fs],Reg}=I0|Is], D, Acc0) -> - case btb_reaches_match(Is, [Reg], D) of - {error,Reason} -> - Comment = btb_comment_no_opt(Reason, Fs), - btb_opt_1(Is, D, [Comment,I0|Acc0]); - {ok,MustSave} -> - Comment = btb_comment_opt(Fs), - Acc1 = btb_gen_save(MustSave, Reg, [Comment|Acc0]), - Acc = case U of - 1 -> Acc1; - _ -> [{test,bs_test_unit,F,[Reg,U]}|Acc1] - end, - btb_opt_1(Is, D, Acc) - end; -btb_opt_1([{test,bs_get_binary2,F,_,[Ctx,{atom,all},U,Fs],Dst}=I0|Is0], D, Acc0) -> - case btb_reaches_match(Is0, [Ctx,Dst], D) of - {error,Reason} -> - Comment = btb_comment_no_opt(Reason, Fs), - btb_opt_1(Is0, D, [Comment,I0|Acc0]); - {ok,MustSave} when U =:= 1 -> - Comment = btb_comment_opt(Fs), - Acc = btb_gen_save(MustSave, Ctx, [Comment|Acc0]), - Is = prepend_move(Ctx, Dst, Is0), - btb_opt_1(Is, D, Acc); - {ok,MustSave} -> - Comment = btb_comment_opt(Fs), - Acc1 = btb_gen_save(MustSave, Ctx, [Comment|Acc0]), - Acc = [{test,bs_test_unit,F,[Ctx,U]}|Acc1], - Is = prepend_move(Ctx, Dst, Is0), - btb_opt_1(Is, D, Acc) - end; -btb_opt_1([I|Is], D, Acc) -> - %%io:format("~p\n", [I]), - btb_opt_1(Is, D, [I|Acc]); -btb_opt_1([], _, Acc) -> - reverse(Acc). - -btb_gen_save(true, Reg, Acc) -> - [{bs_save2,Reg,{atom,start}}|Acc]; -btb_gen_save(false, _, Acc) -> Acc. - -prepend_move(Ctx, Dst, [{block,Bl0}|Is]) -> - Bl = [{set,[Dst],[Ctx],move}|Bl0], - [{block,Bl}|Is]; -prepend_move(Ctx, Dst, Is) -> - [{move,Ctx,Dst}|Is]. - -%% btb_reaches_match([Instruction], [Register], D) -> -%% {ok,MustSave}|{error,Reason} -%% -%% The list of Registers should be a list of registers referencing a -%% match context. The Register may contain one element if the -%% bs_get_binary2 instruction looks like -%% -%% test bs_get_binary2 {f,...} Ctx all _ _ Ctx -%% -%% or two elements if the instruction looks like -%% -%% test bs_get_binary2 {f,...} Ctx all _ _ Dst -%% -%% This function determines whether the bs_get_binary2 instruction -%% can be omitted (retaining the match context instead of creating -%% a sub binary). -%% -%% The rule is that the match context ultimately must end up at a -%% bs_start_match2 instruction and nowhere else. That it, it must not -%% be passed to BIFs, or copied or put into data structures. There -%% must only be one copy alive when the match context reaches the -%% bs_start_match2 instruction. -%% -%% At a branch, we must follow all branches and make sure that the above -%% rule is followed (or that the branch kills the match context). -%% -%% The MustSave return value will be true if control may end up at -%% bs_context_to_binary instruction. Since that instruction uses the -%% saved start position, we must use "bs_save2 Ctx start" to -%% update the saved start position. An additional complication is that -%% "bs_save2 Ctx start" must not be used if Dst and Ctx are -%% different registers and both registers may be passed to -%% a bs_context_to_binary instruction. -%% - -btb_reaches_match(Is, RegList, D) -> - try - Regs = btb_regs_from_list(RegList), - #btb{must_not_save=MustNotSave,must_save=MustSave} = - btb_reaches_match_1(Is, Regs, D), - case MustNotSave andalso MustSave of - true -> btb_error(must_and_must_not_save); - false -> {ok,MustSave} - end - catch - throw:{error,_}=Error -> Error - end. - -btb_reaches_match_1(Is, Regs, D) -> - case btb_are_registers_empty(Regs) of - false -> - btb_reaches_match_2(Is, Regs, D); - true -> - %% The context was killed, which is OK. - D - end. - -btb_reaches_match_2([{block,Bl}|Is], Regs0, D) -> - Regs = btb_reaches_match_block(Bl, Regs0), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{call,Arity,{f,Lbl}}|Is], Regs0, D) -> - case is_tail_call(Is) of - true -> - Regs1 = btb_kill_not_live(Arity, Regs0), - Regs = btb_kill_yregs(Regs1), - btb_tail_call(Lbl, Regs, D); - false -> - btb_call(Arity, Lbl, Regs0, Is, D) - end; -btb_reaches_match_2([{apply,Arity}|Is], Regs, D) -> - btb_call(Arity+2, apply, Regs, Is, D); -btb_reaches_match_2([{call_fun,Live}=I|Is], Regs, D) -> - btb_ensure_not_used([{x,Live}], I, Regs), - btb_call(Live, I, Regs, Is, D); -btb_reaches_match_2([{make_fun2,_,_,_,Live}|Is], Regs, D) -> - btb_call(Live, make_fun2, Regs, Is, D); -btb_reaches_match_2([{call_ext,Arity,Func}=I|Is], Regs0, D) -> - %% Allow us scanning beyond the call in case the match - %% context is saved on the stack. - case beam_jump:is_exit_instruction(I) of - false -> - btb_call(Arity, Func, Regs0, Is, D); - true -> - Regs = btb_kill_not_live(Arity, Regs0), - btb_tail_call(Func, Regs, D) - end; -btb_reaches_match_2([{kill,Y}|Is], Regs, D) -> - btb_reaches_match_1(Is, btb_kill([Y], Regs), D); -btb_reaches_match_2([{deallocate,_}|Is], Regs0, D) -> - Regs = btb_kill_yregs(Regs0), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([return=I|_], Regs0, D) -> - btb_ensure_not_used([{x,0}], I, Regs0), - D; -btb_reaches_match_2([{gc_bif,_,{f,F},Live,Ss,Dst}=I|Is], Regs0, D0) -> - btb_ensure_not_used(Ss, I, Regs0), - Regs1 = btb_kill_not_live(Live, Regs0), - Regs = btb_kill([Dst], Regs1), - D = btb_follow_branch(F, Regs, D0), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{bif,_,{f,F},Ss,Dst}=I|Is], Regs0, D0) -> - btb_ensure_not_used(Ss, I, Regs0), - Regs = btb_kill([Dst], Regs0), - D = btb_follow_branch(F, Regs, D0), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{get_map_elements,{f,F},Src,{list,Ls}}=I|Is], Regs0, D0) -> - {Ss,Ds} = beam_utils:split_even(Ls), - btb_ensure_not_used([Src|Ss], I, Regs0), - Regs = btb_kill(Ds, Regs0), - D = btb_follow_branch(F, Regs, D0), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{test,bs_start_match2,{f,F},Live,[Ctx,_],Ctx}=I|Is], - Regs0, D0) -> - CtxRegs = btb_context_regs(Regs0), - case member(Ctx, CtxRegs) of - false -> - %% This bs_start_match2 instruction does not use "our" - %% match state. Therefore we can continue the search - %% for another bs_start_match2 instruction. - D = btb_follow_branch(F, Regs0, D0), - Regs = btb_kill_not_live(Live, Regs0), - btb_reaches_match_2(Is, Regs, D); - true -> - %% OK. This instruction will use "our" match state, - %% but we must make sure that all other copies of the - %% match state are killed in the code that follows - %% the instruction. (We know that the fail branch cannot - %% be taken in this case.) - OtherCtxRegs = CtxRegs -- [Ctx], - case btb_are_all_unused(OtherCtxRegs, Is, D0) of - false -> btb_error({OtherCtxRegs,not_all_unused_after,I}); - true -> D0 - end - end; -btb_reaches_match_2([{test,bs_start_match2,{f,F},Live,[Bin,_],Ctx}|Is], - Regs0, D0) -> - CtxRegs = btb_context_regs(Regs0), - case member(Bin, CtxRegs) orelse member(Ctx, CtxRegs) of - false -> - %% This bs_start_match2 does not reference any copy of the - %% match state. Therefore it can safely be passed on the - %% way to another (perhaps more suitable) bs_start_match2 - %% instruction. - D = btb_follow_branch(F, Regs0, D0), - Regs = btb_kill_not_live(Live, Regs0), - btb_reaches_match_2(Is, Regs, D); - true -> - %% This variant of the bs_start_match2 instruction does - %% not accept a match state as source. - btb_error(unsuitable_bs_start_match) - end; -btb_reaches_match_2([{test,_,{f,F},Ss}=I|Is], Regs, D0) -> - btb_ensure_not_used(Ss, I, Regs), - D1 = btb_follow_branch(F, Regs, D0), - D = case Is of - [{bs_context_to_binary,_}|_] -> - %% bs_context_to_binary following a test instruction - %% probably needs the current position to be saved as - %% the new start position, but we can't be sure. - %% Therefore, conservatively disable the optimization - %% (instead of forcing a saving of the position). - D1#btb{must_save=true,must_not_save=true}; - _ -> - D1 - end, - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{test,_,{f,F},_,Ss,_}=I|Is], Regs, D0) -> - btb_ensure_not_used(Ss, I, Regs), - D = btb_follow_branch(F, Regs, D0), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{select,_,Src,{f,F},Conds}=I|Is], Regs, D0) -> - btb_ensure_not_used([Src], I, Regs), - D1 = btb_follow_branch(F, Regs, D0), - D = btb_follow_branches(Conds, Regs, D1), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{jump,{f,Lbl}}|_], Regs, #btb{index=Li}=D) -> - Is = fetch_code_at(Lbl, Li), - btb_reaches_match_2(Is, Regs, D); -btb_reaches_match_2([{label,_}|Is], Regs, D) -> - btb_reaches_match_2(Is, Regs, D); -btb_reaches_match_2([{bs_init,{f,0},_,_,Ss,Dst}=I|Is], Regs, D) -> - btb_ensure_not_used(Ss, I, Regs), - btb_reaches_match_1(Is, btb_kill([Dst], Regs), D); -btb_reaches_match_2([{bs_put,{f,0},_,Ss}=I|Is], Regs, D) -> - btb_ensure_not_used(Ss, I, Regs), - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([{bs_restore2,Src,_}=I|Is], Regs0, D) -> - case btb_contains_context(Src, Regs0) of - false -> - btb_reaches_match_1(Is, Regs0, D); - true -> - %% Check that all other copies of the context registers - %% are unused by the following instructions. - Regs = btb_kill([Src], Regs0), - CtxRegs = btb_context_regs(Regs), - case btb_are_all_unused(CtxRegs, Is, D) of - false -> btb_error({CtxRegs,not_all_unused_after,I}); - true -> D#btb{must_not_save=true} - end - end; -btb_reaches_match_2([{bs_context_to_binary,Src}=I|Is], Regs0, D) -> - case btb_contains_context(Src, Regs0) of - false -> - btb_reaches_match_1(Is, Regs0, D); - true -> - %% Check that all other copies of the context registers - %% are unused by the following instructions. - Regs = btb_kill([Src], Regs0), - CtxRegs = btb_context_regs(Regs), - case btb_are_all_unused(CtxRegs, Is, D) of - false -> btb_error({CtxRegs,not_all_unused_after,I}); - true -> D#btb{must_not_save=true} - end - end; -btb_reaches_match_2([{badmatch,Src}=I|_], Regs, D) -> - btb_ensure_not_used([Src], I, Regs), - D; -btb_reaches_match_2([{case_end,Src}=I|_], Regs, D) -> - btb_ensure_not_used([Src], I, Regs), - D; -btb_reaches_match_2([if_end|_], _Regs, D) -> - D; -btb_reaches_match_2([{func_info,_,_,Arity}=I|_], Regs0, D) -> - Regs = btb_kill_yregs(btb_kill_not_live(Arity, Regs0)), - case btb_context_regs(Regs) of - [] -> D; - _ -> {binary_used_in,I} - end; -btb_reaches_match_2([{line,_}|Is], Regs, D) -> - btb_reaches_match_1(Is, Regs, D); -btb_reaches_match_2([I|_], Regs, _) -> - btb_error({btb_context_regs(Regs),I,not_handled}). - -is_tail_call([{deallocate,_}|_]) -> true; -is_tail_call([return|_]) -> true; -is_tail_call(_) -> false. - -btb_call(Arity, Lbl, Regs0, Is, D0) -> - Regs = btb_kill_not_live(Arity, Regs0), - case btb_are_x_registers_empty(Regs) of - false -> - %% There is a match context in one of the x registers. - %% First handle the call as if it were a tail call. - D = btb_tail_call(Lbl, Regs, D0), - - %% No problem so far (the called function can handle a - %% match context). Now we must make sure that we don't - %% have any copies of the match context tucked away in an - %% y register. - RegList = btb_context_regs(Regs), - case [R || {y,_}=R <- RegList] of - [] -> - D; - [_|_] -> - btb_error({multiple_uses,RegList}) - end; - true -> - %% No match context in any x register. It could have been - %% saved to an y register, so continue to scan the code following - %% the call. - btb_reaches_match_1(Is, Regs, D0) - end. - -btb_tail_call(Lbl, Regs, #btb{f=Ftree,must_save=MustSave0}=D) -> - %% Ignore any y registers here. - case [R || {x,_}=R <- btb_context_regs(Regs)] of - [] -> - D; - [{x,_}=Reg] -> - case gb_trees:lookup(Lbl, Ftree) of - {value,{Reg,MustSave}} -> - D#btb{must_save=MustSave0 or MustSave}; - _ when is_integer(Lbl) -> - btb_error({{label,Lbl},no_suitable_bs_start_match}); - _ -> - btb_error({binary_used_in,Lbl}) - end; - [_|_] when not is_integer(Lbl) -> - btb_error({binary_used_in,Lbl}); - [_|_]=RegList -> - btb_error({multiple_uses,RegList}) - end. - -%% btb_follow_branches([Cond], Regs, D) -> D' -%% Recursively follow all the branches. - -btb_follow_branches([{f,Lbl}|T], Regs, D0) -> - D = btb_follow_branch(Lbl, Regs, D0), - btb_follow_branches(T, Regs, D); -btb_follow_branches([_|T], Regs, D) -> - btb_follow_branches(T, Regs, D); -btb_follow_branches([], _, D) -> D. - -%% btb_follow_branch(Lbl, Regs, D) -> D' -%% Recursively follow the branch. - -btb_follow_branch(0, _Regs, D) -> D; -btb_follow_branch(Lbl, Regs, #btb{ok_br=Br0,index=Li}=D) -> - Key = {Lbl,Regs}, - case gb_sets:is_member(Key, Br0) of - true -> - %% We have already followed this branch and it was OK. - D; - false -> - %% New branch. Try it. - Is = fetch_code_at(Lbl, Li), - #btb{ok_br=Br,must_not_save=MustNotSave,must_save=MustSave} = - btb_reaches_match_1(Is, Regs, D), - - %% Since we got back, this branch is OK. - D#btb{ok_br=gb_sets:insert(Key, Br),must_not_save=MustNotSave, - must_save=MustSave} - end. - -btb_reaches_match_block([{set,Ds,Ss,{alloc,Live,_}}=I|Is], Regs0) -> - %% An allocation instruction or a GC bif. We'll kill all registers - %% if any copy of the context is used as the source to the BIF. - btb_ensure_not_used(Ss, I, Regs0), - Regs1 = btb_kill_not_live(Live, Regs0), - Regs = btb_kill(Ds, Regs1), - btb_reaches_match_block(Is, Regs); -btb_reaches_match_block([{set,[Dst]=Ds,[Src],move}|Is], Regs0) -> - Regs1 = btb_kill(Ds, Regs0), - Regs = case btb_contains_context(Src, Regs1) of - false -> Regs1; - true -> btb_set_context(Dst, Regs1) - end, - btb_reaches_match_block(Is, Regs); -btb_reaches_match_block([{set,Ds,Ss,_}=I|Is], Regs0) -> - btb_ensure_not_used(Ss, I, Regs0), - Regs = btb_kill(Ds, Regs0), - btb_reaches_match_block(Is, Regs); -btb_reaches_match_block([], Regs) -> - Regs. - -%% btb_are_all_killed([Register], [Instruction], D) -> true|false -%% Test whether all of the register are unused in the instruction stream. - -btb_are_all_unused(RegList, Is, #btb{index=Li}) -> - all(fun(R) -> - beam_utils:is_not_used(R, Is, Li) - end, RegList). - -%% btp_regs_from_list([Register]) -> RegisterSet. -%% Create a register set from a list of registers. - -btb_regs_from_list(L) -> - foldl(fun(R, Regs) -> - btb_set_context(R, Regs) - end, {0,0}, L). - -%% btb_set_context(Register, RegisterSet) -> RegisterSet' -%% Update RegisterSet to indicate that Register contains the matching context. - -btb_set_context({x,N}, {Xregs,Yregs}) -> - {Xregs bor (1 bsl N),Yregs}; -btb_set_context({y,N}, {Xregs,Yregs}) -> - {Xregs,Yregs bor (1 bsl N)}. - -%% btb_ensure_not_used([Register], Instruction, RegisterSet) -> ok -%% If any register in RegisterSet (the register(s) known to contain -%% the match context) is used in the list of registers, generate an error. - -btb_ensure_not_used(Rs, I, Regs) -> - case lists:any(fun(R) -> btb_contains_context(R, Regs) end, Rs) of - true -> btb_error({binary_used_in,I}); - false -> ok - end. - -%% btb_kill([Register], RegisterSet) -> RegisterSet' -%% Kill all registers mentioned in the list of registers. - -btb_kill([{x,N}|Rs], {Xregs,Yregs}) -> - btb_kill(Rs, {Xregs band (bnot (1 bsl N)),Yregs}); -btb_kill([{y,N}|Rs], {Xregs,Yregs}) -> - btb_kill(Rs, {Xregs,Yregs band (bnot (1 bsl N))}); -btb_kill([{fr,_}|Rs], Regs) -> - btb_kill(Rs, Regs); -btb_kill([], Regs) -> Regs. - -%% btb_kill_not_live(Live, RegisterSet) -> RegisterSet' -%% Kill all registers indicated not live by Live. - -btb_kill_not_live(Live, {Xregs,Yregs}) -> - {Xregs band ((1 bsl Live)-1),Yregs}. - -%% btb_kill(Regs0) -> Regs -%% Kill all y registers. - -btb_kill_yregs({Xregs,_}) -> {Xregs,0}. - -%% btb_are_registers_empty(RegisterSet) -> true|false -%% Test whether the register set is empty. - -btb_are_registers_empty({0,0}) -> true; -btb_are_registers_empty({_,_}) -> false. - -%% btb_are_x_registers_empty(Regs) -> true|false -%% Test whether the x registers are empty. - -btb_are_x_registers_empty({0,_}) -> true; -btb_are_x_registers_empty({_,_}) -> false. - -%% btb_contains_context(Register, RegisterSet) -> true|false -%% Test whether Register contains the context. - -btb_contains_context({x,N}, {Regs,_}) -> Regs band (1 bsl N) =/= 0; -btb_contains_context({y,N}, {_,Regs}) -> Regs band (1 bsl N) =/= 0; -btb_contains_context(_, _) -> false. - -%% btb_context_regs(RegisterSet) -> [Register] -%% Convert the register set to an explicit list of registers. -btb_context_regs({Xregs,Yregs}) -> - btb_context_regs_1(Xregs, 0, x, btb_context_regs_1(Yregs, 0, y, [])). - -btb_context_regs_1(0, _, _, Acc) -> - Acc; -btb_context_regs_1(Regs, N, Tag, Acc) when (Regs band 1) =:= 1 -> - btb_context_regs_1(Regs bsr 1, N+1, Tag, [{Tag,N}|Acc]); -btb_context_regs_1(Regs, N, Tag, Acc) -> - btb_context_regs_1(Regs bsr 1, N+1, Tag, Acc). - -%% btb_index([Function]) -> GbTree({EntryLabel,{Register,MustSave}}) -%% Build an index of functions that accept a match context instead of -%% a binary. MustSave is true if the function may pass the match -%% context to the bs_context_to_binary instruction (in which case -%% the current position in the binary must have saved into the -%% start position using "bs_save_2 Ctx start"). - -btb_index(Fs) -> - btb_index_1(Fs, []). - -btb_index_1([{function,_,_,Entry,Is0}|Fs], Acc0) -> - Is = drop_to_label(Is0, Entry), - Acc = btb_index_2(Is, Entry, false, Acc0), - btb_index_1(Fs, Acc); -btb_index_1([], Acc) -> gb_trees:from_orddict(sort(Acc)). - -btb_index_2([{test,bs_start_match2,{f,_},_,[Reg,_],Reg}|_], - Entry, MustSave, Acc) -> - [{Entry,{Reg,MustSave}}|Acc]; -btb_index_2(Is0, Entry, _, Acc) -> - try btb_index_find_start_match(Is0) of - Is -> btb_index_2(Is, Entry, true, Acc) - catch - throw:none -> Acc - end. - -drop_to_label([{label,L}|Is], L) -> Is; -drop_to_label([_|Is], L) -> drop_to_label(Is, L). - -btb_index_find_start_match([{test,_,{f,F},_},{bs_context_to_binary,_}|Is]) -> - btb_index_find_label(Is, F); -btb_index_find_start_match(_) -> - throw(none). - -btb_index_find_label([{label,L}|Is], L) -> Is; -btb_index_find_label([_|Is], L) -> btb_index_find_label(Is, L). - -btb_error(Error) -> - throw({error,Error}). - -fetch_code_at(Lbl, Li) -> - case beam_utils:code_at(Lbl, Li) of - Is when is_list(Is) -> Is - end. - -%%% -%%% Compilation information warnings. -%%% - -btb_comment_opt({field_flags,[{anno,A}|_]}) -> - {'%',{bin_opt,A}}; -btb_comment_opt(_) -> - {'%',{bin_opt,[]}}. - -btb_comment_no_opt(Reason, {field_flags,[{anno,A}|_]}) -> - {'%',{no_bin_opt,Reason,A}}; -btb_comment_no_opt(Reason, _) -> - {'%',{no_bin_opt,Reason,[]}}. - -collect_warnings(Fs) -> - D = warning_index_functions(Fs), - foldl(fun(F, A) -> collect_warnings_fun(F, D, A) end, [], Fs). - -collect_warnings_fun({function,_,_,_,Is}, D, A) -> - collect_warnings_instr(Is, D, A). - -collect_warnings_instr([{'%',{bin_opt,Where}}|Is], D, Acc0) -> - Acc = add_warning(bin_opt, Where, Acc0), - collect_warnings_instr(Is, D, Acc); -collect_warnings_instr([{'%',{no_bin_opt,Reason0,Where}}|Is], D, Acc0) -> - Reason = warning_translate_label(Reason0, D), - Acc = add_warning({no_bin_opt,Reason}, Where, Acc0), - collect_warnings_instr(Is, D, Acc); -collect_warnings_instr([_|Is], D, Acc) -> - collect_warnings_instr(Is, D, Acc); -collect_warnings_instr([], _, Acc) -> Acc. - -add_warning(Term, Anno, Ws) -> - Line = get_line(Anno), - File = get_file(Anno), - [{File,[{Line,?MODULE,Term}]}|Ws]. - -warning_translate_label(Term, D) when is_tuple(Term) -> - case element(1, Term) of - {label,F} -> - FA = gb_trees:get(F, D), - setelement(1, Term, FA); - _ -> Term - end; -warning_translate_label(Term, _) -> Term. - -get_line([Line|_]) when is_integer(Line) -> Line; -get_line([_|T]) -> get_line(T); -get_line([]) -> none. - -get_file([{file,File}|_]) -> File; -get_file([_|T]) -> get_file(T); -get_file([]) -> "no_file". % should not happen - -warning_index_functions(Fs) -> - D = [{Entry,{F,A}} || {function,F,A,Entry,_} <- Fs], - gb_trees:from_orddict(sort(D)). - -format_error_1({binary_used_in,{extfunc,M,F,A}}) -> - [io_lib:format("sub binary used by ~p:~p/~p", [M,F,A])| - case {M,F,A} of - {erlang,split_binary,2} -> - "; SUGGEST using binary matching instead of split_binary/2"; - _ -> - "" - end]; -format_error_1({binary_used_in,_}) -> - "sub binary is used or returned"; -format_error_1({multiple_uses,_}) -> - "sub binary is matched or used in more than one place"; -format_error_1(unsuitable_bs_start_match) -> - "the binary matching instruction that follows in the same function " - "have problems that prevent delayed sub binary optimization " - "(probably indicated by INFO warnings)"; -format_error_1({{F,A},no_suitable_bs_start_match}) -> - io_lib:format("called function ~p/~p does not begin with a suitable " - "binary matching instruction", [F,A]); -format_error_1(must_and_must_not_save) -> - "different control paths use different positions in the binary"; -format_error_1({_,I,not_handled}) -> - case I of - {'catch',_,_} -> - "the compiler currently does not attempt the delayed sub binary " - "optimization when catch is used"; - {'try',_,_} -> - "the compiler currently does not attempt the delayed sub binary " - "optimization when try/catch is used"; - _ -> - io_lib:format("compiler limitation: instruction ~p prevents " - "delayed sub binary optimization", [I]) - end; -format_error_1(Term) -> - io_lib:format("~w", [Term]). diff --git a/lib/compiler/src/beam_clean.erl b/lib/compiler/src/beam_clean.erl index f5f0ac2218..7299654476 100644 --- a/lib/compiler/src/beam_clean.erl +++ b/lib/compiler/src/beam_clean.erl @@ -23,17 +23,15 @@ -export([module/2]). -export([clean_labels/1]). --import(lists, [foldl/3]). -spec module(beam_utils:module_code(), [compile:option()]) -> {'ok',beam_utils:module_code()}. module({Mod,Exp,Attr,Fs0,_}, Opts) -> Order = [Lbl || {function,_,_,Lbl,_} <- Fs0], - All = foldl(fun({function,_,_,Lbl,_}=Func,D) -> dict:store(Lbl, Func, D) end, - dict:new(), Fs0), + All = maps:from_list([{Lbl,Func} || {function,_,_,Lbl,_}=Func <- Fs0]), WorkList = rootset(Fs0, Exp, Attr), - Used = find_all_used(WorkList, All, sets:from_list(WorkList)), + Used = find_all_used(WorkList, All, cerl_sets:from_list(WorkList)), Fs1 = remove_unused(Order, Used, All), {Fs2,Lc} = clean_labels(Fs1), Fs = maybe_remove_lines(Fs2, Opts), @@ -55,16 +53,16 @@ rootset(Fs, Root0, Attr) -> %% Remove the unused functions. remove_unused([F|Fs], Used, All) -> - case sets:is_element(F, Used) of + case cerl_sets:is_element(F, Used) of false -> remove_unused(Fs, Used, All); - true -> [dict:fetch(F, All)|remove_unused(Fs, Used, All)] + true -> [map_get(F, All)|remove_unused(Fs, Used, All)] end; remove_unused([], _, _) -> []. - + %% Find all used functions. find_all_used([F|Fs0], All, Used0) -> - {function,_,_,_,Code} = dict:fetch(F, All), + {function,_,_,_,Code} = map_get(F, All), {Fs,Used} = update_work_list(Code, {Fs0,Used0}), find_all_used(Fs, All, Used); find_all_used([], _All, Used) -> Used. @@ -78,9 +76,9 @@ update_work_list([_|Is], Sets) -> update_work_list([], Sets) -> Sets. add_to_work_list(F, {Fs,Used}=Sets) -> - case sets:is_element(F, Used) of + case cerl_sets:is_element(F, Used) of true -> Sets; - false -> {[F|Fs],sets:add_element(F, Used)} + false -> {[F|Fs],cerl_sets:add_element(F, Used)} end. diff --git a/lib/compiler/src/beam_dead.erl b/lib/compiler/src/beam_dead.erl deleted file mode 100644 index 546f0461b9..0000000000 --- a/lib/compiler/src/beam_dead.erl +++ /dev/null @@ -1,881 +0,0 @@ -%% -%% %CopyrightBegin% -%% -%% Copyright Ericsson AB 2002-2018. All Rights Reserved. -%% -%% Licensed under the Apache License, Version 2.0 (the "License"); -%% you may not use this file except in compliance with the License. -%% You may obtain a copy of the License at -%% -%% http://www.apache.org/licenses/LICENSE-2.0 -%% -%% Unless required by applicable law or agreed to in writing, software -%% distributed under the License is distributed on an "AS IS" BASIS, -%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -%% See the License for the specific language governing permissions and -%% limitations under the License. -%% -%% %CopyrightEnd% -%% - --module(beam_dead). - --export([module/2]). - -%%% Dead code is code that is executed but has no effect. This -%%% optimization pass either removes dead code or jumps around it, -%%% potentially making it unreachable and a target for the -%%% the beam_jump pass. - --import(lists, [mapfoldl/3,reverse/1]). - - --spec module(beam_utils:module_code(), [compile:option()]) -> - {'ok',beam_utils:module_code()}. - -module({Mod,Exp,Attr,Fs0,_}, _Opts) -> - {Fs1,Lc1} = beam_clean:clean_labels(Fs0), - {Fs,Lc} = mapfoldl(fun function/2, Lc1, Fs1), - %%{Fs,Lc} = {Fs1,Lc1}, - {ok,{Mod,Exp,Attr,Fs,Lc}}. - -function({function,Name,Arity,CLabel,Is0}, Lc0) -> - try - Is1 = beam_jump:remove_unused_labels(Is0), - - %% Initialize label information with the code - %% for the func_info label. Without it, a register - %% may seem to be live when it is not. - [{label,L}|FiIs] = Is1, - D0 = beam_utils:empty_label_index(), - D = beam_utils:index_label(L, FiIs, D0), - - %% Optimize away dead code. - {Is2,Lc} = forward(Is1, Lc0), - Is3 = backward(Is2, D), - Is = move_move_into_block(Is3, []), - {{function,Name,Arity,CLabel,Is},Lc} - catch - Class:Error:Stack -> - io:fwrite("Function: ~w/~w\n", [Name,Arity]), - erlang:raise(Class, Error, Stack) - end. - -%% 'move' instructions outside of blocks may thwart the jump optimizer. -%% Move them back into the block. - -move_move_into_block([{block,Bl0},{move,S,D}|Is], Acc) -> - Bl = Bl0 ++ [{set,[D],[S],move}], - move_move_into_block([{block,Bl}|Is], Acc); -move_move_into_block([{move,S,D}|Is], Acc) -> - Bl = [{set,[D],[S],move}], - move_move_into_block([{block,Bl}|Is], Acc); -move_move_into_block([I|Is], Acc) -> - move_move_into_block(Is, [I|Acc]); -move_move_into_block([], Acc) -> reverse(Acc). - -%%% -%%% Scan instructions in execution order and remove redundant 'move' -%%% instructions. 'move' instructions are redundant if we know that -%%% the register already contains the value being assigned, as in the -%%% following code: -%%% -%%% test is_eq_exact SomeLabel Src Dst -%%% move Src Dst -%%% -%%% or in: -%%% -%%% select_val Register FailLabel [... Literal => L1...] -%%% . -%%% . -%%% . -%%% L1: move Literal Register -%%% -%%% Also add extra labels to help the second backward pass. -%%% - -forward(Is, Lc) -> - forward(Is, #{}, Lc, []). - -forward([{move,_,_}=Move|[{label,L}|_]=Is], D, Lc, Acc) -> - %% move/2 followed by jump/1 is optimized by backward/3. - forward([Move,{jump,{f,L}}|Is], D, Lc, Acc); -forward([{bif,_,_,_,_}=Bif|[{label,L}|_]=Is], D, Lc, Acc) -> - %% bif/4 followed by jump/1 is optimized by backward/3. - forward([Bif,{jump,{f,L}}|Is], D, Lc, Acc); -forward([{block,[]}|Is], D, Lc, Acc) -> - %% Empty blocks can prevent optimizations. - forward(Is, D, Lc, Acc); -forward([{select,select_val,Reg,_,List}=I|Is], D0, Lc, Acc) -> - D = update_value_dict(List, Reg, D0), - forward(Is, D, Lc, [I|Acc]); -forward([{label,Lbl}=LblI,{block,[{set,[Dst],[Lit],move}|BlkIs]}=Blk|Is], D, Lc, Acc) -> - %% Assumption: The target labels in a select_val/3 instruction - %% cannot be reached in any other way than through the select_val/3 - %% instruction (i.e. there can be no fallthrough to such label and - %% it cannot be referenced by, for example, a jump/1 instruction). - Key = {Lbl,Dst}, - Block = case D of - #{Key := Lit} -> {block,BlkIs}; %Safe to remove move instruction. - _ -> Blk %Must keep move instruction. - end, - forward([Block|Is], D, Lc, [LblI|Acc]); -forward([{label,Lbl}=LblI|[{move,Lit,Dst}|Is1]=Is0], D, Lc, Acc) -> - %% Assumption: The target labels in a select_val/3 instruction - %% cannot be reached in any other way than through the select_val/3 - %% instruction (i.e. there can be no fallthrough to such label and - %% it cannot be referenced by, for example, a jump/1 instruction). - Is = case maps:find({Lbl,Dst}, D) of - {ok,Lit} -> Is1; %Safe to remove move instruction. - _ -> Is0 %Keep move instruction. - end, - forward(Is, D, Lc, [LblI|Acc]); -forward([{test,is_eq_exact,_,[Same,Same]}|Is], D, Lc, Acc) -> - forward(Is, D, Lc, Acc); -forward([{test,is_eq_exact,_,[Dst,Src]}=I, - {block,[{set,[Dst],[Src],move}|Bl]}|Is], D, Lc, Acc) -> - forward([I,{block,Bl}|Is], D, Lc, Acc); -forward([{test,is_eq_exact,_,[Dst,Src]}=I,{move,Src,Dst}|Is], D, Lc, Acc) -> - forward([I|Is], D, Lc, Acc); -forward([{test,_,_,_}=I|Is]=Is0, D, Lc, Acc) -> - %% Help the second, backward pass to by inserting labels after - %% relational operators so that they can be skipped if they are - %% known to be true. - case useful_to_insert_label(Is0) of - false -> forward(Is, D, Lc, [I|Acc]); - true -> forward(Is, D, Lc+1, [{label,Lc},I|Acc]) - end; -forward([I|Is], D, Lc, Acc) -> - forward(Is, D, Lc, [I|Acc]); -forward([], _, Lc, Acc) -> {Acc,Lc}. - -update_value_dict([Lit,{f,Lbl}|T], Reg, D0) -> - Key = {Lbl,Reg}, - D = case D0 of - #{Key := inconsistent} -> D0; - #{Key := _} -> D0#{Key := inconsistent}; - _ -> D0#{Key => Lit} - end, - update_value_dict(T, Reg, D); -update_value_dict([], _, D) -> D. - -useful_to_insert_label([_,{label,_}|_]) -> - false; -useful_to_insert_label([{test,Op,_,_}|_]) -> - case Op of - is_lt -> true; - is_ge -> true; - is_eq_exact -> true; - is_ne_exact -> true; - _ -> false - end. - -%%% -%%% Scan instructions in reverse execution order and try to -%%% shortcut branch instructions. -%%% -%%% For example, in this code: -%%% -%%% move Literal Register -%%% jump L1 -%%% . -%%% . -%%% . -%%% L1: test is_{integer,atom} FailLabel Register -%%% select_val {x,0} FailLabel [... Literal => L2...] -%%% . -%%% . -%%% . -%%% L2: ... -%%% -%%% the 'selectval' instruction will always transfer control to L2, -%%% so we can just as well jump to L2 directly by rewriting the -%%% first part of the sequence like this: -%%% -%%% move Literal Register -%%% jump L2 -%%% -%%% If register Register is killed at label L2, we can remove the -%%% 'move' instruction, leaving just the 'jump' instruction: -%%% -%%% jump L2 -%%% -%%% These transformations may leave parts of the code unreachable. -%%% The beam_jump pass will remove the unreachable code. - -backward(Is, D) -> - backward(Is, D, []). - -backward([{test,is_eq_exact,Fail,[Dst,{integer,Arity}]}=I| - [{bif,tuple_size,Fail,[Reg],Dst}|Is]=Is0], D, Acc) -> - %% Provided that Dst is killed following this sequence, - %% we can rewrite the instructions like this: - %% - %% bif tuple_size Fail Reg Dst ==> is_tuple Fail Reg - %% is_eq_exact Fail Dst Integer test_arity Fail Reg Integer - %% - %% (still two instructions, but they they will be combined to - %% one by the loader). - case beam_utils:is_killed(Dst, Acc, D) andalso (Arity bsr 32) =:= 0 of - false -> - %% Not safe because the register Dst is not killed - %% (probably cannot not happen in practice) or the arity - %% does not fit in 32 bits (the loader will fail to load - %% the module). We must move the first instruction to the - %% accumulator to avoid an infinite loop. - backward(Is0, D, [I|Acc]); - true -> - %% Safe. - backward([{test,test_arity,Fail,[Reg,Arity]}, - {test,is_tuple,Fail,[Reg]}|Is], D, Acc) - end; -backward([{label,Lbl}=L|Is], D, Acc) -> - backward(Is, beam_utils:index_label(Lbl, Acc, D), [L|Acc]); -backward([{select,select_val,Reg,{f,Fail0},List0}|Is], D, Acc) -> - List1 = shortcut_select_list(List0, Reg, D, []), - Fail = shortcut_label(Fail0, D), - List = prune_redundant(List1, Fail), - case List of - [] -> - Jump = {jump,{f,Fail}}, - backward([Jump|Is], D, Acc); - [V,F] -> - Test = {test,is_eq_exact,{f,Fail},[Reg,V]}, - Jump = {jump,F}, - backward([Jump,Test|Is], D, Acc); - [{atom,B1},F,{atom,B2},F] when B1 =:= not B2 -> - Test = {test,is_boolean,{f,Fail},[Reg]}, - Jump = {jump,F}, - backward([Jump,Test|Is], D, Acc); - [_|_] -> - Sel = {select,select_val,Reg,{f,Fail},List}, - backward(Is, D, [Sel|Acc]) - end; -backward([{jump,{f,To0}},{move,Src,Reg}=Move|Is], D, Acc) -> - To = shortcut_select_label(To0, Reg, Src, D), - Jump = {jump,{f,To}}, - case is_killed_at(Reg, To, D) of - false -> backward([Move|Is], D, [Jump|Acc]); - true -> backward([Jump|Is], D, Acc) - end; -backward([{jump,{f,To}}=J|[{bif,Op,{f,BifFail},Ops,Reg}|Is]=Is0], D, Acc) -> - try replace_comp_op(To, Reg, Op, Ops, D) of - {Test,Jump} -> - backward([Jump,Test|Is], D, Acc) - catch - throw:not_possible -> - case To =:= BifFail of - true -> - %% The bif instruction is redundant. See the comment - %% in the next clause for why there is no need to - %% test for liveness of Reg at label To. - backward([J|Is], D, Acc); - false -> - backward(Is0, D, [J|Acc]) - end - end; -backward([{jump,{f,To}}=J|[{gc_bif,_,{f,To},_,_,_Dst}|Is]], D, Acc) -> - %% The gc_bif instruction is redundant, since either the gc_bif - %% instruction itself or the jump instruction will transfer control - %% to label To. Note that a gc_bif instruction does not assign its - %% destination register if the failure branch is taken; therefore, - %% the code at label To is not allowed to assume that the destination - %% register is initialized, and it is therefore no need to test - %% for liveness of the destination register at label To. - backward([J|Is], D, Acc); -backward([{test,bs_start_match2,F,Live,[Src,_]=Args,Ctxt}|Is], D, Acc0) -> - {f,To0} = F, - case test_bs_literal(F, Ctxt, D, Acc0) of - {none,Acc} -> - %% Ctxt killed immediately after bs_start_match2. - To = shortcut_bs_context_to_binary(To0, Src, D), - I = {test,is_bitstr,{f,To},[Src]}, - backward(Is, D, [I|Acc]); - {Literal,Acc} -> - %% Ctxt killed after matching a literal. - To = shortcut_bs_context_to_binary(To0, Src, D), - Eq = {test,is_eq_exact,{f,To},[Src,{literal,Literal}]}, - backward(Is, D, [Eq|Acc]); - not_killed -> - %% Ctxt not killed. Not much to do. - To = shortcut_bs_start_match(To0, Src, D), - I = {test,bs_start_match2,{f,To},Live,Args,Ctxt}, - backward(Is, D, [I|Acc0]) - end; -backward([{test,Op,{f,To0},Ops0}|Is], D, Acc) -> - To1 = shortcut_label(To0, D), - To2 = shortcut_rel_op(To1, Op, Ops0, D), - - %% Try to shortcut a repeated test: - %% - %% test Op {f,Fail1} Operands test Op {f,Fail2} Operands - %% . . . ==> ... - %% Fail1: test Op {f,Fail2} Operands Fail1: test Op {f,Fail2} Operands - %% - To = case beam_utils:code_at(To2, D) of - [{test,Op,{f,To3},Ops}|_] -> - case equal_ops(Ops0, Ops) of - true -> To3; - false -> To2 - end; - _Code -> - To2 - end, - I = case Op of - is_eq_exact -> combine_eqs(To, Ops0, D, Acc); - _ -> {test,Op,{f,To},Ops0} - end, - case I of - {test,_,_,_} -> - %% Still a test instruction. Done. - backward(Is, D, [I|Acc]); - _ -> - %% Rewritten to a select_val. Rescan. - backward([I|Is], D, Acc) - end; -backward([{test,Op,{f,To0},Live,Ops0,Dst}|Is], D, Acc) -> - To1 = shortcut_label(To0, D), - - %% Try to shortcut a repeated test: - %% - %% test Op {f,Fail1} _ Ops _ test Op {f,Fail2} _ Ops _ - %% . . . ==> ... - %% Fail1: test Op {f,Fail2} _ Ops _ Fail1: test Op {f,Fail2} _ Ops _ - %% - To = case beam_utils:code_at(To1, D) of - [{test,Op,{f,To2},_,Ops,_}|_] -> - case equal_ops(Ops0, Ops) of - true -> To2; - false -> To1 - end; - _Code -> - To1 - end, - I = {test,Op,{f,To},Live,Ops0,Dst}, - backward(Is, D, [I|Acc]); -backward([{kill,_}=I|Is], D, [{line,_},Exit|_]=Acc) -> - case beam_jump:is_exit_instruction(Exit) of - false -> backward(Is, D, [I|Acc]); - true -> backward(Is, D, Acc) - end; -backward([{bif,'or',{f,To0},[Dst,{atom,false}],Dst}=I|Is], D, - [{test,is_eq_exact,{f,To},[Dst,{atom,true}]}|_]=Acc) -> - case shortcut_label(To0, D) of - To -> - backward(Is, D, Acc); - _ -> - backward(Is, D, [I|Acc]) - end; -backward([{bif,map_get,{f,FF},[Key,Map],_}=I0, - {test,has_map_fields,{f,FT}=F,[Map|Keys0]}=I1|Is], D, Acc) when FF =/= 0 -> - case shortcut_label(FF, D) of - FT -> - case lists:delete(Key, Keys0) of - [] -> - backward([I0|Is], D, Acc); - Keys -> - Test = {test,has_map_fields,F,[Map|Keys]}, - backward([Test|Is], D, [I0|Acc]) - end; - _ -> - backward([I1|Is], D, [I0|Acc]) - end; -backward([{bif,map_get,{f,FF},[_,Map],_}=I0, - {test,is_map,{f,FT},[Map]}=I1|Is], D, Acc) when FF =/= 0 -> - case shortcut_label(FF, D) of - FT -> backward([I0|Is], D, Acc); - _ -> backward([I1|Is], D, [I0|Acc]) - end; -backward([I|Is], D, Acc) -> - backward(Is, D, [I|Acc]); -backward([], _D, Acc) -> Acc. - -equal_ops([{field_flags,FlA0}|T0], [{field_flags,FlB0}|T1]) -> - FlA = lists:keydelete(anno, 1, FlA0), - FlB = lists:keydelete(anno, 1, FlB0), - FlA =:= FlB andalso equal_ops(T0, T1); -equal_ops([Op|T0], [Op|T1]) -> - equal_ops(T0, T1); -equal_ops([], []) -> true; -equal_ops(_, _) -> false. - -shortcut_select_list([Lit,{f,To0}|T], Reg, D, Acc) -> - To = shortcut_select_label(To0, Reg, Lit, D), - shortcut_select_list(T, Reg, D, [{f,To},Lit|Acc]); -shortcut_select_list([], _, _, Acc) -> reverse(Acc). - -shortcut_label(0, _) -> - 0; -shortcut_label(To0, D) -> - case beam_utils:code_at(To0, D) of - [{jump,{f,To}}|_] -> shortcut_label(To, D); - _ -> To0 - end. - -shortcut_select_label(To, Reg, Lit, D) -> - shortcut_rel_op(To, is_ne_exact, [Reg,Lit], D). - -prune_redundant([_,{f,Fail}|T], Fail) -> - prune_redundant(T, Fail); -prune_redundant([V,F|T], Fail) -> - [V,F|prune_redundant(T, Fail)]; -prune_redundant([], _) -> []. - -%% Replace a comparison operator with a test instruction and a jump. -%% For example, if we have this code: -%% -%% bif '=:=' Fail Src1 Src2 {x,0} -%% jump L1 -%% . -%% . -%% . -%% L1: select_val {x,0} FailLabel [... true => L2..., ...false => L3...] -%% -%% the first two instructions can be replaced with -%% -%% test is_eq_exact L3 Src1 Src2 -%% jump L2 -%% -%% provided that {x,0} is killed at both L2 and L3. - -replace_comp_op(To, Reg, Op, Ops, D) -> - False = comp_op_find_shortcut(To, Reg, {atom,false}, D), - True = comp_op_find_shortcut(To, Reg, {atom,true}, D), - {bif_to_test(Op, Ops, False),{jump,{f,True}}}. - -comp_op_find_shortcut(To0, Reg, Val, D) -> - case shortcut_select_label(To0, Reg, Val, D) of - To0 -> - not_possible(); - To -> - case is_killed_at(Reg, To, D) of - false -> not_possible(); - true -> To - end - end. - -bif_to_test(Name, Args, Fail) -> - try - beam_utils:bif_to_test(Name, Args, {f,Fail}) - catch - error:_ -> not_possible() - end. - -not_possible() -> throw(not_possible). - -%% combine_eqs(To, Operands, Acc) -> Instruction. -%% Combine two is_eq_exact instructions or (an is_eq_exact -%% instruction and a select_val instruction) to a select_val -%% instruction if possible. -%% -%% Example: -%% -%% is_eq_exact F1 Reg Lit1 select_val Reg F2 [ Lit1 L1 -%% L1: . Lit2 L2 ] -%% . -%% . ==> -%% . -%% F1: is_eq_exact F2 Reg Lit2 F1: is_eq_exact F2 Reg Lit2 -%% L2: .... L2: -%% -combine_eqs(To, [Reg,{Type,_}=Lit1]=Ops, D, Acc) - when Type =:= atom; Type =:= integer -> - Next = case Acc of - [{label,Lbl}|_] -> Lbl; - [{jump,{f,Lbl}}|_] -> Lbl - end, - case beam_utils:code_at(To, D) of - [{test,is_eq_exact,{f,F2},[Reg,{Type,_}=Lit2]}, - {label,L2}|_] when Lit1 =/= Lit2 -> - {select,select_val,Reg,{f,F2},[Lit1,{f,Next},Lit2,{f,L2}]}; - [{test,is_eq_exact,{f,F2},[Reg,{Type,_}=Lit2]}, - {jump,{f,L2}}|_] when Lit1 =/= Lit2 -> - {select,select_val,Reg,{f,F2},[Lit1,{f,Next},Lit2,{f,L2}]}; - [{select,select_val,Reg,{f,F2},[{Type,_}|_]=List0}|_] -> - List = remove_from_list(Lit1, List0), - {select,select_val,Reg,{f,F2},[Lit1,{f,Next}|List]}; - _Is -> - {test,is_eq_exact,{f,To},Ops} - end; -combine_eqs(To, Ops, _D, _Acc) -> - {test,is_eq_exact,{f,To},Ops}. - -remove_from_list(Lit, [Lit,{f,_}|T]) -> - T; -remove_from_list(Lit, [Val,{f,_}=Fail|T]) -> - [Val,Fail|remove_from_list(Lit, T)]; -remove_from_list(_, []) -> []. - - -test_bs_literal(F, Ctxt, D, - [{test,bs_match_string,F,[Ctxt,Bs]}, - {test,bs_test_tail2,F,[Ctxt,0]}|Acc]) -> - test_bs_literal_1(Ctxt, Acc, D, Bs); -test_bs_literal(F, Ctxt, D, [{test,bs_test_tail2,F,[Ctxt,0]}|Acc]) -> - test_bs_literal_1(Ctxt, Acc, D, <<>>); -test_bs_literal(_, Ctxt, D, Acc) -> - test_bs_literal_1(Ctxt, Acc, D, none). - -test_bs_literal_1(Ctxt, Is, D, Literal) -> - case beam_utils:is_killed(Ctxt, Is, D) of - true -> {Literal,Is}; - false -> not_killed - end. - -%% shortcut_bs_start_match(TargetLabel, Reg) -> TargetLabel -%% A failing bs_start_match2 instruction means that the source (Reg) -%% cannot be a binary. That means that it is safe to skip -%% bs_context_to_binary instructions operating on Reg, and -%% bs_start_match2 instructions operating on Reg. - -shortcut_bs_start_match(To, Reg, D) -> - shortcut_bs_start_match_1(beam_utils:code_at(To, D), Reg, To, D). - -shortcut_bs_start_match_1([{bs_context_to_binary,Reg}|Is], Reg, To, D) -> - shortcut_bs_start_match_1(Is, Reg, To, D); -shortcut_bs_start_match_1([{jump,{f,To}}|_], Reg, _, D) -> - Code = beam_utils:code_at(To, D), - shortcut_bs_start_match_1(Code, Reg, To, D); -shortcut_bs_start_match_1([{test,bs_start_match2,{f,To},_,[Reg|_],_}|_], - Reg, _, D) -> - Code = beam_utils:code_at(To, D), - shortcut_bs_start_match_1(Code, Reg, To, D); -shortcut_bs_start_match_1(_, _, To, _) -> - To. - -%% shortcut_bs_context_to_binary(TargetLabel, Reg) -> TargetLabel -%% If a bs_start_match2 instruction has been eliminated, the -%% bs_context_to_binary instruction can be eliminated too. - -shortcut_bs_context_to_binary(To, Reg, D) -> - shortcut_bs_ctb_1(beam_utils:code_at(To, D), Reg, To, D). - -shortcut_bs_ctb_1([{bs_context_to_binary,Reg}|Is], Reg, To, D) -> - shortcut_bs_ctb_1(Is, Reg, To, D); -shortcut_bs_ctb_1([{jump,{f,To}}|_], Reg, _, D) -> - Code = beam_utils:code_at(To, D), - shortcut_bs_ctb_1(Code, Reg, To, D); -shortcut_bs_ctb_1(_, _, To, _) -> - To. - -%% shortcut_rel_op(FailLabel, Operator, [Operand], D) -> FailLabel' -%% Try to shortcut the given test instruction. Example: -%% -%% is_ge L1 {x,0} 48 -%% . -%% . -%% . -%% L1: is_ge L2 {x,0} 65 -%% -%% The first test instruction can be rewritten to "is_ge L2 {x,0} 48" -%% since the instruction at L1 will also fail. -%% -%% If there are instructions between L1 and the other test instruction -%% it may still be possible to do the shortcut. For example: -%% -%% L1: is_eq_exact L3 {x,0} 92 -%% is_ge L2 {x,0} 65 -%% -%% Since the first test instruction failed, we know that {x,0} must -%% be less than 48; therefore, we know that {x,0} cannot be equal to -%% 92 and the jump to L3 cannot happen. - -shortcut_rel_op(To, Op, Ops, D) -> - case normalize_op({test,Op,{f,To},Ops}) of - {{NormOp,A,B},_} -> - Normalized = {negate_op(NormOp),A,B}, - shortcut_rel_op_fp(To, Normalized, D); - {_,_} -> - To; - error -> - To - end. - -shortcut_rel_op_fp(To0, Normalized, D) -> - Code = beam_utils:code_at(To0, D), - case shortcut_any_label(Code, Normalized) of - error -> - To0; - To -> - shortcut_rel_op_fp(To, Normalized, D) - end. - -%% shortcut_any_label([Instruction], PrevCondition) -> FailLabel | error -%% Using PrevCondition (a previous condition known to be true), -%% try to shortcut to another failure label. - -shortcut_any_label([{jump,{f,Lbl}}|_], _Prev) -> - Lbl; -shortcut_any_label([{label,Lbl}|_], _Prev) -> - Lbl; -shortcut_any_label([{select,select_val,R,{f,Fail},L}|_], Prev) -> - shortcut_selectval(L, R, Fail, Prev); -shortcut_any_label([I|Is], Prev) -> - case normalize_op(I) of - error -> - error; - {Normalized,Fail} -> - %% We have a relational operator. - case will_succeed(Prev, Normalized) of - no -> - %% This test instruction will always branch - %% to Fail. - Fail; - yes -> - %% This test instruction will never branch, - %% so we will look at the next instruction. - shortcut_any_label(Is, Prev); - maybe -> - %% May or may not branch. From now on, we can only - %% shortcut to the this specific failure label - %% Fail. - shortcut_specific_label(Is, Fail, Prev) - end - end. - -%% shortcut_specific_label([Instruction], FailLabel, PrevCondition) -> -%% FailLabel | error -%% We have previously encountered a test instruction that may or -%% may not branch to FailLabel. Therefore we are only allowed -%% to do the shortcut to the same fail label (FailLabel). - -shortcut_specific_label([{label,_}|Is], Fail, Prev) -> - shortcut_specific_label(Is, Fail, Prev); -shortcut_specific_label([{select,select_val,R,{f,F},L}|_], Fail, Prev) -> - case shortcut_selectval(L, R, F, Prev) of - Fail -> Fail; - _ -> error - end; -shortcut_specific_label([I|Is], Fail, Prev) -> - case normalize_op(I) of - error -> - error; - {Normalized,Fail} -> - case will_succeed(Prev, Normalized) of - no -> - %% Will branch to FailLabel. - Fail; - yes -> - %% Will definitely never branch. - shortcut_specific_label(Is, Fail, Prev); - maybe -> - %% May branch, but still OK since it will branch - %% to FailLabel. - shortcut_specific_label(Is, Fail, Prev) - end; - {Normalized,_} -> - %% This test instruction will branch to a different - %% fail label, if it branches at all. - case will_succeed(Prev, Normalized) of - yes -> - %% Still OK, since the branch will never be - %% taken. - shortcut_specific_label(Is, Fail, Prev); - no -> - %% Give up. The branch will definitely be taken - %% to a different fail label. - error; - maybe -> - %% Give up. If the branch is taken, it will be - %% to a different fail label. - error - end - end. - - -%% shortcut_selectval(List, Reg, Fail, PrevCond) -> FailLabel | error -%% Try to shortcut a selectval instruction. A selectval instruction -%% is equivalent to the following instruction sequence: -%% -%% is_ne_exact L1 Reg Value1 -%% . -%% . -%% . -%% is_ne_exact LN Reg ValueN -%% jump DefaultFailLabel -%% -shortcut_selectval([Val,{f,Lbl}|T], R, Fail, Prev) -> - case will_succeed(Prev, {'=/=',R,get_literal(Val)}) of - yes -> shortcut_selectval(T, R, Fail, Prev); - no -> Lbl; - maybe -> error - end; -shortcut_selectval([], _, Fail, _) -> Fail. - -%% will_succeed(PrevCondition, Condition) -> yes | no | maybe -%% PrevCondition is a condition known to be true. This function -%% will tell whether Condition will succeed. - -will_succeed({Op1,Reg,A}, {Op2,Reg,B}) -> - will_succeed_1(Op1, A, Op2, B); -will_succeed({'=:=',Reg,{literal,A}}, {TypeTest,Reg}) -> - case erlang:TypeTest(A) of - false -> no; - true -> yes - end; -will_succeed({_,_,_}, maybe) -> - maybe; -will_succeed({_,_,_}, Test) when is_tuple(Test) -> - maybe. - -will_succeed_1('=:=', A, '<', B) -> - if - B =< A -> no; - true -> yes - end; -will_succeed_1('=:=', A, '=<', B) -> - if - B < A -> no; - true -> yes - end; -will_succeed_1('=:=', A, '=:=', B) -> - if - A =:= B -> yes; - true -> no - end; -will_succeed_1('=:=', A, '=/=', B) -> - if - A =:= B -> no; - true -> yes - end; -will_succeed_1('=:=', A, '>=', B) -> - if - B > A -> no; - true -> yes - end; -will_succeed_1('=:=', A, '>', B) -> - if - B >= A -> no; - true -> yes - end; - -will_succeed_1('=/=', A, '=/=', B) when A =:= B -> yes; -will_succeed_1('=/=', A, '=:=', B) when A =:= B -> no; - -will_succeed_1('<', A, '=:=', B) when B >= A -> no; -will_succeed_1('<', A, '=/=', B) when B >= A -> yes; -will_succeed_1('<', A, '<', B) when B >= A -> yes; -will_succeed_1('<', A, '=<', B) when B > A -> yes; -will_succeed_1('<', A, '>=', B) when B > A -> no; -will_succeed_1('<', A, '>', B) when B >= A -> no; - -will_succeed_1('=<', A, '=:=', B) when B > A -> no; -will_succeed_1('=<', A, '=/=', B) when B > A -> yes; -will_succeed_1('=<', A, '<', B) when B > A -> yes; -will_succeed_1('=<', A, '=<', B) when B >= A -> yes; -will_succeed_1('=<', A, '>=', B) when B > A -> no; -will_succeed_1('=<', A, '>', B) when B >= A -> no; - -will_succeed_1('>=', A, '=:=', B) when B < A -> no; -will_succeed_1('>=', A, '=/=', B) when B < A -> yes; -will_succeed_1('>=', A, '<', B) when B =< A -> no; -will_succeed_1('>=', A, '=<', B) when B < A -> no; -will_succeed_1('>=', A, '>=', B) when B =< A -> yes; -will_succeed_1('>=', A, '>', B) when B < A -> yes; - -will_succeed_1('>', A, '=:=', B) when B =< A -> no; -will_succeed_1('>', A, '=/=', B) when B =< A -> yes; -will_succeed_1('>', A, '<', B) when B =< A -> no; -will_succeed_1('>', A, '=<', B) when B < A -> no; -will_succeed_1('>', A, '>=', B) when B =< A -> yes; -will_succeed_1('>', A, '>', B) when B < A -> yes; - -will_succeed_1(_, _, _, _) -> maybe. - -%% normalize_op(Instruction) -> {Normalized,FailLabel} | error -%% Normalized = {Operator,Register,Literal} | -%% {TypeTest,Register} | -%% maybe -%% Operation = '<' | '=<' | '=:=' | '=/=' | '>=' | '>' -%% TypeTest = is_atom | is_integer ... -%% Literal = {literal,Term} -%% -%% Normalize a relational operator to facilitate further -%% comparisons between operators. Always make the register -%% operand the first operand. Thus the following instruction: -%% -%% {test,is_ge,{f,99},{integer,13},{x,0}} -%% -%% will be normalized to: -%% -%% {'=<',{x,0},{literal,13}} -%% -%% NOTE: Bit syntax test instructions are scary. They may change the -%% state of match contexts and update registers, so we don't dare -%% mess with them. - -normalize_op({test,is_ge,{f,Fail},Ops}) -> - normalize_op_1('>=', Ops, Fail); -normalize_op({test,is_lt,{f,Fail},Ops}) -> - normalize_op_1('<', Ops, Fail); -normalize_op({test,is_eq_exact,{f,Fail},Ops}) -> - normalize_op_1('=:=', Ops, Fail); -normalize_op({test,is_ne_exact,{f,Fail},Ops}) -> - normalize_op_1('=/=', Ops, Fail); -normalize_op({test,Op,{f,Fail},[R]}) -> - case erl_internal:new_type_test(Op, 1) of - true -> {{Op,R},Fail}; - false -> {maybe,Fail} - end; -normalize_op({test,_,{f,Fail},_}=I) -> - case beam_utils:is_pure_test(I) of - true -> {maybe,Fail}; - false -> error - end; -normalize_op(_) -> - error. - -normalize_op_1(Op, [Op1,Op2], Fail) -> - case {get_literal(Op1),get_literal(Op2)} of - {error,error} -> - %% Both operands are registers. - {maybe,Fail}; - {error,Lit} -> - {{Op,Op1,Lit},Fail}; - {Lit,error} -> - {{turn_op(Op),Op2,Lit},Fail}; - {_,_} -> - %% Both operands are literals. Can probably only - %% happen if the Core Erlang optimizations passes were - %% turned off, so don't bother trying to do something - %% smart here. - {maybe,Fail} - end. - -turn_op('<') -> '>'; -turn_op('>=') -> '=<'; -turn_op('=:='=Op) -> Op; -turn_op('=/='=Op) -> Op. - -negate_op('>=') -> '<'; -negate_op('<') -> '>='; -negate_op('=<') -> '>'; -negate_op('>') -> '=<'; -negate_op('=:=') -> '=/='; -negate_op('=/=') -> '=:='. - -get_literal({atom,Val}) -> - {literal,Val}; -get_literal({integer,Val}) -> - {literal,Val}; -get_literal({float,Val}) -> - {literal,Val}; -get_literal(nil) -> - {literal,[]}; -get_literal({literal,_}=Lit) -> - Lit; -get_literal({_,_}) -> error. - - -%%% -%%% Removing stores to Y registers is not always safe -%%% if there is an instruction that causes an exception -%%% within a catch. In practice, there are few or no -%%% opportunities for removing stores to Y registers anyway -%%% if sys_core_fold has been run. -%%% - -is_killed_at({x,_}=Reg, Lbl, D) -> - beam_utils:is_killed_at(Reg, Lbl, D); -is_killed_at({y,_}, _, _) -> - false. diff --git a/lib/compiler/src/beam_dict.erl b/lib/compiler/src/beam_dict.erl index 990e86062a..b2056332e6 100644 --- a/lib/compiler/src/beam_dict.erl +++ b/lib/compiler/src/beam_dict.erl @@ -126,18 +126,17 @@ import(Mod0, Name0, Arity, #asm{imports=Imp0,next_import=NextIndex}=D0) {NextIndex,D2#asm{imports=Imp,next_import=NextIndex+1}} end. -%% Returns the index for a string in the string table (adding the string to the -%% table if necessary). +%% Returns the index for a binary string in the string table (adding +%% the string to the table if necessary). %% string(String, Dict) -> {Offset, Dict'} --spec string(string(), bdict()) -> {non_neg_integer(), bdict()}. +-spec string(binary(), bdict()) -> {non_neg_integer(), bdict()}. -string(Str, Dict) when is_list(Str) -> +string(BinString, Dict) when is_binary(BinString) -> #asm{strings=Strings,string_offset=NextOffset} = Dict, - StrBin = list_to_binary(Str), - case old_string(StrBin, Strings) of + case old_string(BinString, Strings) of none -> - NewDict = Dict#asm{strings = <<Strings/binary,StrBin/binary>>, - string_offset=NextOffset+byte_size(StrBin)}, + NewDict = Dict#asm{strings = <<Strings/binary,BinString/binary>>, + string_offset=NextOffset+byte_size(BinString)}, {NextOffset,NewDict}; Offset when is_integer(Offset) -> {NextOffset-Offset,Dict} diff --git a/lib/compiler/src/beam_disasm.erl b/lib/compiler/src/beam_disasm.erl index d0be39f520..7d048716e4 100644 --- a/lib/compiler/src/beam_disasm.erl +++ b/lib/compiler/src/beam_disasm.erl @@ -1105,6 +1105,16 @@ resolve_inst({get_hd,[Src,Dst]},_,_,_) -> resolve_inst({get_tl,[Src,Dst]},_,_,_) -> {get_tl,Src,Dst}; +%% OTP 22 +resolve_inst({bs_start_match3,[Fail,Bin,Live,Dst]},_,_,_) -> + {bs_start_match3,Fail,Bin,Live,Dst}; +resolve_inst({bs_get_tail,[Src,Dst,Live]},_,_,_) -> + {bs_get_tail,Src,Dst,Live}; +resolve_inst({bs_get_position,[Src,Dst,Live]},_,_,_) -> + {bs_get_position,Src,Dst,Live}; +resolve_inst({bs_set_position,[Src,Dst]},_,_,_) -> + {bs_set_position,Src,Dst}; + %% %% OTP 22. %% diff --git a/lib/compiler/src/beam_except.erl b/lib/compiler/src/beam_except.erl index 98831d87a7..09925b2872 100644 --- a/lib/compiler/src/beam_except.erl +++ b/lib/compiler/src/beam_except.erl @@ -31,7 +31,7 @@ %%% erlang:error(function_clause, Args) => jump FuncInfoLabel %%% --import(lists, [reverse/1,seq/2]). +-import(lists, [reverse/1,reverse/2,seq/2,splitwith/2]). -spec module(beam_utils:module_code(), [compile:option()]) -> {'ok',beam_utils:module_code()}. @@ -53,7 +53,7 @@ function({function,Name,Arity,CLabel,Is0}) -> -record(st, {lbl :: beam_asm:label(), %func_info label loc :: [_], %location for func_info - arity :: arity() %arity for function + arity :: arity() %arity for function }). function_1(Is0) -> @@ -74,27 +74,33 @@ translate([I|Is], St, Acc) -> translate([], _, Acc) -> reverse(Acc). -translate_1(Ar, I, Is, St, [{line,_}=Line|Acc1]=Acc0) -> - case dig_out(Ar, Acc1) of +translate_1(Ar, I, Is, #st{arity=Arity}=St, [{line,_}=Line|Acc1]=Acc0) -> + case dig_out(Ar, Arity, Acc1) of no -> translate(Is, St, [I|Acc0]); - {yes,{function_clause,Arity},Acc2} -> - case {Line,St} of - {{line,Loc},#st{lbl=Fi,loc=Loc,arity=Arity}} -> + {yes,function_clause,Acc2} -> + case {Is,Line,St} of + {[return|_],{line,Loc},#st{lbl=Fi,loc=Loc}} -> Instr = {jump,{f,Fi}}, translate(Is, St, [Instr|Acc2]); - {_,_} -> - %% This must be "error(function_clause, Args)" in - %% the Erlang source code or a fun. Don't translate. + {_,_,_} -> + %% Not a call_only instruction, or not the same + %% location information as in in the line instruction + %% before the func_info instruction. Not safe + %% to translate to a jump. translate(Is, St, [I|Acc0]) end; {yes,Instr,Acc2} -> translate(Is, St, [Instr,Line|Acc2]) end. -dig_out(Ar, [{kill,_}|Is]) -> - dig_out(Ar, Is); -dig_out(1, [{block,Bl0}|Is]) -> +dig_out(1, _Arity, Is) -> + dig_out(Is); +dig_out(2, Arity, Is) -> + dig_out_fc(Arity, Is); +dig_out(_, _, _) -> no. + +dig_out([{block,Bl0}|Is]) -> case dig_out_block(reverse(Bl0)) of no -> no; {yes,What,[]} -> @@ -102,12 +108,7 @@ dig_out(1, [{block,Bl0}|Is]) -> {yes,What,Bl} -> {yes,What,[{block,Bl}|Is]} end; -dig_out(2, [{block,Bl}|Is]) -> - case dig_out_block_fc(Bl) of - no -> no; - {yes,What} -> {yes,What,Is} - end; -dig_out(_, _) -> no. +dig_out(_) -> no. dig_out_block([{set,[{x,0}],[{atom,if_clause}],move}]) -> {yes,if_end,[]}; @@ -141,36 +142,99 @@ fix_block_1([{set,[],[],{alloc,Live,{F1,F2,Needed0,F3}}}|Is], Words) -> fix_block_1([I|Is], Words) -> [I|fix_block_1(Is, Words)]. -dig_out_block_fc([{set,[],[],{alloc,Live,_}}|Bl]) -> - Regs = maps:from_list([{{x,X},{arg,X}} || X <- seq(0, Live-1)]), - dig_out_fc(Bl, Regs); -dig_out_block_fc(_) -> no. -dig_out_fc([{set,[Dst],[Hd,Tl],put_list}|Is], Regs0) -> - Regs = Regs0#{Dst=>{cons,get_reg(Hd, Regs0),get_reg(Tl, Regs0)}}, - dig_out_fc(Is, Regs); -dig_out_fc([{set,[Dst],[Src],move}|Is], Regs0) -> - Regs = Regs0#{Dst=>get_reg(Src, Regs0)}, - dig_out_fc(Is, Regs); -dig_out_fc([{set,_,_,_}|_], _Regs) -> - %% Unknown instruction. It is not a function_clause error. - no; -dig_out_fc([], Regs) -> +dig_out_fc(Arity, Is0) -> + Regs0 = maps:from_list([{{x,X},{arg,X}} || X <- seq(0, Arity-1)]), + {Is,Acc0} = splitwith(fun({label,_}) -> false; + ({test,_,_,_}) -> false; + (_) -> true + end, Is0), + {Regs,Acc} = dig_out_fc_1(reverse(Is), Regs0, Acc0), case Regs of #{{x,0}:={atom,function_clause},{x,1}:=Args} -> - dig_out_fc_1(Args, 0); + case moves_from_stack(Args, 0, []) of + {Moves,Arity} -> + {yes,function_clause,reverse(Moves, Acc)}; + {_,_} -> + no + end; #{} -> no end. -dig_out_fc_1({cons,{arg,I},T}, I) -> - dig_out_fc_1(T, I+1); -dig_out_fc_1(nil, I) -> - {yes,{function_clause,I}}; -dig_out_fc_1(_, _) -> no. +dig_out_fc_1([{block,Bl}|Is], Regs0, Acc) -> + Regs = dig_out_fc_block(Bl, Regs0), + dig_out_fc_1(Is, Regs, Acc); +dig_out_fc_1([{bs_set_position,_,_}=I|Is], Regs, Acc) -> + dig_out_fc_1(Is, Regs, [I|Acc]); +dig_out_fc_1([{bs_get_tail,Src,Dst,Live0}|Is], Regs0, Acc) -> + Regs = prune_xregs(Live0, Regs0), + Live = dig_out_stack_live(Regs, Live0), + I = {bs_get_tail,Src,Dst,Live}, + dig_out_fc_1(Is, Regs, [I|Acc]); +dig_out_fc_1([_|_], _Regs, _Acc) -> + {#{},[]}; +dig_out_fc_1([], Regs, Acc) -> + {Regs,Acc}. + +dig_out_fc_block([{set,[],[],{alloc,Live,_}}|Is], Regs0) -> + Regs = prune_xregs(Live, Regs0), + dig_out_fc_block(Is, Regs); +dig_out_fc_block([{set,[Dst],[Hd,Tl],put_list}|Is], Regs0) -> + Regs = Regs0#{Dst=>{cons,get_reg(Hd, Regs0),get_reg(Tl, Regs0)}}, + dig_out_fc_block(Is, Regs); +dig_out_fc_block([{set,[Dst],[Src],move}|Is], Regs0) -> + Regs = Regs0#{Dst=>get_reg(Src, Regs0)}, + dig_out_fc_block(Is, Regs); +dig_out_fc_block([{set,_,_,_}|_], _Regs) -> + %% Unknown instruction. Fail. + #{}; +dig_out_fc_block([], Regs) -> Regs. + +dig_out_stack_live(Regs, Default) -> + Reg = {x,2}, + case Regs of + #{Reg:=List} -> + dig_out_stack_live_1(List, Default); + #{} -> + Default + end. + +dig_out_stack_live_1({cons,{arg,N},T}, Live) -> + dig_out_stack_live_1(T, max(N + 1, Live)); +dig_out_stack_live_1({cons,_,T}, Live) -> + dig_out_stack_live_1(T, Live); +dig_out_stack_live_1(nil, Live) -> + Live; +dig_out_stack_live_1(_, Live) -> Live. + +prune_xregs(Live, Regs) -> + maps:filter(fun({x,X}, _) -> X < Live end, Regs). + +moves_from_stack({cons,{arg,N},_}, I, _Acc) when N =/= I -> + %% Wrong argument. Give up. + {[],-1}; +moves_from_stack({cons,H,T}, I, Acc) -> + case H of + {arg,I} -> + moves_from_stack(T, I+1, Acc); + _ -> + moves_from_stack(T, I+1, [{move,H,{x,I}}|Acc]) + end; +moves_from_stack(nil, I, Acc) -> + {reverse(Acc),I}; +moves_from_stack({literal,[H|T]}, I, Acc) -> + Cons = {cons,tag_literal(H),tag_literal(T)}, + moves_from_stack(Cons, I, Acc). get_reg(R, Regs) -> case Regs of #{R:=Val} -> Val; #{} -> R end. + +tag_literal([]) -> nil; +tag_literal(T) when is_atom(T) -> {atom,T}; +tag_literal(T) when is_float(T) -> {float,T}; +tag_literal(T) when is_integer(T) -> {integer,T}; +tag_literal(T) -> {literal,T}. diff --git a/lib/compiler/src/beam_flatten.erl b/lib/compiler/src/beam_flatten.erl index 973d16a1bc..3e6bc1b1ed 100644 --- a/lib/compiler/src/beam_flatten.erl +++ b/lib/compiler/src/beam_flatten.erl @@ -32,8 +32,7 @@ module({Mod,Exp,Attr,Fs,Lc}, _Opt) -> {ok,{Mod,Exp,Attr,[function(F) || F <- Fs],Lc}}. function({function,Name,Arity,CLabel,Is0}) -> - Is1 = block(Is0), - Is = opt(Is1), + Is = block(Is0), {function,Name,Arity,CLabel,Is}. block(Is) -> @@ -43,21 +42,12 @@ block([{block,Is0}|Is1], Acc) -> block(Is1, norm_block(Is0, Acc)); block([I|Is], Acc) -> block(Is, [I|Acc]); block([], Acc) -> reverse(Acc). -norm_block([{set,[],[],{alloc,R,{_,nostack,_,_}=Alloc}}|Is], Acc0) -> - case insert_alloc_in_bs_init(Acc0, Alloc) of - impossible -> - norm_block(Is, reverse(norm_allocate(Alloc, R), Acc0)); - Acc -> - norm_block(Is, Acc) - end; norm_block([{set,[],[],{alloc,R,Alloc}}|Is], Acc0) -> norm_block(Is, reverse(norm_allocate(Alloc, R), Acc0)); -norm_block([{set,[D1],[S],get_hd},{set,[D2],[S],get_tl}|Is], Acc) -> - I = {get_list,S,D1,D2}, - norm_block(Is, [I|Acc]); -norm_block([I|Is], Acc) -> norm_block(Is, [norm(I)|Acc]); +norm_block([I|Is], Acc) -> + norm_block(Is, [norm(I)|Acc]); norm_block([], Acc) -> Acc. - + norm({set,[D],As,{bif,N,F}}) -> {bif,N,F,As,D}; norm({set,[D],As,{alloc,R,{gc_bif,N,F}}}) -> {gc_bif,N,F,R,As,D}; norm({set,[D],[],init}) -> {init,D}; @@ -91,57 +81,3 @@ norm_allocate({nozero,Ns,0,Inits}, Regs) -> [{allocate,Ns,Regs}|Inits]; norm_allocate({nozero,Ns,Nh,Inits}, Regs) -> [{allocate_heap,Ns,Nh,Regs}|Inits]. - -%% insert_alloc_in_bs_init(ReverseInstructionStream, AllocationInfo) -> -%% impossible | ReverseInstructionStream' -%% A bs_init/6 instruction should not be followed by a test heap instruction. -%% Given the AllocationInfo from a test heap instruction, merge the -%% allocation amounts into the previous bs_init/6 instruction (if any). -%% -insert_alloc_in_bs_init([{bs_put,_,_,_}=I|Is], Alloc) -> - %% The instruction sequence ends with an bs_put/4 instruction. - %% We'll need to search backwards for the bs_init/6 instruction. - insert_alloc_1(Is, Alloc, [I]); -insert_alloc_in_bs_init(_, _) -> impossible. - -insert_alloc_1([{bs_init=Op,Fail,Info0,Live,Ss,Dst}|Is], - {_,nostack,Ws2,[]}, Acc) when is_integer(Live) -> - %% The number of extra heap words is always in the second position - %% in the Info tuple. - Ws1 = element(2, Info0), - Al = beam_utils:combine_heap_needs(Ws1, Ws2), - Info = setelement(2, Info0, Al), - I = {Op,Fail,Info,Live,Ss,Dst}, - reverse(Acc, [I|Is]); -insert_alloc_1([{bs_put,_,_,_}=I|Is], Alloc, Acc) -> - insert_alloc_1(Is, Alloc, [I|Acc]). - -%% opt(Is0) -> Is -%% 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(Is) -> - opt_1(Is, []). - -opt_1([{move,_,{x,0}}=I|Is0], Acc0) -> - case move_past_kill(Is0, I, Acc0) of - impossible -> opt_1(Is0, [I|Acc0]); - {Is,Acc} -> opt_1(Is, Acc) - end; -opt_1([I|Is], Acc) -> - opt_1(Is, [I|Acc]); -opt_1([], 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([{trim,N,_}=I|Is], {move,Src,Dst}=Move, Acc) -> - case Src of - {y,Y} when Y < N-> impossible; - {y,Y} -> {Is,[{move,{y,Y-N},Dst},I|Acc]}; - _ -> {Is,[Move,I|Acc]} - end; -move_past_kill(Is, Move, Acc) -> - {Is,[Move|Acc]}. diff --git a/lib/compiler/src/beam_jump.erl b/lib/compiler/src/beam_jump.erl index 42b4cdaf4f..8b0e3e32f8 100644 --- a/lib/compiler/src/beam_jump.erl +++ b/lib/compiler/src/beam_jump.erl @@ -22,7 +22,7 @@ -module(beam_jump). -export([module/2, - is_unreachable_after/1,is_exit_instruction/1, + is_exit_instruction/1, remove_unused_labels/1]). %%% The following optimisations are done: @@ -101,6 +101,10 @@ %%% always keep the label. (beam_clean will remove any unused %%% labels.) %%% +%%% (7) Replace a jump to a return instruction with a return instruction. +%%% Similarly, replace a jump to deallocate + return with those +%%% instructions. +%%% %%% Note: This modules depends on (almost) all branches and jumps only %%% going forward, so that we can remove instructions (including definition %%% of labels) after any label that has not been referenced by the code @@ -128,27 +132,127 @@ %%% on the program state. %%% --import(lists, [reverse/1,reverse/2,foldl/3]). +-import(lists, [foldl/3,mapfoldl/3,reverse/1,reverse/2]). -type instruction() :: beam_utils:instruction(). -spec module(beam_utils:module_code(), [compile:option()]) -> {'ok',beam_utils:module_code()}. -module({Mod,Exp,Attr,Fs0,Lc}, _Opt) -> - Fs = [function(F) || F <- Fs0], +module({Mod,Exp,Attr,Fs0,Lc0}, _Opt) -> + {Fs,Lc} = mapfoldl(fun function/2, Lc0, Fs0), {ok,{Mod,Exp,Attr,Fs,Lc}}. %% function(Function) -> Function' %% Optimize jumps and branches. %% %% NOTE: This function assumes that there are no labels inside blocks. -function({function,Name,Arity,CLabel,Asm0}) -> - Asm1 = share(Asm0), - Asm2 = move(Asm1), - Asm3 = opt(Asm2, CLabel), - Asm = remove_unused_labels(Asm3), - {function,Name,Arity,CLabel,Asm}. +function({function,Name,Arity,CLabel,Asm0}, Lc0) -> + try + Asm1 = eliminate_moves(Asm0), + {Asm2,Lc} = insert_labels(Asm1, Lc0, []), + Asm3 = share(Asm2), + Asm4 = move(Asm3), + Asm5 = opt(Asm4, CLabel), + Asm6 = unshare(Asm5), + Asm = remove_unused_labels(Asm6), + {{function,Name,Arity,CLabel,Asm},Lc} + catch + Class:Error:Stack -> + io:fwrite("Function: ~w/~w\n", [Name,Arity]), + erlang:raise(Class, Error, Stack) + end. + +%%% +%%% Scan instructions in execution order and remove redundant 'move' +%%% instructions. 'move' instructions are redundant if we know that +%%% the register already contains the value being assigned, as in the +%%% following code: +%%% +%%% select_val Register FailLabel [... Literal => L1...] +%%% . +%%% . +%%% . +%%% L1: move Literal Register +%%% + +eliminate_moves(Is) -> + eliminate_moves(Is, #{}, []). + +eliminate_moves([{select,select_val,Reg,_,List}=I|Is], D0, Acc) -> + D = update_value_dict(List, Reg, D0), + eliminate_moves(Is, D, [I|Acc]); +eliminate_moves([{label,Lbl},{block,[{set,[Dst],[Lit],move}|BlkIs]}=Blk0|Is], + D, Acc0) -> + Acc = [{label,Lbl}|Acc0], + case already_has_value(Lit, Lbl, Dst, D) andalso + no_fallthrough(Acc0) of + true -> + %% Remove redundant 'move' instruction. + Blk = {block,BlkIs}, + eliminate_moves([Blk|Is], D, Acc); + false -> + %% Keep 'move' instruction. + eliminate_moves([Blk0|Is], D, Acc) + end; +eliminate_moves([{block,[]}|Is], D, Acc) -> + %% Empty blocks can prevent further jump optimizations. + eliminate_moves(Is, D, Acc); +eliminate_moves([I|Is], D0, Acc) -> + D = update_unsafe_labels(I, D0), + eliminate_moves(Is, D, [I|Acc]); +eliminate_moves([], _, Acc) -> reverse(Acc). + +no_fallthrough([I|_]) -> + is_unreachable_after(I). + +already_has_value(Lit, Lbl, Reg, D) -> + case D of + #{Lbl:={Reg,Lit}} -> + true; + #{} -> + false + end. + +update_value_dict([Lit,{f,Lbl}|T], Reg, D0) -> + D = case D0 of + #{Lbl:=unsafe} -> D0; + #{Lbl:={Reg,Lit}} -> D0; + #{Lbl:=_} -> D0#{Lbl:=unsafe}; + #{} -> D0#{Lbl=>{Reg,Lit}} + end, + update_value_dict(T, Reg, D); +update_value_dict([], _, D) -> D. + +update_unsafe_labels(I, D) -> + Ls = instr_labels(I), + update_unsafe_labels_1(Ls, D). + +update_unsafe_labels_1([L|Ls], D) -> + update_unsafe_labels_1(Ls, D#{L=>unsafe}); +update_unsafe_labels_1([], D) -> D. + +%%% +%%% It seems to be useful to insert extra labels after certain +%%% test instructions. This used to be done by beam_dead. +%%% + +insert_labels([{test,Op,_,_}=I|Is], Lc, Acc) -> + Useful = case Op of + is_lt -> true; + is_ge -> true; + is_eq_exact -> true; + is_ne_exact -> true; + _ -> false + end, + case Useful of + false -> insert_labels(Is, Lc, [I|Acc]); + true -> insert_labels(Is, Lc+1, [{label,Lc},I|Acc]) + end; +insert_labels([I|Is], Lc, Acc) -> + insert_labels(Is, Lc, [I|Acc]); +insert_labels([], Lc, Acc) -> + {reverse(Acc),Lc}. %%% %%% (1) We try to share the code for identical code segments by replacing all @@ -271,8 +375,6 @@ extract_seq([{line,_}=Line|Is], Acc) -> extract_seq(Is, [Line|Acc]); extract_seq([{block,_}=Bl|Is], Acc) -> extract_seq_1(Is, [Bl|Acc]); -extract_seq([{bs_context_to_binary,_}=I|Is], Acc) -> - extract_seq_1(Is, [I|Acc]); extract_seq([{label,_}|_]=Is, Acc) -> extract_seq_1(Is, Acc); extract_seq(_, _) -> no. @@ -296,14 +398,13 @@ extract_seq_1(_, _) -> no. { entry :: beam_asm:label(), %Entry label (must not be moved). replace :: #{beam_asm:label() := beam_asm:label()}, %Labels to replace. - labels :: cerl_sets:set(), %Set of referenced labels. - index :: beam_utils:code_index() | {lazy,[beam_utils:instruction()]} %Index built lazily only if needed + labels :: cerl_sets:set() %Set of referenced labels. }). opt(Is0, CLabel) -> find_fixpoint(fun(Is) -> Lbls = initial_labels(Is), - St = #st{entry=CLabel,replace=#{},labels=Lbls,index={lazy,Is}}, + St = #st{entry=CLabel,replace=#{},labels=Lbls}, opt(Is, [], St) end, Is0). @@ -313,7 +414,7 @@ find_fixpoint(OptFun, Is0) -> Is -> find_fixpoint(OptFun, Is) end. -opt([{test,_,{f,L}=Lbl,_}=I|[{jump,{f,L}}|_]=Is], Acc0, St0) -> +opt([{test,_,{f,L}=Lbl,_}=I|[{jump,{f,L}}|_]=Is], Acc, St) -> %% We have %% Test Label Ops %% jump Label @@ -322,23 +423,10 @@ opt([{test,_,{f,L}=Lbl,_}=I|[{jump,{f,L}}|_]=Is], Acc0, St0) -> case beam_utils:is_pure_test(I) of false -> %% Test is not pure; we must keep it. - opt(Is, [I|Acc0], label_used(Lbl, St0)); + opt(Is, [I|Acc], label_used(Lbl, St)); true -> %% The test is pure and its failure label is the same %% as in the jump that follows -- thus it is not needed. - %% Check if any of the previous instructions could also be eliminated. - {Acc,St} = opt_useless_loads(Acc0, L, St0), - opt(Is, Acc, St) - end; -opt([{test,_,{f,L}=Lbl,_}=I|[{label,L}|_]=Is], Acc0, St0) -> - %% Similar to the above, except we have a fall-through rather than jump - %% Test Label Ops - %% label Label - case beam_utils:is_pure_test(I) of - false -> - opt(Is, [I|Acc0], label_used(Lbl, St0)); - true -> - {Acc,St} = opt_useless_loads(Acc0, L, St0), opt(Is, Acc, St) end; opt([{test,Test0,{f,L}=Lbl,Ops}=I|[{jump,To}|Is]=Is0], Acc, St) -> @@ -405,46 +493,6 @@ normalize_replace([{From,To0}|Rest], Replace, Acc) -> normalize_replace([], _Replace, Acc) -> maps:from_list(Acc). -%% After eliminating a test, it might happen, that a register was only used -%% in this test. Let's check if that was the case and if it was so, we can -%% eliminate the load into the register completely. -opt_useless_loads([{block,_}|_]=Is, L, #st{index={lazy,FIs}}=St) -> - opt_useless_loads(Is, L, St#st{index=beam_utils:index_labels(FIs)}); -opt_useless_loads([{block,Block0}|Is], L, #st{index=Index}=St) -> - case opt_useless_block_loads(Block0, L, Index) of - [] -> - opt_useless_loads(Is, L, St); - [_|_]=Block -> - {[{block,Block}|Is],St} - end; -%% After eliminating the test and useless blocks, it might happen, -%% that the previous test could also be eliminated. -%% It might be that the label was already marked as used, even if ultimately, -%% it never will be - we can't do much about it at that point, though -opt_useless_loads([{test,_,{f,L},_}=I|Is], L, St) -> - case beam_utils:is_pure_test(I) of - false -> - {[I|Is],St}; - true -> - opt_useless_loads(Is, L, St) - end; -opt_useless_loads(Is, _L, St) -> - {Is,St}. - -opt_useless_block_loads([{set,[Dst],_,_}=I|Is], L, Index) -> - BlockJump = [{block,Is},{jump,{f,L}}], - case beam_utils:is_killed(Dst, BlockJump, Index) of - true -> - %% The register is killed and not used, we can remove the load - opt_useless_block_loads(Is, L, Index); - false -> - [I|opt_useless_block_loads(Is, L, Index)] - end; -opt_useless_block_loads([I|Is], L, Index) -> - [I|opt_useless_block_loads(Is, L, Index)]; -opt_useless_block_loads([], _L, _Index) -> - []. - collect_labels(Is, Label, #st{entry=Entry,replace=Replace} = St) -> collect_labels_1(Is, Label, Entry, Replace, St). @@ -571,58 +619,109 @@ drop_upto_label([{label,_}|_]=Is) -> Is; drop_upto_label([_|Is]) -> drop_upto_label(Is); drop_upto_label([]) -> []. -%% ulbl(Instruction, UsedGbSet) -> UsedGbSet' -%% Update the gb_set UsedGbSet with any function-local labels +%% unshare([Instruction]) -> [Instruction]. +%% Replace a jump to a return sequence (a `return` instruction +%% optionally preced by a `deallocate` instruction) with the return +%% sequence. This always saves execution time and may also save code +%% space (depending on the architecture). Eliminating `jump` +%% instructions also gives beam_trim more opportunities to trim the +%% stack. + +unshare(Is) -> + Short = unshare_collect_short(Is, #{}), + unshare_short(Is, Short). + +unshare_collect_short([{label,L},return|Is], Map) -> + unshare_collect_short(Is, Map#{L=>[return]}); +unshare_collect_short([{label,L},{deallocate,_}=D,return|Is], Map) -> + %% `deallocate` and `return` are combined into one instruction by + %% the loader. + unshare_collect_short(Is, Map#{L=>[D,return]}); +unshare_collect_short([_|Is], Map) -> + unshare_collect_short(Is, Map); +unshare_collect_short([], Map) -> Map. + +unshare_short([{jump,{f,F}}=I|Is], Map) -> + case Map of + #{F:=Seq} -> + Seq ++ unshare_short(Is, Map); + #{} -> + [I|unshare_short(Is, Map)] + end; +unshare_short([I|Is], Map) -> + [I|unshare_short(Is, Map)]; +unshare_short([], _Map) -> []. + +%% ulbl(Instruction, UsedCerlSet) -> UsedCerlSet' +%% Update the cerl_set UsedCerlSet with any function-local labels %% (i.e. not with labels in call instructions) referenced by %% the instruction Instruction. %% %% NOTE: This function does NOT look for labels inside blocks. -ulbl({test,_,Fail,_}, Used) -> - mark_used(Fail, Used); -ulbl({test,_,Fail,_,_,_}, Used) -> - mark_used(Fail, Used); -ulbl({select,_,_,Fail,Vls}, Used) -> - mark_used_list(Vls, mark_used(Fail, Used)); -ulbl({'try',_,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl({'catch',_,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl({jump,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl({loop_rec,Lbl,_}, Used) -> - mark_used(Lbl, Used); -ulbl({loop_rec_end,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl({wait,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl({wait_timeout,Lbl,_To}, Used) -> - mark_used(Lbl, Used); -ulbl({bif,_Name,Lbl,_As,_R}, Used) -> - mark_used(Lbl, Used); -ulbl({gc_bif,_Name,Lbl,_Live,_As,_R}, Used) -> - mark_used(Lbl, Used); -ulbl({bs_init,Lbl,_,_,_,_}, Used) -> - mark_used(Lbl, Used); -ulbl({bs_put,Lbl,_,_}, Used) -> - mark_used(Lbl, Used); -ulbl({put_map,Lbl,_Op,_Src,_Dst,_Live,_List}, Used) -> - mark_used(Lbl, Used); -ulbl({get_map_elements,Lbl,_Src,_List}, Used) -> - mark_used(Lbl, Used); -ulbl({recv_mark,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl({recv_set,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl({fcheckerror,Lbl}, Used) -> - mark_used(Lbl, Used); -ulbl(_, Used) -> Used. - -mark_used({f,0}, Used) -> Used; -mark_used({f,L}, Used) -> cerl_sets:add_element(L, Used). - -mark_used_list([{f,L}|T], Used) -> - mark_used_list(T, cerl_sets:add_element(L, Used)); -mark_used_list([_|T], Used) -> - mark_used_list(T, Used); -mark_used_list([], Used) -> Used. +ulbl(I, Used) -> + case instr_labels(I) of + [] -> + Used; + [Lbl] -> + cerl_sets:add_element(Lbl, Used); + [_|_]=L -> + ulbl_list(L, Used) + end. + +ulbl_list([L|Ls], Used) -> + ulbl_list(Ls, cerl_sets:add_element(L, Used)); +ulbl_list([], Used) -> Used. + +-spec instr_labels(Instruction) -> Labels when + Instruction :: instruction(), + Labels :: [beam_asm:label()]. + +instr_labels({test,_,Fail,_}) -> + do_instr_labels(Fail); +instr_labels({test,_,Fail,_,_,_}) -> + do_instr_labels(Fail); +instr_labels({select,_,_,Fail,Vls}) -> + do_instr_labels_list(Vls, do_instr_labels(Fail)); +instr_labels({'try',_,Lbl}) -> + do_instr_labels(Lbl); +instr_labels({'catch',_,Lbl}) -> + do_instr_labels(Lbl); +instr_labels({jump,Lbl}) -> + do_instr_labels(Lbl); +instr_labels({loop_rec,Lbl,_}) -> + do_instr_labels(Lbl); +instr_labels({loop_rec_end,Lbl}) -> + do_instr_labels(Lbl); +instr_labels({wait,Lbl}) -> + do_instr_labels(Lbl); +instr_labels({wait_timeout,Lbl,_To}) -> + do_instr_labels(Lbl); +instr_labels({bif,_Name,Lbl,_As,_R}) -> + do_instr_labels(Lbl); +instr_labels({gc_bif,_Name,Lbl,_Live,_As,_R}) -> + do_instr_labels(Lbl); +instr_labels({bs_init,Lbl,_,_,_,_}) -> + do_instr_labels(Lbl); +instr_labels({bs_put,Lbl,_,_}) -> + do_instr_labels(Lbl); +instr_labels({put_map,Lbl,_Op,_Src,_Dst,_Live,_List}) -> + do_instr_labels(Lbl); +instr_labels({get_map_elements,Lbl,_Src,_List}) -> + do_instr_labels(Lbl); +instr_labels({recv_mark,Lbl}) -> + do_instr_labels(Lbl); +instr_labels({recv_set,Lbl}) -> + do_instr_labels(Lbl); +instr_labels({fcheckerror,Lbl}) -> + do_instr_labels(Lbl); +instr_labels(_) -> []. + +do_instr_labels({f,0}) -> []; +do_instr_labels({f,F}) -> [F]. + +do_instr_labels_list([{f,F}|T], Acc) -> + do_instr_labels_list(T, [F|Acc]); +do_instr_labels_list([_|T], Acc) -> + do_instr_labels_list(T, Acc); +do_instr_labels_list([], Acc) -> Acc. diff --git a/lib/compiler/src/beam_kernel_to_ssa.erl b/lib/compiler/src/beam_kernel_to_ssa.erl index c55a57ab32..d6e675ae72 100644 --- a/lib/compiler/src/beam_kernel_to_ssa.erl +++ b/lib/compiler/src/beam_kernel_to_ssa.erl @@ -276,12 +276,11 @@ select_nil(#k_val_clause{val=#k_nil{},body=B}, V, Tf, Vf, St0) -> {Is ++ Bis,St}. select_binary(#k_val_clause{val=#k_binary{segs=#k_var{name=Ctx0}},body=B}, - #k_var{anno=Anno0}=Src, Tf, Vf, St0) -> - Anno = #{reuse_for_context=>member(reuse_for_context, Anno0)}, + #k_var{}=Src, Tf, Vf, St0) -> {Ctx,St1} = new_ssa_var(Ctx0, St0), {Bis0,St2} = match_cg(B, Vf, St1), {TestIs,St} = make_cond_branch(succeeded, [Ctx], Tf, St2), - Bis1 = [#b_set{anno=Anno,op=bs_start_match,dst=Ctx, + Bis1 = [#b_set{op=bs_start_match,dst=Ctx, args=[ssa_arg(Src, St)]}] ++ TestIs ++ Bis0, Bis = finish_bs_matching(Bis1), {Bis,St}. @@ -708,10 +707,6 @@ bif_cg(#k_bif{op=#k_remote{mod=#k_atom{val=erlang},name=#k_atom{val=Name}}, %% internal_cg(Bif, [Arg], [Ret], Le, State) -> %% {[Ainstr],State}. -internal_cg(bs_context_to_binary, [Src0], [], _Le, St) -> - Src = ssa_arg(Src0, St), - Set = #b_set{op=context_to_binary,args=[Src]}, - {[Set],St}; internal_cg(dsetelement, [Index0,Tuple0,New0], _Rs, _Le, St) -> [New,Tuple,#b_literal{val=Index1}] = ssa_args([New0,Tuple0,Index0], St), Index = #b_literal{val=Index1-1}, diff --git a/lib/compiler/src/beam_listing.erl b/lib/compiler/src/beam_listing.erl index 8a0ce5b50a..6121593b11 100644 --- a/lib/compiler/src/beam_listing.erl +++ b/lib/compiler/src/beam_listing.erl @@ -66,7 +66,7 @@ module(Stream, [_|_]=Fs) -> foreach(fun (F) -> io:format(Stream, "~p.\n", [F]) end, Fs). format_asm([{label,L}|Is]) -> - [" {label,",integer_to_list(L),"}.\n"|format_asm(Is)]; + [io_lib:format(" {label,~p}.\n", [L])|format_asm(Is)]; format_asm([I|Is]) -> [io_lib:format(" ~p", [I]),".\n"|format_asm(Is)]; format_asm([]) -> []. diff --git a/lib/compiler/src/beam_peep.erl b/lib/compiler/src/beam_peep.erl index 74da6aa704..5730e9704e 100644 --- a/lib/compiler/src/beam_peep.erl +++ b/lib/compiler/src/beam_peep.erl @@ -94,26 +94,26 @@ peep([{gc_bif,_,_,_,_,Dst}=I|Is], SeenTests0, Acc) -> peep([{jump,{f,L}},{label,L}=I|Is], _, Acc) -> %% Sometimes beam_jump has missed this optimization. peep(Is, gb_sets:empty(), [I|Acc]); -peep([{select,Op,R,F,Vls0}|Is], SeenTests0, Acc0) -> +peep([{select,select_val,R,F,Vls0}|Is], SeenTests0, Acc0) -> case prune_redundant_values(Vls0, F) of [] -> %% No values left. Must convert to plain jump. I = {jump,F}, peep([I|Is], gb_sets:empty(), Acc0); - [{atom,_}=Value,Lbl] when Op =:= select_val -> + [{atom,_}=Value,Lbl] -> %% Single value left. Convert to regular test. Is1 = [{test,is_eq_exact,F,[R,Value]},{jump,Lbl}|Is], peep(Is1, SeenTests0, Acc0); - [{integer,_}=Value,Lbl] when Op =:= select_val -> + [{integer,_}=Value,Lbl] -> %% Single value left. Convert to regular test. Is1 = [{test,is_eq_exact,F,[R,Value]},{jump,Lbl}|Is], peep(Is1, SeenTests0, Acc0); - [Arity,Lbl] when Op =:= select_tuple_arity -> - %% Single value left. Convert to regular test - Is1 = [{test,test_arity,F,[R,Arity]},{jump,Lbl}|Is], + [{atom,B1},Lbl,{atom,B2},Lbl] when B1 =:= not B2 -> + %% Replace with is_boolean test. + Is1 = [{test,is_boolean,F,[R]},{jump,Lbl}|Is], peep(Is1, SeenTests0, Acc0); [_|_]=Vls -> - I = {select,Op,R,F,Vls}, + I = {select,select_val,R,F,Vls}, peep(Is, gb_sets:empty(), [I|Acc0]) end; peep([{get_map_elements,Fail,Src,List}=I|Is], _SeenTests, Acc0) -> diff --git a/lib/compiler/src/beam_split.erl b/lib/compiler/src/beam_split.erl deleted file mode 100644 index 7b18946ae0..0000000000 --- a/lib/compiler/src/beam_split.erl +++ /dev/null @@ -1,90 +0,0 @@ -%% -%% %CopyrightBegin% -%% -%% Copyright Ericsson AB 2011-2018. All Rights Reserved. -%% -%% Licensed under the Apache License, Version 2.0 (the "License"); -%% you may not use this file except in compliance with the License. -%% You may obtain a copy of the License at -%% -%% http://www.apache.org/licenses/LICENSE-2.0 -%% -%% Unless required by applicable law or agreed to in writing, software -%% distributed under the License is distributed on an "AS IS" BASIS, -%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -%% See the License for the specific language governing permissions and -%% limitations under the License. -%% -%% %CopyrightEnd% -%% - --module(beam_split). --export([module/2]). - --import(lists, [reverse/1]). - --spec module(beam_utils:module_code(), [compile:option()]) -> - {'ok',beam_utils:module_code()}. - -module({Mod,Exp,Attr,Fs0,Lc}, _Opts) -> - Fs = [split_blocks(F) || F <- Fs0], - {ok,{Mod,Exp,Attr,Fs,Lc}}. - -%% We must split the basic block when we encounter instructions with labels, -%% such as catches and BIFs. All labels must be visible outside the blocks. - -split_blocks({function,Name,Arity,CLabel,Is0}) -> - Is = split_blocks(Is0, []), - {function,Name,Arity,CLabel,Is}. - -split_blocks([{block,Bl}|Is], Acc0) -> - Acc = split_block(Bl, [], Acc0), - split_blocks(Is, Acc); -split_blocks([I|Is], Acc) -> - split_blocks(Is, [I|Acc]); -split_blocks([], Acc) -> reverse(Acc). - -split_block([{set,[R],As,{bif,N,{f,Lbl}=Fail}}|Is], Bl, Acc) when Lbl =/= 0 -> - split_block(Is, [], [{bif,N,Fail,As,R}|make_block(Bl, Acc)]); -split_block([{set,[],[],{line,_}=Line}, - {set,[R],As,{bif,raise,{f,_}=Fail}}|Is], Bl, Acc) -> - split_block(Is, [], [{bif,raise,Fail,As,R},Line|make_block(Bl, Acc)]); -split_block([{set,[R],As,{alloc,Live,{gc_bif,N,{f,Lbl}=Fail}}}|Is], Bl, Acc) - when Lbl =/= 0 -> - split_block(Is, [], [{gc_bif,N,Fail,Live,As,R}|make_block(Bl, Acc)]); -split_block([{set,[D],[S|Puts],{alloc,R,{put_map,Op,{f,Lbl}=Fail}}}|Is], - Bl, Acc) when Lbl =/= 0 -> - split_block(Is, [], [{put_map,Fail,Op,S,D,R,{list,Puts}}| - make_block(Bl, Acc)]); -split_block([{set,[R],[],{try_catch,Op,L}}|Is], Bl, Acc) -> - split_block(Is, [], [{Op,R,L}|make_block(Bl, Acc)]); -split_block([I|Is], Bl, Acc) -> - split_block(Is, [I|Bl], Acc); -split_block([], Bl, Acc) -> make_block(Bl, Acc). - -make_block([], Acc) -> Acc; -make_block([{set,[D],Ss,{bif,Op,Fail}}|Bl]=Bl0, Acc) -> - %% If the last instruction in the block is a comparison or boolean operator - %% (such as '=:='), move it out of the block to facilitate further - %% optimizations. - Arity = length(Ss), - case erl_internal:comp_op(Op, Arity) orelse - erl_internal:new_type_test(Op, Arity) orelse - erl_internal:bool_op(Op, Arity) of - false -> - [{block,reverse(Bl0)}|Acc]; - true -> - I = {bif,Op,Fail,Ss,D}, - case Bl =:= [] of - true -> [I|Acc]; - false -> [I,{block,reverse(Bl)}|Acc] - end - end; -make_block([{set,[Dst],[Src],move}|Bl], Acc) -> - %% Make optimization of {move,Src,Dst}, {jump,...} possible. - I = {move,Src,Dst}, - case Bl =:= [] of - true -> [I|Acc]; - false -> [I,{block,reverse(Bl)}|Acc] - end; -make_block(Bl, Acc) -> [{block,reverse(Bl)}|Acc]. diff --git a/lib/compiler/src/beam_ssa.erl b/lib/compiler/src/beam_ssa.erl index a2766a0501..9c29c98064 100644 --- a/lib/compiler/src/beam_ssa.erl +++ b/lib/compiler/src/beam_ssa.erl @@ -20,26 +20,35 @@ %% Purpose: Type definitions and utilities for the SSA format. -module(beam_ssa). --export([add_anno/3,get_anno/2, - clobbers_xregs/1,def/2,def_used/2,dominators/1, +-export([add_anno/3,get_anno/2,get_anno/3, + clobbers_xregs/1,def/2,def_used/2, + definitions/1, + dominators/1, flatmapfold_instrs_rpo/4, fold_po/3,fold_po/4,fold_rpo/3,fold_rpo/4, fold_instrs_rpo/4, linearize/1, + mapfold_blocks_rpo/4, mapfold_instrs_rpo/4, + normalize/1, + no_side_effect/1, predecessors/1, rename_vars/3, rpo/1,rpo/2, split_blocks/3, successors/1,successors/2, - update_phi_labels/4,used/1]). + trim_unreachable/1, + update_phi_labels/4,used/1, + uses/1,uses/2]). -export_type([b_module/0,b_function/0,b_blk/0,b_set/0, b_ret/0,b_br/0,b_switch/0,terminator/0, b_var/0,b_literal/0,b_remote/0,b_local/0, value/0,argument/0,label/0, var_name/0,var_base/0,literal_value/0, - op/0,anno/0,block_map/0]). + op/0,anno/0,block_map/0,dominator_map/0, + rename_map/0,rename_proplist/0,usage_map/0, + definition_map/0]). -include("beam_ssa.hrl"). @@ -53,6 +62,9 @@ -type b_switch() :: #b_switch{}. -type terminator() :: b_br() | b_ret() | b_switch(). +-type construct() :: b_module() | b_function() | b_blk() | b_set() | + terminator(). + -type b_var() :: #b_var{}. -type b_literal() :: #b_literal{}. -type b_remote() :: #b_remote{}. @@ -73,6 +85,11 @@ -type anno() :: #{atom() := any()}. -type block_map() :: #{label():=b_blk()}. +-type dominator_map() :: #{label():=ordsets:ordset(label())}. +-type usage_map() :: #{b_var():=[{label(),b_set() | terminator()}]}. +-type definition_map() :: #{b_var():=b_set()}. +-type rename_map() :: #{b_var():=value()}. +-type rename_proplist() :: [{b_var(),value()}]. %% Note: By default, dialyzer will collapse this type to atom(). %% To avoid the collapsing, change the value of SET_LIMIT to 50 in the @@ -81,7 +98,7 @@ -type prim_op() :: 'bs_add' | 'bs_extract' | 'bs_init' | 'bs_init_writable' | 'bs_match' | 'bs_put' | 'bs_start_match' | 'bs_test_tail' | 'bs_utf16_size' | 'bs_utf8_size' | 'build_stacktrace' | - 'call' | 'catch_end' | 'context_to_binary' | + 'call' | 'catch_end' | 'extract' | 'get_hd' | 'get_map_element' | 'get_tl' | 'get_tuple_element' | 'has_map_field' | @@ -100,14 +117,14 @@ %% Primops only used internally during code generation. -type cg_prim_op() :: 'bs_get' | 'bs_match_string' | 'bs_restore' | 'bs_skip' | -'copy' | 'put_tuple_arity' | 'put_tuple_element'. + 'copy' | 'put_tuple_arity' | 'put_tuple_element'. --import(lists, [foldl/3,mapfoldl/3,reverse/1]). +-import(lists, [foldl/3,keyfind/3,mapfoldl/3,member/2,reverse/1]). -spec add_anno(Key, Value, Construct) -> Construct when Key :: atom(), Value :: any(), - Construct :: b_function() | b_blk() | b_set() | terminator(). + Construct :: construct(). add_anno(Key, Val, #b_function{anno=Anno}=Bl) -> Bl#b_function{anno=Anno#{Key=>Val}}; @@ -122,11 +139,17 @@ add_anno(Key, Val, #b_ret{anno=Anno}=Bl) -> add_anno(Key, Val, #b_switch{anno=Anno}=Bl) -> Bl#b_switch{anno=Anno#{Key=>Val}}. --spec get_anno(atom(), b_blk()|b_set()|terminator()) -> any(). +-spec get_anno(atom(), construct()) -> any(). get_anno(Key, Construct) -> - maps:get(Key, get_anno(Construct)). + map_get(Key, get_anno(Construct)). + +-spec get_anno(atom(), construct(),any()) -> any(). + +get_anno(Key, Construct, Default) -> + maps:get(Key, get_anno(Construct), Default). +get_anno(#b_function{anno=Anno}) -> Anno; get_anno(#b_blk{anno=Anno}) -> Anno; get_anno(#b_set{anno=Anno}) -> Anno; get_anno(#b_br{anno=Anno}) -> Anno; @@ -150,6 +173,40 @@ clobbers_xregs(#b_set{op=Op}) -> _ -> false end. +%% no_side_effect(#b_set{}) -> true|false. +%% Test whether this instruction has no side effect and thus is safe +%% not to execute if its value is not used. Note that even if `true` +%% is returned, the instruction could still be impure (e.g. bif:get). + +-spec no_side_effect(b_set()) -> boolean(). + +no_side_effect(#b_set{op=Op}) -> + case Op of + {bif,_} -> true; + {float,get} -> true; + bs_init -> true; + bs_extract -> true; + bs_match -> true; + bs_start_match -> true; + bs_test_tail -> true; + bs_get_tail -> true; + bs_put -> true; + extract -> true; + get_hd -> true; + get_tl -> true; + get_map_element -> true; + get_tuple_element -> true; + has_map_field -> true; + is_nonempty_list -> true; + is_tagged_tuple -> true; + make_fun -> true; + put_map -> true; + put_list -> true; + put_tuple -> true; + succeeded -> true; + _ -> false + end. + -spec predecessors(Blocks) -> #{BlockNumber:=[Predecessor]} when Blocks :: block_map(), BlockNumber :: label(), @@ -180,10 +237,73 @@ successors(#b_blk{last=Terminator}) -> [] end. +%% normalize(Instr0) -> Instr. +%% Normalize instructions to help optimizations. +%% +%% For commutative operators (such as '+' and 'or'), always +%% place a variable operand before a literal operand. +%% +%% Normalize #b_br{} to one of the following forms: +%% +%% #b_br{b_literal{val=true},succ=Label,fail=Label} +%% #b_br{b_var{},succ=Label1,fail=Label2} where Label1 =/= Label2 +%% +%% Simplify a #b_switch{} with a literal argument to a #b_br{}. +%% +%% Simplify a #b_switch{} with a variable argument and an empty +%% switch list to a #b_br{}. + +-spec normalize(b_set() | terminator()) -> + b_set() | terminator(). + +normalize(#b_set{op={bif,Bif},args=Args}=Set) -> + case {is_commutative(Bif),Args} of + {false,_} -> + Set; + {true,[#b_literal{}=Lit,#b_var{}=Var]} -> + Set#b_set{args=[Var,Lit]}; + {true,_} -> + Set + end; +normalize(#b_set{}=Set) -> + Set; +normalize(#b_br{}=Br) -> + case Br of + #b_br{bool=Bool,succ=Same,fail=Same} -> + case Bool of + #b_literal{val=true} -> + Br; + _ -> + Br#b_br{bool=#b_literal{val=true}} + end; + #b_br{bool=#b_literal{val=true},succ=Succ} -> + Br#b_br{fail=Succ}; + #b_br{bool=#b_literal{val=false},fail=Fail} -> + Br#b_br{bool=#b_literal{val=true},succ=Fail}; + #b_br{} -> + Br + end; +normalize(#b_switch{arg=Arg,fail=Fail,list=List}=Sw) -> + case Arg of + #b_literal{} -> + case keyfind(Arg, 1, List) of + false -> + #b_br{bool=#b_literal{val=true},succ=Fail,fail=Fail}; + {Arg,L} -> + #b_br{bool=#b_literal{val=true},succ=L,fail=L} + end; + #b_var{} when List =:= [] -> + #b_br{bool=#b_literal{val=true},succ=Fail,fail=Fail}; + #b_var{} -> + Sw + end; +normalize(#b_ret{}=Ret) -> + Ret. + -spec successors(label(), block_map()) -> [label()]. successors(L, Blocks) -> - successors(maps:get(L, Blocks)). + successors(map_get(L, Blocks)). -spec def(Ls, Blocks) -> Def when Ls :: [label()], @@ -192,7 +312,7 @@ successors(L, Blocks) -> def(Ls, Blocks) -> Top = rpo(Ls, Blocks), - Blks = [maps:get(L, Blocks) || L <- Top], + Blks = [map_get(L, Blocks) || L <- Top], def_1(Blks, []). -spec def_used(Ls, Blocks) -> {Def,Used} when @@ -203,18 +323,18 @@ def(Ls, Blocks) -> def_used(Ls, Blocks) -> Top = rpo(Ls, Blocks), - Blks = [maps:get(L, Blocks) || L <- Top], - Preds = gb_sets:from_list(Top), - def_used_1(Blks, Preds, [], gb_sets:empty()). + Blks = [map_get(L, Blocks) || L <- Top], + Preds = cerl_sets:from_list(Top), + def_used_1(Blks, Preds, [], []). -spec dominators(Blocks) -> Result when Blocks :: block_map(), - Result :: #{label():=ordsets:ordset(label())}. + Result :: dominator_map(). dominators(Blocks) -> Preds = predecessors(Blocks), Top0 = rpo(Blocks), - Top = [{L,maps:get(L, Preds)} || L <- Top0], + Top = [{L,map_get(L, Preds)} || L <- Top0], %% The flow graph for an Erlang function is reducible, and %% therefore one traversal in reverse postorder is sufficient. @@ -230,6 +350,26 @@ fold_instrs_rpo(Fun, From, Acc0, Blocks) -> Top = rpo(From, Blocks), fold_instrs_rpo_1(Top, Fun, Blocks, Acc0). +%% Like mapfold_instrs_rpo but at the block level to support lookahead and +%% scope-dependent transformations. +-spec mapfold_blocks_rpo(Fun, From, Acc, Blocks) -> Result when + Fun :: fun((label(), b_blk(), any()) -> {b_blk(), any()}), + From :: [label()], + Acc :: any(), + Blocks :: block_map(), + Result :: {block_map(), any()}. +mapfold_blocks_rpo(Fun, From, Acc, Blocks) -> + Successors = rpo(From, Blocks), + foldl(fun(Lbl, A) -> + mapfold_blocks_rpo_1(Fun, Lbl, A) + end, {Blocks, Acc}, Successors). + +mapfold_blocks_rpo_1(Fun, Lbl, {Blocks0, Acc0}) -> + Block0 = map_get(Lbl, Blocks0), + {Block, Acc} = Fun(Lbl, Block0, Acc0), + Blocks = Blocks0#{Lbl:=Block}, + {Blocks, Acc}. + -spec mapfold_instrs_rpo(Fun, From, Acc0, Blocks0) -> {Blocks,Acc} when Fun :: fun((b_blk()|terminator(), any()) -> any()), From :: [label()], @@ -312,14 +452,19 @@ fold_po(Fun, From, Acc0, Blocks) -> %% linearize(Blocks) -> [{BlockLabel,#b_blk{}}]. %% Linearize the intermediate representation of the code. +%% Unreachable blocks will be discarded, and phi nodes will +%% be adjusted so that they no longer refers to discarded +%% blocks or to blocks that no longer are predecessors of +%% the phi node block. -spec linearize(Blocks) -> Linear when Blocks :: block_map(), Linear :: [{label(),b_blk()}]. linearize(Blocks) -> - Seen = gb_sets:empty(), - {Linear,_} = linearize_1([0], Blocks, Seen, []), + Seen = cerl_sets:new(), + {Linear0,_} = linearize_1([0], Blocks, Seen, []), + Linear = fix_phis(Linear0, #{}), Linear. -spec rpo(Blocks) -> [Label] when @@ -335,18 +480,18 @@ rpo(Blocks) -> Labels :: [label()]. rpo(From, Blocks) -> - Seen = gb_sets:empty(), + Seen = cerl_sets:new(), {Ls,_} = rpo_1(From, Blocks, Seen, []), Ls. -spec rename_vars(Rename, [label()], block_map()) -> block_map() when - Rename :: [{var_name(),value()}] | #{var_name():=value()}. + Rename :: rename_map() | rename_proplist(). rename_vars(Rename, From, Blocks) when is_list(Rename) -> rename_vars(maps:from_list(Rename), From, Blocks); rename_vars(Rename, From, Blocks) when is_map(Rename)-> Top = rpo(From, Blocks), - Preds = gb_sets:from_list(Top), + Preds = cerl_sets:from_list(Top), F = fun(#b_set{op=phi,args=Args0}=Set) -> Args = rename_phi_vars(Args0, Preds, Rename), Set#b_set{args=Args}; @@ -379,6 +524,19 @@ split_blocks(P, Blocks, Count) -> Ls = beam_ssa:rpo(Blocks), split_blocks_1(Ls, P, Blocks, Count). +-spec trim_unreachable(Blocks0) -> Blocks when + Blocks0 :: block_map(), + Blocks :: block_map(). + +%% trim_unreachable(Blocks0) -> Blocks. +%% Remove all unreachable blocks. Adjust all phi nodes so +%% they don't refer to blocks that has been removed or no +%% no longer branch to the phi node in question. + +trim_unreachable(Blocks) -> + %% Could perhaps be optimized if there is any need. + maps:from_list(linearize(Blocks)). + %% update_phi_labels([BlockLabel], Old, New, Blocks0) -> Blocks. %% In the given blocks, replace label Old in with New in all %% phi nodes. This is useful after merging or splitting @@ -408,42 +566,83 @@ update_phi_labels([], _, _, Blocks) -> Blocks. used(#b_blk{is=Is,last=Last}) -> used_1([Last|Is], ordsets:new()); -used(#b_br{bool=#b_var{name=V}}) -> +used(#b_br{bool=#b_var{}=V}) -> [V]; -used(#b_ret{arg=#b_var{name=V}}) -> +used(#b_ret{arg=#b_var{}=V}) -> [V]; used(#b_set{op=phi,args=Args}) -> - ordsets:from_list([V || {#b_var{name=V},_} <- Args]); + ordsets:from_list([V || {#b_var{}=V,_} <- Args]); used(#b_set{args=Args}) -> ordsets:from_list(used_args(Args)); -used(#b_switch{arg=#b_var{name=V}}) -> +used(#b_switch{arg=#b_var{}=V}) -> [V]; used(_) -> []. +-spec definitions(Blocks :: block_map()) -> definition_map(). +definitions(Blocks) -> + fold_instrs_rpo(fun(#b_set{ dst = Var }=I, Acc) -> + Acc#{Var => I}; + (_Terminator, Acc) -> + Acc + end, [0], #{}, Blocks). + +-spec uses(Blocks :: block_map()) -> usage_map(). +uses(Blocks) -> + uses([0], Blocks). + +-spec uses(From, Blocks) -> usage_map() when + From :: [label()], + Blocks :: block_map(). +uses(From, Blocks) -> + fold_rpo(fun fold_uses_block/3, From, #{}, Blocks). + +fold_uses_block(Lbl, #b_blk{is=Is,last=Last}, UseMap0) -> + F = fun(I, UseMap) -> + foldl(fun(Var, Acc) -> + Uses0 = maps:get(Var, Acc, []), + Uses = [{Lbl, I} | Uses0], + maps:put(Var, Uses, Acc) + end, UseMap, used(I)) + end, + F(Last, foldl(F, UseMap0, Is)). + %%% %%% Internal functions. %%% +is_commutative('and') -> true; +is_commutative('or') -> true; +is_commutative('xor') -> true; +is_commutative('band') -> true; +is_commutative('bor') -> true; +is_commutative('bxor') -> true; +is_commutative('+') -> true; +is_commutative('*') -> true; +is_commutative('=:=') -> true; +is_commutative('==') -> true; +is_commutative('=/=') -> true; +is_commutative('/=') -> true; +is_commutative(_) -> false. + def_used_1([#b_blk{is=Is,last=Last}|Bs], Preds, Def0, Used0) -> {Def,Used1} = def_used_is(Is, Preds, Def0, Used0), - Used = gb_sets:union(gb_sets:from_list(used(Last)), Used1), + Used = ordsets:union(used(Last), Used1), def_used_1(Bs, Preds, Def, Used); def_used_1([], _Preds, Def, Used) -> - {ordsets:from_list(Def),gb_sets:to_list(Used)}. + {ordsets:from_list(Def),Used}. -def_used_is([#b_set{op=phi,dst=#b_var{name=Dst},args=Args}|Is], +def_used_is([#b_set{op=phi,dst=Dst,args=Args}|Is], Preds, Def0, Used0) -> Def = [Dst|Def0], %% We must be careful to only include variables that will %% be used when arriving from one of the predecessor blocks %% in Preds. - Used1 = [V || {#b_var{name=V},L} <- Args, - gb_sets:is_member(L, Preds)], - Used = gb_sets:union(gb_sets:from_list(Used1), Used0), + Used1 = [V || {#b_var{}=V,L} <- Args, cerl_sets:is_element(L, Preds)], + Used = ordsets:union(ordsets:from_list(Used1), Used0), def_used_is(Is, Preds, Def, Used); -def_used_is([#b_set{dst=#b_var{name=Dst}}=I|Is], Preds, Def0, Used0) -> +def_used_is([#b_set{dst=Dst}=I|Is], Preds, Def0, Used0) -> Def = [Dst|Def0], - Used = gb_sets:union(gb_sets:from_list(used(I)), Used0), + Used = ordsets:union(used(I), Used0), def_used_is(Is, Preds, Def, Used); def_used_is([], _Preds, Def, Used) -> {Def,Used}. @@ -454,7 +653,7 @@ def_1([#b_blk{is=Is}|Bs], Def0) -> def_1([], Def) -> ordsets:from_list(Def). -def_is([#b_set{dst=#b_var{name=Dst}}|Is], Def) -> +def_is([#b_set{dst=Dst}|Is], Def) -> def_is(Is, [Dst|Def]); def_is([], Def) -> Def. @@ -462,51 +661,51 @@ iter_dominators([{0,[]}|Ls], _Doms) -> Dom = [0], iter_dominators(Ls, #{0=>Dom}); iter_dominators([{L,Preds}|Ls], Doms) -> - DomPreds = [maps:get(P, Doms) || P <- Preds, maps:is_key(P, Doms)], + DomPreds = [map_get(P, Doms) || P <- Preds, is_map_key(P, Doms)], Dom = ordsets:add_element(L, ordsets:intersection(DomPreds)), iter_dominators(Ls, Doms#{L=>Dom}); iter_dominators([], Doms) -> Doms. fold_rpo_1([L|Ls], Fun, Blocks, Acc0) -> - Block = maps:get(L, Blocks), + Block = map_get(L, Blocks), Acc = Fun(L, Block, Acc0), fold_rpo_1(Ls, Fun, Blocks, Acc); fold_rpo_1([], _, _, Acc) -> Acc. fold_instrs_rpo_1([L|Ls], Fun, Blocks, Acc0) -> - #b_blk{is=Is,last=Last} = maps:get(L, Blocks), + #b_blk{is=Is,last=Last} = map_get(L, Blocks), Acc1 = foldl(Fun, Acc0, Is), Acc = Fun(Last, Acc1), fold_instrs_rpo_1(Ls, Fun, Blocks, Acc); fold_instrs_rpo_1([], _, _, Acc) -> Acc. mapfold_instrs_rpo_1([L|Ls], Fun, Blocks0, Acc0) -> - #b_blk{is=Is0,last=Last0} = Block0 = maps:get(L, Blocks0), + #b_blk{is=Is0,last=Last0} = Block0 = map_get(L, Blocks0), {Is,Acc1} = mapfoldl(Fun, Acc0, Is0), {Last,Acc} = Fun(Last0, Acc1), Block = Block0#b_blk{is=Is,last=Last}, - Blocks = maps:put(L, Block, Blocks0), + Blocks = Blocks0#{L:=Block}, mapfold_instrs_rpo_1(Ls, Fun, Blocks, Acc); mapfold_instrs_rpo_1([], _, Blocks, Acc) -> {Blocks,Acc}. flatmapfold_instrs_rpo_1([L|Ls], Fun, Blocks0, Acc0) -> - #b_blk{is=Is0,last=Last0} = Block0 = maps:get(L, Blocks0), + #b_blk{is=Is0,last=Last0} = Block0 = map_get(L, Blocks0), {Is,Acc1} = flatmapfoldl(Fun, Acc0, Is0), {[Last],Acc} = Fun(Last0, Acc1), Block = Block0#b_blk{is=Is,last=Last}, - Blocks = maps:put(L, Block, Blocks0), + Blocks = Blocks0#{L:=Block}, flatmapfold_instrs_rpo_1(Ls, Fun, Blocks, Acc); flatmapfold_instrs_rpo_1([], _, Blocks, Acc) -> {Blocks,Acc}. linearize_1([L|Ls], Blocks, Seen0, Acc0) -> - case gb_sets:is_member(L, Seen0) of + case cerl_sets:is_element(L, Seen0) of true -> linearize_1(Ls, Blocks, Seen0, Acc0); false -> - Seen1 = gb_sets:insert(L, Seen0), - Block = maps:get(L, Blocks), + Seen1 = cerl_sets:add_element(L, Seen0), + Block = map_get(L, Blocks), Successors = successors(Block), {Acc,Seen} = linearize_1(Successors, Blocks, Seen1, Acc0), linearize_1(Ls, Blocks, Seen, [{L,Block}|Acc]) @@ -514,13 +713,40 @@ linearize_1([L|Ls], Blocks, Seen0, Acc0) -> linearize_1([], _, Seen, Acc) -> {Acc,Seen}. +fix_phis([{L,Blk0}|Bs], S) -> + Blk = case Blk0 of + #b_blk{is=[#b_set{op=phi}|_]=Is0} -> + Is = fix_phis_1(Is0, L, S), + Blk0#b_blk{is=Is}; + #b_blk{} -> + Blk0 + end, + Successors = successors(Blk), + [{L,Blk}|fix_phis(Bs, S#{L=>Successors})]; +fix_phis([], _) -> []. + +fix_phis_1([#b_set{op=phi,args=Args0}=I|Is], L, S) -> + Args = [{Val,Pred} || {Val,Pred} <- Args0, + is_successor(L, Pred, S)], + [I#b_set{args=Args}|fix_phis_1(Is, L, S)]; +fix_phis_1(Is, _, _) -> Is. + +is_successor(L, Pred, S) -> + case S of + #{Pred:=Successors} -> + member(L, Successors); + #{} -> + %% This block has been removed. + false + end. + rpo_1([L|Ls], Blocks, Seen0, Acc0) -> - case gb_sets:is_member(L, Seen0) of + case cerl_sets:is_element(L, Seen0) of true -> rpo_1(Ls, Blocks, Seen0, Acc0); false -> - Block = maps:get(L, Blocks), - Seen1 = gb_sets:insert(L, Seen0), + Block = map_get(L, Blocks), + Seen1 = cerl_sets:add_element(L, Seen0), Successors = successors(Block), {Acc,Seen} = rpo_1(Successors, Blocks, Seen1, Acc0), rpo_1(Ls, Blocks, Seen, [L|Acc]) @@ -528,10 +754,10 @@ rpo_1([L|Ls], Blocks, Seen0, Acc0) -> rpo_1([], _, Seen, Acc) -> {Acc,Seen}. -rename_var(#b_var{name=Old}=V, Rename) -> +rename_var(#b_var{}=Old, Rename) -> case Rename of #{Old:=New} -> New; - #{} -> V + #{} -> Old end; rename_var(#b_remote{mod=Mod0,name=Name0}=Remote, Rename) -> Mod = rename_var(Mod0, Rename), @@ -540,7 +766,7 @@ rename_var(#b_remote{mod=Mod0,name=Name0}=Remote, Rename) -> rename_var(Old, _) -> Old. rename_phi_vars([{Var,L}|As], Preds, Ren) -> - case gb_sets:is_member(L, Preds) of + case cerl_sets:is_element(L, Preds) of true -> [{rename_var(Var, Ren),L}|rename_phi_vars(As, Preds, Ren)]; false -> @@ -549,11 +775,11 @@ rename_phi_vars([{Var,L}|As], Preds, Ren) -> rename_phi_vars([], _, _) -> []. map_instrs_1([L|Ls], Fun, Blocks0) -> - #b_blk{is=Is0,last=Last0} = Blk0 = maps:get(L, Blocks0), + #b_blk{is=Is0,last=Last0} = Blk0 = map_get(L, Blocks0), Is = [Fun(I) || I <- Is0], Last = Fun(Last0), Blk = Blk0#b_blk{is=Is,last=Last}, - Blocks = maps:put(L, Blk, Blocks0), + Blocks = Blocks0#{L:=Blk}, map_instrs_1(Ls, Fun, Blocks); map_instrs_1([], _, Blocks) -> Blocks. @@ -564,7 +790,7 @@ flatmapfoldl(F, Accu0, [Hd|Tail]) -> flatmapfoldl(_, Accu, []) -> {[],Accu}. split_blocks_1([L|Ls], P, Blocks0, Count0) -> - #b_blk{is=Is0} = Blk = maps:get(L, Blocks0), + #b_blk{is=Is0} = Blk = map_get(L, Blocks0), case split_blocks_is(Is0, P, []) of {yes,Bef,Aft} -> NewLbl = Count0, @@ -573,8 +799,8 @@ split_blocks_1([L|Ls], P, Blocks0, Count0) -> BefBlk = Blk#b_blk{is=Bef,last=Br}, NewBlk = Blk#b_blk{is=Aft}, Blocks1 = Blocks0#{L:=BefBlk,NewLbl=>NewBlk}, - Successors = beam_ssa:successors(NewBlk), - Blocks = beam_ssa:update_phi_labels(Successors, L, NewLbl, Blocks1), + Successors = successors(NewBlk), + Blocks = update_phi_labels(Successors, L, NewLbl, Blocks1), split_blocks_1([NewLbl|Ls], P, Blocks, Count); no -> split_blocks_1(Ls, P, Blocks0, Count0) @@ -602,7 +828,7 @@ update_phi_labels_is(Is, _, _) -> Is. rename_label(Old, Old, New) -> New; rename_label(Lbl, _Old, _New) -> Lbl. -used_args([#b_var{name=V}|As]) -> +used_args([#b_var{}=V|As]) -> [V|used_args(As)]; used_args([#b_remote{mod=Mod,name=Name}|As]) -> used_args([Mod,Name|As]); diff --git a/lib/compiler/src/beam_ssa_bsm.erl b/lib/compiler/src/beam_ssa_bsm.erl new file mode 100644 index 0000000000..466337db0e --- /dev/null +++ b/lib/compiler/src/beam_ssa_bsm.erl @@ -0,0 +1,1044 @@ +%% +%% %CopyrightBegin% +%% +%% Copyright Ericsson AB 2018. All Rights Reserved. +%% +%% Licensed under the Apache License, Version 2.0 (the "License"); +%% you may not use this file except in compliance with the License. +%% You may obtain a copy of the License at +%% +%% http://www.apache.org/licenses/LICENSE-2.0 +%% +%% Unless required by applicable law or agreed to in writing, software +%% distributed under the License is distributed on an "AS IS" BASIS, +%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +%% See the License for the specific language governing permissions and +%% limitations under the License. +%% +%% %CopyrightEnd% +%% + +%%% +%%% This pass optimizes bit syntax matching, and is centered around the concept +%%% of "match context reuse" which is best explained through example. To put it +%%% shortly we attempt to turn this: +%%% +%%% <<0,B/bits>> = A, +%%% <<1,C/bits>> = B, +%%% <<D,_/bits>> = C, +%%% D. +%%% +%%% ... Into this: +%%% +%%% <<0,1,D,_/bits>>=A, +%%% D. +%%% +%%% Which is much faster as it avoids the creation of intermediate terms. This +%%% is especially noticeable in loops where such garbage is generated on each +%%% iteration. +%%% +%%% The optimization itself is very simple and can be applied whenever there's +%%% matching on the tail end of a binary; instead of creating a new binary and +%%% starting a new match context on it, we reuse the match context used to +%%% extract the tail and avoid the creation of both objects. +%%% +%%% The catch is that a match context isn't a proper type and nothing outside +%%% of bit syntax match operations can handle them. We therefore need to make +%%% sure that they never "leak" into other instructions, and most of the pass +%%% revolves around getting around that limitation. +%%% +%%% Unlike most other passes we look at the whole module so we can combine +%%% matches across function boundaries, greatly increasing the performance of +%%% complex matches and loops. +%%% + +-module(beam_ssa_bsm). + +-export([module/2, format_error/1]). + +-include("beam_ssa.hrl"). + +-import(lists, [member/2, reverse/1, splitwith/2, map/2, foldl/3, mapfoldl/3, + nth/2, max/1, unzip/1]). + +-spec format_error(term()) -> nonempty_string(). + +format_error(OptInfo) -> + format_opt_info(OptInfo). + +-spec module(Module, Options) -> Result when + Module :: beam_ssa:b_module(), + Options :: [compile:option()], + Result :: {ok, beam_ssa:b_module(), list()}. + +-define(PASS(N), {N,fun N/1}). + +module(#b_module{body=Fs0}=Module, Opts) -> + ModInfo = analyze_module(Module), + + %% combine_matches is repeated after accept_context_args as the control + %% flow changes can enable further optimizations, as in the example below: + %% + %% a(<<0,X/binary>>) -> a(X); + %% a(A) when bit_size(A) =:= 52 -> bar; + %% a(<<1,X/binary>>) -> X. %% Match context will be reused here when + %% %% when repeated. + + {Fs, _} = compile:run_sub_passes( + [?PASS(combine_matches), + ?PASS(accept_context_args), + ?PASS(combine_matches), + ?PASS(allow_context_passthrough), + ?PASS(skip_outgoing_tail_extraction), + ?PASS(annotate_context_parameters)], + {Fs0, ModInfo}), + + Ws = case proplists:get_bool(bin_opt_info, Opts) of + true -> collect_opt_info(Fs); + false -> [] + end, + + {ok, Module#b_module{body=Fs}, Ws}. + +-type module_info() :: #{ func_id() => func_info() }. + +-type func_id() :: {Name :: atom(), Arity :: non_neg_integer()}. + +-type func_info() :: #{ has_bsm_ops => boolean(), + parameters => [#b_var{}], + parameter_info => #{ #b_var{} => param_info() } }. + +-type param_info() :: suitable_for_reuse | + {Problem :: atom(), Where :: term()}. + +-spec analyze_module(#b_module{}) -> module_info(). + +analyze_module(#b_module{body=Fs}) -> + foldl(fun(#b_function{args=Parameters}=F, I) -> + FuncInfo = #{ has_bsm_ops => has_bsm_ops(F), + parameters => Parameters, + parameter_info => #{} }, + FuncId = get_fa(F), + I#{ FuncId => FuncInfo } + end, #{}, Fs). + +has_bsm_ops(#b_function{bs=Blocks}) -> + hbo_blocks(maps:to_list(Blocks)). + +hbo_blocks([{_,#b_blk{is=Is}} | Blocks]) -> + case hbo_is(Is) of + false -> hbo_blocks(Blocks); + true -> true + end; +hbo_blocks([]) -> + false. + +hbo_is([#b_set{op=bs_start_match} | _]) -> true; +hbo_is([_I | Is]) -> hbo_is(Is); +hbo_is([]) -> false. + +%% Checks whether it's legal to make a call with the given argument as a match +%% context, returning the param_info() of the relevant parameter. +-spec check_context_call(#b_set{}, Arg, CtxChain, ModInfo) -> param_info() when + Arg :: #b_var{}, + CtxChain :: [#b_var{}], + ModInfo :: module_info(). +check_context_call(#b_set{args=Args}, Arg, CtxChain, ModInfo) -> + Aliases = [Arg | CtxChain], + ccc_1(Args, Arg, Aliases, ModInfo). + +ccc_1([#b_local{}=Call | Args], Ctx, Aliases, ModInfo) -> + %% Matching operations assume that their context isn't aliased (as in + %% pointer aliasing), so we must reject calls whose arguments contain more + %% than one reference to the context. + %% + %% TODO: Try to fall back to passing binaries in these cases. Partial reuse + %% is better than nothing. + UseCount = foldl(fun(Arg, C) -> + case member(Arg, Aliases) of + true -> C + 1; + false -> C + end + end, 0, Args), + if + UseCount =:= 1 -> + #b_local{name=#b_literal{val=Name},arity=Arity} = Call, + Callee = {Name, Arity}, + + ParamInfo = funcinfo_get(Callee, parameter_info, ModInfo), + Parameters = funcinfo_get(Callee, parameters, ModInfo), + Parameter = nth(1 + arg_index(Ctx, Args), Parameters), + + case maps:find(Parameter, ParamInfo) of + {ok, suitable_for_reuse} -> + suitable_for_reuse; + {ok, Other} -> + {unsuitable_call, {Call, Other}}; + error -> + {no_match_on_entry, Call} + end; + UseCount > 1 -> + {multiple_uses_in_call, Call} + end; +ccc_1([#b_remote{}=Call | _Args], _Ctx, _CtxChain, _ModInfo) -> + {remote_call, Call}; +ccc_1([Fun | _Args], _Ctx, _CtxChain, _ModInfo) -> + %% TODO: It may be possible to support this in the future for locally + %% defined funs, including ones with free variables. + {fun_call, Fun}. + +%% Returns the index of Var in Args. +arg_index(Var, Args) -> arg_index_1(Var, Args, 0). + +arg_index_1(Var, [Var | _Args], Index) -> Index; +arg_index_1(Var, [_Arg | Args], Index) -> arg_index_1(Var, Args, Index + 1). + +is_tail_binary(#b_set{op=bs_match,args=[#b_literal{val=binary} | Rest]}) -> + member(#b_literal{val=all}, Rest); +is_tail_binary(#b_set{op=bs_get_tail}) -> + true; +is_tail_binary(_) -> + false. + +is_tail_binary(#b_var{}=Var, Defs) -> + case find_match_definition(Var, Defs) of + {ok, Def} -> is_tail_binary(Def); + _ -> false + end; +is_tail_binary(_Literal, _Defs) -> + false. + +assert_match_context(#b_var{}=Var, Defs) -> + case maps:find(Var, Defs) of + {ok, #b_set{op=bs_match,args=[_,#b_var{}=Ctx|_]}} -> + assert_match_context(Ctx, Defs); + {ok, #b_set{op=bs_start_match}} -> + ok + end. + +find_match_definition(#b_var{}=Var, Defs) -> + case maps:find(Var, Defs) of + {ok, #b_set{op=bs_extract,args=[Ctx]}} -> maps:find(Ctx, Defs); + {ok, #b_set{op=bs_get_tail}=Def} -> {ok, Def}; + _ -> error + end. + +%% Returns a list of all contexts that were used to extract Var. +context_chain_of(#b_var{}=Var, Defs) -> + case maps:find(Var, Defs) of + {ok, #b_set{op=bs_match,args=[_,#b_var{}=Ctx|_]}} -> + [Ctx | context_chain_of(Ctx, Defs)]; + {ok, #b_set{op=bs_get_tail,args=[Ctx]}} -> + [Ctx | context_chain_of(Ctx, Defs)]; + {ok, #b_set{op=bs_extract,args=[Ctx]}} -> + [Ctx | context_chain_of(Ctx, Defs)]; + _ -> + [] + end. + +%% Grabs the match context used to produce the given variable. +match_context_of(#b_var{}=Var, Defs) -> + Ctx = match_context_of_1(Var, Defs), + assert_match_context(Ctx, Defs), + Ctx. + +match_context_of_1(Var, Defs) -> + case maps:get(Var, Defs) of + #b_set{op=bs_extract,args=[#b_var{}=Ctx0]} -> + #b_set{op=bs_match, + args=[_,#b_var{}=Ctx|_]} = maps:get(Ctx0, Defs), + Ctx; + #b_set{op=bs_get_tail,args=[#b_var{}=Ctx]} -> + Ctx + end. + +funcinfo_get(#b_function{}=F, Attribute, ModInfo) -> + funcinfo_get(get_fa(F), Attribute, ModInfo); +funcinfo_get({_,_}=Key, Attribute, ModInfo) -> + FuncInfo = maps:get(Key, ModInfo), + maps:get(Attribute, FuncInfo). + +funcinfo_set(#b_function{}=F, Attribute, Value, ModInfo) -> + funcinfo_set(get_fa(F), Attribute, Value, ModInfo); +funcinfo_set(Key, Attribute, Value, ModInfo) -> + FuncInfo = maps:put(Attribute, Value, maps:get(Key, ModInfo, #{})), + maps:put(Key, FuncInfo, ModInfo). + +get_fa(#b_function{ anno = Anno }) -> + {_,Name,Arity} = maps:get(func_info, Anno), + {Name,Arity}. + +%% Replaces matched-out binaries with aliases that are lazily converted to +%% binary form when used, allowing us to keep the "match path" free of binary +%% creation. + +-spec alias_matched_binaries(Blocks, Counter, AliasMap) -> Result when + Blocks :: beam_ssa:block_map(), + Counter :: non_neg_integer(), + AliasMap :: match_alias_map(), + Result :: {Blocks, Counter}. + +-type match_alias_map() :: + #{ Binary :: #b_var{} => + { %% Replace all uses of Binary with an alias after this + %% label. + AliasAfter :: beam_ssa:label(), + %% The match context whose tail is equal to Binary. + Context :: #b_var{} } }. + +%% Keeps track of the promotions we need to insert. They're partially keyed by +%% location because they may not be valid on all execution paths and we may +%% need to add redundant promotions in some cases. +-type promotion_map() :: + #{ { PromoteAt :: beam_ssa:label(), + Variable :: #b_var{} } => + Instruction :: #b_set{} }. + +-record(amb, { dominators :: beam_ssa:dominator_map(), + match_aliases :: match_alias_map(), + cnt :: non_neg_integer(), + promotions = #{} :: promotion_map() }). + +alias_matched_binaries(Blocks0, Counter, AliasMap) when AliasMap =/= #{} -> + State0 = #amb{ dominators = beam_ssa:dominators(Blocks0), + match_aliases = AliasMap, + cnt = Counter }, + {Blocks, State} = beam_ssa:mapfold_blocks_rpo(fun amb_1/3, [0], State0, + Blocks0), + {amb_insert_promotions(Blocks, State), State#amb.cnt}; +alias_matched_binaries(Blocks, Counter, _AliasMap) -> + {Blocks, Counter}. + +amb_1(Lbl, #b_blk{is=Is0,last=Last0}=Block, State0) -> + {Is, State1} = mapfoldl(fun(I, State) -> + amb_assign_set(I, Lbl, State) + end, State0, Is0), + {Last, State} = amb_assign_last(Last0, Lbl, State1), + {Block#b_blk{is=Is,last=Last}, State}. + +amb_assign_set(#b_set{op=phi,args=Args0}=I, _Lbl, State0) -> + %% Phi node aliases are relative to their source block, not their + %% containing block. + {Args, State} = + mapfoldl(fun({Arg0, Lbl}, Acc) -> + {Arg, State} = amb_get_alias(Arg0, Lbl, Acc), + {{Arg, Lbl}, State} + end, State0, Args0), + {I#b_set{args=Args}, State}; +amb_assign_set(#b_set{args=Args0}=I, Lbl, State0) -> + {Args, State} = mapfoldl(fun(Arg0, Acc) -> + amb_get_alias(Arg0, Lbl, Acc) + end, State0, Args0), + {I#b_set{args=Args}, State}. + +amb_assign_last(#b_ret{arg=Arg0}=T, Lbl, State0) -> + {Arg, State} = amb_get_alias(Arg0, Lbl, State0), + {T#b_ret{arg=Arg}, State}; +amb_assign_last(#b_switch{arg=Arg0}=T, Lbl, State0) -> + {Arg, State} = amb_get_alias(Arg0, Lbl, State0), + {T#b_switch{arg=Arg}, State}; +amb_assign_last(#b_br{bool=Arg0}=T, Lbl, State0) -> + {Arg, State} = amb_get_alias(Arg0, Lbl, State0), + {T#b_br{bool=Arg}, State}. + +amb_get_alias(#b_var{}=Arg, Lbl, State) -> + case maps:find(Arg, State#amb.match_aliases) of + {ok, {AliasAfter, Context}} -> + %% Our context may not have been created yet, so we skip assigning + %% an alias unless the given block is among our dominators. + Dominators = maps:get(Lbl, State#amb.dominators), + case ordsets:is_element(AliasAfter, Dominators) of + true -> amb_create_alias(Arg, Context, Lbl, State); + false -> {Arg, State} + end; + error -> + {Arg, State} + end; +amb_get_alias(#b_remote{mod=Mod0,name=Name0}=Arg0, Lbl, State0) -> + {Mod, State1} = amb_get_alias(Mod0, Lbl, State0), + {Name, State} = amb_get_alias(Name0, Lbl, State1), + Arg = Arg0#b_remote{mod=Mod,name=Name}, + {Arg, State}; +amb_get_alias(Arg, _Lbl, State) -> + {Arg, State}. + +amb_create_alias(#b_var{}=Arg0, Context, Lbl, State0) -> + Dominators = maps:get(Lbl, State0#amb.dominators), + Promotions0 = State0#amb.promotions, + + PrevPromotions = + [maps:get({Dom, Arg0}, Promotions0) + || Dom <- Dominators, is_map_key({Dom, Arg0}, Promotions0)], + + case PrevPromotions of + [_|_] -> + %% We've already created an alias prior to this block, so we'll + %% grab the most recent one to minimize stack use. + + #b_set{dst=Alias} = max(PrevPromotions), + {Alias, State0}; + [] -> + %% If we haven't created an alias we need to do so now. The + %% promotion will be inserted later by amb_insert_promotions/2. + + Counter = State0#amb.cnt, + Alias = #b_var{name={'@ssa_bsm_alias', Counter}}, + Promotion = #b_set{op=bs_get_tail,dst=Alias,args=[Context]}, + + Promotions = maps:put({Lbl, Arg0}, Promotion, Promotions0), + State = State0#amb{ promotions=Promotions, cnt=Counter+1 }, + + {Alias, State} + end. + +amb_insert_promotions(Blocks0, State) -> + F = fun({Lbl, #b_var{}}, Promotion, Blocks) -> + Block = maps:get(Lbl, Blocks), + + Alias = Promotion#b_set.dst, + {Before, After} = splitwith( + fun(#b_set{args=Args}) -> + not is_var_in_args(Alias, Args) + end, Block#b_blk.is), + Is = Before ++ [Promotion | After], + + maps:put(Lbl, Block#b_blk{is=Is}, Blocks) + end, + maps:fold(F, Blocks0, State#amb.promotions). + +is_var_in_args(Var, [Var | _]) -> true; +is_var_in_args(Var, [#b_remote{name=Var} | _]) -> true; +is_var_in_args(Var, [#b_remote{mod=Var} | _]) -> true; +is_var_in_args(Var, [_ | Args]) -> is_var_in_args(Var, Args); +is_var_in_args(_Var, []) -> false. + +%%% +%%% Subpasses +%%% + +%% Removes superflous chained bs_start_match instructions in the same +%% function. When matching on an extracted tail binary, or on a binary we've +%% already matched on, we reuse the original match context. +%% +%% This pass runs first since it makes subsequent optimizations more effective +%% by removing spots where promotion would be required. + +-type prior_match_map() :: + #{ Binary :: #b_var{} => + [{ %% The context and success label of a previous + %% bs_start_match made on this binary. + ValidAfter :: beam_ssa:label(), + Context :: #b_var{} }] }. + +-record(cm, { definitions :: beam_ssa:definition_map(), + dominators :: beam_ssa:dominator_map(), + blocks :: beam_ssa:block_map(), + match_aliases = #{} :: match_alias_map(), + prior_matches = #{} :: prior_match_map(), + renames = #{} :: beam_ssa:rename_map() }). + +combine_matches({Fs0, ModInfo}) -> + Fs = map(fun(F) -> combine_matches(F, ModInfo) end, Fs0), + {Fs, ModInfo}. + +combine_matches(#b_function{bs=Blocks0,cnt=Counter0}=F, ModInfo) -> + case funcinfo_get(F, has_bsm_ops, ModInfo) of + true -> + {Blocks1, State} = + beam_ssa:mapfold_blocks_rpo( + fun(Lbl, #b_blk{is=Is0}=Block0, State0) -> + {Is, State} = cm_1(Is0, [], Lbl, State0), + {Block0#b_blk{is=Is}, State} + end, [0], + #cm{ definitions = beam_ssa:definitions(Blocks0), + dominators = beam_ssa:dominators(Blocks0), + blocks = Blocks0 }, + Blocks0), + + Blocks2 = beam_ssa:rename_vars(State#cm.renames, [0], Blocks1), + + {Blocks, Counter} = alias_matched_binaries(Blocks2, Counter0, + State#cm.match_aliases), + + F#b_function{ bs=Blocks, cnt=Counter }; + false -> + F + end. + +cm_1([#b_set{ op=bs_start_match, + dst=Ctx, + args=[Src] }, + #b_set{ op=succeeded, + dst=Bool, + args=[Ctx] }]=MatchSeq, Acc0, Lbl, State0) -> + Acc = reverse(Acc0), + case is_tail_binary(Src, State0#cm.definitions) of + true -> cm_combine_tail(Src, Ctx, Bool, Acc, State0); + false -> cm_handle_priors(Src, Ctx, Bool, Acc, MatchSeq, Lbl, State0) + end; +cm_1([I | Is], Acc, Lbl, State) -> + cm_1(Is, [I | Acc], Lbl, State); +cm_1([], Acc, _Lbl, State) -> + {reverse(Acc), State}. + +%% If we're dominated by at least one match on the same source, we can reuse +%% the context created by that match. +cm_handle_priors(Src, DstCtx, Bool, Acc, MatchSeq, Lbl, State0) -> + PriorCtxs = case maps:find(Src, State0#cm.prior_matches) of + {ok, Priors} -> + %% We've seen other match contexts on this source, but + %% we can only consider the ones whose success path + %% dominate us. + Dominators = maps:get(Lbl, State0#cm.dominators, []), + [Ctx || {ValidAfter, Ctx} <- Priors, + ordsets:is_element(ValidAfter, Dominators)]; + error -> + [] + end, + case PriorCtxs of + [Ctx|_] -> + Renames0 = State0#cm.renames, + Renames = Renames0#{ Bool => #b_literal{val=true}, DstCtx => Ctx }, + {Acc, State0#cm{ renames = Renames }}; + [] -> + %% Since we lack a prior match, we need to register this one in + %% case we dominate another. + State = cm_register_prior(Src, DstCtx, Lbl, State0), + {Acc ++ MatchSeq, State} + end. + +cm_register_prior(Src, DstCtx, Lbl, State) -> + Block = maps:get(Lbl, State#cm.blocks), + #b_br{succ=ValidAfter} = Block#b_blk.last, + + Priors0 = maps:get(Src, State#cm.prior_matches, []), + Priors = [{ValidAfter, DstCtx} | Priors0], + + PriorMatches = maps:put(Src, Priors, State#cm.prior_matches), + State#cm{ prior_matches = PriorMatches }. + +cm_combine_tail(Src, DstCtx, Bool, Acc, State0) -> + SrcCtx = match_context_of(Src, State0#cm.definitions), + + %% We replace the source with a context alias as it normally won't be used + %% on the happy path after being matched, and the added cost of conversion + %% is negligible if it is. + Aliases = maps:put(Src, {0, SrcCtx}, State0#cm.match_aliases), + + Renames0 = State0#cm.renames, + Renames = Renames0#{ Bool => #b_literal{val=true}, DstCtx => SrcCtx }, + + State = State0#cm{ match_aliases = Aliases, renames = Renames }, + + {Acc, State}. + +%% Lets functions accept match contexts as arguments. The parameter must be +%% unused before the bs_start_match instruction, and it must be matched in the +%% first block. + +-record(aca, { unused_parameters :: ordsets:ordset(#b_var{}), + counter :: non_neg_integer(), + parameter_info = #{} :: #{ #b_var{} => param_info() }, + match_aliases = #{} :: match_alias_map() }). + +accept_context_args({Fs, ModInfo}) -> + mapfoldl(fun accept_context_args/2, ModInfo, Fs). + +accept_context_args(#b_function{bs=Blocks0}=F, ModInfo0) -> + case funcinfo_get(F, has_bsm_ops, ModInfo0) of + true -> + Parameters = ordsets:from_list(funcinfo_get(F, parameters, ModInfo0)), + State0 = #aca{ unused_parameters = Parameters, + counter = F#b_function.cnt }, + + {Blocks1, State} = aca_1(Blocks0, State0), + {Blocks, Counter} = alias_matched_binaries(Blocks1, + State#aca.counter, + State#aca.match_aliases), + + ModInfo = funcinfo_set(F, parameter_info, State#aca.parameter_info, + ModInfo0), + + {F#b_function{bs=Blocks,cnt=Counter}, ModInfo}; + false -> + {F, ModInfo0} + end. + +aca_1(Blocks, State) -> + %% We only handle block 0 as we don't yet support starting a match after a + %% test. This is generally good enough as the sys_core_bsm pass makes the + %% match instruction come first if possible, and it's rare for a function + %% to binary-match several parameters at once. + EntryBlock = maps:get(0, Blocks), + aca_enable_reuse(EntryBlock#b_blk.is, EntryBlock, Blocks, [], State). + +aca_enable_reuse([#b_set{op=bs_start_match,args=[Src]}=I0 | Rest], + EntryBlock, Blocks0, Acc, State0) -> + case aca_is_reuse_safe(Src, State0) of + true -> + {I, Last, Blocks1, State} = + aca_reuse_context(I0, EntryBlock, Blocks0, State0), + + Is = reverse([I|Acc]) ++ Rest, + Blocks = maps:put(0, EntryBlock#b_blk{is=Is,last=Last}, Blocks1), + + {Blocks, State}; + false -> + {Blocks0, State0} + end; +aca_enable_reuse([I | Is], EntryBlock, Blocks, Acc, State0) -> + UnusedParams0 = State0#aca.unused_parameters, + case ordsets:intersection(UnusedParams0, beam_ssa:used(I)) of + [] -> + aca_enable_reuse(Is, EntryBlock, Blocks, [I | Acc], State0); + PrematureUses -> + UnusedParams = ordsets:subtract(UnusedParams0, PrematureUses), + + %% Mark the offending parameters as unsuitable for context reuse. + ParamInfo = foldl(fun(A, Ps) -> + maps:put(A, {used_before_match, I}, Ps) + end, State0#aca.parameter_info, PrematureUses), + + State = State0#aca{ unused_parameters = UnusedParams, + parameter_info = ParamInfo }, + aca_enable_reuse(Is, EntryBlock, Blocks, [I | Acc], State) + end; +aca_enable_reuse([], _EntryBlock, Blocks, _Acc, State) -> + {Blocks, State}. + +aca_is_reuse_safe(Src, State) -> + %% Context reuse is unsafe unless all uses are dominated by the start_match + %% instruction. Since we only process block 0 it's enough to check if + %% they're unused so far. + ordsets:is_element(Src, State#aca.unused_parameters). + +aca_reuse_context(#b_set{dst=Dst, args=[Src]}=I0, Block, Blocks0, State0) -> + %% When matching fails on a reused context it needs to be converted back + %% to a binary. We only need to do this on the success path since it can't + %% be a context on the type failure path, but it's very common for these + %% to converge which requires special handling. + {State1, Last, Blocks} = + aca_handle_convergence(Src, State0, Block#b_blk.last, Blocks0), + + Aliases = maps:put(Src, {Last#b_br.succ, Dst}, State1#aca.match_aliases), + ParamInfo = maps:put(Src, suitable_for_reuse, State1#aca.parameter_info), + + State = State1#aca{ match_aliases = Aliases, + parameter_info = ParamInfo }, + + I = beam_ssa:add_anno(accepts_match_contexts, true, I0), + + {I, Last, Blocks, State}. + +aca_handle_convergence(Src, State0, Last0, Blocks0) -> + #b_br{fail=Fail0,succ=Succ0} = Last0, + + SuccPath = beam_ssa:rpo([Succ0], Blocks0), + FailPath = beam_ssa:rpo([Fail0], Blocks0), + + %% The promotion logic in alias_matched_binaries breaks down if the source + %% is used after the fail/success paths converge, as we have no way to tell + %% whether the source is a match context or something else past that point. + %% + %% We could handle this through clever insertion of phi nodes but it's + %% far simpler to copy either branch in its entirety. It doesn't matter + %% which one as long as they become disjoint. + ConvergedPaths = ordsets:intersection( + ordsets:from_list(SuccPath), + ordsets:from_list(FailPath)), + + case maps:is_key(Src, beam_ssa:uses(ConvergedPaths, Blocks0)) of + true -> + case shortest(SuccPath, FailPath) of + left -> + {Succ, Blocks, Counter} = + aca_copy_successors(Succ0, Blocks0, State0#aca.counter), + State = State0#aca{ counter = Counter }, + {State, Last0#b_br{succ=Succ}, Blocks}; + right -> + {Fail, Blocks, Counter} = + aca_copy_successors(Fail0, Blocks0, State0#aca.counter), + State = State0#aca{ counter = Counter }, + {State, Last0#b_br{fail=Fail}, Blocks} + end; + false -> + {State0, Last0, Blocks0} + end. + +shortest([_|As], [_|Bs]) -> shortest(As, Bs); +shortest([], _) -> left; +shortest(_, []) -> right. + +%% Copies all successor blocks of Lbl, returning the label to the entry block +%% of this copy. Since the copied blocks aren't referenced anywhere else, they +%% are all guaranteed to be dominated by Lbl. +aca_copy_successors(Lbl0, Blocks0, Counter0) -> + %% Building the block rename map up front greatly simplifies phi node + %% handling. + Path = beam_ssa:rpo([Lbl0], Blocks0), + {BRs, Counter1} = aca_cs_build_brs(Path, Counter0, #{}), + {Blocks, Counter} = aca_cs_1(Path, Blocks0, Counter1, #{}, BRs, #{}), + Lbl = maps:get(Lbl0, BRs), + {Lbl, Blocks, Counter}. + +aca_cs_build_brs([Lbl | Path], Counter0, Acc) -> + aca_cs_build_brs(Path, Counter0 + 1, maps:put(Lbl, Counter0, Acc)); +aca_cs_build_brs([], Counter, Acc) -> + {Acc, Counter}. + +aca_cs_1([Lbl0 | Path], Blocks, Counter0, VRs0, BRs, Acc0) -> + Block0 = maps:get(Lbl0, Blocks), + Lbl = maps:get(Lbl0, BRs), + {VRs, Block, Counter} = aca_cs_block(Block0, Counter0, VRs0, BRs), + Acc = maps:put(Lbl, Block, Acc0), + aca_cs_1(Path, Blocks, Counter, VRs, BRs, Acc); +aca_cs_1([], Blocks, Counter, _VRs, _BRs, Acc) -> + {maps:merge(Blocks, Acc), Counter}. + +aca_cs_block(#b_blk{is=Is0,last=Last0}=Block0, Counter0, VRs0, BRs) -> + {VRs, Is, Counter} = aca_cs_is(Is0, Counter0, VRs0, BRs, []), + Last = aca_cs_last(Last0, VRs, BRs), + Block = Block0#b_blk{is=Is,last=Last}, + {VRs, Block, Counter}. + +aca_cs_is([#b_set{op=Op, + dst=Dst0, + args=Args0}=I0 | Is], + Counter0, VRs0, BRs, Acc) -> + Args = case Op of + phi -> aca_cs_args_phi(Args0, VRs0, BRs); + _ -> aca_cs_args(Args0, VRs0) + end, + Counter = Counter0 + 1, + Dst = #b_var{name={'@ssa_bsm_aca',Counter}}, + I = I0#b_set{dst=Dst,args=Args}, + VRs = maps:put(Dst0, Dst, VRs0), + aca_cs_is(Is, Counter, VRs, BRs, [I | Acc]); +aca_cs_is([], Counter, VRs, _BRs, Acc) -> + {VRs, reverse(Acc), Counter}. + +aca_cs_last(#b_switch{arg=Arg0,list=Switch0,fail=Fail0}=Sw, VRs, BRs) -> + Switch = [{Literal, maps:get(Lbl, BRs)} || {Literal, Lbl} <- Switch0], + Sw#b_switch{arg=aca_cs_arg(Arg0, VRs), + fail=maps:get(Fail0, BRs), + list=Switch}; +aca_cs_last(#b_br{bool=Arg0,succ=Succ0,fail=Fail0}=Br, VRs, BRs) -> + Br#b_br{bool=aca_cs_arg(Arg0, VRs), + succ=maps:get(Succ0, BRs), + fail=maps:get(Fail0, BRs)}; +aca_cs_last(#b_ret{arg=Arg0}=Ret, VRs, _BRs) -> + Ret#b_ret{arg=aca_cs_arg(Arg0, VRs)}. + +aca_cs_args_phi([{Arg, Lbl} | Args], VRs, BRs) -> + case BRs of + #{ Lbl := New } -> + [{aca_cs_arg(Arg, VRs), New} | aca_cs_args_phi(Args, VRs, BRs)]; + #{} -> + aca_cs_args_phi(Args, VRs, BRs) + end; +aca_cs_args_phi([], _VRs, _BRs) -> + []. + +aca_cs_args([Arg | Args], VRs) -> + [aca_cs_arg(Arg, VRs) | aca_cs_args(Args, VRs)]; +aca_cs_args([], _VRs) -> + []. + +aca_cs_arg(#b_remote{mod=Mod0,name=Name0}=Rem, VRs) -> + Mod = aca_cs_arg(Mod0, VRs), + Name = aca_cs_arg(Name0, VRs), + Rem#b_remote{mod=Mod,name=Name}; +aca_cs_arg(Arg, VRs) -> + case VRs of + #{ Arg := New } -> New; + #{} -> Arg + end. + +%% Allows contexts to pass through "wrapper functions" where the context is +%% passed directly to a function that accepts match contexts (including other +%% wrappers). +%% +%% This does not alter the function in any way, it only changes parameter info +%% so that skip_outgoing_tail_extraction is aware that it's safe to pass +%% contexts to us. + +allow_context_passthrough({Fs, ModInfo0}) -> + ModInfo = + acp_forward_params([{F, beam_ssa:uses(F#b_function.bs)} || F <- Fs], + ModInfo0), + {Fs, ModInfo}. + +acp_forward_params(FsUses, ModInfo0) -> + F = fun({#b_function{args=Parameters}=Func, UseMap}, ModInfo) -> + ParamInfo = + foldl(fun(Param, ParamInfo) -> + Uses = maps:get(Param, UseMap, []), + acp_1(Param, Uses, ModInfo, ParamInfo) + end, + funcinfo_get(Func, parameter_info, ModInfo), + Parameters), + funcinfo_set(Func, parameter_info, ParamInfo, ModInfo) + end, + %% Allowing context passthrough on one function may make it possible to + %% enable it on another, so it needs to be repeated for maximum effect. + case foldl(F, ModInfo0, FsUses) of + ModInfo0 -> ModInfo0; + Changed -> acp_forward_params(FsUses, Changed) + end. + +%% We have no way to know if an argument is a context, so it's only safe to +%% forward them if they're passed exactly once in the first block. Any other +%% uses are unsafe, including function_clause errors. +acp_1(Param, [{0, #b_set{op=call}=I}], ModInfo, ParamInfo) -> + %% We don't need to provide a context chain as our callers make sure that + %% multiple arguments never reference the same context. + case check_context_call(I, Param, [], ModInfo) of + {no_match_on_entry, _} -> ParamInfo; + Other -> maps:put(Param, Other, ParamInfo) + end; +acp_1(_Param, _Uses, _ModInfo, ParamInfo) -> + ParamInfo. + +%% This is conceptually similar to combine_matches but operates across +%% functions. Whenever a tail binary is passed to a parameter that accepts +%% match contexts we'll pass the context instead, improving performance by +%% avoiding the creation of a new match context in the callee. +%% +%% We also create an alias to delay extraction until it's needed as an actual +%% binary, which is often rare on the happy path. The cost of being wrong is +%% negligible (`bs_test_unit + bs_get_tail` vs `bs_get_binary`) so we're +%% applying it unconditionally to keep things simple. + +-record(sote, { definitions :: beam_ssa:definition_map(), + mod_info :: module_info(), + match_aliases = #{} :: match_alias_map() }). + +skip_outgoing_tail_extraction({Fs0, ModInfo}) -> + Fs = map(fun(F) -> skip_outgoing_tail_extraction(F, ModInfo) end, Fs0), + {Fs, ModInfo}. + +skip_outgoing_tail_extraction(#b_function{bs=Blocks0}=F, ModInfo) -> + case funcinfo_get(F, has_bsm_ops, ModInfo) of + true -> + State0 = #sote{ definitions = beam_ssa:definitions(Blocks0), + mod_info = ModInfo }, + + {Blocks1, State} = beam_ssa:mapfold_instrs_rpo( + fun sote_rewrite_calls/2, [0], State0, Blocks0), + + {Blocks, Counter} = alias_matched_binaries(Blocks1, + F#b_function.cnt, + State#sote.match_aliases), + + F#b_function{bs=Blocks,cnt=Counter}; + false -> + F + end. + +sote_rewrite_calls(#b_set{op=call,args=Args}=Call, State) -> + sote_rewrite_call(Call, Args, [], State); +sote_rewrite_calls(I, State) -> + {I, State}. + +sote_rewrite_call(Call, [], ArgsOut, State) -> + {Call#b_set{args=reverse(ArgsOut)}, State}; +sote_rewrite_call(Call0, [Arg | ArgsIn], ArgsOut, State0) -> + case is_tail_binary(Arg, State0#sote.definitions) of + true -> + CtxChain = context_chain_of(Arg, State0#sote.definitions), + case check_context_call(Call0, Arg, CtxChain, State0#sote.mod_info) of + suitable_for_reuse -> + Ctx = match_context_of(Arg, State0#sote.definitions), + + MatchAliases0 = State0#sote.match_aliases, + MatchAliases = maps:put(Arg, {0, Ctx}, MatchAliases0), + State = State0#sote{ match_aliases = MatchAliases }, + + Call = beam_ssa:add_anno(bsm_info, context_reused, Call0), + sote_rewrite_call(Call, ArgsIn, [Ctx | ArgsOut], State); + Other -> + Call = beam_ssa:add_anno(bsm_info, Other, Call0), + sote_rewrite_call(Call, ArgsIn, [Arg | ArgsOut], State0) + end; + false -> + sote_rewrite_call(Call0, ArgsIn, [Arg | ArgsOut], State0) + end. + +%% Adds parameter_type_info annotations to help the validator determine whether +%% our optimizations were safe. + +annotate_context_parameters({Fs, ModInfo}) -> + mapfoldl(fun annotate_context_parameters/2, ModInfo, Fs). + +annotate_context_parameters(F, ModInfo) -> + ParamInfo = funcinfo_get(F, parameter_info, ModInfo), + TypeAnno0 = beam_ssa:get_anno(parameter_type_info, F, #{}), + TypeAnno = maps:fold(fun(K, _V, Acc) when is_map_key(K, Acc) -> + %% Assertion. + error(conflicting_parameter_types); + (K, suitable_for_reuse, Acc) -> + T = beam_validator:type_anno(match_context), + Acc#{ K => T }; + (_K, _V, Acc) -> + Acc + end, TypeAnno0, ParamInfo), + {beam_ssa:add_anno(parameter_type_info, TypeAnno, F), ModInfo}. + +%%% +%%% +bin_opt_info +%%% + +collect_opt_info(Fs) -> + foldl(fun(#b_function{bs=Blocks}=F, Acc0) -> + UseMap = beam_ssa:uses(Blocks), + Where = beam_ssa:get_anno(location, F, []), + beam_ssa:fold_instrs_rpo( + fun(I, Acc) -> + collect_opt_info_1(I, Where, UseMap, Acc) + end, [0], Acc0, Blocks) + end, [], Fs). + +collect_opt_info_1(#b_set{op=Op,anno=Anno,dst=Dst}=I, Where, UseMap, Acc0) -> + case is_tail_binary(I) of + true when Op =:= bs_match -> + %% The uses include when the context is passed raw, so we discard + %% everything but the bs_extract instruction to limit warnings to + %% unoptimized uses. + Uses0 = maps:get(Dst, UseMap, []), + case [E || {_, #b_set{op=bs_extract}=E} <- Uses0] of + [Use] -> add_unopt_binary_info(Use, false, Where, UseMap, Acc0); + [] -> Acc0 + end; + true -> + %% Add a warning for each use. Note that we don't do anything + %% special if unused as a later pass will remove this instruction + %% anyway. + Uses = maps:get(Dst, UseMap, []), + foldl(fun({_Lbl, Use}, Acc) -> + add_unopt_binary_info(Use, false, Where, UseMap, Acc) + end, Acc0, Uses); + false -> + add_opt_info(Anno, Where, Acc0) + end; +collect_opt_info_1(#b_ret{anno=Anno}, Where, _UseMap, Acc) -> + add_opt_info(Anno, Where, Acc); +collect_opt_info_1(_I, _Where, _Uses, Acc) -> + Acc. + +add_opt_info(Anno, Where, Acc) -> + case maps:find(bsm_info, Anno) of + {ok, Term} -> [make_warning(Term, Anno, Where) | Acc]; + error -> Acc + end. + +%% When an alias is promoted we need to figure out where it goes to ignore +%% warnings for compiler-generated things, and provide more useful warnings in +%% general. +%% +%% We track whether the binary has been used to build another term because it +%% can be helpful when there's no line information. + +add_unopt_binary_info(#b_set{op=Follow,dst=Dst}, _Nested, Where, UseMap, Acc0) + when Follow =:= put_tuple; + Follow =:= put_list; + Follow =:= put_map -> + %% Term-building instructions. + {_, Uses} = unzip(maps:get(Dst, UseMap, [])), + foldl(fun(Use, Acc) -> + add_unopt_binary_info(Use, true, Where, UseMap, Acc) + end, Acc0, Uses); +add_unopt_binary_info(#b_set{op=Follow,dst=Dst}, Nested, Where, UseMap, Acc0) + when Follow =:= bs_extract; + Follow =:= phi -> + %% Non-building instructions that need to be followed. + {_, Uses} = unzip(maps:get(Dst, UseMap, [])), + foldl(fun(Use, Acc) -> + add_unopt_binary_info(Use, Nested, Where, UseMap, Acc) + end, Acc0, Uses); +add_unopt_binary_info(#b_set{op=call, + args=[#b_remote{mod=#b_literal{val=erlang}, + name=#b_literal{val=error}} | + _Ignored]}, + _Nested, _Where, _UseMap, Acc) -> + %% There's no nice way to tell compiler-generated exceptions apart from + %% user ones so we ignore them all. I doubt anyone cares. + Acc; +add_unopt_binary_info(#b_switch{anno=Anno}=I, Nested, Where, _UseMap, Acc) -> + [make_promotion_warning(I, Nested, Anno, Where) | Acc]; +add_unopt_binary_info(#b_set{anno=Anno}=I, Nested, Where, _UseMap, Acc) -> + [make_promotion_warning(I, Nested, Anno, Where) | Acc]; +add_unopt_binary_info(#b_ret{anno=Anno}=I, Nested, Where, _UseMap, Acc) -> + [make_promotion_warning(I, Nested, Anno, Where) | Acc]; +add_unopt_binary_info(#b_br{anno=Anno}=I, Nested, Where, _UseMap, Acc) -> + [make_promotion_warning(I, Nested, Anno, Where) | Acc]. + +make_promotion_warning(I, Nested, Anno, Where) -> + make_warning({binary_created, I, Nested}, Anno, Where). + +make_warning(Term, Anno, Where) -> + {File, Line} = maps:get(location, Anno, Where), + {File,[{Line,?MODULE,Term}]}. + +format_opt_info(context_reused) -> + "OPTIMIZED: match context reused"; +format_opt_info({binary_created, _, _}=Promotion) -> + io_lib:format("BINARY CREATED: ~s", [format_opt_info_1(Promotion)]); +format_opt_info(Other) -> + io_lib:format("NOT OPTIMIZED: ~s", [format_opt_info_1(Other)]). + +format_opt_info_1({binary_created, #b_set{op=call,args=[Call|_]}, false}) -> + io_lib:format("binary is used in call to ~s which doesn't support " + "context reuse", [format_call(Call)]); +format_opt_info_1({binary_created, #b_set{op=call,args=[Call|_]}, true}) -> + io_lib:format("binary is used in term passed to ~s", + [format_call(Call)]); +format_opt_info_1({binary_created, #b_set{op={bif, BIF},args=Args}, false}) -> + io_lib:format("binary is used in ~p/~p which doesn't support context " + "reuse", [BIF, length(Args)]); +format_opt_info_1({binary_created, #b_set{op={bif, BIF},args=Args}, true}) -> + io_lib:format("binary is used in term passed to ~p/~p", + [BIF, length(Args)]); +format_opt_info_1({binary_created, #b_set{op=Op}, false}) -> + io_lib:format("binary is used in '~p' which doesn't support context " + "reuse", [Op]); +format_opt_info_1({binary_created, #b_set{op=Op}, true}) -> + io_lib:format("binary is used in term passed to '~p'", [Op]); +format_opt_info_1({binary_created, #b_ret{}, false}) -> + io_lib:format("binary is returned from the function", []); +format_opt_info_1({binary_created, #b_ret{}, true}) -> + io_lib:format("binary is used in a term that is returned from the " + "function", []); +format_opt_info_1({unsuitable_call, {Call, Inner}}) -> + io_lib:format("binary used in call to ~s, where ~s", + [format_call(Call), format_opt_info_1(Inner)]); +format_opt_info_1({remote_call, Call}) -> + io_lib:format("binary is used in remote call to ~s", [format_call(Call)]); +format_opt_info_1({fun_call, Call}) -> + io_lib:format("binary is used in fun call (~s)", + [format_call(Call)]); +format_opt_info_1({multiple_uses_in_call, Call}) -> + io_lib:format("binary is passed as multiple arguments to ~s", + [format_call(Call)]); +format_opt_info_1({no_match_on_entry, Call}) -> + io_lib:format("binary is used in call to ~s which does not begin with a " + "suitable binary match", [format_call(Call)]); +format_opt_info_1({used_before_match, #b_set{op=call,args=[Call|_]}}) -> + io_lib:format("binary is used in call to ~s before being matched", + [format_call(Call)]); +format_opt_info_1({used_before_match, #b_set{op={bif, BIF},args=Args}}) -> + io_lib:format("binary is used in ~p/~p before being matched", + [BIF, length(Args)]); +format_opt_info_1({used_before_match, #b_set{op=phi}}) -> + io_lib:format("binary is returned from an expression before being " + "matched", []); +format_opt_info_1({used_before_match, #b_set{op=Op}}) -> + io_lib:format("binary is used in '~p' before being matched",[Op]); +format_opt_info_1(Term) -> + io_lib:format("~w", [Term]). + +format_call(#b_local{name=#b_literal{val=F},arity=A}) -> + io_lib:format("~p/~p", [F, A]); +format_call(#b_remote{mod=#b_literal{val=M},name=#b_literal{val=F},arity=A}) -> + io_lib:format("~p:~p/~p", [M, F, A]); +format_call(Fun) -> + io_lib:format("~p", [Fun]). diff --git a/lib/compiler/src/beam_ssa_codegen.erl b/lib/compiler/src/beam_ssa_codegen.erl index 006c41c0e0..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}) -> []; @@ -315,12 +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; @@ -329,7 +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; @@ -592,8 +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), @@ -610,7 +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; @@ -618,7 +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); @@ -641,7 +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), @@ -696,6 +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; @@ -704,7 +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}) -> @@ -728,13 +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]), @@ -749,7 +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 @@ -760,7 +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}}; @@ -795,6 +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 @@ -818,13 +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 -> @@ -995,8 +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. @@ -1023,12 +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,[]}, @@ -1038,17 +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}; @@ -1180,18 +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 -> @@ -1205,6 +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}|_], _) -> @@ -1212,7 +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) -> []; @@ -1259,22 +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}, @@ -1327,6 +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). @@ -1342,10 +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) -> @@ -1479,13 +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; @@ -1521,6 +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,[]}. @@ -1604,12 +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). @@ -1698,7 +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 diff --git a/lib/compiler/src/beam_ssa_dead.erl b/lib/compiler/src/beam_ssa_dead.erl new file mode 100644 index 0000000000..2cca9ebadf --- /dev/null +++ b/lib/compiler/src/beam_ssa_dead.erl @@ -0,0 +1,1038 @@ +%% +%% %CopyrightBegin% +%% +%% Copyright Ericsson AB 2018. All Rights Reserved. +%% +%% Licensed under the Apache License, Version 2.0 (the "License"); +%% you may not use this file except in compliance with the License. +%% You may obtain a copy of the License at +%% +%% http://www.apache.org/licenses/LICENSE-2.0 +%% +%% Unless required by applicable law or agreed to in writing, software +%% distributed under the License is distributed on an "AS IS" BASIS, +%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +%% See the License for the specific language governing permissions and +%% limitations under the License. +%% +%% %CopyrightEnd% +%% +%% Dead code is code that is executed but has no effect. This +%% optimization pass either removes dead code or jumps around it, +%% potentially making it unreachable so that it can be dropped +%% the next time beam_ssa:linearize/1 is called. +%% + +-module(beam_ssa_dead). +-export([opt/1]). + +-include("beam_ssa.hrl"). +-import(lists, [append/1,last/1,member/2,takewhile/2,reverse/1]). + +-type used_vars() :: #{beam_ssa:label():=ordsets:ordset(beam_ssa:var_name())}. + +-type basic_type_test() :: atom() | {'is_tagged_tuple',pos_integer(),atom()}. +-type type_test() :: basic_type_test() | {'not',basic_type_test()}. +-type op_name() :: atom(). +-type basic_rel_op() :: {op_name(),beam_ssa:b_var(),beam_ssa:value()} | + {basic_type_test(),beam_ssa:value()}. +-type rel_op() :: {op_name(),beam_ssa:b_var(),beam_ssa:value()} | + {type_test(),beam_ssa:value()}. + +-record(st, + {bs :: beam_ssa:block_map(), + us :: used_vars(), + skippable :: #{beam_ssa:label():='true'}, + rel_op=none :: 'none' | rel_op(), + target=any :: 'any' | 'one_way' | beam_ssa:label() + }). + +-spec opt([{Label0,Block0}]) -> [{Label,Block}] when + Label0 :: beam_ssa:label(), + Block0 :: beam_ssa:b_blk(), + Label :: beam_ssa:label(), + Block :: beam_ssa:b_blk(). + +opt(Linear) -> + {Used,Skippable} = used_vars(Linear), + Blocks0 = maps:from_list(Linear), + St0 = #st{bs=Blocks0,us=Used,skippable=Skippable}, + St = shortcut_opt(St0), + #st{bs=Blocks} = combine_eqs(St), + beam_ssa:linearize(Blocks). + +%%% +%%% Shortcut br/switch targets. +%%% +%%% A br/switch may branch to another br/switch that in turn always +%%% branches to another target. Rewrite br/switch to refer to the +%%% ultimate targets directly. That will save execution time, but +%%% could also reduce the size of the code if some of the original +%%% targets become unreachable and be deleted. +%%% +%%% When rewriting branches, we must be careful not to skip instructions +%%% that have side effects or that bind variables that will be used +%%% at the new target. +%%% +%%% We must also avoid branching to phi nodes. The reason is +%%% twofold. First, we might create a critical edge which is strictly +%%% forbidden. Second, there will be a branch from a block that is not +%%% listed in the list of predecessors in the phi node. Those +%%% limitations could probably be overcome, but it is not clear how +%%% much that would improve the code. +%%% + +shortcut_opt(#st{bs=Blocks}=St) -> + %% Processing the blocks in reverse post order seems to give more + %% opportunities for optimizations compared to post order. (Based on + %% running scripts/diffable with both PO and RPO and looking at + %% the diff.) + Ls = beam_ssa:rpo(Blocks), + shortcut_opt(Ls, #{from=>0}, St). + +shortcut_opt([L|Ls], Bs0, #st{bs=Blocks0}=St) -> + #b_blk{is=Is,last=Last0} = Blk0 = get_block(L, St), + Bs = Bs0#{from:=L}, + case shortcut_terminator(Last0, Is, Bs, St) of + Last0 -> + %% No change. No need to update the block. + shortcut_opt(Ls, Bs, St); + Last -> + %% The terminator was simplified in some way. + %% Update the block. + Blk = Blk0#b_blk{last=Last}, + Blocks = Blocks0#{L=>Blk}, + shortcut_opt(Ls, Bs, St#st{bs=Blocks}) + end; +shortcut_opt([], _, St) -> St. + +shortcut_terminator(#b_br{bool=#b_literal{val=true},succ=Succ0}, + _Is, Bs, St0) -> + St = St0#st{rel_op=none}, + shortcut(Succ0, Bs, St); +shortcut_terminator(#b_br{bool=#b_var{}=Bool,succ=Succ0,fail=Fail0}=Br, + Is, Bs, St0) -> + St = St0#st{target=one_way}, + RelOp = get_rel_op(Bool, Is), + SuccBs = bind_var(Bool, #b_literal{val=true}, Bs), + BrSucc = shortcut(Succ0, SuccBs, St#st{rel_op=RelOp}), + FailBs = bind_var(Bool, #b_literal{val=false}, Bs), + BrFail = shortcut(Fail0, FailBs, St#st{rel_op=invert_op(RelOp)}), + case {BrSucc,BrFail} of + {#b_br{bool=#b_literal{val=true},succ=Succ}, + #b_br{bool=#b_literal{val=true},succ=Fail}} + when Succ =/= Succ0; Fail =/= Fail0 -> + %% One or both of the targets were cut short. + beam_ssa:normalize(Br#b_br{succ=Succ,fail=Fail}); + {_,_} -> + %% No change. + Br + end; +shortcut_terminator(#b_switch{arg=Bool,list=List0}=Sw, _Is, Bs, St) -> + List = shortcut_switch(List0, Bool, Bs, St), + beam_ssa:normalize(Sw#b_switch{list=List}); +shortcut_terminator(Last, _Is, _Bs, _St) -> + Last. + +shortcut_switch([{Lit,L0}|T], Bool, Bs, St0) -> + RelOp = {'=:=',Bool,Lit}, + St = St0#st{rel_op=RelOp}, + #b_br{bool=#b_literal{val=true},succ=L} = + shortcut(L0, bind_var(Bool, Lit, Bs), St#st{target=one_way}), + [{Lit,L}|shortcut_switch(T, Bool, Bs, St0)]; +shortcut_switch([], _, _, _) -> []. + +shortcut(L, Bs, St) -> + shortcut_1(L, Bs, ordsets:new(), St). + +shortcut_1(L, Bs0, UnsetVars0, St) -> + case shortcut_2(L, Bs0, UnsetVars0, St) of + none -> + %% No more shortcuts found. Package up the previous + %% label in an unconditional branch. + #b_br{bool=#b_literal{val=true},succ=L,fail=L}; + {#b_br{bool=#b_var{}}=Br,_,_} -> + %% This is a two-way branch. We can't do any better. + Br; + {#b_br{bool=#b_literal{val=true},succ=Succ},Bs,UnsetVars} -> + %% This is a safe `br`, but try to find a better one. + shortcut_1(Succ, Bs#{from:=L}, UnsetVars, St) + end. + +%% Try to shortcut this block, branching to a successor. +shortcut_2(L, Bs0, UnsetVars0, St) -> + #b_blk{is=Is,last=Last} = get_block(L, St), + case eval_is(Is, Bs0, St) of + none -> + %% It is not safe to avoid this block because it + %% has instructions with potential side effects. + none; + Bs -> + %% The instructions in the block (if any) don't + %% have any side effects and can be skipped. + %% Evaluate the terminator. + case eval_terminator(Last, Bs, St) of + none -> + %% The terminator is not suitable (could be + %% because it is a switch that can't be simplified + %% or it is a ret instruction). + none; + #b_br{}=Br -> + %% We have a potentially suitable br. + %% Now update the set of variables that will never + %% be set if this block will be skipped. + SetInThisBlock = [V || #b_set{dst=V} <- Is], + UnsetVars = update_unset_vars(L, Br, SetInThisBlock, + UnsetVars0, St), + + %% Continue checking whether this br is suitable. + shortcut_3(Br, Bs#{from:=L}, UnsetVars, St) + end + end. + +shortcut_3(Br, Bs, UnsetVars, #st{target=Target}=St) -> + case is_br_safe(UnsetVars, Br, St) of + false -> + %% Branching using this `br` is unsafe, either because it + %% is an unconditional branch to a phi node, or because + %% one or more of the variables that are not set will be + %% used. Try to follow branches of this `br`, to find a + %% safe `br`. + case Br of + #b_br{bool=#b_literal{val=true},succ=L} -> + case Target of + L -> + %% We have reached the forced target, and it + %% is unsafe. Give up. + none; + _ -> + %% Try following this branch to see whether it + %% leads to a safe `br`. + shortcut_2(L, Bs, UnsetVars, St) + end; + #b_br{bool=#b_var{},succ=Succ,fail=Fail} -> + case {Succ,Fail} of + {L,Target} -> + %% The failure label is the forced target. + %% Try following the success label to see + %% whether it also ultimately ends up at the + %% forced target. + shortcut_2(L, Bs, UnsetVars, St); + {Target,L} -> + %% The success label is the forced target. + %% Try following the failure label to see + %% whether it also ultimately ends up at the + %% forced target. + shortcut_2(L, Bs, UnsetVars, St); + {_,_} -> + case Target of + any -> + %% This two-way branch is unsafe. Try reducing + %% it to a one-way branch. + shortcut_two_way(Br, Bs, UnsetVars, St); + one_way -> + %% This two-way branch is unsafe. Try reducing + %% it to a one-way branch. + shortcut_two_way(Br, Bs, UnsetVars, St); + _ when is_integer(Target) -> + %% This two-way branch is unsafe, and + %% there already is a forced target. + %% Give up. + none + end + end + end; + true -> + %% This `br` instruction is safe. It does not + %% branch to a phi node, and all variables that + %% will be used are guaranteed to be defined. + case Br of + #b_br{bool=#b_literal{val=true},succ=L} -> + %% This is a one-way branch. + case Target of + any -> + %% No forced target. Success! + {Br,Bs,UnsetVars}; + one_way -> + %% The target must be a one-way branch, which this + %% `br` is. Success! + {Br,Bs,UnsetVars}; + L when is_integer(Target) -> + %% The forced target is L. Success! + {Br,Bs,UnsetVars}; + _ when is_integer(Target) -> + %% Wrong forced target. Try following this branch + %% to see if it ultimately ends up at the forced + %% target. + shortcut_2(L, Bs, UnsetVars, St) + end; + #b_br{bool=#b_var{}} -> + %% This is a two-way branch. + if + Target =:= any; Target =:= one_way -> + %% No specific forced target. Try to reduce the + %% two-way branch to an one-way branch. + case shortcut_two_way(Br, Bs, UnsetVars, St) of + none when Target =:= any -> + %% This `br` can't be reduced to a one-way + %% branch. Return the `br` as-is. + {Br,Bs,UnsetVars}; + none when Target =:= one_way -> + %% This `br` can't be reduced to a one-way + %% branch. The caller wants a one-way branch. + %% Give up. + none; + {_,_,_}=Res -> + %% This `br` was successfully reduced to a + %% one-way branch. + Res + end; + is_integer(Target) -> + %% There is a forced target, which can't + %% be reached because this `br` is a two-way + %% branch. Give up. + none + end + end + end. + +update_unset_vars(L, Br, SetInThisBlock, UnsetVars, #st{skippable=Skippable}) -> + case is_map_key(L, Skippable) of + true -> + %% None of the variables used in this block are used in + %% the successors. We can speed up compilation by avoiding + %% adding variables to the UnsetVars if the presence of + %% those variable would not change the outcome of the + %% tests in is_br_safe/2. + case Br of + #b_br{bool=Bool} -> + case member(Bool, SetInThisBlock) of + true -> + %% Bool is a variable defined in this + %% block. It will change the outcome of + %% the `not member(V, UnsetVars)` check in + %% is_br_safe/2. The other variables + %% defined in this block will not. + ordsets:add_element(Bool, UnsetVars); + false -> + %% Bool is either a variable not defined + %% in this block or a literal. Adding it + %% to the UnsetVars set would not change + %% the outcome of the tests in + %% is_br_safe/2. + UnsetVars + end + end; + false -> + ordsets:union(UnsetVars, ordsets:from_list(SetInThisBlock)) + end. + +shortcut_two_way(#b_br{succ=Succ,fail=Fail}, Bs0, UnsetVars0, St) -> + case shortcut_2(Succ, Bs0, UnsetVars0, St#st{target=Fail}) of + {#b_br{bool=#b_literal{},succ=Fail},_,_}=Res -> + Res; + none -> + case shortcut_2(Fail, Bs0, UnsetVars0, St#st{target=Succ}) of + {#b_br{bool=#b_literal{},succ=Succ},_,_}=Res -> + Res; + none -> + none + end + end. + +get_block(L, St) -> + #st{bs=#{L:=Blk}} = St, + Blk. + +is_br_safe(UnsetVars, Br, #st{us=Us}=St) -> + %% Check that none of the unset variables will be used. + case Br of + #b_br{bool=#b_var{}=V,succ=Succ,fail=Fail} -> + #{Succ:=Used0,Fail:=Used1} = Us, + + %% A two-way branch never branches to a phi node, so there + %% is no need to check for phi nodes here. + not member(V, UnsetVars) andalso + ordsets:is_disjoint(Used0, UnsetVars) andalso + ordsets:is_disjoint(Used1, UnsetVars); + #b_br{succ=Same,fail=Same} -> + %% An unconditional branch must not jump to + %% a phi node. + not is_forbidden(Same, St) andalso + ordsets:is_disjoint(map_get(Same, Us), UnsetVars) + end. + +is_forbidden(L, St) -> + case get_block(L, St) of + #b_blk{is=[#b_set{op=phi}|_]} -> true; + #b_blk{is=[#b_set{op=peek_message}|_]} -> true; + #b_blk{} -> false + end. + + +%% Evaluate the instructions in the block. +%% Return the updated bindings, or 'none' if there is +%% any instruction with potential side effects. + +eval_is([#b_set{op=phi,dst=Dst,args=Args}|Is], Bs0, St) -> + From = map_get(from, Bs0), + [Val] = [Val || {Val,Pred} <- Args, Pred =:= From], + Bs = bind_var(Dst, Val, Bs0), + eval_is(Is, Bs, St); +eval_is([#b_set{op={bif,_},dst=Dst}=I0|Is], Bs, St) -> + I = sub(I0, Bs), + case eval_bif(I, St) of + #b_literal{}=Val -> + eval_is(Is, bind_var(Dst, Val, Bs), St); + none -> + eval_is(Is, Bs, St) + end; +eval_is([#b_set{op=Op,dst=Dst}=I|Is], Bs, St) + when Op =:= is_tagged_tuple; Op =:= is_nonempty_list -> + #b_set{args=Args} = sub(I, Bs), + case eval_rel_op(Op, Args, St) of + #b_literal{}=Val -> + eval_is(Is, bind_var(Dst, Val, Bs), St); + none -> + eval_is(Is, Bs, St) + end; +eval_is([#b_set{}=I|Is], Bs, St) -> + case beam_ssa:no_side_effect(I) of + true -> + %% This instruction has no side effects. It can + %% safely be omitted. + eval_is(Is, Bs, St); + false -> + %% This instruction may have some side effect. + %% It is not safe to avoid this instruction. + none + end; +eval_is([], Bs, _St) -> Bs. + +eval_terminator(#b_br{bool=#b_var{}=Bool}=Br, Bs, _St) -> + Val = get_value(Bool, Bs), + beam_ssa:normalize(Br#b_br{bool=Val}); +eval_terminator(#b_br{bool=#b_literal{}}=Br, _Bs, _St) -> + beam_ssa:normalize(Br); +eval_terminator(#b_switch{arg=Arg,fail=Fail,list=List}=Sw, Bs, St) -> + case get_value(Arg, Bs) of + #b_literal{}=Val -> + %% Literal argument. Simplify to a `br`. + beam_ssa:normalize(Sw#b_switch{arg=Val}); + #b_var{} -> + %% Try optimizing the switch. + case eval_switch(List, Arg, St, Fail) of + none -> + none; + To when is_integer(To) -> + %% Either one of the values in the switch + %% matched a previous value in a '=:=' test, or + %% none of the values matched a previous test. + #b_br{bool=#b_literal{val=true},succ=To,fail=To} + end + end; +eval_terminator(#b_ret{}, _Bs, _St) -> + none. + +eval_switch(List, Arg, #st{rel_op={_,Arg,_}=PrevOp}, Fail) -> + %% There is a previous relational operator testing the same variable. + %% Optimization may be possible. + eval_switch_1(List, Arg, PrevOp, Fail); +eval_switch(_, _, _, _) -> + %% There is either no previous relational operator, or it tests + %% a different variable. Nothing to optimize. + none. + +eval_switch_1([{Lit,Lbl}|T], Arg, PrevOp, Fail) -> + RelOp = {'=:=',Arg,Lit}, + case will_succeed(PrevOp, RelOp) of + yes -> + %% Success. This branch will always be taken. + Lbl; + no -> + %% This branch will never be taken. + eval_switch_1(T, Arg, PrevOp, Fail); + maybe -> + %% This label could be reached. + eval_switch_1(T, Arg, PrevOp, none) + end; +eval_switch_1([], _Arg, _PrevOp, Fail) -> + %% Fail is now either the failure label or 'none'. + Fail. + +bind_var(Var, Val0, Bs) -> + Val = get_value(Val0, Bs), + Bs#{Var=>Val}. + +get_value(#b_var{}=Var, Bs) -> + case Bs of + #{Var:=Val} -> get_value(Val, Bs); + #{} -> Var + end; +get_value(#b_literal{}=Lit, _Bs) -> Lit. + +eval_bif(#b_set{op={bif,Bif},args=Args}, St) -> + Arity = length(Args), + case erl_bifs:is_pure(erlang, Bif, Arity) of + false -> + none; + true -> + case [Lit || #b_literal{val=Lit} <- Args] of + LitArgs when length(LitArgs) =:= Arity -> + try apply(erlang, Bif, LitArgs) of + Val -> #b_literal{val=Val} + catch + error:_ -> none + end; + _ -> + %% Not literal arguments. Try to evaluate + %% it based on a previous relational operator. + eval_rel_op({bif,Bif}, Args, St) + end + end. + +%%% +%%% Handling of relational operators. +%%% + +get_rel_op(Bool, [_|_]=Is) -> + case last(Is) of + #b_set{op=Op,dst=Bool,args=Args} -> + normalize_op(Op, Args); + #b_set{} -> + none + end; +get_rel_op(_, []) -> none. + +%% normalize_op(Instruction) -> {Normalized,FailLabel} | error +%% Normalized = {Operator,Variable,Variable|Literal} | +%% {TypeTest,Variable} +%% Operation = '<' | '=<' | '=:=' | '=/=' | '>=' | '>' +%% TypeTest = is_atom | is_integer ... +%% Variable = #b_var{} +%% Literal = #b_literal{} +%% +%% Normalize a relational operator to facilitate further +%% comparisons between operators. Always make the register +%% operand the first operand. If there are two registers, +%% order the registers in lexical order. +%% +%% For example, this instruction: +%% +%% #b_set{op={bif,=<},args=[#b_literal{}, #b_var{}} +%% +%% will be normalized to: +%% +%% {'=<',#b_var{},#b_literal{}} + +-spec normalize_op(Op, Args) -> NormalizedOp | 'none' when + Op :: beam_ssa:op(), + Args :: [beam_ssa:value()], + NormalizedOp :: basic_rel_op(). + +normalize_op(is_tagged_tuple, [Arg,#b_literal{val=Size},#b_literal{val=Tag}]) + when is_integer(Size), is_atom(Tag) -> + {{is_tagged_tuple,Size,Tag},Arg}; +normalize_op(is_nonempty_list, [Arg]) -> + {is_nonempty_list,Arg}; +normalize_op({bif,Bif}, [Arg]) -> + case erl_internal:new_type_test(Bif, 1) of + true -> {Bif,Arg}; + false -> none + end; +normalize_op({bif,Bif}, [_,_]=Args) -> + case erl_internal:comp_op(Bif, 2) of + true -> + normalize_op_1(Bif, Args); + false -> + none + end; +normalize_op(_, _) -> none. + +normalize_op_1(Bif, Args) -> + case Args of + [#b_literal{}=Arg1,#b_var{}=Arg2] -> + {turn_op(Bif),Arg2,Arg1}; + [#b_var{}=Arg1,#b_literal{}=Arg2] -> + {Bif,Arg1,Arg2}; + [#b_var{}=A,#b_var{}=B] -> + if A < B -> {Bif,A,B}; + true -> {turn_op(Bif),B,A} + end; + [#b_literal{},#b_literal{}] -> + none + end. + +-spec invert_op(basic_rel_op() | 'none') -> rel_op() | 'none'. + +invert_op({Op,Arg1,Arg2}) -> + {invert_op_1(Op),Arg1,Arg2}; +invert_op({TypeTest,Arg}) -> + {{'not',TypeTest},Arg}; +invert_op(none) -> none. + +invert_op_1('>=') -> '<'; +invert_op_1('<') -> '>='; +invert_op_1('=<') -> '>'; +invert_op_1('>') -> '=<'; +invert_op_1('=:=') -> '=/='; +invert_op_1('=/=') -> '=:='; +invert_op_1('==') -> '/='; +invert_op_1('/=') -> '=='. + +turn_op('<') -> '>'; +turn_op('=<') -> '>='; +turn_op('>') -> '<'; +turn_op('>=') -> '=<'; +turn_op('=:='=Op) -> Op; +turn_op('=/='=Op) -> Op; +turn_op('=='=Op) -> Op; +turn_op('/='=Op) -> Op. + +eval_rel_op(_Bif, _Args, #st{rel_op=none}) -> + none; +eval_rel_op(Bif, Args, #st{rel_op=Prev}) -> + case normalize_op(Bif, Args) of + none -> + none; + RelOp -> + case will_succeed(Prev, RelOp) of + yes -> #b_literal{val=true}; + no -> #b_literal{val=false}; + maybe -> none + end + end. + +%% will_succeed(PrevCondition, Condition) -> yes | no | maybe +%% PrevCondition is a condition known to be true. This function +%% will tell whether Condition will succeed. + +will_succeed({_Op,_Var,_Value}=Same, {_Op,_Var,_Value}=Same) -> + %% Repeated test. + yes; +will_succeed({Op1,Var,#b_literal{val=A}}, {Op2,Var,#b_literal{val=B}}) -> + will_succeed_1(Op1, A, Op2, B); +will_succeed({Op1,Var,#b_var{}=A}, {Op2,Var,#b_var{}=B}) -> + will_succeed_vars(Op1, A, Op2, B); +will_succeed({'=:=',Var,#b_literal{val=A}}, {TypeTest,Var}) -> + eval_type_test(TypeTest, A); +will_succeed({_,_}=Same, {_,_}=Same) -> + %% Repeated type test. + yes; +will_succeed({Test1,Var}, {Test2,Var}) -> + will_succeed_test(Test1, Test2); +will_succeed({_,_}, {_,_}) -> + maybe; +will_succeed({_,_}, {_,_,_}) -> + maybe; +will_succeed({_,_,_}, {_,_}) -> + maybe; +will_succeed({_,_,_}, {_,_,_}) -> + maybe. + +will_succeed_test({'not',Test1}, Test2) -> + case Test1 =:= Test2 of + true -> no; + false -> maybe + end; +will_succeed_test(is_tuple, {is_tagged_tuple,_,_}) -> + maybe; +will_succeed_test({is_tagged_tuple,_,_}, is_tuple) -> + yes; +will_succeed_test(is_list, is_nonempty_list) -> + maybe; +will_succeed_test(is_nonempty_list, is_list) -> + yes; +will_succeed_test(T1, T2) -> + case is_numeric_test(T1) andalso is_numeric_test(T2) of + true -> maybe; + false -> no + end. + +will_succeed_1('=:=', A, '<', B) -> + if + B =< A -> no; + true -> yes + end; +will_succeed_1('=:=', A, '=<', B) -> + if + B < A -> no; + true -> yes + end; +will_succeed_1('=:=', A, '=:=', B) when A =/= B -> + no; +will_succeed_1('=:=', A, '=/=', B) -> + if + A =:= B -> no; + true -> yes + end; +will_succeed_1('=:=', A, '>=', B) -> + if + B > A -> no; + true -> yes + end; +will_succeed_1('=:=', A, '>', B) -> + if + B >= A -> no; + true -> yes + end; + +will_succeed_1('=/=', A, '=:=', B) when A =:= B -> no; + +will_succeed_1('<', A, '=:=', B) when B >= A -> no; +will_succeed_1('<', A, '=/=', B) when B >= A -> yes; +will_succeed_1('<', A, '<', B) when B >= A -> yes; +will_succeed_1('<', A, '=<', B) when B > A -> yes; +will_succeed_1('<', A, '>=', B) when B > A -> no; +will_succeed_1('<', A, '>', B) when B >= A -> no; + +will_succeed_1('=<', A, '=:=', B) when B > A -> no; +will_succeed_1('=<', A, '=/=', B) when B > A -> yes; +will_succeed_1('=<', A, '<', B) when B > A -> yes; +will_succeed_1('=<', A, '=<', B) when B >= A -> yes; +will_succeed_1('=<', A, '>=', B) when B > A -> no; +will_succeed_1('=<', A, '>', B) when B >= A -> no; + +will_succeed_1('>=', A, '=:=', B) when B < A -> no; +will_succeed_1('>=', A, '=/=', B) when B < A -> yes; +will_succeed_1('>=', A, '<', B) when B =< A -> no; +will_succeed_1('>=', A, '=<', B) when B < A -> no; +will_succeed_1('>=', A, '>=', B) when B =< A -> yes; +will_succeed_1('>=', A, '>', B) when B < A -> yes; + +will_succeed_1('>', A, '=:=', B) when B =< A -> no; +will_succeed_1('>', A, '=/=', B) when B =< A -> yes; +will_succeed_1('>', A, '<', B) when B =< A -> no; +will_succeed_1('>', A, '=<', B) when B < A -> no; +will_succeed_1('>', A, '>=', B) when B =< A -> yes; +will_succeed_1('>', A, '>', B) when B < A -> yes; + +will_succeed_1('==', A, '==', B) -> + if + A == B -> yes; + true -> no + end; +will_succeed_1('==', A, '/=', B) -> + if + A == B -> no; + true -> yes + end; +will_succeed_1('/=', A, '/=', B) when A == B -> yes; +will_succeed_1('/=', A, '==', B) when A == B -> no; + +will_succeed_1(_, _, _, _) -> maybe. + +will_succeed_vars('=/=', Val, '=:=', Val) -> no; +will_succeed_vars('=:=', Val, '=/=', Val) -> no; +will_succeed_vars('=:=', Val, '>=', Val) -> yes; +will_succeed_vars('=:=', Val, '=<', Val) -> yes; + +will_succeed_vars('/=', Val1, '==', Val2) when Val1 == Val2 -> no; +will_succeed_vars('==', Val1, '/=', Val2) when Val1 == Val2 -> no; + +will_succeed_vars(_, _, _, _) -> maybe. + +is_numeric_test(is_float) -> true; +is_numeric_test(is_integer) -> true; +is_numeric_test(is_number) -> true; +is_numeric_test(_) -> false. + +eval_type_test(Test, Arg) -> + case eval_type_test_1(Test, Arg) of + true -> yes; + false -> no + end. + +eval_type_test_1(is_nonempty_list, Arg) -> + case Arg of + [_|_] -> true; + _ -> false + end; +eval_type_test_1({is_tagged_tuple,Sz,Tag}, Arg) -> + if + tuple_size(Arg) =:= Sz, element(1, Arg) =:= Tag -> + true; + true -> + false + end; +eval_type_test_1(Test, Arg) -> + erlang:Test(Arg). + +%%% +%%% Combine bif:'=:=' and switch instructions +%%% to switch instructions. +%%% +%%% Consider this code: +%%% +%%% 0: +%%% @ssa_bool = bif:'=:=' Var, literal 1 +%%% br @ssa_bool, label 2, label 3 +%%% +%%% 2: +%%% ret literal a +%%% +%%% 3: +%%% @ssa_bool:7 = bif:'=:=' Var, literal 2 +%%% br @ssa_bool:7, label 4, label 999 +%%% +%%% 4: +%%% ret literal b +%%% +%%% 999: +%%% . +%%% . +%%% . +%%% +%%% The two bif:'=:=' instructions can be combined +%%% to a switch: +%%% +%%% 0: +%%% switch Var, label 999, [ { literal 1, label 2 }, +%%% { literal 2, label 3 } ] +%%% +%%% 2: +%%% ret literal a +%%% +%%% 4: +%%% ret literal b +%%% +%%% 999: +%%% . +%%% . +%%% . +%%% + +combine_eqs(#st{bs=Blocks}=St) -> + Ls = reverse(beam_ssa:rpo(Blocks)), + combine_eqs_1(Ls, St). + +combine_eqs_1([L|Ls], #st{bs=Blocks0}=St0) -> + case comb_get_sw(L, St0) of + none -> + combine_eqs_1(Ls, St0); + {_,Arg,_,Fail0,List0} -> + case comb_get_sw(Fail0, St0) of + {true,Arg,Fail1,Fail,List1} -> + %% Another switch/br with the same arguments was + %% found. Try combining them. + case combine_lists(Fail1, List0, List1, Blocks0) of + none -> + %% Different types of literals in the lists, + %% or the success cases in the first switch + %% could branch to the second switch + %% (increasing code size and repeating tests). + combine_eqs_1(Ls, St0); + List -> + %% Everything OK! Combine the lists. + Sw0 = #b_switch{arg=Arg,fail=Fail,list=List}, + Sw = beam_ssa:normalize(Sw0), + Blk0 = map_get(L, Blocks0), + Blk = Blk0#b_blk{last=Sw}, + Blocks = Blocks0#{L:=Blk}, + St = St0#st{bs=Blocks}, + combine_eqs_1(Ls, St) + end; + {true,_OtherArg,_,_,_} -> + %% The other switch/br uses a different Arg. + combine_eqs_1(Ls, St0); + {false,_,_,_,_} -> + %% Not safe: Bindings of variables that will be used + %% or execution of instructions with potential + %% side effects will be skipped. + combine_eqs_1(Ls, St0); + none -> + %% No switch/br at this label. + combine_eqs_1(Ls, St0) + end + end; +combine_eqs_1([], St) -> St. + +comb_get_sw(L, Blocks) -> + comb_get_sw(L, true, Blocks). + +comb_get_sw(L, Safe0, #st{bs=Blocks,skippable=Skippable}) -> + #b_blk{is=Is,last=Last} = map_get(L, Blocks), + Safe1 = Safe0 andalso is_map_key(L, Skippable), + case Last of + #b_ret{} -> + none; + #b_br{bool=#b_var{}=Bool,succ=Succ,fail=Fail} -> + case comb_is(Is, Bool, Safe1) of + {none,_} -> + none; + {#b_set{op={bif,'=:='},args=[#b_var{}=Arg,#b_literal{}=Lit]},Safe} -> + {Safe,Arg,L,Fail,[{Lit,Succ}]}; + {#b_set{},_} -> + none + end; + #b_br{} -> + none; + #b_switch{arg=#b_var{}=Arg,fail=Fail,list=List} -> + {none,Safe} = comb_is(Is, none, Safe1), + {Safe,Arg,L,Fail,List} + end. + +comb_is([#b_set{dst=#b_var{}=Bool}=I], Bool, Safe) -> + {I,Safe}; +comb_is([#b_set{}=I|Is], Bool, Safe0) -> + Safe = Safe0 andalso beam_ssa:no_side_effect(I), + comb_is(Is, Bool, Safe); +comb_is([], _Bool, Safe) -> + {none,Safe}. + +%% combine_list(Fail, List1, List2, Blocks) -> List|none. +%% Try to combine two switch lists, returning the combined +%% list or 'none' if not possible. +%% +%% The values in the two lists must be all of the same type. +%% +%% The code reached from the labels in the first list must +%% not reach the failure label (if they do, tests could +%% be repeated). +%% + +combine_lists(Fail, L1, L2, Blocks) -> + Ls = beam_ssa:rpo([Lbl || {_,Lbl} <- L1], Blocks), + case member(Fail, Ls) of + true -> + %% One or more of labels in the first list + %% could reach the failure label. That + %% means that the second switch/br instruction + %% will be retained, increasing code size and + %% potentially also execution time. + none; + false -> + %% The combined switch will replace both original + %% br/switch instructions, leading to a reduction in code + %% size and potentially also in execution time. + combine_lists_1(L1, L2) + end. + +combine_lists_1(List0, List1) -> + case are_lists_compatible(List0, List1) of + true -> + First = maps:from_list(List0), + List0 ++ [{Val,Lbl} || {Val,Lbl} <- List1, + not is_map_key(Val, First)]; + false -> + none + end. + +are_lists_compatible([{#b_literal{val=Val1},_}|_], + [{#b_literal{val=Val2},_}|_]) -> + case lit_type(Val1) of + none -> false; + Type -> Type =:= lit_type(Val2) + end. + +lit_type(Val) -> + if + is_atom(Val) -> atom; + is_float(Val) -> float; + is_integer(Val) -> integer; + true -> none + end. + +%%% +%%% Calculate used variables for each block. +%%% + +used_vars(Linear) -> + used_vars(reverse(Linear), #{}, #{}). + +used_vars([{L,#b_blk{is=Is}=Blk}|Bs], UsedVars0, Skip0) -> + %% Calculate the variables used by each block and its + %% successors. This information is used by + %% shortcut_opt/1. + + Successors = beam_ssa:successors(Blk), + Used0 = used_vars_succ(Successors, L, UsedVars0, []), + Used = used_vars_blk(Blk, Used0), + UsedVars = used_vars_phis(Is, L, Used, UsedVars0), + + %% combine_eqs/1 needs different variable usage information than + %% shortcut_opt/1. The Skip map will have an entry for each block + %% that can be skipped (does not bind any variable used in + %% successor). This information is also useful for speeding up + %% shortcut_opt/1. + + Defined0 = [Def || #b_set{dst=Def} <- Is], + Defined = ordsets:from_list(Defined0), + MaySkip = ordsets:is_disjoint(Defined, Used0), + case MaySkip of + true -> + Skip = Skip0#{L=>true}, + used_vars(Bs, UsedVars, Skip); + false -> + used_vars(Bs, UsedVars, Skip0) + end; +used_vars([], UsedVars, Skip) -> + {UsedVars,Skip}. + +used_vars_succ([S|Ss], L, LiveMap, Live0) -> + Key = {S,L}, + case LiveMap of + #{Key:=Live} -> + %% The successor has a phi node, and the value for + %% this block in the phi node is a variable. + used_vars_succ(Ss, L, LiveMap, ordsets:union(Live, Live0)); + #{S:=Live} -> + %% No phi node in the successor, or the value for + %% this block in the phi node is a literal. + used_vars_succ(Ss, L, LiveMap, ordsets:union(Live, Live0)); + #{} -> + %% A peek_message block which has not been processed yet. + used_vars_succ(Ss, L, LiveMap, Live0) + end; +used_vars_succ([], _, _, Acc) -> Acc. + +used_vars_phis(Is, L, Live0, UsedVars0) -> + UsedVars = UsedVars0#{L=>Live0}, + Phis = takewhile(fun(#b_set{op=Op}) -> Op =:= phi end, Is), + case Phis of + [] -> + UsedVars; + [_|_] -> + PhiArgs = append([Args || #b_set{args=Args} <- Phis]), + case [{P,V} || {#b_var{}=V,P} <- PhiArgs] of + [_|_]=PhiVars -> + PhiLive0 = rel2fam(PhiVars), + PhiLive = [{{L,P},ordsets:union(ordsets:from_list(Vs), Live0)} || + {P,Vs} <- PhiLive0], + maps:merge(UsedVars, maps:from_list(PhiLive)); + [] -> + %% There were only literals in the phi node(s). + UsedVars + end + end. + +used_vars_blk(#b_blk{is=Is,last=Last}, Used0) -> + Used = ordsets:union(Used0, beam_ssa:used(Last)), + used_vars_is(reverse(Is), Used). + +used_vars_is([#b_set{op=phi}|Is], Used) -> + used_vars_is(Is, Used); +used_vars_is([#b_set{dst=Dst}=I|Is], Used0) -> + Used1 = ordsets:union(Used0, beam_ssa:used(I)), + Used = ordsets:del_element(Dst, Used1), + used_vars_is(Is, Used); +used_vars_is([], Used) -> + Used. + +%%% +%%% Common utilities. +%%% + +sub(#b_set{args=Args}=I, Sub) -> + I#b_set{args=[sub_arg(A, Sub) || A <- Args]}. + +sub_arg(Old, Sub) -> + case Sub of + #{Old:=New} -> New; + #{} -> Old + end. + +rel2fam(S0) -> + S1 = sofs:relation(S0), + S = sofs:rel2fam(S1), + sofs:to_external(S). diff --git a/lib/compiler/src/beam_ssa_funs.erl b/lib/compiler/src/beam_ssa_funs.erl new file mode 100644 index 0000000000..e77c00fa89 --- /dev/null +++ b/lib/compiler/src/beam_ssa_funs.erl @@ -0,0 +1,149 @@ +%% +%% %CopyrightBegin% +%% +%% Copyright Ericsson AB 2018. All Rights Reserved. +%% +%% Licensed under the Apache License, Version 2.0 (the "License"); +%% you may not use this file except in compliance with the License. +%% You may obtain a copy of the License at +%% +%% http://www.apache.org/licenses/LICENSE-2.0 +%% +%% Unless required by applicable law or agreed to in writing, software +%% distributed under the License is distributed on an "AS IS" BASIS, +%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +%% See the License for the specific language governing permissions and +%% limitations under the License. +%% +%% %CopyrightEnd% +%% + +%%% +%%% If a fun is defined locally and only used for calls, it can be replaced +%%% with direct calls to the relevant function. This greatly speeds up "named +%%% functions" (which rely on make_fun to recreate themselves) and macros that +%%% wrap their body in a fun. +%%% + +-module(beam_ssa_funs). + +-export([module/2]). + +-include("beam_ssa.hrl"). + +-import(lists, [foldl/3]). + +-spec module(Module, Options) -> Result when + Module :: beam_ssa:b_module(), + Options :: [compile:option()], + Result :: {ok, beam_ssa:b_module()}. + +module(#b_module{body=Fs0}=Module, _Opts) -> + Trampolines = foldl(fun find_trampolines/2, #{}, Fs0), + Fs = [lfo(F, Trampolines) || F <- Fs0], + {ok, Module#b_module{body=Fs}}. + +%% If a function does absolutely nothing beyond calling another function with +%% the same arguments in the same order, we can shave off a call by short- +%% circuiting it. +find_trampolines(#b_function{args=Args,bs=Blocks}=F, Trampolines) -> + case map_get(0, Blocks) of + #b_blk{is=[#b_set{op=call, + args=[#b_local{}=Actual | Args], + dst=Dst}], + last=#b_ret{arg=Dst}} -> + {_, Name, Arity} = beam_ssa:get_anno(func_info, F), + Trampoline = #b_local{name=#b_literal{val=Name},arity=Arity}, + Trampolines#{Trampoline => Actual}; + _ -> + Trampolines + end. + +lfo(#b_function{bs=Blocks0}=F, Trampolines) -> + Linear0 = beam_ssa:linearize(Blocks0), + Linear = lfo_optimize(Linear0, lfo_analyze(Linear0, #{}), Trampolines), + F#b_function{bs=maps:from_list(Linear)}. + +%% Gather a map of the locally defined funs that are only used for calls. +lfo_analyze([{_L,#b_blk{is=Is,last=Last}}|Bs], LFuns0) -> + LFuns = lfo_analyze_last(Last, lfo_analyze_is(Is, LFuns0)), + lfo_analyze(Bs, LFuns); +lfo_analyze([], LFuns) -> + LFuns. + +lfo_analyze_is([#b_set{op=make_fun, + dst=Dst, + args=[#b_local{} | FreeVars]}=Def | Is], + LFuns0) -> + LFuns = maps:put(Dst, Def, maps:without(FreeVars, LFuns0)), + lfo_analyze_is(Is, LFuns); +lfo_analyze_is([#b_set{op=call, + args=[Fun | CallArgs]} | Is], + LFuns) when is_map_key(Fun, LFuns) -> + #b_set{args=[#b_local{arity=Arity} | FreeVars]} = map_get(Fun, LFuns), + case length(CallArgs) + length(FreeVars) of + Arity -> + lfo_analyze_is(Is, maps:without(CallArgs, LFuns)); + _ -> + %% This will `badarity` at runtime, and it's easier to disable the + %% optimization than to simulate it. + lfo_analyze_is(Is, maps:without([Fun | CallArgs], LFuns)) + end; +lfo_analyze_is([#b_set{args=Args} | Is], LFuns) when map_size(LFuns) =/= 0 -> + %% We disqualify funs that are used outside calls because this forces them + %% to be created anyway, and the slight performance gain from direct calls + %% is not enough to offset the potential increase in stack frame size (the + %% free variables need to be kept alive until the call). + %% + %% This is also a kludge to make HiPE work, as the latter will generate + %% code with the assumption that the functions referenced in a make_fun + %% will only be used by funs, which will not be the case if we mix it with + %% direct calls. See cerl_cconv.erl for details. + %% + %% Future optimizations like delaying fun creation until use may require us + %% to copy affected functions so that HiPE gets its own to play with (until + %% HiPE is fixed anyway). + lfo_analyze_is(Is, maps:without(Args, LFuns)); +lfo_analyze_is([_ | Is], LFuns) -> + lfo_analyze_is(Is, LFuns); +lfo_analyze_is([], LFuns) -> + LFuns. + +lfo_analyze_last(#b_switch{arg=Arg}, LFuns) -> + maps:remove(Arg, LFuns); +lfo_analyze_last(#b_ret{arg=Arg}, LFuns) -> + maps:remove(Arg, LFuns); +lfo_analyze_last(_, LFuns) -> + LFuns. + +%% Replace all calls of suitable funs with a direct call to their +%% implementation. Liveness optimization will get rid of the make_fun +%% instruction. +lfo_optimize(Linear, LFuns, _Trampolines) when map_size(LFuns) =:= 0 -> + Linear; +lfo_optimize(Linear, LFuns, Trampolines) -> + lfo_optimize_1(Linear, LFuns, Trampolines). + +lfo_optimize_1([{L,#b_blk{is=Is0}=Blk}|Bs], LFuns, Trampolines) -> + Is = lfo_optimize_is(Is0, LFuns, Trampolines), + [{L,Blk#b_blk{is=Is}} | lfo_optimize_1(Bs, LFuns, Trampolines)]; +lfo_optimize_1([], _LFuns, _Trampolines) -> + []. + +lfo_optimize_is([#b_set{op=call, + args=[Fun | CallArgs]}=Call0 | Is], + LFuns, Trampolines) when is_map_key(Fun, LFuns) -> + #b_set{args=[Local | FreeVars]} = map_get(Fun, LFuns), + Args = [lfo_short_circuit(Local, Trampolines) | CallArgs ++ FreeVars], + Call = beam_ssa:add_anno(local_fun_opt, Fun, Call0#b_set{args=Args}), + [Call | lfo_optimize_is(Is, LFuns, Trampolines)]; +lfo_optimize_is([I | Is], LFuns, Trampolines) -> + [I | lfo_optimize_is(Is, LFuns, Trampolines)]; +lfo_optimize_is([], _LFuns, _Trampolines) -> + []. + +lfo_short_circuit(Call, Trampolines) -> + case maps:find(Call, Trampolines) of + {ok, Other} -> lfo_short_circuit(Other, Trampolines); + error -> Call + end. diff --git a/lib/compiler/src/beam_ssa_opt.erl b/lib/compiler/src/beam_ssa_opt.erl index da466a3316..ca5eefe4fc 100644 --- a/lib/compiler/src/beam_ssa_opt.erl +++ b/lib/compiler/src/beam_ssa_opt.erl @@ -18,49 +18,165 @@ %% %CopyrightEnd% %% +%%% +%%% This is a collection of various optimizations that don't need a separate +%%% pass by themselves and/or are mutually beneficial to other passes. +%%% +%%% The optimizations are applied in "phases," each with a list of sub-passes +%%% to run. These sub-passes are applied on all functions in a module before +%%% moving on to the next phase, which lets us gather module-level information +%%% in one phase and then apply it in the next without having to risk working +%%% with incomplete information. +%%% +%%% Each sub-pass operates on a #st{} record and a func_info_db(), where the +%%% former is just a #b_function{} whose blocks can be represented either in +%%% linear or map form, and the latter is a map with information about all +%%% functions in the module (see beam_ssa_opt.hrl for more details). +%%% + -module(beam_ssa_opt). -export([module/2]). --include("beam_ssa.hrl"). --import(lists, [all/2,append/1,foldl/3,keyfind/3,member/2,reverse/1,reverse/2, - splitwith/2,takewhile/2,unzip/1]). +-include("beam_ssa_opt.hrl"). + +-import(lists, [all/2,append/1,duplicate/2,foldl/3,keyfind/3,member/2, + reverse/1,reverse/2, + splitwith/2,sort/1,takewhile/2,unzip/1]). + +-define(DEFAULT_REPETITIONS, 2). -spec module(beam_ssa:b_module(), [compile:option()]) -> {'ok',beam_ssa:b_module()}. -module(#b_module{body=Fs0}=Module, Opts) -> - Ps = passes(Opts), - Fs = functions(Fs0, Ps), - {ok,Module#b_module{body=Fs}}. +-record(st, {ssa :: [{beam_ssa:label(),beam_ssa:b_blk()}] | + beam_ssa:block_map(), + args :: [beam_ssa:b_var()], + cnt :: beam_ssa:label(), + anno :: beam_ssa:anno()}). +-type st_map() :: #{ func_id() => #st{} }. + +module(Module, Opts) -> + FuncDb0 = case proplists:get_value(no_module_opt, Opts, false) of + false -> build_func_db(Module); + true -> #{} + end, -functions([F|Fs], Ps) -> - [function(F, Ps)|functions(Fs, Ps)]; -functions([], _Ps) -> []. + %% Passes that perform module-level optimizations are often aided by + %% optimizing callers before callees and vice versa, so we optimize all + %% functions in call order, flipping it as required. + StMap0 = build_st_map(Module), + Order = get_call_order_po(StMap0, FuncDb0), --type b_blk() :: beam_ssa:b_blk(). --type b_var() :: beam_ssa:b_var(). --type label() :: beam_ssa:label(). + Phases = + [{Order, prologue_passes(Opts)}] ++ + repeat(Opts, repeated_passes(Opts), Order) ++ + [{Order, epilogue_passes(Opts)}], + + {StMap, _FuncDb} = foldl(fun({FuncIds, Ps}, {StMap, FuncDb}) -> + phase(FuncIds, Ps, StMap, FuncDb) + end, {StMap0, FuncDb0}, Phases), + + {ok, finish(Module, StMap)}. + +phase([FuncId | Ids], Ps, StMap, FuncDb0) -> + try compile:run_sub_passes(Ps, {map_get(FuncId, StMap), FuncDb0}) of + {St, FuncDb} -> + phase(Ids, Ps, StMap#{ FuncId => St }, FuncDb) + catch + Class:Error:Stack -> + #b_local{name=Name,arity=Arity} = FuncId, + io:fwrite("Function: ~w/~w\n", [Name,Arity]), + erlang:raise(Class, Error, Stack) + end; +phase([], _Ps, StMap, FuncDb) -> + {StMap, FuncDb}. + +%% Repeats the given passes, alternating the order between runs to make the +%% type pass more efficient. +repeat(Opts, Ps, OrderA) -> + Repeat = proplists:get_value(ssa_opt_repeat, Opts, ?DEFAULT_REPETITIONS), + OrderB = reverse(OrderA), + repeat_1(Repeat, Ps, OrderA, OrderB). + +repeat_1(0, _Opts, _OrderA, _OrderB) -> + []; +repeat_1(N, Ps, OrderA, OrderB) when N > 0, N rem 2 =:= 0 -> + [{OrderA, Ps} | repeat_1(N - 1, Ps, OrderA, OrderB)]; +repeat_1(N, Ps, OrderA, OrderB) when N > 0, N rem 2 =:= 1 -> + [{OrderB, Ps} | repeat_1(N - 1, Ps, OrderA, OrderB)]. + +%% + +get_func_id(F) -> + {_Mod, Name, Arity} = beam_ssa:get_anno(func_info, F), + #b_local{name=#b_literal{val=Name}, arity=Arity}. + +-spec build_st_map(#b_module{}) -> st_map(). +build_st_map(#b_module{body=Fs}) -> + build_st_map_1(Fs, #{}). + +build_st_map_1([F | Fs], Map) -> + #b_function{anno=Anno,args=Args,cnt=Counter,bs=Bs} = F, + St = #st{anno=Anno,args=Args,cnt=Counter,ssa=Bs}, + build_st_map_1(Fs, Map#{ get_func_id(F) => St }); +build_st_map_1([], Map) -> + Map. + +-spec finish(#b_module{}, st_map()) -> #b_module{}. +finish(#b_module{body=Fs0}=Module, StMap) -> + Module#b_module{body=finish_1(Fs0, StMap)}. + +finish_1([F0 | Fs], StMap) -> + #st{anno=Anno,cnt=Counter,ssa=Blocks} = map_get(get_func_id(F0), StMap), + F = F0#b_function{anno=Anno,bs=Blocks,cnt=Counter}, + [F | finish_1(Fs, StMap)]; +finish_1([], _StMap) -> + []. + +%% --record(st, {ssa :: beam_ssa:block_map() | [{label(),b_blk()}], - args :: [b_var()], - cnt :: label()}). -define(PASS(N), {N,fun N/1}). -passes(Opts0) -> +prologue_passes(Opts) -> Ps = [?PASS(ssa_opt_split_blocks), + ?PASS(ssa_opt_coalesce_phis), + ?PASS(ssa_opt_tail_phis), ?PASS(ssa_opt_element), ?PASS(ssa_opt_linearize), + ?PASS(ssa_opt_tuple_size), ?PASS(ssa_opt_record), + ?PASS(ssa_opt_cse), %Helps the first type pass. + ?PASS(ssa_opt_type_start)], + passes_1(Ps, Opts). + +%% These passes all benefit from each other (in roughly this order), so they +%% are repeated as required. +repeated_passes(Opts) -> + Ps = [?PASS(ssa_opt_live), + ?PASS(ssa_opt_bs_puts), + ?PASS(ssa_opt_dead), ?PASS(ssa_opt_cse), - ?PASS(ssa_opt_type), + ?PASS(ssa_opt_tail_phis), + ?PASS(ssa_opt_type_continue)], %Must run after ssa_opt_dead to + %clean up phi nodes. + passes_1(Ps, Opts). + +epilogue_passes(Opts) -> + Ps = [?PASS(ssa_opt_type_finish), ?PASS(ssa_opt_float), - ?PASS(ssa_opt_live), + ?PASS(ssa_opt_live), %One last time to clean up the + %mess left by the float pass. ?PASS(ssa_opt_bsm), + ?PASS(ssa_opt_bsm_units), ?PASS(ssa_opt_bsm_shortcut), - ?PASS(ssa_opt_misc), + ?PASS(ssa_opt_sw), ?PASS(ssa_opt_blockify), ?PASS(ssa_opt_sink), - ?PASS(ssa_opt_merge_blocks)], + ?PASS(ssa_opt_merge_blocks), + ?PASS(ssa_opt_trim_unreachable)], + passes_1(Ps, Opts). + +passes_1(Ps, Opts0) -> Negations = [{list_to_atom("no_"++atom_to_list(N)),N} || {N,_} <- Ps], Opts = proplists:substitute_negations(Negations, Opts0), @@ -72,30 +188,132 @@ passes(Opts0) -> {NoName,fun(S) -> S end} end || {Name,_}=P <- Ps]. -function(#b_function{anno=Anno,bs=Blocks0,args=Args,cnt=Count0}=F, Ps) -> +%% Builds a function information map with basic information about incoming and +%% outgoing local calls, as well as whether the function is exported. +-spec build_func_db(#b_module{}) -> func_info_db(). +build_func_db(#b_module{body=Fs,exports=Exports}) -> try - St = #st{ssa=Blocks0,args=Args,cnt=Count0}, - #st{ssa=Blocks,cnt=Count} = compile:run_sub_passes(Ps, St), - F#b_function{bs=Blocks,cnt=Count} + fdb_1(Fs, gb_sets:from_list(Exports), #{}) catch - Class:Error:Stack -> - #{func_info:={_,Name,Arity}} = Anno, - io:fwrite("Function: ~w/~w\n", [Name,Arity]), - erlang:raise(Class, Error, Stack) + %% All module-level optimizations are invalid when a NIF can override a + %% function, so we have to bail out. + throw:load_nif -> #{} end. +fdb_1([#b_function{ args=Args,bs=Bs }=F | Fs], Exports, FuncDb0) -> + Id = get_func_id(F), + + #b_local{name=#b_literal{val=Name}, arity=Arity} = Id, + Exported = gb_sets:is_element({Name, Arity}, Exports), + ArgTypes = duplicate(length(Args), #{}), + + FuncDb1 = case FuncDb0 of + %% We may have an entry already if someone's called us. + #{ Id := Info } -> + FuncDb0#{ Id := Info#func_info{ exported=Exported, + arg_types=ArgTypes }}; + #{} -> + FuncDb0#{ Id => #func_info{ exported=Exported, + arg_types=ArgTypes }} + end, + + FuncDb = beam_ssa:fold_rpo(fun(_L, #b_blk{is=Is}, FuncDb) -> + fdb_is(Is, Id, FuncDb) + end, FuncDb1, Bs), + + fdb_1(Fs, Exports, FuncDb); +fdb_1([], _Exports, FuncDb) -> + FuncDb. + +fdb_is([#b_set{op=call, + args=[#b_local{}=Callee | _]} | Is], + Caller, FuncDb) -> + fdb_is(Is, Caller, fdb_update(Caller, Callee, FuncDb)); +fdb_is([#b_set{op=call, + args=[#b_remote{mod=#b_literal{val=erlang}, + name=#b_literal{val=load_nif}}, + _Path, _LoadInfo]} | _Is], _Caller, _FuncDb) -> + throw(load_nif); +fdb_is([_ | Is], Caller, FuncDb) -> + fdb_is(Is, Caller, FuncDb); +fdb_is([], _Caller, FuncDb) -> + FuncDb. + +fdb_update(Caller, Callee, FuncDb) -> + CallerVertex = maps:get(Caller, FuncDb, #func_info{}), + CalleeVertex = maps:get(Callee, FuncDb, #func_info{}), + + Calls = ordsets:add_element(Callee, CallerVertex#func_info.out), + CalledBy = ordsets:add_element(Caller, CalleeVertex#func_info.in), + + FuncDb#{ Caller => CallerVertex#func_info{out=Calls}, + Callee => CalleeVertex#func_info{in=CalledBy} }. + +%% Returns the post-order of all local calls in this module. That is, it starts +%% with the functions that don't call any others and then walks up the call +%% chain. +%% +%% Functions where module-level optimization is disabled are added last in +%% arbitrary order. + +get_call_order_po(StMap, FuncDb) -> + Leaves = maps:fold(fun(Id, #func_info{out=[]}, Acc) -> + [Id | Acc]; + (_, _, Acc) -> + Acc + end, [], FuncDb), + + Order = gco_po_1(sort(Leaves), FuncDb, [], #{}), + + Order ++ maps:fold(fun(K, _V, Acc) -> + case is_map_key(K, FuncDb) of + false -> [K | Acc]; + true -> Acc + end + end, [], StMap). + +gco_po_1([Id | Ids], FuncDb, Children, Seen) when not is_map_key(Id, Seen) -> + [Id | gco_po_1(Ids, FuncDb, [Id | Children], Seen#{ Id => true })]; +gco_po_1([_Id | Ids], FuncDb, Children, Seen) -> + gco_po_1(Ids, FuncDb, Children, Seen); +gco_po_1([], FuncDb, [_|_]=Children, Seen) -> + gco_po_1(gco_po_parents(Children, FuncDb), FuncDb, [], Seen); +gco_po_1([], _FuncDb, [], _Seen) -> + []. + +gco_po_parents([Child | Children], FuncDb) -> + #{ Child := #func_info{in=Parents}} = FuncDb, + Parents ++ gco_po_parents(Children, FuncDb); +gco_po_parents([], _FuncDb) -> + []. + %%% %%% Trivial sub passes. %%% -ssa_opt_linearize(#st{ssa=Blocks}=St) -> - St#st{ssa=beam_ssa:linearize(Blocks)}. +ssa_opt_dead({#st{ssa=Linear}=St, FuncDb}) -> + {St#st{ssa=beam_ssa_dead:opt(Linear)}, FuncDb}. -ssa_opt_type(#st{ssa=Linear,args=Args}=St) -> - St#st{ssa=beam_ssa_type:opt(Linear, Args)}. +ssa_opt_linearize({#st{ssa=Blocks}=St, FuncDb}) -> + {St#st{ssa=beam_ssa:linearize(Blocks)}, FuncDb}. -ssa_opt_blockify(#st{ssa=Linear}=St) -> - St#st{ssa=maps:from_list(Linear)}. +ssa_opt_type_start({#st{ssa=Linear0,args=Args,anno=Anno}=St0, FuncDb0}) -> + {Linear, FuncDb} = beam_ssa_type:opt_start(Linear0, Args, Anno, FuncDb0), + {St0#st{ssa=Linear}, FuncDb}. + +ssa_opt_type_continue({#st{ssa=Linear0,args=Args,anno=Anno}=St0, FuncDb0}) -> + {Linear, FuncDb} = beam_ssa_type:opt_continue(Linear0, Args, Anno, FuncDb0), + {St0#st{ssa=Linear}, FuncDb}. + +ssa_opt_type_finish({#st{args=Args,anno=Anno0}=St0, FuncDb0}) -> + {Anno, FuncDb} = beam_ssa_type:opt_finish(Args, Anno0, FuncDb0), + {St0#st{anno=Anno}, FuncDb}. + +ssa_opt_blockify({#st{ssa=Linear}=St, FuncDb}) -> + {St#st{ssa=maps:from_list(Linear)}, FuncDb}. + +ssa_opt_trim_unreachable({#st{ssa=Blocks}=St, FuncDb}) -> + {St#st{ssa=beam_ssa:trim_unreachable(Blocks)}, FuncDb}. %%% %%% Split blocks before certain instructions to enable more optimizations. @@ -107,14 +325,264 @@ ssa_opt_blockify(#st{ssa=Linear}=St) -> %%% for sinking get_tuple_element instructions. %%% -ssa_opt_split_blocks(#st{ssa=Blocks0,cnt=Count0}=St) -> +ssa_opt_split_blocks({#st{ssa=Blocks0,cnt=Count0}=St, FuncDb}) -> P = fun(#b_set{op={bif,element}}) -> true; (#b_set{op=call}) -> true; (#b_set{op=make_fun}) -> true; (_) -> false end, {Blocks,Count} = beam_ssa:split_blocks(P, Blocks0, Count0), - St#st{ssa=Blocks,cnt=Count}. + {St#st{ssa=Blocks,cnt=Count}, FuncDb}. + +%%% +%%% Coalesce phi nodes. +%%% +%%% Nested cases can led to code such as this: +%%% +%%% 10: +%%% _1 = phi {literal value1, label 8}, {Var, label 9} +%%% br 11 +%%% +%%% 11: +%%% _2 = phi {_1, label 10}, {literal false, label 3} +%%% +%%% The phi nodes can be coalesced like this: +%%% +%%% 11: +%%% _2 = phi {literal value1, label 8}, {Var, label 9}, {literal false, label 3} +%%% +%%% Coalescing can help other optimizations, and can in some cases reduce register +%%% shuffling (if the phi variables for two phi nodes happens to be allocated to +%%% different registers). +%%% + +ssa_opt_coalesce_phis({#st{ssa=Blocks0}=St, FuncDb}) -> + Ls = beam_ssa:rpo(Blocks0), + Blocks = c_phis_1(Ls, Blocks0), + {St#st{ssa=Blocks}, FuncDb}. + +c_phis_1([L|Ls], Blocks0) -> + case map_get(L, Blocks0) of + #b_blk{is=[#b_set{op=phi}|_]}=Blk -> + Blocks = c_phis_2(L, Blk, Blocks0), + c_phis_1(Ls, Blocks); + #b_blk{} -> + c_phis_1(Ls, Blocks0) + end; +c_phis_1([], Blocks) -> Blocks. + +c_phis_2(L, #b_blk{is=Is0}=Blk0, Blocks0) -> + case c_phis_args(Is0, Blocks0) of + none -> + Blocks0; + {_,_,Preds}=Info -> + Is = c_rewrite_phis(Is0, Info), + Blk = Blk0#b_blk{is=Is}, + Blocks = Blocks0#{L:=Blk}, + c_fix_branches(Preds, L, Blocks) + end. + +c_phis_args([#b_set{op=phi,args=Args0}|Is], Blocks) -> + case c_phis_args_1(Args0, Blocks) of + none -> + c_phis_args(Is, Blocks); + Res -> + Res + end; +c_phis_args(_, _Blocks) -> none. + +c_phis_args_1([{Var,Pred}|As], Blocks) -> + case c_get_pred_vars(Var, Pred, Blocks) of + none -> + c_phis_args_1(As, Blocks); + Result -> + Result + end; +c_phis_args_1([], _Blocks) -> none. + +c_get_pred_vars(Var, Pred, Blocks) -> + case map_get(Pred, Blocks) of + #b_blk{is=[#b_set{op=phi,dst=Var,args=Args}]} -> + {Var,Pred,Args}; + #b_blk{} -> + none + end. + +c_rewrite_phis([#b_set{op=phi,args=Args0}=I|Is], Info) -> + Args = c_rewrite_phi(Args0, Info), + [I#b_set{args=Args}|c_rewrite_phis(Is, Info)]; +c_rewrite_phis(Is, _Info) -> Is. + +c_rewrite_phi([{Var,Pred}|As], {Var,Pred,Values}) -> + Values ++ As; +c_rewrite_phi([{Value,Pred}|As], {_,Pred,Values}) -> + [{Value,P} || {_,P} <- Values] ++ As; +c_rewrite_phi([A|As], Info) -> + [A|c_rewrite_phi(As, Info)]; +c_rewrite_phi([], _Info) -> []. + +c_fix_branches([{_,Pred}|As], L, Blocks0) -> + #b_blk{last=Last0} = Blk0 = map_get(Pred, Blocks0), + #b_br{bool=#b_literal{val=true}} = Last0, %Assertion. + Last = Last0#b_br{bool=#b_literal{val=true},succ=L,fail=L}, + Blk = Blk0#b_blk{last=Last}, + Blocks = Blocks0#{Pred:=Blk}, + c_fix_branches(As, L, Blocks); +c_fix_branches([], _, Blocks) -> Blocks. + +%%% +%%% Eliminate phi nodes in the tail of a function. +%%% +%%% Try to eliminate short blocks that starts with a phi node +%%% and end in a return. For example: +%%% +%%% Result = phi { Res1, 4 }, { literal true, 5 } +%%% Ret = put_tuple literal ok, Result +%%% ret Ret +%%% +%%% The code in this block can be inserted at the end blocks 4 and +%%% 5. Thus, the following code can be inserted into block 4: +%%% +%%% Ret:1 = put_tuple literal ok, Res1 +%%% ret Ret:1 +%%% +%%% And the following code into block 5: +%%% +%%% Ret:2 = put_tuple literal ok, literal true +%%% ret Ret:2 +%%% +%%% Which can be further simplified to: +%%% +%%% ret literal {ok, true} +%%% +%%% This transformation may lead to more code improvements: +%%% +%%% - Stack trimming +%%% - Fewer test_heap instructions +%%% - Smaller stack frames +%%% + +ssa_opt_tail_phis({#st{ssa=SSA0,cnt=Count0}=St, FuncDb}) -> + {SSA,Count} = opt_tail_phis(SSA0, Count0), + {St#st{ssa=SSA,cnt=Count}, FuncDb}. + +opt_tail_phis(Blocks, Count) when is_map(Blocks) -> + opt_tail_phis(maps:values(Blocks), Blocks, Count); +opt_tail_phis(Linear0, Count0) when is_list(Linear0) -> + Blocks0 = maps:from_list(Linear0), + {Blocks,Count} = opt_tail_phis(Blocks0, Count0), + {beam_ssa:linearize(Blocks),Count}. + +opt_tail_phis([#b_blk{is=Is0,last=Last}|Bs], Blocks0, Count0) -> + case {Is0,Last} of + {[#b_set{op=phi,args=[_,_|_]}|_],#b_ret{arg=#b_var{}}=Ret} -> + {Phis,Is} = splitwith(fun(#b_set{op=Op}) -> Op =:= phi end, Is0), + case suitable_tail_ops(Is) of + true -> + {Blocks,Count} = opt_tail_phi(Phis, Is, Ret, + Blocks0, Count0), + opt_tail_phis(Bs, Blocks, Count); + false -> + opt_tail_phis(Bs, Blocks0, Count0) + end; + {_,_} -> + opt_tail_phis(Bs, Blocks0, Count0) + end; +opt_tail_phis([], Blocks, Count) -> + {Blocks,Count}. + +opt_tail_phi(Phis0, Is, Ret, Blocks0, Count0) -> + Phis = rel2fam(reduce_phis(Phis0)), + {Blocks,Count,Cost} = + foldl(fun(PhiArg, Acc) -> + opt_tail_phi_arg(PhiArg, Is, Ret, Acc) + end, {Blocks0,Count0,0}, Phis), + MaxCost = length(Phis) * 3 + 2, + if + Cost =< MaxCost -> + %% The transformation would cause at most a slight + %% increase in code size if no more optimizations + %% can be applied. + {Blocks,Count}; + true -> + %% The code size would be increased too much. + {Blocks0,Count0} + end. + +reduce_phis([#b_set{dst=PhiDst,args=PhiArgs}|Is]) -> + [{L,{PhiDst,Val}} || {Val,L} <- PhiArgs] ++ reduce_phis(Is); +reduce_phis([]) -> []. + +opt_tail_phi_arg({PredL,Sub0}, Is0, Ret0, {Blocks0,Count0,Cost0}) -> + Blk0 = map_get(PredL, Blocks0), + #b_blk{is=IsPrefix,last=#b_br{succ=Next,fail=Next}} = Blk0, + case is_exit_bif(IsPrefix) of + false -> + Sub1 = maps:from_list(Sub0), + {Is1,Count,Sub} = new_names(Is0, Sub1, Count0, []), + Is2 = [sub(I, Sub) || I <- Is1], + Cost = build_cost(Is2, Cost0), + Is = IsPrefix ++ Is2, + Ret = sub(Ret0, Sub), + Blk = Blk0#b_blk{is=Is,last=Ret}, + Blocks = Blocks0#{PredL:=Blk}, + {Blocks,Count,Cost}; + true -> + %% The block ends in a call to a function that + %% will cause an exception. + {Blocks0,Count0,Cost0+3} + end. + +is_exit_bif([#b_set{op=call, + args=[#b_remote{mod=#b_literal{val=Mod}, + name=#b_literal{val=Name}}|Args]}]) -> + erl_bifs:is_exit_bif(Mod, Name, length(Args)); +is_exit_bif(_) -> false. + +new_names([#b_set{dst=Dst}=I|Is], Sub0, Count0, Acc) -> + {NewDst,Count} = new_var(Dst, Count0), + Sub = Sub0#{Dst=>NewDst}, + new_names(Is, Sub, Count, [I#b_set{dst=NewDst}|Acc]); +new_names([], Sub, Count, Acc) -> + {reverse(Acc),Count,Sub}. + +suitable_tail_ops(Is) -> + all(fun(#b_set{op=Op}) -> + is_suitable_tail_op(Op) + end, Is). + +is_suitable_tail_op({bif,_}) -> true; +is_suitable_tail_op(put_list) -> true; +is_suitable_tail_op(put_tuple) -> true; +is_suitable_tail_op(_) -> false. + +build_cost([#b_set{op=put_list,args=Args}|Is], Cost) -> + case are_all_literals(Args) of + true -> + build_cost(Is, Cost); + false -> + build_cost(Is, Cost + 1) + end; +build_cost([#b_set{op=put_tuple,args=Args}|Is], Cost) -> + case are_all_literals(Args) of + true -> + build_cost(Is, Cost); + false -> + build_cost(Is, Cost + length(Args) + 1) + end; +build_cost([#b_set{op={bif,_},args=Args}|Is], Cost) -> + case are_all_literals(Args) of + true -> + build_cost(Is, Cost); + false -> + build_cost(Is, Cost + 1) + end; +build_cost([], Cost) -> Cost. + +are_all_literals(Args) -> + all(fun(#b_literal{}) -> true; + (_) -> false + end, Args). %%% %%% Order element/2 calls. @@ -125,7 +593,7 @@ ssa_opt_split_blocks(#st{ssa=Blocks0,cnt=Count0}=St) -> %%% be replaced with get_tuple_element/3 instructions. %%% -ssa_opt_element(#st{ssa=Blocks}=St) -> +ssa_opt_element({#st{ssa=Blocks}=St, FuncDb}) -> %% Collect the information about element instructions in this %% function. GetEls = collect_element_calls(beam_ssa:linearize(Blocks)), @@ -137,7 +605,7 @@ ssa_opt_element(#st{ssa=Blocks}=St) -> %% For each chain, swap the first element call with the %% element call with the highest index. - St#st{ssa=swap_element_calls(Chains, Blocks)}. + {St#st{ssa=swap_element_calls(Chains, Blocks)}, FuncDb}. collect_element_calls([{L,#b_blk{is=Is0,last=Last}}|Bs]) -> case {Is0,Last} of @@ -198,9 +666,9 @@ swap_element_calls_1([], _, Blocks) -> %%% when applicable. %%% -ssa_opt_record(#st{ssa=Linear}=St) -> +ssa_opt_record({#st{ssa=Linear}=St, FuncDb}) -> Blocks = maps:from_list(Linear), - St#st{ssa=record_opt(Linear, Blocks)}. + {St#st{ssa=record_opt(Linear, Blocks)}, FuncDb}. record_opt([{L,#b_blk{is=Is0,last=Last}=Blk0}|Bs], Blocks) -> Is = record_opt_is(Is0, Last, Blocks), @@ -208,8 +676,7 @@ record_opt([{L,#b_blk{is=Is0,last=Last}=Blk0}|Bs], Blocks) -> [{L,Blk}|record_opt(Bs, Blocks)]; record_opt([], _Blocks) -> []. -record_opt_is([#b_set{op={bif,is_tuple},dst=#b_var{name=Bool}, - args=[Tuple]}=Set], +record_opt_is([#b_set{op={bif,is_tuple},dst=Bool,args=[Tuple]}=Set], Last, Blocks) -> case is_tagged_tuple(Tuple, Bool, Last, Blocks) of {yes,Size,Tag} -> @@ -222,25 +689,24 @@ record_opt_is([I|Is], Last, Blocks) -> [I|record_opt_is(Is, Last, Blocks)]; record_opt_is([], _Last, _Blocks) -> []. -is_tagged_tuple(#b_var{name=Tuple}, Bool, - #b_br{bool=#b_var{name=Bool},succ=Succ,fail=Fail}, +is_tagged_tuple(#b_var{}=Tuple, Bool, + #b_br{bool=Bool,succ=Succ,fail=Fail}, Blocks) -> - SuccBlk = maps:get(Succ, Blocks), + SuccBlk = map_get(Succ, Blocks), is_tagged_tuple_1(SuccBlk, Tuple, Fail, Blocks); is_tagged_tuple(_, _, _, _) -> no. is_tagged_tuple_1(#b_blk{is=Is,last=Last}, Tuple, Fail, Blocks) -> case Is of - [#b_set{op={bif,tuple_size},dst=#b_var{name=ArityVar}, - args=[#b_var{name=Tuple}]}, + [#b_set{op={bif,tuple_size},dst=ArityVar, + args=[#b_var{}=Tuple]}, #b_set{op={bif,'=:='}, - dst=#b_var{name=Bool}, - args=[#b_var{name=ArityVar}, - #b_literal{val=ArityVal}=Arity]}] + dst=Bool, + args=[ArityVar, #b_literal{val=ArityVal}=Arity]}] when is_integer(ArityVal) -> case Last of - #b_br{bool=#b_var{name=Bool},succ=Succ,fail=Fail} -> - SuccBlk = maps:get(Succ, Blocks), + #b_br{bool=Bool,succ=Succ,fail=Fail} -> + SuccBlk = map_get(Succ, Blocks), case is_tagged_tuple_2(SuccBlk, Tuple, Fail) of no -> no; @@ -255,22 +721,22 @@ is_tagged_tuple_1(#b_blk{is=Is,last=Last}, Tuple, Fail, Blocks) -> end. is_tagged_tuple_2(#b_blk{is=Is, - last=#b_br{bool=#b_var{name=Bool},fail=Fail}}, + last=#b_br{bool=#b_var{}=Bool,fail=Fail}}, Tuple, Fail) -> is_tagged_tuple_3(Is, Bool, Tuple); is_tagged_tuple_2(#b_blk{}, _, _) -> no. is_tagged_tuple_3([#b_set{op=get_tuple_element, - dst=#b_var{name=TagVar}, - args=[#b_var{name=Tuple},#b_literal{val=0}]}|Is], + dst=TagVar, + args=[#b_var{}=Tuple,#b_literal{val=0}]}|Is], Bool, Tuple) -> is_tagged_tuple_4(Is, Bool, TagVar); is_tagged_tuple_3([_|Is], Bool, Tuple) -> is_tagged_tuple_3(Is, Bool, Tuple); is_tagged_tuple_3([], _, _) -> no. -is_tagged_tuple_4([#b_set{op={bif,'=:='},dst=#b_var{name=Bool}, - args=[#b_var{name=TagVar}, +is_tagged_tuple_4([#b_set{op={bif,'=:='},dst=Bool, + args=[#b_var{}=TagVar, #b_literal{val=TagVal}=Tag]}], Bool, TagVar) when is_atom(TagVal) -> {yes,Tag}; @@ -286,12 +752,12 @@ is_tagged_tuple_4([], _, _) -> no. %%% subexpressions across instructions that clobber the X registers. %%% -ssa_opt_cse(#st{ssa=Linear}=St) -> +ssa_opt_cse({#st{ssa=Linear}=St, FuncDb}) -> M = #{0=>#{}}, - St#st{ssa=cse(Linear, #{}, M)}. + {St#st{ssa=cse(Linear, #{}, M)}, FuncDb}. cse([{L,#b_blk{is=Is0,last=Last0}=Blk}|Bs], Sub0, M0) -> - Es0 = maps:get(L, M0), + Es0 = map_get(L, M0), {Is1,Es,Sub} = cse_is(Is0, Es0, Sub0, []), Last = sub(Last0, Sub), M = cse_successors(Is1, Blk, Es, M0), @@ -318,7 +784,13 @@ cse_successors(_Is, Blk, Es, M) -> cse_successors_1([L|Ls], Es0, M) -> case M of + #{L:=Es1} when map_size(Es1) =:= 0 -> + %% The map is already empty. No need to do anything + %% since the intersection will be empty. + cse_successors_1(Ls, Es0, M); #{L:=Es1} -> + %% Calculate the intersection of the two maps. + %% Both keys and values must match. Es = maps:filter(fun(Key, Value) -> case Es1 of #{Key:=Value} -> true; @@ -381,11 +853,20 @@ cse_suitable(#b_set{op=get_hd}) -> true; cse_suitable(#b_set{op=get_tl}) -> true; cse_suitable(#b_set{op=put_list}) -> true; cse_suitable(#b_set{op=put_tuple}) -> true; -cse_suitable(#b_set{op={bif,Name},args=Args}) -> +cse_suitable(#b_set{op={bif,tuple_size}}) -> + %% Doing CSE for tuple_size/1 can prevent the + %% creation of test_arity and select_tuple_arity + %% instructions. That could decrease performance + %% and beam_validator could fail to understand + %% that tuple operations that follow are safe. + false; +cse_suitable(#b_set{anno=Anno,op={bif,Name},args=Args}) -> + %% Doing CSE for floating point operators is unsafe. %% Doing CSE for comparison operators would prevent %% creation of 'test' instructions. Arity = length(Args), - not (erl_internal:new_type_test(Name, Arity) orelse + not (is_map_key(float_op, Anno) orelse + erl_internal:new_type_test(Name, Arity) orelse erl_internal:comp_op(Name, Arity) orelse erl_internal:bool_op(Name, Arity)); cse_suitable(#b_set{}) -> false. @@ -410,16 +891,17 @@ cse_suitable(#b_set{}) -> false. {s=undefined :: 'undefined' | 'cleared', regs=#{} :: #{beam_ssa:b_var():=beam_ssa:b_var()}, fail=none :: 'none' | beam_ssa:label(), - ren=#{} :: #{beam_ssa:label():=beam_ssa:label()}, - non_guards :: gb_sets:set(beam_ssa:label()) + non_guards :: gb_sets:set(beam_ssa:label()), + bs :: beam_ssa:block_map() }). -ssa_opt_float(#st{ssa=Linear0,cnt=Count0}=St) -> +ssa_opt_float({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) -> NonGuards0 = float_non_guards(Linear0), NonGuards = gb_sets:from_list(NonGuards0), - Fs = #fs{non_guards=NonGuards}, + Blocks = maps:from_list(Linear0), + Fs = #fs{non_guards=NonGuards,bs=Blocks}, {Linear,Count} = float_opt(Linear0, Count0, Fs), - St#st{ssa=Linear,cnt=Count}. + {St#st{ssa=Linear,cnt=Count}, FuncDb}. float_non_guards([{L,#b_blk{is=Is}}|Bs]) -> case Is of @@ -430,42 +912,35 @@ float_non_guards([{L,#b_blk{is=Is}}|Bs]) -> end; float_non_guards([]) -> [?BADARG_BLOCK]. -float_opt([{L,Blk0}|Bs], Count, Fs) -> - Blk = float_rename_phis(Blk0, Fs), - case float_need_flush(Blk, Fs) of - true -> - float_flush(L, Blk, Bs, Count, Fs); - false -> - float_opt_1(L, Blk, Bs, Count, Fs) - end; -float_opt([], Count, _Fs) -> - {[],Count}. - -float_opt_1(L, #b_blk{last=#b_br{fail=F}}=Blk, Bs0, +float_opt([{L,#b_blk{last=#b_br{fail=F}}=Blk}|Bs0], Count0, #fs{non_guards=NonGuards}=Fs) -> case gb_sets:is_member(F, NonGuards) of true -> %% This block is not inside a guard. %% We can do the optimization. - float_opt_2(L, Blk, Bs0, Count0, Fs); + float_opt_1(L, Blk, Bs0, Count0, Fs); false -> %% This block is inside a guard. Don't do %% any floating point optimizations. {Bs,Count} = float_opt(Bs0, Count0, Fs), {[{L,Blk}|Bs],Count} end; -float_opt_1(L, Blk, Bs, Count, Fs) -> - float_opt_2(L, Blk, Bs, Count, Fs). +float_opt([{L,Blk}|Bs], Count, Fs) -> + float_opt_1(L, Blk, Bs, Count, Fs); +float_opt([], Count, _Fs) -> + {[],Count}. -float_opt_2(L, #b_blk{is=Is0}=Blk0, Bs0, Count0, Fs0) -> +float_opt_1(L, #b_blk{is=Is0}=Blk0, Bs0, Count0, Fs0) -> case float_opt_is(Is0, Fs0, Count0, []) of {Is1,Fs1,Count1} -> - Fs = float_fail_label(Blk0, Fs1), - Split = float_split_conv(Is1, Blk0), - {Blks0,Count2} = float_number(Split, L, Count1), - {Blks,Count3} = float_conv(Blks0, Fs#fs.fail, Count2), - {Bs,Count} = float_opt(Bs0, Count3, Fs), - {Blks++Bs,Count}; + Fs2 = float_fail_label(Blk0, Fs1), + Fail = Fs2#fs.fail, + {Flush,Blk,Fs,Count2} = float_maybe_flush(Blk0, Fs2, Count1), + Split = float_split_conv(Is1, Blk), + {Blks0,Count3} = float_number(Split, L, Count2), + {Blks,Count4} = float_conv(Blks0, Fail, Count3), + {Bs,Count} = float_opt(Bs0, Count4, Fs), + {Blks++Flush++Bs,Count}; none -> {Bs,Count} = float_opt(Bs0, Count0, Fs0), {[{L,Blk0}|Bs],Count} @@ -525,14 +1000,42 @@ float_conv([{L,#b_blk{is=Is0}=Blk0}|Bs0], Fail, Count0) -> end end. -float_need_flush(#b_blk{is=Is}, #fs{s=cleared}) -> +float_maybe_flush(Blk0, #fs{s=cleared,fail=Fail,bs=Blocks}=Fs0, Count0) -> + #b_blk{last=#b_br{bool=#b_var{},succ=Succ}=Br} = Blk0, + #b_blk{is=Is} = map_get(Succ, Blocks), case Is of [#b_set{anno=#{float_op:=_}}|_] -> - false; + %% The next operation is also a floating point operation. + %% No flush needed. + {[],Blk0,Fs0,Count0}; _ -> - true + %% Flush needed. + {Bool0,Count1} = new_reg('@ssa_bool', Count0), + Bool = #b_var{name=Bool0}, + + %% Allocate block numbers. + CheckL = Count1, %For checkerror. + FlushL = Count1 + 1, %For flushing of float regs. + Count = Count1 + 2, + Blk = Blk0#b_blk{last=Br#b_br{succ=CheckL}}, + + %% Build the block with the checkerror instruction. + CheckIs = [#b_set{op={float,checkerror},dst=Bool}], + CheckBr = #b_br{bool=Bool,succ=FlushL,fail=Fail}, + CheckBlk = #b_blk{is=CheckIs,last=CheckBr}, + + %% Build the block that flushes all registers. + FlushIs = float_flush_regs(Fs0), + FlushBr = #b_br{bool=#b_literal{val=true},succ=Succ,fail=Succ}, + FlushBlk = #b_blk{is=FlushIs,last=FlushBr}, + + %% Update state and blocks. + Fs = Fs0#fs{s=undefined,regs=#{},fail=none}, + FlushBs = [{CheckL,CheckBlk},{FlushL,FlushBlk}], + {FlushBs,Blk,Fs,Count} end; -float_need_flush(_, _) -> false. +float_maybe_flush(Blk, Fs, Count) -> + {[],Blk,Fs,Count}. float_opt_is([#b_set{op=succeeded,args=[Src]}=I0], #fs{regs=Rs}=Fs, Count, Acc) -> @@ -557,27 +1060,6 @@ float_opt_is([], Fs, _Count, _Acc) -> #fs{s=undefined} = Fs, %Assertion. none. -float_rename_phis(#b_blk{is=Is}=Blk, #fs{ren=Ren}) -> - if - map_size(Ren) =:= 0 -> - Blk; - true -> - Blk#b_blk{is=float_rename_phis_1(Is, Ren)} - end. - -float_rename_phis_1([#b_set{op=phi,args=Args0}=I|Is], Ren) -> - Args = [float_phi_arg(Arg, Ren) || Arg <- Args0], - [I#b_set{args=Args}|float_rename_phis_1(Is, Ren)]; -float_rename_phis_1(Is, _) -> Is. - -float_phi_arg({Var,OldLbl}, Ren) -> - case Ren of - #{OldLbl:=NewLbl} -> - {Var,NewLbl}; - #{} -> - {Var,OldLbl} - end. - float_make_op(#b_set{op={bif,Op},dst=Dst,args=As0}=I0, Ts, #fs{s=S,regs=Rs0}=Fs, Count0) -> {As1,Rs1,Count1} = float_load(As0, Ts, Rs0, Count0, []), @@ -634,37 +1116,6 @@ new_reg(Base, Count) -> Fr = {Base,Count}, {Fr,Count+1}. -float_flush(L, Blk, Bs0, Count0, #fs{s=cleared,fail=Fail,ren=Ren0}=Fs0) -> - {Bool0,Count1} = new_reg('@ssa_bool', Count0), - Bool = #b_var{name=Bool0}, - - %% Insert two blocks before the current block. First allocate - %% block numbers. - FirstL = L, %For checkerror. - MiddleL = Count1, %For flushed float regs. - LastL = Count1 + 1, %For original block. - Count2 = Count1 + 2, - - %% Build the block with the checkerror instruction. - CheckIs = [#b_set{op={float,checkerror},dst=Bool}], - FirstBlk = #b_blk{is=CheckIs,last=#b_br{bool=Bool,succ=MiddleL,fail=Fail}}, - - %% Build the block that flushes all registers. Note that this must be a - %% separate block in case the original block begins with a phi instruction, - %% to avoid embedding a phi instruction in the middle of a block. - FlushIs = float_flush_regs(Fs0), - MiddleBlk = #b_blk{is=FlushIs,last=#b_br{bool=#b_literal{val=true}, - succ=LastL,fail=LastL}}, - - %% The last block is the original unmodified block. - LastBlk = Blk, - - %% Update state and blocks. - Ren = Ren0#{L=>LastL}, - Fs = Fs0#fs{s=undefined,regs=#{},fail=none,ren=Ren}, - Bs1 = [{FirstL,FirstBlk},{MiddleL,MiddleBlk},{LastL,LastBlk}|Bs0], - float_opt(Bs1, Count2, Fs). - float_fail_label(#b_blk{last=Last}, Fs) -> case Last of #b_br{bool=#b_var{},fail=Fail} -> @@ -688,30 +1139,38 @@ float_flush_regs(#fs{regs=Rs}) -> %%% with a cheaper instructions %%% -ssa_opt_live(#st{ssa=Linear}=St) -> - St#st{ssa=live_opt(reverse(Linear), #{}, [])}. - -live_opt([{L,Blk0}|Bs], LiveMap0, Acc) -> - Successors = beam_ssa:successors(Blk0), - Live0 = live_opt_succ(Successors, L, LiveMap0), - {Blk,Live} = live_opt_blk(Blk0, Live0), +ssa_opt_live({#st{ssa=Linear0}=St, FuncDb}) -> + RevLinear = reverse(Linear0), + Blocks0 = maps:from_list(RevLinear), + Blocks = live_opt(RevLinear, #{}, Blocks0), + Linear = beam_ssa:linearize(Blocks), + {St#st{ssa=Linear}, FuncDb}. + +live_opt([{L,Blk0}|Bs], LiveMap0, Blocks) -> + Blk1 = beam_ssa_share:block(Blk0, Blocks), + Successors = beam_ssa:successors(Blk1), + Live0 = live_opt_succ(Successors, L, LiveMap0, gb_sets:empty()), + {Blk,Live} = live_opt_blk(Blk1, Live0), LiveMap = live_opt_phis(Blk#b_blk.is, L, Live, LiveMap0), - live_opt(Bs, LiveMap, [{L,Blk}|Acc]); + live_opt(Bs, LiveMap, Blocks#{L:=Blk}); live_opt([], _, Acc) -> Acc. -live_opt_succ([S|Ss], L, LiveMap) -> - Live0 = live_opt_succ(Ss, L, LiveMap), +live_opt_succ([S|Ss], L, LiveMap, Live0) -> Key = {S,L}, case LiveMap of #{Key:=Live} -> - gb_sets:union(Live, Live0); + %% The successor has a phi node, and the value for + %% this block in the phi node is a variable. + live_opt_succ(Ss, L, LiveMap, gb_sets:union(Live, Live0)); #{S:=Live} -> - gb_sets:union(Live, Live0); + %% No phi node in the successor, or the value for + %% this block in the phi node is a literal. + live_opt_succ(Ss, L, LiveMap, gb_sets:union(Live, Live0)); #{} -> - Live0 + %% A peek_message block which has not been processed yet. + live_opt_succ(Ss, L, LiveMap, Live0) end; -live_opt_succ([], _, _) -> - gb_sets:empty(). +live_opt_succ([], _, _, Acc) -> Acc. live_opt_phis(Is, L, Live0, LiveMap0) -> LiveMap = LiveMap0#{L=>Live0}, @@ -721,11 +1180,16 @@ live_opt_phis(Is, L, Live0, LiveMap0) -> LiveMap; [_|_] -> PhiArgs = append([Args || #b_set{args=Args} <- Phis]), - PhiVars = [{P,V} || {#b_var{name=V},P} <- PhiArgs], - PhiLive0 = rel2fam(PhiVars), - PhiLive = [{{L,P},gb_sets:union(gb_sets:from_list(Vs), Live0)} || - {P,Vs} <- PhiLive0], - maps:merge(LiveMap, maps:from_list(PhiLive)) + case [{P,V} || {#b_var{}=V,P} <- PhiArgs] of + [_|_]=PhiVars -> + PhiLive0 = rel2fam(PhiVars), + PhiLive = [{{L,P},gb_sets:union(gb_sets:from_list(Vs), Live0)} || + {P,Vs} <- PhiLive0], + maps:merge(LiveMap, maps:from_list(PhiLive)); + [] -> + %% There were only literals in the phi node(s). + LiveMap + end end. live_opt_blk(#b_blk{is=Is0,last=Last}=Blk, Live0) -> @@ -733,26 +1197,21 @@ live_opt_blk(#b_blk{is=Is0,last=Last}=Blk, Live0) -> {Is,Live} = live_opt_is(reverse(Is0), Live1, []), {Blk#b_blk{is=Is},Live}. -live_opt_is([#b_set{op=phi,dst=#b_var{name=Dst}}=I|Is], Live, Acc) -> +live_opt_is([#b_set{op=phi,dst=Dst}=I|Is], Live, Acc) -> case gb_sets:is_member(Dst, Live) of true -> live_opt_is(Is, Live, [I|Acc]); false -> live_opt_is(Is, Live, Acc) end; -live_opt_is([#b_set{op=succeeded,dst=#b_var{name=SuccDst}=SuccDstVar, - args=[#b_var{name=Dst}]}=SuccI, - #b_set{dst=#b_var{name=Dst}}=I|Is], Live0, Acc) -> +live_opt_is([#b_set{op=succeeded,dst=SuccDst=SuccDstVar, + args=[Dst]}=SuccI, + #b_set{dst=Dst}=I|Is], Live0, Acc) -> case gb_sets:is_member(Dst, Live0) of true -> - case gb_sets:is_member(SuccDst, Live0) of - true -> - Live1 = gb_sets:add(Dst, Live0), - Live = gb_sets:delete_any(SuccDst, Live1), - live_opt_is([I|Is], Live, [SuccI|Acc]); - false -> - live_opt_is([I|Is], Live0, Acc) - end; + Live1 = gb_sets:add(Dst, Live0), + Live = gb_sets:delete_any(SuccDst, Live1), + live_opt_is([I|Is], Live, [SuccI|Acc]); false -> case live_opt_unused(I) of {replace,NewI0} -> @@ -762,21 +1221,21 @@ live_opt_is([#b_set{op=succeeded,dst=#b_var{name=SuccDst}=SuccDstVar, case gb_sets:is_member(SuccDst, Live0) of true -> Live1 = gb_sets:add(Dst, Live0), - Live = gb_sets:delete_any(SuccDst, Live1), + Live = gb_sets:delete(SuccDst, Live1), live_opt_is([I|Is], Live, [SuccI|Acc]); false -> live_opt_is([I|Is], Live0, Acc) end end end; -live_opt_is([#b_set{op=Op,dst=#b_var{name=Dst}}=I|Is], Live0, Acc) -> +live_opt_is([#b_set{dst=Dst}=I|Is], Live0, Acc) -> case gb_sets:is_member(Dst, Live0) of true -> Live1 = gb_sets:union(Live0, gb_sets:from_ordset(beam_ssa:used(I))), - Live = gb_sets:delete_any(Dst, Live1), + Live = gb_sets:delete(Dst, Live1), live_opt_is(Is, Live, [I|Acc]); false -> - case is_pure(Op) of + case beam_ssa:no_side_effect(I) of true -> live_opt_is(Is, Live0, Acc); false -> @@ -787,42 +1246,37 @@ live_opt_is([#b_set{op=Op,dst=#b_var{name=Dst}}=I|Is], Live0, Acc) -> live_opt_is([], Live, Acc) -> {Acc,Live}. -is_pure({bif,_}) -> true; -is_pure({float,get}) -> true; -is_pure(bs_extract) -> true; -is_pure(extract) -> true; -is_pure(get_hd) -> true; -is_pure(get_tl) -> true; -is_pure(get_tuple_element) -> true; -is_pure(is_nonempty_list) -> true; -is_pure(is_tagged_tuple) -> true; -is_pure(put_list) -> true; -is_pure(put_tuple) -> true; -is_pure(_) -> false. - live_opt_unused(#b_set{op=get_map_element}=Set) -> {replace,Set#b_set{op=has_map_field}}; live_opt_unused(_) -> keep. %%% -%%% Optimize binary matching instructions. +%%% Optimize binary matching. +%%% +%%% * If the value of segment is never extracted, rewrite +%%% to a bs_skip instruction. +%%% +%%% * Coalesce adjacent bs_skip instructions and skip instructions +%%% with bs_test_tail. %%% -ssa_opt_bsm(#st{ssa=Linear}=St) -> +ssa_opt_bsm({#st{ssa=Linear}=St, FuncDb}) -> Extracted0 = bsm_extracted(Linear), Extracted = cerl_sets:from_list(Extracted0), - St#st{ssa=bsm_skip(Linear, Extracted)}. + {St#st{ssa=bsm_skip(Linear, Extracted)}, FuncDb}. -bsm_skip([{L,#b_blk{is=Is0}=Blk}|Bs], Extracted) -> +bsm_skip([{L,#b_blk{is=Is0}=Blk}|Bs0], Extracted) -> + Bs = bsm_skip(Bs0, Extracted), Is = bsm_skip_is(Is0, Extracted), - [{L,Blk#b_blk{is=Is}}|bsm_skip(Bs, Extracted)]; + coalesce_skips({L,Blk#b_blk{is=Is}}, Bs); bsm_skip([], _) -> []. bsm_skip_is([I0|Is], Extracted) -> case I0 of - #b_set{op=bs_match,args=[#b_literal{val=string}|_]} -> - [I0|bsm_skip_is(Is, Extracted)]; - #b_set{op=bs_match,dst=Ctx,args=[Type,PrevCtx|Args0]} -> + #b_set{op=bs_match, + dst=Ctx, + args=[#b_literal{val=T}=Type,PrevCtx|Args0]} + when T =/= string, T =/= skip -> I = case cerl_sets:is_element(Ctx, Extracted) of true -> I0; @@ -846,18 +1300,75 @@ bsm_extracted([{_,#b_blk{is=Is}}|Bs]) -> end; bsm_extracted([]) -> []. +coalesce_skips({L,#b_blk{is=[#b_set{op=bs_extract}=Extract|Is0], + last=Last0}=Blk0}, Bs0) -> + case coalesce_skips_is(Is0, Last0, Bs0) of + not_possible -> + [{L,Blk0}|Bs0]; + {Is,Last,Bs} -> + Blk = Blk0#b_blk{is=[Extract|Is],last=Last}, + [{L,Blk}|Bs] + end; +coalesce_skips({L,#b_blk{is=Is0,last=Last0}=Blk0}, Bs0) -> + case coalesce_skips_is(Is0, Last0, Bs0) of + not_possible -> + [{L,Blk0}|Bs0]; + {Is,Last,Bs} -> + Blk = Blk0#b_blk{is=Is,last=Last}, + [{L,Blk}|Bs] + end. + +coalesce_skips_is([#b_set{op=bs_match, + args=[#b_literal{val=skip}, + Ctx0,Type,Flags, + #b_literal{val=Size0}, + #b_literal{val=Unit0}]}=Skip0, + #b_set{op=succeeded}], + #b_br{succ=L2,fail=Fail}=Br0, + Bs0) when is_integer(Size0) -> + case Bs0 of + [{L2,#b_blk{is=[#b_set{op=bs_match, + dst=SkipDst, + args=[#b_literal{val=skip},_,_,_, + #b_literal{val=Size1}, + #b_literal{val=Unit1}]}, + #b_set{op=succeeded}=Succeeded], + last=#b_br{fail=Fail}=Br}}|Bs] when is_integer(Size1) -> + SkipBits = Size0 * Unit0 + Size1 * Unit1, + Skip = Skip0#b_set{dst=SkipDst, + args=[#b_literal{val=skip},Ctx0, + Type,Flags, + #b_literal{val=SkipBits}, + #b_literal{val=1}]}, + Is = [Skip,Succeeded], + {Is,Br,Bs}; + [{L2,#b_blk{is=[#b_set{op=bs_test_tail, + args=[_Ctx,#b_literal{val=TailSkip}]}], + last=#b_br{succ=NextSucc,fail=Fail}}}|Bs] -> + SkipBits = Size0 * Unit0, + TestTail = Skip0#b_set{op=bs_test_tail, + args=[Ctx0,#b_literal{val=SkipBits+TailSkip}]}, + Br = Br0#b_br{bool=TestTail#b_set.dst,succ=NextSucc}, + Is = [TestTail], + {Is,Br,Bs}; + _ -> + not_possible + end; +coalesce_skips_is(_, _, _) -> + not_possible. + %%% %%% Short-cutting binary matching instructions. %%% -ssa_opt_bsm_shortcut(#st{ssa=Linear}=St) -> +ssa_opt_bsm_shortcut({#st{ssa=Linear}=St, FuncDb}) -> Positions = bsm_positions(Linear, #{}), case map_size(Positions) of 0 -> %% No binary matching instructions. - St; + {St, FuncDb}; _ -> - St#st{ssa=bsm_shortcut(Linear, Positions)} + {St#st{ssa=bsm_shortcut(Linear, Positions)}, FuncDb} end. bsm_positions([{L,#b_blk{is=Is,last=Last}}|Bs], PosMap0) -> @@ -865,7 +1376,7 @@ bsm_positions([{L,#b_blk{is=Is,last=Last}}|Bs], PosMap0) -> case {Is,Last} of {[#b_set{op=bs_test_tail,dst=Bool,args=[Ctx,#b_literal{val=Bits0}]}], #b_br{bool=Bool,fail=Fail}} -> - Bits = Bits0 + maps:get(Ctx, PosMap0), + Bits = Bits0 + map_get(Ctx, PosMap0), bsm_positions(Bs, PosMap#{L=>{Bits,Fail}}); {_,_} -> bsm_positions(Bs, PosMap) @@ -916,70 +1427,510 @@ bsm_shortcut([{L,#b_blk{is=Is,last=Last0}=Blk}|Bs], PosMap) -> bsm_shortcut([], _PosMap) -> []. %%% -%%% Miscellanous optimizations in execution order. +%%% Eliminate redundant bs_test_unit2 instructions. %%% -ssa_opt_misc(#st{ssa=Linear}=St) -> - St#st{ssa=misc_opt(Linear, #{})}. +ssa_opt_bsm_units({#st{ssa=Linear}=St, FuncDb}) -> + {St#st{ssa=bsm_units(Linear, #{})}, FuncDb}. + +bsm_units([{L,#b_blk{last=#b_br{succ=Succ,fail=Fail}}=Block0} | Bs], UnitMaps0) -> + UnitsIn = maps:get(L, UnitMaps0, #{}), + {Block, UnitsOut} = bsm_units_skip(Block0, UnitsIn), + UnitMaps1 = bsm_units_join(Succ, UnitsOut, UnitMaps0), + UnitMaps = bsm_units_join(Fail, UnitsIn, UnitMaps1), + [{L, Block} | bsm_units(Bs, UnitMaps)]; +bsm_units([{L,#b_blk{last=#b_switch{fail=Fail,list=Switch}}=Block} | Bs], UnitMaps0) -> + UnitsIn = maps:get(L, UnitMaps0, #{}), + Labels = [Fail | [Lbl || {_Arg, Lbl} <- Switch]], + UnitMaps = foldl(fun(Lbl, UnitMaps) -> + bsm_units_join(Lbl, UnitsIn, UnitMaps) + end, UnitMaps0, Labels), + [{L, Block} | bsm_units(Bs, UnitMaps)]; +bsm_units([{L, Block} | Bs], UnitMaps) -> + [{L, Block} | bsm_units(Bs, UnitMaps)]; +bsm_units([], _UnitMaps) -> + []. + +bsm_units_skip(Block, Units) -> + bsm_units_skip_1(Block#b_blk.is, Block, Units). + +bsm_units_skip_1([#b_set{op=bs_start_match,dst=New}|_], Block, Units) -> + %% We bail early since there can't be more than one match per block. + {Block, Units#{ New => 1 }}; +bsm_units_skip_1([#b_set{op=bs_match, + dst=New, + args=[#b_literal{val=skip}, + Ctx, + #b_literal{val=binary}, + _Flags, + #b_literal{val=all}, + #b_literal{val=OpUnit}]}=Skip | Test], + Block0, Units) -> + [#b_set{op=succeeded,dst=Bool,args=[New]}] = Test, %Assertion. + #b_br{bool=Bool} = Last0 = Block0#b_blk.last, %Assertion. + CtxUnit = map_get(Ctx, Units), + if + CtxUnit rem OpUnit =:= 0 -> + Is = takewhile(fun(I) -> I =/= Skip end, Block0#b_blk.is), + Last = Last0#b_br{bool=#b_literal{val=true}}, + Block = Block0#b_blk{is=Is,last=Last}, + {Block, Units#{ New => CtxUnit }}; + CtxUnit rem OpUnit =/= 0 -> + {Block0, Units#{ New => OpUnit, Ctx => OpUnit }} + end; +bsm_units_skip_1([#b_set{op=bs_match,dst=New,args=Args}|_], Block, Units) -> + [_,Ctx|_] = Args, + CtxUnit = map_get(Ctx, Units), + OpUnit = bsm_op_unit(Args), + {Block, Units#{ New => gcd(OpUnit, CtxUnit) }}; +bsm_units_skip_1([_I | Is], Block, Units) -> + bsm_units_skip_1(Is, Block, Units); +bsm_units_skip_1([], Block, Units) -> + {Block, Units}. + +bsm_op_unit([_,_,_,Size,#b_literal{val=U}]) -> + case Size of + #b_literal{val=Sz} when is_integer(Sz) -> Sz*U; + _ -> U + end; +bsm_op_unit([#b_literal{val=string},_,#b_literal{val=String}]) -> + bit_size(String); +bsm_op_unit([#b_literal{val=utf8}|_]) -> + 8; +bsm_op_unit([#b_literal{val=utf16}|_]) -> + 16; +bsm_op_unit([#b_literal{val=utf32}|_]) -> + 32; +bsm_op_unit(_) -> + 1. + +%% Several paths can lead to the same match instruction and the inferred units +%% may differ between them, so we can only keep the information that is common +%% to all paths. +bsm_units_join(Lbl, MapA, UnitMaps0) when is_map_key(Lbl, UnitMaps0) -> + MapB = map_get(Lbl, UnitMaps0), + Merged = if + map_size(MapB) =< map_size(MapA) -> + bsm_units_join_1(maps:keys(MapB), MapA, MapB); + map_size(MapB) > map_size(MapA) -> + bsm_units_join_1(maps:keys(MapA), MapB, MapA) + end, + UnitMaps0#{Lbl := Merged}; +bsm_units_join(Lbl, MapA, UnitMaps0) when MapA =/= #{} -> + UnitMaps0#{Lbl => MapA}; +bsm_units_join(_Lbl, _MapA, UnitMaps0) -> + UnitMaps0. + +bsm_units_join_1([Key | Keys], Left, Right) when is_map_key(Key, Left) -> + UnitA = map_get(Key, Left), + UnitB = map_get(Key, Right), + bsm_units_join_1(Keys, Left, Right#{Key := gcd(UnitA, UnitB)}); +bsm_units_join_1([Key | Keys], Left, Right) -> + bsm_units_join_1(Keys, Left, maps:remove(Key, Right)); +bsm_units_join_1([], _MapA, Right) -> + Right. -misc_opt([{L,#b_blk{is=Is0,last=Last0}=Blk0}|Bs], Sub0) -> - {Is,Sub} = misc_opt_is(Is0, Sub0, []), - Last = sub(Last0, Sub), - Blk = Blk0#b_blk{is=Is,last=Last}, - [{L,Blk}|misc_opt(Bs, Sub)]; -misc_opt([], _) -> []. +%%% +%%% Optimize binary construction. +%%% +%%% If an integer segment or a float segment has a literal size and +%%% a literal value, convert to a binary segment. Coalesce adjacent +%%% literal binary segments. Literal binary segments will be converted +%%% to bs_put_string instructions in later pass. +%%% -misc_opt_is([#b_set{op=phi}=I0|Is], Sub0, Acc) -> - #b_set{dst=Dst,args=Args} = I = sub(I0, Sub0), - case all_same(Args) of +ssa_opt_bs_puts({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) -> + {Linear,Count} = opt_bs_puts(Linear0, Count0, []), + {St#st{ssa=Linear,cnt=Count}, FuncDb}. + +opt_bs_puts([{L,#b_blk{is=Is}=Blk0}|Bs], Count0, Acc0) -> + case Is of + [#b_set{op=bs_put}=I0] -> + case opt_bs_put(L, I0, Blk0, Count0, Acc0) of + not_possible -> + opt_bs_puts(Bs, Count0, [{L,Blk0}|Acc0]); + {Count,Acc1} -> + Acc = opt_bs_puts_merge(Acc1), + opt_bs_puts(Bs, Count, Acc) + end; + _ -> + opt_bs_puts(Bs, Count0, [{L,Blk0}|Acc0]) + end; +opt_bs_puts([], Count, Acc) -> + {reverse(Acc),Count}. + +opt_bs_puts_merge([{L1,#b_blk{is=Is}=Blk0},{L2,#b_blk{is=AccIs}}=BAcc|Acc]) -> + case {AccIs,Is} of + {[#b_set{op=bs_put, + args=[#b_literal{val=binary}, + #b_literal{}, + #b_literal{val=Bin0}, + #b_literal{val=all}, + #b_literal{val=1}]}], + [#b_set{op=bs_put, + args=[#b_literal{val=binary}, + #b_literal{}, + #b_literal{val=Bin1}, + #b_literal{val=all}, + #b_literal{val=1}]}=I0]} -> + %% Coalesce the two segments to one. + Bin = <<Bin0/bitstring,Bin1/bitstring>>, + I = I0#b_set{args=bs_put_args(binary, Bin, all)}, + Blk = Blk0#b_blk{is=[I]}, + [{L2,Blk}|Acc]; + {_,_} -> + [{L1,Blk0},BAcc|Acc] + end. + +opt_bs_put(L, I0, #b_blk{last=Br0}=Blk0, Count0, Acc) -> + case opt_bs_put(I0) of + [Bin] when is_bitstring(Bin) -> + Args = bs_put_args(binary, Bin, all), + I = I0#b_set{args=Args}, + Blk = Blk0#b_blk{is=[I]}, + {Count0,[{L,Blk}|Acc]}; + [{int,Int,Size},Bin] when is_bitstring(Bin) -> + %% Construct a bs_put_integer instruction following + %% by a bs_put_binary instruction. + IntArgs = bs_put_args(integer, Int, Size), + BinArgs = bs_put_args(binary, Bin, all), + {BinL,BinVarNum} = {Count0,Count0+1}, + Count = Count0 + 2, + BinVar = #b_var{name={'@ssa_bool',BinVarNum}}, + BinI = I0#b_set{dst=BinVar,args=BinArgs}, + BinBlk = Blk0#b_blk{is=[BinI],last=Br0#b_br{bool=BinVar}}, + IntI = I0#b_set{args=IntArgs}, + IntBlk = Blk0#b_blk{is=[IntI],last=Br0#b_br{succ=BinL}}, + {Count,[{BinL,BinBlk},{L,IntBlk}|Acc]}; + not_possible -> + not_possible + end. + +opt_bs_put(#b_set{args=[#b_literal{val=binary},_,#b_literal{val=Val}, + #b_literal{val=all},#b_literal{val=Unit}]}) + when is_bitstring(Val) -> + if + bit_size(Val) rem Unit =:= 0 -> + [Val]; true -> - %% Eliminate the phi node if there is just one source - %% value or if the values are identical. - [{Val,_}|_] = Args, - Sub = Sub0#{Dst=>Val}, - misc_opt_is(Is, Sub, Acc); - false -> - misc_opt_is(Is, Sub0, [I|Acc]) + not_possible end; -misc_opt_is([#b_set{}=I0|Is], Sub, Acc) -> - #b_set{op=Op,dst=Dst,args=Args} = I = sub(I0, Sub), - case make_literal(Op, Args) of - #b_literal{}=Literal -> - misc_opt_is(Is, Sub#{Dst=>Literal}, Acc); - error -> - misc_opt_is(Is, Sub, [I|Acc]) +opt_bs_put(#b_set{args=[#b_literal{val=Type},#b_literal{val=Flags}, + #b_literal{val=Val},#b_literal{val=Size}, + #b_literal{val=Unit}]}=I0) when is_integer(Size) -> + EffectiveSize = Size * Unit, + if + EffectiveSize > 0 -> + case {Type,opt_bs_put_endian(Flags)} of + {integer,big} when is_integer(Val) -> + if + EffectiveSize < 64 -> + [<<Val:EffectiveSize>>]; + true -> + opt_bs_put_split_int(Val, EffectiveSize) + end; + {integer,little} when is_integer(Val), EffectiveSize < 128 -> + %% To avoid an explosion in code size, we only try + %% to optimize relatively small fields. + <<Int:EffectiveSize>> = <<Val:EffectiveSize/little>>, + Args = bs_put_args(Type, Int, EffectiveSize), + I = I0#b_set{args=Args}, + opt_bs_put(I); + {binary,_} when is_bitstring(Val) -> + <<Bitstring:EffectiveSize/bits,_/bits>> = Val, + [Bitstring]; + {float,Endian} -> + try + [opt_bs_put_float(Val, EffectiveSize, Endian)] + catch error:_ -> + not_possible + end; + {_,_} -> + not_possible + end; + true -> + not_possible end; -misc_opt_is([], Sub, Acc) -> - {reverse(Acc),Sub}. +opt_bs_put(#b_set{}) -> not_possible. -all_same([{H,_}|T]) -> - all(fun({E,_}) -> E =:= H end, T). +opt_bs_put_float(N, Sz, Endian) -> + case Endian of + big -> <<N:Sz/big-float-unit:1>>; + little -> <<N:Sz/little-float-unit:1>> + end. -make_literal(put_tuple, Args) -> - case make_literal_list(Args, []) of +bs_put_args(Type, Val, Size) -> + [#b_literal{val=Type}, + #b_literal{val=[unsigned,big]}, + #b_literal{val=Val}, + #b_literal{val=Size}, + #b_literal{val=1}]. + +opt_bs_put_endian([big=E|_]) -> E; +opt_bs_put_endian([little=E|_]) -> E; +opt_bs_put_endian([native=E|_]) -> E; +opt_bs_put_endian([_|Fs]) -> opt_bs_put_endian(Fs). + +opt_bs_put_split_int(Int, Size) -> + Pos = opt_bs_put_split_int_1(Int, 0, Size - 1), + UpperSize = Size - Pos, + if + Pos =:= 0 -> + %% Value is 0 or -1 -- keep the original instruction. + not_possible; + UpperSize < 64 -> + %% No or few leading zeroes or ones. + [<<Int:Size>>]; + true -> + %% There are 64 or more leading ones or zeroes in + %% the resulting binary. Split into two separate + %% segments to avoid an explosion in code size. + [{int,Int bsr Pos,UpperSize},<<Int:Pos>>] + end. + +opt_bs_put_split_int_1(_Int, L, R) when L > R -> + 8 * ((L + 7) div 8); +opt_bs_put_split_int_1(Int, L, R) -> + Mid = (L + R) div 2, + case Int bsr Mid of + Upper when Upper =:= 0; Upper =:= -1 -> + opt_bs_put_split_int_1(Int, L, Mid - 1); + _ -> + opt_bs_put_split_int_1(Int, Mid + 1, R) + end. + +%%% +%%% Optimize expressions such as "tuple_size(Var) =:= 2". +%%% +%%% Consider this code: +%%% +%%% 0: +%%% . +%%% . +%%% . +%%% Size = bif:tuple_size Var +%%% BoolVar1 = succeeded Size +%%% br BoolVar1, label 4, label 3 +%%% +%%% 4: +%%% BoolVar2 = bif:'=:=' Size, literal 2 +%%% br BoolVar2, label 6, label 3 +%%% +%%% 6: ... %% OK +%%% +%%% 3: ... %% Not a tuple of size 2 +%%% +%%% The BEAM code will look this: +%%% +%%% {bif,tuple_size,{f,3},[{x,0}],{x,0}}. +%%% {test,is_eq_exact,{f,3},[{x,0},{integer,2}]}. +%%% +%%% Better BEAM code will be produced if we transform the +%%% code like this: +%%% +%%% 0: +%%% . +%%% . +%%% . +%%% br label 10 +%%% +%%% 10: +%%% NewBoolVar = bif:is_tuple Var +%%% br NewBoolVar, label 11, label 3 +%%% +%%% 11: +%%% Size = bif:tuple_size Var +%%% br label 4 +%%% +%%% 4: +%%% BoolVar2 = bif:'=:=' Size, literal 2 +%%% br BoolVar2, label 6, label 3 +%%% +%%% (The key part of the transformation is the removal of +%%% the 'succeeded' instruction to signal to the code generator +%%% that the call to tuple_size/1 can't fail.) +%%% +%%% The BEAM code will look like: +%%% +%%% {test,is_tuple,{f,3},[{x,0}]}. +%%% {test_arity,{f,3},[{x,0},2]}. +%%% +%%% Those two instructions will be combined into a single +%%% is_tuple_of_arity instruction by the loader. +%%% + +ssa_opt_tuple_size({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) -> + {Linear,Count} = opt_tup_size(Linear0, Count0, []), + {St#st{ssa=Linear,cnt=Count}, FuncDb}. + +opt_tup_size([{L,#b_blk{is=Is,last=Last}=Blk}|Bs], Count0, Acc0) -> + case {Is,Last} of + {[#b_set{op={bif,'=:='},dst=Bool,args=[#b_var{}=Tup,#b_literal{val=Arity}]}], + #b_br{bool=Bool}} when is_integer(Arity), Arity >= 0 -> + {Acc,Count} = opt_tup_size_1(Tup, L, Count0, Acc0), + opt_tup_size(Bs, Count, [{L,Blk}|Acc]); + {_,_} -> + opt_tup_size(Bs, Count0, [{L,Blk}|Acc0]) + end; +opt_tup_size([], Count, Acc) -> + {reverse(Acc),Count}. + +opt_tup_size_1(Size, EqL, Count0, [{L,Blk0}|Acc]) -> + case Blk0 of + #b_blk{is=Is0,last=#b_br{bool=Bool,succ=EqL,fail=Fail}} -> + case opt_tup_size_is(Is0, Bool, Size, []) of + none -> + {[{L,Blk0}|Acc],Count0}; + {PreIs,TupleSizeIs,Tuple} -> + opt_tup_size_2(PreIs, TupleSizeIs, L, EqL, + Tuple, Fail, Count0, Acc) + end; + #b_blk{} -> + {[{L,Blk0}|Acc],Count0} + end; +opt_tup_size_1(_, _, Count, Acc) -> + {Acc,Count}. + +opt_tup_size_2(PreIs, TupleSizeIs, PreL, EqL, Tuple, Fail, Count0, Acc) -> + IsTupleL = Count0, + TupleSizeL = Count0 + 1, + Bool = #b_var{name={'@ssa_bool',Count0+2}}, + Count = Count0 + 3, + + True = #b_literal{val=true}, + PreBr = #b_br{bool=True,succ=IsTupleL,fail=IsTupleL}, + PreBlk = #b_blk{is=PreIs,last=PreBr}, + + IsTupleIs = [#b_set{op={bif,is_tuple},dst=Bool,args=[Tuple]}], + IsTupleBr = #b_br{bool=Bool,succ=TupleSizeL,fail=Fail}, + IsTupleBlk = #b_blk{is=IsTupleIs,last=IsTupleBr}, + + TupleSizeBr = #b_br{bool=True,succ=EqL,fail=EqL}, + TupleSizeBlk = #b_blk{is=TupleSizeIs,last=TupleSizeBr}, + {[{TupleSizeL,TupleSizeBlk}, + {IsTupleL,IsTupleBlk}, + {PreL,PreBlk}|Acc],Count}. + +opt_tup_size_is([#b_set{op={bif,tuple_size},dst=Size,args=[Tuple]}=I, + #b_set{op=succeeded,dst=Bool,args=[Size]}], + Bool, Size, Acc) -> + {reverse(Acc),[I],Tuple}; +opt_tup_size_is([I|Is], Bool, Size, Acc) -> + opt_tup_size_is(Is, Bool, Size, [I|Acc]); +opt_tup_size_is([], _, _, _Acc) -> none. + +%%% +%%% Optimize #b_switch{} instructions. +%%% +%%% If the argument for a #b_switch{} comes from a phi node with all +%%% literals, any values in the switch list which are not in the phi +%%% node can be removed. +%%% +%%% If the values in the phi node and switch list are the same, +%%% the failure label can't be reached and be eliminated. +%%% +%%% A #b_switch{} with only one value can be rewritten to +%%% a #b_br{}. A switch that only verifies that the argument +%%% is 'true' or 'false' can be rewritten to a is_boolean test. +%%% + +ssa_opt_sw({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) -> + {Linear,Count} = opt_sw(Linear0, #{}, Count0, []), + {St#st{ssa=Linear,cnt=Count}, FuncDb}. + +opt_sw([{L,#b_blk{is=Is,last=#b_switch{}=Last0}=Blk0}|Bs], Phis0, Count0, Acc) -> + Phis = opt_sw_phis(Is, Phis0), + case opt_sw_last(Last0, Phis) of + #b_switch{arg=Arg,fail=Fail,list=[{Lit,Lbl}]} -> + %% Rewrite a single value switch to a br. + Bool = #b_var{name={'@ssa_bool',Count0}}, + Count = Count0 + 1, + IsEq = #b_set{op={bif,'=:='},dst=Bool,args=[Arg,Lit]}, + Br = #b_br{bool=Bool,succ=Lbl,fail=Fail}, + Blk = Blk0#b_blk{is=Is++[IsEq],last=Br}, + opt_sw(Bs, Phis, Count, [{L,Blk}|Acc]); + #b_switch{arg=Arg,fail=Fail, + list=[{#b_literal{val=B1},Lbl},{#b_literal{val=B2},Lbl}]} + when B1 =:= not B2 -> + %% Replace with is_boolean test. + Bool = #b_var{name={'@ssa_bool',Count0}}, + Count = Count0 + 1, + IsBool = #b_set{op={bif,is_boolean},dst=Bool,args=[Arg]}, + Br = #b_br{bool=Bool,succ=Lbl,fail=Fail}, + Blk = Blk0#b_blk{is=Is++[IsBool],last=Br}, + opt_sw(Bs, Phis, Count, [{L,Blk}|Acc]); + Last0 -> + opt_sw(Bs, Phis, Count0, [{L,Blk0}|Acc]); + Last -> + Blk = Blk0#b_blk{last=Last}, + opt_sw(Bs, Phis, Count0, [{L,Blk}|Acc]) + end; +opt_sw([{L,#b_blk{is=Is}=Blk}|Bs], Phis0, Count, Acc) -> + Phis = opt_sw_phis(Is, Phis0), + opt_sw(Bs, Phis, Count, [{L,Blk}|Acc]); +opt_sw([], _Phis, Count, Acc) -> + {reverse(Acc),Count}. + +opt_sw_phis([#b_set{op=phi,dst=Dst,args=Args}|Is], Phis) -> + case opt_sw_literals(Args, []) of error -> - error; - List -> - #b_literal{val=list_to_tuple(List)} + opt_sw_phis(Is, Phis); + Literals -> + opt_sw_phis(Is, Phis#{Dst=>Literals}) end; -make_literal(put_list, [#b_literal{val=H},#b_literal{val=T}]) -> - #b_literal{val=[H|T]}; -make_literal(_, _) -> error. +opt_sw_phis(_, Phis) -> Phis. + +opt_sw_last(#b_switch{arg=Arg,fail=Fail,list=List0}=Sw0, Phis) -> + case Phis of + #{Arg:=Values0} -> + Values = gb_sets:from_list(Values0), + + %% Prune the switch list to only contain the possible values. + List1 = [P || {Lit,_}=P <- List0, gb_sets:is_member(Lit, Values)], + + %% Now test whether the failure label can ever be reached. + Sw = case gb_sets:size(Values) =:= length(List1) of + true -> + %% The switch list has the same number of values as the phi node. + %% The values must be the same, because the values that were not + %% possible were pruned from the switch list. Therefore, the + %% failure label can't possibly be reached, and we can choose a + %% a new failure label by picking a value from the list. + case List1 of + [{#b_literal{},Lbl}|List] -> + Sw0#b_switch{fail=Lbl,list=List}; + [] -> + Sw0#b_switch{list=List1} + end; + false -> + %% There are some values in the phi node that are not in the + %% switch list; thus, the failure label can still be reached. + Sw0 + end, + beam_ssa:normalize(Sw); + #{} -> + %% Ensure that no label in the switch list is the same + %% as the failure label. + List = [{Val,Lbl} || {Val,Lbl} <- List0, Lbl =/= Fail], + Sw = Sw0#b_switch{list=List}, + beam_ssa:normalize(Sw) + end. -make_literal_list([#b_literal{val=H}|T], Acc) -> - make_literal_list(T, [H|Acc]); -make_literal_list([_|_], _) -> +opt_sw_literals([{#b_literal{}=Lit,_}|T], Acc) -> + opt_sw_literals(T, [Lit|Acc]); +opt_sw_literals([_|_], _Acc) -> error; -make_literal_list([], Acc) -> - reverse(Acc). +opt_sw_literals([], Acc) -> Acc. + %%% %%% Merge blocks. %%% -ssa_opt_merge_blocks(#st{ssa=Blocks}=St) -> +ssa_opt_merge_blocks({#st{ssa=Blocks}=St, FuncDb}) -> Preds = beam_ssa:predecessors(Blocks), - St#st{ssa=merge_blocks_1(beam_ssa:rpo(Blocks), Preds, Blocks)}. + Merged = merge_blocks_1(beam_ssa:rpo(Blocks), Preds, Blocks), + {St#st{ssa=Merged}, FuncDb}. merge_blocks_1([L|Ls], Preds0, Blocks0) -> case Preds0 of @@ -989,10 +1940,11 @@ merge_blocks_1([L|Ls], Preds0, Blocks0) -> true -> #b_blk{is=Is0} = Blk0, #b_blk{is=Is1} = Blk1, + verify_merge_is(Is1), Is = Is0 ++ Is1, Blk = Blk1#b_blk{is=Is}, Blocks1 = maps:remove(L, Blocks0), - Blocks2 = maps:put(P, Blk, Blocks1), + Blocks2 = Blocks1#{P:=Blk}, Successors = beam_ssa:successors(Blk), Blocks = beam_ssa:update_phi_labels(Successors, L, P, Blocks2), Preds = merge_update_preds(Successors, L, P, Preds0), @@ -1006,21 +1958,32 @@ merge_blocks_1([L|Ls], Preds0, Blocks0) -> merge_blocks_1([], _Preds, Blocks) -> Blocks. merge_update_preds([L|Ls], From, To, Preds0) -> - Ps = [rename_label(P, From, To) || P <- maps:get(L, Preds0)], - Preds = maps:put(L, Ps, Preds0), + Ps = [rename_label(P, From, To) || P <- map_get(L, Preds0)], + Preds = Preds0#{L:=Ps}, merge_update_preds(Ls, From, To, Preds); merge_update_preds([], _, _, Preds) -> Preds. rename_label(From, From, To) -> To; rename_label(Lbl, _, _) -> Lbl. -is_merge_allowed(_, _, #b_blk{is=[#b_set{op=peek_message}|_]}) -> +verify_merge_is([#b_set{op=Op}|_]) -> + %% The merged block has only one predecessor, so it should not have any phi + %% nodes. + true = Op =/= phi; %Assertion. +verify_merge_is(_) -> + ok. + +is_merge_allowed(_, #b_blk{}, #b_blk{is=[#b_set{op=peek_message}|_]}) -> false; -is_merge_allowed(L, Blk0, #b_blk{}) -> - case beam_ssa:successors(Blk0) of +is_merge_allowed(L, #b_blk{last=#b_br{}}=Blk, #b_blk{}) -> + %% The predecessor block must have exactly one successor (L) for + %% the merge to be safe. + case beam_ssa:successors(Blk) of [L] -> true; [_|_] -> false - end. + end; +is_merge_allowed(_, #b_blk{last=#b_switch{}}, #b_blk{}) -> + false. %%% %%% When a tuple is matched, the pattern matching compiler generates a @@ -1038,13 +2001,21 @@ is_merge_allowed(L, Blk0, #b_blk{}) -> %%% extracted values. %%% -ssa_opt_sink(#st{ssa=Blocks0}=St) -> +ssa_opt_sink({#st{ssa=Blocks0}=St, FuncDb}) -> Linear = beam_ssa:linearize(Blocks0), %% Create a map with all variables that define get_tuple_element %% instructions. The variable name map to the block it is defined in. - Defs = maps:from_list(def_blocks(Linear)), + case def_blocks(Linear) of + [] -> + %% No get_tuple_element instructions, so there is nothing to do. + {St, FuncDb}; + [_|_]=Defs0 -> + Defs = maps:from_list(Defs0), + {do_ssa_opt_sink(Linear, Defs, St), FuncDb} + end. +do_ssa_opt_sink(Linear, Defs, #st{ssa=Blocks0}=St) -> %% Now find all the blocks that use variables defined by get_tuple_element %% instructions. Used = used_blocks(Linear, Defs, []), @@ -1076,7 +2047,7 @@ ssa_opt_sink(#st{ssa=Blocks0}=St) -> %% Now move all suitable get_tuple_element instructions to their %% new blocks. Blocks = foldl(fun({V,To}, A) -> - From = maps:get(V, Defs), + From = map_get(V, Defs), move_defs(V, From, To, A) end, Blocks0, DefLoc), St#st{ssa=Blocks}. @@ -1085,7 +2056,7 @@ def_blocks([{L,#b_blk{is=Is}}|Bs]) -> def_blocks_is(Is, L, def_blocks(Bs)); def_blocks([]) -> []. -def_blocks_is([#b_set{op=get_tuple_element,dst=#b_var{name=Dst}}|Is], L, Acc) -> +def_blocks_is([#b_set{op=get_tuple_element,dst=Dst}|Is], L, Acc) -> def_blocks_is(Is, L, [{Dst,L}|Acc]); def_blocks_is([_|Is], L, Acc) -> def_blocks_is(Is, L, Acc); @@ -1146,11 +2117,11 @@ unsuitable_loop(L, Blocks, Predecessors) -> unsuitable_loop(L, Blocks, Predecessors, []). unsuitable_loop(L, Blocks, Predecessors, Acc) -> - Ps = maps:get(L, Predecessors), + Ps = map_get(L, Predecessors), unsuitable_loop_1(Ps, Blocks, Predecessors, Acc). unsuitable_loop_1([P|Ps], Blocks, Predecessors, Acc0) -> - case maps:get(P, Blocks) of + case map_get(P, Blocks) of #b_blk{is=[#b_set{op=peek_message}|_]} -> unsuitable_loop_1(Ps, Blocks, Predecessors, Acc0); #b_blk{} -> @@ -1174,7 +2145,7 @@ unsuitable_loop_1([], _, _, Acc) -> Acc. %% variable will not be included in the result list. new_def_locations([{V,UsedIn}|Vs], Defs, Dom) -> - DefIn = maps:get(V, Defs), + DefIn = map_get(V, Defs), case common_dom(UsedIn, DefIn, Dom) of [] -> new_def_locations(Vs, Defs, Dom); @@ -1185,27 +2156,27 @@ new_def_locations([{V,UsedIn}|Vs], Defs, Dom) -> new_def_locations([], _, _) -> []. common_dom([L|Ls], DefIn, Dom) -> - DomBy0 = maps:get(L, Dom), - DomBy = ordsets:subtract(DomBy0, maps:get(DefIn, Dom)), + DomBy0 = map_get(L, Dom), + DomBy = ordsets:subtract(DomBy0, map_get(DefIn, Dom)), common_dom_1(Ls, Dom, DomBy). common_dom_1(_, _, []) -> []; common_dom_1([L|Ls], Dom, [_|_]=DomBy0) -> - DomBy1 = maps:get(L, Dom), + DomBy1 = map_get(L, Dom), DomBy = ordsets:intersection(DomBy0, DomBy1), common_dom_1(Ls, Dom, DomBy); common_dom_1([], _, DomBy) -> DomBy. most_dominated([L|Ls], Dom) -> - most_dominated(Ls, L, maps:get(L, Dom), Dom). + most_dominated(Ls, L, map_get(L, Dom), Dom). most_dominated([L|Ls], L0, DomBy, Dom) -> case member(L, DomBy) of true -> most_dominated(Ls, L0, DomBy, Dom); false -> - most_dominated(Ls, L, maps:get(L, Dom), Dom) + most_dominated(Ls, L, map_get(L, Dom), Dom) end; most_dominated([], L, _, _) -> L. @@ -1228,7 +2199,7 @@ remove_def(V, #b_blk{is=Is0}=Blk) -> {Def,Is} = remove_def_is(Is0, V, []), {Def,Blk#b_blk{is=Is}}. -remove_def_is([#b_set{dst=#b_var{name=Dst}}=Def|Is], Dst, Acc) -> +remove_def_is([#b_set{dst=Dst}=Def|Is], Dst, Acc) -> {Def,reverse(Acc, Is)}; remove_def_is([I|Is], Dst, Acc) -> remove_def_is(Is, Dst, [I|Acc]). @@ -1278,25 +2249,34 @@ insert_def_is([], _V, Def) -> %%% Common utilities. %%% +gcd(A, B) -> + case A rem B of + 0 -> B; + X -> gcd(B, X) + end. + rel2fam(S0) -> S1 = sofs:relation(S0), S = sofs:rel2fam(S1), sofs:to_external(S). -sub(#b_set{op=phi,args=Args}=I, Sub) -> +sub(I, Sub) -> + beam_ssa:normalize(sub_1(I, Sub)). + +sub_1(#b_set{op=phi,args=Args}=I, Sub) -> I#b_set{args=[{sub_arg(A, Sub),P} || {A,P} <- Args]}; -sub(#b_set{args=Args}=I, Sub) -> +sub_1(#b_set{args=Args}=I, Sub) -> I#b_set{args=[sub_arg(A, Sub) || A <- Args]}; -sub(#b_br{bool=#b_var{}=Old}=Br, Sub) -> +sub_1(#b_br{bool=#b_var{}=Old}=Br, Sub) -> New = sub_arg(Old, Sub), Br#b_br{bool=New}; -sub(#b_switch{arg=#b_var{}=Old}=Sw, Sub) -> +sub_1(#b_switch{arg=#b_var{}=Old}=Sw, Sub) -> New = sub_arg(Old, Sub), Sw#b_switch{arg=New}; -sub(#b_ret{arg=#b_var{}=Old}=Ret, Sub) -> +sub_1(#b_ret{arg=#b_var{}=Old}=Ret, Sub) -> New = sub_arg(Old, Sub), Ret#b_ret{arg=New}; -sub(Last, _) -> Last. +sub_1(Last, _) -> Last. sub_arg(#b_remote{mod=Mod,name=Name}=Rem, Sub) -> Rem#b_remote{mod=sub_arg(Mod, Sub),name=sub_arg(Name, Sub)}; @@ -1305,3 +2285,9 @@ sub_arg(Old, Sub) -> #{Old:=New} -> New; #{} -> Old end. + +new_var(#b_var{name={Base,N}}, Count) -> + true = is_integer(N), %Assertion. + {#b_var{name={Base,Count}},Count+1}; +new_var(#b_var{name=Base}, Count) -> + {#b_var{name={Base,Count}},Count+1}. diff --git a/lib/compiler/src/beam_ssa_opt.hrl b/lib/compiler/src/beam_ssa_opt.hrl new file mode 100644 index 0000000000..37711a6f48 --- /dev/null +++ b/lib/compiler/src/beam_ssa_opt.hrl @@ -0,0 +1,53 @@ +%% +%% %CopyrightBegin% +%% +%% Copyright Ericsson AB 2019. All Rights Reserved. +%% +%% Licensed under the Apache License, Version 2.0 (the "License"); +%% you may not use this file except in compliance with the License. +%% You may obtain a copy of the License at +%% +%% http://www.apache.org/licenses/LICENSE-2.0 +%% +%% Unless required by applicable law or agreed to in writing, software +%% distributed under the License is distributed on an "AS IS" BASIS, +%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +%% See the License for the specific language governing permissions and +%% limitations under the License. +%% +%% %CopyrightEnd% +%% + +-include("beam_ssa.hrl"). + +-record(func_info, + {%% Local calls going in/out of this function. + in = ordsets:new() :: ordsets:ordset(func_id()), + out = ordsets:new() :: ordsets:ordset(func_id()), + + %% Whether the function is exported or not; some optimizations may + %% need to be suppressed if it is. + exported = true :: boolean(), + + %% The inferred types of each argument (as opposed to parameter), + %% indexed by call site. + %% + %% This is more effective than the naive approach of joining into a + %% "parameter_type" as we go as it lets us narrow parameter types + %% without having to visit all callers on each pass, which helps a lot + %% when dealing with co-recursive functions. + arg_types = [] :: list(arg_type_map()), + + %% The inferred return type of this function, this is either [type()] + %% or [] to note absence. + ret_type = [] :: list()}). + +-type arg_key() :: {CallerId :: func_id(), + CallDst :: beam_ssa:b_var()}. +-type arg_type_map() :: #{ arg_key() => term() }. + +%% Per-function metadata used by various optimization passes to perform +%% module-level optimization. If a function is absent it means that +%% module-level optimization has been turned off for said function. +-type func_id() :: beam_ssa:b_local(). +-type func_info_db() :: #{ func_id() => #func_info{} }. diff --git a/lib/compiler/src/beam_ssa_pp.erl b/lib/compiler/src/beam_ssa_pp.erl index 9daa2c2523..34ac08b32e 100644 --- a/lib/compiler/src/beam_ssa_pp.erl +++ b/lib/compiler/src/beam_ssa_pp.erl @@ -158,7 +158,7 @@ format_op({Prefix,Name}) -> format_op(Name) -> io_lib:format("~p", [Name]). -format_register(#b_var{name=V}, #{registers:=Regs}) -> +format_register(#b_var{}=V, #{registers:=Regs}) -> {Tag,N} = maps:get(V, Regs), io_lib:format("~p~p", [Tag,N]); format_register(_, #{}) -> "". @@ -224,9 +224,9 @@ format_anno_1(Anno) -> [io_lib:format(" %% Anno: ~p\n", [Anno])] end. -format_live_interval(#b_var{name=V}=Dst, #{live_intervals:=Intervals}) -> +format_live_interval(#b_var{}=Dst, #{live_intervals:=Intervals}) -> case Intervals of - #{V:=Rs0} -> + #{Dst:=Rs0} -> Rs1 = [io_lib:format("~p..~p", [Start,End]) || {Start,End} <- Rs0], Rs = lists:join(" ", Rs1), diff --git a/lib/compiler/src/beam_ssa_pre_codegen.erl b/lib/compiler/src/beam_ssa_pre_codegen.erl index 54aa2efaf6..274f78052d 100644 --- a/lib/compiler/src/beam_ssa_pre_codegen.erl +++ b/lib/compiler/src/beam_ssa_pre_codegen.erl @@ -23,11 +23,10 @@ %% it has been annotated and transformed to help the code generator. %% %% * Some instructions are translated to other instructions closer to -%% the BEAM instructions. For example, the put_tuple instruction is -%% broken apart into the put_tuple_arity and put_tuple_elements -%% instructions. Similary, the binary matching instructions are -%% transformed from the optimization-friendly internal format to -%% instruction more similar to the actual BEAM instructions. +%% the BEAM instructions. For example, the binary matching +%% instructions are transformed from the optimization-friendly +%% internal format to instruction more similar to the actual BEAM +%% instructions. %% %% * Blocks that will need an instruction for allocating a stack frame %% are annotated with a {frame_size,Size} annotation. @@ -73,21 +72,20 @@ -import(lists, [all/2,any/2,append/1,duplicate/2, foldl/3,last/1,map/2,member/2,partition/2, - reverse/1,reverse/2,sort/1,zip/2]). + reverse/1,reverse/2,sort/1,splitwith/2,zip/2]). -spec module(beam_ssa:b_module(), [compile:option()]) -> {'ok',beam_ssa:b_module()}. module(#b_module{body=Fs0}=Module, Opts) -> - FixTuples = proplists:get_bool(no_put_tuple2, Opts), - ExtraAnnos = proplists:get_bool(dprecg, Opts), - Ps = passes(FixTuples, ExtraAnnos), - Fs = functions(Fs0, Ps), + UseBSM3 = not proplists:get_bool(no_bsm3, Opts), + Ps = passes(Opts), + Fs = functions(Fs0, Ps, UseBSM3), {ok,Module#b_module{body=Fs}}. -functions([F|Fs], Ps) -> - [function(F, Ps)|functions(Fs, Ps)]; -functions([], _Ps) -> []. +functions([F|Fs], Ps, UseBSM3) -> + [function(F, Ps, UseBSM3)|functions(Fs, Ps, UseBSM3)]; +functions([], _Ps, _UseBSM3) -> []. -type b_var() :: beam_ssa:b_var(). -type var_name() :: beam_ssa:var_name(). @@ -105,16 +103,18 @@ functions([], _Ps) -> []. -record(st, {ssa :: beam_ssa:block_map(), args :: [b_var()], cnt :: beam_ssa:label(), + use_bsm3 :: boolean(), frames=[] :: [beam_ssa:label()], intervals=[] :: [{b_var(),[range()]}], - aliases=[] :: [{b_var(),b_var()}], res=[] :: [{b_var(),reservation()}] | #{b_var():=reservation()}, regs=#{} :: #{b_var():=ssa_register()}, extra_annos=[] :: [{atom(),term()}] }). -define(PASS(N), {N,fun N/1}). -passes(FixTuples, ExtraAnnos) -> +passes(Opts) -> + AddPrecgAnnos = proplists:get_bool(dprecg, Opts), + FixTuples = proplists:get_bool(no_put_tuple2, Opts), Ps = [?PASS(assert_no_critical_edges), %% Preliminaries. @@ -131,6 +131,10 @@ passes(FixTuples, ExtraAnnos) -> ?PASS(find_yregs), ?PASS(reserve_yregs), + %% Handle legacy binary match instruction that don't + %% accept a Y register as destination. + ?PASS(legacy_bs), + %% Improve reuse of Y registers to potentially %% reduce the size of the stack frame. ?PASS(copy_retval), @@ -139,27 +143,25 @@ passes(FixTuples, ExtraAnnos) -> %% Calculate live intervals. ?PASS(number_instructions), ?PASS(live_intervals), - ?PASS(remove_unsuitable_aliases), ?PASS(reserve_regs), - ?PASS(merge_intervals), %% If needed for a .precg file, save the live intervals %% so they can be included in an annotation. - case ExtraAnnos of + case AddPrecgAnnos of false -> ignore; true -> ?PASS(save_live_intervals) end, %% Allocate registers. ?PASS(linear_scan), - ?PASS(fix_aliased_regs), ?PASS(frame_size), ?PASS(turn_yregs)], [P || P <- Ps, P =/= ignore]. -function(#b_function{anno=Anno,args=Args,bs=Blocks0,cnt=Count0}=F0, Ps) -> +function(#b_function{anno=Anno,args=Args,bs=Blocks0,cnt=Count0}=F0, + Ps, UseBSM3) -> try - St0 = #st{ssa=Blocks0,args=Args,cnt=Count0}, + St0 = #st{ssa=Blocks0,args=Args,use_bsm3=UseBSM3,cnt=Count0}, St = compile:run_sub_passes(Ps, St0), #st{ssa=Blocks,cnt=Count,regs=Regs,extra_annos=ExtraAnnos} = St, F1 = add_extra_annos(F0, ExtraAnnos), @@ -175,14 +177,6 @@ function(#b_function{anno=Anno,args=Args,bs=Blocks0,cnt=Count0}=F0, Ps) -> save_live_intervals(#st{intervals=Intervals}=St) -> St#st{extra_annos=[{live_intervals,Intervals}]}. -fix_aliased_regs(#st{aliases=Aliases,regs=Regs}=St) -> - St#st{regs=fix_aliased_regs(Aliases, Regs)}. - -fix_aliased_regs([{Alias,V}|Aliases], Regs) -> - #{V:=Reg} = Regs, - fix_aliased_regs(Aliases, Regs#{Alias=>Reg}); -fix_aliased_regs([], Regs) -> Regs. - %% Add extra annotations when a .precg listing file is being produced. add_extra_annos(F, Annos) -> foldl(fun({Name,Value}, Acc) -> @@ -215,7 +209,7 @@ assert_no_ces(_, _, Blocks) -> Blocks. %% * Combine bs_match and bs_extract instructions to bs_get %% instructions. -fix_bs(#st{ssa=Blocks,cnt=Count0}=St) -> +fix_bs(#st{ssa=Blocks,cnt=Count0,use_bsm3=UseBSM3}=St) -> F = fun(#b_set{op=bs_start_match,dst=Dst}, A) -> %% Mark the root of the match context list. [{Dst,{context,Dst}}|A]; @@ -233,8 +227,13 @@ fix_bs(#st{ssa=Blocks,cnt=Count0}=St) -> CtxChain = maps:from_list(M), Linear0 = beam_ssa:linearize(Blocks), - %% Insert bs_save / bs_restore instructions where needed. - {Linear1,Count} = bs_save_restore(Linear0, CtxChain, Count0), + %% Insert position instructions where needed. + {Linear1,Count} = case UseBSM3 of + true -> + bs_pos_bsm3(Linear0, CtxChain, Count0); + false -> + bs_pos_bsm2(Linear0, CtxChain, Count0) + end, %% Rename instructions. Linear = bs_instrs(Linear1, CtxChain, []), @@ -242,10 +241,54 @@ fix_bs(#st{ssa=Blocks,cnt=Count0}=St) -> St#st{ssa=maps:from_list(Linear),cnt=Count} end. +%% Insert bs_get_position and bs_set_position instructions as needed. +bs_pos_bsm3(Linear0, CtxChain, Count0) -> + Rs0 = bs_restores(Linear0, CtxChain, #{}, #{}), + Rs = maps:values(Rs0), + S0 = sofs:relation(Rs, [{context,save_point}]), + S1 = sofs:relation_to_family(S0), + S = sofs:to_external(S1), -%% Insert bs_save and bs_restore instructions as needed. + {SavePoints,Count1} = make_bs_pos_dict(S, Count0, []), + {Gets,Count2} = make_bs_setpos_map(Rs, SavePoints, Count1, []), + {Sets,Count} = make_bs_getpos_map(maps:to_list(Rs0), SavePoints, Count2, []), -bs_save_restore(Linear0, CtxChain, Count0) -> + %% Now insert all saves and restores. + {bs_insert_bsm3(Linear0, Gets, Sets, SavePoints),Count}. + +make_bs_setpos_map([{Ctx,Save}=Ps|T], SavePoints, Count, Acc) -> + SavePoint = get_savepoint(Ps, SavePoints), + I = #b_set{op=bs_get_position,dst=SavePoint,args=[Ctx]}, + make_bs_setpos_map(T, SavePoints, Count+1, [{Save,I}|Acc]); +make_bs_setpos_map([], _, Count, Acc) -> + {maps:from_list(Acc),Count}. + +make_bs_getpos_map([{Bef,{Ctx,_}=Ps}|T], SavePoints, Count, Acc) -> + Ignored = #b_var{name={'@ssa_ignored',Count}}, + Args = [Ctx, get_savepoint(Ps, SavePoints)], + I = #b_set{op=bs_set_position,dst=Ignored,args=Args}, + make_bs_getpos_map(T, SavePoints, Count+1, [{Bef,I}|Acc]); +make_bs_getpos_map([], _, Count, Acc) -> + {maps:from_list(Acc),Count}. + +get_savepoint({_,_}=Ps, SavePoints) -> + Name = {'@ssa_bs_position', map_get(Ps, SavePoints)}, + #b_var{name=Name}. + +make_bs_pos_dict([{Ctx,Pts}|T], Count0, Acc0) -> + {Acc, Count} = make_bs_pos_dict_1(Pts, Ctx, Count0, Acc0), + make_bs_pos_dict(T, Count, Acc); +make_bs_pos_dict([], Count, Acc) -> + {maps:from_list(Acc), Count}. + +make_bs_pos_dict_1([H|T], Ctx, I, Acc) -> + make_bs_pos_dict_1(T, Ctx, I+1, [{{Ctx,H},I}|Acc]); +make_bs_pos_dict_1([], Ctx, I, Acc) -> + {[{Ctx,I}|Acc], I}. + +%% As bs_position but without OTP-22 instructions. This is only used when +%% cross-compiling to older versions. +bs_pos_bsm2(Linear0, CtxChain, Count0) -> Rs0 = bs_restores(Linear0, CtxChain, #{}, #{}), Rs = maps:values(Rs0), S0 = sofs:relation(Rs, [{context,save_point}]), @@ -256,7 +299,7 @@ bs_save_restore(Linear0, CtxChain, Count0) -> {Restores,Count} = make_restore_map(maps:to_list(Rs0), Slots, Count1, []), %% Now insert all saves and restores. - {bs_insert(Linear0, Saves, Restores, Slots),Count}. + {bs_insert_bsm2(Linear0, Saves, Restores, Slots),Count}. make_save_map([{Ctx,Save}=Ps|T], Slots, Count, Acc) -> Ignored = #b_var{name={'@ssa_ignored',Count}}, @@ -280,7 +323,7 @@ make_restore_map([], _, Count, Acc) -> make_slot({Same,Same}, _Slots) -> #b_literal{val=start}; make_slot({_,_}=Ps, Slots) -> - #b_literal{val=maps:get(Ps, Slots)}. + #b_literal{val=map_get(Ps, Slots)}. make_save_point_dict([{Ctx,Pts}|T], Acc0) -> Acc = make_save_point_dict_1(Pts, Ctx, 0, Acc0), @@ -309,8 +352,7 @@ bs_restores([], _, _, Rs) -> Rs. bs_update_successors(#b_br{succ=Succ,fail=Fail}, SPos, FPos, D) -> join_positions([{Succ,SPos},{Fail,FPos}], D); -bs_update_successors(#b_switch{fail=Fail,list=List}, SPos, FPos, D) -> - SPos = FPos, %Assertion. +bs_update_successors(#b_switch{fail=Fail,list=List}, SPos, _FPos, D) -> Update = [{L,SPos} || {_,L} <- List] ++ [{Fail,SPos}], join_positions(Update, D); bs_update_successors(#b_ret{}, _, _, D) -> D. @@ -389,10 +431,19 @@ bs_restores_is([#b_set{op=bs_extract,args=[FromPos|_]}|Is], Start = bs_subst_ctx(FromPos, CtxChain), #{Start:=FromPos} = PosMap, %Assertion. bs_restores_is(Is, CtxChain, PosMap, Rs); +bs_restores_is([#b_set{op=call,dst=Dst,args=Args}|Is], + CtxChain, PosMap0, Rs0) -> + {Rs,PosMap1} = bs_restore_args(Args, PosMap0, CtxChain, Dst, Rs0), + PosMap = bs_invalidate_pos(Args, PosMap1, CtxChain), + bs_restores_is(Is, CtxChain, PosMap, Rs); +bs_restores_is([#b_set{op=landingpad}|Is], CtxChain, PosMap0, Rs) -> + %% We can land here from any point, so all positions are invalid. + PosMap = maps:map(fun(_Start,_Pos) -> unknown end, PosMap0), + bs_restores_is(Is, CtxChain, PosMap, Rs); bs_restores_is([#b_set{op=Op,dst=Dst,args=Args}|Is], CtxChain, PosMap0, Rs0) when Op =:= bs_test_tail; - Op =:= call -> + Op =:= bs_get_tail -> {Rs,PosMap} = bs_restore_args(Args, PosMap0, CtxChain, Dst, Rs0), bs_restores_is(Is, CtxChain, PosMap, Rs); bs_restores_is([_|Is], CtxChain, PosMap, Rs) -> @@ -400,7 +451,6 @@ bs_restores_is([_|Is], CtxChain, PosMap, Rs) -> bs_restores_is([], _CtxChain, PosMap, Rs) -> {PosMap,Rs}. - bs_match_type(#b_set{args=[#b_literal{val=skip},_Ctx, #b_literal{val=binary},_Flags, #b_literal{val=all},#b_literal{val=U}]}) -> @@ -411,6 +461,23 @@ bs_match_type(#b_set{args=[#b_literal{val=skip},_Ctx, bs_match_type(_) -> plain. +%% Call instructions leave the match position in an undefined state, +%% requiring us to invalidate each affected argument. +bs_invalidate_pos([#b_var{}=Arg|Args], PosMap0, CtxChain) -> + Start = bs_subst_ctx(Arg, CtxChain), + case PosMap0 of + #{Start:=_} -> + PosMap = PosMap0#{Start:=unknown}, + bs_invalidate_pos(Args, PosMap, CtxChain); + #{} -> + %% Not a match context. + bs_invalidate_pos(Args, PosMap0, CtxChain) + end; +bs_invalidate_pos([_|Args], PosMap, CtxChain) -> + bs_invalidate_pos(Args, PosMap, CtxChain); +bs_invalidate_pos([], PosMap, _CtxChain) -> + PosMap. + bs_restore_args([#b_var{}=Arg|Args], PosMap0, CtxChain, Dst, Rs0) -> Start = bs_subst_ctx(Arg, CtxChain), case PosMap0 of @@ -433,33 +500,45 @@ bs_restore_args([], PosMap, _CtxChain, _Dst, Rs) -> %% Insert all bs_save and bs_restore instructions. -bs_insert([{L,#b_blk{is=Is0}=Blk}|Bs0], Saves, Restores, Slots) -> - Is = bs_insert_is_1(Is0, Restores, Slots), +bs_insert_bsm3(Blocks, Saves, Restores, SavePoints) -> + bs_insert_1(Blocks, Saves, Restores, SavePoints, fun(I) -> I end). + +bs_insert_bsm2(Blocks, Saves, Restores, SavePoints) -> + %% The old instructions require bs_start_match to be annotated with the + %% number of position slots it needs. + bs_insert_1(Blocks, Saves, Restores, SavePoints, + fun(#b_set{op=bs_start_match,dst=Dst}=I0) -> + NumSlots = case SavePoints of + #{Dst:=NumSlots0} -> NumSlots0; + #{} -> 0 + end, + beam_ssa:add_anno(num_slots, NumSlots, I0); + (I) -> + I + end). + +bs_insert_1([{L,#b_blk{is=Is0}=Blk}|Bs0], Saves, Restores, Slots, XFrm) -> + Is = bs_insert_is_1(Is0, Restores, Slots, XFrm), Bs = bs_insert_saves(Is, Bs0, Saves), - [{L,Blk#b_blk{is=Is}}|bs_insert(Bs, Saves, Restores, Slots)]; -bs_insert([], _, _, _) -> []. + [{L,Blk#b_blk{is=Is}}|bs_insert_1(Bs, Saves, Restores, Slots, XFrm)]; +bs_insert_1([], _, _, _, _) -> []. -bs_insert_is_1([#b_set{op=Op,dst=Dst}=I0|Is], Restores, Slots) -> +bs_insert_is_1([#b_set{op=Op,dst=Dst}=I0|Is], Restores, SavePoints, XFrm) -> + I = XFrm(I0), if Op =:= bs_test_tail; + Op =:= bs_get_tail; Op =:= bs_match; Op =:= call -> Rs = case Restores of #{Dst:=R} -> [R]; #{} -> [] end, - Rs ++ [I0|bs_insert_is_1(Is, Restores, Slots)]; - Op =:= bs_start_match -> - NumSlots = case Slots of - #{Dst:=NumSlots0} -> NumSlots0; - #{} -> 0 - end, - I = beam_ssa:add_anno(num_slots, NumSlots, I0), - [I|bs_insert_is_1(Is, Restores, Slots)]; + Rs ++ [I|bs_insert_is_1(Is, Restores, SavePoints, XFrm)]; true -> - [I0|bs_insert_is_1(Is, Restores, Slots)] + [I|bs_insert_is_1(Is, Restores, SavePoints, XFrm)] end; -bs_insert_is_1([], _, _) -> []. +bs_insert_is_1([], _, _, _) -> []. bs_insert_saves([#b_set{dst=Dst}|Is], Bs, Saves) -> case Saves of @@ -505,6 +584,8 @@ bs_instrs_is([#b_set{op=Op,args=Args0}=I0|Is], CtxChain, Acc) -> I1#b_set{op=bs_skip,args=[Type,Ctx|As]}; {bs_match,[#b_literal{val=string},Ctx|As]} -> I1#b_set{op=bs_match_string,args=[Ctx|As]}; + {bs_get_tail,[Ctx|As]} -> + I1#b_set{op=bs_get_tail,args=[Ctx|As]}; {_,_} -> I1 end, @@ -535,6 +616,59 @@ bs_subst_ctx(#b_var{}=Var, CtxChain) -> bs_subst_ctx(Other, _CtxChain) -> Other. +%% legacy_bs(St0) -> St. +%% Binary matching instructions in OTP 21 and earlier don't support +%% a Y register as destination. If St#st.use_bsm3 is false, +%% we will need to rewrite those instructions so that the result +%% is first put in an X register and then moved to a Y register +%% if the operation succeeded. + +legacy_bs(#st{use_bsm3=false,ssa=Blocks0,cnt=Count0,res=Res}=St) -> + IsYreg = maps:from_list([{V,true} || {V,{y,_}} <- Res]), + Linear0 = beam_ssa:linearize(Blocks0), + {Linear,Count} = legacy_bs(Linear0, IsYreg, Count0, #{}, []), + Blocks = maps:from_list(Linear), + St#st{ssa=Blocks,cnt=Count}; +legacy_bs(#st{use_bsm3=true}=St) -> St. + +legacy_bs([{L,Blk}|Bs], IsYreg, Count0, Copies0, Acc) -> + #b_blk{is=Is0,last=Last} = Blk, + Is1 = case Copies0 of + #{L:=Copy} -> [Copy|Is0]; + #{} -> Is0 + end, + {Is,Count,Copies} = legacy_bs_is(Is1, Last, IsYreg, Count0, Copies0, []), + legacy_bs(Bs, IsYreg, Count, Copies, [{L,Blk#b_blk{is=Is}}|Acc]); +legacy_bs([], _IsYreg, Count, _Copies, Acc) -> + {Acc,Count}. + +legacy_bs_is([#b_set{op=Op,dst=Dst}=I0, + #b_set{op=succeeded,dst=SuccDst,args=[Dst]}=SuccI0], + Last, IsYreg, Count0, Copies0, Acc) -> + NeedsFix = is_map_key(Dst, IsYreg) andalso + case Op of + bs_get -> true; + bs_init -> true; + _ -> false + end, + case NeedsFix of + true -> + TempDst = #b_var{name={'@bs_temp_dst',Count0}}, + Count = Count0 + 1, + I = I0#b_set{dst=TempDst}, + SuccI = SuccI0#b_set{args=[TempDst]}, + Copy = #b_set{op=copy,dst=Dst,args=[TempDst]}, + #b_br{bool=SuccDst,succ=SuccL} = Last, + Copies = Copies0#{SuccL=>Copy}, + legacy_bs_is([], Last, IsYreg, Count, Copies, [SuccI,I|Acc]); + false -> + legacy_bs_is([], Last, IsYreg, Count0, Copies0, [SuccI0,I0|Acc]) + end; +legacy_bs_is([I|Is], Last, IsYreg, Count, Copies, Acc) -> + legacy_bs_is(Is, Last, IsYreg, Count, Copies, [I|Acc]); +legacy_bs_is([], _Last, _IsYreg, Count, Copies, Acc) -> + {reverse(Acc),Count,Copies}. + %% sanitize(St0) -> St. %% Remove constructs that can cause problems later: %% @@ -550,7 +684,7 @@ sanitize(#st{ssa=Blocks0,cnt=Count0}=St) -> St#st{ssa=Blocks,cnt=Count}. sanitize([L|Ls], Count0, Blocks0, Values0) -> - #b_blk{is=Is0} = Blk0 = maps:get(L, Blocks0), + #b_blk{is=Is0} = Blk0 = map_get(L, Blocks0), case sanitize_is(Is0, Count0, Values0, false, []) of no_change -> sanitize(Ls, Count0, Blocks0, Values0); @@ -575,23 +709,24 @@ sanitize([], Count, Blocks0, Values) -> false -> remove_unreachable(Ls, Blocks, Reachable, []) end,Count}. -sanitize_is([#b_set{op=get_map_element, - args=[#b_literal{}=Map,Key]}=I0|Is], - Count0, Values, _Changed, Acc) -> - {MapVarName,Count} = new_var_name('@ssa_map', Count0), - MapVar = #b_var{name=MapVarName}, - I = I0#b_set{args=[MapVar,Key]}, - Copy = #b_set{op=copy,dst=MapVar,args=[Map]}, - sanitize_is(Is, Count, Values, true, [I,Copy|Acc]); -sanitize_is([#b_set{op=Op,dst=#b_var{name=Dst},args=Args0}=I0|Is0], - Count, Values, Changed, Acc) -> - Args = map(fun(#b_var{name=V}=Var) -> - case Values of - #{V:=New} -> New; - #{} -> Var - end; - (Lit) -> Lit - end, Args0), +sanitize_is([#b_set{op=get_map_element,args=Args0}=I0|Is], + Count0, Values, Changed, Acc) -> + case sanitize_args(Args0, Values) of + [#b_literal{}=Map,Key] -> + %% Bind the literal map to a variable. + {MapVar,Count} = new_var('@ssa_map', Count0), + I = I0#b_set{args=[MapVar,Key]}, + Copy = #b_set{op=copy,dst=MapVar,args=[Map]}, + sanitize_is(Is, Count, Values, true, [I,Copy|Acc]); + [_,_]=Args0 -> + sanitize_is(Is, Count0, Values, Changed, [I0|Acc]); + [_,_]=Args -> + I = I0#b_set{args=Args}, + sanitize_is(Is, Count0, Values, Changed, [I|Acc]) + end; +sanitize_is([#b_set{op=Op,dst=Dst,args=Args0}=I0|Is0], + Count, Values, Changed0, Acc) -> + Args = sanitize_args(Args0, Values), case sanitize_instr(Op, Args, I0) of {value,Value0} -> Value = #b_literal{val=Value0}, @@ -599,7 +734,9 @@ sanitize_is([#b_set{op=Op,dst=#b_var{name=Dst},args=Args0}=I0|Is0], {ok,I} -> sanitize_is(Is0, Count, Values, true, [I|Acc]); ok -> - sanitize_is(Is0, Count, Values, Changed, [I0|Acc]) + I = I0#b_set{args=Args}, + Changed = Changed0 orelse Args =/= Args0, + sanitize_is(Is0, Count, Values, Changed, [I|Acc]) end; sanitize_is([], Count, Values, Changed, Acc) -> case Changed of @@ -609,6 +746,14 @@ sanitize_is([], Count, Values, Changed, Acc) -> no_change end. +sanitize_args(Args, Values) -> + map(fun(Var) -> + case Values of + #{Var:=New} -> New; + #{} -> Var + end + end, Args). + sanitize_instr({bif,Bif}, [#b_literal{val=Lit}], _I) -> case erl_bifs:is_pure(erlang, Bif, 1) of false -> @@ -672,7 +817,7 @@ sanitize_badarg(I) -> I#b_set{op=call,args=[Func,#b_literal{val=badarg}]}. remove_unreachable([L|Ls], Blocks, Reachable, Acc) -> - #b_blk{is=Is0} = Blk0 = maps:get(L, Blocks), + #b_blk{is=Is0} = Blk0 = map_get(L, Blocks), case split_phis(Is0) of {[_|_]=Phis,Rest} -> Is = [prune_phi(Phi, Reachable) || Phi <- Phis] ++ Rest, @@ -703,8 +848,7 @@ prune_phi(#b_set{args=Args0}=Phi, Reachable) -> fix_tuples(#st{ssa=Blocks0,cnt=Count0}=St) -> F = fun (#b_set{op=put_tuple,args=Args}=Put, C0) -> Arity = #b_literal{val=length(Args)}, - {VarName,C} = new_var_name('@ssa_ignore', C0), - Ignore = #b_var{name=VarName}, + {Ignore,C} = new_var('@ssa_ignore', C0), {[Put#b_set{op=put_tuple_arity,args=[Arity]}, #b_set{dst=Ignore,op=put_tuple_elements,args=Args}],C}; (I, C) -> {[I],C} @@ -738,7 +882,7 @@ place_frames(#st{ssa=Blocks}=St) -> St#st{frames=Frames}. place_frames_1([L|Ls], Blocks, Doms, Tried0, Frames0) -> - Blk = maps:get(L, Blocks), + Blk = map_get(L, Blocks), case need_frame(Blk) of true -> %% This block needs a frame. Try to place it here. @@ -849,14 +993,14 @@ place_frame_here(L, Blocks, Doms, Frames) -> %% Return all predecessors referenced in phi nodes. phi_predecessors(L, Blocks) -> - #b_blk{is=Is} = maps:get(L, Blocks), + #b_blk{is=Is} = map_get(L, Blocks), [P || #b_set{op=phi,args=Args} <- Is, {_,P} <- Args]. %% is_dominated_by(Label, DominatedBy, Dominators) -> true|false. %% Test whether block Label is dominated by block DominatedBy. is_dominated_by(L, DomBy, Doms) -> - DominatedBy = maps:get(L, Doms), + DominatedBy = map_get(L, Doms), ordsets:is_element(DomBy, DominatedBy). %% need_frame(#b_blk{}) -> true|false. @@ -867,12 +1011,12 @@ need_frame(#b_blk{is=Is,last=#b_ret{arg=Ret}}) -> need_frame(#b_blk{is=Is}) -> need_frame_1(Is, body). -need_frame_1([#b_set{op=make_fun,dst=#b_var{name=Fun}}|Is], {return,_}=Context) -> +need_frame_1([#b_set{op=make_fun,dst=Fun}|Is], {return,_}=Context) -> %% Since make_fun clobbers X registers, a stack frame is needed if %% any of the following instructions use any other variable than %% the one holding the reference to the created fun. need_frame_1(Is, Context) orelse - case beam_ssa:used(#b_blk{is=Is,last=#b_ret{arg=#b_var{name=Fun}}}) of + case beam_ssa:used(#b_blk{is=Is,last=#b_ret{arg=Fun}}) of [Fun] -> false; [_|_] -> true end; @@ -887,7 +1031,7 @@ need_frame_1([#b_set{op=call,args=[Func|_]}|Is], Context) -> case Func of #b_remote{mod=#b_literal{val=Mod}, name=#b_literal{val=Name}, - arity=Arity} -> + arity=Arity} when is_atom(Mod), is_atom(Name) -> case erl_bifs:is_exit_bif(Mod, Name, Arity) of true -> false; @@ -899,11 +1043,11 @@ need_frame_1([#b_set{op=call,args=[Func|_]}|Is], Context) -> #b_remote{} -> %% This is an apply(), which always needs a frame. true; - #b_var{} -> - %% A fun call always needs a frame. - true; + #b_local{} -> + Context =:= body orelse Is =/= []; _ -> - Context =:= body orelse Is =/= [] + %% A fun call always needs a frame. + true end; need_frame_1([I|Is], Context) -> beam_ssa:clobbers_xregs(I) orelse need_frame_1(Is, Context); @@ -935,11 +1079,11 @@ is_trap_bif(_, _, _) -> false. %%% used during matching. %%% %%% Depending on where variables are defined and used, they must -%%% be handling in two different ways. +%%% be handled in two different ways. %%% %%% Variables that are always defined in the receive (before branching %%% out into the different clauses of the receive) and used after the -%%% receive, must be handled in the following way: Before each +%%% receive must be handled in the following way: Before each %%% remove_message instruction, each such variable must be copied, and %%% all variables must be consolidated using a phi node in the %%% common exit block for the receive. @@ -987,15 +1131,13 @@ recv_common(Defs, Exit, Blocks) -> %% in the exit block following the receive. recv_fix_common([Msg0|T], Exit, Rm, Blocks0, Count0) -> - {Msg1,Count1} = new_var_name('@recv', Count0), - Msg = #b_var{name=Msg1}, + {Msg,Count1} = new_var('@recv', Count0), Blocks1 = beam_ssa:rename_vars(#{Msg0=>Msg}, [Exit], Blocks0), N = length(Rm), - {MsgVars0,Count} = new_var_names(duplicate(N, '@recv'), Count1), - MsgVars = [#b_var{name=V} || V <- MsgVars0], + {MsgVars,Count} = new_vars(duplicate(N, '@recv'), Count1), PhiArgs = fix_exit_phi_args(MsgVars, Rm, Exit, Blocks1), Phi = #b_set{op=phi,dst=Msg,args=PhiArgs}, - ExitBlk0 = maps:get(Exit, Blocks1), + ExitBlk0 = map_get(Exit, Blocks1), ExitBlk = ExitBlk0#b_blk{is=[Phi|ExitBlk0#b_blk.is]}, Blocks2 = Blocks1#{Exit:=ExitBlk}, Blocks = recv_fix_common_1(MsgVars, Rm, Msg0, Blocks2), @@ -1006,8 +1148,8 @@ recv_fix_common([], _, _, Blocks, Count) -> recv_fix_common_1([V|Vs], [Rm|Rms], Msg, Blocks0) -> Ren = #{Msg=>V}, Blocks1 = beam_ssa:rename_vars(Ren, [Rm], Blocks0), - #b_blk{is=Is0} = Blk0 = maps:get(Rm, Blocks1), - Copy = #b_set{op=copy,dst=V,args=[#b_var{name=Msg}]}, + #b_blk{is=Is0} = Blk0 = map_get(Rm, Blocks1), + Copy = #b_set{op=copy,dst=V,args=[Msg]}, Is = insert_after_phis(Is0, [Copy]), Blk = Blk0#b_blk{is=Is}, Blocks = Blocks1#{Rm:=Blk}, @@ -1016,14 +1158,19 @@ recv_fix_common_1([], [], _Msg, Blocks) -> Blocks. fix_exit_phi_args([V|Vs], [Rm|Rms], Exit, Blocks) -> Path = beam_ssa:rpo([Rm], Blocks), - Pred = exit_predecessor(Path, Exit), - [{V,Pred}|fix_exit_phi_args(Vs, Rms, Exit, Blocks)]; + Preds = exit_predecessors(Path, Exit, Blocks), + [{V,Pred} || Pred <- Preds] ++ fix_exit_phi_args(Vs, Rms, Exit, Blocks); fix_exit_phi_args([], [], _, _) -> []. -exit_predecessor([Pred,Exit|_], Exit) -> - Pred; -exit_predecessor([_|Bs], Exit) -> - exit_predecessor(Bs, Exit). +exit_predecessors([L|Ls], Exit, Blocks) -> + Blk = map_get(L, Blocks), + case member(Exit, beam_ssa:successors(Blk)) of + true -> + [L|exit_predecessors(Ls, Exit, Blocks)]; + false -> + exit_predecessors(Ls, Exit, Blocks) + end; +exit_predecessors([], _Exit, _Blocks) -> []. %% fix_receive([Label], Defs, Blocks0, Count0) -> {Blocks,Count}. %% Add a copy instruction for all variables that are matched out and @@ -1033,16 +1180,14 @@ fix_receive([L|Ls], Defs, Blocks0, Count0) -> {RmDefs,Used0} = beam_ssa:def_used([L], Blocks0), Def = ordsets:subtract(Defs, RmDefs), Used = ordsets:intersection(Def, Used0), - {NewVs,Count} = new_var_names(Used, Count0), - NewVars = [#b_var{name=V} || V <- NewVs], + {NewVars,Count} = new_vars([Base || #b_var{name=Base} <- Used], Count0), Ren = zip(Used, NewVars), Blocks1 = beam_ssa:rename_vars(Ren, [L], Blocks0), - #b_blk{is=Is0} = Blk1 = maps:get(L, Blocks1), - CopyIs = [#b_set{op=copy,dst=New,args=[#b_var{name=Old}]} || - {Old,New} <- Ren], + #b_blk{is=Is0} = Blk1 = map_get(L, Blocks1), + CopyIs = [#b_set{op=copy,dst=New,args=[Old]} || {Old,New} <- Ren], Is = insert_after_phis(Is0, CopyIs), Blk = Blk1#b_blk{is=Is}, - Blocks = maps:put(L, Blk, Blocks1), + Blocks = Blocks1#{L:=Blk}, fix_receive(Ls, Defs, Blocks, Count); fix_receive([], _Defs, Blocks, Count) -> {Blocks,Count}. @@ -1067,7 +1212,7 @@ find_loop_exit_1(_, _, Exit) -> Exit. find_rm_blocks(L, Blocks) -> Seen = gb_sets:singleton(L), - Blk = maps:get(L, Blocks), + Blk = map_get(L, Blocks), Succ = beam_ssa:successors(Blk), find_rm_blocks_1(Succ, Seen, Blocks). @@ -1077,7 +1222,7 @@ find_rm_blocks_1([L|Ls], Seen0, Blocks) -> find_rm_blocks_1(Ls, Seen0, Blocks); false -> Seen = gb_sets:insert(L, Seen0), - Blk = maps:get(L, Blocks), + Blk = map_get(L, Blocks), case find_rm_act(Blk#b_blk.is) of prune -> %% Looping back. Don't look at any successors. @@ -1129,7 +1274,7 @@ find_rm_act([]) -> find_yregs(#st{frames=[]}=St) -> St; find_yregs(#st{frames=[_|_]=Frames,args=Args,ssa=Blocks0}=St) -> - FrameDefs = find_defs(Frames, Blocks0, [V || #b_var{name=V} <- Args]), + FrameDefs = find_defs(Frames, Blocks0, [V || #b_var{}=V <- Args]), Blocks = find_yregs_1(FrameDefs, Blocks0), St#st{ssa=Blocks}. @@ -1139,16 +1284,16 @@ find_yregs_1([{F,Defs}|Fs], Blocks0) -> Ls = beam_ssa:rpo([F], Blocks0), Yregs0 = [], Yregs = find_yregs_2(Ls, Blocks0, D0, Yregs0), - Blk0 = maps:get(F, Blocks0), + Blk0 = map_get(F, Blocks0), Blk = beam_ssa:add_anno(yregs, Yregs, Blk0), Blocks = Blocks0#{F:=Blk}, find_yregs_1(Fs, Blocks); find_yregs_1([], Blocks) -> Blocks. find_yregs_2([L|Ls], Blocks0, D0, Yregs0) -> - Blk0 = maps:get(L, Blocks0), + Blk0 = map_get(L, Blocks0), #b_blk{is=Is,last=Last} = Blk0, - Ys0 = maps:get(L, D0), + Ys0 = map_get(L, D0), {Yregs1,Ys} = find_yregs_is(Is, Ys0, Yregs0), Yregs = find_yregs_terminator(Last, Ys, Yregs1), Successors = beam_ssa:successors(Blk0), @@ -1175,7 +1320,7 @@ find_defs_1([L|Ls], Blocks, Frames, Seen0, Defs0, Acc0) -> false -> Seen1 = gb_sets:insert(L, Seen0), {Acc,Seen} = find_defs_1(Ls, Blocks, Frames, Seen1, Defs0, Acc0), - #b_blk{is=Is} = Blk = maps:get(L, Blocks), + #b_blk{is=Is} = Blk = map_get(L, Blocks), Defs = find_defs_is(Is, Defs0), Successors = beam_ssa:successors(Blk), find_defs_1(Successors, Blocks, Frames, Seen, Defs, Acc) @@ -1184,7 +1329,7 @@ find_defs_1([L|Ls], Blocks, Frames, Seen0, Defs0, Acc0) -> find_defs_1([], _, _, Seen, _, Acc) -> {Acc,Seen}. -find_defs_is([#b_set{dst=#b_var{name=Dst}}|Is], Acc) -> +find_defs_is([#b_set{dst=Dst}|Is], Acc) -> find_defs_is(Is, [Dst|Acc]); find_defs_is([], Acc) -> Acc. @@ -1194,15 +1339,15 @@ find_update_succ([S|Ss], #dk{d=Defs0,k=Killed0}=DK0, D0) -> Defs = ordsets:intersection(Defs0, Defs1), Killed = ordsets:union(Killed0, Killed1), DK = #dk{d=Defs,k=Killed}, - D = maps:put(S, DK, D0), + D = D0#{S:=DK}, find_update_succ(Ss, DK0, D); #{} -> - D = maps:put(S, DK0, D0), + D = D0#{S=>DK0}, find_update_succ(Ss, DK0, D) end; find_update_succ([], _, D) -> D. -find_yregs_is([#b_set{dst=#b_var{name=Dst}}=I|Is], #dk{d=Defs0,k=Killed0}=Ys, Yregs0) -> +find_yregs_is([#b_set{dst=Dst}=I|Is], #dk{d=Defs0,k=Killed0}=Ys, Yregs0) -> Used = beam_ssa:used(I), Yregs1 = ordsets:intersection(Used, Killed0), Yregs = ordsets:union(Yregs0, Yregs1), @@ -1287,7 +1432,7 @@ copy_retval(#st{frames=Frames,ssa=Blocks0,cnt=Count0}=St) -> St#st{ssa=Blocks,cnt=Count}. copy_retval_1([F|Fs], Blocks0, Count0) -> - #b_blk{anno=#{yregs:=Yregs0},is=Is} = maps:get(F, Blocks0), + #b_blk{anno=#{yregs:=Yregs0},is=Is} = map_get(F, Blocks0), Yregs1 = gb_sets:from_list(Yregs0), Yregs = collect_yregs(Is, Yregs1), Ls = beam_ssa:rpo([F], Blocks0), @@ -1296,7 +1441,7 @@ copy_retval_1([F|Fs], Blocks0, Count0) -> copy_retval_1([], Blocks, Count) -> {Blocks,Count}. -collect_yregs([#b_set{op=copy,dst=#b_var{name=Y},args=[#b_var{name=X}]}|Is], +collect_yregs([#b_set{op=copy,dst=Y,args=[#b_var{}=X]}|Is], Yregs0) -> true = gb_sets:is_member(X, Yregs0), %Assertion. Yregs = gb_sets:insert(Y, gb_sets:delete(X, Yregs0)), @@ -1306,7 +1451,7 @@ collect_yregs([#b_set{}|Is], Yregs) -> collect_yregs([], Yregs) -> Yregs. copy_retval_2([L|Ls], Yregs, Copy0, Blocks0, Count0) -> - #b_blk{is=Is0,last=Last} = Blk = maps:get(L, Blocks0), + #b_blk{is=Is0,last=Last} = Blk = map_get(L, Blocks0), RC = case {Last,Ls} of {#b_br{succ=Succ,fail=?BADARG_BLOCK},[Succ|_]} -> true; @@ -1333,17 +1478,20 @@ copy_retval_is([#b_set{op=put_tuple_elements,args=Args0}=I0], false, _Yregs, Copy, Count, Acc) -> I = I0#b_set{args=copy_sub_args(Args0, Copy)}, {reverse(Acc, [I|acc_copy([], Copy)]),Count}; +copy_retval_is([#b_set{op=Op}=I0], false, Yregs, Copy, Count0, Acc0) + when Op =:= call; Op =:= make_fun -> + {I,Count,Acc} = place_retval_copy(I0, Yregs, Copy, Count0, Acc0), + {reverse(Acc, [I]),Count}; copy_retval_is([#b_set{}]=Is, false, _Yregs, Copy, Count, Acc) -> {reverse(Acc, acc_copy(Is, Copy)),Count}; copy_retval_is([#b_set{},#b_set{op=succeeded}]=Is, false, _Yregs, Copy, Count, Acc) -> {reverse(Acc, acc_copy(Is, Copy)),Count}; -copy_retval_is([#b_set{op=Op,dst=#b_var{name=RetVal}=Dst}=I0|Is], RC, Yregs, +copy_retval_is([#b_set{op=Op,dst=#b_var{name=RetName}=Dst}=I0|Is], RC, Yregs, Copy0, Count0, Acc0) when Op =:= call; Op =:= make_fun -> {I1,Count1,Acc} = place_retval_copy(I0, Yregs, Copy0, Count0, Acc0), - case gb_sets:is_member(RetVal, Yregs) of + case gb_sets:is_member(Dst, Yregs) of true -> - {NewVarName,Count} = new_var_name(RetVal, Count1), - NewVar = #b_var{name=NewVarName}, + {NewVar,Count} = new_var(RetName, Count1), Copy = #b_set{op=copy,dst=Dst,args=[NewVar]}, I = I1#b_set{dst=NewVar}, copy_retval_is(Is, RC, Yregs, Copy, Count, [I|Acc]); @@ -1393,16 +1541,15 @@ copy_retval_is([], RC, _, Copy, Count, Acc) -> place_retval_copy(I, _Yregs, none, Count, Acc) -> {I,Count,Acc}; place_retval_copy(#b_set{args=[F|Args0]}=I, Yregs, Copy, Count0, Acc0) -> - #b_set{dst=#b_var{name=Avoid}} = Copy, + #b_set{dst=Avoid} = Copy, {Args,Acc1,Count} = copy_func_args(Args0, Yregs, Avoid, Acc0, [], Count0), Acc = [Copy|Acc1], {I#b_set{args=[F|Args]},Count,Acc}. -copy_func_args([#b_var{name=V}=A|As], Yregs, Avoid, CopyAcc, Acc, Count0) -> - case gb_sets:is_member(V, Yregs) of - true when V =/= Avoid -> - {NewVarName,Count} = new_var_name(V, Count0), - NewVar = #b_var{name=NewVarName}, +copy_func_args([#b_var{name=AName}=A|As], Yregs, Avoid, CopyAcc, Acc, Count0) -> + case gb_sets:is_member(A, Yregs) of + true when A =/= Avoid -> + {NewVar,Count} = new_var(AName, Count0), Copy = #b_set{op=copy,dst=NewVar,args=[A]}, copy_func_args(As, Yregs, Avoid, [Copy|CopyAcc], [NewVar|Acc], Count); _ -> @@ -1446,7 +1593,7 @@ opt_get_list(#st{ssa=Blocks,res=Res}=St) -> St#st{ssa=opt_get_list_1(Ls, ResMap, Blocks)}. opt_get_list_1([L|Ls], Res, Blocks0) -> - #b_blk{is=Is0} = Blk = maps:get(L, Blocks0), + #b_blk{is=Is0} = Blk = map_get(L, Blocks0), case opt_get_list_is(Is0, Res, [], false) of no -> opt_get_list_1(Ls, Res, Blocks0); @@ -1456,9 +1603,9 @@ opt_get_list_1([L|Ls], Res, Blocks0) -> end; opt_get_list_1([], _, Blocks) -> Blocks. -opt_get_list_is([#b_set{op=get_hd,dst=#b_var{name=Hd}, +opt_get_list_is([#b_set{op=get_hd,dst=Hd, args=[Cons]}=GetHd, - #b_set{op=get_tl,dst=#b_var{name=Tl}, + #b_set{op=get_tl,dst=Tl, args=[Cons]}=GetTl|Is], Res, Acc, Changed) -> %% Note that when this pass is run, only Y registers have @@ -1500,12 +1647,12 @@ number_instructions(#st{ssa=Blocks0}=St) -> St#st{ssa=number_is_1(Ls, 1, Blocks0)}. number_is_1([L|Ls], N0, Blocks0) -> - #b_blk{is=Is0,last=Last0} = Bl0 = maps:get(L, Blocks0), + #b_blk{is=Is0,last=Last0} = Bl0 = map_get(L, Blocks0), {Is,N1} = number_is_2(Is0, N0, []), Last = beam_ssa:add_anno(n, N1, Last0), N = N1 + 2, Bl = Bl0#b_blk{is=Is,last=Last}, - Blocks = maps:put(L, Bl, Blocks0), + Blocks = Blocks0#{L:=Bl}, number_is_1(Ls, N, Blocks); number_is_1([], _, Blocks) -> Blocks. @@ -1522,13 +1669,13 @@ number_is_2([], N, Acc) -> %%% live_intervals(#st{args=Args,ssa=Blocks}=St) -> - Vars0 = [{V,{0,1}} || #b_var{name=V} <- Args], + Vars0 = [{V,{0,1}} || #b_var{}=V <- Args], F = fun(L, _, A) -> live_interval_blk(L, Blocks, A) end, LiveMap0 = #{}, - Acc0 = {[],[],LiveMap0}, - {Vars,Aliases,_} = beam_ssa:fold_po(F, Acc0, Blocks), + Acc0 = {[],LiveMap0}, + {Vars,_} = beam_ssa:fold_po(F, Acc0, Blocks), Intervals = merge_ranges(rel2fam(Vars0++Vars)), - St#st{intervals=Intervals,aliases=Aliases}. + St#st{intervals=Intervals}. merge_ranges([{V,Rs}|T]) -> [{V,merge_ranges_1(Rs)}|merge_ranges(T)]; @@ -1540,20 +1687,19 @@ merge_ranges_1([R|Rs]) -> [R|merge_ranges_1(Rs)]; merge_ranges_1([]) -> []. -live_interval_blk(L, Blocks, {Vars0,Aliases0,LiveMap0}) -> +live_interval_blk(L, Blocks, {Vars0,LiveMap0}) -> Live0 = [], Successors = beam_ssa:successors(L, Blocks), Live1 = update_successors(Successors, L, Blocks, LiveMap0, Live0), %% Add ranges for all variables that are live in the successors. - #b_blk{is=Is,last=Last} = maps:get(L, Blocks), + #b_blk{is=Is,last=Last} = map_get(L, Blocks), End = beam_ssa:get_anno(n, Last), Use = [{V,{use,End+1}} || V <- Live1], %% Determine used and defined variables in this block. FirstNumber = first_number(Is, Last), - {UseDef0,Aliases} = live_interval_blk_1([Last|reverse(Is)], - FirstNumber, Aliases0, Use), + UseDef0 = live_interval_blk_1([Last|reverse(Is)], FirstNumber, Use), UseDef = rel2fam(UseDef0), %% Update what is live at the beginning of this block and @@ -1566,7 +1712,7 @@ live_interval_blk(L, Blocks, {Vars0,Aliases0,LiveMap0}) -> %% Construct the ranges for this block. Vars = make_block_ranges(UseDef, FirstNumber, Vars0), - {Vars,Aliases,LiveMap}. + {Vars,LiveMap}. make_block_ranges([{V,[{def,Def}]}|Vs], First, Acc) -> make_block_ranges(Vs, First, [{V,{Def,Def}}|Acc]); @@ -1578,37 +1724,29 @@ make_block_ranges([{V,[{use,_}|_]=Uses}|Vs], First, Acc) -> make_block_ranges(Vs, First, [{V,{First,Last}}|Acc]); make_block_ranges([], _, Acc) -> Acc. -live_interval_blk_1([#b_set{op=phi,dst=#b_var{name=Dst}}|Is], - FirstNumber, Aliases, Acc0) -> +live_interval_blk_1([#b_set{op=phi,dst=Dst}|Is], FirstNumber, Acc0) -> Acc = [{Dst,{def,FirstNumber}}|Acc0], - live_interval_blk_1(Is, FirstNumber, Aliases, Acc); -live_interval_blk_1([#b_set{op=bs_start_match}=I|Is], FirstNumber, - Aliases0, Acc0) -> + live_interval_blk_1(Is, FirstNumber, Acc); +live_interval_blk_1([#b_set{op=bs_start_match}=I|Is], + FirstNumber, Acc0) -> N = beam_ssa:get_anno(n, I), - #b_set{dst=#b_var{name=Dst}} = I, + #b_set{dst=Dst} = I, Acc1 = [{Dst,{def,N}}|Acc0], - Aliases = case beam_ssa:get_anno(reuse_for_context, I) of - true -> - #b_set{args=[#b_var{name=Src}]} = I, - [{Dst,Src}|Aliases0]; - false -> - Aliases0 - end, Acc = [{V,{use,N}} || V <- beam_ssa:used(I)] ++ Acc1, - live_interval_blk_1(Is, FirstNumber, Aliases, Acc); -live_interval_blk_1([I|Is], FirstNumber, Aliases, Acc0) -> + live_interval_blk_1(Is, FirstNumber, Acc); +live_interval_blk_1([I|Is], FirstNumber, Acc0) -> N = beam_ssa:get_anno(n, I), Acc1 = case I of - #b_set{dst=#b_var{name=Dst}} -> + #b_set{dst=Dst} -> [{Dst,{def,N}}|Acc0]; _ -> Acc0 end, Used = beam_ssa:used(I), Acc = [{V,{use,N}} || V <- Used] ++ Acc1, - live_interval_blk_1(Is, FirstNumber, Aliases, Acc); -live_interval_blk_1([], _FirstNumber, Aliases, Acc) -> - {Acc,Aliases}. + live_interval_blk_1(Is, FirstNumber, Acc); +live_interval_blk_1([], _FirstNumber, Acc) -> + Acc. %% first_number([#b_set{}]) -> InstructionNumber. %% Return the number for the first instruction for the block. @@ -1624,7 +1762,7 @@ first_number([], Last) -> update_successors([L|Ls], Pred, Blocks, LiveMap, Live0) -> Live1 = ordsets:union(Live0, get_live(L, LiveMap)), - #b_blk{is=Is} = maps:get(L, Blocks), + #b_blk{is=Is} = map_get(L, Blocks), Live = update_live_phis(Is, Pred, Live1), update_successors(Ls, Pred, Blocks, LiveMap, Live); update_successors([], _, _, _, Live) -> Live. @@ -1635,9 +1773,9 @@ get_live(L, LiveMap) -> #{} -> [] end. -update_live_phis([#b_set{op=phi,dst=#b_var{name=Killed},args=Args}|Is], +update_live_phis([#b_set{op=phi,dst=Killed,args=Args}|Is], Pred, Live0) -> - Used = [V || {#b_var{name=V},L} <- Args, L =:= Pred], + Used = [V || {#b_var{}=V,L} <- Args, L =:= Pred], Live1 = ordsets:union(ordsets:from_list(Used), Live0), Live = ordsets:del_element(Killed, Live1), update_live_phis(Is, Pred, Live); @@ -1650,7 +1788,7 @@ update_live_phis(_, _, Live) -> Live. %% reserve_yregs(St0) -> St. %% In each block that allocates a stack frame, insert instructions %% that copy variables that must be in Y registers (given by -%% YRegisters) to new variables. +%% the `yregs` annotation) to new variables. %% %% Also allocate specific Y registers for try and catch tags. %% The outermost try/catch tag is placed in y0, any directly @@ -1662,7 +1800,7 @@ reserve_yregs(#st{frames=Frames}=St0) -> foldl(fun reserve_yregs_1/2, St0, Frames). reserve_yregs_1(L, #st{ssa=Blocks0,cnt=Count0,res=Res0}=St) -> - Blk = maps:get(L, Blocks0), + Blk = map_get(L, Blocks0), Yregs = beam_ssa:get_anno(yregs, Blk), {Def,Used} = beam_ssa:def_used([L], Blocks0), UsedYregs = ordsets:intersection(Yregs, Used), @@ -1688,7 +1826,7 @@ reserve_try_tags_1([L|Ls], Blocks, Seen0, ActMap0) -> reserve_try_tags_1(Ls, Blocks, Seen0, ActMap0); false -> Seen1 = gb_sets:insert(L, Seen0), - #b_blk{is=Is} = Blk = maps:get(L, Blocks), + #b_blk{is=Is} = Blk = map_get(L, Blocks), Active0 = get_active(L, ActMap0), Active = reserve_try_tags_is(Is, Active0), Successors = beam_ssa:successors(Blk), @@ -1705,10 +1843,10 @@ get_active(L, ActMap) -> #{} -> #{} end. -reserve_try_tags_is([#b_set{op=new_try_tag,dst=#b_var{name=V}}|Is], Active) -> +reserve_try_tags_is([#b_set{op=new_try_tag,dst=V}|Is], Active) -> N = map_size(Active), reserve_try_tags_is(Is, Active#{V=>N}); -reserve_try_tags_is([#b_set{op=kill_try_tag,args=[#b_var{name=Tag}]}|Is], Active) -> +reserve_try_tags_is([#b_set{op=kill_try_tag,args=[Tag]}|Is], Active) -> reserve_try_tags_is(Is, maps:remove(Tag, Active)); reserve_try_tags_is([_|Is], Active) -> reserve_try_tags_is(Is, Active); @@ -1728,17 +1866,15 @@ update_act_map([], _, ActMap) -> ActMap. rename_vars([], _, Blocks, Count) -> {[],Blocks,Count}; rename_vars(Vs, L, Blocks0, Count0) -> - {NewVs,Count} = new_var_names(Vs, Count0), - NewVars = [#b_var{name=V} || V <- NewVs], + {NewVars,Count} = new_vars([Base || #b_var{name=Base} <- Vs], Count0), Ren = zip(Vs, NewVars), Blocks1 = beam_ssa:rename_vars(Ren, [L], Blocks0), - #b_blk{is=Is0} = Blk0 = maps:get(L, Blocks1), - CopyIs = [#b_set{op=copy,dst=New,args=[#b_var{name=Old}]} || - {Old,New} <- Ren], + #b_blk{is=Is0} = Blk0 = map_get(L, Blocks1), + CopyIs = [#b_set{op=copy,dst=New,args=[Old]} || {Old,New} <- Ren], Is = insert_after_phis(Is0, CopyIs), Blk = Blk0#b_blk{is=Is}, - Blocks = maps:put(L, Blk, Blocks1), - {NewVs,Blocks,Count}. + Blocks = Blocks1#{L:=Blk}, + {NewVars,Blocks,Count}. insert_after_phis([#b_set{op=phi}=I|Is], InsertIs) -> [I|insert_after_phis(Is, InsertIs)]; @@ -1759,7 +1895,7 @@ frame_size(#st{frames=Frames,regs=Regs,ssa=Blocks0}=St) -> frame_size_1(L, Regs, Blocks0) -> Def = beam_ssa:def([L], Blocks0), - Yregs0 = [maps:get(V, Regs) || V <- Def, is_yreg(maps:get(V, Regs))], + Yregs0 = [map_get(V, Regs) || V <- Def, is_yreg(map_get(V, Regs))], Yregs = ordsets:from_list(Yregs0), FrameSize = length(ordsets:from_list(Yregs)), if @@ -1771,17 +1907,17 @@ frame_size_1(L, Regs, Blocks0) -> true -> ok end, - Blk0 = maps:get(L, Blocks0), + Blk0 = map_get(L, Blocks0), Blk = beam_ssa:add_anno(frame_size, FrameSize, Blk0), %% Insert an annotation for frame deallocation on %% each #b_ret{}. - Blocks = maps:put(L, Blk, Blocks0), + Blocks = Blocks0#{L:=Blk}, Reachable = beam_ssa:rpo([L], Blocks), frame_deallocate(Reachable, FrameSize, Blocks). frame_deallocate([L|Ls], Size, Blocks0) -> - Blk0 = maps:get(L, Blocks0), + Blk0 = map_get(L, Blocks0), Blk = case Blk0 of #b_blk{last=#b_ret{}=Ret0} -> Ret = beam_ssa:add_anno(deallocate, Size, Ret0), @@ -1789,7 +1925,7 @@ frame_deallocate([L|Ls], Size, Blocks0) -> #b_blk{} -> Blk0 end, - Blocks = maps:put(L, Blk, Blocks0), + Blocks = Blocks0#{L:=Blk}, frame_deallocate(Ls, Size, Blocks); frame_deallocate([], _, Blocks) -> Blocks. @@ -1802,7 +1938,7 @@ frame_deallocate([], _, Blocks) -> Blocks. turn_yregs(#st{frames=Frames,regs=Regs0,ssa=Blocks}=St) -> Regs1 = foldl(fun(L, A) -> - Blk = maps:get(L, Blocks), + Blk = map_get(L, Blocks), FrameSize = beam_ssa:get_anno(frame_size, Blk), Def = beam_ssa:def([L], Blocks), [turn_yregs_1(Def, FrameSize, Regs0)|A] @@ -1811,7 +1947,7 @@ turn_yregs(#st{frames=Frames,regs=Regs0,ssa=Blocks}=St) -> St#st{regs=Regs}. turn_yregs_1(Def, FrameSize, Regs) -> - Yregs0 = [{maps:get(V, Regs),V} || V <- Def, is_yreg(maps:get(V, Regs))], + Yregs0 = [{map_get(V, Regs),V} || V <- Def, is_yreg(map_get(V, Regs))], Yregs1 = rel2fam(Yregs0), FrameSize = length(Yregs1), Yregs2 = [{{y,FrameSize-Y-1},Vs} || {{y,Y},Vs} <- Yregs1], @@ -1845,7 +1981,7 @@ reserve_regs(#st{args=Args,ssa=Blocks,intervals=Intervals,res=Res0}=St) -> Res = maps:from_list(Res3), St#st{res=reserve_xregs(Blocks, Res)}. -reserve_arg_regs([#b_var{name=Arg}|Is], N, Acc) -> +reserve_arg_regs([#b_var{}=Arg|Is], N, Acc) -> reserve_arg_regs(Is, N+1, [{Arg,{x,N}}|Acc]); reserve_arg_regs([], _, Acc) -> Acc. @@ -1857,25 +1993,41 @@ reserve_zregs(Blocks, Intervals, Res) -> end, beam_ssa:fold_rpo(F, [0], Res, Blocks). +reserve_zreg([#b_set{op=call,dst=Dst}], + #b_br{bool=Dst}, _ShortLived, A) -> + %% If type optimization has determined that the result of a call can be + %% used directly in a branch, we must avoid reserving a z register or code + %% generation will fail. + A; reserve_zreg([#b_set{op={bif,tuple_size},dst=Dst}, - #b_set{op={bif,'=:='},args=[Dst,Val]}], _Last, ShortLived, A0) -> - case Val of - #b_literal{val=Arity} when Arity bsr 32 =:= 0 -> + #b_set{op={bif,'=:='},args=[Dst,Val]}], Last, ShortLived, A0) -> + case {Val,Last} of + {#b_literal{val=Arity},#b_br{bool=#b_var{}}} when Arity bsr 32 =:= 0 -> %% These two instructions can be combined to a test_arity %% instruction provided that the arity variable is short-lived. reserve_zreg_1(Dst, ShortLived, A0); - _ -> + {_,_} -> + %% Either the arity is too big, or the boolean value is not + %% used in a conditional branch. A0 end; reserve_zreg([#b_set{op={bif,tuple_size},dst=Dst}], #b_switch{}, ShortLived, A) -> reserve_zreg_1(Dst, ShortLived, A); -reserve_zreg([#b_set{op=Op,dst=#b_var{name=Dst}}|Is], Last, ShortLived, A0) -> +reserve_zreg([#b_set{op={bif,'xor'}}], _Last, _ShortLived, A) -> + %% There is no short, easy way to rewrite 'xor' to a series of + %% test instructions. + A; +reserve_zreg([#b_set{op={bif,is_record}}], _Last, _ShortLived, A) -> + %% There is no short, easy way to rewrite is_record/2 to a series of + %% test instructions. + A; +reserve_zreg([#b_set{op=Op,dst=Dst}|Is], Last, ShortLived, A0) -> IsZReg = case Op of - context_to_binary -> true; bs_match_string -> true; - bs_restore -> true; bs_save -> true; + bs_restore -> true; + bs_set_position -> true; {float,clearerror} -> true; kill_try_tag -> true; landingpad -> true; @@ -1896,7 +2048,7 @@ reserve_zreg([], #b_br{bool=Bool}, ShortLived, A) -> reserve_zreg_1(Bool, ShortLived, A); reserve_zreg([], _, _, A) -> A. -reserve_zreg_1(#b_var{name=V}, ShortLived, A) -> +reserve_zreg_1(#b_var{}=V, ShortLived, A) -> case cerl_sets:is_element(V, ShortLived) of true -> [{V,z}|A]; false -> A @@ -1909,7 +2061,7 @@ reserve_fregs(Blocks, Res) -> end, beam_ssa:fold_rpo(F, [0], Res, Blocks). -reserve_freg([#b_set{op={float,Op},dst=#b_var{name=V}}|Is], Res) -> +reserve_freg([#b_set{op={float,Op},dst=V}|Is], Res) -> case Op of get -> reserve_freg(Is, Res); @@ -1934,23 +2086,95 @@ reserve_freg([], Res) -> Res. %% will allocate the lowest free X register for the variable. reserve_xregs(Blocks, Res) -> - F = fun(L, #b_blk{is=Is,last=Last}, R) -> - {Xs0,Used0} = reserve_terminator(L, Last, Blocks, R), - reserve_xregs_is(reverse(Is), R, Xs0, Used0) - end, - beam_ssa:fold_po(F, Res, Blocks). - -reserve_xregs_is([#b_set{op=Op,dst=#b_var{name=Dst},args=Args}=I|Is], Res0, Xs0, Used0) -> - Xs1 = case is_gc_safe(I) of - true -> - Xs0; - false -> - %% There may be a garbage collection after executing this - %% instruction. We will need prune the list of preferred - %% X registers. - res_xregs_prune(Xs0, Used0, Res0) - end, - Res = reserve_xreg(Dst, Xs1, Res0), + Ls = reverse(beam_ssa:rpo(Blocks)), + reserve_xregs(Ls, Blocks, #{}, Res). + +reserve_xregs([L|Ls], Blocks, XsMap0, Res0) -> + #b_blk{anno=Anno,is=Is0,last=Last} = map_get(L, Blocks), + + %% Calculate mapping from variable name to the preferred + %% register. + Xs0 = reserve_terminator(L, Is0, Last, Blocks, XsMap0, Res0), + + %% We need to figure out where the code generator will + %% place instructions that will do a garbage collection. + %% Insert 'gc' markers as pseudo-instructions in the + %% instruction sequence. + Is1 = reverse(Is0), + Is2 = res_place_gc_instrs(Is1, []), + Is = res_place_allocate(Anno, Is2), + + %% Add register hints for variables that are defined + %% in the (reversed) instruction sequence. + {Res,Xs} = reserve_xregs_is(Is, Res0, Xs0, []), + + XsMap = XsMap0#{L=>Xs}, + reserve_xregs(Ls, Blocks, XsMap, Res); +reserve_xregs([], _, _, Res) -> Res. + +%% Insert explicit 'gc' markers points where there will +%% be a garbage collection. (Note that the instruction +%% sequence passed to this function is reversed.) + +res_place_gc_instrs([#b_set{op=phi}=I|Is], Acc) -> + res_place_gc_instrs(Is, [I|Acc]); +res_place_gc_instrs([#b_set{op=Op}=I|Is], Acc) + when Op =:= call; Op =:= make_fun -> + case Acc of + [] -> + res_place_gc_instrs(Is, [I|Acc]); + [GC|_] when GC =:= gc; GC =:= test_heap -> + res_place_gc_instrs(Is, [I,gc|Acc]); + [_|_] -> + res_place_gc_instrs(Is, [I,gc|Acc]) + end; +res_place_gc_instrs([#b_set{op=Op,args=Args}=I|Is], Acc0) -> + case beam_ssa_codegen:classify_heap_need(Op, Args) of + neutral -> + case Acc0 of + [test_heap|Acc] -> + res_place_gc_instrs(Is, [test_heap,I|Acc]); + Acc -> + res_place_gc_instrs(Is, [I|Acc]) + end; + {put,_} -> + case Acc0 of + [test_heap|Acc] -> + res_place_gc_instrs(Is, [test_heap,I|Acc]); + Acc -> + res_place_gc_instrs(Is, [test_heap,I|Acc]) + end; + _ -> + res_place_gc_instrs(Is, [gc,I|Acc0]) + end; +res_place_gc_instrs([], Acc) -> + %% Reverse and replace 'test_heap' markers with 'gc'. + %% (The distinction is no longer useful.) + res_place_gc_instrs_rev(Acc, []). + +res_place_gc_instrs_rev([test_heap|Is], [gc|_]=Acc) -> + res_place_gc_instrs_rev(Is, Acc); +res_place_gc_instrs_rev([test_heap|Is], Acc) -> + res_place_gc_instrs_rev(Is, [gc|Acc]); +res_place_gc_instrs_rev([gc|Is], [gc|_]=Acc) -> + res_place_gc_instrs_rev(Is, Acc); +res_place_gc_instrs_rev([I|Is], Acc) -> + res_place_gc_instrs_rev(Is, [I|Acc]); +res_place_gc_instrs_rev([], Acc) -> Acc. + +res_place_allocate(#{yregs:=_}, Is) -> + %% There will be an 'allocate' instruction inserted here. + Is ++ [gc]; +res_place_allocate(#{}, Is) -> Is. + +reserve_xregs_is([gc|Is], Res, Xs0, Used) -> + %% At this point, the code generator will place an instruction + %% that does a garbage collection. We must prune the remembered + %% registers. + Xs = res_xregs_prune(Xs0, Used, Res), + reserve_xregs_is(Is, Res, Xs, Used); +reserve_xregs_is([#b_set{op=Op,dst=Dst,args=Args}=I|Is], Res0, Xs0, Used0) -> + Res = reserve_xreg(Dst, Xs0, Res0), Used1 = ordsets:union(Used0, beam_ssa:used(I)), Used = ordsets:del_element(Dst, Used1), case Op of @@ -1961,28 +2185,74 @@ reserve_xregs_is([#b_set{op=Op,dst=#b_var{name=Dst},args=Args}=I|Is], Res0, Xs0, Xs = reserve_call_args(tl(Args)), reserve_xregs_is(Is, Res, Xs, Used); _ -> - reserve_xregs_is(Is, Res, Xs1, Used) + reserve_xregs_is(Is, Res, Xs0, Used) end; -reserve_xregs_is([], Res, _Xs, _Used) -> Res. - -reserve_terminator(L, #b_br{bool=#b_literal{val=true},succ=Succ}, Blocks, Res) -> - case maps:get(Succ, Blocks) of +reserve_xregs_is([], Res, Xs, _Used) -> + {Res,Xs}. + +%% Pick up register hints from the successors of this blocks. +reserve_terminator(_L, _Is, #b_br{bool=#b_var{},succ=Succ,fail=?BADARG_BLOCK}, + _Blocks, XsMap, _Res) -> + %% We know that no variables are used at ?BADARG_BLOCK, so + %% any register hints from the success blocks are safe to use. + map_get(Succ, XsMap); +reserve_terminator(L, Is, #b_br{bool=#b_var{},succ=Succ,fail=Fail}, + Blocks, XsMap, Res) when Succ =/= Fail -> + #{Succ:=SuccBlk,Fail:=FailBlk} = Blocks, + case {SuccBlk,FailBlk} of + {#b_blk{is=[],last=#b_br{succ=PhiL,fail=PhiL}}, + #b_blk{is=[],last=#b_br{succ=PhiL,fail=PhiL}}} -> + %% Both branches ultimately transfer to the same + %% block (via two blocks with no instructions). + %% Pick up register hints from the phi nodes + %% in the common block. + #{PhiL:=#b_blk{is=PhiIs}} = Blocks, + Xs = res_xregs_from_phi(PhiIs, Succ, Res, #{}), + res_xregs_from_phi(PhiIs, Fail, Res, Xs); + {_,_} when Is =/= [] -> + case last(Is) of + #b_set{op=succeeded,args=[Arg]} -> + %% We know that Arg will not be used at the failure + %% label, so we can pick up register hints from the + %% success label. + Br = #b_br{bool=#b_literal{val=true},succ=Succ,fail=Succ}, + case reserve_terminator(L, [], Br, Blocks, XsMap, Res) of + #{Arg:=Reg} -> #{Arg=>Reg}; + #{} -> #{} + end; + _ -> + %% Register hints from the success block may not + %% be safe at the failure block, and vice versa. + #{} + end; + {_,_} -> + %% Register hints from the success block may not + %% be safe at the failure block, and vice versa. + #{} + end; +reserve_terminator(L, Is, #b_br{bool=#b_literal{val=true},succ=Succ}, + Blocks, XsMap, Res) -> + case map_get(Succ, Blocks) of #b_blk{is=[],last=Last} -> - reserve_terminator(Succ, Last, Blocks, Res); - #b_blk{is=[_|_]=Is} -> - {res_xregs_from_phi(Is, L, Res, #{}),[]} + reserve_terminator(Succ, Is, Last, Blocks, XsMap, Res); + #b_blk{is=[_|_]=PhiIs} -> + res_xregs_from_phi(PhiIs, L, Res, #{}) end; -reserve_terminator(_, Last, _, _) -> - {#{},beam_ssa:used(Last)}. +reserve_terminator(_, _, _, _, _, _) -> #{}. -res_xregs_from_phi([#b_set{op=phi,dst=#b_var{name=Dst},args=Args}|Is], +%% Pick up a reservation from a phi node. +res_xregs_from_phi([#b_set{op=phi,dst=Dst,args=Args}|Is], Pred, Res, Acc) -> - case [V || {#b_var{name=V},L} <- Args, L =:= Pred] of + case [V || {#b_var{}=V,L} <- Args, L =:= Pred] of [] -> + %% The value of the phi node for this predecessor + %% is a literal. Nothing to do here. res_xregs_from_phi(Is, Pred, Res, Acc); [V] -> case Res of #{Dst:={prefer,Reg}} -> + %% Try placing V in the same register as for + %% the phi node. res_xregs_from_phi(Is, Pred, Res, Acc#{V=>Reg}); #{Dst:=_} -> res_xregs_from_phi(Is, Pred, Res, Acc) @@ -1993,8 +2263,8 @@ res_xregs_from_phi(_, _, _, Acc) -> Acc. reserve_call_args(Args) -> reserve_call_args(Args, 0, #{}). -reserve_call_args([#b_var{name=Name}|As], X, Xs) -> - reserve_call_args(As, X+1, Xs#{Name=>{x,X}}); +reserve_call_args([#b_var{}=Var|As], X, Xs) -> + reserve_call_args(As, X+1, Xs#{Var=>{x,X}}); reserve_call_args([#b_literal{}|As], X, Xs) -> reserve_call_args(As, X+1, Xs); reserve_call_args([], _, Xs) -> Xs. @@ -2002,12 +2272,12 @@ reserve_call_args([], _, Xs) -> Xs. reserve_xreg(V, Xs, Res) -> case Res of #{V:=_} -> - %% Already reserved. + %% Already reserved (but not as an X register). Res; #{} -> case Xs of #{V:=X} -> - %% Add a hint that a specific X register is + %% Add a hint that this specific X register is %% preferred, unless it is already in use. Res#{V=>{prefer,X}}; #{} -> @@ -2016,23 +2286,15 @@ reserve_xreg(V, Xs, Res) -> end end. -is_gc_safe(#b_set{op=phi}) -> - false; -is_gc_safe(#b_set{op=Op,args=Args}) -> - case beam_ssa_codegen:classify_heap_need(Op, Args) of - neutral -> true; - {put,_} -> true; - _ -> false - end. - %% res_xregs_prune(PreferredRegs, Used, Res) -> PreferredRegs. -%% Prune the list of preferred to only include X registers that -%% are guaranteed to survice a garbage collection. +%% Prune the list of preferred registers, to make sure that +%% there are no "holes" (uninitialized X registers) when +%% invoking the garbage collector. -res_xregs_prune(Xs, Used, Res) -> +res_xregs_prune(Xs, Used, Res) when map_size(Xs) =/= 0 -> %% The number of safe registers is the number of the X registers %% used after this point. The actual number of safe registers may - %% be highter than this number, but this is a conservative safe + %% be higher than this number, but this is a conservative safe %% estimate. NumSafe = foldl(fun(V, N) -> case Res of @@ -2044,83 +2306,8 @@ res_xregs_prune(Xs, Used, Res) -> %% Remove unsafe registers from the list of potential %% preferred registers. - maps:filter(fun(_, {x,X}) -> X < NumSafe end, Xs). - -%%% -%%% Remove unsuitable aliases. -%%% -%%% If a binary is matched more than once, we must not put the -%%% the match context in the same register as the binary to -%%% avoid the following situation: -%%% -%%% {test,bs_start_match2,{f,3},1,[{x,0},0],{x,0}}. -%%% . -%%% . -%%% . -%%% {test,bs_start_match2,{f,6},1,[{x,0},0],{x,1}}. %% ILLEGAL! -%%% -%%% The second instruction is illegal because a match context source -%%% is only allowed if source and destination registers are identical. -%%% - -remove_unsuitable_aliases(#st{aliases=[_|_]=Aliases0,ssa=Blocks}=St) -> - R = rem_unsuitable(maps:values(Blocks)), - Unsuitable0 = [V || {V,[_,_|_]} <- rel2fam(R)], - Unsuitable = gb_sets:from_list(Unsuitable0), - Aliases =[P || {_,V}=P <- Aliases0, - not gb_sets:is_member(V, Unsuitable)], - St#st{aliases=Aliases}; -remove_unsuitable_aliases(#st{aliases=[]}=St) -> St. - -rem_unsuitable([#b_blk{is=Is}|Bs]) -> - Vs = [{V,Dst} || - #b_set{op=bs_start_match,dst=#b_var{name=Dst}, - args=[#b_var{name=V}]} <- Is], - Vs ++ rem_unsuitable(Bs); -rem_unsuitable([]) -> []. - -%%% -%%% Merge intervals. -%%% - -merge_intervals(#st{aliases=Aliases0,intervals=Intervals0, - res=Reserved}=St) -> - Aliases1 = [A || A <- Aliases0, - is_suitable_alias(A, Reserved)], - case Aliases1 of - [] -> - St#st{aliases=Aliases1}; - [_|_] -> - Intervals1 = maps:from_list(Intervals0), - {Intervals,Aliases} = - merge_intervals_1(Aliases1, Intervals1, []), - St#st{aliases=Aliases,intervals=Intervals} - end. - -merge_intervals_1([{Alias,V}|Vs], Intervals0, Acc) -> - #{Alias:=Int1,V:=Int2} = Intervals0, - Int3 = lists:merge(Int1, Int2), - Int = merge_intervals_2(Int3), - Intervals1 = maps:remove(Alias, Intervals0), - Intervals = Intervals1#{V:=Int}, - merge_intervals_1(Vs, Intervals, [{Alias,V}|Acc]); -merge_intervals_1([], Intervals, Acc) -> - {maps:to_list(Intervals),Acc}. - -merge_intervals_2([{A1,B1},{A2,B2}|Is]) when A2 =< B1 -> - merge_intervals_2([{min(A1, A2),max(B1, B2)}|Is]); -merge_intervals_2([{_A1,B1}=R|[{A2,_B2}|_]=Is]) when B1 < A2 -> - [R|merge_intervals_2(Is)]; -merge_intervals_2([_]=Is) -> Is. - -is_suitable_alias({V1,V2}, Reserved) -> - #{V1:=Res1,V2:=Res2} = Reserved, - case {Res1,Res2} of - {x,x} -> true; - {x,{x,_}} -> true; - {{x,_},x} -> true; - {_,_} -> false - end. + maps:filter(fun(_, {x,X}) -> X < NumSafe end, Xs); +res_xregs_prune(Xs, _Used, _Res) -> Xs. %%% %%% Register allocation using linear scan. @@ -2155,7 +2342,13 @@ linear_scan(#st{intervals=Intervals0,res=Res}=St0) -> Free = init_free(maps:to_list(Res)), Intervals1 = [init_interval(Int, Res) || Int <- Intervals0], Intervals = sort(Intervals1), - IsReserved = fun (#i{reg=Reg}) -> Reg =/= none end, + IsReserved = fun(#i{reg=Reg}) -> + case Reg of + none -> false; + {prefer,{_,_}} -> false; + {_,_} -> true + end + end, {UnhandledRes,Unhandled} = partition(IsReserved, Intervals), L = #l{unhandled_res=UnhandledRes, unhandled_any=Unhandled,free=Free}, @@ -2163,7 +2356,7 @@ linear_scan(#st{intervals=Intervals0,res=Res}=St0) -> St#st{regs=maps:from_list(Regs)}. init_interval({V,[{Start,_}|_]=Rs}, Res) -> - Info = maps:get(V, Res), + Info = map_get(V, Res), Pool = case Info of {prefer,{x,_}} -> x; x -> x; @@ -2364,16 +2557,16 @@ free_reg(#i{reg={_,_}=Reg}=I, L) -> update_pool(I, FreeRegs, L). get_pool(#i{pool=Pool}, #l{free=Free}) -> - maps:get(Pool, Free). + map_get(Pool, Free). update_pool(#i{pool=Pool}, New, #l{free=Free0}=L) -> - Free = maps:put(Pool, New, Free0), + Free = Free0#{Pool:=New}, L#l{free=Free}. get_next_free(#i{pool=Pool}, #l{free=Free0}=L0) -> K = {next,Pool}, - N = maps:get(K, Free0), - Free = maps:put(K, N+1, Free0), + N = map_get(K, Free0), + Free = Free0#{K:=N+1}, L = L0#l{free=Free}, if is_integer(Pool) -> {{y,N},L}; @@ -2409,7 +2602,7 @@ are_overlapping_1({_,_}, []) -> false. is_loop_header(L, Blocks) -> %% We KNOW that a loop header must start with a peek_message %% instruction. - case maps:get(L, Blocks) of + case map_get(L, Blocks) of #b_blk{is=[#b_set{op=peek_message}|_]} -> true; _ -> false end. @@ -2420,21 +2613,21 @@ rel2fam(S0) -> sofs:to_external(S). split_phis(Is) -> - partition(fun(#b_set{op=Op}) -> Op =:= phi end, Is). + splitwith(fun(#b_set{op=Op}) -> Op =:= phi end, Is). is_yreg({y,_}) -> true; is_yreg({x,_}) -> false; is_yreg({z,_}) -> false; is_yreg({fr,_}) -> false. -new_var_names([V0|Vs0], Count0) -> - {V,Count1} = new_var_name(V0, Count0), - {Vs,Count} = new_var_names(Vs0, Count1), +new_vars([Base|Vs0], Count0) -> + {V,Count1} = new_var(Base, Count0), + {Vs,Count} = new_vars(Vs0, Count1), {[V|Vs],Count}; -new_var_names([], Count) -> {[],Count}. +new_vars([], Count) -> {[],Count}. -new_var_name({Base,Int}, Count) -> +new_var({Base,Int}, Count) -> true = is_integer(Int), %Assertion. - {{Base,Count},Count+1}; -new_var_name(Base, Count) -> - {{Base,Count},Count+1}. + {#b_var{name={Base,Count}},Count+1}; +new_var(Base, Count) -> + {#b_var{name={Base,Count}},Count+1}. diff --git a/lib/compiler/src/beam_ssa_recv.erl b/lib/compiler/src/beam_ssa_recv.erl index 82fe006487..1e0e1ecac2 100644 --- a/lib/compiler/src/beam_ssa_recv.erl +++ b/lib/compiler/src/beam_ssa_recv.erl @@ -101,7 +101,7 @@ opt([{L,#b_blk{is=[#b_set{op=peek_message}|_]}=Blk0}|Bs], Blocks0, Preds) -> case recv_opt(Preds, L, Blocks0) of {yes,Blocks1} -> Blk = beam_ssa:add_anno(recv_set, L, Blk0), - Blocks = maps:put(L, Blk, Blocks1), + Blocks = Blocks1#{L:=Blk}, opt(Bs, Blocks, []); no -> opt(Bs, Blocks0, []) @@ -111,11 +111,11 @@ opt([{L,_}|Bs], Blocks, Preds) -> opt([], Blocks, _) -> Blocks. recv_opt([L|Ls], RecvLbl, Blocks) -> - #b_blk{is=Is0} = Blk0 = maps:get(L, Blocks), + #b_blk{is=Is0} = Blk0 = map_get(L, Blocks), case recv_opt_is(Is0, RecvLbl, Blocks, []) of {yes,Is} -> Blk = Blk0#b_blk{is=Is}, - {yes,maps:put(L, Blk, Blocks)}; + {yes,Blocks#{L:=Blk}}; no -> recv_opt(Ls, RecvLbl, Blocks) end; @@ -138,7 +138,7 @@ recv_opt_is([I|Is], RecvLbl, Blocks, Acc) -> recv_opt_is(Is, RecvLbl, Blocks, [I|Acc]); recv_opt_is([], _, _, _) -> no. -makes_ref(#b_set{dst=#b_var{name=Dst},args=[Func0|_]}, Blocks) -> +makes_ref(#b_set{dst=Dst,args=[Func0|_]}, Blocks) -> Func = case Func0 of #b_remote{mod=#b_literal{val=erlang}, name=#b_literal{val=Name},arity=A0} -> @@ -158,15 +158,15 @@ makes_ref(#b_set{dst=#b_var{name=Dst},args=[Func0|_]}, Blocks) -> end. ref_in_tuple(Tuple, Blocks) -> - F = fun(#b_set{op=get_tuple_element,dst=#b_var{name=Ref}, - args=[#b_var{name=Tup},#b_literal{val=1}]}, no) + F = fun(#b_set{op=get_tuple_element,dst=Ref, + args=[#b_var{}=Tup,#b_literal{val=1}]}, no) when Tup =:= Tuple -> {yes,Ref}; (_, A) -> A end, beam_ssa:fold_instrs_rpo(F, [0], no, Blocks). opt_ref_used(RecvLbl, Ref, Blocks) -> - Vs = #{{var,Ref}=>ref,ref=>Ref,ref_matched=>false}, + Vs = #{Ref=>ref,ref=>Ref,ref_matched=>false}, case opt_ref_used_1(RecvLbl, Vs, Blocks) of used -> true; not_used -> false; @@ -174,7 +174,7 @@ opt_ref_used(RecvLbl, Ref, Blocks) -> end. opt_ref_used_1(L, Vs0, Blocks) -> - #b_blk{is=Is} = Blk = maps:get(L, Blocks), + #b_blk{is=Is} = Blk = map_get(L, Blocks), case opt_ref_used_is(Is, Vs0) of #{}=Vs -> opt_ref_used_last(Blk, Vs, Blocks); @@ -182,10 +182,10 @@ opt_ref_used_1(L, Vs0, Blocks) -> Result end. -opt_ref_used_is([#b_set{op=peek_message,dst=#b_var{name=M}}|Is], Vs0) -> - Vs = Vs0#{{var,M}=>message}, +opt_ref_used_is([#b_set{op=peek_message,dst=Msg}|Is], Vs0) -> + Vs = Vs0#{Msg=>message}, opt_ref_used_is(Is, Vs); -opt_ref_used_is([#b_set{op={bif,Bif},args=Args,dst=#b_var{name=B}}=I|Is], +opt_ref_used_is([#b_set{op={bif,Bif},args=Args,dst=Dst}=I|Is], Vs0) -> S = case Bif of '=:=' -> true; @@ -199,7 +199,7 @@ opt_ref_used_is([#b_set{op={bif,Bif},args=Args,dst=#b_var{name=B}}=I|Is], Bool when is_boolean(Bool) -> case is_ref_msg_comparison(Args, Vs0) of true -> - Vs = Vs0#{B=>{is_ref,Bool}}, + Vs = Vs0#{Dst=>{is_ref,Bool}}, opt_ref_used_is(Is, Vs); false -> opt_ref_used_is(Is, Vs0) @@ -225,7 +225,7 @@ opt_ref_used_is([], Vs) -> Vs. opt_ref_used_last(#b_blk{last=Last}=Blk, Vs, Blocks) -> case Last of - #b_br{bool=#b_var{name=Bool},succ=Succ,fail=Fail} -> + #b_br{bool=#b_var{}=Bool,succ=Succ,fail=Fail} -> case Vs of #{Bool:={is_ref,Matched}} -> ref_used_in([{Succ,Vs#{ref_matched:=Matched}}, @@ -252,17 +252,16 @@ ref_used_in([{L,Vs0}|Ls], Blocks) -> end; ref_used_in([], _) -> done. -update_vars(#b_set{args=Args,dst=#b_var{name=B}}, Vs) -> - Vars = [V || #b_var{name=V} <- Args], - All = all(fun(V) -> - Var = {var,V}, +update_vars(#b_set{args=Args,dst=Dst}, Vs) -> + Vars = [V || #b_var{}=V <- Args], + All = all(fun(Var) -> case Vs of #{Var:=message} -> true; #{} -> false end end, Vars), case All of - true -> Vs#{{var,B}=>message}; + true -> Vs#{Dst=>message}; false -> Vs end. @@ -270,9 +269,7 @@ update_vars(#b_set{args=Args,dst=#b_var{name=B}}, Vs) -> %% Return 'true' if Args denotes a comparison between the %% reference and message or part of the message. -is_ref_msg_comparison([#b_var{name=A1},#b_var{name=A2}], Vs) -> - V1 = {var,A1}, - V2 = {var,A2}, +is_ref_msg_comparison([#b_var{}=V1,#b_var{}=V2], Vs) -> case Vs of #{V1:=ref,V2:=message} -> true; #{V1:=message,V2:=ref} -> true; diff --git a/lib/compiler/src/beam_ssa_share.erl b/lib/compiler/src/beam_ssa_share.erl new file mode 100644 index 0000000000..426efa2cc9 --- /dev/null +++ b/lib/compiler/src/beam_ssa_share.erl @@ -0,0 +1,370 @@ +%% +%% %CopyrightBegin% +%% +%% Copyright Ericsson AB 2018. All Rights Reserved. +%% +%% Licensed under the Apache License, Version 2.0 (the "License"); +%% you may not use this file except in compliance with the License. +%% You may obtain a copy of the License at +%% +%% http://www.apache.org/licenses/LICENSE-2.0 +%% +%% Unless required by applicable law or agreed to in writing, software +%% distributed under the License is distributed on an "AS IS" BASIS, +%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +%% See the License for the specific language governing permissions and +%% limitations under the License. +%% +%% %CopyrightEnd% +%% + +%% +%% Share code for semantically equivalent blocks referred to +%% to by `br` and `switch` instructions. +%% +%% A similar optimization is done in beam_jump, but doing it here as +%% well is beneficial as it may enable other optimizations. If there +%% are many semantically equivalent clauses, this optimization can +%% substanstially decrease compilation times. +%% +%% block/2 is called from the liveness optimization pass in +%% beam_ssa_opt, as code sharing helps the liveness pass and vice +%% versa. +%% + +-module(beam_ssa_share). +-export([module/2,block/2]). + +-include("beam_ssa.hrl"). + +-import(lists, [keyfind/3,reverse/1,sort/1]). + +-spec module(beam_ssa:b_module(), [compile:option()]) -> + {'ok',beam_ssa:b_module()}. + +module(#b_module{body=Fs0}=Module, _Opts) -> + Fs = [function(F) || F <- Fs0], + {ok,Module#b_module{body=Fs}}. + +-spec block(Blk0, Blocks0) -> Blk when + Blk0 :: beam_ssa:b_blk(), + Blocks0 :: beam_ssa:block_map(), + Blk :: beam_ssa:b_blk(). + +block(#b_blk{last=Last0}=Blk, Blocks) -> + case share_terminator(Last0, Blocks) of + none -> Blk; + Last -> Blk#b_blk{last=beam_ssa:normalize(Last)} + end. + +%%% +%%% Local functions. +%%% + +function(#b_function{anno=Anno,bs=Blocks0}=F) -> + try + PO = reverse(beam_ssa:rpo(Blocks0)), + {Blocks1,Changed} = blocks(PO, Blocks0, false), + Blocks = case Changed of + true -> + beam_ssa:trim_unreachable(Blocks1); + false -> + Blocks0 + end, + F#b_function{bs=Blocks} + catch + Class:Error:Stack -> + #{func_info:={_,Name,Arity}} = Anno, + io:fwrite("Function: ~w/~w\n", [Name,Arity]), + erlang:raise(Class, Error, Stack) + end. + +blocks([L|Ls], Blocks, Changed) -> + #b_blk{last=Last0} = Blk0 = map_get(L, Blocks), + case block(Blk0, Blocks) of + #b_blk{last=Last0} -> + blocks(Ls, Blocks, Changed); + #b_blk{}=Blk -> + blocks(Ls, Blocks#{L:=Blk}, true) + end; +blocks([], Blocks, Changed) -> + {Blocks,Changed}. + +share_terminator(#b_br{bool=#b_var{},succ=Succ0,fail=Fail0}=Br, Blocks) -> + {Succ,SuccBlk} = shortcut_nonempty_block(Succ0, Blocks), + {Fail,FailBlk} = shortcut_nonempty_block(Fail0, Blocks), + case are_equivalent(Succ, SuccBlk, Fail, FailBlk, Blocks) of + true -> + %% The blocks are semantically equivalent. + Br#b_br{succ=Succ,fail=Succ}; + false -> + if + Succ =:= Succ0, Fail =:= Fail0 -> + %% None of blocks were cut short. + none; + true -> + %% One or both labels were cut short + %% to avoid jumping to an empty block. + Br#b_br{succ=Succ,fail=Fail} + end + end; +share_terminator(#b_switch{}=Sw, Blocks) -> + share_switch(Sw, Blocks); +share_terminator(_Last, _Blocks) -> none. + +%% Test whether the two blocks are semantically equivalent. This +%% function is specially optimized to return `false` as fast as +%% possible if the blocks are not equivalent, as that is the common +%% case. + +are_equivalent(_Succ, _, ?BADARG_BLOCK, _, _Blocks) -> + %% ?BADARG_BLOCK is special. Sharing could be incorrect. + false; +are_equivalent(_Succ, #b_blk{is=Is1,last=#b_ret{arg=RetVal1}=Ret1}, + _Fail, #b_blk{is=Is2,last=#b_ret{arg=RetVal2}=Ret2}, _Blocks) -> + case {RetVal1,RetVal2} of + {#b_literal{},#b_literal{}} -> + case RetVal1 =:= RetVal2 of + true -> + %% The return values are identical literals. We + %% only need to compare the canonicalized bodies. + Can1 = canonical_is(Is1), + Can2 = canonical_is(Is2), + Can1 =:= Can2; + false -> + %% Non-equal literals. + false + end; + {#b_var{},#b_var{}} -> + %% The return values are varibles. We must canonicalize + %% the blocks (including returns) and compare them. + Can1 = canonical_is(Is1 ++ [Ret1]), + Can2 = canonical_is(Is2 ++ [Ret2]), + Can1 =:= Can2; + {_,_} -> + %% One literal and one variable. + false + end; +are_equivalent(Succ, + #b_blk{is=Is1, + last=#b_br{bool=#b_literal{val=true}, + succ=Target}}, + Fail, + #b_blk{is=Is2, + last=#b_br{bool=#b_literal{val=true}, + succ=Target}}, + Blocks) -> + %% Both blocks end with an unconditional branch to the + %% same target block. If the target block has phi nodes, + %% we must pick up the values from the phi nodes and + %% compare them. + #b_blk{is=Is} = map_get(Target, Blocks), + Phis1 = canonical_terminator_phis(Is, Succ), + Phis2 = canonical_terminator_phis(Is, Fail), + case {Phis1,Phis2} of + {[#b_set{args=[#b_literal{}]}|_],_} when Phis1 =/= Phis2 -> + %% Different values are used in the phi nodes. + false; + {_,[#b_set{args=[#b_literal{}]}|_]} when Phis1 =/= Phis2 -> + %% Different values are used in the phi nodes. + false; + {_,_} -> + %% The values in the phi nodes are variables or identical + %% literals. We must canonicalize the blocks and compare + %% them. + Can1 = canonical_is(Is1 ++ Phis1), + Can2 = canonical_is(Is2 ++ Phis2), + Can1 =:= Can2 + end; +are_equivalent(Succ0, #b_blk{is=Is1,last=#b_br{bool=#b_var{},fail=Same}}, + Fail0, #b_blk{is=Is2,last=#b_br{bool=#b_var{},fail=Same}}, + Blocks) -> + %% Two-way branches with identical failure labels. First compare the + %% canonicalized bodies of the blocks. + case canonical_is(Is1) =:= canonical_is(Is2) of + false -> + %% Different bodies. + false; + true -> + %% Bodies were equal. That is fairly uncommon, so to keep + %% the code simple we will rewrite the `br` to a `switch` + %% and let share_switch/2 do the work of following the + %% branches. + Sw = #b_switch{arg=#b_var{name=not_used},fail=Fail0, + list=[{#b_literal{},Succ0}]}, + #b_switch{fail=Fail,list=[{_,Succ}]} = share_switch(Sw, Blocks), + Fail =:= Succ + end; +are_equivalent(_, _, _, _, _) -> false. + +share_switch(#b_switch{fail=Fail0,list=List0}=Sw, Blocks) -> + Prep = share_prepare_sw([{value,Fail0}|List0], Blocks, 0, []), + Res = do_share_switch(Prep, Blocks, []), + [{_,Fail}|List] = [VL || {_,VL} <- sort(Res)], + Sw#b_switch{fail=Fail,list=List}. + +share_prepare_sw([{V,L0}|T], Blocks, N, Acc) -> + {L,_Blk} = shortcut_nonempty_block(L0, Blocks), + share_prepare_sw(T, Blocks, N+1, [{{L,#{}},{N,{V,L}}}|Acc]); +share_prepare_sw([], _, _, Acc) -> Acc. + +do_share_switch(Prep, Blocks, Acc) -> + Map = share_switch_1(Prep, Blocks, #{}), + share_switch_2(maps:values(Map), Blocks, Acc). + +share_switch_1([{Next0,Res}|T], Blocks, Map) -> + {Can,Next} = canonical_block(Next0, Blocks), + case Map of + #{Can:=Ls} -> + share_switch_1(T, Blocks, Map#{Can:=[{Next,Res}|Ls]}); + #{} -> + share_switch_1(T, Blocks, Map#{Can=>[{Next,Res}]}) + end; +share_switch_1([], _Blocks, Map) -> Map. + +share_switch_2([[{_,{N,Res}}]|T], Blocks, Acc) -> + %% This block is not equivalent to any other block. + share_switch_2(T, Blocks, [{N,Res}|Acc]); +share_switch_2([[{done,{_,{_,Common}}}|_]=Eqs|T], Blocks, Acc0) -> + %% Two or more blocks are semantically equivalent, and all blocks + %% are either terminated with a `ret` or a `br` to the same target + %% block. Replace the labels in the `switch` for all of those + %% blocks with the label for the first of the blocks. + Acc = [{N,{V,Common}} || {done,{N,{V,_}}} <- Eqs] ++ Acc0, + share_switch_2(T, Blocks, Acc); +share_switch_2([[{_,_}|_]=Prep|T], Blocks, Acc0) -> + %% Two or more blocks are semantically equivalent, but they have + %% different successful successor blocks. Now we must check + %% recursively whether the successor blocks are equivalent too. + Acc = do_share_switch(Prep, Blocks, Acc0), + share_switch_2(T, Blocks, Acc); +share_switch_2([], _, Acc) -> Acc. + +canonical_block({L,VarMap0}, Blocks) -> + #b_blk{is=Is,last=Last0} = map_get(L, Blocks), + case canonical_terminator(L, Last0, Blocks) of + none -> + %% The block has a terminator that we don't handle. + {{none,L},done}; + {Last,done} -> + %% The block ends with a `ret` or an unconditional `br` to + %% another block. + {Can,_VarMap} = canonical_is(Is ++ Last, VarMap0, []), + {Can,done}; + {Last,Next} -> + %% The block ends with a conditional branch. + {Can,VarMap} = canonical_is(Is ++ Last, VarMap0, []), + {Can,{Next,VarMap}} + end. + +%% Translate a sequence of instructions to a canonical representation. If the +%% canonical representation of two blocks compare equal, the blocks are +%% semantically equivalent. The following translations are done: +%% +%% * Variables defined in the instruction sequence are replaced with +%% {var,0}, {var,1}, and so on. Free variables are not changed. +%% +%% * `location` annotations that would produce a `line` instruction are +%% kept. All other annotations are cleared. +%% +%% * Instructions are repackaged into tuples instead of into the +%% usual records. The main reason is to avoid violating the types for +%% the SSA records. We can simplify things a little by linking the +%% instructions directly instead of putting them into a list. + +canonical_is(Is) -> + {Can,_} = canonical_is(Is, #{}, []), + Can. + +canonical_is([#b_set{op=Op,dst=Dst,args=Args0}=I|Is], VarMap0, Acc) -> + Args = [canonical_arg(Arg, VarMap0) || Arg <-Args0], + Var = {var,map_size(VarMap0)}, + VarMap = VarMap0#{Dst=>Var}, + LineAnno = case Op of + bs_match -> + %% The location annotation for a bs_match instruction + %% is only used in warnings, never to emit a `line` + %% instruction. Therefore, it should not be included. + []; + _ -> + %% The location annotation will be used in a `line` + %% instruction. It must be included. + beam_ssa:get_anno(location, I, none) + end, + canonical_is(Is, VarMap, {Op,LineAnno,Var,Args,Acc}); +canonical_is([#b_ret{arg=Arg}], VarMap, Acc0) -> + Acc1 = case Acc0 of + {call,_Anno,Var,[#b_local{}|_]=Args,PrevAcc} -> + %% This is a tail-recursive call to a local function. + %% There will be no line instruction generated; + %% thus, the annotation is not significant. + {call,[],Var,Args,PrevAcc}; + _ -> + Acc0 + end, + {{ret,canonical_arg(Arg, VarMap),Acc1},VarMap}; +canonical_is([#b_br{bool=#b_var{},fail=Fail}], VarMap, Acc) -> + {{br,succ,Fail,Acc},VarMap}; +canonical_is([#b_br{succ=Succ}], VarMap, Acc) -> + {{br,Succ,Acc},VarMap}; +canonical_is([], VarMap, Acc) -> + {Acc,VarMap}. + +canonical_terminator(_L, #b_ret{}=Ret, _Blocks) -> + {[Ret],done}; +canonical_terminator(L, #b_br{bool=#b_literal{val=true},succ=Succ}=Br, Blocks) -> + #b_blk{is=Is} = map_get(Succ, Blocks), + case canonical_terminator_phis(Is, L) of + [] -> + {[],Succ}; + [_|_]=Phis -> + {Phis ++ [Br],done} + end; +canonical_terminator(_L, #b_br{bool=#b_var{},succ=Succ}=Br, _Blocks) -> + {[Br],Succ}; +canonical_terminator(_, _, _) -> none. + +canonical_terminator_phis([#b_set{op=phi,args=PhiArgs}=Phi|Is], L) -> + {Value,L} = keyfind(L, 2, PhiArgs), + [Phi#b_set{op=copy,args=[Value]}|canonical_terminator_phis(Is, L)]; +canonical_terminator_phis([#b_set{op=peek_message}=I|_], L) -> + %% We could get stuck into an infinite loop if we allowed the + %% comparisons to continue into this block. Force an unequal + %% compare with all other predecessors of this block. + [I#b_set{op=copy,args=[#b_literal{val=L}]}]; +canonical_terminator_phis(_, _) -> []. + +canonical_arg(#b_var{}=Var, VarMap) -> + case VarMap of + #{Var:=CanonicalVar} -> + CanonicalVar; + #{} -> + Var + end; +canonical_arg(#b_remote{mod=Mod,name=Name}, VarMap) -> + {remote,canonical_arg(Mod, VarMap), + canonical_arg(Name, VarMap)}; +canonical_arg(Other, _VarMap) -> Other. + +%% Shortcut branches to empty blocks if safe. + +shortcut_nonempty_block(L, Blocks) -> + case map_get(L, Blocks) of + #b_blk{is=[],last=#b_br{bool=#b_literal{val=true},succ=Succ}}=Blk -> + %% This block is empty. + case is_forbidden(Succ, Blocks) of + false -> + shortcut_nonempty_block(Succ, Blocks); + true -> + {L,Blk} + end; + #b_blk{}=Blk -> + {L,Blk} + end. + +is_forbidden(L, Blocks) -> + case map_get(L, Blocks) of + #b_blk{is=[#b_set{op=phi}|_]} -> true; + #b_blk{is=[#b_set{op=peek_message}|_]} -> true; + #b_blk{} -> false + end. diff --git a/lib/compiler/src/beam_ssa_type.erl b/lib/compiler/src/beam_ssa_type.erl index e5f15da836..32583f5abf 100644 --- a/lib/compiler/src/beam_ssa_type.erl +++ b/lib/compiler/src/beam_ssa_type.erl @@ -19,16 +19,22 @@ %% -module(beam_ssa_type). --export([opt/2]). +-export([opt_start/4, opt_continue/4, opt_finish/3]). --include("beam_ssa.hrl"). --import(lists, [any/2,droplast/1,foldl/3,last/1,member/2, - reverse/1,search/2,sort/1]). +-include("beam_ssa_opt.hrl"). +-import(lists, [all/2,any/2,droplast/1,foldl/3,last/1,member/2, + partition/2,reverse/1,sort/1]). -define(UNICODE_INT, #t_integer{elements={0,16#10FFFF}}). --record(d, {ds :: #{beam_ssa:var_name():=beam_ssa:b_set()}, - ls :: #{beam_ssa:label():=type_db()}}). +-record(d, + {ds :: #{beam_ssa:b_var():=beam_ssa:b_set()}, + ls :: #{beam_ssa:label():=type_db()}, + once :: cerl_sets:set(beam_ssa:b_var()), + func_id :: func_id(), + func_db :: func_info_db(), + sub = #{} :: #{beam_ssa:b_var():=beam_ssa:value()}, + ret_type = [] :: [type()]}). -define(ATOM_SET_SIZE, 5). @@ -46,34 +52,155 @@ {'binary',pos_integer()} | 'cons' | 'float' | 'list' | 'map' | 'nil' |'number'. -type type_db() :: #{beam_ssa:var_name():=type()}. --spec opt([{Label0,Block0}], Args) -> [{Label,Block}] when - Label0 :: beam_ssa:label(), - Block0 :: beam_ssa:b_blk(), +-spec opt_start(Linear, Args, Anno, FuncDb) -> {Linear, FuncDb} when + Linear :: [{non_neg_integer(), beam_ssa:b_blk()}], Args :: [beam_ssa:b_var()], - Label :: beam_ssa:label(), - Block :: beam_ssa:b_blk(). + Anno :: beam_ssa:anno(), + FuncDb :: func_info_db(). +opt_start(Linear, Args, Anno, FuncDb) -> + %% This is the first run through the module, so our arg_types can be + %% incomplete as we may not have visited all call sites at least once. + Ts = maps:from_list([{V,any} || #b_var{}=V <- Args]), + opt_continue_1(Linear, Args, get_func_id(Anno), Ts, FuncDb). + +-spec opt_continue(Linear, Args, Anno, FuncDb) -> {Linear, FuncDb} when + Linear :: [{non_neg_integer(), beam_ssa:b_blk()}], + Args :: [beam_ssa:b_var()], + Anno :: beam_ssa:anno(), + FuncDb :: func_info_db(). +opt_continue(Linear, Args, Anno, FuncDb) -> + Id = get_func_id(Anno), + case FuncDb of + #{ Id := #func_info{exported=false,arg_types=ArgTypes} } -> + %% This is a local function and we're guaranteed to have visited + %% every call site at least once, so we know that the parameter + %% types are at least as narrow as the join of all argument types. + Ts = join_arg_types(Args, ArgTypes, Anno), + opt_continue_1(Linear, Args, Id, Ts, FuncDb); + #{} -> + %% We can't infer the parameter types of exported functions, nor + %% the ones where module-level optimization is disabled, but + %% running the pass again could still help other functions. + Ts = maps:from_list([{V,any} || #b_var{}=V <- Args]), + opt_continue_1(Linear, Args, Id, Ts, FuncDb) + end. -opt(Linear, Args) -> - Ts = maps:from_list([{V,any} || #b_var{name=V} <- Args]), +join_arg_types(Args, ArgTypes, Anno) -> + %% We suppress type optimization for parameters that have already been + %% optimized by another pass, as they may have done things we have no idea + %% how to interpret and running them over could generate incorrect code. + ParamTypes = maps:get(parameter_type_info, Anno, #{}), + Ts0 = join_arg_types_1(Args, ArgTypes, #{}), + maps:fold(fun(Arg, _V, Ts) -> + maps:put(Arg, any, Ts) + end, Ts0, ParamTypes). + +join_arg_types_1([Arg | Args], [TM | TMs], Ts) when map_size(TM) =/= 0 -> + join_arg_types_1(Args, TMs, Ts#{ Arg => join(maps:values(TM))}); +join_arg_types_1([Arg | Args], [_TM | TMs], Ts) -> + join_arg_types_1(Args, TMs, Ts#{ Arg => any }); +join_arg_types_1([], [], Ts) -> + Ts. + +-spec opt_continue_1(Linear, Args, Id, Ts, FuncDb) -> Result when + Linear :: [{non_neg_integer(), beam_ssa:b_blk()}], + Args :: [beam_ssa:b_var()], + Id :: func_id(), + Ts :: type_db(), + FuncDb :: func_info_db(), + Result :: {Linear, FuncDb}. +opt_continue_1(Linear0, Args, Id, Ts, FuncDb0) -> + UsedOnce = used_once(Linear0, Args), FakeCall = #b_set{op=call,args=[#b_remote{mod=#b_literal{val=unknown}, name=#b_literal{val=unknown}, arity=0}]}, - Defs = maps:from_list([{V,FakeCall#b_set{dst=Var}} || - #b_var{name=V}=Var <- Args]), - D = #d{ds=Defs,ls=#{0=>Ts}}, - opt_1(Linear, D). + Defs = maps:from_list([{Var,FakeCall#b_set{dst=Var}} || + #b_var{}=Var <- Args]), + + D = #d{ func_db=FuncDb0, + func_id=Id, + ds=Defs, + ls=#{0=>Ts,?BADARG_BLOCK=>#{}}, + once=UsedOnce }, + + {Linear, FuncDb, NewRet} = opt_1(Linear0, D, []), + + case FuncDb of + #{ Id := Entry0 } -> + Entry = Entry0#func_info{ret_type=NewRet}, + {Linear, FuncDb#{ Id := Entry }}; + #{} -> + %% Module-level optimizations have been turned off for this + %% function. + {Linear, FuncDb} + end. -opt_1([{L,Blk}|Bs], #d{ls=Ls}=D) -> +-spec opt_finish(Args, Anno, FuncDb) -> {Anno, FuncDb} when + Args :: [beam_ssa:b_var()], + Anno :: beam_ssa:anno(), + FuncDb :: func_info_db(). +opt_finish(Args, Anno, FuncDb) -> + Id = get_func_id(Anno), + case FuncDb of + #{ Id := #func_info{exported=false,arg_types=ArgTypes} } -> + ParamInfo0 = maps:get(parameter_type_info, Anno, #{}), + ParamInfo = opt_finish_1(Args, ArgTypes, ParamInfo0), + {Anno#{ parameter_type_info => ParamInfo }, FuncDb}; + #{} -> + {Anno, FuncDb} + end. + +opt_finish_1([Arg | Args], [TypeMap | TypeMaps], ParamInfo) + when is_map_key(Arg, ParamInfo); %% See join_arg_types/3 + map_size(TypeMap) =:= 0 -> + opt_finish_1(Args, TypeMaps, ParamInfo); +opt_finish_1([Arg | Args], [TypeMap | TypeMaps], ParamInfo0) -> + case join(maps:values(TypeMap)) of + any -> + opt_finish_1(Args, TypeMaps, ParamInfo0); + JoinedType -> + JoinedType = verified_type(JoinedType), + ParamInfo = ParamInfo0#{ Arg => validator_anno(JoinedType) }, + opt_finish_1(Args, TypeMaps, ParamInfo) + end; +opt_finish_1([], [], ParamInfo) -> + ParamInfo. + +validator_anno(#t_tuple{size=Size,exact=Exact}) -> + beam_validator:type_anno(tuple, Size, Exact); +validator_anno(#t_integer{elements={Same,Same}}) -> + beam_validator:type_anno(integer, Same); +validator_anno(#t_integer{}) -> + beam_validator:type_anno(integer); +validator_anno(float) -> + beam_validator:type_anno(float); +validator_anno(#t_atom{elements=[Val]}) -> + beam_validator:type_anno(atom, Val); +validator_anno(#t_atom{}=A) -> + case t_is_boolean(A) of + true -> beam_validator:type_anno(bool); + false -> beam_validator:type_anno(atom) + end; +validator_anno(T) -> + beam_validator:type_anno(T). + +get_func_id(Anno) -> + #{func_info:={_Mod, Name, Arity}} = Anno, + #b_local{name=#b_literal{val=Name}, arity=Arity}. + +opt_1([{L,Blk}|Bs], #d{ls=Ls}=D, Acc) -> case Ls of #{L:=Ts} -> - opt_2(L, Blk, Bs, Ts, D); + opt_2(L, Blk, Bs, Ts, D, Acc); #{} -> %% This block is never reached. Discard it. - opt_1(Bs, D) + opt_1(Bs, D, Acc) end; -opt_1([], _) -> []. +opt_1([], D, Acc) -> + #d{func_db=FuncDb,ret_type=NewRet} = D, + {reverse(Acc), FuncDb, NewRet}. -opt_2(L, #b_blk{is=Is0}=Blk0, Bs, Ts, D0) -> +opt_2(L, #b_blk{is=Is0}=Blk0, Bs, Ts, #d{sub=Sub}=D0, Acc) -> case Is0 of [#b_set{op=call,dst=Dst, args=[#b_remote{mod=#b_literal{val=Mod}, @@ -84,75 +211,249 @@ opt_2(L, #b_blk{is=Is0}=Blk0, Bs, Ts, D0) -> %% Rewrite the terminator to a 'ret', and remove %% all type information for this label. That will %% simplify the phi node in the former successor. - Args = [simplify_arg(Arg, Ts) || Arg <- Args0], + Args = simplify_args(Args0, Sub, Ts), I = I0#b_set{args=[Rem|Args]}, Ret = #b_ret{arg=Dst}, Blk = Blk0#b_blk{is=[I],last=Ret}, Ls = maps:remove(L, D0#d.ls), - D = D0#d{ls=Ls}, - [{L,Blk}|opt_1(Bs, D)]; + + %% We potentially lack a return value. + RetType = join([none | D0#d.ret_type]), + + D = D0#d{ls=Ls,ret_type=[RetType]}, + opt_1(Bs, D, [{L,Blk} | Acc]); false -> - opt_3(L, Blk0, Bs, Ts, D0) + opt_3(L, Blk0, Bs, Ts, D0, Acc) end; _ -> - opt_3(L, Blk0, Bs, Ts, D0) + opt_3(L, Blk0, Bs, Ts, D0, Acc) end. -opt_3(L, #b_blk{is=Is0,last=Last0}=Blk0, Bs, Ts0, #d{ds=Ds0,ls=Ls0}=D0) -> - {Is,Ts,Ds} = opt_is(Is0, Ts0, Ds0, Ls0, []), - D1 = D0#d{ds=Ds}, - Last = opt_terminator(Last0, Ts, Ds), +opt_3(L, #b_blk{is=Is0,last=Last0}=Blk0, Bs, Ts0, + #d{ds=Ds0,ls=Ls0,sub=Sub0,func_db=Fdb0}=D0, Acc) -> + {Is,Ts,Ds,Fdb,Sub} = opt_is(Is0, Ts0, Ds0, Fdb0, Ls0, D0, Sub0, []), + D1 = D0#d{ds=Ds,sub=Sub,func_db=Fdb}, + Last1 = simplify_terminator(Last0, Sub, Ts, Ds), + Last = opt_terminator(Last1, Ts, Ds), D = update_successors(Last, Ts, D1), Blk = Blk0#b_blk{is=Is,last=Last}, - [{L,Blk}|opt_1(Bs, D)]. + opt_1(Bs, D, [{L,Blk} | Acc]). + +simplify_terminator(#b_br{bool=Bool}=Br, Sub, Ts, _Ds) -> + Br#b_br{bool=simplify_arg(Bool, Sub, Ts)}; +simplify_terminator(#b_switch{arg=Arg}=Sw, Sub, Ts, _Ds) -> + Sw#b_switch{arg=simplify_arg(Arg, Sub, Ts)}; +simplify_terminator(#b_ret{arg=Arg}=Ret, Sub, Ts, Ds) -> + %% Reducing the result of a call to a literal (fairly common for 'ok') + %% breaks tail call optimization. + case Ds of + #{ Arg := #b_set{op=call}} -> Ret; + #{} -> Ret#b_ret{arg=simplify_arg(Arg, Sub, Ts)} + end. -opt_is([#b_set{op=phi,dst=#b_var{name=Dst},args=Args0}=I0|Is], Ts0, Ds0, Ls, Acc) -> +opt_is([#b_set{op=phi,dst=Dst,args=Args0}=I0|Is], + Ts0, Ds0, Fdb, Ls, D, Sub0, Acc) -> %% Simplify the phi node by removing all predecessor blocks that no %% longer exists or no longer branches to this block. - Args = [P || {_,From}=P <- Args0, maps:is_key(From, Ls)], - I = I0#b_set{args=Args}, - Ts = update_types(I, Ts0, Ds0), - Ds = Ds0#{Dst=>I}, - opt_is(Is, Ts, Ds, Ls, [I|Acc]); -opt_is([#b_set{dst=#b_var{name=Dst}}=I0|Is], Ts0, Ds0, Ls, Acc) -> - I = simplify(I0, Ts0), + Args = [{simplify_arg(Arg, Sub0, Ts0),From} || + {Arg,From} <- Args0, maps:is_key(From, Ls)], + case all_same(Args) of + true -> + %% Eliminate the phi node if there is just one source + %% value or if the values are identical. + [{Val,_}|_] = Args, + Sub = Sub0#{Dst=>Val}, + opt_is(Is, Ts0, Ds0, Fdb, Ls, D, Sub, Acc); + false -> + I = I0#b_set{args=Args}, + Ts = update_types(I, Ts0, Ds0), + Ds = Ds0#{Dst=>I}, + opt_is(Is, Ts, Ds, Fdb, Ls, D, Sub0, [I|Acc]) + end; +opt_is([#b_set{op=call,args=Args0,dst=Dst}=I0 | Is], + Ts0, Ds0, Fdb0, Ls, D, Sub, Acc) -> + Args = simplify_args(Args0, Sub, Ts0), + I1 = beam_ssa:normalize(I0#b_set{args=Args}), + + %% This is a bit of a kludge; we know that any instruction whose return + %% type is 'none' will fail at runtime, but we don't yet have a way to cut + %% a block short so we move on like nothing nothing happened. + %% + %% This complicates argument type optimization as unreachable calls can + %% add types that will never occur, so we skip optimizing this call if + %% the type of any of its arguments is 'none'. + [_Callee | Rest] = Args, + case all(fun(Arg) -> get_type(Arg, Ts0) =/= none end, Rest) of + true -> + {Ts, Ds, Fdb, I} = opt_call(I1, D, Ts0, Ds0, Fdb0), + opt_is(Is, Ts, Ds, Fdb, Ls, D, Sub, [I|Acc]); + false -> + Ts = Ts0#{ Dst => any }, + Ds = Ds0#{ Dst => I1 }, + opt_is(Is, Ts, Ds, Fdb0, Ls, D, Sub, [I1|Acc]) + end; +opt_is([#b_set{op=succeeded,args=[Arg],dst=Dst}=I], + Ts0, Ds0, Fdb, Ls, D, Sub0, Acc) -> + case Ds0 of + #{ Arg := #b_set{op=call} } -> + %% The success check of a call is part of exception handling and + %% must not be optimized away. We still have to update its type + %% though. + Ts = update_types(I, Ts0, Ds0), + Ds = Ds0#{Dst=>I}, + + opt_is([], Ts, Ds, Fdb, Ls, D, Sub0, [I|Acc]); + #{} -> + Args = simplify_args([Arg], Sub0, Ts0), + Type = type(succeeded, Args, Ts0, Ds0), + case get_literal_from_type(Type) of + #b_literal{}=Lit -> + Sub = Sub0#{Dst=>Lit}, + opt_is([], Ts0, Ds0, Fdb, Ls, D, Sub, Acc); + none -> + Ts = Ts0#{Dst=>Type}, + Ds = Ds0#{Dst=>I}, + opt_is([], Ts, Ds, Fdb, Ls, D, Sub0, [I|Acc]) + end + end; +opt_is([#b_set{args=Args0,dst=Dst}=I0|Is], + Ts0, Ds0, Fdb, Ls, D, Sub0, Acc) -> + Args = simplify_args(Args0, Sub0, Ts0), + I1 = beam_ssa:normalize(I0#b_set{args=Args}), + case simplify(I1, Ts0) of + #b_set{}=I2 -> + I = beam_ssa:normalize(I2), + Ts = update_types(I, Ts0, Ds0), + Ds = Ds0#{Dst=>I}, + opt_is(Is, Ts, Ds, Fdb, Ls, D, Sub0, [I|Acc]); + #b_literal{}=Lit -> + Sub = Sub0#{Dst=>Lit}, + opt_is(Is, Ts0, Ds0, Fdb, Ls, D, Sub, Acc); + #b_var{}=Var -> + case Is of + [#b_set{op=succeeded,dst=SuccDst,args=[Dst]}] -> + %% We must remove this 'succeeded' instruction. + Sub = Sub0#{Dst=>Var,SuccDst=>#b_literal{val=true}}, + opt_is([], Ts0, Ds0, Fdb, Ls, D, Sub, Acc); + _ -> + Sub = Sub0#{Dst=>Var}, + opt_is(Is, Ts0, Ds0, Fdb, Ls, D, Sub, Acc) + end + end; +opt_is([], Ts, Ds, Fdb, _Ls, _D, Sub, Acc) -> + {reverse(Acc), Ts, Ds, Fdb, Sub}. + +opt_call(#b_set{dst=Dst,args=[#b_local{}=Callee|Args]}=I0, D, Ts0, Ds0, Fdb0) -> + {Ts, Ds, I} = opt_local_call(I0, Ts0, Ds0, Fdb0), + case Fdb0 of + #{ Callee := #func_info{exported=false,arg_types=ArgTypes0}=Info } -> + %% Update the argument types of *this exact call*, the types + %% will be joined later when the callee is optimized. + CallId = {D#d.func_id, Dst}, + ArgTypes = update_arg_types(Args, ArgTypes0, CallId, Ts0), + + Fdb = Fdb0#{ Callee => Info#func_info{arg_types=ArgTypes} }, + {Ts, Ds, Fdb, I}; + #{} -> + %% We can't narrow the argument types of exported functions as they + %% can receive anything as part of an external call. + {Ts, Ds, Fdb0, I} + end; +opt_call(#b_set{dst=Dst}=I, _D, Ts0, Ds0, Fdb) -> Ts = update_types(I, Ts0, Ds0), - Ds = Ds0#{Dst=>I}, - opt_is(Is, Ts, Ds, Ls, [I|Acc]); -opt_is([], Ts, Ds, _Ls, Acc) -> - {reverse(Acc),Ts,Ds}. + Ds = Ds0#{ Dst => I }, + {Ts, Ds, Fdb, I}. + +opt_local_call(#b_set{dst=Dst,args=[Id|_]}=I0, Ts0, Ds0, Fdb) -> + %% We skip propagating 'none' as we don't yet have a good way to cut a + %% block short. + Type = case Fdb of + #{ Id := #func_info{ret_type=[T]} } when T =/= none -> T; + #{} -> any + end, + I = case Type of + any -> I0; + _ -> beam_ssa:add_anno(result_type, validator_anno(Type), I0) + end, + Ts = Ts0#{ Dst => Type }, + Ds = Ds0#{ Dst => I }, + {Ts, Ds, I}. + +update_arg_types([Arg | Args], [TypeMap0 | TypeMaps], CallId, Ts) -> + %% Match contexts are treated as bitstrings when optimizing arguments, as + %% we don't yet support removing the "bs_start_match3" instruction. + NewType = case get_type(Arg, Ts) of + #t_bs_match{} -> {binary, 1}; + Type -> Type + end, + PrevType = maps:get(CallId, TypeMap0, NewType), + + %% The new type must be narrower than the old one. + true = meet(NewType, PrevType) =/= none, %Assertion. + + TypeMap = TypeMap0#{ CallId => NewType }, + [TypeMap | update_arg_types(Args, TypeMaps, CallId, Ts)]; +update_arg_types([], [], _CallId, _Ts) -> + []. +simplify(#b_set{op={bif,'and'},args=Args}=I, Ts) -> + case is_safe_bool_op(Args, Ts) of + true -> + case Args of + [_,#b_literal{val=false}=Res] -> Res; + [Res,#b_literal{val=true}] -> Res; + _ -> eval_bif(I, Ts) + end; + false -> + I + end; +simplify(#b_set{op={bif,'or'},args=Args}=I, Ts) -> + case is_safe_bool_op(Args, Ts) of + true -> + case Args of + [Res,#b_literal{val=false}] -> Res; + [_,#b_literal{val=true}=Res] -> Res; + _ -> eval_bif(I, Ts) + end; + false -> + I + end; simplify(#b_set{op={bif,element},args=[#b_literal{val=Index},Tuple]}=I, Ts) -> case t_tuple_size(get_type(Tuple, Ts)) of {_,Size} when is_integer(Index), 1 =< Index, Index =< Size -> I#b_set{op=get_tuple_element,args=[Tuple,#b_literal{val=Index-1}]}; _ -> - I + eval_bif(I, Ts) end; simplify(#b_set{op={bif,hd},args=[List]}=I, Ts) -> case get_type(List, Ts) of cons -> I#b_set{op=get_hd}; _ -> - I + eval_bif(I, Ts) end; simplify(#b_set{op={bif,tl},args=[List]}=I, Ts) -> case get_type(List, Ts) of cons -> I#b_set{op=get_tl}; _ -> - I + eval_bif(I, Ts) end; simplify(#b_set{op={bif,size},args=[Term]}=I, Ts) -> case get_type(Term, Ts) of #t_tuple{} -> - I#b_set{op={bif,tuple_size}}; + simplify(I#b_set{op={bif,tuple_size}}, Ts); + _ -> + eval_bif(I, Ts) + end; +simplify(#b_set{op={bif,tuple_size},args=[Term]}=I, Ts) -> + case get_type(Term, Ts) of + #t_tuple{size=Size,exact=true} -> + #b_literal{val=Size}; _ -> I end; -simplify(#b_set{op={bif,'=='},args=Args0}=I0, Ts) -> - Args = [simplify_arg(Arg, Ts) || Arg <- Args0], - I = I0#b_set{args=Args}, +simplify(#b_set{op={bif,'=='},args=Args}=I, Ts) -> Types = get_types(Args, Ts), EqEq = case {meet(Types),join(Types)} of {none,any} -> true; @@ -164,50 +465,147 @@ simplify(#b_set{op={bif,'=='},args=Args0}=I0, Ts) -> end, case EqEq of true -> - I#b_set{op={bif,'=:='}}; + simplify(I#b_set{op={bif,'=:='}}, Ts); false -> - I + eval_bif(I, Ts) end; -simplify(#b_set{op={bif,Op},args=Args0}=I0, Ts) -> - Args = [simplify_arg(Arg, Ts) || Arg <- Args0], - I = I0#b_set{args=Args}, +simplify(#b_set{op={bif,'=:='},args=[Same,Same]}, _Ts) -> + #b_literal{val=true}; +simplify(#b_set{op={bif,'=:='},args=Args}=I, Ts) -> + case meet(get_types(Args, Ts)) of + none -> #b_literal{val=false}; + _ -> eval_bif(I, Ts) + end; +simplify(#b_set{op={bif,Op},args=Args}=I, Ts) -> Types = get_types(Args, Ts), case is_float_op(Op, Types) of false -> - I; + eval_bif(I, Ts); true -> AnnoArgs = [anno_float_arg(A) || A <- Types], - beam_ssa:add_anno(float_op, AnnoArgs, I) + eval_bif(beam_ssa:add_anno(float_op, AnnoArgs, I), Ts) + end; +simplify(#b_set{op=get_tuple_element,args=[Tuple,#b_literal{val=0}]}=I, Ts) -> + case get_type(Tuple, Ts) of + #t_tuple{elements=[First]} -> + #b_literal{val=First}; + #t_tuple{} -> + I end; -simplify(#b_set{op=wait_timeout,args=[Timeout0]}=I, Ts) -> - case simplify_arg(Timeout0, Ts) of - #b_literal{val=infinity} -> - I#b_set{op=wait,args=[]}; - Timeout -> - I#b_set{args=[Timeout]} +simplify(#b_set{op=is_nonempty_list,args=[Src]}=I, Ts) -> + case get_type(Src, Ts) of + any -> I; + list -> I; + cons -> #b_literal{val=true}; + _ -> #b_literal{val=false} end; -simplify(#b_set{op=Op,args=Args0}=I, Ts) -> - Safe = case Op of - call -> true; - put_list -> true; - put_tuple -> true; - _ -> false - end, - case Safe of - true -> - Args = [simplify_arg(Arg, Ts) || Arg <- Args0], - I#b_set{args=Args}; +simplify(#b_set{op=is_tagged_tuple, + args=[Src,#b_literal{val=Size},#b_literal{val=Tag}]}=I, Ts) -> + case get_type(Src, Ts) of + #t_tuple{exact=true,size=Size,elements=[Tag]} -> + #b_literal{val=true}; + #t_tuple{exact=true,size=ActualSize,elements=[]} -> + if + Size =/= ActualSize -> + #b_literal{val=false}; + true -> + I + end; + #t_tuple{exact=false} -> + I; + any -> + I; + _ -> + #b_literal{val=false} + end; +simplify(#b_set{op=put_list,args=[#b_literal{val=H}, + #b_literal{val=T}]}, _Ts) -> + #b_literal{val=[H|T]}; +simplify(#b_set{op=put_tuple,args=Args}=I, _Ts) -> + case make_literal_list(Args) of + none -> I; + List -> #b_literal{val=list_to_tuple(List)} + end; +simplify(#b_set{op=wait_timeout,args=[#b_literal{val=0}]}, _Ts) -> + #b_literal{val=true}; +simplify(#b_set{op=wait_timeout,args=[#b_literal{val=infinity}]}=I, _Ts) -> + I#b_set{op=wait,args=[]}; +simplify(I, _Ts) -> I. + +make_literal_list(Args) -> + make_literal_list(Args, []). + +make_literal_list([#b_literal{val=H}|T], Acc) -> + make_literal_list(T, [H|Acc]); +make_literal_list([_|_], _) -> + none; +make_literal_list([], Acc) -> + reverse(Acc). + +is_safe_bool_op(Args, Ts) -> + [T1,T2] = get_types(Args, Ts), + t_is_boolean(T1) andalso t_is_boolean(T2). + +all_same([{H,_}|T]) -> + all(fun({E,_}) -> E =:= H end, T). + +eval_bif(#b_set{op={bif,Bif},args=Args}=I, Ts) -> + Arity = length(Args), + case erl_bifs:is_pure(erlang, Bif, Arity) of false -> - I + I; + true -> + case make_literal_list(Args) of + none -> + case get_types(Args, Ts) of + [any] -> + I; + [Type] -> + case will_succeed(Bif, Type) of + yes -> + #b_literal{val=true}; + no -> + #b_literal{val=false}; + maybe -> + I + end; + _ -> + I + end; + LitArgs -> + try apply(erlang, Bif, LitArgs) of + Val -> #b_literal{val=Val} + catch + error:_ -> I + end + + end end. -simplify_arg(#b_var{}=Arg, Ts) -> - Type = get_type(Arg, Ts), - case get_literal_from_type(Type) of - none -> Arg; - #b_literal{}=Lit -> Lit +simplify_args(Args, Sub, Ts) -> + [simplify_arg(Arg, Sub, Ts) || Arg <- Args]. + +simplify_arg(#b_var{}=Arg0, Sub, Ts) -> + case sub_arg(Arg0, Sub) of + #b_literal{}=LitArg -> + LitArg; + #b_var{}=Arg -> + Type = get_type(Arg, Ts), + case get_literal_from_type(Type) of + none -> Arg; + #b_literal{}=Lit -> Lit + end end; -simplify_arg(Arg, _Ts) -> Arg. +simplify_arg(#b_remote{mod=Mod,name=Name}=Rem, Sub, Ts) -> + Rem#b_remote{mod=simplify_arg(Mod, Sub, Ts), + name=simplify_arg(Name, Sub, Ts)}; +simplify_arg(Arg, _Sub, _Ts) -> Arg. + +sub_arg(#b_var{}=Old, Sub) -> + case Sub of + #{Old:=New} -> New; + #{} -> Old + end. is_float_op('-', [float]) -> true; @@ -228,75 +626,121 @@ anno_float_arg(float) -> float; anno_float_arg(_) -> convert. opt_terminator(#b_br{bool=#b_literal{}}=Br, _Ts, _Ds) -> - Br; -opt_terminator(#b_br{bool=#b_var{name=V}=Var}=Br, Ts, Ds) -> - BoolType = get_type(Var, Ts), - case get_literal_from_type(BoolType) of - #b_literal{}=BoolLit -> - Br#b_br{bool=BoolLit}; - none -> - #{V:=Set} = Ds, - case Set of - #b_set{op={bif,'=:='},args=[Bool,#b_literal{val=true}]} -> - case t_is_boolean(get_type(Bool, Ts)) of - true -> - %% Bool =:= true ==> Bool - simplify_not(Br#b_br{bool=Bool}, Ts, Ds); - false -> - Br - end; - #b_set{} -> - simplify_not(Br, Ts, Ds) - end + beam_ssa:normalize(Br); +opt_terminator(#b_br{bool=#b_var{}=V}=Br, Ts, Ds) -> + #{V:=Set} = Ds, + case Set of + #b_set{op={bif,'=:='},args=[Bool,#b_literal{val=true}]} -> + case t_is_boolean(get_type(Bool, Ts)) of + true -> + %% Bool =:= true ==> Bool + simplify_not(Br#b_br{bool=Bool}, Ts, Ds); + false -> + Br + end; + #b_set{} -> + simplify_not(Br, Ts, Ds) end; -opt_terminator(#b_switch{arg=#b_literal{val=Val0}=Arg,fail=Fail,list=List}, - _Ts, _Ds) -> - {value,{_,L}} = search(fun({#b_literal{val=Val1},_}) -> - Val1 =:= Val0 - end, List ++ [{Arg,Fail}]), - #b_br{bool=#b_literal{val=true},succ=L,fail=L}; -opt_terminator(#b_switch{arg=V}=Sw0, Ts, Ds) -> - case get_literal_from_type(Ts) of - #b_literal{}=Lit -> - Sw = Sw0#b_switch{arg=Lit}, - opt_terminator(Sw, Ts, Ds); - none -> - case get_type(V, Ts) of - #t_integer{elements={_,_}=Range} -> - simplify_switch_int(Sw0, Range); - Type -> - case t_is_boolean(Type) of - true -> - case simplify_switch_bool(Sw0, Ts, Ds) of - #b_br{}=Br -> - opt_terminator(Br, Ts, Ds); - Sw -> - Sw - end; - false -> - Sw0 - end +opt_terminator(#b_switch{arg=#b_literal{}}=Sw, _Ts, _Ds) -> + beam_ssa:normalize(Sw); +opt_terminator(#b_switch{arg=#b_var{}=V}=Sw0, Ts, Ds) -> + Type = get_type(V, Ts), + case Type of + #t_integer{elements={_,_}=Range} -> + simplify_switch_int(Sw0, Range); + _ -> + case t_is_boolean(Type) of + true -> + case simplify_switch_bool(Sw0, Ts, Ds) of + #b_br{}=Br -> + opt_terminator(Br, Ts, Ds); + Sw -> + beam_ssa:normalize(Sw) + end; + false -> + beam_ssa:normalize(Sw0) end end; -opt_terminator(#b_ret{}=Ret, _Ts, _Ds) -> - Ret. +opt_terminator(#b_ret{}=Ret, _Ts, _Ds) -> Ret. -update_successors(#b_br{bool=#b_literal{val=false},fail=S}, Ts, D) -> - update_successor(S, Ts, D); update_successors(#b_br{bool=#b_literal{val=true},succ=S}, Ts, D) -> update_successor(S, Ts, D); -update_successors(#b_br{bool=#b_var{name=V},succ=Succ,fail=Fail}, Ts, D0) -> - D = update_successor(Fail, Ts#{V:=t_atom(false)}, D0), - SuccTs = infer_types(V, Ts, D0), - update_successor(Succ, SuccTs#{V:=t_atom(true)}, D); -update_successors(#b_switch{arg=#b_var{name=V},fail=Fail,list=List}, Ts, D0) -> - D = update_successor(Fail, Ts, D0), - foldl(fun({Val,S}, A) -> - T = get_type(Val, Ts), - update_successor(S, Ts#{V=>T}, A) - end, D, List); -update_successors(#b_ret{}, _Ts, D) -> D. +update_successors(#b_br{bool=#b_var{}=Bool,succ=Succ,fail=Fail}, Ts0, D0) -> + case cerl_sets:is_element(Bool, D0#d.once) of + true -> + %% This variable is defined in this block and is only + %% referenced by this br terminator. Therefore, there is + %% no need to include the type database passed on to the + %% successors of this block. + Ts = maps:remove(Bool, Ts0), + {SuccTs,FailTs} = infer_types(Bool, Ts, D0), + D = update_successor(Fail, FailTs, D0), + update_successor(Succ, SuccTs, D); + false -> + {SuccTs,FailTs} = infer_types(Bool, Ts0, D0), + D = update_successor_bool(Bool, false, Fail, FailTs, D0), + update_successor_bool(Bool, true, Succ, SuccTs, D) + end; +update_successors(#b_switch{arg=#b_var{}=V,fail=Fail,list=List}, Ts0, D0) -> + case cerl_sets:is_element(V, D0#d.once) of + true -> + %% This variable is defined in this block and is only + %% referenced by this switch terminator. Therefore, there is + %% no need to include the type database passed on to the + %% successors of this block. + Ts = maps:remove(V, Ts0), + D = update_successor(Fail, Ts, D0), + F = fun({_Val,S}, A) -> + update_successor(S, Ts, A) + end, + foldl(F, D, List); + false -> + %% V can not be equal to any of the values in List at the fail + %% block. + FailTs = subtract_sw_list(V, List, Ts0), + D = update_successor(Fail, FailTs, D0), + F = fun({Val,S}, A) -> + T = get_type(Val, Ts0), + update_successor(S, Ts0#{V=>T}, A) + end, + foldl(F, D, List) + end; +update_successors(#b_ret{arg=Arg}, Ts, D) -> + FuncId = D#d.func_id, + case D#d.ds of + #{ Arg := #b_set{op=call,args=[FuncId | _]} } -> + %% Returning a call to ourselves doesn't affect our own return + %% type. + D; + #{} -> + RetType = join([get_type(Arg, Ts) | D#d.ret_type]), + D#d{ret_type=[RetType]} + end. +subtract_sw_list(V, List, Ts) -> + Ts#{ V := sub_sw_list_1(get_type(V, Ts), List, Ts) }. + +sub_sw_list_1(Type, [{Val,_}|T], Ts) -> + ValType = get_type(Val, Ts), + sub_sw_list_1(subtract(Type, ValType), T, Ts); +sub_sw_list_1(Type, [], _Ts) -> + Type. + +update_successor_bool(#b_var{}=Var, BoolValue, S, Ts, D) -> + case t_is_boolean(get_type(Var, Ts)) of + true -> + update_successor(S, Ts#{Var:=t_atom(BoolValue)}, D); + false -> + %% The `br` terminator is preceeded by an instruction that + %% does not produce a boolean value, such a `new_try_tag`. + update_successor(S, Ts, D) + end. + +update_successor(?BADARG_BLOCK, _Ts, #d{}=D) -> + %% We KNOW that no variables are used in the ?BADARG_BLOCK, + %% so there is no need to update the type information. That + %% can be a huge timesaver for huge functions. + D; update_successor(S, Ts0, #d{ls=Ls}=D) -> case Ls of #{S:=Ts1} -> @@ -306,48 +750,15 @@ update_successor(S, Ts0, #d{ls=Ls}=D) -> D#d{ls=Ls#{S=>Ts0}} end. -update_types(#b_set{op=Op,dst=#b_var{name=Dst},args=Args}, Ts, Ds) -> +update_types(#b_set{op=Op,dst=Dst,args=Args}, Ts, Ds) -> T = type(Op, Args, Ts, Ds), Ts#{Dst=>T}. type(phi, Args, Ts, _Ds) -> Types = [get_type(A, Ts) || {A,_} <- Args], join(Types); -type({bif,'=:='}, [Same,Same], _Ts, _Ds) -> - t_atom(true); -type({bif,'=:='}, [_,_]=Args, Ts, _Ds) -> - case get_literals(Args, Ts) of - [#b_literal{val=Lit1},#b_literal{val=Lit2}] -> - t_atom(Lit1 =:= Lit2); - [_,_] -> - case meet(get_types(Args, Ts)) of - none -> t_atom(false); - _ -> t_boolean() - end - end; -type({bif,tuple_size}, [Src], Ts, _Ds) -> - case t_tuple_size(get_type(Src, Ts)) of - {exact,Size} -> - t_integer(Size); - _ -> - t_integer() - end; type({bif,'band'}, Args, Ts, _Ds) -> band_type(Args, Ts); -type({bif,Bif}, [Src]=Args, Ts, _Ds) -> - case get_type(Src, Ts) of - any -> - bif_type(Bif, Args); - Type -> - case will_succeed(Bif, Type) of - yes -> - t_atom(true); - no -> - t_atom(false); - maybe -> - bif_type(Bif, Args) - end - end; type({bif,Bif}, Args, Ts, _Ds) -> case bif_type(Bif, Args) of number -> @@ -369,6 +780,8 @@ type(bs_extract, [Ctx], Ts, _Ds) -> Type; type(bs_match, Args, _Ts, _Ds) -> #t_bs_match{type=bs_match_type(Args)}; +type(bs_get_tail, _Args, _Ts, _Ds) -> + {binary, 1}; type(call, [#b_remote{mod=#b_literal{val=Mod}, name=#b_literal{val=Name}}|Args], Ts, _Ds) -> case {Mod,Name,Args} of @@ -388,6 +801,14 @@ type(call, [#b_remote{mod=#b_literal{val=Mod}, {_,_} -> #t_tuple{} end; + {erlang,'++',[List1,List2]} -> + case get_type(List1, Ts) =:= cons orelse + get_type(List2, Ts) =:= cons of + true -> cons; + false -> list + end; + {erlang,'--',[_,_]} -> + list; {math,_,_} -> case is_math_bif(Name, length(Args)) of false -> any; @@ -399,42 +820,12 @@ type(call, [#b_remote{mod=#b_literal{val=Mod}, false -> any end end; -type(get_tuple_element, [Tuple,#b_literal{val=0}], Ts, _Ds) -> - case get_type(Tuple, Ts) of - #t_tuple{elements=[First]} -> - get_type(#b_literal{val=First}, Ts); - #t_tuple{} -> - any - end; -type(is_nonempty_list, [Src], Ts, _Ds) -> - case get_type(Src, Ts) of - any -> - t_boolean(); - list -> - t_boolean(); - cons -> - t_atom(true); - _ -> - t_atom(false) - end; -type(is_tagged_tuple, [Src,#b_literal{val=Size},#b_literal{val=Tag}], Ts, _Ds) -> - case get_type(Src, Ts) of - #t_tuple{exact=true,size=Size,elements=[Tag]} -> - t_atom(true); - #t_tuple{exact=true,size=ActualSize,elements=[]} -> - if - Size =/= ActualSize -> - t_atom(false); - true -> - t_boolean() - end; - #t_tuple{exact=false} -> - t_boolean(); - any -> - t_boolean(); - _ -> - t_atom(false) - end; +type(is_nonempty_list, [_], _Ts, _Ds) -> + t_boolean(); +type(is_tagged_tuple, [_,#b_literal{},#b_literal{}], _Ts, _Ds) -> + t_boolean(); +type(put_map, _Args, _Ts, _Ds) -> + map; type(put_list, _Args, _Ts, _Ds) -> cons; type(put_tuple, Args, _Ts, _Ds) -> @@ -444,11 +835,16 @@ type(put_tuple, Args, _Ts, _Ds) -> _ -> #t_tuple{exact=true,size=length(Args)} end; -type(succeeded, [#b_var{name=Src}], Ts, Ds) -> +type(succeeded, [#b_var{}=Src], Ts, Ds) -> case maps:get(Src, Ds) of #b_set{op={bif,Bif},args=BifArgs} -> Types = get_types(BifArgs, Ts), case {Bif,Types} of + {BoolOp,[T1,T2]} when BoolOp =:= 'and'; BoolOp =:= 'or' -> + case t_is_boolean(T1) andalso t_is_boolean(T2) of + true -> t_atom(true); + false -> t_boolean() + end; {byte_size,[{binary,_}]} -> t_atom(true); {bit_size,[{binary,_}]} -> @@ -478,6 +874,8 @@ type(succeeded, [#b_var{name=Src}], Ts, Ds) -> #b_set{} -> t_boolean() end; +type(succeeded, [#b_literal{}], _Ts, _Ds) -> + t_atom(true); type(_, _, _, _) -> any. arith_op_type(Args, Ts) -> @@ -493,7 +891,6 @@ arith_op_type(Args, Ts) -> (number, #t_integer{}) -> number; (number, float) -> float; (any, _) -> number; - (_, any) -> number; (Same, Same) -> Same; (_, _) -> none end, unknown, Types). @@ -554,7 +951,6 @@ will_succeed(is_list, Type) -> case Type of list -> yes; cons -> yes; - nil -> yes; _ -> no end; will_succeed(is_map, Type) -> @@ -577,8 +973,6 @@ will_succeed(is_tuple, Type) -> will_succeed(_, _) -> maybe. -band_type([#b_literal{val=Int},Other], Ts) when is_integer(Int) -> - band_type_1(Int, Other, Ts); band_type([Other,#b_literal{val=Int}], Ts) when is_integer(Int) -> band_type_1(Int, Other, Ts); band_type([_,_], _) -> t_integer(). @@ -662,28 +1056,82 @@ simplify_switch_bool(#b_switch{arg=B,list=List0}=Sw, Ts, Ds) -> Sw end. -simplify_not(#b_br{bool=#b_var{name=V},succ=Succ,fail=Fail}=Br, Ts, Ds) -> +simplify_not(#b_br{bool=#b_var{}=V,succ=Succ,fail=Fail}=Br0, Ts, Ds) -> case Ds of #{V:=#b_set{op={bif,'not'},args=[Bool]}} -> case t_is_boolean(get_type(Bool, Ts)) of true -> - Br#b_br{bool=Bool,succ=Fail,fail=Succ}; + Br = Br0#b_br{bool=Bool,succ=Fail,fail=Succ}, + beam_ssa:normalize(Br); false -> - Br + Br0 end; #{} -> - Br + Br0 end. -get_literals(Values, Ts) -> - [get_literal_from_type(get_type(Val, Ts)) || Val <- Values]. +%%% +%%% Calculate the set of variables that are only used once in the +%%% terminator of the block that defines them. That will allow us to +%%% discard type information for variables that will never be +%%% referenced by the successor blocks, potentially improving +%%% compilation times. +%%% + +used_once(Linear, Args) -> + Map0 = used_once_1(reverse(Linear), #{}), + Map = maps:without(Args, Map0), + cerl_sets:from_list(maps:keys(Map)). + +used_once_1([{L,#b_blk{is=Is,last=Last}}|Bs], Uses0) -> + Uses1 = used_once_last_uses(beam_ssa:used(Last), L, Uses0), + Uses = used_once_2(reverse(Is), L, Uses1), + used_once_1(Bs, Uses); +used_once_1([], Uses) -> Uses. + +used_once_2([#b_set{dst=Dst}=I|Is], L, Uses0) -> + Uses = used_once_uses(beam_ssa:used(I), L, Uses0), + case Uses of + #{Dst:=[L]} -> + used_once_2(Is, L, Uses); + #{} -> + %% Used more than once or used once in + %% in another block. + used_once_2(Is, L, maps:remove(Dst, Uses)) + end; +used_once_2([], _, Uses) -> Uses. + +used_once_uses([V|Vs], L, Uses) -> + case Uses of + #{V:=more_than_once} -> + used_once_uses(Vs, L, Uses); + #{} -> + %% Already used or first use is not in + %% a terminator. + used_once_uses(Vs, L, Uses#{V=>more_than_once}) + end; +used_once_uses([], _, Uses) -> Uses. + +used_once_last_uses([V|Vs], L, Uses) -> + case Uses of + #{V:=[_]} -> + %% Second time this variable is used. + used_once_last_uses(Vs, L, Uses#{V:=more_than_once}); + #{V:=more_than_once} -> + %% Used at least twice before. + used_once_last_uses(Vs, L, Uses); + #{} -> + %% First time this variable is used. + used_once_last_uses(Vs, L, Uses#{V=>[L]}) + end; +used_once_last_uses([], _, Uses) -> Uses. + get_types(Values, Ts) -> [get_type(Val, Ts) || Val <- Values]. - -spec get_type(beam_ssa:value(), type_db()) -> type(). -get_type(#b_var{name=V}, Ts) -> +get_type(#b_var{}=V, Ts) -> #{V:=T} = Ts, T; get_type(#b_literal{val=Val}, _Ts) -> @@ -709,40 +1157,117 @@ get_type(#b_literal{val=Val}, _Ts) -> any end. -infer_types(V, Ts, #d{ds=Ds}) -> +%% infer_types(Var, Types, #d{}) -> {SuccTypes,FailTypes} +%% Looking at the expression that defines the variable Var, infer +%% the types for the variables in the arguments. Return the updated +%% type database for the case that the expression evaluates to +%% true, and and for the case that it evaluates to false. +%% +%% Here is an example. The variable being asked about is +%% the variable Bool, which is defined like this: +%% +%% Bool = is_nonempty_list L +%% +%% If 'is_nonempty_list L' evaluates to 'true', L must +%% must be cons. The meet of the previously known type of L and 'cons' +%% will be added to SuccTypes. +%% +%% On the other hand, if 'is_nonempty_list L' evaluates to false, L +%% is not cons and cons can be subtracted from the previously known +%% type for L. For example, if L was known to be 'list', subtracting +%% 'cons' would give 'nil' as the only possible type. The result of the +%% subtraction for L will be added to FailTypes. +%% +%% Here is another example, asking about the variable Bool: +%% +%% Head = bif:hd L +%% Bool = succeeded Head +%% +%% 'succeeded Head' will evaluate to 'true' if the instrution that +%% defined Head succeeded. In this case, it is the 'bif:hd L' +%% instruction, which will succeed if L is 'cons'. Thus, the meet of +%% the previous type for L and 'cons' will be added to SuccTypes. +%% +%% If 'succeeded Head' evaluates to 'false', it means that 'bif:hd L' +%% failed and that L is not 'cons'. 'cons' can be subtracted from the +%% previously known type for L and the result put in FailTypes. + +infer_types(#b_var{}=V, Ts, #d{ds=Ds}) -> #{V:=#b_set{op=Op,args=Args}} = Ds, - Types = infer_type(Op, Args, Ds), - meet_types(Types, Ts). + Types0 = infer_type(Op, Args, Ds), + + %% We must be careful with types inferred from '=:='. + %% + %% If we have seen L =:= [a], we know that L is 'cons' if the + %% comparison succeeds. However, if the comparison fails, L could + %% still be 'cons'. Therefore, we must not subtract 'cons' from the + %% previous type of L. + %% + %% However, it is safe to subtract a type inferred from '=:=' if + %% it is single-valued, e.g. if it is [] or the atom 'true'. + EqTypes0 = infer_eq_type(Op, Args, Ts, Ds), + {Types1,EqTypes} = partition(fun({_,T}) -> + is_singleton_type(T) + end, EqTypes0), + + Types = Types1 ++ Types0, + {meet_types(EqTypes++Types, Ts),subtract_types(Types, Ts)}. + +infer_eq_type({bif,'=:='}, [#b_var{}=Src,#b_literal{}=Lit], Ts, Ds) -> + Def = maps:get(Src, Ds), + Type = get_type(Lit, Ts), + [{Src,Type}|infer_tuple_size(Def, Lit) ++ + infer_first_element(Def, Lit)]; +infer_eq_type({bif,'=:='}, [#b_var{}=Arg0,#b_var{}=Arg1], Ts, _Ds) -> + %% As an example, assume that L1 is known to be 'list', and L2 is + %% known to be 'cons'. Then if 'L1 =:= L2' evaluates to 'true', it can + %% be inferred that L1 is 'cons' (the meet of 'cons' and 'list'). + Type0 = get_type(Arg0, Ts), + Type1 = get_type(Arg1, Ts), + Type = meet(Type0, Type1), + [{V,MeetType} || + {V,OrigType,MeetType} <- + [{Arg0,Type0,Type},{Arg1,Type1,Type}], + OrigType =/= MeetType]; +infer_eq_type(_Op, _Args, _Ts, _Ds) -> + []. -infer_type({bif,element}, [#b_literal{val=Pos},#b_var{name=Tuple}], _Ds) -> +infer_type({bif,element}, [#b_literal{val=Pos},#b_var{}=Tuple], _Ds) -> if is_integer(Pos), 1 =< Pos -> [{Tuple,#t_tuple{size=Pos}}]; true -> [] end; -infer_type({bif,'=:='}, [#b_var{name=Src},#b_literal{}=Lit], Ds) -> - Def = maps:get(Src, Ds), - Type = get_type(Lit, #{}), - [{Src,Type}|infer_tuple_size(Def, Lit) ++ - infer_first_element(Def, Lit)]; -infer_type({bif,Bif}, [#b_var{name=Src}]=Args, _Ds) -> +infer_type({bif,element}, [#b_var{}=Position,#b_var{}=Tuple], _Ds) -> + [{Position,t_integer()},{Tuple,#t_tuple{}}]; +infer_type({bif,Bif}, [#b_var{}=Src]=Args, _Ds) -> case inferred_bif_type(Bif, Args) of any -> []; T -> [{Src,T}] end; -infer_type({bif,is_map_key}, [_,#b_var{name=Src}], _Ds) -> +infer_type({bif,binary_part}, [#b_var{}=Src,_], _Ds) -> + [{Src,{binary,8}}]; +infer_type({bif,is_map_key}, [_,#b_var{}=Src], _Ds) -> [{Src,map}]; -infer_type({bif,map_get}, [_,#b_var{name=Src}], _Ds) -> +infer_type({bif,map_get}, [_,#b_var{}=Src], _Ds) -> [{Src,map}]; -infer_type(bs_start_match, [#b_var{name=Bin}], _Ds) -> +infer_type({bif,Bif}, [_,_]=Args, _Ds) -> + case inferred_bif_type(Bif, Args) of + any -> []; + T -> [{A,T} || #b_var{}=A <- Args] + end; +infer_type({bif,binary_part}, [#b_var{}=Src,Pos,Len], _Ds) -> + [{Src,{binary,8}}| + [{V,t_integer()} || #b_var{}=V <- [Pos,Len]]]; +infer_type(bs_start_match, [#b_var{}=Bin], _Ds) -> [{Bin,{binary,1}}]; -infer_type(is_nonempty_list, [#b_var{name=Src}], _Ds) -> +infer_type(is_nonempty_list, [#b_var{}=Src], _Ds) -> [{Src,cons}]; -infer_type(is_tagged_tuple, [#b_var{name=Src},#b_literal{val=Size}, +infer_type(is_tagged_tuple, [#b_var{}=Src,#b_literal{val=Size}, #b_literal{val=Tag}], _Ds) -> [{Src,#t_tuple{exact=true,size=Size,elements=[Tag]}}]; -infer_type(succeeded, [#b_var{name=Src}], Ds) -> +infer_type(succeeded, [#b_var{}=Src], Ds) -> #b_set{op=Op,args=Args} = maps:get(Src, Ds), infer_type(Op, Args, Ds); infer_type(_Op, _Args, _Ds) -> @@ -755,7 +1280,6 @@ infer_type(_Op, _Args, _Ds) -> %% Note that that the following BIFs are handle elsewhere: %% %% band/2 -%% tuple_size/1 bif_type(abs, [_]) -> number; bif_type(bit_size, [_]) -> t_integer(); @@ -766,9 +1290,12 @@ bif_type(floor, [_]) -> t_integer(); bif_type(is_map_key, [_,_]) -> t_boolean(); bif_type(length, [_]) -> t_integer(); bif_type(map_size, [_]) -> t_integer(); +bif_type(node, []) -> #t_atom{}; +bif_type(node, [_]) -> #t_atom{}; bif_type(round, [_]) -> t_integer(); bif_type(size, [_]) -> t_integer(); bif_type(trunc, [_]) -> t_integer(); +bif_type(tuple_size, [_]) -> t_integer(); bif_type('bnot', [_]) -> t_integer(); bif_type('bor', [_,_]) -> t_integer(); bif_type('bsl', [_,_]) -> t_integer(); @@ -803,22 +1330,42 @@ inferred_bif_type(is_number, [_]) -> number; inferred_bif_type(is_tuple, [_]) -> #t_tuple{}; inferred_bif_type(abs, [_]) -> number; inferred_bif_type(bit_size, [_]) -> {binary,1}; +inferred_bif_type('bnot', [_]) -> t_integer(); inferred_bif_type(byte_size, [_]) -> {binary,1}; inferred_bif_type(ceil, [_]) -> number; inferred_bif_type(float, [_]) -> number; inferred_bif_type(floor, [_]) -> number; +inferred_bif_type(hd, [_]) -> cons; +inferred_bif_type(length, [_]) -> list; +inferred_bif_type(map_size, [_]) -> map; +inferred_bif_type('not', [_]) -> t_boolean(); inferred_bif_type(round, [_]) -> number; inferred_bif_type(trunc, [_]) -> number; +inferred_bif_type(tl, [_]) -> cons; inferred_bif_type(tuple_size, [_]) -> #t_tuple{}; +inferred_bif_type('and', [_,_]) -> t_boolean(); +inferred_bif_type('or', [_,_]) -> t_boolean(); +inferred_bif_type('xor', [_,_]) -> t_boolean(); +inferred_bif_type('band', [_,_]) -> t_integer(); +inferred_bif_type('bor', [_,_]) -> t_integer(); +inferred_bif_type('bsl', [_,_]) -> t_integer(); +inferred_bif_type('bsr', [_,_]) -> t_integer(); +inferred_bif_type('bxor', [_,_]) -> t_integer(); +inferred_bif_type('div', [_,_]) -> t_integer(); +inferred_bif_type('rem', [_,_]) -> t_integer(); +inferred_bif_type('+', [_,_]) -> number; +inferred_bif_type('-', [_,_]) -> number; +inferred_bif_type('*', [_,_]) -> number; +inferred_bif_type('/', [_,_]) -> number; inferred_bif_type(_, _) -> any. -infer_tuple_size(#b_set{op={bif,tuple_size},args=[#b_var{name=Tuple}]}, +infer_tuple_size(#b_set{op={bif,tuple_size},args=[#b_var{}=Tuple]}, #b_literal{val=Size}) when is_integer(Size) -> [{Tuple,#t_tuple{exact=true,size=Size}}]; infer_tuple_size(_, _) -> []. infer_first_element(#b_set{op=get_tuple_element, - args=[#b_var{name=Tuple},#b_literal{val=0}]}, + args=[#b_var{}=Tuple,#b_literal{val=0}]}, #b_literal{val=First}) -> [{Tuple,#t_tuple{size=1,elements=[First]}}]; infer_first_element(_, _) -> []. @@ -918,6 +1465,9 @@ t_tuple_size(#t_tuple{size=Size,exact=true}) -> t_tuple_size(_) -> none. +is_singleton_type(Type) -> + get_literal_from_type(Type) =/= none. + %% join(Type1, Type2) -> Type %% Return the "join" of Type1 and Type2. The join is a more general %% type than Type1 and Type2. For example: @@ -982,14 +1532,40 @@ gcd(A, B) -> meet_types([{V,T0}|Vs], Ts) -> #{V:=T1} = Ts, - T = meet(T0, T1), - meet_types(Vs, Ts#{V:=T}); + case meet(T0, T1) of + T1 -> meet_types(Vs, Ts); + T -> meet_types(Vs, Ts#{V:=T}) + end; meet_types([], Ts) -> Ts. meet([T1,T2|Ts]) -> meet([meet(T1, T2)|Ts]); meet([T]) -> T. +subtract_types([{V,T0}|Vs], Ts) -> + #{V:=T1} = Ts, + case subtract(T1, T0) of + T1 -> subtract_types(Vs, Ts); + T -> subtract_types(Vs, Ts#{V:=T}) + end; +subtract_types([], Ts) -> Ts. + +%% subtract(Type1, Type2) -> Type. +%% Subtract Type2 from Type1. Example: +%% +%% subtract(list, cons) -> nil + +subtract(#t_atom{elements=[_|_]=Set0}, #t_atom{elements=[_|_]=Set1}) -> + case ordsets:subtract(Set0, Set1) of + [] -> none; + [_|_]=Set -> #t_atom{elements=Set} + end; +subtract(number, float) -> #t_integer{}; +subtract(number, #t_integer{elements=any}) -> float; +subtract(list, cons) -> nil; +subtract(list, nil) -> cons; +subtract(T, _) -> T. + %% meet(Type1, Type2) -> Type %% Return the "meet" of Type1 and Type2. The meet is a narrower %% type than Type1 and Type2. For example: @@ -1025,13 +1601,9 @@ meet(#t_integer{elements={Min1,Max1}}, #t_integer{elements={Min2,Max2}}) -> #t_integer{elements={max(Min1, Min2),min(Max1, Max2)}}; meet(#t_integer{}=T, number) -> T; -meet(float, number) -> float; -meet(#t_integer{}=T, number) -> T; -meet(float, number) -> float; +meet(float=T, number) -> T; meet(number, #t_integer{}=T) -> T; -meet(#t_integer{}=T, number) -> T; meet(number, float=T) -> T; -meet(float=T, number) -> T; meet(list, cons) -> cons; meet(list, nil) -> nil; meet(cons, list) -> cons; diff --git a/lib/compiler/src/beam_trim.erl b/lib/compiler/src/beam_trim.erl index 1acbedd45b..acf3838da4 100644 --- a/lib/compiler/src/beam_trim.erl +++ b/lib/compiler/src/beam_trim.erl @@ -21,12 +21,11 @@ -module(beam_trim). -export([module/2]). --import(lists, [reverse/1,reverse/2,splitwith/2,sort/1]). +-import(lists, [any/2,member/2,reverse/1,reverse/2,splitwith/2,sort/1]). -record(st, - {safe :: gb_sets:set(beam_asm:label()), %Safe labels. - lbl :: beam_utils:code_index() %Code at each label. - }). + {safe :: cerl_sets:set(beam_asm:label()) %Safe labels. + }). -spec module(beam_utils:module_code(), [compile:option()]) -> {'ok',beam_utils:module_code()}. @@ -36,10 +35,15 @@ module({Mod,Exp,Attr,Fs0,Lc}, _Opts) -> {ok,{Mod,Exp,Attr,Fs,Lc}}. function({function,Name,Arity,CLabel,Is0}) -> - %%ok = io:fwrite("~w: ~p\n", [?LINE,{Name,Arity}]), - St = #st{safe=safe_labels(Is0, []),lbl=beam_utils:index_labels(Is0)}, - Is = trim(Is0, St, []), - {function,Name,Arity,CLabel,Is}. + try + St = #st{safe=safe_labels(Is0, [])}, + Is = trim(Is0, St, []), + {function,Name,Arity,CLabel,Is} + catch + Class:Error:Stack -> + io:fwrite("Function: ~w/~w\n", [Name,Arity]), + erlang:raise(Class, Error, Stack) + end. trim([{kill,_}|_]=Is0, St, Acc) -> {Kills0,Is1} = splitwith(fun({kill,_}) -> true; @@ -47,14 +51,33 @@ trim([{kill,_}|_]=Is0, St, Acc) -> end, Is0), Kills = sort(Kills0), try - {FrameSize,Layout} = frame_layout(Is1, Kills, St), - Configs = trim_instructions(Layout), - try_remap(Configs, Is1, FrameSize) - of + %% Find out the size and layout of the stack frame. + %% Example of a layout: + %% + %% [{kill,{y,0}},{dead,{y,1},{live,{y,2}},{kill,{y,3}}] + %% + %% That means that y0 and y3 are to be killed, that y1 + %% has been killed previously, and that y2 is live. + {FrameSize,Layout} = frame_layout(Is1, Kills, St), + + %% Calculate all recipes that are not worse in terms + %% of estimated execution time. The recipes are ordered + %% in descending order from how much they trim. + Recipes = trim_recipes(Layout), + + %% Try the recipes in order. A recipe may not work out because + %% a register that was previously killed may be + %% resurrected. If that happens, the next recipe, which trims + %% less, will be tried. + try_remap(Recipes, Is1, FrameSize) + of {Is,TrimInstr} -> + %% One of the recipes was applied. trim(Is, St, reverse(TrimInstr)++Acc) catch not_possible -> + %% No recipe worked out. Use the original kill + %% instructions. trim(Is1, St, reverse(Kills, Acc)) end; trim([I|Is], St, Acc) -> @@ -62,34 +85,42 @@ trim([I|Is], St, Acc) -> trim([], _, Acc) -> reverse(Acc). -%% trim_instructions([{kill,R}|{live,R}|{dead,R}]) -> {[Instruction],MapFun} -%% Figure out the sequence of moves and trim to use. +%% trim_recipes([{kill,R}|{live,R}|{dead,R}]) -> [Recipe]. +%% Recipe = {Kills,NumberToTrim,Moves} +%% Kills = [{kill,Y}] +%% Moves = [{move,SrcY,DstY}] +%% +%% Calculate how to best trim the stack and kill the correct +%% Y registers. Return a list of possible recipes. The best +%% recipe (the one that trims the most) is first in the list. +%% All of the recipes are no worse in estimated execution time +%% than the original sequences of kill instructions. -trim_instructions(Layout) -> +trim_recipes(Layout) -> Cost = length([I || {kill,_}=I <- Layout]), - trim_instructions_1(Layout, 0, [], {Cost,[]}). + trim_recipes_1(Layout, 0, [], {Cost,[]}). -trim_instructions_1([{kill,{y,Trim0}}|Ks], Trim0, Moves, Config0) -> +trim_recipes_1([{kill,{y,Trim0}}|Ks], Trim0, Moves, Recipes0) -> Trim = Trim0 + 1, - Config = save_config(Ks, Trim, Moves, Config0), - trim_instructions_1(Ks, Trim, Moves, Config); -trim_instructions_1([{dead,{y,Trim0}}|Ks], Trim0, Moves, Config0) -> + Recipes = save_recipe(Ks, Trim, Moves, Recipes0), + trim_recipes_1(Ks, Trim, Moves, Recipes); +trim_recipes_1([{dead,{y,Trim0}}|Ks], Trim0, Moves, Recipes0) -> Trim = Trim0 + 1, - Config = save_config(Ks, Trim, Moves, Config0), - trim_instructions_1(Ks, Trim, Moves, Config); -trim_instructions_1([{live,{y,Trim0}=Src}|Ks0], Trim0, Moves0, Config0) -> + Recipes = save_recipe(Ks, Trim, Moves, Recipes0), + trim_recipes_1(Ks, Trim, Moves, Recipes); +trim_recipes_1([{live,{y,Trim0}=Src}|Ks0], Trim0, Moves0, Recipes0) -> case take_last_dead(Ks0) of none -> - {_,ConfigList} = Config0, - ConfigList; + {_,RecipesList} = Recipes0, + RecipesList; {Dst,Ks} -> Trim = Trim0 + 1, Moves = [{move,Src,Dst}|Moves0], - Config = save_config(Ks, Trim, Moves, Config0), - trim_instructions_1(Ks, Trim, Moves, Config) + Recipes = save_recipe(Ks, Trim, Moves, Recipes0), + trim_recipes_1(Ks, Trim, Moves, Recipes) end; -trim_instructions_1([], _, _, {_,ConfigList}) -> - ConfigList. +trim_recipes_1([], _, _, {_,RecipesList}) -> + RecipesList. take_last_dead(L) -> take_last_dead_1(reverse(L)). @@ -100,28 +131,48 @@ take_last_dead_1([{dead,Reg}|Is]) -> {Reg,reverse(Is)}; take_last_dead_1(_) -> none. -save_config(Ks, Trim, Moves, {MaxCost,Acc}=Config) -> - case config_cost(Ks, Moves) of - Cost when Cost =< MaxCost -> - {MaxCost,[{Ks,Trim,Moves}|Acc]}; +save_recipe(Ks, Trim, Moves, {MaxCost,Acc}=Recipes) -> + case recipe_cost(Ks, Moves) of + Cost when Cost =< MaxCost -> + %% The price is right. + {MaxCost,[{Ks,Trim,Moves}|Acc]}; _Cost -> - Config + %% Too expensive. + Recipes end. -config_cost(Ks, Moves) -> +recipe_cost(Ks, Moves) -> %% We estimate that a {move,{y,_},{y,_}} instruction is roughly twice as %% expensive as a {kill,{y,_}} instruction. A {trim,_} instruction is %% roughly as expensive as a {kill,{y,_}} instruction. - config_cost_1(Ks, 1+2*length(Moves)). + recipe_cost_1(Ks, 1+2*length(Moves)). -config_cost_1([{kill,_}|Ks], Cost) -> - config_cost_1(Ks, Cost+1); -config_cost_1([_|Ks], Cost) -> - config_cost_1(Ks, Cost); -config_cost_1([], Cost) -> Cost. +recipe_cost_1([{kill,_}|Ks], Cost) -> + recipe_cost_1(Ks, Cost+1); +recipe_cost_1([_|Ks], Cost) -> + recipe_cost_1(Ks, Cost); +recipe_cost_1([], Cost) -> Cost. -expand_config({Layout,Trim,Moves}, FrameSize) -> +%% try_remap([Recipe], [Instruction], FrameSize) -> +%% {[Instruction],[TrimInstruction]}. +%% Try to renumber Y registers in the instruction stream. The +%% first rececipe that works will be used. +%% +%% This function will issue a `not_possible` exception if none +%% of the recipes were possible to apply. + +try_remap([R|Rs], Is, FrameSize) -> + {TrimInstr,Map} = expand_recipe(R, FrameSize), + try + {remap(Is, Map, []),TrimInstr} + catch + throw:not_possible -> + try_remap(Rs, Is, FrameSize) + end; +try_remap([], _, _) -> throw(not_possible). + +expand_recipe({Layout,Trim,Moves}, FrameSize) -> Kills = [Kill || {kill,_}=Kill <- Layout], {Kills++reverse(Moves, [{trim,Trim,FrameSize-Trim}]),create_map(Trim, Moves)}. @@ -132,16 +183,16 @@ create_map(Trim, []) -> (Any) -> Any end; create_map(Trim, Moves) -> - GbTree0 = [{Src,Dst-Trim} || {move,{y,Src},{y,Dst}} <- Moves], - GbTree = gb_trees:from_orddict(sort(GbTree0)), - IllegalTargets = gb_sets:from_list([Dst || {move,_,{y,Dst}} <- Moves]), + Map0 = [{Src,Dst-Trim} || {move,{y,Src},{y,Dst}} <- Moves], + Map = maps:from_list(Map0), + IllegalTargets = cerl_sets:from_list([Dst || {move,_,{y,Dst}} <- Moves]), fun({y,Y0}) when Y0 < Trim -> - case gb_trees:lookup(Y0, GbTree) of - {value,Y} -> {y,Y}; - none -> throw(not_possible) - end; + case Map of + #{Y0:=Y} -> {y,Y}; + #{} -> throw(not_possible) + end; ({y,Y}) -> - case gb_sets:is_element(Y, IllegalTargets) of + case cerl_sets:is_element(Y, IllegalTargets) of true -> throw(not_possible); false -> {y,Y-Trim} end; @@ -149,19 +200,17 @@ create_map(Trim, Moves) -> (Any) -> Any end. -try_remap([C|Cs], Is, FrameSize) -> - {TrimInstr,Map} = expand_config(C, FrameSize), - try - {remap(Is, Map, []),TrimInstr} - catch - throw:not_possible -> - try_remap(Cs, Is, FrameSize) - end; -try_remap([], _, _) -> throw(not_possible). - +remap([{'%',_}=I|Is], Map, Acc) -> + remap(Is, Map, [I|Acc]); remap([{block,Bl0}|Is], Map, Acc) -> Bl = remap_block(Bl0, Map, []), remap(Is, Map, [{block,Bl}|Acc]); +remap([{bs_get_tail,Src,Dst,Live}|Is], Map, Acc) -> + I = {bs_get_tail,Map(Src),Map(Dst),Live}, + remap(Is, Map, [I|Acc]); +remap([{bs_set_position,Src1,Src2}|Is], Map, Acc) -> + I = {bs_set_position,Map(Src1),Map(Src2)}, + remap(Is, Map, [I|Acc]); remap([{call_fun,_}=I|Is], Map, Acc) -> remap(Is, Map, [I|Acc]); remap([{call,_,_}=I|Is], Map, Acc) -> @@ -205,35 +254,68 @@ remap([return|_]=Is, _, Acc) -> reverse(Acc, Is); remap([{line,_}=I|Is], Map, Acc) -> remap(Is, Map, [I|Acc]). - + remap_block([{set,Ds0,Ss0,Info}|Is], Map, Acc) -> Ds = [Map(D) || D <- Ds0], Ss = [Map(S) || S <- Ss0], remap_block(Is, Map, [{set,Ds,Ss,Info}|Acc]); remap_block([], _, Acc) -> reverse(Acc). - -safe_labels([{label,L},{line,_},{badmatch,{Tag,_}}|Is], Acc) when Tag =/= y -> - safe_labels(Is, [L|Acc]); -safe_labels([{label,L},{line,_},{case_end,{Tag,_}}|Is], Acc) when Tag =/= y -> - safe_labels(Is, [L|Acc]); -safe_labels([{label,L},{line,_},if_end|Is], Acc) -> - safe_labels(Is, [L|Acc]); -safe_labels([{label,L}, - {block,[{set,[{x,0}],[{Tag,_}],move}]}, - {line,_}, - {call_ext,1,{extfunc,erlang,error,1}}|Is], Acc) when Tag =/= y -> - safe_labels(Is, [L|Acc]); + +%% safe_labels([Instruction], Accumulator) -> gb_set() +%% Build a gb_set of safe labels. The code at a safe +%% label does not depend on the values in a specific +%% Y register, only that all Y registers are initialized +%% so that it safe to scan the stack when an exception +%% is generated. +%% +%% In other words, code at a safe label will continue +%% to work if Y registers have been renumbered and +%% the size of the stack frame has changed. + +safe_labels([{label,L}|Is], Acc) -> + case is_safe_label(Is) of + true -> safe_labels(Is, [L|Acc]); + false -> safe_labels(Is, Acc) + end; safe_labels([_|Is], Acc) -> safe_labels(Is, Acc); -safe_labels([], Acc) -> gb_sets:from_list(Acc). +safe_labels([], Acc) -> cerl_sets:from_list(Acc). + +is_safe_label([{'%',_}|Is]) -> + is_safe_label(Is); +is_safe_label([{line,_}|Is]) -> + is_safe_label(Is); +is_safe_label([{badmatch,{Tag,_}}|_]) -> + Tag =/= y; +is_safe_label([{case_end,{Tag,_}}|_]) -> + Tag =/= y; +is_safe_label([{try_case_end,{Tag,_}}|_]) -> + Tag =/= y; +is_safe_label([if_end|_]) -> + true; +is_safe_label([{block,Bl}|Is]) -> + is_safe_label_block(Bl) andalso is_safe_label(Is); +is_safe_label([{call_ext,_,{extfunc,M,F,A}}|_]) -> + erl_bifs:is_exit_bif(M, F, A); +is_safe_label(_) -> false. + +is_safe_label_block([{set,Ds,Ss,_}|Is]) -> + IsYreg = fun({y,_}) -> true; + (_) -> false + end, + %% This instruction is safe if the instruction + %% neither reads or writes Y registers. + not (any(IsYreg, Ss) orelse any(IsYreg, Ds)) andalso + is_safe_label_block(Is); +is_safe_label_block([]) -> true. %% frame_layout([Instruction], [{kill,_}], St) -> %% [{kill,Reg} | {live,Reg} | {dead,Reg}] %% Figure out the layout of the stack frame. -frame_layout(Is, Kills, #st{safe=Safe,lbl=D}) -> +frame_layout(Is, Kills, #st{safe=Safe}) -> N = frame_size(Is, Safe), - IsKilled = fun(R) -> beam_utils:is_not_used(R, Is, D) end, + IsKilled = fun(R) -> is_not_used(R, Is) end, {N,frame_layout_1(Kills, 0, N, IsKilled, [])}. frame_layout_1([{kill,{y,Y}}=I|Ks], Y, N, IsKilled, Acc) -> @@ -253,7 +335,14 @@ frame_layout_2(Is) -> reverse(Is). %% frame_size([Instruction], SafeLabels) -> FrameSize %% Find out the frame size by looking at the code that follows. +%% +%% Implicitly, also check that the instructions are a straight +%% sequence of code that ends in a return. Any branches are +%% to safe labels (i.e., the code at those labels don't depend +%% on the contents of any Y register). +frame_size([{'%',_}|Is], Safe) -> + frame_size(Is, Safe); frame_size([{block,_}|Is], Safe) -> frame_size(Is, Safe); frame_size([{call_fun,_}|Is], Safe) -> @@ -285,15 +374,94 @@ frame_size([{make_fun2,_,_,_,_}|Is], Safe) -> frame_size(Is, Safe); frame_size([{get_map_elements,{f,L},_,_}|Is], Safe) -> frame_size_branch(L, Is, Safe); -frame_size([{deallocate,N}|_], _) -> N; +frame_size([{deallocate,N}|_], _) -> + N; frame_size([{line,_}|Is], Safe) -> frame_size(Is, Safe); +frame_size([{bs_set_position,_,_}|Is], Safe) -> + frame_size(Is, Safe); +frame_size([{bs_get_tail,_,_,_}|Is], Safe) -> + frame_size(Is, Safe); frame_size(_, _) -> throw(not_possible). frame_size_branch(0, Is, Safe) -> frame_size(Is, Safe); frame_size_branch(L, Is, Safe) -> - case gb_sets:is_member(L, Safe) of + case cerl_sets:is_element(L, Safe) of false -> throw(not_possible); true -> frame_size(Is, Safe) end. + +%% is_not_used(Y, [Instruction]) -> true|false. +%% Test whether the value of Y is unused in the instruction sequence. +%% Return true if the value of Y is not used, and false if it is used. +%% +%% This function handles the same instructions as frame_size/2. It +%% assumes that any labels in the instructions are safe labels. + +is_not_used(Y, [{'%',_}|Is]) -> + is_not_used(Y, Is); +is_not_used(Y, [{apply,_}|Is]) -> + is_not_used(Y, Is); +is_not_used(Y, [{bif,_,{f,_},Ss,Dst}|Is]) -> + is_not_used_ss_dst(Y, Ss, Dst, Is); +is_not_used(Y, [{block,Bl}|Is]) -> + case is_not_used_block(Y, Bl) of + used -> false; + killed -> true; + transparent -> is_not_used(Y, Is) + end; +is_not_used(Y, [{bs_get_tail,Src,Dst,_}|Is]) -> + is_not_used_ss_dst(Y, [Src], Dst, Is); +is_not_used(Y, [{bs_init,_,_,_,Ss,Dst}|Is]) -> + is_not_used_ss_dst(Y, Ss, Dst, Is); +is_not_used(Y, [{bs_put,{f,_},_,Ss}|Is]) -> + not member(Y, Ss) andalso is_not_used(Y, Is); +is_not_used(Y, [{bs_set_position,Src1,Src2}|Is]) -> + Y =/= Src1 andalso Y =/= Src2 andalso + is_not_used(Y, Is); +is_not_used(Y, [{call,_,_}|Is]) -> + is_not_used(Y, Is); +is_not_used(Y, [{call_ext,_,_}=I|Is]) -> + beam_jump:is_exit_instruction(I) orelse is_not_used(Y, Is); +is_not_used(Y, [{call_fun,_}|Is]) -> + is_not_used(Y, Is); +is_not_used(_Y, [{deallocate,_}|_]) -> + true; +is_not_used(Y, [{gc_bif,_,{f,_},_Live,Ss,Dst}|Is]) -> + is_not_used_ss_dst(Y, Ss, Dst, Is); +is_not_used(Y, [{get_map_elements,{f,_},S,{list,List}}|Is]) -> + {Ss,Ds} = beam_utils:split_even(List), + case member(Y, [S|Ss]) of + true -> + false; + false -> + member(Y, Ds) orelse is_not_used(Y, Is) + end; +is_not_used(Y, [{kill,Yreg}|Is]) -> + Y =:= Yreg orelse is_not_used(Y, Is); +is_not_used(Y, [{line,_}|Is]) -> + is_not_used(Y, Is); +is_not_used(Y, [{make_fun2,_,_,_,_}|Is]) -> + is_not_used(Y, Is); +is_not_used(Y, [{test,_,_,Ss}|Is]) -> + not member(Y, Ss) andalso is_not_used(Y, Is); +is_not_used(Y, [{test,_Op,{f,_},_Live,Ss,Dst}|Is]) -> + is_not_used_ss_dst(Y, Ss, Dst, Is). + +is_not_used_block(Y, [{set,Ds,Ss,_}|Is]) -> + case member(Y, Ss) of + true -> + used; + false -> + case member(Y, Ds) of + true -> + killed; + false -> + is_not_used_block(Y, Is) + end + end; +is_not_used_block(_Y, []) -> transparent. + +is_not_used_ss_dst(Y, Ss, Dst, Is) -> + not member(Y, Ss) andalso (Y =:= Dst orelse is_not_used(Y, Is)). diff --git a/lib/compiler/src/beam_utils.erl b/lib/compiler/src/beam_utils.erl index 686d314c2d..6e6574c0b3 100644 --- a/lib/compiler/src/beam_utils.erl +++ b/lib/compiler/src/beam_utils.erl @@ -18,27 +18,16 @@ %% %CopyrightEnd% %% %% Purpose : Common utilities used by several optimization passes. -%% +%% -module(beam_utils). --export([is_killed/3,is_killed_at/3,is_not_used/3, - empty_label_index/0,index_label/3,index_labels/1,replace_labels/4, - code_at/2,bif_to_test/3,is_pure_test/1, - combine_heap_needs/2, - split_even/1 - ]). +-export([replace_labels/4,is_pure_test/1,split_even/1]). -export_type([code_index/0,module_code/0,instruction/0]). --import(lists, [flatmap/2,map/2,member/2,sort/1,reverse/1]). - --define(is_const(Val), (Val =:= nil orelse - element(1, Val) =:= integer orelse - element(1, Val) =:= float orelse - element(1, Val) =:= atom orelse - element(1, Val) =:= literal)). +-import(lists, [map/2,reverse/1]). -%% instruction() describes all instructions that are used during optimzation +%% instruction() describes all instructions that are used during optimization %% (from beam_a to beam_z). -type instruction() :: atom() | tuple(). @@ -54,97 +43,6 @@ -type fail() :: beam_asm:fail() | 'fail'. -type test() :: {'test',atom(),fail(),[beam_asm:src()]} | {'test',atom(),fail(),integer(),list(),beam_asm:reg()}. --type result_cache() :: gb_trees:tree(beam_asm:label(), 'killed' | 'used'). - --record(live, - {lbl :: code_index(), %Label to code index. - res :: result_cache()}). %Result cache for each label. - -%% is_killed(Register, [Instruction], State) -> true|false -%% Determine whether a register is killed by the instruction sequence. -%% If true is returned, it means that the register will not be -%% referenced in ANY way (not even indirectly by an allocate instruction); -%% i.e. it is OK to enter the instruction sequence with Register -%% containing garbage. -%% -%% The state (constructed by index_instructions/1) is used to allow us -%% to determine the kill state across branches. - --spec is_killed(beam_asm:reg(), [instruction()], code_index()) -> boolean(). - -is_killed(R, Is, D) -> - St = #live{lbl=D,res=gb_trees:empty()}, - case check_liveness(R, Is, St) of - {killed,_} -> true; - {exit_not_used,_} -> false; - {_,_} -> false - end. - -%% is_killed_at(Reg, Lbl, State) -> true|false -%% Determine whether Reg is killed at label Lbl. - --spec is_killed_at(beam_asm:reg(), beam_asm:label(), code_index()) -> boolean(). - -is_killed_at(R, Lbl, D) when is_integer(Lbl) -> - St0 = #live{lbl=D,res=gb_trees:empty()}, - case check_liveness_at(R, Lbl, St0) of - {killed,_} -> true; - {exit_not_used,_} -> false; - {_,_} -> false - end. - -%% is_not_used(Register, [Instruction], State) -> true|false -%% Determine whether a register is never used in the instruction sequence -%% (it could still be referenced by an allocate instruction, meaning that -%% it MUST be initialized, but that its value does not matter). -%% The state is used to allow us to determine the usage state -%% across branches. - --spec is_not_used(beam_asm:reg(), [instruction()], code_index()) -> boolean(). - -is_not_used(R, Is, D) -> - St = #live{lbl=D,res=gb_trees:empty()}, - case check_liveness(R, Is, St) of - {used,_} -> false; - {exit_not_used,_} -> true; - {_,_} -> true - end. - -%% index_labels(FunctionIs) -> State -%% Index the instruction sequence so that we can quickly -%% look up the instruction following a specific label. - --spec index_labels([instruction()]) -> code_index(). - -index_labels(Is) -> - index_labels_1(Is, []). - -%% empty_label_index() -> State -%% Create an empty label index. - --spec empty_label_index() -> code_index(). - -empty_label_index() -> - gb_trees:empty(). - -%% index_label(Label, [Instruction], State) -> State -%% Add an index for a label. - --spec index_label(beam_asm:label(), [instruction()], code_index()) -> - code_index(). - -index_label(Lbl, Is0, Acc) -> - Is = drop_labels(Is0), - gb_trees:enter(Lbl, Is, Acc). - - -%% code_at(Label, State) -> [I]. -%% Retrieve the code at the given label. - --spec code_at(beam_asm:label(), code_index()) -> [instruction()]. - -code_at(L, Ll) -> - gb_trees:get(L, Ll). %% replace_labels(FunctionIs, Tail, ReplaceDb, Fallback) -> FunctionIs. %% Replace all labels in instructions according to the ReplaceDb. @@ -158,44 +56,6 @@ code_at(L, Ll) -> replace_labels(Is, Acc, D, Fb) -> replace_labels_1(Is, Acc, D, Fb). -%% bif_to_test(Bif, [Op], Fail) -> {test,Test,Fail,[Op]} -%% Convert a BIF to a test. Fail if not possible. - --spec bif_to_test(atom(), list(), fail()) -> test(). - -bif_to_test(is_atom, [_]=Ops, Fail) -> {test,is_atom,Fail,Ops}; -bif_to_test(is_boolean, [_]=Ops, Fail) -> {test,is_boolean,Fail,Ops}; -bif_to_test(is_binary, [_]=Ops, Fail) -> {test,is_binary,Fail,Ops}; -bif_to_test(is_bitstring,[_]=Ops, Fail) -> {test,is_bitstr,Fail,Ops}; -bif_to_test(is_float, [_]=Ops, Fail) -> {test,is_float,Fail,Ops}; -bif_to_test(is_function, [_]=Ops, Fail) -> {test,is_function,Fail,Ops}; -bif_to_test(is_function, [_,_]=Ops, Fail) -> {test,is_function2,Fail,Ops}; -bif_to_test(is_integer, [_]=Ops, Fail) -> {test,is_integer,Fail,Ops}; -bif_to_test(is_list, [_]=Ops, Fail) -> {test,is_list,Fail,Ops}; -bif_to_test(is_map, [_]=Ops, Fail) -> {test,is_map,Fail,Ops}; -bif_to_test(is_number, [_]=Ops, Fail) -> {test,is_number,Fail,Ops}; -bif_to_test(is_pid, [_]=Ops, Fail) -> {test,is_pid,Fail,Ops}; -bif_to_test(is_port, [_]=Ops, Fail) -> {test,is_port,Fail,Ops}; -bif_to_test(is_reference, [_]=Ops, Fail) -> {test,is_reference,Fail,Ops}; -bif_to_test(is_tuple, [_]=Ops, Fail) -> {test,is_tuple,Fail,Ops}; -bif_to_test('=<', [A,B], Fail) -> {test,is_ge,Fail,[B,A]}; -bif_to_test('>', [A,B], Fail) -> {test,is_lt,Fail,[B,A]}; -bif_to_test('<', [_,_]=Ops, Fail) -> {test,is_lt,Fail,Ops}; -bif_to_test('>=', [_,_]=Ops, Fail) -> {test,is_ge,Fail,Ops}; -bif_to_test('==', [C,A], Fail) when ?is_const(C) -> - {test,is_eq,Fail,[A,C]}; -bif_to_test('==', [_,_]=Ops, Fail) -> {test,is_eq,Fail,Ops}; -bif_to_test('/=', [C,A], Fail) when ?is_const(C) -> - {test,is_ne,Fail,[A,C]}; -bif_to_test('/=', [_,_]=Ops, Fail) -> {test,is_ne,Fail,Ops}; -bif_to_test('=:=', [C,A], Fail) when ?is_const(C) -> - {test,is_eq_exact,Fail,[A,C]}; -bif_to_test('=:=', [_,_]=Ops, Fail) -> {test,is_eq_exact,Fail,Ops}; -bif_to_test('=/=', [C,A], Fail) when ?is_const(C) -> - {test,is_ne_exact,Fail,[A,C]}; -bif_to_test('=/=', [_,_]=Ops, Fail) -> {test,is_ne_exact,Fail,Ops}. - - %% is_pure_test({test,Op,Fail,Ops}) -> true|false. %% Return 'true' if the test instruction does not modify any %% registers and/or bit syntax matching state. @@ -215,22 +75,9 @@ is_pure_test({test,test_arity,_,[_,_]}) -> true; is_pure_test({test,has_map_fields,_,[_|_]}) -> true; is_pure_test({test,is_bitstr,_,[_]}) -> true; is_pure_test({test,is_function2,_,[_,_]}) -> true; -is_pure_test({test,Op,_,Ops}) -> +is_pure_test({test,Op,_,Ops}) -> erl_internal:new_type_test(Op, length(Ops)). -%% combine_heap_needs(HeapNeed1, HeapNeed2) -> HeapNeed -%% Combine the heap need for two allocation instructions. - --type heap_need_tag() :: 'floats' | 'words'. --type heap_need() :: non_neg_integer() | - {'alloc',[{heap_need_tag(),non_neg_integer()}]}. --spec combine_heap_needs(heap_need(), heap_need()) -> heap_need(). - -combine_heap_needs(H1, H2) when is_integer(H1), is_integer(H2) -> - H1 + H2; -combine_heap_needs(H1, H2) -> - {alloc,combine_alloc_lists([H1,H2])}. - %% split_even/1 %% [1,2,3,4,5,6] -> {[1,3,5],[2,4,6]} @@ -242,443 +89,6 @@ split_even(Rs) -> split_even(Rs, [], []). %%% Local functions. %%% - -%% check_liveness(Reg, [Instruction], #live{}) -> -%% {killed | not_used | used, #live{}} -%% Find out whether Reg is used or killed in instruction sequence. -%% -%% killed - Reg is assigned or killed by an allocation instruction. -%% not_used - the value of Reg is not used, but Reg must not be garbage -%% exit_not_used - the value of Reg is not used, but must not be garbage -%% because the stack will be scanned because an -%% exit BIF will raise an exception -%% used - Reg is used - -check_liveness({fr,_}, _, St) -> - %% Conservatively always consider the floating point register used. - {used,St}; -check_liveness(R, [{block,Blk}|Is], St0) -> - case check_liveness_block(R, Blk, St0) of - {transparent,St1} -> - check_liveness(R, Is, St1); - {alloc_used,St1} -> - %% Used by an allocating instruction, but value not referenced. - %% Must check the rest of the instructions. - not_used(check_liveness(R, Is, St1)); - {Other,_}=Res when is_atom(Other) -> - Res - end; -check_liveness(R, [{label,_}|Is], St) -> - check_liveness(R, Is, St); -check_liveness(R, [{test,_,{f,Fail},As}|Is], St0) -> - case member(R, As) of - true -> - {used,St0}; - false -> - case check_liveness_at(R, Fail, St0) of - {killed,St1} -> - check_liveness(R, Is, St1); - {exit_not_used,St1} -> - not_used(check_liveness(R, Is, St1)); - {not_used,St1} -> - not_used(check_liveness(R, Is, St1)); - {used,_}=Used -> - Used - end - end; -check_liveness(R, [{test,Op,Fail,Live,Ss,Dst}|Is], St) -> - %% Check this instruction as a block to get a less conservative - %% result if the caller is is_not_used/3. - Block = [{set,[Dst],Ss,{alloc,Live,{bif,Op,Fail}}}], - check_liveness(R, [{block,Block}|Is], St); -check_liveness(R, [{select,_,R,_,_}|_], St) -> - {used,St}; -check_liveness(R, [{select,_,_,Fail,Branches}|_], St) -> - check_liveness_everywhere(R, [Fail|Branches], St); -check_liveness(R, [{jump,{f,F}}|_], St) -> - check_liveness_at(R, F, St); -check_liveness(R, [{case_end,Used}|_], St) -> - check_liveness_exit(R, Used, St); -check_liveness(R, [{try_case_end,Used}|_], St) -> - check_liveness_exit(R, Used, St); -check_liveness(R, [{badmatch,Used}|_], St) -> - check_liveness_exit(R, Used, St); -check_liveness(R, [if_end|_], St) -> - check_liveness_exit(R, ignore, St); -check_liveness(R, [{func_info,_,_,Ar}|_], St) -> - case R of - {x,X} when X < Ar -> {used,St}; - _ -> {killed,St} - end; -check_liveness(R, [{kill,R}|_], St) -> - {killed,St}; -check_liveness(R, [{kill,_}|Is], St) -> - check_liveness(R, Is, St); -check_liveness(R, [{bs_init,_,_,none,Ss,Dst}|Is], St) -> - case member(R, Ss) of - true -> - {used,St}; - false -> - if - R =:= Dst -> {killed,St}; - true -> check_liveness(R, Is, St) - end - end; -check_liveness(R, [{bs_init,_,_,Live,Ss,Dst}|Is], St) -> - case R of - {x,X} -> - case member(R, Ss) of - true -> - {used,St}; - false -> - if - X < Live -> - not_used(check_liveness(R, Is, St)); - true -> - {killed,St} - end - end; - {y,_} -> - case member(R, Ss) of - true -> {used,St}; - false -> - %% If the exception is taken, the stack may - %% be scanned. Therefore the register is not - %% guaranteed to be killed. - if - R =:= Dst -> {not_used,St}; - true -> not_used(check_liveness(R, Is, St)) - end - end - end; -check_liveness(R, [{deallocate,_}|Is], St) -> - case R of - {y,_} -> {killed,St}; - _ -> check_liveness(R, Is, St) - end; -check_liveness({x,_}=R, [return|_], St) -> - case R of - {x,0} -> {used,St}; - {x,_} -> {killed,St} - end; -check_liveness(R, [{call,Live,_}|Is], St) -> - case R of - {x,X} when X < Live -> {used,St}; - {x,_} -> {killed,St}; - {y,_} -> not_used(check_liveness(R, Is, St)) - end; -check_liveness(R, [{call_ext,Live,_}=I|Is], St) -> - case R of - {x,X} when X < Live -> - {used,St}; - {x,_} -> - {killed,St}; - {y,_} -> - case beam_jump:is_exit_instruction(I) of - false -> - not_used(check_liveness(R, Is, St)); - true -> - %% We must make sure we don't check beyond this - %% instruction or we will fall through into random - %% unrelated code and get stuck in a loop. - {exit_not_used,St} - end - end; -check_liveness(R, [{call_fun,Live}|Is], St) -> - case R of - {x,X} when X =< Live -> {used,St}; - {x,_} -> {killed,St}; - {y,_} -> not_used(check_liveness(R, Is, St)) - end; -check_liveness(R, [{apply,Args}|Is], St) -> - case R of - {x,X} when X < Args+2 -> {used,St}; - {x,_} -> {killed,St}; - {y,_} -> not_used(check_liveness(R, Is, St)) - end; -check_liveness(R, [{bif,Op,Fail,Ss,D}|Is], St) -> - Set = {set,[D],Ss,{bif,Op,Fail}}, - check_liveness(R, [{block,[Set]}|Is], St); -check_liveness(R, [{gc_bif,Op,{f,Fail},Live,Ss,D}|Is], St) -> - Set = {set,[D],Ss,{alloc,Live,{gc_bif,Op,Fail}}}, - check_liveness(R, [{block,[Set]}|Is], St); -check_liveness(R, [{bs_put,{f,0},_,Ss}|Is], St) -> - case member(R, Ss) of - true -> {used,St}; - false -> check_liveness(R, Is, St) - end; -check_liveness(R, [{bs_restore2,S,_}|Is], St) -> - case R of - S -> {used,St}; - _ -> check_liveness(R, Is, St) - end; -check_liveness(R, [{bs_save2,S,_}|Is], St) -> - case R of - S -> {used,St}; - _ -> check_liveness(R, Is, St) - end; -check_liveness(R, [{move,S,D}|Is], St) -> - case R of - S -> {used,St}; - D -> {killed,St}; - _ -> check_liveness(R, Is, St) - end; -check_liveness(R, [{make_fun2,_,_,_,NumFree}|Is], St) -> - case R of - {x,X} when X < NumFree -> {used,St}; - {x,_} -> {killed,St}; - {y,_} -> not_used(check_liveness(R, Is, St)) - end; -check_liveness(R, [{'catch'=Op,Y,Fail}|Is], St) -> - Set = {set,[Y],[],{try_catch,Op,Fail}}, - check_liveness(R, [{block,[Set]}|Is], St); -check_liveness(R, [{'try'=Op,Y,Fail}|Is], St) -> - Set = {set,[Y],[],{try_catch,Op,Fail}}, - check_liveness(R, [{block,[Set]}|Is], St); -check_liveness(R, [{try_end,Y}|Is], St) -> - case R of - Y -> - {killed,St}; - {y,_} -> - %% y registers will be used if an exception occurs and - %% control transfers to the label given in the previous - %% try/2 instruction. - {used,St}; - _ -> - check_liveness(R, Is, St) - end; -check_liveness(R, [{catch_end,Y}|Is], St) -> - case R of - Y -> {killed,St}; - _ -> check_liveness(R, Is, St) - end; -check_liveness(R, [{get_tuple_element,S,_,D}|Is], St) -> - case R of - S -> {used,St}; - D -> {killed,St}; - _ -> check_liveness(R, Is, St) - end; -check_liveness(R, [{bs_context_to_binary,S}|Is], St) -> - case R of - S -> {used,St}; - _ -> check_liveness(R, Is, St) - end; -check_liveness(R, [{loop_rec,{f,_},{x,0}}|_], St) -> - case R of - {x,_} -> - {killed,St}; - _ -> - %% y register. Rarely happens. Be very conversative and - %% assume it's used. - {used,St} - end; -check_liveness(R, [{loop_rec_end,{f,Fail}}|_], St) -> - check_liveness_at(R, Fail, St); -check_liveness(R, [{line,_}|Is], St) -> - check_liveness(R, Is, St); -check_liveness(R, [{get_map_elements,{f,Fail},S,{list,L}}|Is], St0) -> - {Ss,Ds} = split_even(L), - case member(R, [S|Ss]) of - true -> - {used,St0}; - false -> - case check_liveness_at(R, Fail, St0) of - {killed,St}=Killed -> - case member(R, Ds) of - true -> Killed; - false -> check_liveness(R, Is, St) - end; - Other -> - Other - end - end; -check_liveness(R, [{put_map,F,Op,S,D,Live,{list,Puts}}|Is], St) -> - Set = {set,[D],[S|Puts],{alloc,Live,{put_map,Op,F}}}, - check_liveness(R, [{block,[Set]}||Is], St); -check_liveness(R, [{put_tuple,Ar,D}|Is], St) -> - Set = {set,[D],[],{put_tuple,Ar}}, - check_liveness(R, [{block,[Set]}||Is], St); -check_liveness(R, [{put_list,S1,S2,D}|Is], St) -> - Set = {set,[D],[S1,S2],put_list}, - check_liveness(R, [{block,[Set]}||Is], St); -check_liveness(R, [{test_heap,N,Live}|Is], St) -> - I = {block,[{set,[],[],{alloc,Live,{nozero,nostack,N,[]}}}]}, - check_liveness(R, [I|Is], St); -check_liveness(R, [{allocate_zero,N,Live}|Is], St) -> - I = {block,[{set,[],[],{alloc,Live,{zero,N,0,[]}}}]}, - check_liveness(R, [I|Is], St); -check_liveness(R, [{get_hd,S,D}|Is], St) -> - I = {block,[{set,[D],[S],get_hd}]}, - check_liveness(R, [I|Is], St); -check_liveness(R, [{get_tl,S,D}|Is], St) -> - I = {block,[{set,[D],[S],get_tl}]}, - check_liveness(R, [I|Is], St); -check_liveness(R, [remove_message|Is], St) -> - check_liveness(R, Is, St); -check_liveness({x,X}, [build_stacktrace|_], St) when X > 0 -> - {killed,St}; -check_liveness(R, [{recv_mark,_}|Is], St) -> - check_liveness(R, Is, St); -check_liveness(R, [{recv_set,_}|Is], St) -> - check_liveness(R, Is, St); -check_liveness(R, [{'%',_}|Is], St) -> - check_liveness(R, Is, St); -check_liveness(_R, Is, St) when is_list(Is) -> - %% Not implemented. Conservatively assume that the register is used. - {used,St}. - -check_liveness_everywhere(R, Lbls, St0) -> - check_liveness_everywhere_1(R, Lbls, killed, St0). - -check_liveness_everywhere_1(R, [{f,Lbl}|T], Res0, St0) -> - {Res1,St} = check_liveness_at(R, Lbl, St0), - Res = case Res1 of - killed -> Res0; - _ -> Res1 - end, - case Res of - used -> {used,St}; - _ -> check_liveness_everywhere_1(R, T, Res, St) - end; -check_liveness_everywhere_1(R, [_|T], Res, St) -> - check_liveness_everywhere_1(R, T, Res, St); -check_liveness_everywhere_1(_, [], Res, St) -> - {Res,St}. - -check_liveness_at(R, Lbl, #live{lbl=Ll,res=ResMemorized}=St0) -> - case gb_trees:lookup(Lbl, ResMemorized) of - {value,Res} -> - {Res,St0}; - none -> - {Res,St} = case gb_trees:lookup(Lbl, Ll) of - {value,Is} -> check_liveness(R, Is, St0); - none -> {used,St0} - end, - {Res,St#live{res=gb_trees:insert(Lbl, Res, St#live.res)}} - end. - -not_used({used,_}=Res) -> Res; -not_used({_,St}) -> {not_used,St}. - -check_liveness_exit(R, R, St) -> {used,St}; -check_liveness_exit({x,_}, _, St) -> {killed,St}; -check_liveness_exit({y,_}, _, St) -> {exit_not_used,St}. - -%% check_liveness_block(Reg, [Instruction], State) -> -%% {killed | not_used | used | alloc_used | transparent,State'} -%% Finds out how Reg is used in the instruction sequence inside a block. -%% Returns one of: -%% killed - Reg is assigned a new value or killed by an -%% allocation instruction -%% not_used - The value is not used, but the register is referenced -%% e.g. by an allocation instruction -%% transparent - Reg is neither used nor killed -%% alloc_used - Used only in an allocate instruction -%% used - Reg is explicitly used by an instruction -%% -%% Annotations are not allowed. -%% -%% (Unknown instructions will cause an exception.) - -check_liveness_block({x,X}=R, [{set,Ds,Ss,{alloc,Live,Op}}|Is], St0) -> - if - X >= Live -> - {killed,St0}; - true -> - case check_liveness_block_1(R, Ss, Ds, Op, Is, St0) of - {transparent,St} -> {alloc_used,St}; - {_,_}=Res -> not_used(Res) - end - end; -check_liveness_block({y,_}=R, [{set,Ds,Ss,{alloc,_Live,Op}}|Is], St0) -> - case check_liveness_block_1(R, Ss, Ds, Op, Is, St0) of - {transparent,St} -> {alloc_used,St}; - {_,_}=Res -> not_used(Res) - end; -check_liveness_block({y,_}=R, [{set,Ds,Ss,{try_catch,_,Op}}|Is], St0) -> - case Ds of - [R] -> - {killed,St0}; - _ -> - case check_liveness_block_1(R, Ss, Ds, Op, Is, St0) of - {exit_not_used,St} -> - {used,St}; - {transparent,St} -> - %% Conservatively assumed that it is used. - {used,St}; - {_,_}=Res -> - Res - end - end; -check_liveness_block(R, [{set,Ds,Ss,Op}|Is], St) -> - check_liveness_block_1(R, Ss, Ds, Op, Is, St); -check_liveness_block(_, [], St) -> {transparent,St}. - -check_liveness_block_1(R, Ss, Ds, Op, Is, St0) -> - case member(R, Ss) of - true -> - {used,St0}; - false -> - case check_liveness_block_2(R, Op, Ss, St0) of - {killed,St} -> - case member(R, Ds) of - true -> {killed,St}; - false -> check_liveness_block(R, Is, St) - end; - {exit_not_used,St} -> - case member(R, Ds) of - true -> {exit_not_used,St}; - false -> check_liveness_block(R, Is, St) - end; - {not_used,St} -> - not_used(case member(R, Ds) of - true -> {killed,St}; - false -> check_liveness_block(R, Is, St) - end); - {used,St} -> - {used,St} - end - end. - -check_liveness_block_2(R, {gc_bif,Op,{f,Lbl}}, Ss, St) -> - check_liveness_block_3(R, Lbl, {Op,length(Ss)}, St); -check_liveness_block_2(R, {bif,Op,{f,Lbl}}, Ss, St) -> - Arity = length(Ss), - case erl_internal:comp_op(Op, Arity) orelse - erl_internal:new_type_test(Op, Arity) of - true -> - {killed,St}; - false -> - check_liveness_block_3(R, Lbl, {Op,length(Ss)}, St) - end; -check_liveness_block_2(R, {put_map,_Op,{f,Lbl}}, _Ss, St) -> - check_liveness_block_3(R, Lbl, {unsafe,0}, St); -check_liveness_block_2(_, _, _, St) -> - {killed,St}. - -check_liveness_block_3({x,_}, 0, _FA, St) -> - {killed,St}; -check_liveness_block_3({y,_}, 0, {F,A}, St) -> - %% If the exception is thrown, the stack may be scanned, - %% thus implicitly using the y register. - case erl_bifs:is_safe(erlang, F, A) of - true -> {killed,St}; - false -> {used,St} - end; -check_liveness_block_3(R, Lbl, _FA, St0) -> - check_liveness_at(R, Lbl, St0). - -index_labels_1([{label,Lbl}|Is0], Acc) -> - Is = drop_labels(Is0), - index_labels_1(Is0, [{Lbl,Is}|Acc]); -index_labels_1([_|Is], Acc) -> - index_labels_1(Is, Acc); -index_labels_1([], Acc) -> gb_trees:from_orddict(sort(Acc)). - -drop_labels([{label,_}|Is]) -> drop_labels(Is); -drop_labels(Is) -> Is. - - replace_labels_1([{test,Test,{f,Lbl},Ops}|Is], Acc, D, Fb) -> replace_labels_1(Is, [{test,Test,{f,label(Lbl, D, Fb)},Ops}|Acc], D, Fb); replace_labels_1([{test,Test,{f,Lbl},Live,Ops,Dst}|Is], Acc, D, Fb) -> @@ -734,21 +144,6 @@ label(Old, D, Fb) -> _ -> Fb(Old) end. -%% Help function for combine_heap_needs. - -combine_alloc_lists(Al0) -> - Al1 = flatmap(fun(Words) when is_integer(Words) -> - [{words,Words}]; - ({alloc,List}) -> - List - end, Al0), - Al2 = sofs:relation(Al1), - Al3 = sofs:relation_to_family(Al2), - Al4 = sofs:to_external(Al3), - [{Tag,lists:sum(L)} || {Tag,L} <- Al4]. - -%% live_opt/4. - split_even([], Ss, Ds) -> {reverse(Ss),reverse(Ds)}; split_even([S,D|Rs], Ss, Ds) -> diff --git a/lib/compiler/src/beam_validator.erl b/lib/compiler/src/beam_validator.erl index 953aced66e..b56d53d4ce 100644 --- a/lib/compiler/src/beam_validator.erl +++ b/lib/compiler/src/beam_validator.erl @@ -26,6 +26,7 @@ %% Interface for compiler. -export([module/2, format_error/1]). +-export([type_anno/1, type_anno/2, type_anno/3]). -import(lists, [any/2,dropwhile/2,foldl/3,map/2,foreach/2,reverse/1]). @@ -44,6 +45,33 @@ module({Mod,Exp,Attr,Fs,Lc}=Code, _Opts) {error,[{atom_to_list(Mod),Es}]} end. +%% Provides a stable interface for type annotations, used by certain passes to +%% indicate that we can safely assume that a register has a given type. +-spec type_anno(term()) -> term(). +type_anno(atom) -> {atom,[]}; +type_anno(bool) -> bool; +type_anno({binary,_}) -> term; +type_anno(cons) -> cons; +type_anno(float) -> {float,[]}; +type_anno(integer) -> {integer,[]}; +type_anno(list) -> list; +type_anno(map) -> map; +type_anno(match_context) -> match_context; +type_anno(number) -> number; +type_anno(nil) -> nil. + +-spec type_anno(term(), term()) -> term(). +type_anno(atom, Value) -> {atom, Value}; +type_anno(float, Value) -> {float, Value}; +type_anno(integer, Value) -> {integer, Value}. + +-spec type_anno(term(), term(), term()) -> term(). +type_anno(tuple, Size, Exact) when is_integer(Size) -> + case Exact of + true -> {tuple, Size}; + false -> {tuple, [Size]} + end. + -spec format_error(term()) -> iolist(). format_error({{_M,F,A},{I,Off,limit}}) -> @@ -90,34 +118,9 @@ format_error(Error) -> %% format as used in the compiler and in .S files. validate(Module, Fs) -> - Ft = index_bs_start_match(Fs, []), + Ft = index_parameter_types(Fs, []), validate_0(Module, Fs, Ft). -index_bs_start_match([{function,_,_,Entry,Code0}|Fs], Acc0) -> - Code = dropwhile(fun({label,L}) when L =:= Entry -> false; - (_) -> true - end, Code0), - case Code of - [{label,Entry}|Is] -> - Acc = index_bs_start_match_1(Is, Entry, Acc0), - index_bs_start_match(Fs, Acc); - _ -> - %% Something serious is wrong. Ignore it for now. - %% It will be detected and diagnosed later. - index_bs_start_match(Fs, Acc0) - end; -index_bs_start_match([], Acc) -> - gb_trees:from_orddict(lists:sort(Acc)). - -index_bs_start_match_1([{test,bs_start_match2,_,_,_,_}=I|_], Entry, Acc) -> - [{Entry,[I]}|Acc]; -index_bs_start_match_1([{test,_,{f,F},_},{bs_context_to_binary,_}|Is0], Entry, Acc) -> - [{label,F}|Is] = dropwhile(fun({label,L}) when L =:= F -> false; - (_) -> true - end, Is0), - index_bs_start_match_1(Is, Entry, Acc); -index_bs_start_match_1(_, _, Acc) -> Acc. - validate_0(_Module, [], _) -> []; validate_0(Module, [{function,Name,Ar,Entry,Code}|Fs], Ft) -> try validate_1(Code, Name, Ar, Entry, Ft) of @@ -170,6 +173,32 @@ validate_0(Module, [{function,Name,Ar,Entry,Code}|Fs], Ft) -> slots=0 :: non_neg_integer() %Number of slots }). +index_parameter_types([{function,_,_,Entry,Code0}|Fs], Acc0) -> + Code = dropwhile(fun({label,L}) when L =:= Entry -> false; + (_) -> true + end, Code0), + case Code of + [{label,Entry}|Is] -> + Acc = index_parameter_types_1(Is, Entry, Acc0), + index_parameter_types(Fs, Acc); + _ -> + %% Something serious is wrong. Ignore it for now. + %% It will be detected and diagnosed later. + index_parameter_types(Fs, Acc0) + end; +index_parameter_types([], Acc) -> + gb_trees:from_orddict(lists:sort(Acc)). + +index_parameter_types_1([{'%', {type_info, Reg, Type0}} | Is], Entry, Acc) -> + Type = case Type0 of + match_context -> #ms{}; + _ -> Type0 + end, + Key = {Entry, Reg}, + index_parameter_types_1(Is, Entry, [{Key, Type} | Acc]); +index_parameter_types_1(_, _, Acc) -> + Acc. + validate_1(Is, Name, Arity, Entry, Ft) -> validate_2(labels(Is), Name, Arity, Entry, Ft). @@ -274,11 +303,11 @@ valfun_1(_I, #vst{current=none}=Vst) -> %% the original R10B compiler thought would return. Vst; valfun_1({badmatch,Src}, Vst) -> - assert_term(Src, Vst), + assert_not_fragile(Src, Vst), verify_y_init(Vst), kill_state(Vst); valfun_1({case_end,Src}, Vst) -> - assert_term(Src, Vst), + assert_not_fragile(Src, Vst), verify_y_init(Vst), kill_state(Vst); valfun_1(if_end, Vst) -> @@ -286,36 +315,21 @@ valfun_1(if_end, Vst) -> kill_state(Vst); valfun_1({try_case_end,Src}, Vst) -> verify_y_init(Vst), - assert_term(Src, Vst), + assert_not_fragile(Src, Vst), kill_state(Vst); %% Instructions that cannot cause exceptions -valfun_1({bs_context_to_binary,Ctx}, #vst{current=#st{x=Xs}}=Vst) -> - case Ctx of - {Tag,X} when Tag =:= x; Tag =:= y -> - Type = case gb_trees:lookup(X, Xs) of - {value,#ms{}} -> term; - _ -> get_term_type(Ctx, Vst) - end, - set_type_reg(Type, Ctx, Vst); - _ -> - error({bad_source,Ctx}) - end; +valfun_1({bs_get_tail,Ctx,Dst,Live}, Vst0) -> + bsm_validate_context(Ctx, Vst0), + verify_live(Live, Vst0), + verify_y_init(Vst0), + Vst = prune_x_regs(Live, Vst0), + extract_term(binary, [Ctx], Dst, Vst, Vst0); valfun_1(bs_init_writable=I, Vst) -> call(I, 1, Vst); valfun_1(build_stacktrace=I, Vst) -> call(I, 1, Vst); -valfun_1({move,{y,_}=Src,{y,_}=Dst}, Vst) -> - %% The stack trimming optimization may generate a move from an initialized - %% but unassigned Y register to another Y register. - case get_term_type_1(Src, Vst) of - {catchtag,_} -> error({catchtag,Src}); - {trytag,_} -> error({trytag,Src}); - Type -> set_type_reg(Type, Dst, Vst) - end; -valfun_1({move,Src,Dst}, Vst0) -> - Type = get_move_term_type(Src, Vst0), - Vst = set_type_reg(Type, Dst, Vst0), - set_alias(Src, Dst, Vst); +valfun_1({move,Src,Dst}, Vst) -> + assign(Src, Dst, Vst); valfun_1({fmove,Src,{fr,_}=Dst}, Vst) -> assert_type(float, Src, Vst), set_freg(Dst, Vst); @@ -323,7 +337,7 @@ valfun_1({fmove,{fr,_}=Src,Dst}, Vst0) -> assert_freg_set(Src, Vst0), assert_fls(checked, Vst0), Vst = eat_heap_float(Vst0), - set_type_reg({float,[]}, Dst, Vst); + create_term({float,[]}, Dst, Vst); valfun_1({kill,{y,_}=Reg}, Vst) -> set_type_y(initialized, Reg, Vst); valfun_1({init,{y,_}=Reg}, Vst) -> @@ -345,24 +359,24 @@ valfun_1({bif,Op,{f,_},Src,Dst}=I, Vst) -> end; %% Put instructions. valfun_1({put_list,A,B,Dst}, Vst0) -> - assert_term(A, Vst0), - assert_term(B, Vst0), + assert_not_fragile(A, Vst0), + assert_not_fragile(B, Vst0), Vst = eat_heap(2, Vst0), - set_type_reg(cons, Dst, Vst); + create_term(cons, Dst, Vst); valfun_1({put_tuple2,Dst,{list,Elements}}, Vst0) -> - _ = [assert_term(El, Vst0) || El <- Elements], + _ = [assert_not_fragile(El, Vst0) || El <- Elements], Size = length(Elements), Vst = eat_heap(Size+1, Vst0), Type = {tuple,Size}, - set_type_reg(Type, Dst, Vst); + create_term(Type, Dst, Vst); valfun_1({put_tuple,Sz,Dst}, Vst0) when is_integer(Sz) -> Vst1 = eat_heap(1, Vst0), - Vst = set_type_reg(tuple_in_progress, Dst, Vst1), + Vst = create_term(tuple_in_progress, Dst, Vst1), #vst{current=St0} = Vst, St = St0#st{puts_left={Sz,{Dst,{tuple,Sz}}}}, Vst#vst{current=St}; valfun_1({put,Src}, Vst0) -> - assert_term(Src, Vst0), + assert_not_fragile(Src, Vst0), Vst = eat_heap(1, Vst0), #vst{current=St0} = Vst, case St0 of @@ -370,7 +384,7 @@ valfun_1({put,Src}, Vst0) -> error(not_building_a_tuple); #st{puts_left={1,{Dst,Type}}} -> St = St0#st{puts_left=none}, - set_type_reg(Type, Dst, Vst#vst{current=St}); + create_term(Type, Dst, Vst#vst{current=St}); #st{puts_left={PutsLeft,Info}} when is_integer(PutsLeft) -> St = St0#st{puts_left={PutsLeft-1,Info}}, Vst#vst{current=St} @@ -385,6 +399,13 @@ valfun_1(remove_message, Vst) -> %% The message term is no longer fragile. It can be used %% without restrictions. remove_fragility(Vst); +valfun_1({'%', {type_info, Reg, match_context}}, Vst) -> + update_type(fun meet/2, #ms{}, Reg, Vst); +valfun_1({'%', {type_info, Reg, Type}}, Vst) -> + %% Explicit type information inserted by optimization passes to indicate + %% that Reg has a certain type, so that we can accept cross-function type + %% optimizations. + update_type(fun meet/2, Type, Reg, Vst); valfun_1({'%',_}, Vst) -> Vst; valfun_1({line,_}, Vst) -> @@ -459,18 +480,22 @@ valfun_1({try_case,Reg}, #vst{current=#st{ct=[Fail|Fails]}}=Vst0) -> error({bad_type,Type}) end; valfun_1({get_list,Src,D1,D2}, Vst0) -> + assert_not_literal(Src), assert_type(cons, Src, Vst0), - Vst = set_type_reg(term, Src, D1, Vst0), - set_type_reg(term, Src, D2, Vst); + Vst = extract_term(term, [Src], D1, Vst0), + extract_term(term, [Src], D2, Vst); valfun_1({get_hd,Src,Dst}, Vst) -> + assert_not_literal(Src), assert_type(cons, Src, Vst), - set_type_reg(term, Src, Dst, Vst); + extract_term(term, [Src], Dst, Vst); valfun_1({get_tl,Src,Dst}, Vst) -> + assert_not_literal(Src), assert_type(cons, Src, Vst), - set_type_reg(term, Src, Dst, Vst); + extract_term(term, [Src], Dst, Vst); valfun_1({get_tuple_element,Src,I,Dst}, Vst) -> + assert_not_literal(Src), assert_type({tuple_element,I+1}, Src, Vst), - set_type_reg(term, Src, Dst, Vst); + extract_term(term, [Src], Dst, Vst); valfun_1({jump,{f,Lbl}}, Vst) -> kill_state(branch_state(Lbl, Vst)); valfun_1(I, Vst) -> @@ -569,54 +594,59 @@ valfun_4({make_fun2,_,_,_,Live}, Vst) -> call(make_fun, Live, Vst); %% Other BIFs valfun_4({bif,tuple_size,{f,Fail},[Tuple],Dst}=I, Vst0) -> - TupleType0 = get_term_type(Tuple, Vst0), Vst1 = branch_state(Fail, Vst0), - TupleType = upgrade_tuple_type({tuple,[0]}, TupleType0), - Vst = set_type(TupleType, Tuple, Vst1), + Vst = update_type(fun meet/2, {tuple,[0]}, Tuple, Vst1), set_type_reg_expr({integer,[]}, I, Dst, Vst); valfun_4({bif,element,{f,Fail},[Pos,Tuple],Dst}, Vst0) -> - TupleType0 = get_term_type(Tuple, Vst0), - PosType = get_term_type(Pos, Vst0), + PosType = get_durable_term_type(Pos, Vst0), Vst1 = branch_state(Fail, Vst0), - TupleType = upgrade_tuple_type({tuple,[get_tuple_size(PosType)]}, TupleType0), - Vst = set_aliased_type(TupleType, Tuple, Vst1), - set_type_reg(term, Tuple, Dst, Vst); + Type = {tuple,[get_tuple_size(PosType)]}, + Vst = update_type(fun meet/2, Type, Tuple, Vst1), + extract_term(term, [Tuple], Dst, Vst); valfun_4({bif,raise,{f,0},Src,_Dst}, Vst) -> validate_src(Src, Vst), kill_state(Vst); valfun_4(raw_raise=I, Vst) -> call(I, 3, Vst); -valfun_4({bif,map_get,{f,Fail},[_Key,Map]=Src,Dst}, Vst0) -> - validate_src(Src, Vst0), +valfun_4({bif,map_get,{f,Fail},[_Key,Map]=Ss,Dst}, Vst0) -> + validate_src(Ss, Vst0), + Vst1 = branch_state(Fail, Vst0), + Vst = update_type(fun meet/2, map, Map, Vst1), + extract_term(term, Ss, Dst, Vst); +valfun_4({bif,is_map_key,{f,Fail},[_Key,Map]=Ss,Dst}, Vst0) -> + validate_src(Ss, Vst0), Vst1 = branch_state(Fail, Vst0), - Vst = set_type(map, Map, Vst1), - Type = propagate_fragility(term, Src, Vst), - set_type_reg(Type, Dst, Vst); -valfun_4({bif,is_map_key,{f,Fail},[_Key,Map]=Src,Dst}, Vst0) -> - validate_src(Src, Vst0), + Vst = update_type(fun meet/2, map, Map, Vst1), + extract_term(bool, Ss, Dst, Vst); +valfun_4({bif,Op,{f,Fail},[Cons]=Ss,Dst}, Vst0) + when Op =:= hd; Op =:= tl -> + validate_src(Ss, Vst0), Vst1 = branch_state(Fail, Vst0), - Vst = set_type(map, Map, Vst1), - Type = propagate_fragility(bool, Src, Vst), - set_type_reg(Type, Dst, Vst); -valfun_4({bif,Op,{f,Fail},Src,Dst}, Vst0) -> - validate_src(Src, Vst0), + Vst = update_type(fun meet/2, cons, Cons, Vst1), + Type = bif_type(Op, Ss, Vst), + extract_term(Type, Ss, Dst, Vst); +valfun_4({bif,Op,{f,Fail},Ss,Dst}, Vst0) -> + validate_src(Ss, Vst0), Vst = branch_state(Fail, Vst0), - Type0 = bif_type(Op, Src, Vst), - Type = propagate_fragility(Type0, Src, Vst), - set_type_reg(Type, Dst, Vst); -valfun_4({gc_bif,Op,{f,Fail},Live,Src,Dst}, #vst{current=St0}=Vst0) -> + Type = bif_type(Op, Ss, Vst), + extract_term(Type, Ss, Dst, Vst); +valfun_4({gc_bif,Op,{f,Fail},Live,Ss,Dst}, #vst{current=St0}=Vst0) -> + validate_src(Ss, Vst0), verify_live(Live, Vst0), verify_y_init(Vst0), St = kill_heap_allocation(St0), Vst1 = Vst0#vst{current=St}, Vst2 = branch_state(Fail, Vst1), - Vst = prune_x_regs(Live, Vst2), - validate_src(Src, Vst), - Type0 = bif_type(Op, Src, Vst), - Type = propagate_fragility(Type0, Src, Vst), - set_type_reg(Type, Dst, Vst); + Vst3 = case Op of + length -> update_type(fun meet/2, list, hd(Ss), Vst2); + map_size -> update_type(fun meet/2, map, hd(Ss), Vst2); + _ -> Vst2 + end, + Type = bif_type(Op, Ss, Vst3), + Vst = prune_x_regs(Live, Vst3), + extract_term(Type, Ss, Dst, Vst, Vst0); valfun_4(return, #vst{current=#st{numy=none}}=Vst) -> - assert_term({x,0}, Vst), + assert_not_fragile({x,0}, Vst), kill_state(Vst); valfun_4(return, #vst{current=#st{numy=NumY}}) -> error({stack_frame,NumY}); @@ -626,7 +656,7 @@ valfun_4({loop_rec,{f,Fail},Dst}, Vst0) -> %% remove_message/0 is executed. If control transfers %% to the loop_rec_end/1 instruction, no part of %% this term must be stored in a Y register. - set_type_reg({fragile,term}, Dst, Vst); + create_term({fragile,term}, Dst, Vst); valfun_4({wait,_}, Vst) -> verify_y_init(Vst), kill_state(Vst); @@ -642,50 +672,27 @@ valfun_4(timeout, #vst{current=St}=Vst) -> valfun_4(send, Vst) -> call(send, 2, Vst); valfun_4({set_tuple_element,Src,Tuple,I}, Vst) -> - assert_term(Src, Vst), + assert_not_fragile(Src, Vst), assert_type({tuple_element,I+1}, Tuple, Vst), Vst; %% Match instructions. valfun_4({select_val,Src,{f,Fail},{list,Choices}}, Vst0) -> assert_term(Src, Vst0), + assert_choices(Choices), Vst = branch_state(Fail, Vst0), kill_state(select_val_branches(Src, Choices, Vst)); valfun_4({select_tuple_arity,Tuple,{f,Fail},{list,Choices}}, Vst) -> assert_type(tuple, Tuple, Vst), - TupleType = case get_term_type(Tuple, Vst) of - {fragile,TupleType0} -> TupleType0; - TupleType0 -> TupleType0 - end, + assert_arities(Choices), + TupleType = get_durable_term_type(Tuple, Vst), kill_state(branch_arities(Choices, Tuple, TupleType, branch_state(Fail, Vst))); %% New bit syntax matching instructions. -valfun_4({test,bs_start_match2,{f,Fail},Live,[Ctx,NeedSlots],Ctx}, Vst0) -> - %% If source and destination registers are the same, match state - %% is OK as input. - CtxType = get_move_term_type(Ctx, Vst0), - verify_live(Live, Vst0), - verify_y_init(Vst0), - Vst1 = prune_x_regs(Live, Vst0), - BranchVst = case CtxType of - #ms{} -> - %% The failure branch will never be taken when Ctx - %% is a match context. Therefore, the type for Ctx - %% at the failure label must not be match_context - %% (or we could reject legal code). - set_type_reg(term, Ctx, Vst1); - _ -> - Vst1 - end, - Vst = branch_state(Fail, BranchVst), - set_type_reg(bsm_match_state(NeedSlots), Ctx, Vst); -valfun_4({test,bs_start_match2,{f,Fail},Live,[Src,Slots],Dst}, Vst0) -> - assert_term(Src, Vst0), - verify_live(Live, Vst0), - verify_y_init(Vst0), - Vst1 = prune_x_regs(Live, Vst0), - Vst = branch_state(Fail, Vst1), - set_type_reg(bsm_match_state(Slots), Src, Dst, Vst); +valfun_4({test,bs_start_match3,{f,Fail},Live,[Src],Dst}, Vst) -> + validate_bs_start_match(Fail, Live, bsm_match_state(), Src, Dst, Vst); +valfun_4({test,bs_start_match2,{f,Fail},Live,[Src,Slots],Dst}, Vst) -> + validate_bs_start_match(Fail, Live, bsm_match_state(Slots), Src, Dst, Vst); valfun_4({test,bs_match_string,{f,Fail},[Ctx,_,_]}, Vst) -> bsm_validate_context(Ctx, Vst), branch_state(Fail, Vst); @@ -722,69 +729,81 @@ valfun_4({bs_save2,Ctx,SavePoint}, Vst) -> bsm_save(Ctx, SavePoint, Vst); valfun_4({bs_restore2,Ctx,SavePoint}, Vst) -> bsm_restore(Ctx, SavePoint, Vst); +valfun_4({bs_get_position, Ctx, Dst, Live}, Vst0) -> + bsm_validate_context(Ctx, Vst0), + verify_live(Live, Vst0), + verify_y_init(Vst0), + Vst = prune_x_regs(Live, Vst0), + create_term(bs_position, Dst, Vst); +valfun_4({bs_set_position, Ctx, Pos}, Vst) -> + bsm_validate_context(Ctx, Vst), + assert_type(bs_position, Pos, Vst), + Vst; %% Other test instructions. -valfun_4({test,is_float,{f,Lbl},[Float]}, Vst) -> - assert_term(Float, Vst), - set_type({float,[]}, Float, branch_state(Lbl, Vst)); -valfun_4({test,is_tuple,{f,Lbl},[Tuple]}, Vst) -> - Type0 = get_term_type(Tuple, Vst), - Type = upgrade_tuple_type({tuple,[0]}, Type0), - set_aliased_type(Type, Tuple, branch_state(Lbl, Vst)); -valfun_4({test,is_nonempty_list,{f,Lbl},[Cons]}, Vst) -> - assert_term(Cons, Vst), - Type = cons, - set_aliased_type(Type, Cons, branch_state(Lbl, Vst)); +valfun_4({test,is_atom,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, {atom,[]}, Src, Vst); +valfun_4({test,is_boolean,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, bool, Src, Vst); +valfun_4({test,is_float,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, {float,[]}, Src, Vst); +valfun_4({test,is_tuple,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, {tuple,[0]}, Src, Vst); +valfun_4({test,is_integer,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, {integer,[]}, Src, Vst); +valfun_4({test,is_nonempty_list,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, cons, Src, Vst); +valfun_4({test,is_list,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, list, Src, Vst); +valfun_4({test,is_nil,{f,Lbl},[Src]}, Vst) -> + type_test(Lbl, nil, Src, Vst); +valfun_4({test,is_map,{f,Lbl},[Src]}, Vst) -> + case Src of + {Tag,_} when Tag =:= x; Tag =:= y -> + type_test(Lbl, map, Src, Vst); + {literal,Map} when is_map(Map) -> + Vst; + _ -> + assert_term(Src, Vst), + kill_state(Vst) + end; valfun_4({test,test_arity,{f,Lbl},[Tuple,Sz]}, Vst) when is_integer(Sz) -> assert_type(tuple, Tuple, Vst), - Type = {tuple,Sz}, - set_aliased_type(Type, Tuple, branch_state(Lbl, Vst)); + update_type(fun meet/2, {tuple,Sz}, Tuple, branch_state(Lbl, Vst)); valfun_4({test,is_tagged_tuple,{f,Lbl},[Src,Sz,_Atom]}, Vst) -> - validate_src([Src], Vst), - Type = {tuple,Sz}, - set_aliased_type(Type, Src, branch_state(Lbl, Vst)); + assert_term(Src, Vst), + update_type(fun meet/2, {tuple,Sz}, Src, branch_state(Lbl, Vst)); valfun_4({test,has_map_fields,{f,Lbl},Src,{list,List}}, Vst) -> assert_type(map, Src, Vst), assert_unique_map_keys(List), branch_state(Lbl, Vst); -valfun_4({test,is_map,{f,Lbl},[Src]}, Vst0) -> - Vst = branch_state(Lbl, Vst0), - case Src of - {Tag,_} when Tag =:= x; Tag =:= y -> - Type = map, - set_aliased_type(Type, Src, Vst); - {literal,Map} when is_map(Map) -> - Vst0; - _ -> - kill_state(Vst0) - end; valfun_4({test,is_eq_exact,{f,Lbl},[Src,Val]=Ss}, Vst0) -> validate_src(Ss, Vst0), Infer = infer_types(Src, Vst0), Vst1 = Infer(Val, Vst0), - Vst = branch_state(Lbl, Vst1), - case Val of - {literal,Tuple} when is_tuple(Tuple) -> - Type0 = get_term_type(Val, Vst), - Type = upgrade_tuple_type({tuple,tuple_size(Tuple)}, - Type0), - set_aliased_type(Type, Src, Vst); - _ -> - Vst - end; + Vst2 = update_ne_types(Src, Val, Vst1), + Vst3 = branch_state(Lbl, Vst2), + Vst = Vst3#vst{current=Vst1#vst.current}, + update_eq_types(Src, Val, Vst); +valfun_4({test,is_ne_exact,{f,Lbl},[Src,Val]=Ss}, Vst0) -> + validate_src(Ss, Vst0), + Vst1 = update_eq_types(Src, Val, Vst0), + Vst2 = branch_state(Lbl, Vst1), + Vst = Vst2#vst{current=Vst0#vst.current}, + update_ne_types(Src, Val, Vst); valfun_4({test,_Op,{f,Lbl},Src}, Vst) -> validate_src(Src, Vst), branch_state(Lbl, Vst); valfun_4({bs_add,{f,Fail},[A,B,_],Dst}, Vst) -> - assert_term(A, Vst), - assert_term(B, Vst), - set_type_reg({integer,[]}, Dst, branch_state(Fail, Vst)); + assert_not_fragile(A, Vst), + assert_not_fragile(B, Vst), + create_term({integer,[]}, Dst, branch_state(Fail, Vst)); valfun_4({bs_utf8_size,{f,Fail},A,Dst}, Vst) -> assert_term(A, Vst), - set_type_reg({integer,[]}, Dst, branch_state(Fail, Vst)); + create_term({integer,[]}, Dst, branch_state(Fail, Vst)); valfun_4({bs_utf16_size,{f,Fail},A,Dst}, Vst) -> assert_term(A, Vst), - set_type_reg({integer,[]}, Dst, branch_state(Fail, Vst)); + create_term({integer,[]}, Dst, branch_state(Fail, Vst)); valfun_4({bs_init2,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> verify_live(Live, Vst0), verify_y_init(Vst0), @@ -792,12 +811,12 @@ valfun_4({bs_init2,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> is_integer(Sz) -> ok; true -> - assert_term(Sz, Vst0) + assert_not_fragile(Sz, Vst0) end, Vst1 = heap_alloc(Heap, Vst0), Vst2 = branch_state(Fail, Vst1), Vst = prune_x_regs(Live, Vst2), - set_type_reg(binary, Dst, Vst); + create_term(binary, Dst, Vst); valfun_4({bs_init_bits,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> verify_live(Live, Vst0), verify_y_init(Vst0), @@ -810,43 +829,43 @@ valfun_4({bs_init_bits,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> Vst1 = heap_alloc(Heap, Vst0), Vst2 = branch_state(Fail, Vst1), Vst = prune_x_regs(Live, Vst2), - set_type_reg(binary, Dst, Vst); + create_term(binary, Dst, Vst); valfun_4({bs_append,{f,Fail},Bits,Heap,Live,_Unit,Bin,_Flags,Dst}, Vst0) -> verify_live(Live, Vst0), verify_y_init(Vst0), - assert_term(Bits, Vst0), - assert_term(Bin, Vst0), + assert_not_fragile(Bits, Vst0), + assert_not_fragile(Bin, Vst0), Vst1 = heap_alloc(Heap, Vst0), Vst2 = branch_state(Fail, Vst1), Vst = prune_x_regs(Live, Vst2), - set_type_reg(binary, Dst, Vst); + create_term(binary, Dst, Vst); valfun_4({bs_private_append,{f,Fail},Bits,_Unit,Bin,_Flags,Dst}, Vst0) -> - assert_term(Bits, Vst0), - assert_term(Bin, Vst0), + assert_not_fragile(Bits, Vst0), + assert_not_fragile(Bin, Vst0), Vst = branch_state(Fail, Vst0), - set_type_reg(binary, Dst, Vst); + create_term(binary, Dst, Vst); valfun_4({bs_put_string,Sz,_}, Vst) when is_integer(Sz) -> Vst; valfun_4({bs_put_binary,{f,Fail},Sz,_,_,Src}, Vst) -> - assert_term(Sz, Vst), - assert_term(Src, Vst), + assert_not_fragile(Sz, Vst), + assert_not_fragile(Src, Vst), branch_state(Fail, Vst); valfun_4({bs_put_float,{f,Fail},Sz,_,_,Src}, Vst) -> - assert_term(Sz, Vst), - assert_term(Src, Vst), + assert_not_fragile(Sz, Vst), + assert_not_fragile(Src, Vst), branch_state(Fail, Vst); valfun_4({bs_put_integer,{f,Fail},Sz,_,_,Src}, Vst) -> - assert_term(Sz, Vst), - assert_term(Src, Vst), + assert_not_fragile(Sz, Vst), + assert_not_fragile(Src, Vst), branch_state(Fail, Vst); valfun_4({bs_put_utf8,{f,Fail},_,Src}, Vst) -> - assert_term(Src, Vst), + assert_not_fragile(Src, Vst), branch_state(Fail, Vst); valfun_4({bs_put_utf16,{f,Fail},_,Src}, Vst) -> - assert_term(Src, Vst), + assert_not_fragile(Src, Vst), branch_state(Fail, Vst); valfun_4({bs_put_utf32,{f,Fail},_,Src}, Vst) -> - assert_term(Src, Vst), + assert_not_fragile(Src, Vst), branch_state(Fail, Vst); %% Map instructions. valfun_4({put_map_assoc,{f,Fail},Src,Dst,Live,{list,List}}, Vst) -> @@ -859,10 +878,11 @@ valfun_4(_, _) -> error(unknown_instruction). verify_get_map(Fail, Src, List, Vst0) -> + assert_not_literal(Src), %OTP 22. assert_type(map, Src, Vst0), Vst1 = foldl(fun(D, Vsti) -> case is_reg_defined(D,Vsti) of - true -> set_type_reg(term,D,Vsti); + true -> create_term(term, D, Vsti); false -> Vsti end end, Vst0, extract_map_vals(List)), @@ -881,7 +901,7 @@ extract_map_keys([]) -> []. verify_get_map_pair([Src,Dst|Vs], Map, Vst0, Vsti0) -> assert_term(Src, Vst0), - Vsti = set_type_reg(term, Map, Dst, Vsti0), + Vsti = extract_term(term, [Map], Dst, Vsti0), verify_get_map_pair(Vs, Map, Vst0, Vsti); verify_get_map_pair([], _Map, _Vst0, Vst) -> Vst. @@ -889,13 +909,29 @@ verify_put_map(Fail, Src, Dst, Live, List, Vst0) -> assert_type(map, Src, Vst0), verify_live(Live, Vst0), verify_y_init(Vst0), - foreach(fun (Term) -> assert_term(Term, Vst0) end, List), + foreach(fun (Term) -> assert_not_fragile(Term, Vst0) end, List), Vst1 = heap_alloc(0, Vst0), Vst2 = branch_state(Fail, Vst1), Vst = prune_x_regs(Live, Vst2), Keys = extract_map_keys(List), assert_unique_map_keys(Keys), - set_type_reg(map, Dst, Vst). + create_term(map, Dst, Vst). + +%% +%% Common code for validating bs_start_match* instructions. +%% + +validate_bs_start_match(Fail, Live, Type, Src, Dst, Vst0) -> + verify_live(Live, Vst0), + verify_y_init(Vst0), + + %% #ms{} can represent either a match context or a term, so we have to mark + %% the source as a term if it fails, and retain the incoming type if it + %% succeeds (match context or not). + Vst1 = set_aliased_type(term, Src, Vst0), + Vst2 = prune_x_regs(Live, Vst1), + Vst3 = branch_state(Fail, Vst2), + extract_term(Type, [Src], Dst, Vst3, Vst0). %% %% Common code for validating bs_get* instructions. @@ -906,7 +942,7 @@ validate_bs_get(Fail, Ctx, Live, Type, Dst, Vst0) -> verify_y_init(Vst0), Vst1 = prune_x_regs(Live, Vst0), Vst = branch_state(Fail, Vst1), - set_type_reg(Type, Dst, Vst). + create_term(Type, Dst, Vst). %% %% Common code for validating bs_skip_utf* instructions. @@ -979,55 +1015,62 @@ verify_call_args(_, Live, _) -> verify_call_args_1(0, _) -> ok; verify_call_args_1(N, Vst) -> X = N - 1, - get_term_type({x,X}, Vst), + assert_not_fragile({x,X}, Vst), verify_call_args_1(X, Vst). verify_local_call(Lbl, Live, Vst) -> - case all_ms_in_x_regs(Live, Vst) of - [{R,Ctx}] -> - %% Verify that there is a suitable bs_start_match2 instruction. - verify_call_match_context(Lbl, R, Vst), - - %% Since the callee has consumed the match context, - %% there must be no additional copies in Y registers. - #ms{id=Id} = Ctx, - case ms_in_y_regs(Id, Vst) of - [] -> - ok; - [_|_]=Ys -> - error({multiple_match_contexts,[R|Ys]}) - end; - [_,_|_]=Xs0 -> - Xs = [R || {R,_} <- Xs0], - error({multiple_match_contexts,Xs}); - [] -> - ok - end. + F = fun({R, Type}) -> + verify_arg_type(Lbl, R, Type, Vst) + end, + TRegs = typed_call_regs(Live, Vst), + verify_no_ms_aliases(TRegs), + foreach(F, TRegs). -all_ms_in_x_regs(0, _Vst) -> +typed_call_regs(0, _Vst) -> []; -all_ms_in_x_regs(Live0, Vst) -> +typed_call_regs(Live0, Vst) -> Live = Live0 - 1, R = {x,Live}, - case get_move_term_type(R, Vst) of - #ms{}=M -> - [{R,M}|all_ms_in_x_regs(Live, Vst)]; - _ -> - all_ms_in_x_regs(Live, Vst) - end. + [{R, get_move_term_type(R, Vst)} | typed_call_regs(Live, Vst)]. -ms_in_y_regs(Id, #vst{current=#st{y=Ys0}}) -> - Ys = gb_trees:to_list(Ys0), - [{y,Y} || {Y,#ms{id=OtherId}} <- Ys, OtherId =:= Id]. +%% Verifies that the same match context isn't present twice. +verify_no_ms_aliases(Regs) -> + CtxIds = [Id || {_, #ms{id=Id}} <- Regs], + UniqueCtxIds = ordsets:from_list(CtxIds), + if + length(UniqueCtxIds) < length(CtxIds) -> + error({multiple_match_contexts, Regs}); + length(UniqueCtxIds) =:= length(CtxIds) -> + ok + end. -verify_call_match_context(Lbl, Ctx, #vst{ft=Ft}) -> - case gb_trees:lookup(Lbl, Ft) of - none -> - error(no_bs_start_match2); - {value,[{test,bs_start_match2,_,_,[Ctx,_],Ctx}|_]} -> - ok; - {value,[{test,bs_start_match2,_,_,_,_}=I|_]} -> - error({unsuitable_bs_start_match2,I}) +%% Verifies that the given argument narrows to what the function expects. +verify_arg_type(Lbl, Reg, #ms{}, #vst{ft=Ft}) -> + %% Match contexts require explicit support, and may not be passed to a + %% function that accepts arbitrary terms. + case gb_trees:lookup({Lbl, Reg}, Ft) of + {value, #ms{}} -> ok; + _ -> error(no_bs_start_match2) + end; +verify_arg_type(Lbl, Reg, GivenType, #vst{ft=Ft}) -> + case gb_trees:lookup({Lbl, Reg}, Ft) of + {value, bool} when GivenType =:= {atom, true}; + GivenType =:= {atom, false}; + GivenType =:= {atom, []} -> + %% We don't yet support upgrading true/false to bool, so we + %% assume unknown atoms can be bools when validating calls. + ok; + {value, #ms{}} -> + %% Functions that accept match contexts also accept all other + %% terms. This will change once we support union types. + ok; + {value, RequiredType} -> + case meet(GivenType, RequiredType) of + none -> error({bad_arg_type, Reg, GivenType, RequiredType}); + _ -> ok + end; + none -> + ok end. allocate(Zero, Stk, Heap, Live, #vst{current=#st{numy=none}}=Vst0) -> @@ -1088,6 +1131,29 @@ prune_x_regs(Live, #vst{current=St0}=Vst) St = St0#st{x=gb_trees:from_orddict(Xs),defs=Defs,aliases=Aliases}, Vst#vst{current=St}. +%% All choices in a select_val list must be integers, floats, or atoms. +%% All must be of the same type. +assert_choices([{Tag,_},{f,_}|T]) -> + if + Tag =:= atom; Tag =:= float; Tag =:= integer -> + assert_choices_1(T, Tag); + true -> + error(bad_select_list) + end; +assert_choices([]) -> ok. + +assert_choices_1([{Tag,_},{f,_}|T], Tag) -> + assert_choices_1(T, Tag); +assert_choices_1([_,{f,_}|_], _Tag) -> + error(bad_select_list); +assert_choices_1([], _Tag) -> ok. + +assert_arities([Arity,{f,_}|T]) when is_integer(Arity) -> + assert_arities(T); +assert_arities([]) -> ok; +assert_arities(_) -> error(bad_tuple_arity_list). + + %%% %%% Floating point checking. %%% @@ -1173,6 +1239,8 @@ assert_unique_map_keys([_,_|_]=Ls) -> %%% New binary matching instructions. %%% +bsm_match_state() -> + #ms{}. bsm_match_state(Slots) -> #ms{slots=Slots}. @@ -1186,6 +1254,12 @@ bsm_get_context({x,X}=Reg, #vst{current=#st{x=Xs}}=_Vst) when is_integer(X) -> {value,{fragile,#ms{}=Ctx}} -> Ctx; _ -> error({no_bsm_context,Reg}) end; +bsm_get_context({y,Y}=Reg, #vst{current=#st{y=Ys}}=_Vst) when is_integer(Y) -> + case gb_trees:lookup(Y, Ys) of + {value,#ms{}=Ctx} -> Ctx; + {value,{fragile,#ms{}=Ctx}} -> Ctx; + _ -> error({no_bsm_context,Reg}) + end; bsm_get_context(Reg, _) -> error({bad_source,Reg}). bsm_save(Reg, {atom,start}, Vst) -> @@ -1218,24 +1292,23 @@ bsm_restore(Reg, SavePoint, Vst) -> select_val_branches(Src, Choices, Vst) -> Infer = infer_types(Src, Vst), - select_val_branches_1(Choices, Infer, Vst). + select_val_branches_1(Choices, Src, Infer, Vst). -select_val_branches_1([Val,{f,L}|T], Infer, Vst0) -> - Vst = branch_state(L, Infer(Val, Vst0)), - select_val_branches_1(T, Infer, Vst); -select_val_branches_1([], _, Vst) -> Vst. +select_val_branches_1([Val,{f,L}|T], Src, Infer, Vst0) -> + Vst1 = set_aliased_type(Val, Src, Infer(Val, Vst0)), + Vst = branch_state(L, Vst1), + select_val_branches_1(T, Src, Infer, Vst); +select_val_branches_1([], _, _, Vst) -> Vst. infer_types(Src, Vst) -> case get_def(Src, Vst) of {bif,is_map,{f,_},[Map],_} -> - fun({atom,true}, S) -> set_type_reg(map, Map, S); + fun({atom,true}, S) -> update_type(fun meet/2, map, Map, S); (_, S) -> S end; {bif,tuple_size,{f,_},[Tuple],_} -> fun({integer,Arity}, S) -> - Type0 = get_term_type(Tuple, S), - Type = upgrade_tuple_type({tuple,Arity}, Type0), - set_type(Type, Tuple, S); + update_type(fun meet/2, {tuple,Arity}, Tuple, S); (_, S) -> S end; {bif,'=:=',{f,_},[ArityReg,{integer,_}=Val],_} when ArityReg =/= Src -> @@ -1252,11 +1325,93 @@ infer_types(Src, Vst) -> %%% Keeping track of types. %%% -set_alias(Reg1, Reg2, #vst{current=St0}=Vst) -> - case Reg1 of +%% Assigns Src to Dst and marks them as aliasing each other. +assign({y,_}=Src, {y,_}=Dst, Vst) -> + %% The stack trimming optimization may generate a move from an initialized + %% but unassigned Y register to another Y register. + case get_term_type_1(Src, Vst) of + initialized -> set_type_reg(initialized, Dst, Vst); + _ -> assign_1(Src, Dst, Vst) + end; +assign({Kind,_}=Reg, Dst, Vst) when Kind =:= x; Kind =:= y -> + assign_1(Reg, Dst, Vst); +assign(Literal, Dst, Vst) -> + create_term(get_term_type(Literal, Vst), Dst, Vst). + +%% Creates a completely new term with the given type. +create_term(Type, Dst, Vst) -> + set_type_reg(Type, Dst, Vst). + +%% Extracts a term from Ss, propagating fragility. +extract_term(Type, Ss, Dst, Vst) -> + extract_term(Type, Ss, Dst, Vst, Vst). + +%% As extract_term/4, but uses the incoming Vst for fragility in case x-regs +%% have been pruned and the sources can no longer be found. +extract_term(Type0, Ss, Dst, Vst, OrigVst) -> + Type = propagate_fragility(Type0, Ss, OrigVst), + set_type_reg(Type, Dst, Vst). + +%% Helper function for simple "is_type" tests. +type_test(Fail, Type, Reg, Vst0) -> + assert_term(Reg, Vst0), + Vst = branch_state(Fail, update_type(fun subtract/2, Type, Reg, Vst0)), + update_type(fun meet/2, Type, Reg, Vst). + +%% This is used when linear code finds out more and more information about a +%% type, so that the type gets more specialized. +update_type(Merge, Type0, Reg, Vst) -> + %% If the old type can't be merged with the new one, the type information + %% is inconsistent and we know that some instructions will never be + %% executed at run-time. For example: + %% + %% {test,is_list,Fail,[Reg]}. + %% {test,is_tuple,Fail,[Reg]}. + %% {test,test_arity,Fail,[Reg,5]}. + %% + %% Note that the test_arity instruction can never be reached, so we use the + %% new type instead of 'none'. + Type = case Merge(get_durable_term_type(Reg, Vst), Type0) of + none -> Type0; + T -> T + end, + set_aliased_type(propagate_fragility(Type, [Reg], Vst), Reg, Vst). + +update_ne_types(LHS, RHS, Vst) -> + T1 = get_durable_term_type(LHS, Vst), + T2 = get_durable_term_type(RHS, Vst), + Type = propagate_fragility(subtract(T1, T2), [LHS], Vst), + set_aliased_type(Type, LHS, Vst). + +update_eq_types(LHS, RHS, Vst0) -> + T1 = get_durable_term_type(LHS, Vst0), + T2 = get_durable_term_type(RHS, Vst0), + Meet = meet(T1, T2), + Vst = case T1 =/= Meet of + true -> + LType = propagate_fragility(Meet, [LHS], Vst0), + set_aliased_type(LType, LHS, Vst0); + false -> + Vst0 + end, + case T2 =/= Meet of + true -> + RType = propagate_fragility(Meet, [RHS], Vst0), + set_aliased_type(RType, RHS, Vst); + false -> + Vst + end. + +%% Helper functions for the above. + +assign_1(Src, Dst, Vst0) -> + Type = get_move_term_type(Src, Vst0), + Vst = set_type_reg(Type, Dst, Vst0), + case Src of {Kind,_} when Kind =:= x; Kind =:= y -> + #vst{current=St0} = Vst, #st{aliases=Aliases0} = St0, - Aliases = Aliases0#{Reg1=>Reg2,Reg2=>Reg1}, + Aliases = Aliases0#{Src=>Dst,Dst=>Src}, St = St0#st{aliases=Aliases}, Vst#vst{current=St}; _ -> @@ -1283,8 +1438,10 @@ kill_aliases(Reg, #st{aliases=Aliases0}=St) -> St end. -set_type(Type, {x,_}=Reg, Vst) -> set_type_reg(Type, Reg, Vst); -set_type(Type, {y,_}=Reg, Vst) -> set_type_y(Type, Reg, Vst); +set_type(Type, {x,_}=Reg, Vst) -> + set_type_reg(Type, Reg, Reg, Vst); +set_type(Type, {y,_}=Reg, Vst) -> + set_type_reg(Type, Reg, Reg, Vst); set_type(_, _, #vst{}=Vst) -> Vst. set_type_reg(Type, Src, Dst, Vst) -> @@ -1388,6 +1545,16 @@ assert_term(Src, Vst) -> get_term_type(Src, Vst), ok. +assert_not_fragile(Src, Vst) -> + case get_term_type(Src, Vst) of + {fragile, _} -> error({fragile_message_reference, Src}); + _ -> ok + end. + +assert_not_literal({x,_}) -> ok; +assert_not_literal({y,_}) -> ok; +assert_not_literal(Literal) -> error({literal_not_allowed,Literal}). + %% The possible types. %% %% First non-term types: @@ -1420,12 +1587,16 @@ assert_term(Src, Vst) -> %% %% term Any valid Erlang (but not of the special types above). %% +%% binary Binary or bitstring. +%% %% bool The atom 'true' or the atom 'false'. %% %% cons Cons cell: [_|_] %% %% nil Empty list: [] %% +%% list List: [] or [_|_] +%% %% {tuple,[Sz]} Tuple. An element has been accessed using %% element/2 or setelement/3 so that it is known that %% the type is a tuple of size at least Sz. @@ -1446,7 +1617,7 @@ assert_term(Src, Vst) -> %% %% map Map. %% -%% +%% none A conflict in types. There will be an exception at runtime. %% %% FRAGILITY %% --------- @@ -1459,14 +1630,55 @@ assert_term(Src, Vst) -> %% Such terms are wrapped in a {fragile,Type} tuple, where Type is one %% of the types described above. -assert_type(WantedType, Term, Vst) -> - case get_term_type(Term, Vst) of - {fragile,Type} -> - assert_type(WantedType, Type); - Type -> - assert_type(WantedType, Type) +%% meet(Type1, Type2) -> Type +%% Return the meet of two types. The meet is a more specific type. +%% It will be 'none' if the types are in conflict. + +meet(Same, Same) -> + Same; +meet(term, Other) -> + Other; +meet(Other, term) -> + Other; +meet(T1, T2) -> + case {erlang:min(T1, T2),erlang:max(T1, T2)} of + {{atom,_}=A,{atom,[]}} -> A; + {bool,{atom,B}=Atom} when is_boolean(B) -> Atom; + {bool,{atom,[]}} -> bool; + {cons,list} -> cons; + {{float,_}=T,{float,[]}} -> T; + {{integer,_}=T,{integer,[]}} -> T; + {list,nil} -> nil; + {number,{integer,_}=T} -> T; + {number,{float,_}=T} -> T; + {{tuple,Size1},{tuple,Size2}} -> + case {Size1,Size2} of + {[Sz1],[Sz2]} -> + {tuple,[erlang:max(Sz1, Sz2)]}; + {Sz1,[Sz2]} when Sz2 =< Sz1 -> + {tuple,Sz1}; + {_,_} -> + none + end; + {_,_} -> none end. +%% subtract(Type1, Type2) -> Type +%% Subtract Type2 from Type2. Example: +%% subtract(list, nil) -> cons + +subtract(list, nil) -> cons; +subtract(list, cons) -> nil; +subtract(number, {integer,[]}) -> {float,[]}; +subtract(number, {float,[]}) -> {integer,[]}; +subtract(bool, {atom,false}) -> {atom, true}; +subtract(bool, {atom,true}) -> {atom, false}; +subtract(Type, _) -> Type. + +assert_type(WantedType, Term, Vst) -> + Type = get_durable_term_type(Term, Vst), + assert_type(WantedType, Type). + assert_type(Correct, Correct) -> ok; assert_type(float, {float,_}) -> ok; assert_type(tuple, {tuple,_}) -> ok; @@ -1484,34 +1696,6 @@ assert_type(cons, {literal,[_|_]}) -> assert_type(Needed, Actual) -> error({bad_type,{needed,Needed},{actual,Actual}}). -%% upgrade_tuple_type(NewTupleType, OldType) -> TupleType. -%% upgrade_tuple_type/2 is used when linear code finds out more and -%% more information about a tuple type, so that the type gets more -%% specialized. If OldType is not a tuple type, the type information -%% is inconsistent, and we know that some instructions will never -%% be executed at run-time. - -upgrade_tuple_type(NewType, {fragile,OldType}) -> - make_fragile(upgrade_tuple_type_1(NewType, OldType)); -upgrade_tuple_type(NewType, OldType) -> - upgrade_tuple_type_1(NewType, OldType). - -upgrade_tuple_type_1({tuple,[Sz]}, {tuple,[OldSz]}=T) when Sz < OldSz -> - %% The old type has a higher value for the least tuple size. - T; -upgrade_tuple_type_1({tuple,[Sz]}, {tuple,OldSz}=T) - when is_integer(Sz), is_integer(OldSz), Sz =< OldSz -> - %% The old size is exact, and the new size is smaller than the old size. - T; -upgrade_tuple_type_1({tuple,_}=T, _) -> - %% The new type information is exact or has a higher value for - %% the least tuple size. - %% Note that inconsistencies are also handled in this - %% clause, e.g. if the old type was an integer or a tuple accessed - %% outside its size; inconsistences will generally cause an exception - %% at run-time but are safe from our point of view. - T. - get_tuple_size({integer,[]}) -> 0; get_tuple_size({integer,Sz}) -> Sz; get_tuple_size(_) -> 0. @@ -1519,6 +1703,17 @@ get_tuple_size(_) -> 0. validate_src(Ss, Vst) when is_list(Ss) -> foreach(fun(S) -> get_term_type(S, Vst) end, Ss). +%% get_durable_term_type(Src, ValidatorState) -> Type +%% Get the type of the source Src. The returned type Type will be +%% a standard Erlang type (no catch/try tags or match contexts). +%% Fragility will be stripped. + +get_durable_term_type(Src, Vst) -> + case get_term_type(Src, Vst) of + {fragile,Type} -> Type; + Type -> Type + end. + %% get_move_term_type(Src, ValidatorState) -> Type %% Get the type of the source Src. The returned type Type will be %% a standard Erlang type (no catch/try tags). Match contexts are OK. @@ -1552,6 +1747,8 @@ get_term_type_1(nil=T, _) -> T; get_term_type_1({atom,A}=T, _) when is_atom(A) -> T; get_term_type_1({float,F}=T, _) when is_float(F) -> T; get_term_type_1({integer,I}=T, _) when is_integer(I) -> T; +get_term_type_1({literal,[_|_]}, _) -> cons; +get_term_type_1({literal,Bitstring}, _) when is_bitstring(Bitstring) -> binary; get_term_type_1({literal,Map}, _) when is_map(Map) -> map; get_term_type_1({literal,Tuple}, _) when is_tuple(Tuple) -> {tuple,tuple_size(Tuple)}; @@ -1657,7 +1854,7 @@ merge_regs_1([{R1,_}|Rs1], [{R2,_}|_]=Rs2) when R1 < R2 -> merge_regs_1([{R1,_}|_]=Rs1, [{R2,_}|Rs2]) when R1 > R2 -> merge_regs_1(Rs1, Rs2); merge_regs_1([{R,Type1}|Rs1], [{R,Type2}|Rs2]) -> - [{R,merge_types(Type1, Type2)}|merge_regs_1(Rs1, Rs2)]; + [{R,join(Type1, Type2)}|merge_regs_1(Rs1, Rs2)]; merge_regs_1([], []) -> []; merge_regs_1([], [_|_]) -> []; merge_regs_1([_|_], []) -> []. @@ -1676,63 +1873,90 @@ merge_y_regs_1(Y, S, Regs0) when Y >= 0 -> Type0 -> merge_y_regs_1(Y-1, S, Regs0); Type1 -> - Type = merge_types(Type0, Type1), + Type = join(Type0, Type1), Regs = gb_trees:update(Y, Type, Regs0), merge_y_regs_1(Y-1, S, Regs) end; merge_y_regs_1(_, _, Regs) -> Regs. -%% merge_types(Type1, Type2) -> Type +%% join(Type1, Type2) -> Type %% Return the most specific type possible. %% Note: Type1 must NOT be the same as Type2. -merge_types({fragile,Same}=Type, Same) -> +join({literal,_}=T1, T2) -> + join_literal(T1, T2); +join(T1, {literal,_}=T2) -> + join_literal(T2, T1); +join({fragile,Same}=Type, Same) -> Type; -merge_types({fragile,T1}, T2) -> - make_fragile(merge_types(T1, T2)); -merge_types(Same, {fragile,Same}=Type) -> +join({fragile,T1}, T2) -> + make_fragile(join(T1, T2)); +join(Same, {fragile,Same}=Type) -> Type; -merge_types(T1, {fragile,T2}) -> - make_fragile(merge_types(T1, T2)); -merge_types(uninitialized=I, _) -> I; -merge_types(_, uninitialized=I) -> I; -merge_types(initialized=I, _) -> I; -merge_types(_, initialized=I) -> I; -merge_types({catchtag,T0},{catchtag,T1}) -> +join(T1, {fragile,T2}) -> + make_fragile(join(T1, T2)); +join(uninitialized=I, _) -> I; +join(_, uninitialized=I) -> I; +join(initialized=I, _) -> I; +join(_, initialized=I) -> I; +join({catchtag,T0},{catchtag,T1}) -> {catchtag,ordsets:from_list(T0++T1)}; -merge_types({trytag,T0},{trytag,T1}) -> +join({trytag,T0},{trytag,T1}) -> {trytag,ordsets:from_list(T0++T1)}; -merge_types({tuple,A}, {tuple,B}) -> +join({tuple,A}, {tuple,B}) -> {tuple,[min(tuple_sz(A), tuple_sz(B))]}; -merge_types({Type,A}, {Type,B}) +join({Type,A}, {Type,B}) when Type =:= atom; Type =:= integer; Type =:= float -> if A =:= B -> {Type,A}; true -> {Type,[]} end; -merge_types({Type,_}, number) +join({Type,_}, number) when Type =:= integer; Type =:= float -> number; -merge_types(number, {Type,_}) +join(number, {Type,_}) when Type =:= integer; Type =:= float -> number; -merge_types(bool, {atom,A}) -> +join(bool, {atom,A}) -> merge_bool(A); -merge_types({atom,A}, bool) -> +join({atom,A}, bool) -> merge_bool(A); -merge_types(cons, {literal,[_|_]}) -> - cons; -merge_types({literal,[_|_]}, cons) -> - cons; -merge_types({literal,[_|_]}, {literal,[_|_]}) -> - cons; -merge_types(#ms{id=Id1,valid=B1,slots=Slots1}, +join({atom,_}, {atom,_}) -> + {atom,[]}; +join(#ms{id=Id1,valid=B1,slots=Slots1}, #ms{id=Id2,valid=B2,slots=Slots2}) -> Id = if Id1 =:= Id2 -> Id1; true -> make_ref() end, #ms{id=Id,valid=B1 band B2,slots=min(Slots1, Slots2)}; -merge_types(T1, T2) when T1 =/= T2 -> - %% Too different. All we know is that the type is a 'term'. +join(T1, T2) when T1 =/= T2 -> + %% We've exhaused all other options, so the type must either be a list or + %% a 'term'. + join_list(T1, T2). + +%% Merges types of literals. Note that the left argument must either be a +%% literal or exactly equal to the second argument. +join_literal(Same, Same) -> + Same; +join_literal({literal,[_|_]}, T) -> + join_literal(T, cons); +join_literal({literal,#{}}, T) -> + join_literal(T, map); +join_literal({literal,Tuple}, T) when is_tuple(Tuple) -> + join_literal(T, {tuple, tuple_size(Tuple)}); +join_literal({literal,_}, T) -> + %% Bitstring, fun, or similar. + join_literal(T, term); +join_literal(T1, T2) -> + %% We're done extracting the types, try merging them again. + join(T1, T2). + +join_list(nil, cons) -> list; +join_list(nil, list) -> list; +join_list(cons, list) -> list; +join_list(T, nil) -> join_list(nil, T); +join_list(T, cons) -> join_list(cons, T); +join_list(_, _) -> + %% Not a list, so it must be a term. term. tuple_sz([Sz]) -> Sz; @@ -1836,13 +2060,16 @@ bif_type('+', Src, Vst) -> bif_type('*', Src, Vst) -> arith_type(Src, Vst); bif_type(abs, [Num], Vst) -> - case get_term_type(Num, Vst) of + case get_durable_term_type(Num, Vst) of {float,_}=T -> T; {integer,_}=T -> T; _ -> number end; bif_type(float, _, _) -> {float,[]}; bif_type('/', _, _) -> {float,[]}; +%% Binary operations +bif_type('byte_size', _, _) -> {integer,[]}; +bif_type('bit_size', _, _) -> {integer,[]}; %% Integer operations. bif_type(ceil, [_], _) -> {integer,[]}; bif_type('div', [_,_], _) -> {integer,[]}; @@ -1885,6 +2112,7 @@ bif_type(is_port, [_], _) -> bool; bif_type(is_reference, [_], _) -> bool; bif_type(is_tuple, [_], _) -> bool; %% Misc. +bif_type(tuple_size, [_], _) -> {integer,[]}; bif_type(node, [], _) -> {atom,[]}; bif_type(node, [_], _) -> {atom,[]}; bif_type(hd, [_], _) -> term; @@ -1919,8 +2147,18 @@ is_bif_safe(self, 0) -> true; is_bif_safe(node, 0) -> true; is_bif_safe(_, _) -> false. +arith_type([A], Vst) -> + %% Unary '+' or '-'. + case get_durable_term_type(A, Vst) of + {integer,_} -> {integer,[]}; + {float,_} -> {float,[]}; + _ -> number + end; arith_type([A,B], Vst) -> - case {get_term_type(A, Vst),get_term_type(B, Vst)} of + TypeA = get_durable_term_type(A, Vst), + TypeB = get_durable_term_type(B, Vst), + case {TypeA, TypeB} of + {{integer,_},{integer,_}} -> {integer,[]}; {{float,_},_} -> {float,[]}; {_,{float,_}} -> {float,[]}; {_,_} -> number @@ -1942,10 +2180,24 @@ return_type_1(erlang, setelement, 3, Vst) -> {tuple,[0]} end, case get_term_type({x,0}, Vst) of - {integer,[]} -> TupleType; - {integer,I} -> upgrade_tuple_type({tuple,[I]}, TupleType); - _ -> TupleType + {integer,[]} -> + TupleType; + {integer,I} -> + case meet({tuple,[I]}, TupleType) of + none -> TupleType; + T -> T + end; + _ -> + TupleType + end; +return_type_1(erlang, '++', 2, Vst) -> + case get_term_type({x,0}, Vst) =:= cons orelse + get_term_type({x,1}, Vst) =:= cons of + true -> cons; + false -> list end; +return_type_1(erlang, '--', 2, _Vst) -> + list; return_type_1(erlang, F, A, _) -> return_type_erl(F, A); return_type_1(math, F, A, _) -> diff --git a/lib/compiler/src/beam_z.erl b/lib/compiler/src/beam_z.erl index 677094b3cd..415b579240 100644 --- a/lib/compiler/src/beam_z.erl +++ b/lib/compiler/src/beam_z.erl @@ -71,6 +71,31 @@ undo_renames([{get_hd,Src,Dst1},{get_tl,Src,Dst2}|Is]) -> [{get_list,Src,Dst1,Dst2}|undo_renames(Is)]; undo_renames([{get_tl,Src,Dst2},{get_hd,Src,Dst1}|Is]) -> [{get_list,Src,Dst1,Dst2}|undo_renames(Is)]; +undo_renames([{bs_put,_,{bs_put_binary,1,_}, + [{atom,all},{literal,<<>>}]}|Is]) -> + undo_renames(Is); +undo_renames([{bs_put,Fail,{bs_put_binary,1,_Flags}, + [{atom,all},{literal,BinString}]}|Is0]) -> + Bits = bit_size(BinString), + Bytes = Bits div 8, + case Bits rem 8 of + 0 -> + I = {bs_put_string,byte_size(BinString), + {string,BinString}}, + [undo_rename(I)|undo_renames(Is0)]; + Rem -> + <<Binary:Bytes/bytes,Int:Rem>> = BinString, + PutInt = {bs_put_integer,Fail,{integer,Rem},1, + {field_flags,[unsigned,big]},{integer,Int}}, + Is = [PutInt|undo_renames(Is0)], + case Binary of + <<>> -> + Is; + _ -> + [{bs_put_string,byte_size(Binary), + {string,Binary}}|Is] + end + end; undo_renames([I|Is]) -> [undo_rename(I)|undo_renames(Is)]; undo_renames([]) -> []. @@ -79,8 +104,6 @@ undo_rename({bs_put,F,{I,U,Fl},[Sz,Src]}) -> {I,F,Sz,U,Fl,Src}; undo_rename({bs_put,F,{I,Fl},[Src]}) -> {I,F,Fl,Src}; -undo_rename({bs_put,{f,0},{bs_put_string,_,_}=I,[]}) -> - I; undo_rename({bif,bs_add=I,F,[Src1,Src2,{integer,U}],Dst}) -> {I,F,[Src1,Src2,U],Dst}; undo_rename({bif,bs_utf8_size=I,F,[Src],Dst}) -> @@ -101,7 +124,7 @@ undo_rename({test,bs_match_string=Op,F,[Ctx,Bin0]}) -> 0 -> Bin0; Rem -> <<Bin0/bitstring,0:(8-Rem)>> end, - {test,Op,F,[Ctx,Bits,{string,binary_to_list(Bin)}]}; + {test,Op,F,[Ctx,Bits,{string,Bin}]}; undo_rename({put_map,Fail,assoc,S,D,R,L}) -> {put_map_assoc,Fail,S,D,R,L}; undo_rename({put_map,Fail,exact,S,D,R,L}) -> diff --git a/lib/compiler/src/compile.erl b/lib/compiler/src/compile.erl index 883a5d8a1f..53d3cec2d7 100644 --- a/lib/compiler/src/compile.erl +++ b/lib/compiler/src/compile.erl @@ -210,8 +210,11 @@ do_compile(Input, Opts0) -> {error,Reason} end end, - %% Dialyzer has already spawned workers. - case lists:member(dialyzer, Opts) of + %% Some tools, like Dialyzer, has already spawned workers + %% and spawning extra workers actually slow the compilation + %% down instead of speeding it up, so we provide a mechanism + %% to bypass the compiler process. + case lists:member(no_spawn_compiler_process, Opts) of true -> IntFun(); false -> @@ -248,6 +251,9 @@ expand_opt(report, Os) -> [report_errors,report_warnings|Os]; expand_opt(return, Os) -> [return_errors,return_warnings|Os]; +expand_opt(no_bsm3, Os) -> + %% The new bsm pass requires bsm3 instructions. + [no_bsm3,no_bsm_opt|Os]; expand_opt(r16, Os) -> expand_opt_before_21(Os); expand_opt(r17, Os) -> @@ -259,13 +265,18 @@ expand_opt(r19, Os) -> expand_opt(r20, Os) -> expand_opt_before_21(Os); expand_opt(r21, Os) -> - [no_put_tuple2|Os]; + [no_put_tuple2 | expand_opt(no_bsm3, Os)]; expand_opt({debug_info_key,_}=O, Os) -> [encrypt_debug_info,O|Os]; +expand_opt(no_type_opt, Os) -> + [no_ssa_opt_type_start, + no_ssa_opt_type_continue, + no_ssa_opt_type_finish | Os]; expand_opt(O, Os) -> [O|Os]. expand_opt_before_21(Os) -> - [no_put_tuple2,no_get_hd_tl,no_ssa_opt_record,no_utf8_atoms|Os]. + [no_put_tuple2, no_get_hd_tl, no_ssa_opt_record, + no_utf8_atoms | expand_opt(no_bsm3, Os)]. %% format_error(ErrorDescriptor) -> string() @@ -816,6 +827,15 @@ kernel_passes() -> {pass,beam_kernel_to_ssa}, {iff,dssa,{listing,"ssa"}}, {iff,ssalint,{pass,beam_ssa_lint}}, + {unless,no_share_opt,{pass,beam_ssa_share}}, + {iff,dssashare,{listing,"ssashare"}}, + {iff,ssalint,{pass,beam_ssa_lint}}, + {unless,no_bsm_opt,{pass,beam_ssa_bsm}}, + {iff,dssabsm,{listing,"ssabsm"}}, + {iff,ssalint,{pass,beam_ssa_lint}}, + {unless,no_fun_opt,{pass,beam_ssa_funs}}, + {iff,dssafuns,{listing,"ssafuns"}}, + {iff,ssalint,{pass,beam_ssa_lint}}, {unless,no_ssa_opt,{pass,beam_ssa_opt}}, {iff,dssaopt,{listing,"ssaopt"}}, {iff,ssalint,{pass,beam_ssa_lint}}, @@ -839,20 +859,12 @@ asm_passes() -> {iff,dblk,{listing,"block"}}, {unless,no_except,{pass,beam_except}}, {iff,dexcept,{listing,"except"}}, - {unless,no_bs_opt,{pass,beam_bs}}, - {iff,dbs,{listing,"bs"}}, - {pass,beam_split}, - {iff,dsplit,{listing,"split"}}, - {unless,no_dead,{pass,beam_dead}}, - {iff,ddead,{listing,"dead"}}, {unless,no_jopt,{pass,beam_jump}}, {iff,djmp,{listing,"jump"}}, {unless,no_peep_opt,{pass,beam_peep}}, {iff,dpeep,{listing,"peep"}}, {pass,beam_clean}, {iff,dclean,{listing,"clean"}}, - {unless,no_bsm_opt,{pass,beam_bsm}}, - {iff,dbsm,{listing,"bsm"}}, {unless,no_stack_trimming,{pass,beam_trim}}, {iff,dtrim,{listing,"trim"}}, {pass,beam_flatten}]}, @@ -861,7 +873,9 @@ asm_passes() -> %% need to do a few clean-ups to code. {iff,no_postopt,[{pass,beam_clean}]}, + {iff,diffable,?pass(diffable)}, {pass,beam_z}, + {iff,diffable,{listing,"S"}}, {iff,dz,{listing,"z"}}, {iff,dopt,{listing,"optimize"}}, {iff,'S',{listing,"S"}}, @@ -1005,11 +1019,17 @@ parse_module(_Code, St0) -> end. do_parse_module(DefEncoding, #compile{ifile=File,options=Opts,dir=Dir}=St) -> + SourceName0 = proplists:get_value(source, Opts, File), + SourceName = case member(deterministic, Opts) of + true -> filename:basename(SourceName0); + false -> SourceName0 + end, R = epp:parse_file(File, - [{includes,[".",Dir|inc_paths(Opts)]}, - {macros,pre_defs(Opts)}, - {default_encoding,DefEncoding}, - extra]), + [{includes,[".",Dir|inc_paths(Opts)]}, + {source_name, SourceName}, + {macros,pre_defs(Opts)}, + {default_encoding,DefEncoding}, + extra]), case R of {ok,Forms,Extra} -> Encoding = proplists:get_value(encoding, Extra), @@ -1913,6 +1933,39 @@ restore_expand_module([F|Fs]) -> [F|restore_expand_module(Fs)]; restore_expand_module([]) -> []. +%%% +%%% Transform the BEAM code to make it more friendly for +%%% diffing: using function names instead of labels for +%%% local calls and number labels relative to each function. +%%% + +diffable(Code0, St) -> + {Mod,Exp,Attr,Fs0,NumLabels} = Code0, + EntryLabels0 = [{Entry,{Name,Arity}} || + {function,Name,Arity,Entry,_} <- Fs0], + EntryLabels = maps:from_list(EntryLabels0), + Fs = [diffable_fix_function(F, EntryLabels) || F <- Fs0], + Code = {Mod,Exp,Attr,Fs,NumLabels}, + {ok,Code,St}. + +diffable_fix_function({function,Name,Arity,Entry0,Is0}, LabelMap0) -> + Entry = maps:get(Entry0, LabelMap0), + {Is1,LabelMap} = diffable_label_map(Is0, 1, LabelMap0, []), + Fb = fun(Old) -> error({no_fb,Old}) end, + Is = beam_utils:replace_labels(Is1, [], LabelMap, Fb), + {function,Name,Arity,Entry,Is}. + +diffable_label_map([{label,Old}|Is], New, Map, Acc) -> + case Map of + #{Old:=NewLabel} -> + diffable_label_map(Is, New, Map, [{label,NewLabel}|Acc]); + #{} -> + diffable_label_map(Is, New+1, Map#{Old=>New}, [{label,New}|Acc]) + end; +diffable_label_map([I|Is], New, Map, Acc) -> + diffable_label_map(Is, New, Map, [I|Acc]); +diffable_label_map([], _New, Map, Acc) -> + {Acc,Map}. -spec options() -> 'ok'. @@ -2033,10 +2086,7 @@ pre_load() -> L = [beam_a, beam_asm, beam_block, - beam_bs, - beam_bsm, beam_clean, - beam_dead, beam_dict, beam_except, beam_flatten, @@ -2045,12 +2095,15 @@ pre_load() -> beam_opcodes, beam_peep, beam_ssa, + beam_ssa_bsm, beam_ssa_codegen, + beam_ssa_dead, + beam_ssa_funs, beam_ssa_opt, beam_ssa_pre_codegen, beam_ssa_recv, + beam_ssa_share, beam_ssa_type, - beam_split, beam_trim, beam_utils, beam_validator, diff --git a/lib/compiler/src/compiler.app.src b/lib/compiler/src/compiler.app.src index 74529f7fef..108a0ca100 100644 --- a/lib/compiler/src/compiler.app.src +++ b/lib/compiler/src/compiler.app.src @@ -24,10 +24,7 @@ beam_a, beam_asm, beam_block, - beam_bs, - beam_bsm, beam_clean, - beam_dead, beam_dict, beam_disasm, beam_except, @@ -38,14 +35,17 @@ beam_opcodes, beam_peep, beam_ssa, + beam_ssa_bsm, beam_ssa_codegen, + beam_ssa_dead, + beam_ssa_funs, beam_ssa_lint, beam_ssa_opt, beam_ssa_pp, beam_ssa_pre_codegen, beam_ssa_recv, + beam_ssa_share, beam_ssa_type, - beam_split, beam_trim, beam_utils, beam_validator, diff --git a/lib/compiler/src/erl_bifs.erl b/lib/compiler/src/erl_bifs.erl index 71ab0e872a..d925decce6 100644 --- a/lib/compiler/src/erl_bifs.erl +++ b/lib/compiler/src/erl_bifs.erl @@ -108,6 +108,7 @@ is_pure(erlang, list_to_atom, 1) -> true; is_pure(erlang, list_to_binary, 1) -> true; is_pure(erlang, list_to_float, 1) -> true; is_pure(erlang, list_to_integer, 1) -> true; +is_pure(erlang, list_to_integer, 2) -> true; is_pure(erlang, list_to_pid, 1) -> true; is_pure(erlang, list_to_tuple, 1) -> true; is_pure(erlang, max, 2) -> true; @@ -194,6 +195,7 @@ is_safe(erlang, is_float, 1) -> true; is_safe(erlang, is_function, 1) -> true; is_safe(erlang, is_integer, 1) -> true; is_safe(erlang, is_list, 1) -> true; +is_safe(erlang, is_map, 1) -> true; is_safe(erlang, is_number, 1) -> true; is_safe(erlang, is_pid, 1) -> true; is_safe(erlang, is_port, 1) -> true; diff --git a/lib/compiler/src/genop.tab b/lib/compiler/src/genop.tab index 8e34e3e291..86590fad87 100755 --- a/lib/compiler/src/genop.tab +++ b/lib/compiler/src/genop.tab @@ -574,7 +574,25 @@ BEAM_FORMAT_NUMBER=0 ## put it into the register Tail. 163: get_tl/2 +# OTP 22 + ## @spec put_tuple2 Destination Elements ## @doc Build a tuple with the elements in the list Elements and put it ## put into register Destination. 164: put_tuple2/2 + +## @spec bs_get_tail Ctx Dst Live +## @doc Sets Dst to the tail of Ctx at the current position +165: bs_get_tail/3 + +## @spec bs_start_match3 Fail Bin Live Dst +## @doc Starts a binary match sequence +166: bs_start_match3/4 + +## @spec bs_get_position Ctx Dst Live +## @doc Sets Dst to the current position of Ctx +167: bs_get_position/3 + +## @spec bs_set_positon Ctx Pos +## @doc Sets the current position of Ctx to Pos +168: bs_set_position/2 diff --git a/lib/compiler/src/sys_core_bsm.erl b/lib/compiler/src/sys_core_bsm.erl index d7b26c3a56..685e807e65 100644 --- a/lib/compiler/src/sys_core_bsm.erl +++ b/lib/compiler/src/sys_core_bsm.erl @@ -24,161 +24,52 @@ -export([module/2,format_error/1]). -include("core_parse.hrl"). --import(lists, [member/2,reverse/1,usort/1]). -spec module(cerl:c_module(), [compile:option()]) -> {'ok', cerl:c_module()}. -module(#c_module{defs=Ds0}=Mod, Opts) -> - {Ds,Ws0} = function(Ds0, [], []), - case member(bin_opt_info, Opts) of - false -> - {ok,Mod#c_module{defs=Ds}}; - true -> - Ws1 = [make_warning(Where, What) || {Where,What} <- Ws0], - Ws = usort(Ws1), - {ok,Mod#c_module{defs=Ds},Ws} - end. +module(#c_module{defs=Ds}=Mod, _Opts) -> + {ok,Mod#c_module{defs=function(Ds)}}. -function([{#c_var{name={F,Arity}}=Name,B0}|Fs], FsAcc, Ws0) -> - try cerl_trees:mapfold(fun bsm_an/2, Ws0, B0) of - {B,Ws} -> - function(Fs, [{Name,B}|FsAcc], Ws) +function([{#c_var{name={F,Arity}}=Name,B0}|Fs]) -> + try cerl_trees:map(fun bsm_reorder/1, B0) of + B -> [{Name,B} | function(Fs)] catch Class:Error:Stack -> - io:fwrite("Function: ~w/~w\n", [F,Arity]), - erlang:raise(Class, Error, Stack) + io:fwrite("Function: ~w/~w\n", [F,Arity]), + erlang:raise(Class, Error, Stack) end; -function([], Fs, Ws) -> - {reverse(Fs),Ws}. +function([]) -> + []. -type error() :: atom(). -spec format_error(error()) -> nonempty_string(). -format_error(bin_opt_alias) -> - "INFO: the '=' operator will prevent delayed sub binary optimization"; -format_error(bin_partition) -> - "INFO: matching non-variables after a previous clause matching a variable " - "will prevent delayed sub binary optimization"; -format_error(bin_var_used) -> - "INFO: using a matched out sub binary will prevent " - "delayed sub binary optimization"; -format_error(orig_bin_var_used_in_guard) -> - "INFO: using the original binary variable in a guard will prevent " - "delayed sub binary optimization"; -format_error(bin_var_used_in_guard) -> - "INFO: using a matched out sub binary in a guard will prevent " - "delayed sub binary optimization". - +format_error(_) -> error(badarg). -%%% -%%% Annotate bit syntax matching to faciliate optimization in further passes. -%%% +%%% Reorder bit syntax matching to faciliate optimization in further passes. -bsm_an(Core0, Ws0) -> - case bsm_an(Core0) of - {ok,Core} -> - {Core,Ws0}; - {ok,Core,W} -> - {Core,[W|Ws0]} - end. +bsm_reorder(#c_case{arg=#c_var{}=V}=Case) -> + bsm_reorder_1([V], Case); +bsm_reorder(#c_case{arg=#c_values{es=Es}}=Case) -> + bsm_reorder_1(Es, Case); +bsm_reorder(Core) -> + Core. -bsm_an(#c_case{arg=#c_var{}=V}=Case) -> - bsm_an_1([V], Case); -bsm_an(#c_case{arg=#c_values{es=Es}}=Case) -> - bsm_an_1(Es, Case); -bsm_an(Other) -> - {ok,Other}. - -bsm_an_1(Vs0, #c_case{clauses=Cs0}=Case) -> +bsm_reorder_1(Vs0, #c_case{clauses=Cs0}=Case) -> case bsm_leftmost(Cs0) of - none -> - {ok,Case}; - 1 -> - bsm_an_2(Vs0, Cs0, Case); - Pos -> - Vs = move_from_col(Pos, Vs0), - Cs = [C#c_clause{pats=move_from_col(Pos, Ps)} || - #c_clause{pats=Ps}=C <- Cs0], - bsm_an_2(Vs, Cs, Case) - end. - -bsm_an_2(Vs, Cs, Case) -> - try - bsm_ensure_no_partition(Cs), - {ok,bsm_do_an(Vs, Cs, Case)} - catch - throw:{problem,Where,What} -> - {ok,Case,{Where,What}} + Pos when Pos > 0, Pos =/= none -> + Vs = core_lib:make_values(move_from_col(Pos, Vs0)), + Cs = [C#c_clause{pats=move_from_col(Pos, Ps)} + || #c_clause{pats=Ps}=C <- Cs0], + Case#c_case{arg=Vs,clauses=Cs}; + _ -> + Case end. move_from_col(Pos, L) -> {First,[Col|Rest]} = lists:split(Pos - 1, L), [Col|First] ++ Rest. -bsm_do_an([#c_var{name=Vname}=V0|Vs0], Cs0, Case) -> - Cs = bsm_do_an_var(Vname, Cs0), - V = bsm_annotate_for_reuse(V0), - Vs = core_lib:make_values([V|Vs0]), - Case#c_case{arg=Vs,clauses=Cs}; -bsm_do_an(_Vs, _Cs, Case) -> Case. - -bsm_do_an_var(V, [#c_clause{pats=[P|_],guard=G,body=B0}=C0|Cs]) -> - case P of - #c_var{name=VarName} -> - case core_lib:is_var_used(V, G) of - true -> bsm_problem(C0, orig_bin_var_used_in_guard); - false -> ok - end, - case core_lib:is_var_used(VarName, G) of - true -> bsm_problem(C0, bin_var_used_in_guard); - false -> ok - end, - B1 = bsm_maybe_ctx_to_binary(VarName, B0), - B = bsm_maybe_ctx_to_binary(V, B1), - C = C0#c_clause{body=B}, - [C|bsm_do_an_var(V, Cs)]; - #c_alias{} -> - case bsm_could_match_binary(P) of - false -> - [C0|bsm_do_an_var(V, Cs)]; - true -> - bsm_problem(C0, bin_opt_alias) - end; - _ -> - case bsm_could_match_binary(P) andalso bsm_is_var_used(V, G, B0) of - false -> - [C0|bsm_do_an_var(V, Cs)]; - true -> - bsm_problem(C0, bin_var_used) - end - end; -bsm_do_an_var(_, []) -> []. - -bsm_annotate_for_reuse(#c_var{anno=Anno}=Var) -> - Var#c_var{anno=[reuse_for_context|Anno]}. - -bsm_is_var_used(V, G, B) -> - core_lib:is_var_used(V, G) orelse core_lib:is_var_used(V, B). - -bsm_maybe_ctx_to_binary(V, B) -> - case core_lib:is_var_used(V, B) andalso not previous_ctx_to_binary(V, B) of - false -> - B; - true -> - #c_seq{arg=#c_primop{name=#c_literal{val=bs_context_to_binary}, - args=[#c_var{name=V}]}, - body=B} - end. - -previous_ctx_to_binary(V, Core) -> - case Core of - #c_seq{arg=#c_primop{name=#c_literal{val=bs_context_to_binary}, - args=[#c_var{name=V}]}} -> - true; - _ -> - false - end. - %% bsm_leftmost(Cs) -> none | ArgumentNumber %% Find the leftmost argument that matches a nonempty binary. %% Return either 'none' or the argument number (1-N). @@ -200,94 +91,3 @@ bsm_leftmost_2([_|Ps], Cs, N, Pos) -> bsm_leftmost_2(Ps, Cs, N+1, Pos); bsm_leftmost_2([], Cs, _, Pos) -> bsm_leftmost_1(Cs, Pos). - -%% bsm_ensure_no_partition(Cs) -> ok (exception if problem) -%% There must only be a single bs_start_match2 instruction if we -%% are to reuse the binary variable for the match context. -%% -%% To make sure that there is only a single bs_start_match2 -%% instruction, we will check for partitions such as: -%% -%% foo(<<...>>) -> ... -%% foo(<Variable>) when ... -> ... -%% foo(<Non-variable pattern>) -> -%% -%% If there is such partition, we reject the optimization. - -bsm_ensure_no_partition(Cs) -> - bsm_ensure_no_partition_1(Cs, before). - -%% Loop through each clause. -bsm_ensure_no_partition_1([#c_clause{pats=Ps,guard=G}|Cs], State0) -> - State = bsm_ensure_no_partition_2(Ps, G, State0), - case State of - 'after' -> - bsm_ensure_no_partition_after(Cs); - _ -> - ok - end, - bsm_ensure_no_partition_1(Cs, State); -bsm_ensure_no_partition_1([], _) -> ok. - -bsm_ensure_no_partition_2([#c_binary{}|_], _, _State) -> - within; -bsm_ensure_no_partition_2([#c_alias{}=Alias|_], N, State) -> - %% Retrieve the real pattern that the alias refers to and check that. - P = bsm_real_pattern(Alias), - bsm_ensure_no_partition_2([P], N, State); -bsm_ensure_no_partition_2([_|_], _, before=State) -> - %% No binary matching yet - therefore no partition. - State; -bsm_ensure_no_partition_2([P|_], _, State) -> - case bsm_could_match_binary(P) of - false -> - State; - true -> - %% The pattern P *may* match a binary, so we must update the state. - %% (P must be a variable.) - 'after' - end. - -bsm_ensure_no_partition_after([#c_clause{pats=Ps}=C|Cs]) -> - case Ps of - [#c_var{}|_] -> - bsm_ensure_no_partition_after(Cs); - _ -> - bsm_problem(C, bin_partition) - end; -bsm_ensure_no_partition_after([]) -> ok. - -bsm_could_match_binary(#c_alias{pat=P}) -> bsm_could_match_binary(P); -bsm_could_match_binary(#c_cons{}) -> false; -bsm_could_match_binary(#c_tuple{}) -> false; -bsm_could_match_binary(#c_literal{val=Lit}) -> is_bitstring(Lit); -bsm_could_match_binary(_) -> true. - -bsm_real_pattern(#c_alias{pat=P}) -> bsm_real_pattern(P); -bsm_real_pattern(P) -> P. - -bsm_problem(Where, What) -> - throw({problem,Where,What}). - -make_warning(Core, Term) -> - case should_suppress_warning(Core) of - true -> - ok; - false -> - Anno = cerl:get_ann(Core), - Line = get_line(Anno), - File = get_file(Anno), - {File,[{Line,?MODULE,Term}]} - end. - -should_suppress_warning(Core) -> - Ann = cerl:get_ann(Core), - member(compiler_generated, Ann). - -get_line([Line|_]) when is_integer(Line) -> Line; -get_line([_|T]) -> get_line(T); -get_line([]) -> none. - -get_file([{file,File}|_]) -> File; -get_file([_|T]) -> get_file(T); -get_file([]) -> "no_file". % should not happen diff --git a/lib/compiler/src/sys_core_fold.erl b/lib/compiler/src/sys_core_fold.erl index d848cd8f19..43c99be982 100644 --- a/lib/compiler/src/sys_core_fold.erl +++ b/lib/compiler/src/sys_core_fold.erl @@ -2667,12 +2667,20 @@ opt_build_stacktrace(#c_let{vars=[#c_var{name=Cooked}], #c_call{module=#c_literal{val=erlang}, name=#c_literal{val=raise}, args=[Class,Exp,#c_var{name=Cooked}]} -> - %% The stacktrace is only used in a call to erlang:raise/3. - %% There is no need to build the stacktrace. Replace the - %% call to erlang:raise/3 with the the raw_raise/3 instruction, - %% which will use a raw stacktrace. - #c_primop{name=#c_literal{val=raw_raise}, - args=[Class,Exp,RawStk]}; + case core_lib:is_var_used(Cooked, #c_cons{hd=Class,tl=Exp}) of + true -> + %% Not safe. The stacktrace is used in the class or + %% reason. + Let; + false -> + %% The stacktrace is only used in the last + %% argument for erlang:raise/3. There is no need + %% to build the stacktrace. Replace the call to + %% erlang:raise/3 with the the raw_raise/3 + %% instruction, which will use a raw stacktrace. + #c_primop{name=#c_literal{val=raw_raise}, + args=[Class,Exp,RawStk]} + end; #c_let{vars=[#c_var{name=V}],arg=Arg,body=B0} when V =/= Cooked -> case core_lib:is_var_used(Cooked, Arg) of false -> diff --git a/lib/compiler/src/sys_core_fold_lists.erl b/lib/compiler/src/sys_core_fold_lists.erl index 9867fab46a..e93b435011 100644 --- a/lib/compiler/src/sys_core_fold_lists.erl +++ b/lib/compiler/src/sys_core_fold_lists.erl @@ -37,22 +37,27 @@ call(#c_call{anno=Anno}, lists, all, [Arg1,Arg2]) -> Xs = #c_var{name='Xs'}, X = #c_var{name='X'}, Err1 = #c_tuple{es=[#c_literal{val='case_clause'}, X]}, - CC1 = #c_clause{pats=[#c_literal{val=true}], guard=#c_literal{val=true}, + CC1 = #c_clause{anno=Anno, + pats=[#c_literal{val=true}], guard=#c_literal{val=true}, body=#c_apply{anno=Anno, op=Loop, args=[Xs]}}, - CC2 = #c_clause{pats=[#c_literal{val=false}], guard=#c_literal{val=true}, + CC2 = #c_clause{anno=Anno, + pats=[#c_literal{val=false}], guard=#c_literal{val=true}, body=#c_literal{val=false}}, - CC3 = #c_clause{pats=[X], guard=#c_literal{val=true}, + CC3 = #c_clause{anno=Anno, + pats=[X], guard=#c_literal{val=true}, body=match_fail(Anno, Err1)}, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=#c_case{arg=#c_apply{anno=Anno, op=F, args=[X]}, clauses = [CC1, CC2, CC3]}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, + pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=1}]}, body=#c_literal{val=true}}, Err2 = #c_tuple{es=[#c_literal{val='function_clause'}, F, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^all',1}}|Anno], Err2)}, Fun = #c_fun{vars=[Xs], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, @@ -66,16 +71,21 @@ call(#c_call{anno=Anno}, lists, any, [Arg1,Arg2]) -> Xs = #c_var{name='Xs'}, X = #c_var{name='X'}, Err1 = #c_tuple{es=[#c_literal{val='case_clause'}, X]}, - CC1 = #c_clause{pats=[#c_literal{val=true}], guard=#c_literal{val=true}, + CC1 = #c_clause{anno=Anno, + pats=[#c_literal{val=true}], guard=#c_literal{val=true}, body=#c_literal{val=true}}, - CC2 = #c_clause{pats=[#c_literal{val=false}], guard=#c_literal{val=true}, + CC2 = #c_clause{anno=Anno, + pats=[#c_literal{val=false}], guard=#c_literal{val=true}, body=#c_apply{anno=Anno, op=Loop, args=[Xs]}}, - CC3 = #c_clause{pats=[X], guard=#c_literal{val=true}, + CC3 = #c_clause{anno=Anno, + pats=[X], guard=#c_literal{val=true}, body=match_fail(Anno, Err1)}, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=#c_case{arg=#c_apply{anno=Anno, op=F, args=[X]}, clauses = [CC1, CC2, CC3]}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, + pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=1}]}, @@ -94,16 +104,17 @@ call(#c_call{anno=Anno}, lists, foreach, [Arg1,Arg2]) -> F = #c_var{name='F'}, Xs = #c_var{name='Xs'}, X = #c_var{name='X'}, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=#c_seq{arg=#c_apply{anno=Anno, op=F, args=[X]}, body=#c_apply{anno=Anno, op=Loop, args=[Xs]}}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=1}]}, body=#c_literal{val=ok}}, Err = #c_tuple{es=[#c_literal{val='function_clause'}, F, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^foreach',1}}|Anno], Err)}, Fun = #c_fun{vars=[Xs], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, @@ -117,7 +128,8 @@ call(#c_call{anno=Anno}, lists, map, [Arg1,Arg2]) -> Xs = #c_var{name='Xs'}, X = #c_var{name='X'}, H = #c_var{name='H'}, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=#c_let{vars=[H], arg=#c_apply{anno=Anno, op=F, args=[X]}, @@ -126,7 +138,7 @@ call(#c_call{anno=Anno}, lists, map, [Arg1,Arg2]) -> tl=#c_apply{anno=Anno, op=Loop, args=[Xs]}}}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=1}]}, @@ -146,7 +158,8 @@ call(#c_call{anno=Anno}, lists, flatmap, [Arg1,Arg2]) -> Xs = #c_var{name='Xs'}, X = #c_var{name='X'}, H = #c_var{name='H'}, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=#c_let{vars=[H], arg=#c_apply{anno=Anno, op=F, args=[X]}, body=#c_call{anno=[compiler_generated|Anno], @@ -156,13 +169,13 @@ call(#c_call{anno=Anno}, lists, flatmap, [Arg1,Arg2]) -> #c_apply{anno=Anno, op=Loop, args=[Xs]}]}}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=1}]}, body=#c_literal{val=[]}}, Err = #c_tuple{es=[#c_literal{val='function_clause'}, F, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^flatmap',1}}|Anno], Err)}, Fun = #c_fun{vars=[Xs], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, @@ -177,11 +190,13 @@ call(#c_call{anno=Anno}, lists, filter, [Arg1,Arg2]) -> X = #c_var{name='X'}, B = #c_var{name='B'}, Err1 = #c_tuple{es=[#c_literal{val='case_clause'}, X]}, - CC1 = #c_clause{pats=[#c_literal{val=true}], guard=#c_literal{val=true}, + CC1 = #c_clause{anno=Anno, + pats=[#c_literal{val=true}], guard=#c_literal{val=true}, body=#c_cons{anno=[compiler_generated], hd=X, tl=Xs}}, - CC2 = #c_clause{pats=[#c_literal{val=false}], guard=#c_literal{val=true}, + CC2 = #c_clause{anno=Anno, + pats=[#c_literal{val=false}], guard=#c_literal{val=true}, body=Xs}, - CC3 = #c_clause{pats=[X], guard=#c_literal{val=true}, + CC3 = #c_clause{anno=Anno, pats=[X], guard=#c_literal{val=true}, body=match_fail(Anno, Err1)}, Case = #c_case{arg=B, clauses = [CC1, CC2, CC3]}, C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, @@ -192,13 +207,15 @@ call(#c_call{anno=Anno}, lists, filter, [Arg1,Arg2]) -> op=Loop, args=[Xs]}, body=Case}}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, + pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=1}]}, body=#c_literal{val=[]}}, Err2 = #c_tuple{es=[#c_literal{val='function_clause'}, F, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, + pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^filter',1}}|Anno], Err2)}, Fun = #c_fun{vars=[Xs], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, @@ -212,19 +229,20 @@ call(#c_call{anno=Anno}, lists, foldl, [Arg1,Arg2,Arg3]) -> Xs = #c_var{name='Xs'}, X = #c_var{name='X'}, A = #c_var{name='A'}, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=#c_apply{anno=Anno, op=Loop, args=[Xs, #c_apply{anno=Anno, op=F, args=[X, A]}]}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=2}]}, body=A}, Err = #c_tuple{es=[#c_literal{val='function_clause'}, F, A, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^foldl',2}}|Anno], Err)}, Fun = #c_fun{vars=[Xs, A], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, @@ -238,19 +256,20 @@ call(#c_call{anno=Anno}, lists, foldr, [Arg1,Arg2,Arg3]) -> Xs = #c_var{name='Xs'}, X = #c_var{name='X'}, A = #c_var{name='A'}, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=#c_apply{anno=Anno, op=F, args=[X, #c_apply{anno=Anno, op=Loop, args=[Xs, A]}]}}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=2}]}, body=A}, Err = #c_tuple{es=[#c_literal{val='function_clause'}, F, A, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^foldr',2}}|Anno], Err)}, Fun = #c_fun{vars=[Xs, A], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, @@ -266,13 +285,14 @@ call(#c_call{anno=Anno}, lists, mapfoldl, [Arg1,Arg2,Arg3]) -> Avar = #c_var{name='A'}, Match = fun (A, P, E) -> - C1 = #c_clause{pats=[P], guard=#c_literal{val=true}, body=E}, + C1 = #c_clause{anno=Anno, pats=[P], guard=#c_literal{val=true}, body=E}, Err = #c_tuple{es=[#c_literal{val='badmatch'}, X]}, - C2 = #c_clause{pats=[X], guard=#c_literal{val=true}, + C2 = #c_clause{anno=Anno, pats=[X], guard=#c_literal{val=true}, body=match_fail(Anno, Err)}, #c_case{arg=A, clauses=[C1, C2]} end, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, + pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, body=Match(#c_apply{anno=Anno, op=F, args=[X, Avar]}, #c_tuple{es=[X, Avar]}, %%% Tuple passing version @@ -292,7 +312,7 @@ call(#c_call{anno=Anno}, lists, mapfoldl, [Arg1,Arg2,Arg3]) -> %%% body=#c_values{es=[#c_cons{hd=X, tl=Xs}, %%% A]}} )}, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=2}]}, @@ -302,7 +322,7 @@ call(#c_call{anno=Anno}, lists, mapfoldl, [Arg1,Arg2,Arg3]) -> %%% Multiple-value version %%% body=#c_values{es=[#c_literal{val=[]}, A]}}, Err = #c_tuple{es=[#c_literal{val='function_clause'}, F, Avar, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^mapfoldl',2}}|Anno], Err)}, Fun = #c_fun{vars=[Xs, Avar], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, @@ -326,13 +346,13 @@ call(#c_call{anno=Anno}, lists, mapfoldr, [Arg1,Arg2,Arg3]) -> Avar = #c_var{name='A'}, Match = fun (A, P, E) -> - C1 = #c_clause{pats=[P], guard=#c_literal{val=true}, body=E}, + C1 = #c_clause{anno=Anno, pats=[P], guard=#c_literal{val=true}, body=E}, Err = #c_tuple{es=[#c_literal{val='badmatch'}, X]}, - C2 = #c_clause{pats=[X], guard=#c_literal{val=true}, + C2 = #c_clause{anno=Anno, pats=[X], guard=#c_literal{val=true}, body=match_fail(Anno, Err)}, #c_case{arg=A, clauses=[C1, C2]} end, - C1 = #c_clause{pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, + C1 = #c_clause{anno=Anno, pats=[#c_cons{hd=X, tl=Xs}], guard=#c_literal{val=true}, %%% Tuple passing version body=Match(#c_apply{anno=Anno, op=Loop, @@ -352,7 +372,8 @@ call(#c_call{anno=Anno}, lists, mapfoldr, [Arg1,Arg2,Arg3]) -> %%% #c_values{es=[#c_cons{hd=X, tl=Xs}, %%% A]})} }, - C2 = #c_clause{pats=[#c_literal{val=[]}], + C2 = #c_clause{anno=Anno, + pats=[#c_literal{val=[]}], guard=#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[F, #c_literal{val=2}]}, @@ -362,7 +383,7 @@ call(#c_call{anno=Anno}, lists, mapfoldr, [Arg1,Arg2,Arg3]) -> %%% Multiple-value version %%% body=#c_values{es=[#c_literal{val=[]}, A]}}, Err = #c_tuple{es=[#c_literal{val='function_clause'}, F, Avar, Xs]}, - C3 = #c_clause{pats=[Xs], guard=#c_literal{val=true}, + C3 = #c_clause{anno=Anno, pats=[Xs], guard=#c_literal{val=true}, body=match_fail([{function_name,{'lists^mapfoldr',2}}|Anno], Err)}, Fun = #c_fun{vars=[Xs, Avar], body=#c_case{arg=Xs, clauses=[C1, C2, C3]}}, diff --git a/lib/compiler/src/sys_core_inline.erl b/lib/compiler/src/sys_core_inline.erl index 5a6cc45e4a..3380e3f1bd 100644 --- a/lib/compiler/src/sys_core_inline.erl +++ b/lib/compiler/src/sys_core_inline.erl @@ -195,6 +195,9 @@ kill_id_anns(Body) -> cerl_trees:map(fun(#c_fun{anno=A0}=CFun) -> A = kill_id_anns_1(A0), CFun#c_fun{anno=A}; + (#c_var{anno=A0}=Var) -> + A = kill_id_anns_1(A0), + Var#c_var{anno=A}; (Expr) -> %% Mark everything as compiler generated to %% suppress bogus warnings. diff --git a/lib/compiler/src/v3_core.erl b/lib/compiler/src/v3_core.erl index 45e0ed5088..34930c3afe 100644 --- a/lib/compiler/src/v3_core.erl +++ b/lib/compiler/src/v3_core.erl @@ -2627,7 +2627,8 @@ cfun(#ifun{anno=A,id=Id,vars=Args,clauses=Lcs,fc=Lfc}, _As, St0) -> [],A#a.us,St2}. c_call_erl(Fun, Args) -> - cerl:c_call(cerl:c_atom(erlang), cerl:c_atom(Fun), Args). + As = [compiler_generated], + cerl:ann_c_call(As, cerl:c_atom(erlang), cerl:c_atom(Fun), Args). %% lit_vars(Literal) -> [Var]. diff --git a/lib/compiler/src/v3_kernel.erl b/lib/compiler/src/v3_kernel.erl index fe8e252e5a..f7ca66b1da 100644 --- a/lib/compiler/src/v3_kernel.erl +++ b/lib/compiler/src/v3_kernel.erl @@ -82,8 +82,7 @@ -export([module/2,format_error/1]). -import(lists, [map/2,foldl/3,foldr/3,mapfoldl/3,splitwith/2,member/2, - keymember/3,keyfind/3,partition/2,droplast/1,last/1,sort/1, - reverse/1]). + keyfind/3,partition/2,droplast/1,last/1,sort/1,reverse/1]). -import(ordsets, [add_element/2,del_element/2,union/2,union/1,subtract/2]). -import(cerl, [c_tuple/1]). @@ -1416,7 +1415,6 @@ is_remote_bif(_, _, _) -> false. %% called for effect only. bif_vals(dsetelement, 3) -> 0; -bif_vals(bs_context_to_binary, 1) -> 0; bif_vals(_, _) -> 1. bif_vals(_, _, _) -> 1. @@ -2337,8 +2335,7 @@ uexpr(#k_bif{anno=A,op=Op,args=As}=Bif, {break,Rs}, St0) -> {Brs,St1} = bif_returns(Op, Rs, St0), {Bif#k_bif{anno=#k{us=Used,ns=lit_list_vars(Brs),a=A},ret=Brs}, Used,St1}; -uexpr(#k_match{anno=A,vars=Vs0,body=B0}, Br, St0) -> - Vs = handle_reuse_annos(Vs0, St0), +uexpr(#k_match{anno=A,vars=Vs,body=B0}, Br, St0) -> Rs = break_rets(Br), {B1,Bu,St1} = umatch(B0, Br, St0), case is_in_guard(St1) of @@ -2441,33 +2438,6 @@ make_fdef(Anno, Name, Arity, Vs, Body) -> vars=Vs,body=Body,ret=[]}, #k_fdef{anno=Anno,func=Name,arity=Arity,vars=Vs,body=Match}. - -%% handle_reuse_annos([#k_var{}], State) -> State. -%% In general, it is only safe to reuse a variable for a match context -%% if the original value of the variable will no longer be needed. -%% -%% If a variable has been bound in an outer letrec and is therefore -%% free in the current function, the variable may still be used. -%% We don't bother to check whether the variable is actually used, -%% but simply clears the 'reuse_for_context' annotation for any variable -%% that is free. -handle_reuse_annos(Vs, St) -> - [handle_reuse_anno(V, St) || V <- Vs]. - -handle_reuse_anno(#k_var{anno=A}=V, St) -> - case member(reuse_for_context, A) of - false -> V; - true -> handle_reuse_anno_1(V, St) - end. - -handle_reuse_anno_1(#k_var{anno=Anno,name=Vname}=V, #kern{ff={F,A}}=St) -> - FreeVs = get_free(F, A, St), - case keymember(Vname, #k_var.name, FreeVs) of - true -> V#k_var{anno=Anno--[reuse_for_context]}; - false -> V - end; -handle_reuse_anno_1(V, _St) -> V. - %% get_free(Name, Arity, State) -> [Free]. %% store_free(Name, Arity, [Free], State) -> State. @@ -2511,8 +2481,7 @@ umatch(#k_alt{anno=A,first=F0,then=T0}, Br, St0) -> Used = union(Fu, Tu), {#k_alt{anno=#k{us=Used,ns=[],a=A},first=F1,then=T1}, Used,St2}; -umatch(#k_select{anno=A,var=V0,types=Ts0}, Br, St0) -> - V = handle_reuse_anno(V0, St0), +umatch(#k_select{anno=A,var=V,types=Ts0}, Br, St0) -> {Ts1,Tus,St1} = umatch_list(Ts0, Br, St0), Used = case member(no_usage, get_kanno(V)) of true -> Tus; |