diff options
Diffstat (limited to 'lib/compiler/src')
29 files changed, 3783 insertions, 3377 deletions
diff --git a/lib/compiler/src/Makefile b/lib/compiler/src/Makefile index 87b0d345f2..f253f31d13 100644 --- a/lib/compiler/src/Makefile +++ b/lib/compiler/src/Makefile @@ -49,10 +49,10 @@ MODULES = \ beam_a \ beam_asm \ beam_block \ + beam_call_types \ beam_clean \ beam_dict \ beam_disasm \ - beam_except \ beam_flatten \ beam_jump \ beam_listing \ @@ -72,6 +72,7 @@ MODULES = \ beam_ssa_type \ beam_kernel_to_ssa \ beam_trim \ + beam_types \ beam_utils \ beam_validator \ beam_z \ @@ -104,6 +105,7 @@ HRL_FILES= \ beam_disasm.hrl \ beam_ssa_opt.hrl \ beam_ssa.hrl \ + beam_types.hrl \ core_parse.hrl \ v3_kernel.hrl @@ -190,6 +192,7 @@ release_docs_spec: # Dependencies -- alphabetically, please # ---------------------------------------------------- +$(EBIN)/beam_call_types.beam: beam_types.hrl $(EBIN)/beam_disasm.beam: $(EGEN)/beam_opcodes.hrl beam_disasm.hrl $(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 @@ -204,7 +207,8 @@ $(EBIN)/beam_ssa_pp.beam: beam_ssa.hrl $(EBIN)/beam_ssa_pre_codegen.beam: beam_ssa.hrl $(EBIN)/beam_ssa_recv.beam: beam_ssa.hrl $(EBIN)/beam_ssa_share.beam: beam_ssa.hrl -$(EBIN)/beam_ssa_type.beam: beam_ssa.hrl +$(EBIN)/beam_ssa_type.beam: beam_ssa.hrl beam_types.hrl +$(EBIN)/beam_types.beam: beam_types.hrl $(EBIN)/cerl.beam: core_parse.hrl $(EBIN)/compile.beam: core_parse.hrl ../../stdlib/include/erl_compile.hrl $(EBIN)/core_lib.beam: core_parse.hrl diff --git a/lib/compiler/src/beam_asm.erl b/lib/compiler/src/beam_asm.erl index df09dcb06c..60e19ec596 100644 --- a/lib/compiler/src/beam_asm.erl +++ b/lib/compiler/src/beam_asm.erl @@ -64,11 +64,30 @@ module(Code, ExtraChunks, CompileInfo, CompilerOpts) -> assemble({Mod,Exp0,Attr0,Asm0,NumLabels}, ExtraChunks, CompileInfo, CompilerOpts) -> {1,Dict0} = beam_dict:atom(Mod, beam_dict:new()), {0,Dict1} = beam_dict:fname(atom_to_list(Mod) ++ ".erl", Dict0), + Dict2 = shared_fun_wrappers(CompilerOpts, Dict1), NumFuncs = length(Asm0), {Asm,Attr} = on_load(Asm0, Attr0), Exp = cerl_sets:from_list(Exp0), - {Code,Dict2} = assemble_1(Asm, Exp, Dict1, []), - build_file(Code, Attr, Dict2, NumLabels, NumFuncs, ExtraChunks, CompileInfo, CompilerOpts). + {Code,Dict} = assemble_1(Asm, Exp, Dict2, []), + build_file(Code, Attr, Dict, NumLabels, NumFuncs, + ExtraChunks, CompileInfo, CompilerOpts). + +shared_fun_wrappers(Opts, Dict) -> + case proplists:get_bool(no_shared_fun_wrappers, Opts) of + false -> + %% The compiler in OTP 23 depends on the on the loader + %% using the new indices in funs and being able to have + %% multiple make_fun2 instructions referring to the same + %% fun entry. Artificially set the highest opcode for the + %% module to ensure that it cannot be loaded in OTP 22 + %% and earlier. + Swap = beam_opcodes:opcode(swap, 2), + beam_dict:opcode(Swap, Dict); + true -> + %% Fun wrappers are not shared for compatibility with a + %% previous OTP release. + Dict + end. on_load(Fs0, Attr0) -> case proplists:get_value(on_load, Attr0) of diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl index 707974b2c1..a734ca3a10 100644 --- a/lib/compiler/src/beam_block.erl +++ b/lib/compiler/src/beam_block.erl @@ -33,8 +33,9 @@ module({Mod,Exp,Attr,Fs0,Lc}, _Opts) -> function({function,Name,Arity,CLabel,Is0}) -> try - Is1 = blockify(Is0), - Is = embed_lines(Is1), + Is1 = swap_opt(Is0), + Is2 = blockify(Is1), + Is = embed_lines(Is2), {function,Name,Arity,CLabel,Is} catch Class:Error:Stack -> @@ -42,6 +43,40 @@ function({function,Name,Arity,CLabel,Is0}) -> erlang:raise(Class, Error, Stack) end. +%%% +%%% Try to use a `swap` instruction instead of a sequence of moves. +%%% +%%% Note that beam_ssa_codegen generates `swap` instructions only for +%%% the moves within a single SSA instruction (such as `call`), not +%%% for the moves generated by a sequence of SSA instructions. +%%% Therefore, this optimization is needed. +%%% + +swap_opt([{move,Reg1,{x,X}=Temp}=Move1, + {move,Reg2,Reg1}=Move2, + {move,Temp,Reg2}=Move3|Is]) when Reg1 =/= Temp -> + case is_unused(X, Is) of + true -> + [{swap,Reg1,Reg2}|swap_opt(Is)]; + false -> + [Move1|swap_opt([Move2,Move3|Is])] + end; +swap_opt([I|Is]) -> + [I|swap_opt(Is)]; +swap_opt([]) -> []. + +is_unused(X, [{call,A,_}|_]) when A =< X -> true; +is_unused(X, [{call_ext,A,_}|_]) when A =< X -> true; +is_unused(X, [{make_fun2,_,_,_,A}|_]) when A =< X -> true; +is_unused(X, [{move,Src,Dst}|Is]) -> + case {Src,Dst} of + {{x,X},_} -> false; + {_,{x,X}} -> true; + {_,_} -> is_unused(X, Is) + end; +is_unused(X, [{line,_}|Is]) -> is_unused(X, Is); +is_unused(_, _) -> false. + %% blockify(Instructions0) -> Instructions %% Collect sequences of instructions to basic blocks. %% Also do some simple optimations on instructions outside the blocks. diff --git a/lib/compiler/src/beam_call_types.erl b/lib/compiler/src/beam_call_types.erl new file mode 100644 index 0000000000..904d82a62d --- /dev/null +++ b/lib/compiler/src/beam_call_types.erl @@ -0,0 +1,551 @@ +%% +%% %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% +%% + +-module(beam_call_types). + +-include("beam_types.hrl"). + +-import(lists, [duplicate/2,foldl/3]). + +-export([will_succeed/3, types/3]). + +%% +%% Returns whether a call will succeed or not. +%% +%% Note that it only answers 'yes' for functions in the 'erlang' module as +%% calls to other modules may fail due to not being loaded, even if we consider +%% the module to be known. +%% + +-spec will_succeed(Mod, Func, ArgTypes) -> Result when + Mod :: atom(), + Func :: atom(), + ArgTypes :: [normal_type()], + Result :: yes | no | maybe. + +will_succeed(erlang, BoolOp, [LHS, RHS]) when BoolOp =:= 'and'; + BoolOp =:= 'or' -> + case {succeeds_if_type(LHS, beam_types:make_boolean()), + succeeds_if_type(RHS, beam_types:make_boolean())} of + {yes, yes} -> yes; + {no, _} -> no; + {_, no} -> no; + {_, _} -> maybe + end; +will_succeed(erlang, bit_size, [Arg]) -> + succeeds_if_type(Arg, #t_bitstring{}); +will_succeed(erlang, byte_size, [Arg]) -> + succeeds_if_type(Arg, #t_bitstring{}); +will_succeed(erlang, map_size, [Arg]) -> + succeeds_if_type(Arg, #t_map{}); +will_succeed(erlang, 'not', [Arg]) -> + succeeds_if_type(Arg, beam_types:make_boolean()); +will_succeed(erlang, setelement, [#t_integer{elements={Min,Max}}, + #t_tuple{exact=Exact,size=Size}, _]) -> + case Min >= 1 andalso Max =< Size of + true -> yes; + false when Exact -> no; + false -> maybe + end; +will_succeed(erlang, size, [Arg]) -> + succeeds_if_type(Arg, #t_bitstring{}); +will_succeed(erlang, tuple_size, [Arg]) -> + succeeds_if_type(Arg, #t_tuple{}); +will_succeed(Mod, Func, Args) -> + Arity = length(Args), + case erl_bifs:is_safe(Mod, Func, Arity) of + true -> + yes; + false -> + case erl_bifs:is_exit_bif(Mod, Func, Arity) of + true -> no; + false -> maybe + end + end. + +succeeds_if_type(ArgType, Required) -> + case beam_types:meet(ArgType, Required) of + ArgType -> yes; + none -> no; + _ -> maybe + end. + +%% +%% Returns the inferred return and argument types for known functions, and +%% whether it's safe to subtract argument types on failure. +%% +%% Note that the return type will be 'none' if we can statically determine that +%% the function will fail at runtime. +%% + +-spec types(Mod, Func, ArgTypes) -> {RetType, ArgTypes, CanSubtract} when + Mod :: atom(), + Func :: atom(), + ArgTypes :: [normal_type()], + RetType :: type(), + CanSubtract :: boolean(). + +%% Functions that only fail due to bad argument *types*, meaning it's safe to +%% subtract argument types on failure. +%% +%% Note that these are all from the erlang module; suitable functions in other +%% modules could fail due to the module not being loaded. +types(erlang, 'map_size', [_]) -> + sub_safe(#t_integer{}, [#t_map{}]); +types(erlang, 'tuple_size', [_]) -> + sub_safe(#t_integer{}, [#t_tuple{}]); +types(erlang, 'bit_size', [_]) -> + sub_safe(#t_integer{}, [#t_bitstring{}]); +types(erlang, 'byte_size', [_]) -> + sub_safe(#t_integer{}, [#t_bitstring{}]); +types(erlang, 'hd', [_]) -> + sub_safe(any, [cons]); +types(erlang, 'tl', [_]) -> + sub_safe(any, [cons]); +types(erlang, 'length', [_]) -> + sub_safe(#t_integer{}, [list]); +types(erlang, 'not', [_]) -> + Bool = beam_types:make_boolean(), + sub_safe(Bool, [Bool]); + +%% Boolean ops +types(erlang, 'and', [_,_]) -> + Bool = beam_types:make_boolean(), + sub_unsafe(Bool, [Bool, Bool]); +types(erlang, 'or', [_,_]) -> + Bool = beam_types:make_boolean(), + sub_unsafe(Bool, [Bool, Bool]); +types(erlang, 'xor', [_,_]) -> + Bool = beam_types:make_boolean(), + sub_unsafe(Bool, [Bool, Bool]); + +%% Bitwise ops +types(erlang, 'band', [_,_]=Args) -> + sub_unsafe(band_return_type(Args), [#t_integer{}, #t_integer{}]); +types(erlang, 'bor', [_,_]) -> + sub_unsafe(#t_integer{}, [#t_integer{}, #t_integer{}]); +types(erlang, 'bxor', [_,_]) -> + sub_unsafe(#t_integer{}, [#t_integer{}, #t_integer{}]); +types(erlang, 'bsl', [_,_]) -> + sub_unsafe(#t_integer{}, [#t_integer{}, #t_integer{}]); +types(erlang, 'bsr', [_,_]) -> + sub_unsafe(#t_integer{}, [#t_integer{}, #t_integer{}]); +types(erlang, 'bnot', [_]) -> + sub_unsafe(#t_integer{}, [#t_integer{}]); + +%% Fixed-type arithmetic +types(erlang, 'float', [_]) -> + sub_unsafe(float, [number]); +types(erlang, 'round', [_]) -> + sub_unsafe(#t_integer{}, [number]); +types(erlang, 'floor', [_]) -> + sub_unsafe(#t_integer{}, [number]); +types(erlang, 'ceil', [_]) -> + sub_unsafe(#t_integer{}, [number]); +types(erlang, 'trunc', [_]) -> + sub_unsafe(#t_integer{}, [number]); +types(erlang, '/', [_,_]) -> + sub_unsafe(float, [number, number]); +types(erlang, 'div', [_,_]) -> + sub_unsafe(#t_integer{}, [#t_integer{}, #t_integer{}]); +types(erlang, 'rem', [_,_]) -> + sub_unsafe(#t_integer{}, [#t_integer{}, #t_integer{}]); + +%% Mixed-type arithmetic; '+'/2 and friends are handled in the catch-all +%% clause for the 'erlang' module. +types(erlang, 'abs', [_]=Args) -> + mixed_arith_types(Args); + +%% List operations +types(erlang, '++', [LHS,RHS]) -> + %% `[] ++ RHS` yields RHS, even if RHS is not a list. + RetType = case {LHS, RHS} of + {cons, _} -> cons; + {_, cons} -> cons; + _ -> beam_types:join(list, RHS) + end, + sub_unsafe(RetType, [list, any]); +types(erlang, '--', [_,_]) -> + sub_unsafe(list, [list, list]); + +%% Misc ops. +types(erlang, 'binary_part', [_, _]) -> + PosLen = make_two_tuple(#t_integer{}, #t_integer{}), + Binary = #t_bitstring{unit=8}, + sub_unsafe(Binary, [Binary, PosLen]); +types(erlang, 'binary_part', [_, _, _]) -> + Binary = #t_bitstring{unit=8}, + sub_unsafe(Binary, [Binary, #t_integer{}, #t_integer{}]); +types(erlang, 'is_map_key', [_,_]) -> + sub_unsafe(beam_types:make_boolean(), [any,#t_map{}]); +types(erlang, 'map_get', [_,_]) -> + sub_unsafe(any, [any,#t_map{}]); +types(erlang, 'node', [_]) -> + sub_unsafe(#t_atom{}, [any]); +types(erlang, 'node', []) -> + sub_unsafe(#t_atom{}, []); +types(erlang, 'size', [_]) -> + sub_unsafe(#t_integer{}, [any]); +types(erlang, 'size', [_]) -> + sub_unsafe(#t_integer{}, [any]); + +%% Tuple element ops +types(erlang, element, [PosType, TupleType]) -> + Index = case PosType of + #t_integer{elements={Same,Same}} when is_integer(Same) -> + Same; + _ -> + 0 + end, + + RetType = case TupleType of + #t_tuple{size=Sz,elements=Es} when Index =< Sz, + Index >= 1 -> + beam_types:get_element_type(Index, Es); + _ -> + any + end, + + sub_unsafe(RetType, [#t_integer{}, #t_tuple{size=Index}]); +types(erlang, setelement, [PosType, TupleType, ArgType]) -> + RetType = case {PosType,TupleType} of + {#t_integer{elements={Index,Index}}, + #t_tuple{elements=Es0,size=Size}=T} when Index >= 1 -> + %% This is an exact index, update the type of said + %% element or return 'none' if it's known to be out of + %% bounds. + Es = beam_types:set_element_type(Index, ArgType, Es0), + case T#t_tuple.exact of + false -> + T#t_tuple{size=max(Index, Size),elements=Es}; + true when Index =< Size -> + T#t_tuple{elements=Es}; + true -> + none + end; + {#t_integer{elements={Min,Max}}, + #t_tuple{elements=Es0,size=Size}=T} when Min >= 1 -> + %% We know this will land between Min and Max, so kill + %% the types for those indexes. + Es = discard_tuple_element_info(Min, Max, Es0), + case T#t_tuple.exact of + false -> + T#t_tuple{elements=Es,size=max(Min, Size)}; + true when Min =< Size -> + T#t_tuple{elements=Es,size=Size}; + true -> + none + end; + {_,#t_tuple{}=T} -> + %% Position unknown, so we have to discard all element + %% information. + T#t_tuple{elements=#{}}; + {#t_integer{elements={Min,_Max}},_} -> + #t_tuple{size=Min}; + {_,_} -> + #t_tuple{} + end, + sub_unsafe(RetType, [#t_integer{}, #t_tuple{}, any]); + +types(erlang, make_fun, [_,_,Arity0]) -> + Type = case Arity0 of + #t_integer{elements={Arity,Arity}} when Arity >= 0 -> + #t_fun{arity=Arity}; + _ -> + #t_fun{} + end, + sub_unsafe(Type, [#t_atom{}, #t_atom{}, #t_integer{}]); + +types(erlang, Name, Args) -> + Arity = length(Args), + + case erl_bifs:is_exit_bif(erlang, Name, Arity) of + true -> + {none, Args, false}; + false -> + case erl_internal:arith_op(Name, Arity) of + true -> + mixed_arith_types(Args); + false -> + IsTest = + erl_internal:new_type_test(Name, Arity) orelse + erl_internal:comp_op(Name, Arity), + + RetType = case IsTest of + true -> beam_types:make_boolean(); + false -> any + end, + + sub_unsafe(RetType, duplicate(Arity, any)) + end + end; + +%% +%% Math BIFs +%% + +types(math, cos, [_]) -> + sub_unsafe(float, [number]); +types(math, cosh, [_]) -> + sub_unsafe(float, [number]); +types(math, sin, [_]) -> + sub_unsafe(float, [number]); +types(math, sinh, [_]) -> + sub_unsafe(float, [number]); +types(math, tan, [_]) -> + sub_unsafe(float, [number]); +types(math, tanh, [_]) -> + sub_unsafe(float, [number]); +types(math, acos, [_]) -> + sub_unsafe(float, [number]); +types(math, acosh, [_]) -> + sub_unsafe(float, [number]); +types(math, asin, [_]) -> + sub_unsafe(float, [number]); +types(math, asinh, [_]) -> + sub_unsafe(float, [number]); +types(math, atan, [_]) -> + sub_unsafe(float, [number]); +types(math, atanh, [_]) -> + sub_unsafe(float, [number]); +types(math, erf, [_]) -> + sub_unsafe(float, [number]); +types(math, erfc, [_]) -> + sub_unsafe(float, [number]); +types(math, exp, [_]) -> + sub_unsafe(float, [number]); +types(math, log, [_]) -> + sub_unsafe(float, [number]); +types(math, log2, [_]) -> + sub_unsafe(float, [number]); +types(math, log10, [_]) -> + sub_unsafe(float, [number]); +types(math, sqrt, [_]) -> + sub_unsafe(float, [number]); +types(math, atan2, [_,_]) -> + sub_unsafe(float, [number, number]); +types(math, pow, [_,_]) -> + sub_unsafe(float, [number, number]); +types(math, ceil, [_]) -> + sub_unsafe(float, [number]); +types(math, floor, [_]) -> + sub_unsafe(float, [number]); +types(math, fmod, [_,_]) -> + sub_unsafe(float, [number, number]); +types(math, pi, []) -> + sub_unsafe(float, []); + +%% +%% List functions +%% + +%% Operator aliases. +types(lists, append, [_,_]=Args) -> + types(erlang, '++', Args); +types(lists, append, [_]) -> + %% This is implemented through folding the list over erlang:'++'/2, so it + %% can hypothetically return anything, but we can infer that its argument + %% is a list on success. + sub_unsafe(any, [list]); +types(lists, subtract, [_,_]) -> + sub_unsafe(list, [list, list]); + +%% Functions returning booleans. +types(lists, all, [_,_]) -> + sub_unsafe(beam_types:make_boolean(), [#t_fun{arity=1}, list]); +types(lists, any, [_,_]) -> + sub_unsafe(beam_types:make_boolean(), [#t_fun{arity=1}, list]); +types(lists, keymember, [_,_,_]) -> + sub_unsafe(beam_types:make_boolean(), [any, #t_integer{}, list]); +types(lists, member, [_,_]) -> + sub_unsafe(beam_types:make_boolean(), [any, list]); +types(lists, prefix, [_,_]) -> + sub_unsafe(beam_types:make_boolean(), [list, list]); +types(lists, suffix, [_,_]) -> + sub_unsafe(beam_types:make_boolean(), [list, list]); + +%% Functions returning plain lists. +types(lists, dropwhile, [_,_]) -> + sub_unsafe(list, [#t_fun{arity=1}, list]); +types(lists, duplicate, [_,_]) -> + sub_unsafe(list, [#t_integer{}, any]); +types(lists, filter, [_,_]) -> + sub_unsafe(list, [#t_fun{arity=1}, list]); +types(lists, flatten, [_]) -> + sub_unsafe(list, [list]); +types(lists, map, [_Fun, List]) -> + sub_unsafe(same_length_type(List), [#t_fun{arity=1}, list]); +types(lists, reverse, [List]) -> + sub_unsafe(same_length_type(List), [list]); +types(lists, sort, [List]) -> + sub_unsafe(same_length_type(List), [list]); +types(lists, takewhile, [_,_]) -> + sub_unsafe(list, [#t_fun{arity=1}, list]); +types(lists, usort, [List]) -> + sub_unsafe(same_length_type(List), [list]); +types(lists, zip, [A,B]) -> + ZipType = lists_zip_type([A,B]), + sub_unsafe(ZipType, [ZipType, ZipType]); +types(lists, zip3, [A,B,C]) -> + ZipType = lists_zip_type([A,B,C]), + sub_unsafe(ZipType, [ZipType, ZipType, ZipType]); +types(lists, zipwith, [_,A,B]) -> + ZipType = lists_zip_type([A,B]), + sub_unsafe(ZipType, [#t_fun{arity=2}, ZipType, ZipType]); +types(lists, zipwith3, [_,A,B,C]) -> + ZipType = lists_zip_type([A,B,C]), + sub_unsafe(ZipType, [#t_fun{arity=3}, ZipType, ZipType, ZipType]); + +%% Functions with complex return values. +types(lists, keyfind, [KeyType,PosType,_]) -> + TupleType = case PosType of + #t_integer{elements={Index,Index}} when is_integer(Index), + Index >= 1 -> + Es = beam_types:set_element_type(Index, KeyType, #{}), + #t_tuple{size=Index,elements=Es}; + _ -> + #t_tuple{} + end, + RetType = beam_types:join(TupleType, beam_types:make_atom(false)), + sub_unsafe(RetType, [any, #t_integer{}, list]); +types(lists, MapFold, [_Fun, _Init, List]) + when MapFold =:= mapfoldl; MapFold =:= mapfoldr -> + RetType = make_two_tuple(same_length_type(List), any), + sub_unsafe(RetType, [#t_fun{arity=2}, any, list]); +types(lists, partition, [_,_]) -> + sub_unsafe(make_two_tuple(list, list), [#t_fun{arity=1}, list]); +types(lists, search, [_,_]) -> + TupleType = make_two_tuple(beam_types:make_atom(value), any), + RetType = beam_types:join(TupleType, beam_types:make_atom(false)), + sub_unsafe(RetType, [#t_fun{arity=1}, list]); +types(lists, splitwith, [_,_]) -> + sub_unsafe(make_two_tuple(list, list), [#t_fun{arity=1}, list]); +types(lists, unzip, [List]) -> + ListType = same_length_type(List), + RetType = make_two_tuple(ListType, ListType), + sub_unsafe(RetType, [list]); + +%% Catch-all clause for unknown functions. + +types(_, _, Args) -> + sub_unsafe(any, [any || _ <- Args]). + +%% +%% Helpers +%% + +sub_unsafe(none, ArgTypes) -> + %% This is known to fail at runtime, but the type optimization pass + %% doesn't yet support cutting a block short at any point, so we + %% pretend it's raining instead. + %% + %% Actual exit BIFs get special treatment in the catch-all clause + %% for the 'erlang' module. + sub_unsafe(any, ArgTypes); +sub_unsafe(RetType, ArgTypes) -> + {RetType, ArgTypes, false}. + +sub_safe(RetType, ArgTypes) -> + {RetType, ArgTypes, true}. + +mixed_arith_types([FirstType | _]=Args0) -> + RetType = foldl(fun(#t_integer{}, #t_integer{}) -> #t_integer{}; + (#t_integer{}, number) -> number; + (#t_integer{}, float) -> float; + (float, #t_integer{}) -> float; + (float, number) -> float; + (float, float) -> float; + (number, #t_integer{}) -> number; + (number, float) -> float; + (number, number) -> number; + (any, _) -> number; + (_, _) -> none + end, FirstType, Args0), + sub_unsafe(RetType, [number || _ <- Args0]). + +band_return_type([#t_integer{elements={Int,Int}}, RHS]) when is_integer(Int) -> + band_return_type_1(RHS, Int); +band_return_type([LHS, #t_integer{elements={Int,Int}}]) when is_integer(Int) -> + band_return_type_1(LHS, Int); +band_return_type(_) -> + #t_integer{}. + +band_return_type_1(LHS, Int) -> + case LHS of + #t_integer{elements={Min0,Max0}} when Max0 - Min0 < 1 bsl 256 -> + {Intersection, Union} = range_masks(Min0, Max0), + + Min = Intersection band Int, + Max = min(Max0, Union band Int), + + #t_integer{elements={Min,Max}}; + _ when Int >= 0 -> + %% The range is either unknown or too wide, conservatively assume + %% that the new range is 0 .. Int. + #t_integer{elements={0,Int}}; + _ when Int < 0 -> + %% We can't infer boundaries when the range is unknown and the + %% other operand is a negative number, as the latter sign-extends + %% to infinity and we can't express an inverted range at the + %% moment (cf. X band -8; either less than -7 or greater than 7). + #t_integer{} + end. + +%% Returns two bitmasks describing all possible values between From and To. +%% +%% The first contains the bits that are common to all values, and the second +%% contains the bits that are set by any value in the range. +range_masks(From, To) when From =< To -> + range_masks_1(From, To, 0, -1, 0). + +range_masks_1(From, To, BitPos, Intersection, Union) when From < To -> + range_masks_1(From + (1 bsl BitPos), To, BitPos + 1, + Intersection band From, Union bor From); +range_masks_1(_From, To, _BitPos, Intersection0, Union0) -> + Intersection = To band Intersection0, + Union = To bor Union0, + {Intersection, Union}. + +discard_tuple_element_info(Min, Max, Es) -> + foldl(fun(El, Acc) when Min =< El, El =< Max -> + maps:remove(El, Acc); + (_El, Acc) -> Acc + end, Es, maps:keys(Es)). + +%% For a lists function that return a list of the same length as the input +%% list, return the type of the list. +same_length_type(cons) -> cons; +same_length_type(nil) -> nil; +same_length_type(_) -> list. + +%% lists:zip/2 and friends only succeed when all arguments have the same +%% length, so if one of them is cons, we can infer that all of them are cons +%% on success. +lists_zip_type(Types) -> + foldl(fun(cons, _) -> cons; + (_, cons) -> cons; + (nil, _) -> nil; + (_, T) -> T + end, list, Types). + +make_two_tuple(Type1, Type2) -> + Es0 = beam_types:set_element_type(1, Type1, #{}), + Es = beam_types:set_element_type(2, Type2, Es0), + #t_tuple{size=2,exact=true,elements=Es}. diff --git a/lib/compiler/src/beam_clean.erl b/lib/compiler/src/beam_clean.erl index 7299654476..6b2b2ce085 100644 --- a/lib/compiler/src/beam_clean.erl +++ b/lib/compiler/src/beam_clean.erl @@ -34,7 +34,8 @@ module({Mod,Exp,Attr,Fs0,_}, Opts) -> 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), + Fs3 = fix_swap(Fs2, Opts), + Fs = maybe_remove_lines(Fs3, Opts), {ok,{Mod,Exp,Attr,Fs,Lc}}. %% Determine the rootset, i.e. exported functions and @@ -137,31 +138,54 @@ function_replace([{function,Name,Arity,Entry,Asm0}|Fs], Dict, Acc) -> function_replace([], _, Acc) -> Acc. %%% +%%% If compatibility with a previous release (OTP 22 or earlier) has +%%% been requested, replace swap instructions with a sequence of moves. +%%% + +fix_swap(Fs, Opts) -> + case proplists:get_bool(no_swap, Opts) of + false -> Fs; + true -> fold_functions(fun swap_moves/1, Fs) + end. + +swap_moves([{swap,Reg1,Reg2}|Is]) -> + Temp = {x,1022}, + [{move,Reg1,Temp},{move,Reg2,Reg1},{move,Temp,Reg2}|swap_moves(Is)]; +swap_moves([I|Is]) -> + [I|swap_moves(Is)]; +swap_moves([]) -> []. + +%%% %%% Remove line instructions if requested. %%% maybe_remove_lines(Fs, Opts) -> case proplists:get_bool(no_line_info, Opts) of false -> Fs; - true -> remove_lines(Fs) + true -> fold_functions(fun remove_lines/1, Fs) end. -remove_lines([{function,N,A,Lbl,Is0}|T]) -> - Is = remove_lines_fun(Is0), - [{function,N,A,Lbl,Is}|remove_lines(T)]; -remove_lines([]) -> []. - -remove_lines_fun([{line,_}|Is]) -> - remove_lines_fun(Is); -remove_lines_fun([{block,Bl0}|Is]) -> +remove_lines([{line,_}|Is]) -> + remove_lines(Is); +remove_lines([{block,Bl0}|Is]) -> Bl = remove_lines_block(Bl0), - [{block,Bl}|remove_lines_fun(Is)]; -remove_lines_fun([I|Is]) -> - [I|remove_lines_fun(Is)]; -remove_lines_fun([]) -> []. + [{block,Bl}|remove_lines(Is)]; +remove_lines([I|Is]) -> + [I|remove_lines(Is)]; +remove_lines([]) -> []. remove_lines_block([{set,_,_,{line,_}}|Is]) -> remove_lines_block(Is); remove_lines_block([I|Is]) -> [I|remove_lines_block(Is)]; remove_lines_block([]) -> []. + + +%%% +%%% Helpers. +%%% + +fold_functions(F, [{function,N,A,Lbl,Is0}|T]) -> + Is = F(Is0), + [{function,N,A,Lbl,Is}|fold_functions(F, T)]; +fold_functions(_F, []) -> []. diff --git a/lib/compiler/src/beam_dict.erl b/lib/compiler/src/beam_dict.erl index b2056332e6..4d0cec857d 100644 --- a/lib/compiler/src/beam_dict.erl +++ b/lib/compiler/src/beam_dict.erl @@ -40,6 +40,7 @@ -type lambda_info() :: {label(),{index(),label(),non_neg_integer()}}. -type lambda_tab() :: {non_neg_integer(),[lambda_info()]}. +-type wrapper() :: #{label() => index()}. -record(asm, {atoms = #{} :: atom_tab(), @@ -48,6 +49,7 @@ imports = gb_trees:empty() :: import_tab(), strings = <<>> :: binary(), %String pool lambdas = {0,[]} :: lambda_tab(), + wrappers = #{} :: wrapper(), literals = dict:new() :: literal_tab(), fnames = #{} :: fname_tab(), lines = #{} :: line_tab(), @@ -147,11 +149,21 @@ string(BinString, Dict) when is_binary(BinString) -> -spec lambda(label(), non_neg_integer(), bdict()) -> {non_neg_integer(), bdict()}. -lambda(Lbl, NumFree, #asm{lambdas={OldIndex,Lambdas0}}=Dict) -> - %% Set Index the same as OldIndex. - Index = OldIndex, - Lambdas = [{Lbl,{Index,Lbl,NumFree}}|Lambdas0], - {OldIndex,Dict#asm{lambdas={OldIndex+1,Lambdas}}}. +lambda(Lbl, NumFree, #asm{wrappers=Wrappers0, + lambdas={OldIndex,Lambdas0}}=Dict) -> + case Wrappers0 of + #{Lbl:=Index} -> + %% OTP 23: There old is a fun entry for this wrapper function. + %% Share the fun entry. + {Index,Dict}; + #{} -> + %% Set Index the same as OldIndex. + Index = OldIndex, + Wrappers = Wrappers0#{Lbl=>Index}, + Lambdas = [{Lbl,{Index,Lbl,NumFree}}|Lambdas0], + {OldIndex,Dict#asm{wrappers=Wrappers, + lambdas={OldIndex+1,Lambdas}}} + end. %% Returns the index for a literal (adding it to the literal table if necessary). %% literal(Literal, Dict) -> {Index,Dict'} diff --git a/lib/compiler/src/beam_disasm.erl b/lib/compiler/src/beam_disasm.erl index 7d048716e4..45b69d7e95 100644 --- a/lib/compiler/src/beam_disasm.erl +++ b/lib/compiler/src/beam_disasm.erl @@ -1123,6 +1123,13 @@ resolve_inst({put_tuple2,[Dst,{{z,1},{u,_},List0}]},_,_,_) -> {put_tuple2,Dst,{list,List}}; %% +%% OTP 23. +%% +resolve_inst({swap,[_,_]=List},_,_,_) -> + [R1,R2] = resolve_args(List), + {swap,R1,R2}; + +%% %% Catches instructions that are not yet handled. %% resolve_inst(X,_,_,_) -> ?exit({resolve_inst,X}). diff --git a/lib/compiler/src/beam_except.erl b/lib/compiler/src/beam_except.erl deleted file mode 100644 index 2b9c1b0cf5..0000000000 --- a/lib/compiler/src/beam_except.erl +++ /dev/null @@ -1,255 +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_except). --export([module/2]). - -%%% Rewrite certain calls to erlang:error/{1,2} to specialized -%%% instructions: -%%% -%%% erlang:error({badmatch,Value}) => badmatch Value -%%% erlang:error({case_clause,Value}) => case_end Value -%%% erlang:error({try_clause,Value}) => try_case_end Value -%%% erlang:error(if_clause) => if_end -%%% erlang:error(function_clause, Args) => jump FuncInfoLabel -%%% - --import(lists, [reverse/1,reverse/2,seq/2,splitwith/2]). - --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 = function_1(Is0), - {function,Name,Arity,CLabel,Is} - catch - Class:Error:Stack -> - io:fwrite("Function: ~w/~w\n", [Name,Arity]), - erlang:raise(Class, Error, Stack) - end. - --record(st, - {lbl :: beam_asm:label(), %func_info label - loc :: [_], %location for func_info - arity :: arity() %arity for function - }). - -function_1(Is0) -> - case Is0 of - [{label,Lbl},{line,Loc},{func_info,_,_,Arity}|_] -> - St = #st{lbl=Lbl,loc=Loc,arity=Arity}, - translate(Is0, St, []); - [{label,_}|_] -> - %% No line numbers. The source must be a .S file. - %% There is no need to do anything. - Is0 - end. - -translate([{call_ext,Ar,{extfunc,erlang,error,Ar}}=I|Is], St, Acc) -> - translate_1(Ar, I, Is, St, Acc); -translate([I|Is], St, Acc) -> - translate(Is, St, [I|Acc]); -translate([], _, Acc) -> - reverse(Acc). - -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,Acc2} -> - case {Is,Line,St} of - {[return|_],{line,Loc},#st{lbl=Fi,loc=Loc}} -> - Instr = {jump,{f,Fi}}, - translate(Is, St, [Instr|Acc2]); - {_,_,_} -> - %% 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(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,[]} -> - {yes,What,Is}; - {yes,What,Bl} -> - {yes,What,[{block,Bl}|Is]} - end; -dig_out(_) -> no. - -dig_out_block([{set,[{x,0}],[{atom,if_clause}],move}]) -> - {yes,if_end,[]}; -dig_out_block([{set,[{x,0}],[{literal,{Exc,Value}}],move}|Is]) -> - translate_exception(Exc, {literal,Value}, Is, 0); -dig_out_block([{set,[{x,0}],[{atom,Exc},Value],put_tuple2}|Is]) -> - translate_exception(Exc, Value, Is, 3); -dig_out_block(_) -> no. - -translate_exception(badmatch, Val, Is, Words) -> - {yes,{badmatch,Val},fix_block(Is, Words)}; -translate_exception(case_clause, Val, Is, Words) -> - {yes,{case_end,Val},fix_block(Is, Words)}; -translate_exception(try_clause, Val, Is, Words) -> - {yes,{try_case_end,Val},fix_block(Is, Words)}; -translate_exception(_, _, _, _) -> no. - -fix_block(Is, 0) -> - reverse(Is); -fix_block(Is, Words) -> - reverse(fix_block_1(Is, Words)). - -fix_block_1([{set,[],[],{alloc,Live,{F1,F2,Needed0,F3}}}|Is], Words) -> - case Needed0 - Words of - 0 -> - Is; - Needed -> - true = Needed >= 0, %Assertion. - [{set,[],[],{alloc,Live,{F1,F2,Needed,F3}}}|Is] - end; -fix_block_1([I|Is], Words) -> - [I|fix_block_1(Is, Words)]; -fix_block_1([], _Words) -> - %% Rare. The heap allocation was probably done by a binary - %% construction instruction. - []. - -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), Arity, Regs0, Acc0), - case Regs of - #{{x,0}:={atom,function_clause},{x,1}:=Args} -> - case moves_from_stack(Args, 0, []) of - {Moves,Arity} -> - {yes,function_clause,reverse(Moves, Acc)}; - {_,_} -> - no - end; - #{} -> - no - end. - -dig_out_fc_1([{block,Bl}|Is], Arity, Regs0, Acc) -> - Regs = dig_out_fc_block(Bl, Regs0), - dig_out_fc_1(Is, Arity, Regs, Acc); -dig_out_fc_1([{bs_set_position,_,_}=I|Is], Arity, Regs, Acc) -> - dig_out_fc_1(Is, Arity, Regs, [I|Acc]); -dig_out_fc_1([{bs_get_tail,Src,Dst,Live0}|Is], Arity, Regs0, Acc) -> - case Src of - {x,X} when X < Arity -> - %% The heuristic for determining the number of live - %% registers is likely to give an incorrect result. - %% Give up. - {#{},[]}; - _ -> - Regs = prune_xregs(Live0, Regs0), - Live = dig_out_stack_live(Regs, Live0), - I = {bs_get_tail,Src,Dst,Live}, - dig_out_fc_1(Is, Arity, Regs, [I|Acc]) - end; -dig_out_fc_1([_|_], _Arity, _Regs, _Acc) -> - {#{},[]}; -dig_out_fc_1([], _Arity, 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); -moves_from_stack(_, _, _) -> - %% Not understood. Give up. - {[],-1}. - - -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_kernel_to_ssa.erl b/lib/compiler/src/beam_kernel_to_ssa.erl index df95749fb3..2406a634e6 100644 --- a/lib/compiler/src/beam_kernel_to_ssa.erl +++ b/lib/compiler/src/beam_kernel_to_ssa.erl @@ -34,7 +34,7 @@ -type label() :: beam_ssa:label(). %% Main codegen structure. --record(cg, {lcount=1 :: label(), %Label counter +-record(cg, {lcount=1 :: label(), %Label counter bfail=1 :: label(), catch_label=none :: 'none' | label(), vars=#{} :: map(), %Defined variables. @@ -83,6 +83,7 @@ function(#k_fdef{anno=Anno0,func=Name,arity=Arity, cg_fun(Ke, St0) -> {UltimateFail,FailIs,St1} = make_failure(badarg, St0), + ?EXCEPTION_BLOCK = UltimateFail, %Assertion. St2 = St1#cg{bfail=UltimateFail,ultimate_failure=UltimateFail}, {B,St} = cg(Ke, St2), Asm = [{label,0}|B++FailIs], @@ -279,7 +280,7 @@ select_binary(#k_val_clause{val=#k_binary{segs=#k_var{name=Ctx0}},body=B}, #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), + {TestIs,St} = make_succeeded(Ctx, {guard, Tf}, St2), Bis1 = [#b_set{op=bs_start_match,dst=Ctx, args=[ssa_arg(Src, St)]}] ++ TestIs ++ Bis0, Bis = finish_bs_matching(Bis1), @@ -311,6 +312,35 @@ make_cond_branch(Cond, Args, Fail, St0) -> make_uncond_branch(Fail) -> #b_br{bool=#b_literal{val=true},succ=Fail,fail=Fail}. +%% +%% The 'succeeded' instruction needs special treatment in catch blocks to +%% prevent the checked operation from being optimized away if a later pass +%% determines that it always fails. +%% + +make_succeeded(Var, {in_catch, CatchLbl}, St0) -> + {Bool, St1} = new_ssa_var('@ssa_bool', St0), + {Succ, St2} = new_label(St1), + {Fail, St} = new_label(St2), + + Check = [#b_set{op=succeeded,dst=Bool,args=[Var]}, + #b_br{bool=Bool,succ=Succ,fail=Fail}], + + %% Add a dummy block that references the checked variable, ensuring it + %% stays alive and that it won't be merged with the landing pad. + Trampoline = [{label,Fail}, + #b_set{op=exception_trampoline,args=[Var]}, + make_uncond_branch(CatchLbl)], + + {Check ++ Trampoline ++ [{label,Succ}], St}; +make_succeeded(Var, {no_catch, Fail}, St) -> + %% Ultimate failure raises an exception, so we must treat it as if it were + %% in a catch to keep it from being optimized out. + #cg{ultimate_failure=Fail} = St, %Assertion + make_succeeded(Var, {in_catch, Fail}, St); +make_succeeded(Var, {guard, Fail}, St) -> + make_cond_branch(succeeded, [Var], Fail, St). + %% Instructions for selection of binary segments. select_bin_segs(Scs, Ivar, Tf, St) -> @@ -394,7 +424,7 @@ select_extract_int(#k_var{name=Tl}, Val, #k_int{val=Sz}, U, Fs, Vf, <<Val:Bits/little>> end, Bits = bit_size(Bin), %Assertion. - {TestIs,St} = make_cond_branch(succeeded, [Dst], Vf, St1), + {TestIs,St} = make_succeeded(Dst, {guard, Vf}, St1), Set = #b_set{op=bs_match,dst=Dst, args=[#b_literal{val=string},Ctx,#b_literal{val=Bin}]}, {[Set|TestIs],St}. @@ -412,7 +442,7 @@ build_bs_instr(Anno, Type, Fail, Ctx, Size, Unit0, Flags0, Dst, St0) -> #b_set{anno=Anno,op=bs_match,dst=Dst, args=[TypeArg,Ctx,Flags]} end, - {Is,St} = make_cond_branch(succeeded, [Dst], Fail, St0), + {Is,St} = make_succeeded(Dst, {guard, Fail}, St0), {[Get|Is],St}. select_val(#k_val_clause{val=#k_tuple{es=Es},body=B}, V, Vf, St0) -> @@ -475,7 +505,7 @@ select_extract_map([P|Ps], Src, Fail, St0) -> Key = ssa_arg(Key0, St0), {Dst,St1} = new_ssa_var(Dst0, St0), Set = #b_set{op=get_map_element,dst=Dst,args=[MapSrc,Key]}, - {TestIs,St2} = make_cond_branch(succeeded, [Dst], Fail, St1), + {TestIs,St2} = make_succeeded(Dst, {guard, Fail}, St1), {Is,St} = select_extract_map(Ps, Src, Fail, St2), {[Set|TestIs]++Is,St}; select_extract_map([], _, _, St) -> @@ -596,7 +626,7 @@ match_fmf(F, LastFail, St0, [H|T]) -> {Rs,St3} = match_fmf(F, LastFail, St2, T), {R ++ [{label,Fail}] ++ Rs,St3}. -%% fail_label(State) -> {Where,FailureLabel}. +%% fail_context(State) -> {Where,FailureLabel}. %% Where = guard | no_catch | in_catch %% Return an indication of which part of a function code is %% being generated for and the appropriate failure label to @@ -609,7 +639,7 @@ match_fmf(F, LastFail, St0, [H|T]) -> %% a try/catch or catch. %% in_catch - In the scope of a try/catch or catch. -fail_label(#cg{catch_label=Catch,bfail=Fail,ultimate_failure=Ult}) -> +fail_context(#cg{catch_label=Catch,bfail=Fail,ultimate_failure=Ult}) -> if Fail =/= Ult -> {guard,Fail}; @@ -619,14 +649,6 @@ fail_label(#cg{catch_label=Catch,bfail=Fail,ultimate_failure=Ult}) -> {in_catch,Catch} end. -%% bif_fail_label(State) -> FailureLabel. -%% Return the appropriate failure label for a guard BIF call or -%% primop that fails. - -bif_fail_label(St) -> - {_,Fail} = fail_label(St), - Fail. - %% call_cg(Func, [Arg], [Ret], Le, State) -> %% {[Ainstr],State}. %% enter_cg(Func, [Arg], Le, St) -> {[Ainstr],St}. @@ -635,7 +657,7 @@ bif_fail_label(St) -> call_cg(Func, As, [], Le, St) -> call_cg(Func, As, [#k_var{name='@ssa_ignored'}], Le, St); call_cg(Func0, As, [#k_var{name=R}|MoreRs]=Rs, Le, St0) -> - case fail_label(St0) of + case fail_context(St0) of {guard,Fail} -> %% Inside a guard. The only allowed function call is to %% erlang:error/1,2. We will generate a branch to the @@ -645,7 +667,7 @@ call_cg(Func0, As, [#k_var{name=R}|MoreRs]=Rs, Le, St0) -> [#k_var{name=DestVar}] = Rs, St = set_ssa_var(DestVar, #b_literal{val=unused}, St0), {[make_uncond_branch(Fail),#cg_unreachable{}],St}; - {Catch,Fail} -> + FailCtx -> %% Ordinary function call in a function body. Args = ssa_args(As, St0), {Ret,St1} = new_ssa_var(R, St0), @@ -657,13 +679,9 @@ call_cg(Func0, As, [#k_var{name=R}|MoreRs]=Rs, Le, St0) -> St2 = foldl(fun(#k_var{name=Dummy}, S) -> set_ssa_var(Dummy, #b_literal{val=unused}, S) end, St1, MoreRs), - case Catch of - no_catch -> - {[Call],St2}; - in_catch -> - {TestIs,St} = make_cond_branch(succeeded, [Ret], Fail, St2), - {[Call|TestIs],St} - end + + {TestIs,St} = make_succeeded(Ret, FailCtx, St2), + {[Call|TestIs],St} end. enter_cg(Func0, As0, Le, St0) -> @@ -748,8 +766,8 @@ bif_cg(Bif, As0, [#k_var{name=Dst0}], Le, St0) -> I = #b_set{anno=line_anno(Le),op={bif,Bif},dst=Dst,args=As}, case erl_bifs:is_safe(erlang, Bif, length(As)) of false -> - Fail = bif_fail_label(St1), - {Is,St} = make_cond_branch(succeeded, [Dst], Fail, St1), + FailCtx = fail_context(St1), + {Is,St} = make_succeeded(Dst, FailCtx, St1), {[I|Is],St}; true-> {[I],St1} @@ -797,7 +815,7 @@ cg_recv_mesg(#k_var{name=R}, Rm, Tl, Le, St0) -> {Dst,St1} = new_ssa_var(R, St0), {Mis,St2} = match_cg(Rm, none, St1), RecvLbl = St1#cg.recv, - {TestIs,St} = make_cond_branch(succeeded, [Dst], Tl, St2), + {TestIs,St} = make_succeeded(Dst, {guard, Tl}, St2), Is = [#b_br{anno=line_anno(Le),bool=#b_literal{val=true}, succ=RecvLbl,fail=RecvLbl}, {label,RecvLbl}, @@ -813,7 +831,7 @@ cg_recv_wait(Te, Es, St0) -> {Tis,St1} = cg(Es, St0), Args = [ssa_arg(Te, St1)], {WaitDst,St2} = new_ssa_var('@ssa_wait', St1), - {WaitIs,St} = make_cond_branch(succeeded, [WaitDst], St1#cg.recv, St2), + {WaitIs,St} = make_succeeded(WaitDst, {guard, St1#cg.recv}, St2), %% Infinite timeout will be optimized later. Is = [#b_set{op=wait_timeout,dst=WaitDst,args=Args}] ++ WaitIs ++ [#b_set{op=timeout}] ++ Tis, @@ -924,9 +942,9 @@ put_cg([#k_var{name=R}], #k_tuple{es=Es}, _Le, St0) -> PutTuple = #b_set{op=put_tuple,dst=Ret,args=Args}, {[PutTuple],St}; put_cg([#k_var{name=R}], #k_binary{segs=Segs}, Le, St0) -> - Fail = bif_fail_label(St0), + FailCtx = fail_context(St0), {Dst,St1} = new_ssa_var(R, St0), - cg_binary(Dst, Segs, Fail, Le, St1); + cg_binary(Dst, Segs, FailCtx, Le, St1); put_cg([#k_var{name=R}], #k_map{op=Op,var=Map, es=[#k_map_pair{key=#k_var{}=K,val=V}]}, Le, St0) -> @@ -955,14 +973,14 @@ put_cg([#k_var{name=R}], Con0, _Le, St0) -> {[],St}. put_cg_map(LineAnno, Op, SrcMap, Dst, List, St0) -> - Fail = bif_fail_label(St0), Args = [#b_literal{val=Op},SrcMap|List], PutMap = #b_set{anno=LineAnno,op=put_map,dst=Dst,args=Args}, if Op =:= assoc -> {[PutMap],St0}; true -> - {Is,St} = make_cond_branch(succeeded, [Dst], Fail, St0), + FailCtx = fail_context(St0), + {Is,St} = make_succeeded(Dst, FailCtx, St0), {[PutMap|Is],St} end. @@ -970,8 +988,8 @@ put_cg_map(LineAnno, Op, SrcMap, Dst, List, St0) -> %%% Code generation for constructing binaries. %%% -cg_binary(Dst, Segs0, Fail, Le, St0) -> - {PutCode0,SzCalc0,St1} = cg_bin_put(Segs0, Fail, St0), +cg_binary(Dst, Segs0, FailCtx, Le, St0) -> + {PutCode0,SzCalc0,St1} = cg_bin_put(Segs0, FailCtx, St0), LineAnno = line_anno(Le), Anno = Le#k.a, case PutCode0 of @@ -980,8 +998,8 @@ cg_binary(Dst, Segs0, Fail, Le, St0) -> {label,_}|_] -> #k_bin_seg{unit=Unit0,next=Segs} = Segs0, Unit = #b_literal{val=Unit0}, - {PutCode,SzCalc1,St2} = cg_bin_put(Segs, Fail, St1), - {_,SzVar,SzCode0,St3} = cg_size_calc(1, SzCalc1, Fail, St2), + {PutCode,SzCalc1,St2} = cg_bin_put(Segs, FailCtx, St1), + {_,SzVar,SzCode0,St3} = cg_size_calc(1, SzCalc1, FailCtx, St2), SzCode = cg_bin_anno(SzCode0, LineAnno), Args = case member(single_use, Anno) of true -> @@ -990,14 +1008,14 @@ cg_binary(Dst, Segs0, Fail, Le, St0) -> [#b_literal{val=append},Src,SzVar,Unit] end, BsInit = #b_set{anno=LineAnno,op=bs_init,dst=Dst,args=Args}, - {TestIs,St} = make_cond_branch(succeeded, [Dst], Fail, St3), + {TestIs,St} = make_succeeded(Dst, FailCtx, St3), {SzCode ++ [BsInit] ++ TestIs ++ PutCode,St}; [#b_set{op=bs_put}|_] -> - {Unit,SzVar,SzCode0,St2} = cg_size_calc(8, SzCalc0, Fail, St1), + {Unit,SzVar,SzCode0,St2} = cg_size_calc(8, SzCalc0, FailCtx, St1), SzCode = cg_bin_anno(SzCode0, LineAnno), Args = [#b_literal{val=new},SzVar,Unit], BsInit = #b_set{anno=LineAnno,op=bs_init,dst=Dst,args=Args}, - {TestIs,St} = make_cond_branch(succeeded, [Dst], Fail, St2), + {TestIs,St} = make_succeeded(Dst, FailCtx, St2), {SzCode ++ [BsInit] ++ TestIs ++ PutCode0,St} end. @@ -1005,18 +1023,18 @@ cg_bin_anno([Set|Sets], Anno) -> [Set#b_set{anno=Anno}|Sets]; cg_bin_anno([], _) -> []. -%% cg_size_calc(PreferredUnit, SzCalc, Fail, St0) -> +%% cg_size_calc(PreferredUnit, SzCalc, FailCtx, St0) -> %% {ActualUnit,SizeVariable,SizeCode,St}. %% Generate size calculation code. -cg_size_calc(Unit, error, _Fail, St) -> +cg_size_calc(Unit, error, _FailCtx, St) -> {#b_literal{val=Unit},#b_literal{val=badarg},[],St}; -cg_size_calc(8, [{1,_}|_]=SzCalc, Fail, St) -> - cg_size_calc(1, SzCalc, Fail, St); -cg_size_calc(8, SzCalc, Fail, St0) -> - {Var,Pre,St} = cg_size_calc_1(SzCalc, Fail, St0), +cg_size_calc(8, [{1,_}|_]=SzCalc, FailCtx, St) -> + cg_size_calc(1, SzCalc, FailCtx, St); +cg_size_calc(8, SzCalc, FailCtx, St0) -> + {Var,Pre,St} = cg_size_calc_1(SzCalc, FailCtx, St0), {#b_literal{val=8},Var,Pre,St}; -cg_size_calc(1, SzCalc0, Fail, St0) -> +cg_size_calc(1, SzCalc0, FailCtx, St0) -> SzCalc = map(fun({8,#b_literal{val=Size}}) -> {1,#b_literal{val=8*Size}}; ({8,{{bif,byte_size},Src}}) -> @@ -1026,54 +1044,54 @@ cg_size_calc(1, SzCalc0, Fail, St0) -> ({_,_}=Pair) -> Pair end, SzCalc0), - {Var,Pre,St} = cg_size_calc_1(SzCalc, Fail, St0), + {Var,Pre,St} = cg_size_calc_1(SzCalc, FailCtx, St0), {#b_literal{val=1},Var,Pre,St}. -cg_size_calc_1(SzCalc, Fail, St0) -> - cg_size_calc_2(SzCalc, #b_literal{val=0}, Fail, St0). +cg_size_calc_1(SzCalc, FailCtx, St0) -> + cg_size_calc_2(SzCalc, #b_literal{val=0}, FailCtx, St0). -cg_size_calc_2([{_,{'*',Unit,{_,_}=Bif}}|T], Sum0, Fail, St0) -> - {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, Fail, St0), - {BifDst,Pre1,St2} = cg_size_bif(Bif, Fail, St1), - {Sum,Pre2,St} = cg_size_add(Sum1, BifDst, Unit, Fail, St2), +cg_size_calc_2([{_,{'*',Unit,{_,_}=Bif}}|T], Sum0, FailCtx, St0) -> + {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, FailCtx, St0), + {BifDst,Pre1,St2} = cg_size_bif(Bif, FailCtx, St1), + {Sum,Pre2,St} = cg_size_add(Sum1, BifDst, Unit, FailCtx, St2), {Sum,Pre0++Pre1++Pre2,St}; -cg_size_calc_2([{_,#b_literal{}=Sz}|T], Sum0, Fail, St0) -> - {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, Fail, St0), - {Sum,Pre,St} = cg_size_add(Sum1, Sz, #b_literal{val=1}, Fail, St1), +cg_size_calc_2([{_,#b_literal{}=Sz}|T], Sum0, FailCtx, St0) -> + {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, FailCtx, St0), + {Sum,Pre,St} = cg_size_add(Sum1, Sz, #b_literal{val=1}, FailCtx, St1), {Sum,Pre0++Pre,St}; -cg_size_calc_2([{_,#b_var{}=Sz}|T], Sum0, Fail, St0) -> - {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, Fail, St0), - {Sum,Pre,St} = cg_size_add(Sum1, Sz, #b_literal{val=1}, Fail, St1), +cg_size_calc_2([{_,#b_var{}=Sz}|T], Sum0, FailCtx, St0) -> + {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, FailCtx, St0), + {Sum,Pre,St} = cg_size_add(Sum1, Sz, #b_literal{val=1}, FailCtx, St1), {Sum,Pre0++Pre,St}; -cg_size_calc_2([{_,{_,_}=Bif}|T], Sum0, Fail, St0) -> - {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, Fail, St0), - {BifDst,Pre1,St2} = cg_size_bif(Bif, Fail, St1), - {Sum,Pre2,St} = cg_size_add(Sum1, BifDst, #b_literal{val=1}, Fail, St2), +cg_size_calc_2([{_,{_,_}=Bif}|T], Sum0, FailCtx, St0) -> + {Sum1,Pre0,St1} = cg_size_calc_2(T, Sum0, FailCtx, St0), + {BifDst,Pre1,St2} = cg_size_bif(Bif, FailCtx, St1), + {Sum,Pre2,St} = cg_size_add(Sum1, BifDst, #b_literal{val=1}, FailCtx, St2), {Sum,Pre0++Pre1++Pre2,St}; -cg_size_calc_2([], Sum, _Fail, St) -> +cg_size_calc_2([], Sum, _FailCtx, St) -> {Sum,[],St}. -cg_size_bif(#b_var{}=Var, _Fail, St) -> +cg_size_bif(#b_var{}=Var, _FailCtx, St) -> {Var,[],St}; -cg_size_bif({Name,Src}, Fail, St0) -> +cg_size_bif({Name,Src}, FailCtx, St0) -> {Dst,St1} = new_ssa_var('@ssa_bif', St0), Bif = #b_set{op=Name,dst=Dst,args=[Src]}, - {TestIs,St} = make_cond_branch(succeeded, [Dst], Fail, St1), + {TestIs,St} = make_succeeded(Dst, FailCtx, St1), {Dst,[Bif|TestIs],St}. -cg_size_add(#b_literal{val=0}, Val, #b_literal{val=1}, _Fail, St) -> +cg_size_add(#b_literal{val=0}, Val, #b_literal{val=1}, _FailCtx, St) -> {Val,[],St}; -cg_size_add(A, B, Unit, Fail, St0) -> +cg_size_add(A, B, Unit, FailCtx, St0) -> {Dst,St1} = new_ssa_var('@ssa_sum', St0), - {TestIs,St} = make_cond_branch(succeeded, [Dst], Fail, St1), + {TestIs,St} = make_succeeded(Dst, FailCtx, St1), BsAdd = #b_set{op=bs_add,dst=Dst,args=[A,B,Unit]}, {Dst,[BsAdd|TestIs],St}. -cg_bin_put(Seg, Fail, St) -> - cg_bin_put_1(Seg, Fail, [], [], St). +cg_bin_put(Seg, FailCtx, St) -> + cg_bin_put_1(Seg, FailCtx, [], [], St). cg_bin_put_1(#k_bin_seg{size=Size0,unit=U,type=T,flags=Fs,seg=Src0,next=Next}, - Fail, Acc, SzCalcAcc, St0) -> + FailCtx, Acc, SzCalcAcc, St0) -> [Src,Size] = ssa_args([Src0,Size0], St0), NeedSize = bs_need_size(T), TypeArg = #b_literal{val=T}, @@ -1083,9 +1101,12 @@ cg_bin_put_1(#k_bin_seg{size=Size0,unit=U,type=T,flags=Fs,seg=Src0,next=Next}, true -> [TypeArg,Flags,Src,Size,Unit]; false -> [TypeArg,Flags,Src] end, - {Is,St} = make_cond_branch(bs_put, Args, Fail, St0), + %% bs_put has its own 'succeeded' logic, and should always jump directly to + %% the fail label regardless of whether it's in a catch or not. + {_, FailLbl} = FailCtx, + {Is,St} = make_cond_branch(bs_put, Args, FailLbl, St0), SzCalc = bin_size_calc(T, Src, Size, U), - cg_bin_put_1(Next, Fail, reverse(Is, Acc), [SzCalc|SzCalcAcc], St); + cg_bin_put_1(Next, FailCtx, reverse(Is, Acc), [SzCalc|SzCalcAcc], St); cg_bin_put_1(#k_bin_end{}, _, Acc, SzCalcAcc, St) -> SzCalc = fold_size_calc(SzCalcAcc, 0, []), {reverse(Acc),SzCalc,St}. diff --git a/lib/compiler/src/beam_ssa.erl b/lib/compiler/src/beam_ssa.erl index 6492d1e1bf..77619368c7 100644 --- a/lib/compiler/src/beam_ssa.erl +++ b/lib/compiler/src/beam_ssa.erl @@ -21,7 +21,7 @@ -module(beam_ssa). -export([add_anno/3,get_anno/2,get_anno/3, - clobbers_xregs/1,def/2,def_used/2, + clobbers_xregs/1,def/2,def_unused/3, definitions/1, dominators/1,common_dominators/3, flatmapfold_instrs_rpo/4, @@ -79,7 +79,7 @@ -type var_base() :: atom() | non_neg_integer(). -type literal_value() :: atom() | integer() | float() | list() | - nil() | tuple() | map() | binary(). + nil() | tuple() | map() | binary() | fun(). -type op() :: {'bif',atom()} | {'float',float_op()} | prim_op() | cg_prim_op(). -type anno() :: #{atom() := any()}. @@ -101,7 +101,7 @@ 'bs_match' | 'bs_put' | 'bs_start_match' | 'bs_test_tail' | 'bs_utf16_size' | 'bs_utf8_size' | 'build_stacktrace' | 'call' | 'catch_end' | - 'extract' | + 'extract' | 'exception_trampoline' | 'get_hd' | 'get_map_element' | 'get_tl' | 'get_tuple_element' | 'has_map_field' | 'is_nonempty_list' | 'is_tagged_tuple' | @@ -120,10 +120,11 @@ %% Primops only used internally during code generation. -type cg_prim_op() :: 'bs_get' | 'bs_get_position' | 'bs_match_string' | 'bs_restore' | 'bs_save' | 'bs_set_position' | 'bs_skip' | - 'copy' | 'put_tuple_arity' | 'put_tuple_element' | - 'put_tuple_elements' | 'set_tuple_element'. + 'copy' | 'match_fail' | 'put_tuple_arity' | + 'put_tuple_element' | 'put_tuple_elements' | + 'set_tuple_element'. --import(lists, [foldl/3,keyfind/3,mapfoldl/3,member/2,reverse/1,umerge/1]). +-import(lists, [foldl/3,keyfind/3,mapfoldl/3,member/2,reverse/1]). -spec add_anno(Key, Value, Construct) -> Construct when Key :: atom(), @@ -319,17 +320,18 @@ def(Ls, Blocks) -> Blks = [map_get(L, Blocks) || L <- Top], def_1(Blks, []). --spec def_used(Ls, Blocks) -> {Def,Used} when +-spec def_unused(Ls, Used, Blocks) -> {Def,Unused} when Ls :: [label()], + Used :: ordsets:ordset(var_name()), Blocks :: block_map(), Def :: ordsets:ordset(var_name()), - Used :: ordsets:ordset(var_name()). + Unused :: ordsets:ordset(var_name()). -def_used(Ls, Blocks) -> +def_unused(Ls, Unused, Blocks) -> Top = rpo(Ls, Blocks), Blks = [map_get(L, Blocks) || L <- Top], Preds = cerl_sets:from_list(Top), - def_used_1(Blks, Preds, [], []). + def_unused_1(Blks, Preds, [], Unused). %% dominators(BlockMap) -> {Dominators,Numbering}. %% Calculate the dominator tree, returning a map where each entry @@ -651,34 +653,28 @@ is_commutative('=/=') -> true; is_commutative('/=') -> true; is_commutative(_) -> false. -def_used_1([#b_blk{is=Is,last=Last}|Bs], Preds, Def0, UsedAcc) -> - {Def,Used} = def_used_is(Is, Preds, Def0, used(Last)), - case Used of - [] -> - def_used_1(Bs, Preds, Def, UsedAcc); - [_|_] -> - def_used_1(Bs, Preds, Def, [Used|UsedAcc]) - end; -def_used_1([], _Preds, Def0, UsedAcc) -> - Def = ordsets:from_list(Def0), - Used = umerge(UsedAcc), - {Def,Used}. +def_unused_1([#b_blk{is=Is,last=Last}|Bs], Preds, Def0, Unused0) -> + Unused1 = ordsets:subtract(Unused0, used(Last)), + {Def,Unused} = def_unused_is(Is, Preds, Def0, Unused1), + def_unused_1(Bs, Preds, Def, Unused); +def_unused_1([], _Preds, Def, Unused) -> + {ordsets:from_list(Def), Unused}. -def_used_is([#b_set{op=phi,dst=Dst,args=Args}|Is], - Preds, Def0, Used0) -> +def_unused_is([#b_set{op=phi,dst=Dst,args=Args}|Is], + Preds, Def0, Unused0) -> 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{}=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=Dst}=I|Is], Preds, Def0, Used0) -> + Unused1 = [V || {#b_var{}=V,L} <- Args, cerl_sets:is_element(L, Preds)], + Unused = ordsets:subtract(Unused0, ordsets:from_list(Unused1)), + def_unused_is(Is, Preds, Def, Unused); +def_unused_is([#b_set{dst=Dst}=I|Is], Preds, Def0, Unused0) -> Def = [Dst|Def0], - Used = ordsets:union(used(I), Used0), - def_used_is(Is, Preds, Def, Used); -def_used_is([], _Preds, Def, Used) -> - {Def,Used}. + Unused = ordsets:subtract(Unused0, used(I)), + def_unused_is(Is, Preds, Def, Unused); +def_unused_is([], _Preds, Def, Unused) -> + {Def,Unused}. def_1([#b_blk{is=Is}|Bs], Def0) -> Def = def_is(Is, Def0), diff --git a/lib/compiler/src/beam_ssa.hrl b/lib/compiler/src/beam_ssa.hrl index fa76b08453..509a94135e 100644 --- a/lib/compiler/src/beam_ssa.hrl +++ b/lib/compiler/src/beam_ssa.hrl @@ -62,5 +62,13 @@ -record(b_local, {name :: beam_ssa:b_literal(), arity :: non_neg_integer()}). -%% If this block exists, it calls erlang:error(badarg). --define(BADARG_BLOCK, 1). +%% This is a psuedo-block used to express that certain instructions and BIFs +%% throw exceptions on failure. The code generator rewrites all branches to +%% this block to {f,0} which causes the instruction to throw an exception +%% instead of branching. +%% +%% Since this is not an ordinary block, it's illegal to merge it with other +%% blocks, and jumps are only valid when we know that an exception will be +%% thrown by the operation that branches here; the *block itself* does not +%% throw an exception. +-define(EXCEPTION_BLOCK, 1). diff --git a/lib/compiler/src/beam_ssa_bsm.erl b/lib/compiler/src/beam_ssa_bsm.erl index 1ac9e6a3bb..7a8dc127d7 100644 --- a/lib/compiler/src/beam_ssa_bsm.erl +++ b/lib/compiler/src/beam_ssa_bsm.erl @@ -57,6 +57,7 @@ -export([module/2, format_error/1]). -include("beam_ssa.hrl"). +-include("beam_types.hrl"). -import(lists, [member/2, reverse/1, splitwith/2, map/2, foldl/3, mapfoldl/3, nth/2, max/1, unzip/1]). @@ -683,9 +684,9 @@ aca_copy_successors(Lbl0, Blocks0, Counter0) -> Lbl = maps:get(Lbl0, BRs), {Lbl, Blocks, Counter}. -aca_cs_build_brs([?BADARG_BLOCK=Lbl | Path], Counter, Acc) -> - %% ?BADARG_BLOCK is a marker and not an actual block, so renaming it will - %% break exception handling. +aca_cs_build_brs([?EXCEPTION_BLOCK=Lbl | Path], Counter, Acc) -> + %% ?EXCEPTION_BLOCK is a marker and not an actual block, so renaming it + %% will break exception handling. aca_cs_build_brs(Path, Counter, Acc#{ Lbl => Lbl }); aca_cs_build_brs([Lbl | Path], Counter0, Acc) -> aca_cs_build_brs(Path, Counter0 + 1, Acc#{ Lbl => Counter0 }); @@ -883,8 +884,7 @@ annotate_context_parameters(F, ModInfo) -> %% Assertion. error(conflicting_parameter_types); (K, suitable_for_reuse, Acc) -> - T = beam_validator:type_anno(match_context), - Acc#{ K => T }; + Acc#{ K => #t_bs_context{} }; (_K, _V, Acc) -> Acc end, TypeAnno0, ParamInfo), diff --git a/lib/compiler/src/beam_ssa_codegen.erl b/lib/compiler/src/beam_ssa_codegen.erl index 08641e2abc..ff880c6296 100644 --- a/lib/compiler/src/beam_ssa_codegen.erl +++ b/lib/compiler/src/beam_ssa_codegen.erl @@ -28,7 +28,7 @@ -include("beam_ssa.hrl"). --import(lists, [foldl/3,keymember/3,keysort/2,last/1,map/2,mapfoldl/3, +-import(lists, [foldl/3,keymember/3,keysort/2,map/2,mapfoldl/3, reverse/1,reverse/2,sort/1,splitwith/2,takewhile/2]). -record(cg, {lcount=1 :: beam_label(), %Label counter @@ -37,7 +37,8 @@ used_labels=gb_sets:empty() :: gb_sets:set(ssa_label()), regs=#{} :: #{beam_ssa:var_name()=>ssa_register()}, ultimate_fail=1 :: beam_label(), - catches=gb_sets:empty() :: gb_sets:set(ssa_label()) + catches=gb_sets:empty() :: gb_sets:set(ssa_label()), + fc_label=1 :: beam_label() }). -spec module(beam_ssa:b_module(), [compile:option()]) -> @@ -114,17 +115,17 @@ functions(Forms, AtomMod) -> function(#b_function{anno=Anno,bs=Blocks}, AtomMod, St0) -> #{func_info:={_,Name,Arity}} = Anno, try - assert_badarg_block(Blocks), %Assertion. + assert_exception_block(Blocks), %Assertion. Regs = maps:get(registers, Anno), St1 = St0#cg{labels=#{},used_labels=gb_sets:empty(), regs=Regs}, {Fi,St2} = new_label(St1), %FuncInfo label {Entry,St3} = local_func_label(Name, Arity, St2), {Ult,St4} = new_label(St3), %Ultimate failure - Labels = (St4#cg.labels)#{0=>Entry,?BADARG_BLOCK=>0}, + Labels = (St4#cg.labels)#{0=>Entry,?EXCEPTION_BLOCK=>0}, St5 = St4#cg{labels=Labels,used_labels=gb_sets:singleton(Entry), ultimate_fail=Ult}, - {Body,St} = cg_fun(Blocks, St5), + {Body,St} = cg_fun(Blocks, St5#cg{fc_label=Fi}), Asm = [{label,Fi},line(Anno), {func_info,AtomMod,{atom,Name},Arity}] ++ add_parameter_annos(Body, Anno) ++ @@ -137,10 +138,10 @@ function(#b_function{anno=Anno,bs=Blocks}, AtomMod, St0) -> erlang:raise(Class, Error, Stack) end. -assert_badarg_block(Blocks) -> - %% Assertion: ?BADARG_BLOCK must be the call erlang:error(badarg). +assert_exception_block(Blocks) -> + %% Assertion: ?EXCEPTION_BLOCK must be a call erlang:error(badarg). case Blocks of - #{?BADARG_BLOCK:=Blk} -> + #{?EXCEPTION_BLOCK:=Blk} -> #b_blk{is=[#b_set{op=call,dst=Ret, args=[#b_remote{mod=#b_literal{val=erlang}, name=#b_literal{val=error}}, @@ -148,7 +149,7 @@ assert_badarg_block(Blocks) -> last=#b_ret{arg=Ret}} = Blk, ok; #{} -> - %% ?BADARG_BLOCK has been removed because it was never used. + %% ?EXCEPTION_BLOCK has been removed because it was never used. ok end. @@ -384,6 +385,7 @@ classify_heap_need(is_tagged_tuple) -> neutral; classify_heap_need(kill_try_tag) -> gc; classify_heap_need(landingpad) -> gc; classify_heap_need(make_fun) -> gc; +classify_heap_need(match_fail) -> gc; classify_heap_need(new_try_tag) -> gc; classify_heap_need(peek_message) -> gc; classify_heap_need(put_map) -> gc; @@ -629,7 +631,7 @@ liveness_get(S, LiveMap) -> end. liveness_successors(Terminator) -> - successors(Terminator) -- [?BADARG_BLOCK]. + successors(Terminator) -- [?EXCEPTION_BLOCK]. liveness_is([#cg_alloc{}=I0|Is], Regs, Live, Acc) -> I = I0#cg_alloc{live=num_live(Live, Regs)}, @@ -973,6 +975,12 @@ cg_block(Is0, Last, Next, St0) -> case Last of #cg_br{succ=Next,fail=Next} -> cg_block(Is0, none, St0); + #cg_br{succ=Same,fail=Same} when Same =:= ?EXCEPTION_BLOCK -> + %% An expression in this block *always* throws an exception, so we + %% terminate it with an 'if_end' to make sure the validator knows + %% that the following instructions won't actually be reached. + {Is,St} = cg_block(Is0, none, St0), + {Is++[if_end],St}; #cg_br{succ=Same,fail=Same} -> {Fail,St1} = use_block_label(Same, St0), {Is,St} = cg_block(Is0, none, St1), @@ -1178,6 +1186,10 @@ cg_block([#cg_set{op=call}=I, #cg_set{op=succeeded,dst=Bool}], {Bool,_Fail}, St) -> %% A call in try/catch block. cg_block([I], none, St); +cg_block([#cg_set{op=match_fail}=I, + #cg_set{op=succeeded,dst=Bool}], {Bool,_Fail}, St) -> + %% A match_fail instruction in a try/catch block. + cg_block([I], none, St); cg_block([#cg_set{op=get_map_element,dst=Dst0,args=Args0}, #cg_set{op=succeeded,dst=Bool}], {Bool,Fail0}, St) -> [Dst,Map,Key] = beam_args([Dst0|Args0], St), @@ -1239,6 +1251,28 @@ cg_block([#cg_set{op=copy}|_]=T0, Context, St0) -> no -> {Is,St} end; +cg_block([#cg_set{op=match_fail,args=Args0,anno=Anno}], none, St) -> + Args = beam_args(Args0, St), + Is = cg_match_fail(Args, line(Anno), none), + {Is,St}; +cg_block([#cg_set{op=match_fail,args=Args0,anno=Anno}|T], Context, St0) -> + FcLabel = case Context of + {return,_,none} -> + %% There is no stack frame. If this is a function_clause + %% exception, it is safe to jump to the label of the + %% func_info instruction. + St0#cg.fc_label; + _ -> + %% This is most probably not a function_clause. + %% If this is a function_clause exception + %% (rare), it is not safe to jump to the + %% func_info label. + none + end, + Args = beam_args(Args0, St0), + Is0 = cg_match_fail(Args, line(Anno), FcLabel), + {Is1,St} = cg_block(T, Context, St0), + {Is0++Is1,St}; cg_block([#cg_set{op=Op,dst=Dst0,args=Args0}=Set], none, St) -> [Dst|Args] = beam_args([Dst0|Args0], St), Is = cg_instr(Op, Args, Dst, Set), @@ -1270,8 +1304,7 @@ cg_copy(T0, St) -> end, T0), Moves0 = cg_copy_1(Copies, St), Moves1 = [Move || {move,Src,Dst}=Move <- Moves0, Src =/= Dst], - Scratch = {x,1022}, - Moves = order_moves(Moves1, Scratch), + Moves = order_moves(Moves1), {Moves,T}. cg_copy_1([#cg_set{dst=Dst0,args=Args}|T], St) -> @@ -1512,6 +1545,42 @@ cg_call(#cg_set{anno=Anno,op=call,dst=Dst0,args=Args0}, Is = setup_args(Args++[Func], Anno, Context, St) ++ Line ++ Call, {Is,St}. +cg_match_fail([{atom,function_clause}|Args], Line, Fc) -> + case Fc of + none -> + %% There is a stack frame (probably because of inlining). + %% Jumping to the func_info label is not allowed by + %% beam_validator. Rewrite the instruction as a call to + %% erlang:error/2. + make_fc(Args, Line); + _ -> + setup_args(Args) ++ [{jump,{f,Fc}}] + end; +cg_match_fail([{atom,Op}], Line, _Fc) -> + [Line,Op]; +cg_match_fail([{atom,Op},Val], Line, _Fc) -> + [Line,{Op,Val}]. + +make_fc(Args, Line) -> + %% Recreate the original call to erlang:error/2. + Live = foldl(fun({x,X}, A) -> max(X+1, A); + (_, A) -> A + end, 0, Args), + TmpReg = {x,Live}, + StkMoves = build_stk(reverse(Args), TmpReg, nil), + [{test_heap,2*length(Args),Live}|StkMoves] ++ + [{move,{atom,function_clause},{x,0}}, + Line, + {call_ext,2,{extfunc,erlang,error,2}}]. + +build_stk([V], _TmpReg, Tail) -> + [{put_list,V,Tail,{x,1}}]; +build_stk([V|Vs], TmpReg, Tail) -> + I = {put_list,V,Tail,TmpReg}, + [I|build_stk(Vs, TmpReg, TmpReg)]; +build_stk([], _TmpReg, nil) -> + [{move,nil,{x,1}}]. + build_call(call_fun, Arity, _Func, none, Dst) -> [{call_fun,Arity}|copy({x,0}, Dst)]; build_call(call_fun, Arity, _Func, {return,Dst,N}, Dst) when is_integer(N) -> @@ -1550,15 +1619,15 @@ build_apply(Arity, {return,Val,N}, _Dst) when is_integer(N) -> build_apply(Arity, none, Dst) -> [{apply,Arity}|copy({x,0}, 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(put_map, [{atom,assoc},SrcMap|Ss], Dst, Set) -> + Live = get_live(Set), + [{put_map_assoc,{f,0},SrcMap,Dst,Live,{list,Ss}}]; cg_instr(Op, Args, Dst, _Set) -> cg_instr(Op, Args, Dst). @@ -1728,7 +1797,7 @@ cg_catch(Agg, T0, Context, St0) -> cg_try(Agg, Tag, T0, Context, St0) -> {Moves0,T1} = cg_extract(T0, Agg, St0), - Moves = order_moves(Moves0, {x,3}), + Moves = order_moves(Moves0), [#cg_set{op=kill_try_tag}|T2] = T1, {T,St} = cg_block(T2, Context, St0), {[{try_case,Tag}|Moves++T],St}. @@ -1780,7 +1849,7 @@ linearize(Blocks) -> Linear = beam_ssa:linearize(Blocks), linearize_1(Linear, Blocks). -linearize_1([{?BADARG_BLOCK,_}|Ls], Blocks) -> +linearize_1([{?EXCEPTION_BLOCK,_}|Ls], Blocks) -> linearize_1(Ls, Blocks); linearize_1([{L,Block0}|Ls], Blocks) -> Block = translate_block(L, Block0, Blocks), @@ -1884,8 +1953,7 @@ setup_args([]) -> []; setup_args([_|_]=Args) -> Moves = gen_moves(Args, 0, []), - Scratch = {x,1+last(sort([length(Args)-1|[X || {x,X} <- Args]]))}, - order_moves(Moves, Scratch). + order_moves(Moves). %% kill_yregs(Anno, #cg{}) -> [{kill,{y,Y}}]. %% Kill Y registers that will not be used again. @@ -1905,47 +1973,48 @@ gen_moves([A|As], I, Acc) -> gen_moves([], _, Acc) -> keysort(3, Acc). -%% order_moves([Move], ScratchReg) -> [Move] +%% order_moves([Move]) -> [Move] %% Orders move instruction so that source registers are not %% destroyed before they are used. If there are cycles %% (such as {move,{x,0},{x,1}}, {move,{x,1},{x,1}}), -%% the scratch register is used to break up the cycle. -%% If possible, the first move of the input list is placed +%% swap instructions will be used to break up the cycle. +%% +%% If possible, the first move of the input list is placed %% last in the result list (to make the move to {x,0} occur %% just before the call to allow the Beam loader to coalesce %% the instructions). -order_moves(Ms, Scr) -> order_moves(Ms, Scr, []). +order_moves(Ms) -> order_moves(Ms, []). -order_moves([{move,_,_}=M|Ms0], ScrReg, Acc0) -> - {Chain,Ms} = collect_chain(Ms0, [M], ScrReg), +order_moves([{move,_,_}=M|Ms0], Acc0) -> + {Chain,Ms} = collect_chain(Ms0, [M]), Acc = reverse(Chain, Acc0), - order_moves(Ms, ScrReg, Acc); -order_moves([], _, Acc) -> Acc. + order_moves(Ms, Acc); +order_moves([], Acc) -> Acc. -collect_chain(Ms, Path, ScrReg) -> - collect_chain(Ms, Path, [], ScrReg). +collect_chain(Ms, Path) -> + collect_chain(Ms, Path, []). -collect_chain([{move,Src,Same}=M|Ms0], [{move,Same,_}|_]=Path, Others, ScrReg) -> +collect_chain([{move,Src,Same}=M|Ms0], [{move,Same,_}|_]=Path, Others) -> case keymember(Src, 3, Path) of false -> - collect_chain(reverse(Others, Ms0), [M|Path], [], ScrReg); + collect_chain(reverse(Others, Ms0), [M|Path], []); true -> - %% There is a cycle, which we must break up. - {break_up_cycle(M, Path, ScrReg),reverse(Others, Ms0)} + %% There is a cycle. + {break_up_cycle(M, Path),reverse(Others, Ms0)} end; -collect_chain([M|Ms], Path, Others, ScrReg) -> - collect_chain(Ms, Path, [M|Others], ScrReg); -collect_chain([], Path, Others, _) -> +collect_chain([M|Ms], Path, Others) -> + collect_chain(Ms, Path, [M|Others]); +collect_chain([], Path, Others) -> {Path,Others}. -break_up_cycle({move,Src,_}=M, Path, ScrReg) -> - [{move,ScrReg,Src},M|break_up_cycle1(Src, Path, ScrReg)]. +break_up_cycle({move,Src,_Dst}=M, Path) -> + break_up_cycle_1(Src, [M|Path], []). -break_up_cycle1(Dst, [{move,Src,Dst}|Path], ScrReg) -> - [{move,Src,ScrReg}|Path]; -break_up_cycle1(Dst, [M|Path], LastMove) -> - [M|break_up_cycle1(Dst, Path, LastMove)]. +break_up_cycle_1(Dst, [{move,_Src,Dst}|Path], Acc) -> + reverse(Acc, Path); +break_up_cycle_1(Dst, [{move,S,D}|Path], Acc) -> + break_up_cycle_1(Dst, Path, [{swap,S,D}|Acc]). %%% %%% General utility functions. diff --git a/lib/compiler/src/beam_ssa_dead.erl b/lib/compiler/src/beam_ssa_dead.erl index e78e4647a8..55ded77d43 100644 --- a/lib/compiler/src/beam_ssa_dead.erl +++ b/lib/compiler/src/beam_ssa_dead.erl @@ -760,8 +760,8 @@ 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 -> 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; @@ -781,9 +781,9 @@ 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 -> 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) -> if diff --git a/lib/compiler/src/beam_ssa_lint.erl b/lib/compiler/src/beam_ssa_lint.erl index a003607dab..224095d4c4 100644 --- a/lib/compiler/src/beam_ssa_lint.erl +++ b/lib/compiler/src/beam_ssa_lint.erl @@ -65,13 +65,19 @@ format_error({{_M,F,A},{phi_inside_block, Name, Id}}) -> [F, A, format_var(Name), Id]); format_error({{_M,F,A},{undefined_label_in_phi, Label, I}}) -> io_lib:format("~p/~p: Unknown block label ~p in phi node ~ts", - [F, A, Label, format_instr(I)]). + [F, A, Label, format_instr(I)]); +format_error({{_M,F,A},{succeeded_not_preceded, I}}) -> + io_lib:format("~p/~p: ~ts does not reference the preceding instruction", + [F, A, format_instr(I)]); +format_error({{_M,F,A},{succeeded_not_last, I}}) -> + io_lib:format("~p/~p: ~ts is not the last instruction in its block", + [F, A, format_instr(I)]). format_instr(I) -> [$',beam_ssa_pp:format_instr(I),$']. format_var(V) -> - beam_ssa_pp:format_var(#b_var{name=V}). + beam_ssa_pp:format_var(V). validate_function(F) -> try @@ -86,34 +92,36 @@ validate_function(F) -> erlang:raise(Class, Error, Stack) end. --type defined_vars() :: gb_sets:set(beam_ssa:var_name()). +-type defined_vars() :: gb_sets:set(beam_ssa:argument()). -record(vvars, {blocks :: #{ beam_ssa:label() => beam_ssa:b_blk() }, branch_def_vars :: #{ - %% Describes the variable state at the time of this exact branch (phi - %% node validation). - {From :: beam_ssa:label(), To :: beam_ssa:label()} => defined_vars(), - %% Describes the variable state common to all branches leading to this - %% label (un/redefined variable validation). - beam_ssa:label() => defined_vars() }, + %% Describes the variable state at the time of + %% this exact branch (phi node validation). + {From :: beam_ssa:label(), + To :: beam_ssa:label()} => defined_vars(), + %% Describes the variable state common to all + %% branches leading to this label (un/redefined + %% variable validation). + beam_ssa:label() => defined_vars() }, defined_vars :: defined_vars()}). -spec validate_variables(beam_ssa:b_function()) -> ok. validate_variables(#b_function{ args = Args, bs = Blocks }) -> %% Prefill the mapping with function arguments. - ArgNames = vvars_get_varnames(Args), - DefVars = gb_sets:from_list(ArgNames), + Args = vvars_get_variables(Args), + DefVars = gb_sets:from_list(Args), Entry = 0, State = #vvars{blocks = Blocks, branch_def_vars = #{ Entry => DefVars }, defined_vars = DefVars}, - ok = vvars_assert_unique(Blocks, ArgNames), + ok = vvars_assert_unique(Blocks, Args), vvars_phi_nodes(vvars_block(Entry, State)). %% Checks the uniqueness of all variables across all blocks. --spec vvars_assert_unique(Blocks, [beam_ssa:var_name()]) -> ok when +-spec vvars_assert_unique(Blocks, [beam_ssa:b_var()]) -> ok when Blocks :: #{ beam_ssa:label() => beam_ssa:b_blk() }. vvars_assert_unique(Blocks, Args) -> BlockIs = [Is || #b_blk{is=Is} <- maps:values(Blocks)], @@ -124,12 +132,12 @@ vvars_assert_unique(Blocks, Args) -> ok. -spec vvars_assert_unique_1(Is, Defined) -> ok when - Is :: list(beam_ssa:b_set()), - Defined :: #{ beam_ssa:var_name() => beam_ssa:b_set() }. -vvars_assert_unique_1([#b_set{dst=#b_var{name=DstName}}=I|Is], Defined) -> + Is :: list(beam_ssa:b_set()), + Defined :: #{ beam_ssa:b_var() => beam_ssa:b_set() }. +vvars_assert_unique_1([#b_set{dst=Dst}=I|Is], Defined) -> case Defined of - #{DstName:=Old} -> throw({redefined_variable, DstName, Old, I}); - _ -> vvars_assert_unique_1(Is, Defined#{DstName=>I}) + #{Dst:=Old} -> throw({redefined_variable, Dst, Old, I}); + _ -> vvars_assert_unique_1(Is, Defined#{Dst=>I}) end; vvars_assert_unique_1([], Defined) -> Defined. @@ -141,17 +149,17 @@ vvars_phi_nodes(#vvars{ blocks = Blocks }=State) -> ok. -spec vvars_phi_nodes_1(Is, Id, State) -> ok when - Is :: list(beam_ssa:b_set()), - Id :: beam_ssa:label(), - State :: #vvars{}. + Is :: list(beam_ssa:b_set()), + Id :: beam_ssa:label(), + State :: #vvars{}. vvars_phi_nodes_1([#b_set{ op = phi, args = Phis }=I | Is], Id, State) -> ok = vvars_assert_phi_paths(Phis, I, Id, State), ok = vvars_assert_phi_vars(Phis, I, Id, State), vvars_phi_nodes_1(Is, Id, State); vvars_phi_nodes_1([_ | Is], Id, _State) -> - case [Dst || #b_set{op=phi,dst=#b_var{name=Dst}} <- Is] of - [Name|_] -> - throw({phi_inside_block, Name, Id}); + case [Dst || #b_set{op=phi,dst=Dst} <- Is] of + [Var|_] -> + throw({phi_inside_block, Var, Id}); [] -> ok end; @@ -161,10 +169,10 @@ vvars_phi_nodes_1([], _Id, _State) -> %% Checks whether all paths leading to this phi node are represented, and that %% it doesn't reference any non-existent paths. -spec vvars_assert_phi_paths(Phis, I, Id, State) -> ok when - Phis :: list({beam_ssa:argument(), beam_ssa:label()}), - Id :: beam_ssa:label(), - I :: beam_ssa:b_set(), - State :: #vvars{}. + Phis :: list({beam_ssa:argument(), beam_ssa:label()}), + Id :: beam_ssa:label(), + I :: beam_ssa:b_set(), + State :: #vvars{}. vvars_assert_phi_paths(Phis, I, Id, State) -> BranchKeys = maps:keys(State#vvars.branch_def_vars), RequiredPaths = ordsets:from_list([From || {From, To} <- BranchKeys, To =:= Id]), @@ -173,34 +181,34 @@ vvars_assert_phi_paths(Phis, I, Id, State) -> [_|_]=MissingPaths -> throw({missing_phi_paths, MissingPaths, I}); [] -> ok end. - %% %% The following test is sometimes useful to find missing optimizations. - %% %% It is commented out, though, because it can be triggered by - %% %% by weird but legal code. - %% case ordsets:subtract(ProvidedPaths, RequiredPaths) of - %% [_|_]=GarbagePaths -> throw({garbage_phi_paths, GarbagePaths, I}); - %% [] -> ok - %% end. +%% %% The following test is sometimes useful to find missing optimizations. +%% %% It is commented out, though, because it can be triggered by +%% %% by weird but legal code. +%% case ordsets:subtract(ProvidedPaths, RequiredPaths) of +%% [_|_]=GarbagePaths -> throw({garbage_phi_paths, GarbagePaths, I}); +%% [] -> ok +%% end. %% Checks whether all variables used in this phi node are defined in the branch %% they arrived on. -spec vvars_assert_phi_vars(Phis, I, Id, State) -> ok when - Phis :: list({beam_ssa:argument(), beam_ssa:label()}), - Id :: beam_ssa:label(), - I :: beam_ssa:b_set(), - State :: #vvars{}. + Phis :: list({beam_ssa:argument(), beam_ssa:label()}), + Id :: beam_ssa:label(), + I :: beam_ssa:b_set(), + State :: #vvars{}. vvars_assert_phi_vars(Phis, I, Id, #vvars{blocks=Blocks, branch_def_vars=BranchDefVars}) -> Vars = [{Var, From} || {#b_var{}=Var, From} <- Phis], - foreach(fun({#b_var{name=VarName}, From}) -> + foreach(fun({Var, From}) -> BranchKey = {From, Id}, case BranchDefVars of #{BranchKey:=DefVars} -> - case gb_sets:is_member(VarName, DefVars) of + case gb_sets:is_member(Var, DefVars) of true -> ok; - false -> throw({unknown_variable, VarName, I}) + false -> throw({unknown_variable, Var, I}) end; #{} -> - throw({unknown_phi_variable, VarName, BranchKey, I}) + throw({unknown_phi_variable, Var, BranchKey, I}) end end, Vars), Labels = [From || {#b_literal{},From} <- Phis], @@ -214,32 +222,44 @@ vvars_assert_phi_vars(Phis, I, Id, #vvars{blocks=Blocks, end, Labels). -spec vvars_block(Id, State) -> #vvars{} when - Id :: beam_ssa:label(), - State :: #vvars{}. + Id :: beam_ssa:label(), + State :: #vvars{}. vvars_block(Id, State0) -> #{ Id := #b_blk{ is = Is, last = Terminator} } = State0#vvars.blocks, #{ Id := DefVars } = State0#vvars.branch_def_vars, State = State0#vvars{ defined_vars = DefVars }, vvars_terminator(Terminator, Id, vvars_block_1(Is, State)). --spec vvars_block_1(Blocks, State) -> #vvars{} when - Blocks :: list(beam_ssa:b_blk()), - State :: #vvars{}. +-spec vvars_block_1(Is, State) -> #vvars{} when + Is :: list(#b_set{}), + State :: #vvars{}. vvars_block_1([], State) -> State; -vvars_block_1([#b_set{ dst = #b_var{ name = DstName }, op = phi } | Is], State0) -> +vvars_block_1([#b_set{dst=OpVar,args=OpArgs}=I, + #b_set{op=succeeded,args=[OpVar],dst=SuccVar}], State) -> + ok = vvars_assert_args(OpArgs, I, State), + vvars_save_var(SuccVar, vvars_save_var(OpVar, State)); +vvars_block_1([#b_set{op=succeeded,args=Args}=I | [_|_]], State) -> + ok = vvars_assert_args(Args, I, State), + %% 'succeeded' must be the last instruction in its block. + throw({succeeded_not_last, I}); +vvars_block_1([#b_set{op=succeeded,args=Args}=I], State)-> + ok = vvars_assert_args(Args, I, State), + %% 'succeeded' must be be directly preceded by the operation it checks. + throw({succeeded_not_preceded, I}); +vvars_block_1([#b_set{ dst = Dst, op = phi } | Is], State) -> %% We don't check phi node arguments at this point since we may not have %% visited their definition yet. They'll be handled later on in %% vvars_phi_nodes/1 after all blocks are processed. - vvars_block_1(Is, vvars_save_var(DstName, State0)); -vvars_block_1([#b_set{ dst = #b_var{ name = DstName }, args = Args }=I | Is], State0) -> - ok = vvars_assert_args(Args, I, State0), - vvars_block_1(Is, vvars_save_var(DstName, State0)). + vvars_block_1(Is, vvars_save_var(Dst, State)); +vvars_block_1([#b_set{ dst = Dst, args = Args }=I | Is], State) -> + ok = vvars_assert_args(Args, I, State), + vvars_block_1(Is, vvars_save_var(Dst, State)). -spec vvars_terminator(Terminator, From, State) -> #vvars{} when - Terminator :: beam_ssa:terminator(), - From :: beam_ssa:label(), - State :: #vvars{}. + Terminator :: beam_ssa:terminator(), + From :: beam_ssa:label(), + State :: #vvars{}. vvars_terminator(#b_ret{ arg = Arg }=I, _From, State) -> ok = vvars_assert_args([Arg], I, State), State; @@ -264,62 +284,62 @@ vvars_terminator(#b_br{ bool = Arg, succ = Succ, fail = Fail }=I, From, State) - vvars_terminator_1(Labels, From, State). -spec vvars_terminator_1(Labels, From, State) -> #vvars{} when - Labels :: list(beam_ssa:label()), - From :: beam_ssa:label(), - State :: #vvars{}. + Labels :: list(beam_ssa:label()), + From :: beam_ssa:label(), + State :: #vvars{}. vvars_terminator_1(Labels0, From, State0) -> %% Filter out all branches that have already been taken. This should result %% in either all of Labels0 or an empty list. Labels = [To || To <- Labels0, - not maps:is_key({From, To}, State0#vvars.branch_def_vars)], + not maps:is_key({From, To}, State0#vvars.branch_def_vars)], true = Labels =:= Labels0 orelse Labels =:= [], %Assertion State1 = foldl(fun(To, State) -> - vvars_save_branch(From, To, State) + vvars_save_branch(From, To, State) end, State0, Labels), foldl(fun(To, State) -> - vvars_block(To, State) + vvars_block(To, State) end, State1, Labels). %% Gets all variable names in args, ignoring literals etc --spec vvars_get_varnames(Args) -> list(beam_ssa:var_name()) when - Args :: list(beam_ssa:argument()). -vvars_get_varnames(Args) -> - [Name || #b_var{ name = Name } <- Args]. +-spec vvars_get_variables(Args) -> list(beam_ssa:b_var()) when + Args :: list(beam_ssa:argument()). +vvars_get_variables(Args) -> + [Var || #b_var{}=Var <- Args]. %% Checks that all variables in Args are defined in all paths leading to the %% current State. -spec vvars_assert_args(Args, I, State) -> ok when - Args :: list(beam_ssa:argument()), - I :: beam_ssa:terminator() | beam_ssa:b_set(), - State :: #vvars{}. + Args :: list(beam_ssa:argument()), + I :: beam_ssa:terminator() | beam_ssa:b_set(), + State :: #vvars{}. vvars_assert_args(Args, I, #vvars{defined_vars=DefVars}=State) -> foreach(fun(#b_remote{mod=Mod,name=Name}) -> vvars_assert_args([Mod,Name], I, State); - (#b_var{name=Name}) -> - case gb_sets:is_member(Name, DefVars) of + (#b_var{}=Var) -> + case gb_sets:is_member(Var, DefVars) of true -> ok; - false -> throw({unknown_variable,Name,I}) + false -> throw({unknown_variable,Var,I}) end; (_) -> ok end, Args). %% Checks that all given labels are defined in State. -spec vvars_assert_labels(Labels, I, State) -> ok when - Labels :: list(beam_ssa:label()), - I :: beam_ssa:terminator(), - State :: #vvars{}. + Labels :: list(beam_ssa:label()), + I :: beam_ssa:terminator(), + State :: #vvars{}. vvars_assert_labels(Labels, I, #vvars{blocks=Blocks}) -> foreach(fun(Label) -> - case maps:is_key(Label, Blocks) of - false -> throw({unknown_block, Label, I}); - true -> ok - end + case maps:is_key(Label, Blocks) of + false -> throw({unknown_block, Label, I}); + true -> ok + end end, Labels). -spec vvars_save_branch(From, To, State) -> #vvars{} when - From :: beam_ssa:label(), - To :: beam_ssa:label(), - State :: #vvars{}. + From :: beam_ssa:label(), + To :: beam_ssa:label(), + State :: #vvars{}. vvars_save_branch(From, To, State) -> DefVars = State#vvars.defined_vars, Branches0 = State#vvars.branch_def_vars, @@ -335,15 +355,15 @@ vvars_save_branch(From, To, State) -> end. -spec vvars_merge_branches(New, Existing) -> defined_vars() when - New :: defined_vars(), - Existing :: defined_vars(). + New :: defined_vars(), + Existing :: defined_vars(). vvars_merge_branches(New, Existing) -> gb_sets:intersection(New, Existing). --spec vvars_save_var(VarName, State) -> #vvars{} when - VarName :: beam_ssa:var_name(), - State :: #vvars{}. -vvars_save_var(VarName, State0) -> +-spec vvars_save_var(Var, State) -> #vvars{} when + Var :: #b_var{}, + State :: #vvars{}. +vvars_save_var(Var, State0) -> %% vvars_assert_unique guarantees that variables are never set twice. - DefVars = gb_sets:insert(VarName, State0#vvars.defined_vars), + DefVars = gb_sets:insert(Var, State0#vvars.defined_vars), State0#vvars{ defined_vars = DefVars }. diff --git a/lib/compiler/src/beam_ssa_opt.erl b/lib/compiler/src/beam_ssa_opt.erl index d87c66c272..580abf4ed9 100644 --- a/lib/compiler/src/beam_ssa_opt.erl +++ b/lib/compiler/src/beam_ssa_opt.erl @@ -145,7 +145,8 @@ prologue_passes(Opts) -> ?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_cse), % Helps the first type pass. + ?PASS(ssa_opt_live), % ... ?PASS(ssa_opt_type_start)], passes_1(Ps, Opts). @@ -157,6 +158,9 @@ repeated_passes(Opts) -> ?PASS(ssa_opt_dead), ?PASS(ssa_opt_cse), ?PASS(ssa_opt_tail_phis), + ?PASS(ssa_opt_sink), + ?PASS(ssa_opt_tuple_size), + ?PASS(ssa_opt_record), ?PASS(ssa_opt_type_continue)], %Must run after ssa_opt_dead to %clean up phi nodes. passes_1(Ps, Opts). @@ -172,8 +176,8 @@ epilogue_passes(Opts) -> ?PASS(ssa_opt_bsm), ?PASS(ssa_opt_bsm_units), ?PASS(ssa_opt_bsm_shortcut), - ?PASS(ssa_opt_blockify), ?PASS(ssa_opt_sink), + ?PASS(ssa_opt_blockify), ?PASS(ssa_opt_merge_blocks), ?PASS(ssa_opt_get_tuple_element), ?PASS(ssa_opt_trim_unreachable)], @@ -898,6 +902,7 @@ cse_suitable(#b_set{}) -> false. -record(fs, {s=undefined :: 'undefined' | 'cleared', regs=#{} :: #{beam_ssa:b_var():=beam_ssa:b_var()}, + vars=cerl_sets:new() :: cerl_sets:set(), fail=none :: 'none' | beam_ssa:label(), non_guards :: gb_sets:set(beam_ssa:label()), bs :: beam_ssa:block_map() @@ -910,22 +915,39 @@ ssa_opt_float({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) -> {Linear,Count} = float_opt(Linear0, Count0, Fs), {St#st{ssa=Linear,cnt=Count}, FuncDb}. -float_blk_is_in_guard(#b_blk{last=#b_br{fail=F}}, #fs{non_guards=NonGuards}) -> - not gb_sets:is_member(F, NonGuards); -float_blk_is_in_guard(#b_blk{}, #fs{}) -> +%% The fconv instruction doesn't support jumping to a fail label, so we have to +%% skip this optimization if the fail block is a guard. +%% +%% We also skip the optimization in blocks that always fail, as it's both +%% difficult and pointless to rewrite them to use float ops. +float_can_optimize_blk(#b_blk{last=#b_br{bool=#b_var{},fail=F}}, + #fs{non_guards=NonGuards}) -> + gb_sets:is_member(F, NonGuards); +float_can_optimize_blk(#b_blk{}, #fs{}) -> false. +float_opt([{L,#b_blk{is=[#b_set{op=exception_trampoline,args=[Var]}]}=Blk0} | + Bs0], Count0, Fs) -> + %% If we've replaced a BIF with float operations, we'll have a lot of extra + %% blocks that jump to the same failure block, which may have a trampoline + %% that refers to the original operation. + %% + %% Since the point of the trampoline is to keep the BIF from being removed + %% by liveness optimization, we can discard it as the liveness pass leaves + %% floats alone. + Blk = case cerl_sets:is_element(Var, Fs#fs.vars) of + true -> Blk0#b_blk{is=[]}; + false -> Blk0 + end, + {Bs, Count} = float_opt(Bs0, Count0, Fs), + {[{L,Blk}|Bs],Count}; float_opt([{L,Blk}|Bs0], Count0, Fs) -> - case float_blk_is_in_guard(Blk, Fs) of + case float_can_optimize_blk(Blk, Fs) of true -> - %% This block is inside a guard. Don't do - %% any floating point optimizations. - {Bs,Count} = float_opt(Bs0, Count0, Fs), - {[{L,Blk}|Bs],Count}; + float_opt_1(L, Blk, Bs0, Count0, Fs); false -> - %% This block is not inside a guard. - %% We can do the optimization. - float_opt_1(L, Blk, Bs0, Count0, Fs) + {Bs,Count} = float_opt(Bs0, Count0, Fs), + {[{L,Blk}|Bs],Count} end; float_opt([], Count, _Fs) -> {[],Count}. @@ -1004,12 +1026,12 @@ 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, %% If the success block starts with a floating point operation, we can - %% defer flushing to that block as long as it isn't a guard. + %% defer flushing to that block as long as it's suitable for optimization. #b_blk{is=Is} = SuccBlk = map_get(Succ, Blocks), - SuccIsGuard = float_blk_is_in_guard(SuccBlk, Fs0), + CanOptimizeSucc = float_can_optimize_blk(SuccBlk, Fs0), case Is of - [#b_set{anno=#{float_op:=_}}|_] when not SuccIsGuard -> + [#b_set{anno=#{float_op:=_}}|_] when CanOptimizeSucc -> %% No flush needed. {[],Blk0,Fs0,Count0}; _ -> @@ -1065,21 +1087,22 @@ float_opt_is([], Fs, _Count, _Acc) -> none. float_make_op(#b_set{op={bif,Op},dst=Dst,args=As0}=I0, - Ts, #fs{s=S,regs=Rs0}=Fs, Count0) -> + Ts, #fs{s=S,regs=Rs0,vars=Vs0}=Fs, Count0) -> {As1,Rs1,Count1} = float_load(As0, Ts, Rs0, Count0, []), {As,Is0} = unzip(As1), {Fr,Count2} = new_reg('@fr', Count1), FrDst = #b_var{name=Fr}, I = I0#b_set{op={float,Op},dst=FrDst,args=As}, + Vs = cerl_sets:add_element(Dst, Vs0), Rs = Rs1#{Dst=>FrDst}, Is = append(Is0) ++ [I], case S of undefined -> {Ignore,Count} = new_reg('@ssa_ignore', Count2), C = #b_set{op={float,clearerror},dst=#b_var{name=Ignore}}, - {[C|Is],Fs#fs{s=cleared,regs=Rs},Count}; + {[C|Is],Fs#fs{s=cleared,regs=Rs,vars=Vs},Count}; cleared -> - {Is,Fs#fs{regs=Rs},Count2} + {Is,Fs#fs{regs=Rs,vars=Vs},Count2} end. float_load([A|As], [T|Ts], Rs0, Count0, Acc) -> @@ -1208,34 +1231,31 @@ live_opt_is([#b_set{op=phi,dst=Dst}=I|Is], Live, Acc) -> false -> live_opt_is(Is, Live, Acc) end; -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 -> - 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} -> - NewI = NewI0#b_set{dst=SuccDstVar}, - live_opt_is([NewI|Is], Live0, Acc); - keep -> - case gb_sets:is_member(SuccDst, Live0) of - true -> - Live1 = gb_sets:add(Dst, Live0), - Live = gb_sets:delete(SuccDst, Live1), - live_opt_is([I|Is], Live, [SuccI|Acc]); - false -> - live_opt_is([I|Is], Live0, Acc) - end - end +live_opt_is([#b_set{op=succeeded,dst=SuccDst,args=[MapDst]}=SuccI, + #b_set{op=get_map_element,dst=MapDst}=MapI | Is], + Live0, Acc) -> + case {gb_sets:is_member(SuccDst, Live0), + gb_sets:is_member(MapDst, Live0)} of + {true, true} -> + Live = gb_sets:delete(SuccDst, Live0), + live_opt_is([MapI | Is], Live, [SuccI | Acc]); + {true, false} -> + %% 'get_map_element' is unused; replace 'succeeded' with + %% 'has_map_field' + NewI = MapI#b_set{op=has_map_field,dst=SuccDst}, + live_opt_is([NewI | Is], Live0, Acc); + {false, true} -> + %% 'succeeded' is unused (we know it will succeed); discard it and + %% keep 'get_map_element' + live_opt_is([MapI | Is], Live0, Acc); + {false, false} -> + live_opt_is(Is, Live0, Acc) end; 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))), + LiveUsed = gb_sets:from_ordset(beam_ssa:used(I)), + Live1 = gb_sets:union(Live0, LiveUsed), Live = gb_sets:delete(Dst, Live1), live_opt_is(Is, Live, [I|Acc]); false -> @@ -1243,17 +1263,14 @@ live_opt_is([#b_set{dst=Dst}=I|Is], Live0, Acc) -> true -> live_opt_is(Is, Live0, Acc); false -> - Live = gb_sets:union(Live0, gb_sets:from_ordset(beam_ssa:used(I))), + LiveUsed = gb_sets:from_ordset(beam_ssa:used(I)), + Live = gb_sets:union(Live0, LiveUsed), live_opt_is(Is, Live, [I|Acc]) end end; live_opt_is([], Live, Acc) -> {Acc,Live}. -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. %%% @@ -1939,6 +1956,10 @@ verify_merge_is(_) -> is_merge_allowed(_, #b_blk{}, #b_blk{is=[#b_set{op=peek_message}|_]}) -> false; +is_merge_allowed(_, #b_blk{}, #b_blk{is=[#b_set{op=exception_trampoline}|_]}) -> + false; +is_merge_allowed(_, #b_blk{is=[#b_set{op=exception_trampoline}|_]}, #b_blk{}) -> + false; is_merge_allowed(L, #b_blk{last=#b_br{}}=Blk, #b_blk{is=Is}) -> %% The predecessor block must have exactly one successor (L) for %% the merge to be safe. @@ -1977,9 +1998,7 @@ is_merge_allowed(_, #b_blk{last=#b_switch{}}, #b_blk{}) -> %%% extracted values. %%% -ssa_opt_sink({#st{ssa=Blocks0}=St, FuncDb}) -> - Linear = beam_ssa:linearize(Blocks0), - +ssa_opt_sink({#st{ssa=Linear}=St, FuncDb}) -> %% Create a map with all variables that define get_tuple_element %% instructions. The variable name map to the block it is defined in. case def_blocks(Linear) of @@ -1988,10 +2007,12 @@ ssa_opt_sink({#st{ssa=Blocks0}=St, FuncDb}) -> {St, FuncDb}; [_|_]=Defs0 -> Defs = maps:from_list(Defs0), - {do_ssa_opt_sink(Linear, Defs, St), FuncDb} + {do_ssa_opt_sink(Defs, St), FuncDb} end. -do_ssa_opt_sink(Linear, Defs, #st{ssa=Blocks0}=St) -> +do_ssa_opt_sink(Defs, #st{ssa=Linear}=St) -> + Blocks0 = maps:from_list(Linear), + %% Now find all the blocks that use variables defined by get_tuple_element %% instructions. Used = used_blocks(Linear, Defs, []), @@ -2016,7 +2037,8 @@ do_ssa_opt_sink(Linear, Defs, #st{ssa=Blocks0}=St) -> From = map_get(V, Defs), move_defs(V, From, To, A) end, Blocks0, DefLoc), - St#st{ssa=Blocks}. + + St#st{ssa=beam_ssa:linearize(Blocks)}. def_blocks([{L,#b_blk{is=Is}}|Bs]) -> def_blocks_is(Is, L, def_blocks(Bs)); @@ -2049,6 +2071,7 @@ unsuitable_1([{L,#b_blk{is=[#b_set{op=Op}|_]}}|Bs]) -> Unsuitable = case Op of bs_extract -> true; bs_put -> true; + exception_trampoline -> true; {float,_} -> true; landingpad -> true; peek_message -> true; @@ -2257,13 +2280,15 @@ non_guards(Linear) -> non_guards_1([{L,#b_blk{is=Is}}|Bs]) -> case Is of + [#b_set{op=exception_trampoline}|_] -> + [L | non_guards_1(Bs)]; [#b_set{op=landingpad}|_] -> [L | non_guards_1(Bs)]; _ -> non_guards_1(Bs) end; non_guards_1([]) -> - [?BADARG_BLOCK]. + [?EXCEPTION_BLOCK]. rel2fam(S0) -> S1 = sofs:relation(S0), diff --git a/lib/compiler/src/beam_ssa_pre_codegen.erl b/lib/compiler/src/beam_ssa_pre_codegen.erl index 7f816b9802..61c42fdb6d 100644 --- a/lib/compiler/src/beam_ssa_pre_codegen.erl +++ b/lib/compiler/src/beam_ssa_pre_codegen.erl @@ -108,7 +108,8 @@ functions([], _Ps, _UseBSM3) -> []. intervals=[] :: [{b_var(),[range()]}], res=[] :: [{b_var(),reservation()}] | #{b_var():=reservation()}, regs=#{} :: #{b_var():=ssa_register()}, - extra_annos=[] :: [{atom(),term()}] + extra_annos=[] :: [{atom(),term()}], + location :: term() }). -define(PASS(N), {N,fun N/1}). @@ -119,7 +120,9 @@ passes(Opts) -> %% Preliminaries. ?PASS(fix_bs), + ?PASS(exception_trampolines), ?PASS(sanitize), + ?PASS(match_fail_instructions), case FixTuples of false -> ignore; true -> ?PASS(fix_tuples) @@ -164,7 +167,9 @@ passes(Opts) -> function(#b_function{anno=Anno,args=Args,bs=Blocks0,cnt=Count0}=F0, Ps, UseBSM3) -> try - St0 = #st{ssa=Blocks0,args=Args,use_bsm3=UseBSM3,cnt=Count0}, + Location = maps:get(location, Anno, none), + St0 = #st{ssa=Blocks0,args=Args,use_bsm3=UseBSM3, + cnt=Count0,location=Location}, St = compile:run_sub_passes(Ps, St0), #st{ssa=Blocks,cnt=Count,regs=Regs,extra_annos=ExtraAnnos} = St, F1 = add_extra_annos(F0, ExtraAnnos), @@ -598,6 +603,10 @@ bs_instrs([{L,#b_blk{is=Is0}=Blk}|Bs], CtxChain, Acc0) -> bs_instrs([], _, Acc) -> reverse(Acc). +bs_instrs_is([#b_set{op=succeeded}=I|Is], CtxChain, Acc) -> + %% This instruction refers to a specific operation, so we must not + %% substitute the context argument. + bs_instrs_is(Is, CtxChain, [I | Acc]); bs_instrs_is([#b_set{op=Op,args=Args0}=I0|Is], CtxChain, Acc) -> Args = [bs_subst_ctx(A, CtxChain) || A <- Args0], I1 = I0#b_set{args=Args}, @@ -691,6 +700,44 @@ legacy_bs_is([I|Is], Last, IsYreg, Count, Copies, Acc) -> legacy_bs_is([], _Last, _IsYreg, Count, Copies, Acc) -> {reverse(Acc),Count,Copies}. +%% exception_trampolines(St0) -> St. +%% +%% Removes the "exception trampolines" that were added to prevent exceptions +%% from being optimized away. + +exception_trampolines(#st{ssa=Blocks0}=St) -> + RPO = reverse(beam_ssa:rpo(Blocks0)), + Blocks = et_1(RPO, #{}, Blocks0), + St#st{ssa=Blocks}. + +et_1([L | Ls], Trampolines, Blocks) -> + #{ L := #b_blk{is=Is,last=Last0}=Block0 } = Blocks, + case {Is, Last0} of + {[#b_set{op=exception_trampoline}], #b_br{succ=Succ}} -> + et_1(Ls, Trampolines#{ L => Succ }, maps:remove(L, Blocks)); + {_, #b_br{succ=Same,fail=Same}} when Same =:= ?EXCEPTION_BLOCK -> + %% The exception block is just a marker saying that we should raise + %% an exception (= {f,0}) instead of jumping to a particular fail + %% block. Since it's not a reachable block we can't allow + %% unconditional jumps to it except through a trampoline. + error({illegal_jump_to_exception_block, L}); + {_, #b_br{succ=Succ0,fail=Fail0}} -> + Succ = maps:get(Succ0, Trampolines, Succ0), + Fail = maps:get(Fail0, Trampolines, Fail0), + if + Succ =/= Succ0; Fail =/= Fail0 -> + Last = Last0#b_br{succ=Succ,fail=Fail}, + Block = Block0#b_blk{last=Last}, + et_1(Ls, Trampolines, Blocks#{ L := Block }); + Succ =:= Succ0, Fail =:= Fail0 -> + et_1(Ls, Trampolines, Blocks) + end; + {_, _} -> + et_1(Ls, Trampolines, Blocks) + end; +et_1([], _Trampolines, Blocks) -> + Blocks. + %% sanitize(St0) -> St. %% Remove constructs that can cause problems later: %% @@ -856,6 +903,114 @@ prune_phi(#b_set{args=Args0}=Phi, Reachable) -> gb_sets:is_element(Pred, Reachable)], Phi#b_set{args=Args}. +%%% Rewrite certain calls to erlang:error/{1,2} to specialized +%%% instructions: +%%% +%%% erlang:error({badmatch,Value}) => badmatch Value +%%% erlang:error({case_clause,Value}) => case_end Value +%%% erlang:error({try_clause,Value}) => try_case_end Value +%%% erlang:error(if_clause) => if_end +%%% erlang:error(function_clause, Args) => jump FuncInfoLabel +%%% +%%% In SSA code, we represent those instructions as a 'match_fail' +%%% instruction with the name of the BEAM instruction as the first +%%% argument. + +match_fail_instructions(#st{ssa=Blocks0,args=Args,location=Location}=St) -> + Ls = maps:to_list(Blocks0), + Info = {length(Args),Location}, + Blocks = match_fail_instrs_1(Ls, Info, Blocks0), + St#st{ssa=Blocks}. + +match_fail_instrs_1([{L,#b_blk{is=Is0}=Blk}|Bs], Arity, Blocks0) -> + case match_fail_instrs_blk(Is0, Arity, []) of + none -> + match_fail_instrs_1(Bs, Arity, Blocks0); + Is -> + Blocks = Blocks0#{L:=Blk#b_blk{is=Is}}, + match_fail_instrs_1(Bs, Arity, Blocks) + end; +match_fail_instrs_1([], _Arity, Blocks) -> Blocks. + +match_fail_instrs_blk([#b_set{op=put_tuple,dst=Dst, + args=[#b_literal{val=Tag},Val]}, + #b_set{op=call, + args=[#b_remote{mod=#b_literal{val=erlang}, + name=#b_literal{val=error}}, + Dst]}=Call|Is], + _Arity, Acc) -> + match_fail_instr(Call, Tag, Val, Is, Acc); +match_fail_instrs_blk([#b_set{op=call, + args=[#b_remote{mod=#b_literal{val=erlang}, + name=#b_literal{val=error}}, + #b_literal{val={Tag,Val}}]}=Call|Is], + _Arity, Acc) -> + match_fail_instr(Call, Tag, #b_literal{val=Val}, Is, Acc); +match_fail_instrs_blk([#b_set{op=call, + args=[#b_remote{mod=#b_literal{val=erlang}, + name=#b_literal{val=error}}, + #b_literal{val=if_clause}]}=Call|Is], + _Arity, Acc) -> + I = Call#b_set{op=match_fail,args=[#b_literal{val=if_end}]}, + reverse(Acc, [I|Is]); +match_fail_instrs_blk([#b_set{op=call,anno=Anno, + args=[#b_remote{mod=#b_literal{val=erlang}, + name=#b_literal{val=error}}, + #b_literal{val=function_clause}, + Stk]}=Call], + {Arity,Location}, Acc) -> + case match_fail_stk(Stk, Acc, [], []) of + {[_|_]=Vars,Is} when length(Vars) =:= Arity -> + case maps:get(location, Anno, none) of + Location -> + I = Call#b_set{op=match_fail, + args=[#b_literal{val=function_clause}|Vars]}, + Is ++ [I]; + _ -> + %% erlang:error/2 has a different location than the + %% func_info instruction at the beginning of the function + %% (probably because of inlining). Keep the original call. + reverse(Acc, [Call]) + end; + _ -> + %% Either the stacktrace could not be picked apart (for example, + %% if the call to erlang:error/2 was handwritten) or the number + %% of arguments in the stacktrace was different from the arity + %% of the host function (because it is the implementation of a + %% fun). Keep the original call. + reverse(Acc, [Call]) + end; +match_fail_instrs_blk([I|Is], Arity, Acc) -> + match_fail_instrs_blk(Is, Arity, [I|Acc]); +match_fail_instrs_blk(_, _, _) -> + none. + +match_fail_instr(Call, Tag, Val, Is, Acc) -> + Op = case Tag of + badmatch -> Tag; + case_clause -> case_end; + try_clause -> try_case_end; + _ -> none + end, + case Op of + none -> + none; + _ -> + I = Call#b_set{op=match_fail,args=[#b_literal{val=Op},Val]}, + reverse(Acc, [I|Is]) + end. + +match_fail_stk(#b_var{}=V, [#b_set{op=put_list,dst=V,args=[H,T]}|Is], IAcc, VAcc) -> + match_fail_stk(T, Is, IAcc, [H|VAcc]); +match_fail_stk(#b_literal{val=[H|T]}, Is, IAcc, VAcc) -> + match_fail_stk(#b_literal{val=T}, Is, IAcc, [#b_literal{val=H}|VAcc]); +match_fail_stk(#b_literal{val=[]}, [], IAcc, VAcc) -> + {reverse(VAcc),IAcc}; +match_fail_stk(T, [#b_set{op=Op}=I|Is], IAcc, VAcc) + when Op =:= bs_get_tail; Op =:= bs_set_position -> + match_fail_stk(T, Is, [I|IAcc], VAcc); +match_fail_stk(_, _, _, _) -> none. + %%% %%% Fix tuples. %%% @@ -911,9 +1066,8 @@ use_set_tuple_element(#st{ssa=Blocks0}=St) -> Blocks = use_ste_1(RPO, Uses, Blocks0), St#st{ssa=Blocks}. -use_ste_1([L|Ls], Uses, Blocks0) -> - {Blk0,Blocks} = use_ste_across(L, Uses, Blocks0), - #b_blk{is=Is0} = Blk0, +use_ste_1([L|Ls], Uses, Blocks) -> + #b_blk{is=Is0} = Blk0 = map_get(L, Blocks), case use_ste_is(Is0, Uses) of Is0 -> use_ste_1(Ls, Uses, Blocks); @@ -976,69 +1130,6 @@ extract_ste(#b_set{op=call,dst=Dst, end; extract_ste(#b_set{}) -> none. -%%% Optimize accross blocks within a try/catch block. - -use_ste_across(L, Uses, Blocks) -> - case map_get(L, Blocks) of - #b_blk{last=#b_br{bool=#b_var{}}}=Blk -> - try - use_ste_across_1(L, Blk, Uses, Blocks) - catch - throw:not_possible -> - {Blk,Blocks} - end; - #b_blk{}=Blk -> - {Blk,Blocks} - end. - -use_ste_across_1(L, Blk0, Uses, Blocks0) -> - #b_blk{is=IsThis,last=#b_br{bool=Bool,succ=Next}} = Blk0, - case reverse(IsThis) of - [#b_set{op=succeeded,dst=Bool,args=[Result]}=Succ0, - #b_set{op=call,args=[#b_remote{}|_],dst=Result}=Call1|Prefix] -> - case is_single_use(Bool, Uses) andalso - is_n_uses(2, Result, Uses) of - true -> ok; - false -> throw(not_possible) - end, - Call2 = use_ste_across_next(Next, Uses, Blocks0), - Is = [Call1,Call2], - case use_ste_is(Is, decrement_uses(Result, Uses)) of - [#b_set{}=Call,#b_set{op=set_tuple_element}=Ste] -> - Blocks1 = use_ste_fix_next(Ste, Next, Blocks0), - Succ = Succ0#b_set{args=[Call#b_set.dst]}, - Blk = Blk0#b_blk{is=reverse(Prefix, [Call,Succ])}, - Blocks = Blocks1#{L:=Blk}, - {Blk,Blocks}; - _ -> - throw(not_possible) - end; - _ -> - throw(not_possible) - end. - -use_ste_across_next(Next, Uses, Blocks) -> - case map_get(Next, Blocks) of - #b_blk{is=[#b_set{op=call,dst=Result,args=[#b_remote{}|_]}=Call, - #b_set{op=succeeded,dst=Bool,args=[Result]}], - last=#b_br{bool=Bool}} -> - case is_single_use(Bool, Uses) andalso - is_n_uses(2, Result, Uses) of - true -> ok; - false -> throw(not_possible) - end, - Call; - #b_blk{} -> - throw(not_possible) - end. - -use_ste_fix_next(Ste, Next, Blocks) -> - Blk0 = map_get(Next, Blocks), - #b_blk{is=[#b_set{op=call},#b_set{op=succeeded}],last=Br0} = Blk0, - Br = beam_ssa:normalize(Br0#b_br{bool=#b_literal{val=true}}), - Blk = Blk0#b_blk{is=[Ste],last=Br}, - Blocks#{Next:=Blk}. - %% Count how many times each variable is used. count_uses(Blocks) -> @@ -1048,7 +1139,7 @@ count_uses_blk([#b_blk{is=Is,last=Last}|Bs], CountMap0) -> F = fun(I, CountMap) -> foldl(fun(Var, Acc) -> case Acc of - #{Var:=3} -> Acc; + #{Var:=2} -> Acc; #{Var:=C} -> Acc#{Var:=C+1}; #{} -> Acc#{Var=>1} end @@ -1058,16 +1149,6 @@ count_uses_blk([#b_blk{is=Is,last=Last}|Bs], CountMap0) -> count_uses_blk(Bs, CountMap); count_uses_blk([], CountMap) -> CountMap. -decrement_uses(V, Uses) -> - #{V:=C} = Uses, - Uses#{V:=C-1}. - -is_n_uses(N, V, Uses) -> - case Uses of - #{V:=N} -> true; - #{} -> false - end. - is_single_use(V, Uses) -> case Uses of #{V:=1} -> true; @@ -1189,10 +1270,10 @@ place_frame_here(L, Blocks, Doms, Frames) -> Descendants = beam_ssa:rpo([L], Blocks), PhiPredecessors = phi_predecessors(L, Blocks), MustDominate = ordsets:from_list(PhiPredecessors ++ Descendants), - Dominates = all(fun(?BADARG_BLOCK) -> + Dominates = all(fun(?EXCEPTION_BLOCK) -> %% This block defines no variables and calls %% erlang:error(badarg). It does not matter - %% whether L dominates ?BADARG_BLOCK or not; + %% whether L dominates ?EXCEPTION_BLOCK or not; %% it is still safe to put the frame in L. true; (Bl) -> @@ -1340,9 +1421,9 @@ recv_common(_Defs, none, _Blocks) -> %% in the tail position of a function. []; recv_common(Defs, Exit, Blocks) -> - {ExitDefs,ExitUsed} = beam_ssa:def_used([Exit], Blocks), + {ExitDefs,ExitUnused} = beam_ssa:def_unused([Exit], Defs, Blocks), Def = ordsets:subtract(Defs, ExitDefs), - ordsets:intersection(Def, ExitUsed). + ordsets:subtract(Def, ExitUnused). %% recv_crit_edges([RemoveMessageLabel], LoopExit, %% Blocks0, Count0) -> {Blocks,Count}. @@ -1447,9 +1528,9 @@ exit_predecessors([], _Exit, _Blocks) -> []. %% later used within a clause of the receive. fix_receive([L|Ls], Defs, Blocks0, Count0) -> - {RmDefs,Used0} = beam_ssa:def_used([L], Blocks0), + {RmDefs,Unused} = beam_ssa:def_unused([L], Defs, Blocks0), Def = ordsets:subtract(Defs, RmDefs), - Used = ordsets:intersection(Def, Used0), + Used = ordsets:subtract(Def, Unused), {NewVars,Count} = new_vars([Base || #b_var{name=Base} <- Used], Count0), Ren = zip(Used, NewVars), Blocks1 = beam_ssa:rename_vars(Ren, [L], Blocks0), @@ -1483,8 +1564,8 @@ find_loop_exit(_, _) -> %% loop exit block. none. -find_loop_exit_1([?BADARG_BLOCK|Ls], RmSet, Dominators) -> - %% ?BADARG_BLOCK is a marker and not an actual block, so it is not +find_loop_exit_1([?EXCEPTION_BLOCK|Ls], RmSet, Dominators) -> + %% ?EXCEPTION_BLOCK is a marker and not an actual block, so it is not %% the block we are looking for. find_loop_exit_1(Ls, RmSet, Dominators); find_loop_exit_1([L|Ls], RmSet, Dominators) -> @@ -1756,7 +1837,7 @@ collect_yregs([], Yregs) -> Yregs. copy_retval_2([L|Ls], Yregs, Copy0, Blocks0, Count0) -> #b_blk{is=Is0,last=Last} = Blk = map_get(L, Blocks0), RC = case {Last,Ls} of - {#b_br{succ=Succ,fail=?BADARG_BLOCK},[Succ|_]} -> + {#b_br{succ=Succ,fail=?EXCEPTION_BLOCK},[Succ|_]} -> true; {_,_} -> false @@ -2105,8 +2186,8 @@ reserve_yregs(#st{frames=Frames}=St0) -> reserve_yregs_1(L, #st{ssa=Blocks0,cnt=Count0,res=Res0}=St) -> 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), + {Def,Unused} = beam_ssa:def_unused([L], Yregs, Blocks0), + UsedYregs = ordsets:subtract(Yregs, Unused), DefBefore = ordsets:subtract(UsedYregs, Def), {BeforeVars,Blocks,Count} = rename_vars(DefBefore, L, Blocks0, Count0), InsideVars = ordsets:subtract(UsedYregs, DefBefore), @@ -2495,9 +2576,9 @@ 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}, +reserve_terminator(_L, _Is, #b_br{bool=#b_var{},succ=Succ,fail=?EXCEPTION_BLOCK}, _Blocks, XsMap, _Res) -> - %% We know that no variables are used at ?BADARG_BLOCK, so + %% We know that no variables are used at ?EXCEPTION_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}, diff --git a/lib/compiler/src/beam_ssa_share.erl b/lib/compiler/src/beam_ssa_share.erl index 73983bd34a..fa728992f8 100644 --- a/lib/compiler/src/beam_ssa_share.erl +++ b/lib/compiler/src/beam_ssa_share.erl @@ -117,8 +117,8 @@ share_terminator(_Last, _Blocks) -> none. %% 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. +are_equivalent(_Succ, _, ?EXCEPTION_BLOCK, _, _Blocks) -> + %% ?EXCEPTION_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) -> diff --git a/lib/compiler/src/beam_ssa_type.erl b/lib/compiler/src/beam_ssa_type.erl index 3c06c83e2e..d93191c689 100644 --- a/lib/compiler/src/beam_ssa_type.erl +++ b/lib/compiler/src/beam_ssa_type.erl @@ -22,11 +22,12 @@ -export([opt_start/4, opt_continue/4, opt_finish/3]). -include("beam_ssa_opt.hrl"). --import(lists, [all/2,any/2,droplast/1,foldl/3,last/1,member/2, - keyfind/3,reverse/1,reverse/2, - sort/1,split/2]). +-include("beam_types.hrl"). --define(UNICODE_INT, #t_integer{elements={0,16#10FFFF}}). +-import(lists, [all/2,any/2,droplast/1,duplicate/2,foldl/3,last/1,member/2, + keyfind/3,reverse/1,sort/1,split/2,zip/2]). + +-define(UNICODE_MAX, (16#10FFFF)). -record(d, {ds :: #{beam_ssa:b_var():=beam_ssa:b_set()}, @@ -37,21 +38,6 @@ sub = #{} :: #{beam_ssa:b_var():=beam_ssa:value()}, ret_type = [] :: [type()]}). --define(ATOM_SET_SIZE, 5). - -%% Records that represent type information. --record(t_atom, {elements=any :: 'any' | [atom()]}). --record(t_integer, {elements=any :: 'any' | {integer(),integer()}}). --record(t_bs_match, {type :: type()}). --record(t_tuple, {size=0 :: integer(), - exact=false :: boolean(), - %% Known element types (1-based index), unknown elements are - %% are assumed to be 'any'. - elements=#{} :: #{ non_neg_integer() => type() }}). - --type type() :: 'any' | 'none' | - #t_atom{} | #t_integer{} | #t_bs_match{} | #t_tuple{} | - {'binary',pos_integer()} | 'cons' | 'float' | 'list' | 'map' | 'nil' | 'number'. -type type_db() :: #{beam_ssa:var_name():=type()}. -spec opt_start(Linear, Args, Anno, FuncDb) -> {Linear, FuncDb} when @@ -98,7 +84,8 @@ join_arg_types(Args, ArgTypes, Anno) -> 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))}); + Type = beam_types:join(maps:values(TM)), + join_arg_types_1(Args, TMs, Ts#{ Arg => Type }); join_arg_types_1([Arg | Args], [_TM | TMs], Ts) -> join_arg_types_1(Args, TMs, Ts#{ Arg => any }); join_arg_types_1([], [], Ts) -> @@ -122,7 +109,7 @@ opt_continue_1(Linear0, Args, Id, Ts, FuncDb0) -> D = #d{ func_db=FuncDb0, func_id=Id, ds=Defs, - ls=#{0=>Ts,?BADARG_BLOCK=>#{}}, + ls=#{0=>Ts,?EXCEPTION_BLOCK=>#{}}, once=UsedOnce }, {Linear, FuncDb, NewRet} = opt(Linear0, D, []), @@ -157,43 +144,15 @@ opt_finish_1([Arg | Args], [TypeMap | TypeMaps], ParamInfo) 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); - none -> - %% This function will never be called. Pretend that we don't - %% know the type for this argument. - opt_finish_1(Args, TypeMaps, ParamInfo0); - JoinedType -> - JoinedType = verified_type(JoinedType), - ParamInfo = ParamInfo0#{ Arg => validator_anno(JoinedType) }, - opt_finish_1(Args, TypeMaps, ParamInfo) - end; + JoinedType = beam_types:join(maps:values(TypeMap)), + ParamInfo = case JoinedType of + any -> ParamInfo0; + _ -> ParamInfo0#{ Arg => JoinedType } + end, + opt_finish_1(Args, TypeMaps, ParamInfo); opt_finish_1([], [], ParamInfo) -> ParamInfo. -validator_anno(#t_tuple{size=Size,exact=Exact,elements=Elements0}) -> - Elements = maps:fold(fun(Index, Type, Acc) -> - Key = beam_validator:type_anno(integer, Index), - Acc#{ Key => validator_anno(Type) } - end, #{}, Elements0), - beam_validator:type_anno(tuple, Size, Exact, Elements); -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}. @@ -212,27 +171,17 @@ opt([], D, Acc) -> opt_1(L, #b_blk{is=Is0,last=Last0}=Blk0, Bs, Ts0, #d{ds=Ds0,sub=Sub0,func_db=Fdb0}=D0, Acc) -> - case opt_is(Is0, Ts0, Ds0, Fdb0, D0, Sub0, []) of - {Is,Ts,Ds,Fdb,Sub} -> - 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}, - opt(Bs, D, [{L,Blk}|Acc]); - {no_return,Ret,Is,Ds,Fdb,Sub} -> - %% This call will never reach the successor block. - %% Rewrite the terminator to a 'ret', and remove - %% all type information for this label. That can - %% potentially narrow the type of the phi node - %% in the former successor. - Ls = maps:remove(L, D0#d.ls), - RetType = join([none|D0#d.ret_type]), - D = D0#d{ds=Ds,ls=Ls,sub=Sub, - func_db=Fdb,ret_type=[RetType]}, - Blk = Blk0#b_blk{is=Is,last=Ret}, - opt(Bs, D, [{L,Blk}|Acc]) - end. + {Is,Ts,Ds,Fdb,Sub} = opt_is(Is0, Ts0, Ds0, Fdb0, D0, Sub0, []), + + D1 = D0#d{ds=Ds,sub=Sub,func_db=Fdb}, + + Last1 = simplify_terminator(Last0, Sub, Ts, Ds), + Last2 = opt_terminator(Last1, Ts, Ds), + + {Last, D} = update_successors(Last2, Ts, D1), + + Blk = Blk0#b_blk{is=Is,last=Last}, + opt(Bs, D, [{L,Blk}|Acc]). simplify_terminator(#b_br{bool=Bool}=Br, Sub, Ts, _Ds) -> Br#b_br{bool=simplify_arg(Bool, Sub, Ts)}; @@ -265,203 +214,138 @@ opt_is([#b_set{op=phi,dst=Dst,args=Args0}=I0|Is], Ds = Ds0#{Dst=>I}, opt_is(Is, Ts, Ds, Fdb, D, Sub0, [I|Acc]) end; -opt_is([#b_set{op=call,args=Args0,dst=Dst}=I0|Is], +opt_is([#b_set{op=call,args=Args0}=I0|Is], + Ts, Ds, Fdb0, D, Sub, Acc) -> + Args = simplify_args(Args0, Sub, Ts), + I1 = beam_ssa:normalize(I0#b_set{args=Args}), + {I, Fdb} = opt_call(I1, Ts, Fdb0, D), + opt_simplify(I, Is, Ts, Ds, Fdb, D, Sub, Acc); +opt_is([#b_set{op=make_fun,args=Args0}=I0|Is], Ts0, Ds0, Fdb0, D, Sub0, Acc) -> Args = simplify_args(Args0, Sub0, Ts0), I1 = beam_ssa:normalize(I0#b_set{args=Args}), - {Ts1,Ds,Fdb,I2} = opt_call(I1, D, Ts0, Ds0, Fdb0), - case {map_get(Dst, Ts1),Is} of - {Type,[#b_set{op=succeeded}]} when Type =/= none -> - %% This call instruction is inside a try/catch - %% block. Don't attempt to simplify it. - opt_is(Is, Ts1, Ds, Fdb, D, Sub0, [I2|Acc]); - {none,[#b_set{op=succeeded}]} -> - %% This call instruction is inside a try/catch - %% block, but we know it will never return and - %% later optimizations may try to exploit that. - %% - %% For example, if we have an expression that - %% either returns this call or a tuple, we know - %% that the expression always returns a tuple - %% and can turn a later element/3 into - %% get_tuple_element. - %% - %% This is sound but difficult to validate in a - %% meaningful way as try/catch currently forces - %% us to maintain the illusion that the success - %% block is reachable even when its not, so we - %% disable the optimization to keep things - %% simple. - Ts = Ts1#{ Dst := any }, - opt_is(Is, Ts, Ds, Fdb, D, Sub0, [I2|Acc]); - {none,_} -> - %% This call never returns. The rest of the - %% instructions will not be executed. - Ret = #b_ret{arg=Dst}, - {no_return,Ret,reverse(Acc, [I2]),Ds,Fdb,Sub0}; - {_,_} -> - case simplify_call(I2) of - #b_set{}=I -> - opt_is(Is, Ts1, Ds, Fdb, D, Sub0, [I|Acc]); - #b_literal{}=Lit -> - Sub = Sub0#{Dst=>Lit}, - Ts = maps:remove(Dst, Ts1), - opt_is(Is, Ts, Ds0, Fdb, D, Sub, Acc); - #b_var{}=Var -> - Ts = maps:remove(Dst, Ts1), - Sub = Sub0#{Dst=>Var}, - opt_is(Is, Ts, Ds0, Fdb, D, Sub, Acc) - end - end; -opt_is([#b_set{op=succeeded,args=[Arg],dst=Dst}=I], + {Ts,Ds,Fdb,I} = opt_make_fun(I1, D, Ts0, Ds0, Fdb0), + opt_is(Is, Ts, Ds, Fdb, D, Sub0, [I|Acc]); +opt_is([#b_set{op=succeeded,args=[Arg],dst=Dst,anno=Anno}=I], Ts0, Ds0, Fdb, 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, 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, D, Sub, Acc); - none -> - Ts = Ts0#{Dst=>Type}, - Ds = Ds0#{Dst=>I}, - opt_is([], Ts, Ds, Fdb, D, Sub0, [I|Acc]) - end + Type = case Ds0 of + #{ Arg := #b_set{op=call} } -> + %% Calls can always throw exceptions and their return types + %% are what they return on success, so we must avoid + %% simplifying arguments in case `Arg` would become a + %% literal, which would trick 'succeeded' into thinking it + %% can't fail. + type(succeeded, [Arg], Anno, Ts0, Ds0); + #{} -> + Args = simplify_args([Arg], Sub0, Ts0), + type(succeeded, Args, Anno, Ts0, Ds0) + end, + case beam_types:get_singleton_value(Type) of + {ok, Lit} -> + Sub = Sub0#{ Dst => #b_literal{val=Lit} }, + opt_is([], Ts0, Ds0, Fdb, D, Sub, Acc); + error -> + Ts = Ts0#{ Dst => Type }, + Ds = Ds0#{ Dst => I }, + opt_is([], Ts, Ds, Fdb, D, Sub0, [I | Acc]) end; -opt_is([#b_set{args=Args0,dst=Dst}=I0|Is], - Ts0, Ds0, Fdb, D, Sub0, Acc) -> - Args = simplify_args(Args0, Sub0, Ts0), - I1 = beam_ssa:normalize(I0#b_set{args=Args}), - case simplify(I1, Ts0) of +opt_is([#b_set{args=Args0}=I0|Is], + Ts, Ds, Fdb, D, Sub, Acc) -> + Args = simplify_args(Args0, Sub, Ts), + I = beam_ssa:normalize(I0#b_set{args=Args}), + opt_simplify(I, Is, Ts, Ds, Fdb, D, Sub, Acc); +opt_is([], Ts, Ds, Fdb, _D, Sub, Acc) -> + {reverse(Acc), Ts, Ds, Fdb, Sub}. + +opt_simplify(#b_set{dst=Dst}=I0, Is, Ts0, Ds0, Fdb, D, Sub0, Acc) -> + case simplify(I0, Ts0) of #b_set{}=I2 -> I = beam_ssa:normalize(I2), Ts = update_types(I, Ts0, Ds0), - Ds = Ds0#{Dst=>I}, + Ds = Ds0#{ Dst => I }, opt_is(Is, Ts, Ds, Fdb, D, Sub0, [I|Acc]); #b_literal{}=Lit -> - Sub = Sub0#{Dst=>Lit}, + Sub = Sub0#{ Dst => Lit }, opt_is(Is, Ts0, Ds0, Fdb, 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}}, + %% We must remove this 'succeeded' instruction since the + %% variable it checks is gone. + Sub = Sub0#{ Dst => Var, SuccDst => #b_literal{val=true} }, opt_is([], Ts0, Ds0, Fdb, D, Sub, Acc); _ -> - Sub = Sub0#{Dst=>Var}, + Sub = Sub0#{ Dst => Var}, opt_is(Is, Ts0, Ds0, Fdb, D, Sub, Acc) end - end; -opt_is([], Ts, Ds, Fdb, _D, Sub, Acc) -> - {reverse(Acc), Ts, Ds, Fdb, Sub}. - -simplify_call(#b_set{op=call,args=[#b_remote{}=Rem|Args]}=I) -> - case Rem of - #b_remote{mod=#b_literal{val=Mod}, - name=#b_literal{val=Name}} -> - case erl_bifs:is_pure(Mod, Name, length(Args)) of - true -> - simplify_remote_call(Mod, Name, Args, I); - false -> - I - end; - #b_remote{} -> - I - end; -simplify_call(I) -> I. - -%% Simplify a remote call to a pure BIF. -simplify_remote_call(erlang, '++', [#b_literal{val=[]},Tl], _I) -> - Tl; -simplify_remote_call(erlang, setelement, - [#b_literal{val=Pos}, - #b_literal{val=Tuple}, - #b_var{}=Value], I) - when is_integer(Pos), 1 =< Pos, Pos =< tuple_size(Tuple) -> - %% Position is a literal integer and the shape of the - %% tuple is known. - Els0 = [#b_literal{val=El} || El <- tuple_to_list(Tuple)], - {Bef,[_|Aft]} = split(Pos - 1, Els0), - Els = Bef ++ [Value|Aft], - I#b_set{op=put_tuple,args=Els}; -simplify_remote_call(Mod, Name, Args0, I) -> - case make_literal_list(Args0) of - none -> - I; - Args -> - %% The arguments are literals. Try to evaluate the BIF. - try apply(Mod, Name, Args) of - Val -> - case cerl:is_literal_term(Val) of - true -> - #b_literal{val=Val}; - false -> - %% The value can't be expressed as a literal - %% (e.g. a pid). - I - end - catch - _:_ -> - %% Failed. Don't bother trying to optimize - %% the call. - I - end end. -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), +opt_call(#b_set{dst=Dst,args=[#b_local{}=Callee|Args]}=I0, Ts, Fdb0, D) -> + I = opt_local_call_return(I0, Callee, Fdb0), case Fdb0 of #{ Callee := #func_info{exported=false,arg_types=ArgTypes0}=Info } -> + %% Match contexts are treated as bitstrings when optimizing + %% arguments, as we don't yet support removing the + %% "bs_start_match3" instruction. + Types = [case raw_type(Arg, Ts) of + #t_bs_context{} -> #t_bitstring{}; + Type -> Type + end || Arg <- Args], + %% 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), + ArgTypes = update_arg_types(Types, ArgTypes0, CallId), Fdb = Fdb0#{ Callee => Info#func_info{arg_types=ArgTypes} }, - {Ts, Ds, Fdb, I}; + {I, Fdb}; #{} -> %% 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} + {I, Fdb0} end; -opt_call(#b_set{dst=Dst}=I, _D, Ts0, Ds0, Fdb) -> +opt_call(I, _Ts, Fdb, _D) -> + {I, Fdb}. + +opt_local_call_return(I, Callee, Fdb) -> + case Fdb of + #{ Callee := #func_info{ret_type=[Type]} } when Type =/= any -> + beam_ssa:add_anno(result_type, Type, I); + #{} -> + I + end. + +%% While we have no way to know which arguments a fun will be called with, we +%% do know its free variables and can update their types as if this were a +%% local call. +opt_make_fun(#b_set{op=make_fun, + dst=Dst, + args=[#b_local{}=Callee | FreeVars]}=I, + D, Ts0, Ds0, Fdb0) -> Ts = update_types(I, Ts0, Ds0), Ds = Ds0#{ Dst => I }, - {Ts, Ds, Fdb, I}. + case Fdb0 of + #{ Callee := #func_info{exported=false,arg_types=ArgTypes0}=Info } -> + ArgCount = Callee#b_local.arity - length(FreeVars), -opt_local_call(#b_set{dst=Dst,args=[Id|_]}=I0, Ts0, Ds0, Fdb) -> - Type = case Fdb of - #{ Id := #func_info{ret_type=[T]} } -> T; - #{} -> any - end, - I = case Type of - any -> I0; - none -> 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, - TypeMap = TypeMap0#{ CallId => NewType }, - [TypeMap | update_arg_types(Args, TypeMaps, CallId, Ts)]; -update_arg_types([], [], _CallId, _Ts) -> + FVTypes = [raw_type(FreeVar, Ts) || FreeVar <- FreeVars], + Types = duplicate(ArgCount, any) ++ FVTypes, + + CallId = {D#d.func_id, Dst}, + ArgTypes = update_arg_types(Types, ArgTypes0, CallId), + + 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. + +update_arg_types([ArgType | ArgTypes], [TypeMap0 | TypeMaps], CallId) -> + TypeMap = TypeMap0#{ CallId => ArgType }, + [TypeMap | update_arg_types(ArgTypes, TypeMaps, CallId)]; +update_arg_types([], [], _CallId) -> []. simplify(#b_set{op={bif,'and'},args=Args}=I, Ts) -> @@ -487,8 +371,10 @@ simplify(#b_set{op={bif,'or'},args=Args}=I, Ts) -> I end; simplify(#b_set{op={bif,element},args=[#b_literal{val=Index},Tuple]}=I0, Ts) -> - case t_tuple_size(get_type(Tuple, Ts)) of - {_,Size} when is_integer(Index), 1 =< Index, Index =< Size -> + case normalized_type(Tuple, Ts) of + #t_tuple{size=Size} when is_integer(Index), + 1 =< Index, + Index =< Size -> I = I0#b_set{op=get_tuple_element, args=[Tuple,#b_literal{val=Index-1}]}, simplify(I, Ts); @@ -496,67 +382,97 @@ simplify(#b_set{op={bif,element},args=[#b_literal{val=Index},Tuple]}=I0, Ts) -> eval_bif(I0, Ts) end; simplify(#b_set{op={bif,hd},args=[List]}=I, Ts) -> - case get_type(List, Ts) of + case normalized_type(List, Ts) of cons -> I#b_set{op=get_hd}; _ -> eval_bif(I, Ts) end; simplify(#b_set{op={bif,tl},args=[List]}=I, Ts) -> - case get_type(List, Ts) of + case normalized_type(List, Ts) of cons -> I#b_set{op=get_tl}; _ -> eval_bif(I, Ts) end; simplify(#b_set{op={bif,size},args=[Term]}=I, Ts) -> - case get_type(Term, Ts) of + case normalized_type(Term, Ts) of #t_tuple{} -> 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 + case normalized_type(Term, Ts) of #t_tuple{size=Size,exact=true} -> #b_literal{val=Size}; _ -> I end; -simplify(#b_set{op={bif,'=='},args=Args}=I, Ts) -> - Types = get_types(Args, Ts), - EqEq = case {meet(Types),join(Types)} of - {none,any} -> true; - {#t_integer{},#t_integer{}} -> true; - {float,float} -> true; - {{binary,_},_} -> true; - {#t_atom{},_} -> true; - {_,_} -> false - end, +simplify(#b_set{op={bif,is_function},args=[Fun,#b_literal{val=Arity}]}=I, Ts) + when is_integer(Arity), Arity >= 0 -> + case normalized_type(Fun, Ts) of + #t_fun{arity=any} -> + I; + #t_fun{arity=Arity} -> + #b_literal{val=true}; + any -> + I; + _ -> + #b_literal{val=false} + end; +simplify(#b_set{op={bif,Op0},args=Args}=I, Ts) when Op0 =:= '=='; Op0 =:= '/=' -> + Types = normalized_types(Args, Ts), + EqEq0 = case {beam_types:meet(Types),beam_types:join(Types)} of + {none,any} -> true; + {#t_integer{},#t_integer{}} -> true; + {float,float} -> true; + {#t_bitstring{},_} -> true; + {#t_atom{},_} -> true; + {_,_} -> false + end, + EqEq = EqEq0 orelse any_non_numeric_argument(Args, Ts), case EqEq of true -> - simplify(I#b_set{op={bif,'=:='}}, Ts); + Op = case Op0 of + '==' -> '=:='; + '/=' -> '=/=' + end, + simplify(I#b_set{op={bif,Op}}, Ts); false -> eval_bif(I, Ts) end; simplify(#b_set{op={bif,'=:='},args=[Same,Same]}, _Ts) -> #b_literal{val=true}; -simplify(#b_set{op={bif,'=:='},args=[A1,_A2]=Args}=I, Ts) -> - [T1,T2] = get_types(Args, Ts), - case meet(T1, T2) of +simplify(#b_set{op={bif,'=:='},args=[LHS,RHS]}=I, Ts) -> + LType = raw_type(LHS, Ts), + RType = raw_type(RHS, Ts), + case beam_types:meet(LType, RType) of none -> #b_literal{val=false}; _ -> - case {t_is_boolean(T1),T2} of + case {beam_types:is_boolean_type(LType), + beam_types:normalize(RType)} of {true,#t_atom{elements=[true]}} -> %% Bool =:= true ==> Bool - A1; + LHS; + {true,#t_atom{elements=[false]}} -> + %% Bool =:= false ==> not Bool + %% + %% This will be further optimized to eliminate the + %% 'not', swapping the success and failure + %% branches in the br instruction. If LHS comes + %% from a type test (such as is_atom/1) or a + %% comparison operator (such as >=) that can be + %% translated to test instruction, this + %% optimization will eliminate one instruction. + simplify(I#b_set{op={bif,'not'},args=[LHS]}, Ts); {_,_} -> eval_bif(I, Ts) end end; simplify(#b_set{op={bif,Op},args=Args}=I, Ts) -> - Types = get_types(Args, Ts), + Types = normalized_types(Args, Ts), case is_float_op(Op, Types) of false -> eval_bif(I, Ts); @@ -565,20 +481,15 @@ simplify(#b_set{op={bif,Op},args=Args}=I, Ts) -> eval_bif(beam_ssa:add_anno(float_op, AnnoArgs, I), Ts) end; simplify(#b_set{op=get_tuple_element,args=[Tuple,#b_literal{val=N}]}=I, Ts) -> - case get_type(Tuple, Ts) of - #t_tuple{size=Size,elements=Es} when Size > N -> - ElemType = get_element_type(N + 1, Es), - case get_literal_from_type(ElemType) of - #b_literal{}=Lit -> Lit; - none -> I - end; - none -> - %% Will never be executed because of type conflict. - %% #b_literal{val=ignored}; - I + #t_tuple{size=Size,elements=Es} = normalized_type(Tuple, Ts), + true = Size > N, %Assertion. + ElemType = beam_types:get_element_type(N + 1, Es), + case beam_types:get_singleton_value(ElemType) of + {ok, Val} -> #b_literal{val=Val}; + error -> I end; simplify(#b_set{op=is_nonempty_list,args=[Src]}=I, Ts) -> - case get_type(Src, Ts) of + case normalized_type(Src, Ts) of any -> I; list -> I; cons -> #b_literal{val=true}; @@ -586,7 +497,7 @@ simplify(#b_set{op=is_nonempty_list,args=[Src]}=I, Ts) -> end; simplify(#b_set{op=is_tagged_tuple, args=[Src,#b_literal{val=Size},#b_literal{}=Tag]}=I, Ts) -> - simplify_is_record(I, get_type(Src, Ts), Size, Tag, Ts); + simplify_is_record(I, normalized_type(Src, Ts), Size, Tag, Ts); simplify(#b_set{op=put_list,args=[#b_literal{val=H}, #b_literal{val=T}]}, _Ts) -> #b_literal{val=[H|T]}; @@ -599,8 +510,97 @@ 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(#b_set{op=call,args=[#b_remote{}=Rem|Args]}=I, _Ts) -> + case Rem of + #b_remote{mod=#b_literal{val=Mod}, + name=#b_literal{val=Name}} -> + case erl_bifs:is_pure(Mod, Name, length(Args)) of + true -> + simplify_remote_call(Mod, Name, Args, I); + false -> + I + end; + #b_remote{} -> + I + end; simplify(I, _Ts) -> I. +%% Simplify a remote call to a pure BIF. +simplify_remote_call(erlang, '++', [#b_literal{val=[]},Tl], _I) -> + Tl; +simplify_remote_call(erlang, setelement, + [#b_literal{val=Pos}, + #b_literal{val=Tuple}, + #b_var{}=Value], I) + when is_integer(Pos), 1 =< Pos, Pos =< tuple_size(Tuple) -> + %% Position is a literal integer and the shape of the + %% tuple is known. + Els0 = [#b_literal{val=El} || El <- tuple_to_list(Tuple)], + {Bef,[_|Aft]} = split(Pos - 1, Els0), + Els = Bef ++ [Value|Aft], + I#b_set{op=put_tuple,args=Els}; +simplify_remote_call(Mod, Name, Args0, I) -> + case make_literal_list(Args0) of + none -> + I; + Args -> + %% The arguments are literals. Try to evaluate the BIF. + try apply(Mod, Name, Args) of + Val -> + case cerl:is_literal_term(Val) of + true -> + #b_literal{val=Val}; + false -> + %% The value can't be expressed as a literal + %% (e.g. a pid). + I + end + catch + _:_ -> + %% Failed. Don't bother trying to optimize + %% the call. + I + end + end. + +any_non_numeric_argument([#b_literal{val=Lit}|_], _Ts) -> + is_non_numeric(Lit); +any_non_numeric_argument([#b_var{}=V|T], Ts) -> + is_non_numeric_type(raw_type(V, Ts)) orelse any_non_numeric_argument(T, Ts); +any_non_numeric_argument([], _Ts) -> false. + +is_non_numeric([H|T]) -> + is_non_numeric(H) andalso is_non_numeric(T); +is_non_numeric(Tuple) when is_tuple(Tuple) -> + is_non_numeric_tuple(Tuple, tuple_size(Tuple)); +is_non_numeric(Map) when is_map(Map) -> + %% Note that 17.x and 18.x compare keys in different ways. + %% Be very conservative -- require that both keys and values + %% are non-numeric. + is_non_numeric(maps:to_list(Map)); +is_non_numeric(Num) when is_number(Num) -> + false; +is_non_numeric(_) -> true. + +is_non_numeric_tuple(Tuple, El) when El >= 1 -> + is_non_numeric(element(El, Tuple)) andalso + is_non_numeric_tuple(Tuple, El-1); +is_non_numeric_tuple(_Tuple, 0) -> true. + +is_non_numeric_type(#t_atom{}) -> true; +is_non_numeric_type(#t_bitstring{}) -> true; +is_non_numeric_type(nil) -> true; +is_non_numeric_type(#t_tuple{size=Size,exact=true,elements=Types}) + when map_size(Types) =:= Size -> + is_non_numeric_tuple_type(Size, Types); +is_non_numeric_type(_) -> false. + +is_non_numeric_tuple_type(0, _Types) -> + true; +is_non_numeric_tuple_type(Pos, Types) -> + is_non_numeric_type(map_get(Pos, Types)) andalso + is_non_numeric_tuple_type(Pos - 1, Types). + make_literal_list(Args) -> make_literal_list(Args, []). @@ -611,9 +611,11 @@ make_literal_list([_|_], _) -> 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). +is_safe_bool_op([LHS, RHS], Ts) -> + LType = raw_type(LHS, Ts), + RType = raw_type(RHS, Ts), + beam_types:is_boolean_type(LType) andalso + beam_types:is_boolean_type(RType). all_same([{H,_}|T]) -> all(fun({E,_}) -> E =:= H end, T). @@ -626,7 +628,7 @@ eval_bif(#b_set{op={bif,Bif},args=Args}=I, Ts) -> true -> case make_literal_list(Args) of none -> - case get_types(Args, Ts) of + case normalized_types(Args, Ts) of [any] -> I; [Type] -> @@ -659,10 +661,9 @@ simplify_arg(#b_var{}=Arg0, Sub, Ts) -> #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 + case beam_types:get_singleton_value(raw_type(Arg, Ts)) of + {ok, Val} -> #b_literal{val=Val}; + error -> Arg end end; simplify_arg(#b_remote{mod=Mod,name=Name}=Rem, Sub, Ts) -> @@ -701,7 +702,7 @@ opt_terminator(#b_br{bool=#b_var{}}=Br, Ts, Ds) -> opt_terminator(#b_switch{arg=#b_literal{}}=Sw, _Ts, _Ds) -> beam_ssa:normalize(Sw); opt_terminator(#b_switch{arg=#b_var{}=V}=Sw, Ts, Ds) -> - case get_type(V, Ts) of + case normalized_type(V, Ts) of any -> beam_ssa:normalize(Sw); Type -> @@ -717,7 +718,7 @@ opt_switch(#b_switch{fail=Fail,list=List0}=Sw0, Type, Ts, Ds) -> #t_integer{elements={_,_}=Range} -> simplify_switch_int(Sw1, Range); #t_atom{elements=[_|_]} -> - case t_is_boolean(Type) of + case beam_types:is_boolean_type(Type) of true -> #b_br{} = Br = simplify_switch_bool(Sw1, Ts, Ds), opt_terminator(Br, Ts, Ds); @@ -731,7 +732,7 @@ opt_switch(#b_switch{fail=Fail,list=List0}=Sw0, Type, Ts, Ds) -> prune_switch_list([{_,Fail}|T], Fail, Type, Ts) -> prune_switch_list(T, Fail, Type, Ts); prune_switch_list([{Arg,_}=Pair|T], Fail, Type, Ts) -> - case meet(get_type(Arg, Ts), Type) of + case beam_types:meet(raw_type(Arg, Ts), Type) of none -> %% Different types. This value can never match. prune_switch_list(T, Fail, Type, Ts); @@ -740,82 +741,91 @@ prune_switch_list([{Arg,_}=Pair|T], Fail, Type, Ts) -> end; prune_switch_list([], _, _, _) -> []. -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{}=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 it in the type database passed on to - %% the successors of this block. - Ts = maps:remove(Bool, Ts0), - {SuccTs,FailTs} = infer_types_br(Bool, Ts, D0), - D = update_successor(Fail, FailTs, D0), - update_successor(Succ, SuccTs, D); - false -> - {SuccTs,FailTs} = infer_types_br(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}, Ts, 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 it in the type database passed on to - %% the successors of this block. - D = update_successor(Fail, Ts, D0), - F = fun({Val,S}, A) -> - SuccTs0 = infer_types_switch(V, Val, Ts, D), - SuccTs = maps:remove(V, SuccTs0), - update_successor(S, SuccTs, 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, Ts), - D = update_successor(Fail, FailTs, D0), - F = fun({Val,S}, A) -> - SuccTs = infer_types_switch(V, Val, Ts, D), - update_successor(S, SuccTs, A) - end, - foldl(F, D, List) +update_successors(#b_br{bool=#b_literal{val=true},succ=Succ}=Last, Ts, D0) -> + {Last, update_successor(Succ, Ts, D0)}; +update_successors(#b_br{bool=#b_var{}=Bool,succ=Succ,fail=Fail}=Last0, + Ts, D0) -> + UsedOnce = cerl_sets:is_element(Bool, D0#d.once), + case infer_types_br(Bool, Ts, UsedOnce, D0) of + {#{}=SuccTs, #{}=FailTs} -> + D1 = update_successor(Succ, SuccTs, D0), + D = update_successor(Fail, FailTs, D1), + {Last0, D}; + {#{}=SuccTs, none} -> + Last = Last0#b_br{bool=#b_literal{val=true},fail=Succ}, + {Last, update_successor(Succ, SuccTs, D0)}; + {none, #{}=FailTs} -> + Last = Last0#b_br{bool=#b_literal{val=true},succ=Fail}, + {Last, update_successor(Fail, FailTs, D0)} 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; +update_successors(#b_switch{arg=#b_var{}=V,fail=Fail0,list=List0}=Last0, + Ts, D0) -> + UsedOnce = cerl_sets:is_element(V, D0#d.once), + + {List1, D1} = update_switch(List0, V, Ts, UsedOnce, [], D0), + FailTs = update_switch_failure(V, List0, Ts, UsedOnce, D1), + + case FailTs of + none -> + %% The fail block is unreachable; swap it with one of the choices. + [{_, Fail} | List] = List1, + Last = Last0#b_switch{fail=Fail,list=List}, + {Last, D1}; #{} -> - RetType = join([get_type(Arg, Ts) | D#d.ret_type]), - D#d{ret_type=[RetType]} - end. + D = update_successor(Fail0, FailTs, D1), + Last = Last0#b_switch{list=List1}, + {Last, D} + end; +update_successors(#b_ret{arg=Arg}=Last, Ts, D0) -> + FuncId = D0#d.func_id, + D = case D0#d.ds of + #{ Arg := #b_set{op=call,args=[FuncId | _]} } -> + %% Returning a call to ourselves doesn't affect our own return + %% type. + D0; + #{} -> + RetType = beam_types:join([raw_type(Arg, Ts) | D0#d.ret_type]), + D0#d{ret_type=[RetType]} + end, + {Last, D}. + +update_switch([{Val, Lbl}=Sw | List], V, Ts, UsedOnce, Acc, D0) -> + case infer_types_switch(V, Val, Ts, UsedOnce, D0) of + none -> + update_switch(List, V, Ts, UsedOnce, Acc, D0); + SwTs -> + D = update_successor(Lbl, SwTs, D0), + update_switch(List, V, Ts, UsedOnce, [Sw | Acc], D) + end; +update_switch([], _V, _Ts, _UsedOnce, Acc, D) -> + {reverse(Acc), D}. -subtract_sw_list(V, List, Ts) -> - Ts#{ V := sub_sw_list_1(get_type(V, Ts), List, Ts) }. +update_switch_failure(V, List, Ts, UsedOnce, D) -> + case sub_sw_list_1(raw_type(V, Ts), List, Ts) of + none -> + none; + FailType -> + case beam_types:get_singleton_value(FailType) of + {ok, Value} -> + %% This is the only possible value at the fail label, so we + %% can infer types as if we matched it directly. + Lit = #b_literal{val=Value}, + infer_types_switch(V, Lit, Ts, UsedOnce, D); + error when UsedOnce -> + ts_remove_var(V, Ts); + error -> + Ts + end + end. sub_sw_list_1(Type, [{Val,_}|T], Ts) -> - ValType = get_type(Val, Ts), - sub_sw_list_1(subtract(Type, ValType), T, Ts); + ValType = raw_type(Val, Ts), + sub_sw_list_1(beam_types: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, +update_successor(?EXCEPTION_BLOCK, _Ts, #d{}=D) -> + %% We KNOW that no variables are used in the ?EXCEPTION_BLOCK, %% so there is no need to update the type information. That %% can be a huge timesaver for huge functions. D; @@ -828,247 +838,102 @@ update_successor(S, Ts0, #d{ls=Ls}=D) -> D#d{ls=Ls#{S=>Ts0}} end. -update_types(#b_set{op=Op,dst=Dst,args=Args}, Ts, Ds) -> - T = type(Op, Args, Ts, Ds), +update_types(#b_set{op=Op,dst=Dst,anno=Anno,args=Args}, Ts, Ds) -> + T = type(Op, Args, Anno, Ts, Ds), Ts#{Dst=>T}. -type(phi, Args, Ts, _Ds) -> - Types = [get_type(A, Ts) || {A,_} <- Args], - join(Types); -type({bif,'band'}, Args, Ts, _Ds) -> - band_type(Args, Ts); -type({bif,Bif}, Args, Ts, _Ds) -> - case bif_type(Bif, Args) of - number -> - arith_op_type(Args, Ts); - Type -> - Type +type(phi, Args, _Anno, Ts, _Ds) -> + Types = [raw_type(A, Ts) || {A,_} <- Args], + beam_types:join(Types); +type({bif,Bif}, Args, _Anno, Ts, _Ds) -> + ArgTypes = normalized_types(Args, Ts), + {RetType, _, _} = beam_call_types:types(erlang, Bif, ArgTypes), + RetType; +type(bs_init, _Args, _Anno, _Ts, _Ds) -> + #t_bitstring{}; +type(bs_extract, [Ctx], _Anno, _Ts, Ds) -> + #b_set{op=bs_match,args=Args} = map_get(Ctx, Ds), + bs_match_type(Args); +type(bs_match, _Args, _Anno, _Ts, _Ds) -> + #t_bs_context{}; +type(bs_get_tail, _Args, _Anno, _Ts, _Ds) -> + #t_bitstring{}; +type(call, [#b_local{} | _Args], Anno, _Ts, _Ds) -> + case Anno of + #{ result_type := Type } -> Type; + #{} -> any end; -type(bs_init, _Args, _Ts, _Ds) -> - {binary, 1}; -type(bs_extract, [Ctx], Ts, _Ds) -> - #t_bs_match{type=Type} = get_type(Ctx, Ts), - 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 - {erlang,setelement,[Pos,Tuple,Arg]} -> - case {get_type(Pos, Ts),get_type(Tuple, Ts)} of - {#t_integer{elements={Index,Index}}, - #t_tuple{elements=Es0,size=Size}=T} -> - %% This is an exact index, update the type of said element - %% or return 'none' if it's known to be out of bounds. - Es = set_element_type(Index, get_type(Arg, Ts), Es0), - case T#t_tuple.exact of - false -> - T#t_tuple{size=max(Index, Size),elements=Es}; - true when Index =< Size -> - T#t_tuple{elements=Es}; - true -> - none - end; - {#t_integer{elements={Min,_}}=IntType, - #t_tuple{elements=Es0,size=Size}=T} -> - %% Remove type information for all indices that - %% falls into the range of the integer. - Es = remove_element_info(IntType, Es0), - case T#t_tuple.exact of - false -> - T#t_tuple{elements=Es,size=max(Min, Size)}; - true when Min =< Size -> - T#t_tuple{elements=Es,size=Size}; - true -> - none - end; - {_,#t_tuple{}=T} -> - %% Position unknown, so we have to discard all element - %% information. - T#t_tuple{elements=#{}}; - {#t_integer{elements={Min,_Max}},_} -> - #t_tuple{size=Min}; - {_,_} -> - #t_tuple{} - end; - {erlang,'++',[LHS,RHS]} -> - LType = get_type(LHS, Ts), - RType = get_type(RHS, Ts), - case LType =:= cons orelse RType =:= cons of - true -> - cons; - false -> - %% `[] ++ RHS` yields RHS, even if RHS is not a list. - join(list, RType) - end; - {erlang,'--',[_,_]} -> - list; - {lists,F,Args} -> - Types = get_types(Args, Ts), - lists_function_type(F, Types); - {math,_,_} -> - case is_math_bif(Name, length(Args)) of - false -> any; - true -> float - end; - {_,_,_} -> - case erl_bifs:is_exit_bif(Mod, Name, length(Args)) of - true -> none; - false -> any - end - end; -type(get_tuple_element, [Tuple, Offset], Ts, _Ds) -> - #t_tuple{size=Size,elements=Es} = get_type(Tuple, Ts), + name=#b_literal{val=Name}}|Args], _Anno, Ts, _Ds) -> + ArgTypes = normalized_types(Args, Ts), + {RetType, _, _} = beam_call_types:types(Mod, Name, ArgTypes), + RetType; +type(get_tuple_element, [Tuple, Offset], _Anno, Ts, _Ds) -> + #t_tuple{size=Size,elements=Es} = normalized_type(Tuple, Ts), #b_literal{val=N} = Offset, true = Size > N, %Assertion. - get_element_type(N + 1, Es); -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) -> + beam_types:get_element_type(N + 1, Es); +type(is_nonempty_list, [_], _Anno, _Ts, _Ds) -> + beam_types:make_boolean(); +type(is_tagged_tuple, [_,#b_literal{},#b_literal{}], _Anno, _Ts, _Ds) -> + beam_types:make_boolean(); +type(make_fun, [#b_local{arity=TotalArity}|Env], _Anno, _Ts, _Ds) -> + #t_fun{arity=TotalArity-length(Env)}; +type(put_map, _Args, _Anno, _Ts, _Ds) -> + #t_map{}; +type(put_list, _Args, _Anno, _Ts, _Ds) -> cons; -type(put_tuple, Args, Ts, _Ds) -> +type(put_tuple, Args, _Anno, Ts, _Ds) -> {Es, _} = foldl(fun(Arg, {Es0, Index}) -> - Type = get_type(Arg, Ts), - Es = set_element_type(Index, Type, Es0), - {Es, Index + 1} + Type = raw_type(Arg, Ts), + Es = beam_types:set_element_type(Index, Type, Es0), + {Es, Index + 1} end, {#{}, 1}, Args), #t_tuple{exact=true,size=length(Args),elements=Es}; -type(succeeded, [#b_var{}=Src], Ts, Ds) -> +type(succeeded, [#b_var{}=Src], _Anno, Ts, _Ds) + when map_get(Src, Ts) =:= none -> + beam_types:make_atom(false); +type(succeeded, [#b_var{}=Src], _Anno, 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,_}]} -> - t_atom(true); - {map_size,[map]} -> - t_atom(true); - {'not',[Type]} -> - case t_is_boolean(Type) of - true -> t_atom(true); - false -> t_boolean() - end; - {size,[{binary,_}]} -> - t_atom(true); - {tuple_size,[#t_tuple{}]} -> - t_atom(true); - {_,_} -> - t_boolean() + ArgTypes = normalized_types(BifArgs, Ts), + case beam_call_types:will_succeed(erlang, Bif, ArgTypes) of + yes -> beam_types:make_atom(true); + no -> beam_types:make_atom(false); + maybe -> beam_types:make_boolean() + end; + #b_set{op=call,args=[#b_remote{mod=#b_literal{val=Mod}, + name=#b_literal{val=Func}} | + CallArgs]} -> + ArgTypes = normalized_types(CallArgs, Ts), + case beam_call_types:will_succeed(Mod, Func, ArgTypes) of + yes -> beam_types:make_atom(true); + no -> beam_types:make_atom(false); + maybe -> beam_types:make_boolean() end; #b_set{op=get_hd} -> - t_atom(true); + beam_types:make_atom(true); #b_set{op=get_tl} -> - t_atom(true); + beam_types:make_atom(true); #b_set{op=get_tuple_element} -> - t_atom(true); + beam_types:make_atom(true); + #b_set{op=put_tuple} -> + beam_types:make_atom(true); #b_set{op=wait} -> - t_atom(false); + beam_types:make_atom(false); #b_set{} -> - t_boolean() + beam_types:make_boolean() end; -type(succeeded, [#b_literal{}], _Ts, _Ds) -> - t_atom(true); -type(_, _, _, _) -> any. - -arith_op_type(Args, Ts) -> - Types = get_types(Args, Ts), - foldl(fun(#t_integer{}, unknown) -> t_integer(); - (#t_integer{}, number) -> number; - (#t_integer{}, float) -> float; - (#t_integer{}, #t_integer{}) -> t_integer(); - (float, unknown) -> float; - (float, #t_integer{}) -> float; - (float, number) -> float; - (number, unknown) -> number; - (number, #t_integer{}) -> number; - (number, float) -> float; - (any, _) -> number; - (Same, Same) -> Same; - (_, _) -> none - end, unknown, Types). - -lists_function_type(F, Types) -> - case {F,Types} of - %% Functions that return booleans. - {all,[_,_]} -> - t_boolean(); - {any,[_,_]} -> - t_boolean(); - {keymember,[_,_,_]} -> - t_boolean(); - {member,[_,_]} -> - t_boolean(); - {prefix,[_,_]} -> - t_boolean(); - {suffix,[_,_]} -> - t_boolean(); - - %% Functions that return lists. - {dropwhile,[_,_]} -> - list; - {duplicate,[_,_]} -> - list; - {filter,[_,_]} -> - list; - {flatten,[_]} -> - list; - {map,[_Fun,List]} -> - same_length_type(List); - {MapFold,[_Fun,_Acc,List]} when MapFold =:= mapfoldl; - MapFold =:= mapfoldr -> - #t_tuple{size=2,exact=true, - elements=#{1=>same_length_type(List)}}; - {partition,[_,_]} -> - t_two_tuple(list, list); - {reverse,[List]} -> - same_length_type(List); - {sort,[List]} -> - same_length_type(List); - {splitwith,[_,_]} -> - t_two_tuple(list, list); - {takewhile,[_,_]} -> - list; - {unzip,[List]} -> - ListType = same_length_type(List), - t_two_tuple(ListType, ListType); - {usort,[List]} -> - same_length_type(List); - {zip,[_,_]} -> - list; - {zipwith,[_,_,_]} -> - list; - {_,_} -> - any - end. - -%% For a lists function that return a list of the same -%% length as the input list, return the type of the list. -same_length_type(cons) -> cons; -same_length_type(nil) -> nil; -same_length_type(_) -> list. - -t_two_tuple(Type1, Type2) -> - #t_tuple{size=2,exact=true, - elements=#{1=>Type1,2=>Type2}}. +type(succeeded, [#b_literal{}], _Anno, _Ts, _Ds) -> + beam_types:make_atom(true); +type(_, _, _, _, _) -> any. %% will_succeed(TestOperation, Type) -> yes|no|maybe. %% Test whether TestOperation applied to an argument of type Type %% will succeed. Return yes, no, or maybe. %% -%% Type is a type as described in the comment for verified_type/1 at -%% the very end of this file, but it will *never* be 'any'. +%% Type can be any type as described in beam_types.hrl, but it must *never* be +%% any. will_succeed(is_atom, Type) -> case Type of @@ -1077,13 +942,13 @@ will_succeed(is_atom, Type) -> end; will_succeed(is_binary, Type) -> case Type of - {binary,U} when U rem 8 =:= 0 -> yes; - {binary,_} -> maybe; + #t_bitstring{unit=U} when U rem 8 =:= 0 -> yes; + #t_bitstring{} -> maybe; _ -> no end; will_succeed(is_bitstring, Type) -> case Type of - {binary,_} -> yes; + #t_bitstring{} -> yes; _ -> no end; will_succeed(is_boolean, Type) -> @@ -1091,7 +956,7 @@ will_succeed(is_boolean, Type) -> #t_atom{elements=any} -> maybe; #t_atom{elements=Es} -> - case t_is_boolean(Type) of + case beam_types:is_boolean_type(Type) of true -> yes; false -> @@ -1109,6 +974,11 @@ will_succeed(is_float, Type) -> number -> maybe; _ -> no end; +will_succeed(is_function, Type) -> + case Type of + #t_fun{} -> yes; + _ -> no + end; will_succeed(is_integer, Type) -> case Type of #t_integer{} -> yes; @@ -1123,7 +993,7 @@ will_succeed(is_list, Type) -> end; will_succeed(is_map, Type) -> case Type of - map -> yes; + #t_map{} -> yes; _ -> no end; will_succeed(is_number, Type) -> @@ -1140,35 +1010,12 @@ will_succeed(is_tuple, Type) -> end; will_succeed(_, _) -> maybe. - -band_type([Other,#b_literal{val=Int}], Ts) when is_integer(Int) -> - band_type_1(Int, Other, Ts); -band_type([_,_], _) -> t_integer(). - -band_type_1(Int, OtherSrc, Ts) -> - Type = band_type_2(Int, 0), - OtherType = get_type(OtherSrc, Ts), - meet(Type, OtherType). - -band_type_2(N, Bits) when Bits < 64 -> - case 1 bsl Bits of - P when P =:= N + 1 -> - t_integer(0, N); - P when P > N + 1 -> - t_integer(); - _ -> - band_type_2(N, Bits+1) - end; -band_type_2(_, _) -> - %% Negative or large positive number. Give up. - t_integer(). - bs_match_type([#b_literal{val=Type}|Args]) -> bs_match_type(Type, Args). bs_match_type(binary, Args) -> [_,_,_,#b_literal{val=U}] = Args, - {binary,U}; + #t_bitstring{unit=U}; bs_match_type(float, _) -> float; bs_match_type(integer, Args) -> @@ -1180,24 +1027,24 @@ bs_match_type(integer, Args) -> NumBits = Size * Unit, case member(unsigned, Flags) of true -> - t_integer(0, (1 bsl NumBits)-1); + beam_types:make_integer(0, (1 bsl NumBits)-1); false -> %% Signed integer. Don't bother. - t_integer() + #t_integer{} end; [_|_] -> - t_integer() + #t_integer{} end; bs_match_type(skip, _) -> any; bs_match_type(string, _) -> any; bs_match_type(utf8, _) -> - ?UNICODE_INT; + beam_types:make_integer(0, ?UNICODE_MAX); bs_match_type(utf16, _) -> - ?UNICODE_INT; + beam_types:make_integer(0, ?UNICODE_MAX); bs_match_type(utf32, _) -> - ?UNICODE_INT. + beam_types:make_integer(0, ?UNICODE_MAX). simplify_switch_atom(#t_atom{elements=Atoms}, #b_switch{list=List0}=Sw) -> case sort([A || {#b_literal{val=A},_} <- List0]) of @@ -1229,14 +1076,14 @@ eq_ranges(_, _, _) -> false. simplify_is_record(I, #t_tuple{exact=Exact, size=Size, elements=Es}, - RecSize, RecTag, Ts) -> + RecSize, #b_literal{val=TagVal}=RecTag, Ts) -> TagType = maps:get(1, Es, any), - TagMatch = case get_literal_from_type(TagType) of - #b_literal{}=RecTag -> yes; - #b_literal{} -> no; - none -> + TagMatch = case beam_types:get_singleton_value(TagType) of + {ok, TagVal} -> yes; + {ok, _} -> no; + error -> %% Is it at all possible for the tag to match? - case meet(get_type(RecTag, Ts), TagType) of + case beam_types:meet(raw_type(RecTag, Ts), TagType) of none -> no; _ -> maybe end @@ -1266,7 +1113,7 @@ simplify_switch_bool(#b_switch{arg=B,fail=Fail,list=List0}, 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 + case beam_types:is_boolean_type(raw_type(Bool, Ts)) of true -> Br = Br0#b_br{bool=Bool,succ=Fail,fail=Succ}, beam_ssa:normalize(Br); @@ -1334,40 +1181,18 @@ used_once_last_uses([V|Vs], L, Uses) -> end; used_once_last_uses([], _, Uses) -> Uses. +normalized_types(Values, Ts) -> + [normalized_type(Val, Ts) || Val <- Values]. -get_types(Values, Ts) -> - [get_type(Val, Ts) || Val <- Values]. --spec get_type(beam_ssa:value(), type_db()) -> type(). +normalized_type(V, Ts) -> + beam_types:normalize(raw_type(V, Ts)). -get_type(#b_var{}=V, Ts) -> - #{V:=T} = Ts, - T; -get_type(#b_literal{val=Val}, _Ts) -> - if - is_atom(Val) -> - t_atom(Val); - is_float(Val) -> - float; - is_integer(Val) -> - t_integer(Val); - is_list(Val), Val =/= [] -> - cons; - is_map(Val) -> - map; - Val =:= {} -> - #t_tuple{exact=true}; - is_tuple(Val) -> - {Es, _} = foldl(fun(E, {Es0, Index}) -> - Type = get_type(#b_literal{val=E}, #{}), - Es = set_element_type(Index, Type, Es0), - {Es, Index + 1} - end, {#{}, 1}, tuple_to_list(Val)), - #t_tuple{exact=true,size=tuple_size(Val),elements=Es}; - Val =:= [] -> - nil; - true -> - any - end. +-spec raw_type(beam_ssa:value(), type_db()) -> type(). + +raw_type(#b_literal{val=Value}, _Ts) -> + beam_types:make_type_from_value(Value); +raw_type(V, Ts) -> + map_get(V, Ts). %% infer_types(Var, Types, #d{}) -> {SuccTypes,FailTypes} %% Looking at the expression that defines the variable Var, infer @@ -1390,10 +1215,107 @@ get_type(#b_literal{val=Val}, _Ts) -> %% 'cons' would give 'nil' as the only possible type. The result of the %% subtraction for L will be added to FailTypes. -infer_types_br(#b_var{}=V, Ts, #d{ds=Ds}) -> +infer_types_br(#b_var{}=V, Ts, UsedOnce, #d{ds=Ds}) -> #{V:=#b_set{op=Op,args=Args}} = Ds, - PosTypes0 = infer_type(Op, Args, Ds), - NegTypes0 = infer_type_negative(Op, Args, Ds), + + {PosTypes, NegTypes} = infer_type(Op, Args, Ts, Ds), + + SuccTs0 = meet_types(PosTypes, Ts), + FailTs0 = subtract_types(NegTypes, Ts), + + case UsedOnce of + true -> + %% The branch variable is defined in this block and is only + %% referenced by this terminator. Therefore, there is no need to + %% include it in the type database passed on to the successors of + %% of this block. + SuccTs = ts_remove_var(V, SuccTs0), + FailTs = ts_remove_var(V, FailTs0), + {SuccTs, FailTs}; + false -> + SuccTs = infer_br_value(V, true, SuccTs0), + FailTs = infer_br_value(V, false, FailTs0), + {SuccTs, FailTs} + end. + +infer_br_value(_V, _Bool, none) -> + none; +infer_br_value(V, Bool, NewTs) -> + #{ V := T } = NewTs, + case beam_types:is_boolean_type(T) of + true -> + NewTs#{ V := beam_types:make_atom(Bool) }; + false -> + %% V is a try/catch tag or similar, leave it alone. + NewTs + end. + +infer_types_switch(V, Lit, Ts0, UsedOnce, #d{ds=Ds}) -> + {PosTypes, _} = infer_type({bif,'=:='}, [V, Lit], Ts0, Ds), + Ts = meet_types(PosTypes, Ts0), + case UsedOnce of + true -> ts_remove_var(V, Ts); + false -> Ts + end. + +ts_remove_var(_V, none) -> none; +ts_remove_var(V, Ts) -> maps:remove(V, Ts). + +infer_type(succeeded, [#b_var{}=Src], Ts, Ds) -> + #b_set{op=Op,args=Args} = maps:get(Src, Ds), + infer_success_type(Op, Args, Ts, Ds); + +%% Type tests are handled separately from other BIFs as we're inferring types +%% based on their result, so we know that subtraction is safe even if we're +%% not branching on 'succeeded'. +infer_type(is_tagged_tuple, [#b_var{}=Src,#b_literal{val=Size}, + #b_literal{}=Tag], _Ts, _Ds) -> + Es = beam_types:set_element_type(1, raw_type(Tag, #{}), #{}), + T = {Src,#t_tuple{exact=true,size=Size,elements=Es}}, + {[T], [T]}; +infer_type(is_nonempty_list, [#b_var{}=Src], _Ts, _Ds) -> + T = {Src,cons}, + {[T], [T]}; +infer_type({bif,is_atom}, [Arg], _Ts, _Ds) -> + T = {Arg, #t_atom{}}, + {[T], [T]}; +infer_type({bif,is_binary}, [Arg], _Ts, _Ds) -> + T = {Arg, #t_bitstring{unit=8}}, + {[T], [T]}; +infer_type({bif,is_bitstring}, [Arg], _Ts, _Ds) -> + T = {Arg, #t_bitstring{}}, + {[T], [T]}; +infer_type({bif,is_boolean}, [Arg], _Ts, _Ds) -> + T = {Arg, beam_types:make_boolean()}, + {[T], [T]}; +infer_type({bif,is_float}, [Arg], _Ts, _Ds) -> + T = {Arg, float}, + {[T], [T]}; +infer_type({bif,is_integer}, [Arg], _Ts, _Ds) -> + T = {Arg, #t_integer{}}, + {[T], [T]}; +infer_type({bif,is_list}, [Arg], _Ts, _Ds) -> + T = {Arg, list}, + {[T], [T]}; +infer_type({bif,is_map}, [Arg], _Ts, _Ds) -> + T = {Arg, #t_map{}}, + {[T], [T]}; +infer_type({bif,is_number}, [Arg], _Ts, _Ds) -> + T = {Arg, number}, + {[T], [T]}; +infer_type({bif,is_tuple}, [Arg], _Ts, _Ds) -> + T = {Arg, #t_tuple{}}, + {[T], [T]}; +infer_type({bif,'=:='}, [#b_var{}=LHS,#b_var{}=RHS], 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'). + LType = raw_type(LHS, Ts), + RType = raw_type(RHS, Ts), + Type = beam_types:meet(LType, RType), + + PosTypes = [{V,Type} || {V, OrigType} <- [{LHS, LType}, {RHS, RType}], + OrigType =/= Type], %% We must be careful with types inferred from '=:='. %% @@ -1404,39 +1326,39 @@ infer_types_br(#b_var{}=V, Ts, #d{ds=Ds}) -> %% %% However, it is safe to subtract a type inferred from '=:=' if %% it is single-valued, e.g. if it is [] or the atom 'true'. + NegTypes = case beam_types:is_singleton_type(Type) of + true -> PosTypes; + false -> [] + end, - EqTypes = infer_eq_type(Op, Args, Ts, Ds), - NegTypes1 = [P || {_,T}=P <- EqTypes, is_singleton_type(T)], - - PosTypes = EqTypes ++ PosTypes0, - SuccTs = meet_types(PosTypes, Ts), - - NegTypes = NegTypes0 ++ NegTypes1, - FailTs = subtract_types(NegTypes, Ts), - - {SuccTs,FailTs}. - -infer_types_switch(V, Lit, Ts, #d{ds=Ds}) -> - Types = infer_eq_type({bif,'=:='}, [V, Lit], Ts, Ds), - meet_types(Types, Ts). - -infer_eq_type({bif,'=:='}, [#b_var{}=Src,#b_literal{}=Lit], Ts, Ds) -> + {PosTypes, NegTypes}; +infer_type({bif,'=:='}, [#b_var{}=Src,#b_literal{}=Lit], Ts, Ds) -> Def = maps:get(Src, Ds), - Type = get_type(Lit, Ts), - [{Src,Type} | infer_eq_lit(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) -> - []. + Type = raw_type(Lit, Ts), + EqLitTypes = infer_eq_lit(Def, Lit), + PosTypes = [{Src,Type} | EqLitTypes], + {PosTypes, EqLitTypes}; +infer_type(_Op, _Args, _Ts, _Ds) -> + {[], []}. + +infer_success_type({bif,Op}, Args, Ts, _Ds) -> + ArgTypes = normalized_types(Args, Ts), + + {_, PosTypes0, CanSubtract} = beam_call_types:types(erlang, Op, ArgTypes), + PosTypes = [T || {#b_var{},_}=T <- zip(Args, PosTypes0)], + + case CanSubtract of + true -> {PosTypes, PosTypes}; + false -> {PosTypes, []} + end; +infer_success_type(call, [#b_var{}=Fun|Args], _Ts, _Ds) -> + T = {Fun, #t_fun{arity=length(Args)}}, + {[T], []}; +infer_success_type(bs_start_match, [#b_var{}=Bin], _Ts, _Ds) -> + T = {Bin,#t_bitstring{}}, + {[T], [T]}; +infer_success_type(_Op, _Args, _Ts, _Ds) -> + {[], []}. infer_eq_lit(#b_set{op={bif,tuple_size},args=[#b_var{}=Tuple]}, #b_literal{val=Size}) when is_integer(Size) -> @@ -1445,178 +1367,11 @@ infer_eq_lit(#b_set{op=get_tuple_element, args=[#b_var{}=Tuple,#b_literal{val=N}]}, #b_literal{}=Lit) -> Index = N + 1, - Es = set_element_type(Index, get_type(Lit, #{}), #{}), + Es = beam_types:set_element_type(Index, raw_type(Lit, #{}), #{}), [{Tuple,#t_tuple{size=Index,elements=Es}}]; -infer_eq_lit(_, _) -> []. - -infer_type_negative(Op, Args, Ds) -> - case is_negative_inference_safe(Op, Args) of - true -> - infer_type(Op, Args, Ds); - false -> - [] - end. - -%% Conservative list of instructions for which negative -%% inference is safe. -is_negative_inference_safe(is_nonempty_list, _Args) -> true; -is_negative_inference_safe(_, _) -> false. - -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,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,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{}=Src], _Ds) -> - [{Src,map}]; -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{}=Src], _Ds) -> - [{Src,cons}]; -infer_type(is_tagged_tuple, [#b_var{}=Src,#b_literal{val=Size}, - #b_literal{}=Tag], _Ds) -> - Es = set_element_type(1, get_type(Tag, #{}), #{}), - [{Src,#t_tuple{exact=true,size=Size,elements=Es}}]; -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) -> +infer_eq_lit(_, _) -> []. -%% bif_type(Name, Args) -> Type -%% Return the return type for the guard BIF or operator Name with -%% arguments Args. -%% -%% Note that that the following BIFs are handle elsewhere: -%% -%% band/2 - -bif_type(abs, [_]) -> number; -bif_type(bit_size, [_]) -> t_integer(); -bif_type(byte_size, [_]) -> t_integer(); -bif_type(ceil, [_]) -> t_integer(); -bif_type(float, [_]) -> float; -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(); -bif_type('bsr', [_,_]) -> t_integer(); -bif_type('bxor', [_,_]) -> t_integer(); -bif_type('div', [_,_]) -> t_integer(); -bif_type('rem', [_,_]) -> t_integer(); -bif_type('/', [_,_]) -> float; -bif_type(Name, Args) -> - Arity = length(Args), - case erl_internal:new_type_test(Name, Arity) orelse - erl_internal:bool_op(Name, Arity) orelse - erl_internal:comp_op(Name, Arity) of - true -> - t_boolean(); - false -> - case erl_internal:arith_op(Name, Arity) of - true -> number; - false -> any - end - end. - -inferred_bif_type(is_atom, [_]) -> t_atom(); -inferred_bif_type(is_binary, [_]) -> {binary,8}; -inferred_bif_type(is_bitstring, [_]) -> {binary,1}; -inferred_bif_type(is_boolean, [_]) -> t_boolean(); -inferred_bif_type(is_float, [_]) -> float; -inferred_bif_type(is_integer, [_]) -> t_integer(); -inferred_bif_type(is_list, [_]) -> list; -inferred_bif_type(is_map, [_]) -> map; -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. - -is_math_bif(cos, 1) -> true; -is_math_bif(cosh, 1) -> true; -is_math_bif(sin, 1) -> true; -is_math_bif(sinh, 1) -> true; -is_math_bif(tan, 1) -> true; -is_math_bif(tanh, 1) -> true; -is_math_bif(acos, 1) -> true; -is_math_bif(acosh, 1) -> true; -is_math_bif(asin, 1) -> true; -is_math_bif(asinh, 1) -> true; -is_math_bif(atan, 1) -> true; -is_math_bif(atanh, 1) -> true; -is_math_bif(erf, 1) -> true; -is_math_bif(erfc, 1) -> true; -is_math_bif(exp, 1) -> true; -is_math_bif(log, 1) -> true; -is_math_bif(log2, 1) -> true; -is_math_bif(log10, 1) -> true; -is_math_bif(sqrt, 1) -> true; -is_math_bif(atan2, 2) -> true; -is_math_bif(pow, 2) -> true; -is_math_bif(ceil, 1) -> true; -is_math_bif(floor, 1) -> true; -is_math_bif(fmod, 2) -> true; -is_math_bif(pi, 0) -> true; -is_math_bif(_, _) -> false. - join_types(Ts0, Ts1) -> if map_size(Ts0) < map_size(Ts1) -> @@ -1630,7 +1385,7 @@ join_types_1([V|Vs], Ts0, Ts1) -> {#{V:=Same},#{V:=Same}} -> join_types_1(Vs, Ts0, Ts1); {#{V:=T0},#{V:=T1}} -> - case join(T0, T1) of + case beam_types:join(T0, T1) of T1 -> join_types_1(Vs, Ts0, Ts1); T -> @@ -1642,326 +1397,21 @@ join_types_1([V|Vs], Ts0, Ts1) -> join_types_1([], Ts0, Ts1) -> maps:merge(Ts0, Ts1). -join([T1,T2|Ts]) -> - join([join(T1, T2)|Ts]); -join([T]) -> T. - -get_literal_from_type(#t_atom{elements=[Atom]}) -> - #b_literal{val=Atom}; -get_literal_from_type(#t_integer{elements={Int,Int}}) -> - #b_literal{val=Int}; -get_literal_from_type(nil) -> - #b_literal{val=[]}; -get_literal_from_type(_) -> none. - -remove_element_info(#t_integer{elements={Min,Max}}, Es) -> - foldl(fun(El, Acc) when Min =< El, El =< Max -> - maps:remove(El, Acc); - (_El, Acc) -> Acc - end, Es, maps:keys(Es)). - -t_atom() -> - #t_atom{elements=any}. - -t_atom(Atom) when is_atom(Atom) -> - #t_atom{elements=[Atom]}. - -t_boolean() -> - #t_atom{elements=[false,true]}. - -t_integer() -> - #t_integer{elements=any}. - -t_integer(Int) when is_integer(Int) -> - #t_integer{elements={Int,Int}}. - -t_integer(Min, Max) when is_integer(Min), is_integer(Max) -> - #t_integer{elements={Min,Max}}. - -t_is_boolean(#t_atom{elements=[F,T]}) -> - F =:= false andalso T =:= true; -t_is_boolean(#t_atom{elements=[B]}) -> - is_boolean(B); -t_is_boolean(_) -> false. - -t_tuple_size(#t_tuple{size=Size,exact=false}) -> - {at_least,Size}; -t_tuple_size(#t_tuple{size=Size,exact=true}) -> - {exact,Size}; -t_tuple_size(_) -> - none. - -is_singleton_type(Type) -> - get_literal_from_type(Type) =/= none. - -get_element_type(Index, Es) -> - case Es of - #{ Index := T } -> T; - #{} -> any - end. - -set_element_type(_Key, none, Es) -> - Es; -set_element_type(Key, any, Es) -> - maps:remove(Key, Es); -set_element_type(Key, Type, Es) -> - Es#{ Key => Type }. - -%% join(Type1, Type2) -> Type -%% Return the "join" of Type1 and Type2. The join is a more general -%% type than Type1 and Type2. For example: -%% -%% join(#t_integer{elements=any}, #t_integer=elements={0,3}}) -> -%% #t_integer{} -%% -%% The join for two different types result in 'any', which is -%% the top element for our type lattice: -%% -%% join(#t_integer{}, map) -> any - --spec join(type(), type()) -> type(). - -join(T, T) -> - verified_type(T); -join(none, T) -> - verified_type(T); -join(T, none) -> - verified_type(T); -join(any, _) -> any; -join(_, any) -> any; -join(#t_atom{elements=[_|_]=Set1}, #t_atom{elements=[_|_]=Set2}) -> - Set = ordsets:union(Set1, Set2), - case ordsets:size(Set) of - Size when Size =< ?ATOM_SET_SIZE -> - #t_atom{elements=Set}; - _Size -> - #t_atom{elements=any} - end; -join(#t_atom{elements=any}=T, #t_atom{elements=[_|_]}) -> T; -join(#t_atom{elements=[_|_]}, #t_atom{elements=any}=T) -> T; -join({binary,U1}, {binary,U2}) -> - {binary,gcd(U1, U2)}; -join(#t_integer{}, #t_integer{}) -> t_integer(); -join(list, cons) -> list; -join(cons, list) -> list; -join(nil, cons) -> list; -join(cons, nil) -> list; -join(nil, list) -> list; -join(list, nil) -> list; -join(#t_integer{}, float) -> number; -join(float, #t_integer{}) -> number; -join(#t_integer{}, number) -> number; -join(number, #t_integer{}) -> number; -join(float, number) -> number; -join(number, float) -> number; -join(#t_tuple{size=Sz,exact=ExactA,elements=EsA}, - #t_tuple{size=Sz,exact=ExactB,elements=EsB}) -> - Exact = ExactA and ExactB, - Es = join_tuple_elements(Sz, EsA, EsB), - #t_tuple{size=Sz,exact=Exact,elements=Es}; -join(#t_tuple{size=SzA,elements=EsA}, #t_tuple{size=SzB,elements=EsB}) -> - Sz = min(SzA, SzB), - Es = join_tuple_elements(Sz, EsA, EsB), - #t_tuple{size=Sz,elements=Es}; -join(_T1, _T2) -> - %%io:format("~p ~p\n", [_T1,_T2]), - any. - -join_tuple_elements(MinSize, EsA, EsB) -> - Es0 = join_elements(EsA, EsB), - maps:filter(fun(Index, _Type) -> Index =< MinSize end, Es0). - -join_elements(Es1, Es2) -> - Keys = if - map_size(Es1) =< map_size(Es2) -> maps:keys(Es1); - map_size(Es1) > map_size(Es2) -> maps:keys(Es2) - end, - join_elements_1(Keys, Es1, Es2, #{}). - -join_elements_1([Key | Keys], Es1, Es2, Acc0) -> - case {Es1, Es2} of - {#{ Key := Type1 }, #{ Key := Type2 }} -> - Acc = set_element_type(Key, join(Type1, Type2), Acc0), - join_elements_1(Keys, Es1, Es2, Acc); - {#{}, #{}} -> - join_elements_1(Keys, Es1, Es2, Acc0) - end; -join_elements_1([], _Es1, _Es2, Acc) -> - Acc. - -gcd(A, B) -> - case A rem B of - 0 -> B; - X -> gcd(B, X) - end. - meet_types([{V,T0}|Vs], Ts) -> #{V:=T1} = Ts, - case meet(T0, T1) of + case beam_types:meet(T0, T1) of + none -> none; 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 + case beam_types:subtract(T1, T0) of + none -> none; 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: -%% -%% meet(#t_integer{elements=any}, #t_integer{elements={0,3}}) -> -%% #t_integer{elements={0,3}} -%% -%% The meet for two different types result in 'none', which is -%% the bottom element for our type lattice: -%% -%% meet(#t_integer{}, map) -> none - --spec meet(type(), type()) -> type(). - -meet(T, T) -> - verified_type(T); -meet(#t_atom{elements=[_|_]=Set1}, #t_atom{elements=[_|_]=Set2}) -> - case ordsets:intersection(Set1, Set2) of - [] -> - none; - [_|_]=Set -> - #t_atom{elements=Set} - end; -meet(#t_atom{elements=[_|_]}=T, #t_atom{elements=any}) -> - T; -meet(#t_atom{elements=any}, #t_atom{elements=[_|_]}=T) -> - T; -meet(#t_integer{elements={_,_}}=T, #t_integer{elements=any}) -> - T; -meet(#t_integer{elements=any}, #t_integer{elements={_,_}}=T) -> - T; -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=T, number) -> T; -meet(number, #t_integer{}=T) -> T; -meet(number, float=T) -> T; -meet(list, cons) -> cons; -meet(list, nil) -> nil; -meet(cons, list) -> cons; -meet(nil, list) -> nil; -meet(#t_tuple{}=T1, #t_tuple{}=T2) -> - meet_tuples(T1, T2); -meet({binary,U1}, {binary,U2}) -> - {binary,max(U1, U2)}; -meet(any, T) -> - verified_type(T); -meet(T, any) -> - verified_type(T); -meet(_, _) -> - %% Inconsistent types. There will be an exception at runtime. - none. - -meet_tuples(#t_tuple{size=Sz1,exact=true}, - #t_tuple{size=Sz2,exact=true}) when Sz1 =/= Sz2 -> - none; -meet_tuples(#t_tuple{size=Sz1,exact=Ex1,elements=Es1}, - #t_tuple{size=Sz2,exact=Ex2,elements=Es2}) -> - Size = max(Sz1, Sz2), - Exact = Ex1 or Ex2, - case meet_elements(Es1, Es2) of - none -> - none; - Es -> - #t_tuple{size=Size,exact=Exact,elements=Es} - end. - -meet_elements(Es1, Es2) -> - Keys = maps:keys(Es1) ++ maps:keys(Es2), - meet_elements_1(Keys, Es1, Es2, #{}). - -meet_elements_1([Key | Keys], Es1, Es2, Acc) -> - case {Es1, Es2} of - {#{ Key := Type1 }, #{ Key := Type2 }} -> - case meet(Type1, Type2) of - none -> none; - Type -> meet_elements_1(Keys, Es1, Es2, Acc#{ Key => Type }) - end; - {#{ Key := Type1 }, _} -> - meet_elements_1(Keys, Es1, Es2, Acc#{ Key => Type1 }); - {_, #{ Key := Type2 }} -> - meet_elements_1(Keys, Es1, Es2, Acc#{ Key => Type2 }) - end; -meet_elements_1([], _Es1, _Es2, Acc) -> - Acc. - -%% verified_type(Type) -> Type -%% Returns the passed in type if it is one of the defined types. -%% Crashes if there is anything wrong with the type. -%% -%% Here are all possible types: -%% -%% any Any Erlang term (top element for the type lattice). -%% -%% #t_atom{} Any atom or some specific atoms. -%% {binary,Unit} Binary/bitstring aligned to unit Unit. -%% float Floating point number. -%% #t_integer{} Integer -%% list Empty or nonempty list. -%% map Map. -%% nil Empty list. -%% cons Cons (nonempty list). -%% number A number (float or integer). -%% #t_tuple{} Tuple. -%% -%% none No type (bottom element for the type lattice). - --spec verified_type(T) -> T when - T :: type(). - -verified_type(any=T) -> T; -verified_type(none=T) -> T; -verified_type(#t_atom{elements=any}=T) -> T; -verified_type(#t_atom{elements=[_|_]}=T) -> T; -verified_type({binary,U}=T) when is_integer(U) -> T; -verified_type(#t_integer{elements=any}=T) -> T; -verified_type(#t_integer{elements={Min,Max}}=T) - when is_integer(Min), is_integer(Max) -> T; -verified_type(list=T) -> T; -verified_type(map=T) -> T; -verified_type(nil=T) -> T; -verified_type(cons=T) -> T; -verified_type(number=T) -> T; -verified_type(#t_tuple{size=Size,elements=Es}=T) -> - %% All known elements must have a valid index and type. 'any' is prohibited - %% since it's implicit and should never be present in the map. - maps:fold(fun(Index, Element, _) when is_integer(Index), - 1 =< Index, Index =< Size, - Element =/= any, Element =/= none -> - verified_type(Element) - end, [], Es), - T; -verified_type(float=T) -> T. diff --git a/lib/compiler/src/beam_trim.erl b/lib/compiler/src/beam_trim.erl index acf3838da4..ad8839cc7d 100644 --- a/lib/compiler/src/beam_trim.erl +++ b/lib/compiler/src/beam_trim.erl @@ -244,6 +244,9 @@ remap([{make_fun2,_,_,_,_}=I|T], Map, Acc) -> remap([{deallocate,N}|Is], Map, Acc) -> I = {deallocate,Map({frame_size,N})}, remap(Is, Map, [I|Acc]); +remap([{swap,Reg1,Reg2}|Is], Map, Acc) -> + I = {swap,Map(Reg1),Map(Reg2)}, + remap(Is, Map, [I|Acc]); remap([{test,Name,Fail,Ss}|Is], Map, Acc) -> I = {test,Name,Fail,[Map(S) || S <- Ss]}, remap(Is, Map, [I|Acc]); @@ -382,6 +385,8 @@ frame_size([{bs_set_position,_,_}|Is], Safe) -> frame_size(Is, Safe); frame_size([{bs_get_tail,_,_,_}|Is], Safe) -> frame_size(Is, Safe); +frame_size([{swap,_,_}|Is], Safe) -> + frame_size(Is, Safe); frame_size(_, _) -> throw(not_possible). frame_size_branch(0, Is, Safe) -> @@ -444,6 +449,8 @@ 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, [{swap,Reg1,Reg2}|Is]) -> + Y =/= Reg1 andalso Y =/= Reg2 andalso 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]) -> diff --git a/lib/compiler/src/beam_types.erl b/lib/compiler/src/beam_types.erl new file mode 100644 index 0000000000..821ccd31bb --- /dev/null +++ b/lib/compiler/src/beam_types.erl @@ -0,0 +1,778 @@ +%% +%% %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% +%% + +-module(beam_types). + +-include("beam_types.hrl"). + +-import(lists, [foldl/3, reverse/1, reverse/2]). + +-export([meet/1, meet/2, join/1, join/2, subtract/2]). + +-export([get_singleton_value/1, + is_singleton_type/1, + is_boolean_type/1, + normalize/1]). + +-export([get_element_type/2, set_element_type/3]). + +-export([make_type_from_value/1]). + +-export([make_atom/1, + make_boolean/0, + make_integer/1, + make_integer/2]). + +-define(IS_LIST_TYPE(N), + N =:= list orelse + N =:= cons orelse + N =:= nil). + +-define(IS_NUMBER_TYPE(N), + N =:= number orelse + N =:= float orelse + is_record(N, t_integer)). + +-define(TUPLE_SET_LIMIT, 20). + +%% Folds meet/2 over a list. + +-spec meet([type()]) -> type(). + +meet([T1, T2 | Ts]) -> + meet([meet(T1, T2) | Ts]); +meet([T]) -> T. + +%% Return the "meet" of Type1 and Type2, which is more general than Type1 and +%% Type2. This is identical to glb/2 but can operate on and produce unions. +%% +%% A = #t_union{list=nil, number=[number], other=[#t_map{}]} +%% B = #t_union{number=[#t_integer{}], other=[#t_map{}]} +%% +%% meet(A, B) -> +%% #t_union{number=[#t_integer{}], other=[#t_map{}]} +%% +%% The meet of two different types result in 'none', which is the bottom +%% element for our type lattice: +%% +%% meet(#t_integer{}, #t_map{}) -> none + +-spec meet(type(), type()) -> type(). + +meet(T, T) -> + verified_type(T); +meet(any, T) -> + verified_type(T); +meet(T, any) -> + verified_type(T); +meet(#t_union{}=A, B) -> + meet_unions(A, B); +meet(A, #t_union{}=B) -> + meet_unions(B, A); +meet(A, B) -> + glb(A, B). + +meet_unions(#t_union{atom=AtomA,list=ListA,number=NumberA, + tuple_set=TSetA,other=OtherA}, + #t_union{atom=AtomB,list=ListB,number=NumberB, + tuple_set=TSetB,other=OtherB}) -> + Union = #t_union{atom=glb(AtomA, AtomB), + list=glb(ListA, ListB), + number=glb(NumberA, NumberB), + tuple_set=meet_tuple_sets(TSetA, TSetB), + other=glb(OtherA, OtherB)}, + shrink_union(Union); +meet_unions(#t_union{atom=AtomA}, #t_atom{}=B) -> + case glb(AtomA, B) of + none -> none; + Atom -> Atom + end; +meet_unions(#t_union{number=NumberA}, B) when ?IS_NUMBER_TYPE(B) -> + case glb(NumberA, B) of + none -> none; + Number -> Number + end; +meet_unions(#t_union{list=ListA}, B) when ?IS_LIST_TYPE(B) -> + case glb(ListA, B) of + none -> none; + List -> List + end; +meet_unions(#t_union{tuple_set=Tuples}, #t_tuple{}=B) -> + Set = meet_tuple_sets(Tuples, new_tuple_set(B)), + shrink_union(#t_union{tuple_set=Set}); +meet_unions(#t_union{other=OtherA}, OtherB) -> + case glb(OtherA, OtherB) of + none -> none; + Other -> Other + end. + +meet_tuple_sets(none, _) -> + none; +meet_tuple_sets(_, none) -> + none; +meet_tuple_sets(#t_tuple{}=A, #t_tuple{}=B) -> + new_tuple_set(glb(A, B)); +meet_tuple_sets(#t_tuple{}=Tuple, Records) -> + mts_tuple(Records, Tuple, []); +meet_tuple_sets(Records, #t_tuple{}=Tuple) -> + meet_tuple_sets(Tuple, Records); +meet_tuple_sets(RecordsA, RecordsB) -> + mts_records(RecordsA, RecordsB). + +mts_tuple([{Key, Type} | Records], Tuple, Acc) -> + case glb(Type, Tuple) of + none -> mts_tuple(Records, Tuple, Acc); + T -> mts_tuple(Records, Tuple, [{Key, T} | Acc]) + end; +mts_tuple([], _Tuple, [_|_]=Acc) -> + reverse(Acc); +mts_tuple([], _Tuple, []) -> + none. + +mts_records(RecordsA, RecordsB) -> + mts_records(RecordsA, RecordsB, []). + +mts_records([{Key, A} | RsA], [{Key, B} | RsB], Acc) -> + case glb(A, B) of + none -> mts_records(RsA, RsB, Acc); + T -> mts_records(RsA, RsB, [{Key, T} | Acc]) + end; +mts_records([{KeyA, _} | _ ]=RsA, [{KeyB, _} | RsB], Acc) when KeyA > KeyB -> + mts_records(RsA, RsB, Acc); +mts_records([{KeyA, _} | RsA], [{KeyB, _} | _] = RsB, Acc) when KeyA < KeyB -> + mts_records(RsA, RsB, Acc); +mts_records(_RsA, [], [_|_]=Acc) -> + reverse(Acc); +mts_records([], _RsB, [_|_]=Acc) -> + reverse(Acc); +mts_records(_RsA, _RsB, []) -> + none. + +%% Folds join/2 over a list. + +-spec join([type()]) -> type(). + +join([T1, T2| Ts]) -> + join([join(T1, T2) | Ts]); +join([T]) -> T. + +%% Return the "join" of Type1 and Type2, which is more general than Type1 and +%% Type2. This is identical to lub/2 but can operate on and produce unions. +%% +%% join(#t_integer{}, #t_map{}) -> #t_union{number=[#t_integer{}], +%% other=[#t_map{}]} + +-spec join(type(), type()) -> type(). + +join(T, T) -> T; +join(_T, any) -> any; +join(any, _T) -> any; +join(T, none) -> T; +join(none, T) -> T; + +join(#t_union{}=A, B) -> + join_unions(A, B); +join(A, #t_union{}=B) -> + join_unions(B, A); + +%% Union creation... +join(#t_atom{}=A, #t_atom{}=B) -> + lub(A, B); +join(#t_atom{}=A, B) when ?IS_LIST_TYPE(B) -> + #t_union{atom=A,list=B}; +join(#t_atom{}=A, B) when ?IS_NUMBER_TYPE(B) -> + #t_union{atom=A,number=B}; +join(#t_atom{}=A, #t_tuple{}=B) -> + #t_union{atom=A,tuple_set=new_tuple_set(B)}; +join(#t_atom{}=A, B) -> + #t_union{atom=A,other=B}; +join(A, #t_atom{}=B) -> + join(B, A); + +join(A, B) when ?IS_LIST_TYPE(A), ?IS_LIST_TYPE(B) -> + lub(A, B); +join(A, B) when ?IS_LIST_TYPE(A), ?IS_NUMBER_TYPE(B) -> + #t_union{list=A,number=B}; +join(A, #t_tuple{}=B) when ?IS_LIST_TYPE(A) -> + #t_union{list=A,tuple_set=new_tuple_set(B)}; +join(A, B) when ?IS_LIST_TYPE(A) -> + #t_union{list=A,other=B}; +join(A, B) when ?IS_LIST_TYPE(B) -> + join(B, A); + +join(A, B) when ?IS_NUMBER_TYPE(A), ?IS_NUMBER_TYPE(B) -> + lub(A, B); +join(A, #t_tuple{}=B) when ?IS_NUMBER_TYPE(A) -> + #t_union{number=A,tuple_set=new_tuple_set(B)}; +join(A, B) when ?IS_NUMBER_TYPE(A) -> + #t_union{number=A,other=B}; +join(A, B) when ?IS_NUMBER_TYPE(B) -> + join(B, A); + +join(#t_tuple{}=A, #t_tuple{}=B) -> + case {record_key(A), record_key(B)} of + {Same, Same} -> + lub(A, B); + {none, _Key} -> + lub(A, B); + {_Key, none} -> + lub(A, B); + {KeyA, KeyB} when KeyA < KeyB -> + #t_union{tuple_set=[{KeyA, A}, {KeyB, B}]}; + {KeyA, KeyB} when KeyA > KeyB -> + #t_union{tuple_set=[{KeyB, B}, {KeyA, A}]} + end; +join(#t_tuple{}=A, B) -> + %% All other combinations have been tried already, so B must be 'other' + #t_union{tuple_set=new_tuple_set(A),other=B}; +join(A, #t_tuple{}=B) -> + join(B, A); + +join(A, B) -> + lub(A, B). + +join_unions(#t_union{atom=AtomA,list=ListA,number=NumberA, + tuple_set=TSetA,other=OtherA}, + #t_union{atom=AtomB,list=ListB,number=NumberB, + tuple_set=TSetB,other=OtherB}) -> + Union = #t_union{atom=lub(AtomA, AtomB), + list=lub(ListA, ListB), + number=lub(NumberA, NumberB), + tuple_set=join_tuple_sets(TSetA, TSetB), + other=lub(OtherA, OtherB)}, + shrink_union(Union); +join_unions(#t_union{atom=AtomA}=A, #t_atom{}=B) -> + A#t_union{atom=lub(AtomA, B)}; +join_unions(#t_union{list=ListA}=A, B) when ?IS_LIST_TYPE(B) -> + A#t_union{list=lub(ListA, B)}; +join_unions(#t_union{number=NumberA}=A, B) when ?IS_NUMBER_TYPE(B) -> + A#t_union{number=lub(NumberA, B)}; +join_unions(#t_union{tuple_set=TSetA}=A, #t_tuple{}=B) -> + Set = join_tuple_sets(TSetA, new_tuple_set(B)), + shrink_union(A#t_union{tuple_set=Set}); +join_unions(#t_union{other=OtherA}=A, B) -> + case lub(OtherA, B) of + any -> any; + T -> A#t_union{other=T} + end. + +join_tuple_sets(A, none) -> + A; +join_tuple_sets(none, B) -> + B; +join_tuple_sets(#t_tuple{}=A, #t_tuple{}=B) -> + lub(A, B); +join_tuple_sets(#t_tuple{}=Tuple, Records) -> + jts_tuple(Records, Tuple); +join_tuple_sets(Records, #t_tuple{}=Tuple) -> + join_tuple_sets(Tuple, Records); +join_tuple_sets(RecordsA, RecordsB) -> + jts_records(RecordsA, RecordsB). + +jts_tuple([{_Key, Tuple} | Records], Acc) -> + jts_tuple(Records, lub(Tuple, Acc)); +jts_tuple([], Acc) -> + Acc. + +jts_records(RsA, RsB) -> + jts_records(RsA, RsB, 0, []). + +jts_records(RsA, RsB, N, Acc) when N > ?TUPLE_SET_LIMIT -> + A = normalize_tuple_set(RsA, none), + B = normalize_tuple_set(RsB, A), + #t_tuple{} = normalize_tuple_set(Acc, B); +jts_records([{Key, A} | RsA], [{Key, B} | RsB], N, Acc) -> + jts_records(RsA, RsB, N + 1, [{Key, lub(A, B)} | Acc]); +jts_records([{KeyA, _} | _]=RsA, [{KeyB, B} | RsB], N, Acc) when KeyA > KeyB -> + jts_records(RsA, RsB, N + 1, [{KeyB, B} | Acc]); +jts_records([{KeyA, A} | RsA], [{KeyB, _} | _] = RsB, N, Acc) when KeyA < KeyB -> + jts_records(RsA, RsB, N + 1, [{KeyA, A} | Acc]); +jts_records([], RsB, _N, Acc) -> + reverse(Acc, RsB); +jts_records(RsA, [], _N, Acc) -> + reverse(Acc, RsA). + +%% Subtract Type2 from Type1. Example: +%% subtract(list, cons) -> nil + +-spec subtract(type(), type()) -> type(). + +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_union{atom=Atom}=A, #t_atom{}=B)-> + shrink_union(A#t_union{atom=subtract(Atom, B)}); +subtract(#t_union{number=Number}=A, B) when ?IS_NUMBER_TYPE(B) -> + shrink_union(A#t_union{number=subtract(Number, B)}); +subtract(#t_union{list=List}=A, B) when ?IS_LIST_TYPE(B) -> + shrink_union(A#t_union{list=subtract(List, B)}); +subtract(#t_union{tuple_set=[_|_]=Records0}=A, #t_tuple{}=B) -> + %% Filter out all records that are strictly more specific than B. + NewSet = case [{Key, T} || {Key, T} <- Records0, meet(T, B) =/= T] of + [_|_]=Records -> Records; + [] -> none + end, + shrink_union(A#t_union{tuple_set=NewSet}); +subtract(#t_union{tuple_set=#t_tuple{}=Tuple}=A, #t_tuple{}=B) -> + %% Exclude Tuple if it's strictly more specific than B. + case meet(Tuple, B) of + Tuple -> shrink_union(A#t_union{tuple_set=none}); + _ -> A + end; + +subtract(T, _) -> T. + +%%% +%%% Type operators +%%% + +-spec get_singleton_value(Type) -> Result when + Type :: type(), + Result :: {ok, term()} | error. +get_singleton_value(#t_atom{elements=[Atom]}) -> + {ok, Atom}; +get_singleton_value(#t_integer{elements={Int,Int}}) -> + {ok, Int}; +get_singleton_value(nil) -> + {ok, []}; +get_singleton_value(_) -> + error. + +-spec is_boolean_type(type()) -> boolean(). +is_boolean_type(#t_atom{elements=[F,T]}) -> + F =:= false andalso T =:= true; +is_boolean_type(#t_atom{elements=[B]}) -> + is_boolean(B); +is_boolean_type(#t_union{}=T) -> + is_boolean_type(normalize(T)); +is_boolean_type(_) -> + false. + +-spec is_singleton_type(type()) -> boolean(). +is_singleton_type(Type) -> + get_singleton_value(Type) =/= error. + +-spec set_element_type(Key, Type, Elements) -> Elements when + Key :: term(), + Type :: type(), + Elements :: elements(). +set_element_type(_Key, none, Es) -> + Es; +set_element_type(Key, any, Es) -> + maps:remove(Key, Es); +set_element_type(Key, Type, Es) -> + Es#{ Key => Type }. + +-spec get_element_type(Key, Elements) -> type() when + Key :: term(), + Elements :: elements(). +get_element_type(Index, Es) -> + case Es of + #{ Index := T } -> T; + #{} -> any + end. + +-spec normalize(type()) -> normal_type(). +normalize(#t_union{atom=Atom,list=List,number=Number, + tuple_set=Tuples,other=Other}) -> + A = lub(Atom, List), + B = lub(A, Number), + C = lub(B, Other), + normalize_tuple_set(Tuples, C); +normalize(T) -> + verified_normal_type(T). + +normalize_tuple_set([{_, A} | Records], B) -> + normalize_tuple_set(Records, lub(A, B)); +normalize_tuple_set([], B) -> + B; +normalize_tuple_set(A, B) -> + lub(A, B). + +%%% +%%% Type constructors +%%% + +-spec make_type_from_value(term()) -> type(). +make_type_from_value(Value) -> + mtfv_1(Value). + +mtfv_1([]) -> nil; +mtfv_1([_|_]) -> cons; +mtfv_1(A) when is_atom(A) -> #t_atom{elements=[A]}; +mtfv_1(B) when is_binary(B) -> #t_bitstring{unit=8}; +mtfv_1(B) when is_bitstring(B) -> #t_bitstring{}; +mtfv_1(F) when is_float(F) -> float; +mtfv_1(F) when is_function(F) -> + {arity, Arity} = erlang:fun_info(F, arity), + #t_fun{arity=Arity}; +mtfv_1(I) when is_integer(I) -> make_integer(I); +mtfv_1(M) when is_map(M) -> #t_map{}; +mtfv_1(T) when is_tuple(T) -> + {Es,_} = foldl(fun(Val, {Es0, Index}) -> + Type = mtfv_1(Val), + Es = set_element_type(Index, Type, Es0), + {Es, Index + 1} + end, {#{}, 1}, tuple_to_list(T)), + #t_tuple{exact=true,size=tuple_size(T),elements=Es}; +mtfv_1(_Term) -> + any. + +-spec make_atom(atom()) -> type(). +make_atom(Atom) when is_atom(Atom) -> + #t_atom{elements=[Atom]}. + +-spec make_boolean() -> type(). +make_boolean() -> + #t_atom{elements=[false,true]}. + +-spec make_integer(integer()) -> type(). +make_integer(Int) when is_integer(Int) -> + make_integer(Int, Int). + +-spec make_integer(Min, Max) -> type() when + Min :: integer(), + Max :: integer(). +make_integer(Min, Max) when is_integer(Min), is_integer(Max), Min =< Max -> + #t_integer{elements={Min,Max}}. + +%%% +%%% Helpers +%%% + +%% Return the greatest lower bound of the types Type1 and Type2. The GLB is a +%% more specific type than Type1 and Type2, and is always a normal type. +%% +%% glb(#t_integer{elements=any}, #t_integer{elements={0,3}}) -> +%% #t_integer{elements={0,3}} +%% +%% The GLB of two different types result in 'none', which is the bottom +%% element for our type lattice: +%% +%% glb(#t_integer{}, #t_map{}) -> none + +-spec glb(normal_type(), normal_type()) -> normal_type(). + +glb(T, T) -> + verified_normal_type(T); +glb(any, T) -> + verified_normal_type(T); +glb(T, any) -> + verified_normal_type(T); +glb(#t_atom{elements=[_|_]=Set1}, #t_atom{elements=[_|_]=Set2}) -> + case ordsets:intersection(Set1, Set2) of + [] -> + none; + [_|_]=Set -> + #t_atom{elements=Set} + end; +glb(#t_atom{elements=[_|_]}=T, #t_atom{elements=any}) -> + T; +glb(#t_atom{elements=any}, #t_atom{elements=[_|_]}=T) -> + T; +glb(#t_bs_context{slots=SlotCountA,valid=ValidSlotsA}, + #t_bs_context{slots=SlotCountB,valid=ValidSlotsB}) -> + CommonSlotMask = (1 bsl min(SlotCountA, SlotCountB)) - 1, + CommonSlotsA = ValidSlotsA band CommonSlotMask, + CommonSlotsB = ValidSlotsB band CommonSlotMask, + if + CommonSlotsA =:= CommonSlotsB -> + #t_bs_context{slots=max(SlotCountA, SlotCountB), + valid=ValidSlotsA bor ValidSlotsB}; + CommonSlotsA =/= CommonSlotsB -> + none + end; +glb(#t_fun{arity=any}, #t_fun{}=T) -> + T; +glb(#t_fun{}=T, #t_fun{arity=any}) -> + T; +glb(#t_integer{elements={_,_}}=T, #t_integer{elements=any}) -> + T; +glb(#t_integer{elements=any}, #t_integer{elements={_,_}}=T) -> + T; +glb(#t_integer{elements={MinA,MaxA}}, #t_integer{elements={MinB,MaxB}}) + when MinA >= MinB, MinA =< MaxB; + MinB >= MinA, MinB =< MaxA -> + true = MinA =< MaxA andalso MinB =< MaxB, %Assertion. + #t_integer{elements={max(MinA, MinB),min(MaxA, MaxB)}}; +glb(#t_integer{}=T, number) -> T; +glb(float=T, number) -> T; +glb(number, #t_integer{}=T) -> T; +glb(number, float=T) -> T; +glb(list, cons) -> cons; +glb(list, nil) -> nil; +glb(cons, list) -> cons; +glb(nil, list) -> nil; +glb(#t_tuple{}=T1, #t_tuple{}=T2) -> + glb_tuples(T1, T2); +glb(#t_bitstring{unit=U1}, #t_bitstring{unit=U2}) -> + #t_bitstring{unit=U1 * U2 div gcd(U1, U2)}; +glb(_, _) -> + %% Inconsistent types. There will be an exception at runtime. + none. + +glb_tuples(#t_tuple{size=Sz1,exact=true}, + #t_tuple{size=Sz2,exact=true}) when Sz1 =/= Sz2 -> + none; +glb_tuples(#t_tuple{size=Sz1,exact=Ex1,elements=Es1}, + #t_tuple{size=Sz2,exact=Ex2,elements=Es2}) -> + Size = max(Sz1, Sz2), + Exact = Ex1 or Ex2, + case glb_elements(Es1, Es2) of + none -> + none; + Es -> + #t_tuple{size=Size,exact=Exact,elements=Es} + end. + +glb_elements(Es1, Es2) -> + Keys = maps:keys(Es1) ++ maps:keys(Es2), + glb_elements_1(Keys, Es1, Es2, #{}). + +glb_elements_1([Key | Keys], Es1, Es2, Acc) -> + case {Es1, Es2} of + {#{ Key := Type1 }, #{ Key := Type2 }} -> + %% Note the use of meet/2; elements don't need to be normal types. + case meet(Type1, Type2) of + none -> none; + Type -> glb_elements_1(Keys, Es1, Es2, Acc#{ Key => Type }) + end; + {#{ Key := Type1 }, _} -> + glb_elements_1(Keys, Es1, Es2, Acc#{ Key => Type1 }); + {_, #{ Key := Type2 }} -> + glb_elements_1(Keys, Es1, Es2, Acc#{ Key => Type2 }) + end; +glb_elements_1([], _Es1, _Es2, Acc) -> + Acc. + +%% Return the least upper bound of the types Type1 and Type2. The LUB is a more +%% general type than Type1 and Type2, and is always a normal type. +%% +%% For example: +%% +%% lub(#t_integer{elements=any}, #t_integer=elements={0,3}}) -> +%% #t_integer{} +%% +%% The LUB for two different types result in 'any' (not a union type!), which +%% is the top element for our type lattice: +%% +%% lub(#t_integer{}, #t_map{}) -> any + +-spec lub(normal_type(), normal_type()) -> normal_type(). + +lub(T, T) -> + verified_normal_type(T); +lub(none, T) -> + verified_normal_type(T); +lub(T, none) -> + verified_normal_type(T); +lub(any, _) -> + any; +lub(_, any) -> + any; +lub(#t_atom{elements=[_|_]=Set1}, #t_atom{elements=[_|_]=Set2}) -> + Set = ordsets:union(Set1, Set2), + case ordsets:size(Set) of + Size when Size =< ?ATOM_SET_SIZE -> + #t_atom{elements=Set}; + _Size -> + #t_atom{elements=any} + end; +lub(#t_atom{elements=any}=T, #t_atom{elements=[_|_]}) -> T; +lub(#t_atom{elements=[_|_]}, #t_atom{elements=any}=T) -> T; +lub(#t_bitstring{unit=U1}, #t_bitstring{unit=U2}) -> + #t_bitstring{unit=gcd(U1, U2)}; +lub(#t_fun{}, #t_fun{}) -> + #t_fun{}; +lub(#t_integer{elements={MinA,MaxA}}, + #t_integer{elements={MinB,MaxB}}) -> + #t_integer{elements={min(MinA,MinB),max(MaxA,MaxB)}}; +lub(#t_bs_context{slots=SlotsA,valid=ValidA}, + #t_bs_context{slots=SlotsB,valid=ValidB}) -> + #t_bs_context{slots=min(SlotsA, SlotsB), + valid=ValidA band ValidB}; +lub(#t_integer{}, #t_integer{}) -> #t_integer{}; +lub(list, cons) -> list; +lub(cons, list) -> list; +lub(nil, cons) -> list; +lub(cons, nil) -> list; +lub(nil, list) -> list; +lub(list, nil) -> list; +lub(#t_integer{}, float) -> number; +lub(float, #t_integer{}) -> number; +lub(#t_integer{}, number) -> number; +lub(number, #t_integer{}) -> number; +lub(float, number) -> number; +lub(number, float) -> number; +lub(#t_tuple{size=Sz,exact=ExactA,elements=EsA}, + #t_tuple{size=Sz,exact=ExactB,elements=EsB}) -> + Exact = ExactA and ExactB, + Es = lub_tuple_elements(Sz, EsA, EsB), + #t_tuple{size=Sz,exact=Exact,elements=Es}; +lub(#t_tuple{size=SzA,elements=EsA}, #t_tuple{size=SzB,elements=EsB}) -> + Sz = min(SzA, SzB), + Es = lub_tuple_elements(Sz, EsA, EsB), + #t_tuple{size=Sz,elements=Es}; +lub(_T1, _T2) -> + %%io:format("~p ~p\n", [_T1,_T2]), + any. + +lub_tuple_elements(MinSize, EsA, EsB) -> + Es0 = lub_elements(EsA, EsB), + maps:filter(fun(Index, _Type) -> Index =< MinSize end, Es0). + +lub_elements(Es1, Es2) -> + Keys = if + map_size(Es1) =< map_size(Es2) -> maps:keys(Es1); + map_size(Es1) > map_size(Es2) -> maps:keys(Es2) + end, + lub_elements_1(Keys, Es1, Es2, #{}). + +lub_elements_1([Key | Keys], Es1, Es2, Acc0) -> + case {Es1, Es2} of + {#{ Key := Type1 }, #{ Key := Type2 }} -> + %% Note the use of join/2; elements don't need to be normal types. + Acc = set_element_type(Key, join(Type1, Type2), Acc0), + lub_elements_1(Keys, Es1, Es2, Acc); + {#{}, #{}} -> + lub_elements_1(Keys, Es1, Es2, Acc0) + end; +lub_elements_1([], _Es1, _Es2, Acc) -> + Acc. + +%% + +gcd(A, B) -> + case A rem B of + 0 -> B; + X -> gcd(B, X) + end. + +%% + +record_key(#t_tuple{exact=true,size=Size,elements=#{ 1 := Tag }}) -> + case is_singleton_type(Tag) of + true -> {Size, Tag}; + false -> none + end; +record_key(_) -> + none. + +new_tuple_set(T) -> + case record_key(T) of + none -> T; + Key -> [{Key, T}] + end. + +%% + +shrink_union(#t_union{other=any}) -> + any; +shrink_union(#t_union{atom=Atom,list=none,number=none, + tuple_set=none,other=none}) -> + Atom; +shrink_union(#t_union{atom=none,list=List,number=none, + tuple_set=none,other=none}) -> + List; +shrink_union(#t_union{atom=none,list=none,number=Number, + tuple_set=none,other=none}) -> + Number; +shrink_union(#t_union{atom=none,list=none,number=none, + tuple_set=#t_tuple{}=Tuple,other=none}) -> + Tuple; +shrink_union(#t_union{atom=none,list=none,number=none, + tuple_set=[{_Key, Record}],other=none}) -> + #t_tuple{} = Record; %Assertion. +shrink_union(#t_union{atom=none,list=none,number=none, + tuple_set=none,other=Other}) -> + Other; +shrink_union(#t_union{}=T) -> + T. + +%% Verifies that the given type is well-formed. + +-spec verified_type(T) -> T when + T :: type(). + +verified_type(#t_union{atom=Atom, + list=List, + number=Number, + tuple_set=TSet, + other=Other}=T) -> + _ = verified_normal_type(Atom), + _ = verified_normal_type(List), + _ = verified_normal_type(Number), + _ = verify_tuple_set(TSet), + _ = verified_normal_type(Other), + T; +verified_type(T) -> + verified_normal_type(T). + +verify_tuple_set([_|_]=T) -> + _ = [verified_normal_type(Rec) || {_, Rec} <- T], + T; +verify_tuple_set(#t_tuple{}=T) -> + none = record_key(T), %Assertion. + T; +verify_tuple_set(none=T) -> + T. + +-spec verified_normal_type(T) -> T when + T :: normal_type(). + +verified_normal_type(any=T) -> T; +verified_normal_type(none=T) -> T; +verified_normal_type(#t_atom{elements=any}=T) -> T; +verified_normal_type(#t_atom{elements=[_|_]}=T) -> T; +verified_normal_type(#t_bitstring{unit=U}=T) + when is_integer(U), U >= 1 -> + T; +verified_normal_type(#t_bs_context{}=T) -> T; +verified_normal_type(#t_fun{arity=Arity}=T) + when Arity =:= any; is_integer(Arity) -> + T; +verified_normal_type(float=T) -> T; +verified_normal_type(#t_integer{elements=any}=T) -> T; +verified_normal_type(#t_integer{elements={Min,Max}}=T) + when is_integer(Min), is_integer(Max), Min =< Max -> + T; +verified_normal_type(list=T) -> T; +verified_normal_type(#t_map{}=T) -> T; +verified_normal_type(nil=T) -> T; +verified_normal_type(cons=T) -> T; +verified_normal_type(number=T) -> T; +verified_normal_type(#t_tuple{size=Size,elements=Es}=T) -> + %% All known elements must have a valid index and type (which may be a + %% union). 'any' is prohibited since it's implicit and should never be + %% present in the map, and a 'none' element ought to have reduced the + %% entire tuple to 'none'. + maps:fold(fun(Index, Element, _) when is_integer(Index), + 1 =< Index, Index =< Size, + Element =/= any, Element =/= none -> + verified_type(Element) + end, [], Es), + T. diff --git a/lib/compiler/src/beam_types.hrl b/lib/compiler/src/beam_types.hrl new file mode 100644 index 0000000000..09f87d61ba --- /dev/null +++ b/lib/compiler/src/beam_types.hrl @@ -0,0 +1,88 @@ +%% +%% %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% +%% + +%% Common term types for passes operating on beam SSA and assembly. Helper +%% functions for wrangling these can be found in beam_types.erl +%% +%% The type lattice is as follows: +%% +%% any Any Erlang term (top element). +%% +%% - #t_atom{} Atom, or a set thereof. +%% - #t_bitstring{} Bitstring. +%% - #t_bs_context{} Match context. +%% - #t_fun{} Fun. +%% - #t_map{} Map. +%% - number Any number. +%% -- float Floating point number. +%% -- integer Integer. +%% - list Any list. +%% -- cons Cons (nonempty list). +%% -- nil The empty list. +%% - #t_tuple{} Tuple. +%% +%% none No type (bottom element). +%% +%% We also use #t_union{} to represent conflicting types produced by certain +%% expressions, e.g. the "#t_atom{} or #t_tuple{}" of lists:keyfind/3, which is +%% very useful for preserving type information when we would otherwise have +%% reduced it to 'any'. Since few operations can make direct use of this extra +%% type information, types should generally be normalized to one of the above +%% before use. + +-define(ATOM_SET_SIZE, 5). + +-record(t_atom, {elements=any :: 'any' | [atom()]}). +-record(t_fun, {arity=any :: arity() | 'any'}). +-record(t_integer, {elements=any :: 'any' | {integer(),integer()}}). +-record(t_bitstring, {unit=1 :: pos_integer()}). +-record(t_bs_context, {slots=0 :: non_neg_integer(), + valid=0 :: non_neg_integer()}). +-record(t_map, {elements=#{} :: map_elements()}). +-record(t_tuple, {size=0 :: integer(), + exact=false :: boolean(), + elements=#{} :: tuple_elements()}). + +%% Known element types, unknown elements are assumed to be 'any'. The key is +%% a 1-based integer index for tuples, and a plain literal for maps (that is, +%% not wrapped in a #b_literal{}, just the value itself). + +-type tuple_elements() :: #{ Key :: pos_integer() => type() }. +-type map_elements() :: #{ Key :: term() => type() }. + +-type elements() :: tuple_elements() | map_elements(). + +-type normal_type() :: any | none | + list | number | + #t_atom{} | #t_bitstring{} | #t_bs_context{} | + #t_fun{} | #t_integer{} | #t_map{} | #t_tuple{} | + 'cons' | 'float' | 'nil'. + +-type record_key() :: {Arity :: integer(), Tag :: normal_type() }. +-type record_set() :: ordsets:ordset({record_key(), #t_tuple{}}). +-type tuple_set() :: #t_tuple{} | record_set(). + +-record(t_union, {atom=none :: none | #t_atom{}, + list=none :: none | list | cons | nil, + number=none :: none | number | float | #t_integer{}, + tuple_set=none :: none | tuple_set(), + other=none :: normal_type()}). + +-type type() :: #t_union{} | normal_type(). diff --git a/lib/compiler/src/beam_validator.erl b/lib/compiler/src/beam_validator.erl index 349d74eb58..911b5eb777 100644 --- a/lib/compiler/src/beam_validator.erl +++ b/lib/compiler/src/beam_validator.erl @@ -19,6 +19,10 @@ -module(beam_validator). +-include("beam_types.hrl"). + +-define(UNICODE_MAX, (16#10FFFF)). + -compile({no_auto_import,[min/2]}). %% Avoid warning for local function error/1 clashing with autoimported BIF. @@ -26,7 +30,6 @@ %% Interface for compiler. -export([module/2, format_error/1]). --export([type_anno/1, type_anno/2, type_anno/4]). -import(lists, [dropwhile/2,foldl/3,member/2,reverse/1,sort/1,zip/2]). @@ -45,34 +48,6 @@ 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,_}) -> binary; -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) when is_atom(Value) -> {atom, Value}; -type_anno(float, Value) when is_float(Value) -> {float, Value}; -type_anno(integer, Value) when is_integer(Value) -> {integer, Value}. - --spec type_anno(term(), term(), term(), term()) -> term(). -type_anno(tuple, Size, Exact, Elements) when is_integer(Size), Size >= 0, - is_map(Elements) -> - case Exact of - true -> {tuple, Size, Elements}; - false -> {tuple, [Size], Elements} - end. - -spec format_error(term()) -> iolist(). format_error({{_M,F,A},{I,Off,limit}}) -> @@ -119,7 +94,7 @@ format_error(Error) -> %% format as used in the compiler and in .S files. validate(Module, Fs) -> - Ft = index_parameter_types(Fs, []), + Ft = build_function_table(Fs, []), validate_0(Module, Fs, Ft). validate_0(_Module, [], _) -> []; @@ -136,8 +111,15 @@ validate_0(Module, [{function,Name,Ar,Entry,Code}|Fs], Ft) -> erlang:raise(Class, Error, Stack) end. +-record(t_abstract, {kind}). + +%% The types are the same as in 'beam_types.hrl', with the addition of +%% #t_abstract{} that describes tuples under construction, match context +%% positions, and so on. +-type validator_type() :: #t_abstract{} | type(). + -record(value_ref, {id :: index()}). --record(value, {op :: term(), args :: [argument()], type :: type()}). +-record(value, {op :: term(), args :: [argument()], type :: validator_type()}). -type argument() :: #value_ref{} | literal(). @@ -149,34 +131,28 @@ validate_0(Module, [{function,Name,Ar,Entry,Code}|Fs], Ft) -> {literal, term()} | nil. --type tuple_sz() :: [non_neg_integer()] | %% Inexact - non_neg_integer(). %% Exact. - -%% Match context type. --record(ms, - {id=make_ref() :: reference(), %Unique ID. - valid=0 :: non_neg_integer(), %Valid slots - slots=0 :: non_neg_integer() %Number of slots - }). - --type type() :: binary | - cons | - list | - map | - nil | - #ms{} | - ms_position | - none | - number | - term | - tuple_in_progress | - {tuple, tuple_sz(), #{ literal() => type() }} | - literal(). - +%% Register tags describe the state of the register rather than the value they +%% contain (if any). +%% +%% initialized The register has been initialized with some valid term +%% so that it is safe to pass to the garbage collector. +%% NOT safe to use in any other way (will not crash the +%% emulator, but clearly points to a bug in the compiler). +%% +%% uninitialized The register contains any old garbage and can not be +%% passed to the garbage collector. +%% +%% {catchtag,[Lbl]} A special term used within a catch. Must only be used +%% by the catch instructions; NOT safe to use in other +%% instructions. +%% +%% {trytag,[Lbl]} A special term used within a try block. Must only be +%% used by the catch instructions; NOT safe to use in other +%% instructions. -type tag() :: initialized | uninitialized | - {catchtag, [label()]} | - {trytag, [label()]}. + {catchtag, ordsets:ordset(label())} | + {trytag, ordsets:ordset(label())}. -type x_regs() :: #{ {x, index()} => #value_ref{} }. -type y_regs() :: #{ {y, index()} => tag() | #value_ref{} }. @@ -200,11 +176,11 @@ validate_0(Module, [{function,Name,Ar,Entry,Code}|Fs], Ft) -> numy=none :: none | undecided | index(), %% Available heap size. h=0, - %Available heap size for floats. + %%Available heap size for floats. hf=0, %% Floating point state. fls=undefined, - %% List of hot catch/try labels + %% List of hot catch/try tags ct=[], %% Previous instruction was setelement/3. setelem=false, @@ -225,36 +201,32 @@ validate_0(Module, [{function,Name,Ar,Entry,Code}|Fs], Ft) -> branched=gb_trees:empty() :: branched_tab(), %% All defined labels labels=gb_sets:empty() :: label_set(), - %% Argument information of other functions in the module + %% Information of other functions in the module ft=gb_trees:empty() :: ft_tab(), %% Counter for #value_ref{} creation ref_ctr=0 :: index() }). -index_parameter_types([{function,_,_,Entry,Code0}|Fs], Acc0) -> +build_function_table([{function,_,Arity,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); + Info = #{ arity => Arity, + parameter_types => find_parameter_types(Is, #{}) }, + build_function_table(Fs, [{Entry, Info} | Acc0]); _ -> - %% Something serious is wrong. Ignore it for now. + %% Something is seriously wrong. Ignore it for now. %% It will be detected and diagnosed later. - index_parameter_types(Fs, Acc0) + build_function_table(Fs, Acc0) end; -index_parameter_types([], Acc) -> +build_function_table([], Acc) -> gb_trees:from_orddict(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) -> +find_parameter_types([{'%', {type_info, Reg, Type}} | Is], Acc) -> + find_parameter_types(Is, Acc#{ Reg => Type }); +find_parameter_types(_, Acc) -> Acc. validate_1(Is, Name, Arity, Entry, Ft) -> @@ -326,7 +298,7 @@ init_vst(Arity, Ls1, Ls2, Ft) -> init_function_args(-1, Vst) -> Vst; init_function_args(X, Vst) -> - init_function_args(X - 1, create_term(term, argument, [], {x,X}, Vst)). + init_function_args(X - 1, create_term(any, argument, [], {x,X}, Vst)). kill_heap_allocation(St) -> St#st{h=0,hf=0}. @@ -381,17 +353,34 @@ valfun_1({try_case_end,Src}, Vst) -> kill_state(Vst); %% Instructions that cannot cause exceptions valfun_1({bs_get_tail,Ctx,Dst,Live}, Vst0) -> - bsm_validate_context(Ctx, Vst0), + assert_type(#t_bs_context{}, Ctx, Vst0), verify_live(Live, Vst0), verify_y_init(Vst0), Vst = prune_x_regs(Live, Vst0), - extract_term(binary, bs_get_tail, [Ctx], Dst, Vst, Vst0); + extract_term(#t_bitstring{}, bs_get_tail, [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,Src,Dst}, Vst) -> assign(Src, Dst, Vst); +valfun_1({swap,RegA,RegB}, Vst0) -> + assert_movable(RegA, Vst0), + assert_movable(RegB, Vst0), + + %% We don't expect fragile registers to be swapped. + %% Therefore, we can conservatively make both registers + %% fragile if one of the register is fragile instead of + %% swapping the fragility of the registers. + Sources = [RegA,RegB], + Vst1 = propagate_fragility(RegA, Sources, Vst0), + Vst2 = propagate_fragility(RegB, Sources, Vst1), + + %% Swap the value references. + VrefA = get_reg_vref(RegA, Vst2), + VrefB = get_reg_vref(RegB, Vst2), + Vst = set_reg_vref(VrefB, RegA, Vst2), + set_reg_vref(VrefA, RegB, Vst); valfun_1({fmove,Src,{fr,_}=Dst}, Vst) -> assert_type(float, Src, Vst), set_freg(Dst, Vst); @@ -399,25 +388,36 @@ valfun_1({fmove,{fr,_}=Src,Dst}, Vst0) -> assert_freg_set(Src, Vst0), assert_fls(checked, Vst0), Vst = eat_heap_float(Vst0), - create_term({float,[]}, fmove, [], Dst, Vst); + create_term(float, fmove, [], Dst, Vst); valfun_1({kill,Reg}, Vst) -> create_tag(initialized, kill, [], Reg, Vst); valfun_1({init,Reg}, Vst) -> create_tag(initialized, init, [], Reg, Vst); valfun_1({test_heap,Heap,Live}, Vst) -> test_heap(Heap, Live, Vst); -valfun_1({bif,Op,{f,_},Ss,Dst}=I, Vst) -> - case is_bif_safe(Op, length(Ss)) of - false -> - %% Since the BIF can fail, make sure that any catch state - %% is updated. - valfun_2(I, Vst); - true -> - %% It can't fail, so we finish handling it here (not updating - %% catch state). - validate_src(Ss, Vst), - Type = bif_return_type(Op, Ss, Vst), - extract_term(Type, {bif,Op}, Ss, Dst, Vst) +valfun_1({bif,Op,{f,0},Ss,Dst}=I, Vst) -> + case will_bif_succeed(Op, Ss, Vst) of + yes -> + %% This BIF cannot fail, handle it here without updating catch + %% state. + validate_bif(Op, cannot_fail, Ss, Dst, Vst); + no -> + %% The stack will be scanned, so Y registers must be initialized. + verify_y_init(Vst), + kill_state(Vst); + maybe -> + %% The BIF can fail, make sure that any catch state is updated. + valfun_2(I, Vst) + end; +valfun_1({gc_bif,Op,{f,0},Live,Ss,Dst}=I, Vst) -> + case will_bif_succeed(Op, Ss, Vst) of + yes -> + validate_gc_bif(Op, cannot_fail, Ss, Dst, Live, Vst); + no -> + verify_y_init(Vst), + kill_state(Vst); + maybe -> + valfun_2(I, Vst) end; %% Put instructions. valfun_1({put_list,A,B,Dst}, Vst0) -> @@ -431,14 +431,15 @@ valfun_1({put_tuple2,Dst,{list,Elements}}, Vst0) -> Vst = eat_heap(Size+1, Vst0), {Es,_} = foldl(fun(Val, {Es0, Index}) -> Type = get_term_type(Val, Vst0), - Es = set_element_type({integer,Index}, Type, Es0), + Es = beam_types:set_element_type(Index, Type, Es0), {Es, Index + 1} end, {#{}, 1}, Elements), - Type = {tuple,Size,Es}, + Type = #t_tuple{exact=true,size=Size,elements=Es}, create_term(Type, put_tuple2, [], Dst, Vst); valfun_1({put_tuple,Sz,Dst}, Vst0) when is_integer(Sz) -> Vst1 = eat_heap(1, Vst0), - Vst = create_term(tuple_in_progress, put_tuple, [], Dst, Vst1), + Vst = create_term(#t_abstract{kind=unfinished_tuple}, put_tuple, [], + Dst, Vst1), #vst{current=St0} = Vst, St = St0#st{puts_left={Sz,{Dst,Sz,#{}}}}, Vst#vst{current=St}; @@ -450,15 +451,31 @@ valfun_1({put,Src}, Vst0) -> #st{puts_left=none} -> error(not_building_a_tuple); #st{puts_left={1,{Dst,Sz,Es0}}} -> - Es = Es0#{ {integer,Sz} => get_term_type(Src, Vst0) }, + ElementType = get_term_type(Src, Vst0), + Es = beam_types:set_element_type(Sz, ElementType, Es0), St = St0#st{puts_left=none}, - create_term({tuple,Sz,Es}, put_tuple, [], Dst, Vst#vst{current=St}); + Type = #t_tuple{exact=true,size=Sz,elements=Es}, + create_term(Type, put_tuple, [], Dst, Vst#vst{current=St}); #st{puts_left={PutsLeft,{Dst,Sz,Es0}}} when is_integer(PutsLeft) -> Index = Sz - PutsLeft + 1, - Es = Es0#{ {integer,Index} => get_term_type(Src, Vst0) }, + ElementType = get_term_type(Src, Vst0), + Es = beam_types:set_element_type(Index, ElementType, Es0), St = St0#st{puts_left={PutsLeft-1,{Dst,Sz,Es}}}, Vst#vst{current=St} end; +%% This instruction never fails, though it may be invalid in some contexts; see +%% val_dsetel/2 +valfun_1({set_tuple_element,Src,Tuple,N}, Vst) -> + I = N + 1, + assert_term(Src, Vst), + assert_type(#t_tuple{size=I}, Tuple, Vst), + %% Manually update the tuple type; we can't rely on the ordinary update + %% helpers as we must support overwriting (rather than just widening or + %% narrowing) known elements, and we can't use extract_term either since + %% the source tuple may be aliased. + #t_tuple{elements=Es0}=Type = normalize(get_term_type(Tuple, Vst)), + Es = beam_types:set_element_type(I, get_term_type(Src, Vst), Es0), + override_type(Type#t_tuple{elements=Es}, Tuple, Vst); %% Instructions for optimization of selective receives. valfun_1({recv_mark,{f,Fail}}, Vst) when is_integer(Fail) -> Vst; @@ -469,8 +486,13 @@ 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, none}}, Vst) -> + %% Unreachable code, typically after a call that never returns. + kill_state(Vst); +valfun_1({'%', {type_info, Reg, #t_bs_context{}=Type}}, Vst) -> + %% This is a gross hack, but we'll be rid of it once we have proper union + %% types. + override_type(Type, 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 @@ -490,15 +512,19 @@ valfun_1({line,_}, Vst) -> Vst; %% Exception generating calls valfun_1({call_ext,Live,Func}=I, Vst) -> - case call_return_type(Func, Vst) of - exception -> - verify_live(Live, Vst), - %% The stack will be scanned, so Y registers - %% must be initialized. + case will_call_succeed(Func, Vst) of + yes -> + %% This call cannot fail, handle it here without updating catch + %% state. + call(Func, Live, Vst); + no -> + %% The stack will be scanned, so Y registers must be initialized. + verify_live(Live, Vst), verify_y_init(Vst), - kill_state(Vst); - _ -> - valfun_2(I, Vst) + kill_state(Vst); + maybe -> + %% The call can fail, make sure that any catch state is updated. + valfun_2(I, Vst) end; valfun_1(_I, #vst{current=#st{ct=undecided}}) -> error(unknown_catch_try_state); @@ -530,54 +556,56 @@ valfun_1({'catch',Dst,{f,Fail}}, Vst) when Fail =/= none -> init_try_catch_branch(catchtag, Dst, Fail, Vst); valfun_1({'try',Dst,{f,Fail}}, Vst) when Fail =/= none -> init_try_catch_branch(trytag, Dst, Fail, Vst); -valfun_1({catch_end,Reg}, #vst{current=#st{ct=[Fail|_]}}=Vst0) -> +valfun_1({catch_end,Reg}, #vst{current=#st{ct=[Tag|_]}}=Vst0) -> case get_tag_type(Reg, Vst0) of - {catchtag,Fail} -> + {catchtag,_Fail}=Tag -> %% {x,0} contains the caught term, if any. - create_term(term, catch_end, [], {x,0}, kill_catch_tag(Reg, Vst0)); + create_term(any, catch_end, [], {x,0}, kill_catch_tag(Reg, Vst0)); Type -> error({wrong_tag_type,Type}) end; -valfun_1({try_end,Reg}, #vst{current=#st{ct=[Fail|_]}}=Vst) -> +valfun_1({try_end,Reg}, #vst{current=#st{ct=[Tag|_]}}=Vst) -> case get_tag_type(Reg, Vst) of - {trytag,Fail} -> + {trytag,_Fail}=Tag -> %% Kill the catch tag, note that x registers are unaffected. kill_catch_tag(Reg, Vst); Type -> error({wrong_tag_type,Type}) end; -valfun_1({try_case,Reg}, #vst{current=#st{ct=[Fail|_]}}=Vst0) -> +valfun_1({try_case,Reg}, #vst{current=#st{ct=[Tag|_]}}=Vst0) -> case get_tag_type(Reg, Vst0) of - {trytag,Fail} -> + {trytag,_Fail}=Tag -> %% Kill the catch tag and all x registers. Vst1 = prune_x_regs(0, kill_catch_tag(Reg, Vst0)), %% Class:Error:Stacktrace - Vst2 = create_term({atom,[]}, try_case, [], {x,0}, Vst1), - Vst = create_term(term, try_case, [], {x,1}, Vst2), - create_term(term, try_case, [], {x,2}, Vst); + Vst2 = create_term(#t_atom{}, try_case, [], {x,0}, Vst1), + Vst = create_term(any, try_case, [], {x,1}, Vst2), + create_term(any, try_case, [], {x,2}, Vst); Type -> error({wrong_tag_type,Type}) end; +%% Simple getters that can't fail. valfun_1({get_list,Src,D1,D2}, Vst0) -> assert_not_literal(Src), assert_type(cons, Src, Vst0), - Vst = extract_term(term, get_hd, [Src], D1, Vst0), - extract_term(term, get_tl, [Src], D2, Vst); + Vst = extract_term(any, get_hd, [Src], D1, Vst0), + extract_term(any, get_tl, [Src], D2, Vst); valfun_1({get_hd,Src,Dst}, Vst) -> assert_not_literal(Src), assert_type(cons, Src, Vst), - extract_term(term, get_hd, [Src], Dst, Vst); + extract_term(any, get_hd, [Src], Dst, Vst); valfun_1({get_tl,Src,Dst}, Vst) -> assert_not_literal(Src), assert_type(cons, Src, Vst), - extract_term(term, get_tl, [Src], Dst, Vst); + extract_term(any, get_tl, [Src], Dst, Vst); valfun_1({get_tuple_element,Src,N,Dst}, Vst) -> + Index = N+1, assert_not_literal(Src), - assert_type({tuple_element,N+1}, Src, Vst), - Index = {integer,N+1}, - Type = get_element_type(Index, Src, Vst), - extract_term(Type, {bif,element}, [Index, Src], Dst, Vst); + assert_type(#t_tuple{size=Index}, Src, Vst), + #t_tuple{elements=Es} = normalize(get_term_type(Src, Vst)), + Type = beam_types:get_element_type(Index, Es), + extract_term(Type, {bif,element}, [{integer,Index}, Src], Dst, Vst); valfun_1({jump,{f,Lbl}}, Vst) -> branch(Lbl, Vst, fun(SuccVst) -> @@ -587,38 +615,45 @@ valfun_1({jump,{f,Lbl}}, Vst) -> valfun_1(I, Vst) -> valfun_2(I, Vst). -init_try_catch_branch(Tag, Dst, Fail, Vst0) -> - Vst1 = create_tag({Tag,[Fail]}, 'try_catch', [], Dst, Vst0), - #vst{current=#st{ct=Fails}=St0} = Vst1, - St = St0#st{ct=[[Fail]|Fails]}, - Vst = Vst0#vst{current=St}, +init_try_catch_branch(Kind, Dst, Fail, Vst0) -> + Tag = {Kind, [Fail]}, + Vst = create_tag(Tag, 'try_catch', [], Dst, Vst0), branch(Fail, Vst, - fun(CatchVst) -> - #vst{current=#st{ys=Ys}} = CatchVst, - maps:fold(fun init_catch_handler_1/3, CatchVst, Ys) + fun(CatchVst0) -> + %% We add the tag here because branch/4 rejects jumps to + %% labels referenced by try tags. + #vst{current=#st{ct=Tags,ys=Ys}=St0} = CatchVst0, + St = St0#st{ct=[Tag|Tags]}, + CatchVst = CatchVst0#vst{current=St}, + + maps:fold(fun init_catch_handler_1/3, CatchVst, Ys) end, - fun(SuccVst) -> - %% All potentially-throwing instructions after this - %% one will implicitly branch to the fail label; - %% see valfun_2/2 - SuccVst + fun(SuccVst0) -> + #vst{current=#st{ct=Tags}=St0} = SuccVst0, + St = St0#st{ct=[Tag|Tags]}, + SuccVst = SuccVst0#vst{current=St}, + + %% All potentially-throwing instructions after this one will + %% implicitly branch to the current try/catch handler; see + %% valfun_2/2 + SuccVst end). %% Set the initial state at the try/catch label. Assume that Y registers %% contain terms or try/catch tags. init_catch_handler_1(Reg, initialized, Vst) -> - create_term(term, 'catch_handler', [], Reg, Vst); + create_term(any, 'catch_handler', [], Reg, Vst); init_catch_handler_1(Reg, uninitialized, Vst) -> - create_term(term, 'catch_handler', [], Reg, Vst); + create_term(any, 'catch_handler', [], Reg, Vst); init_catch_handler_1(_, _, Vst) -> Vst. -valfun_2(I, #vst{current=#st{ct=[[Fail]|_]}}=Vst) when is_integer(Fail) -> +valfun_2(I, #vst{current=#st{ct=[{_,[Fail]}|_]}}=Vst) when is_integer(Fail) -> %% We have an active try/catch tag and we can jump there from this %% instruction, so we need to update the branched state of the try/catch %% handler. - valfun_3(I, branch_state(Fail, Vst)); + valfun_3(I, fork_state(Fail, Vst)); valfun_2(I, #vst{current=#st{ct=[]}}=Vst) -> valfun_3(I, Vst); valfun_2(_, _) -> @@ -672,8 +707,16 @@ valfun_4({apply,Live}, Vst) -> valfun_4({apply_last,Live,_}, Vst) -> tail_call(apply, Live+2, Vst); valfun_4({call_fun,Live}, Vst) -> - validate_src([{x,Live}], Vst), - call('fun', Live+1, Vst); + Fun = {x,Live}, + assert_term(Fun, Vst), + + %% An exception is raised on error, hence branching to 0. + branch(0, Vst, + fun(SuccVst0) -> + SuccVst = update_type(fun meet/2, #t_fun{arity=Live}, + Fun, SuccVst0), + call('fun', Live+1, SuccVst) + end); valfun_4({call,Live,Func}, Vst) -> call(Func, Live, Vst); valfun_4({call_ext,Live,Func}, Vst) -> @@ -692,76 +735,28 @@ valfun_4({call_ext_last,Live,Func,StkSize}, tail_call(Func, Live, Vst); valfun_4({call_ext_last,_,_,_}, #vst{current=#st{numy=NumY}}) -> error({allocated,NumY}); -valfun_4({make_fun2,_,_,_,Live}, Vst) -> - call(make_fun, Live, Vst); -%% Other BIFs -valfun_4({bif,element,{f,Fail},[Pos,Src],Dst}, Vst) -> - branch(Fail, Vst, - fun(SuccVst0) -> - PosType = get_term_type(Pos, SuccVst0), - TupleType = {tuple,[get_tuple_size(PosType)],#{}}, +valfun_4({make_fun2,{f,Lbl},_,_,NumFree}, #vst{ft=Ft}=Vst0) -> + #{ arity := Arity0 } = gb_trees:get(Lbl, Ft), + Arity = Arity0 - NumFree, - SuccVst1 = update_type(fun meet/2, TupleType, - Src, SuccVst0), - SuccVst = update_type(fun meet/2, {integer,[]}, - Pos, SuccVst1), + true = Arity >= 0, %Assertion. - ElementType = get_element_type(PosType, Src, SuccVst), - extract_term(ElementType, {bif,element}, [Pos,Src], - Dst, SuccVst) - end); + Vst = prune_x_regs(NumFree, Vst0), + verify_call_args(make_fun, NumFree, Vst), + verify_y_init(Vst), + + create_term(#t_fun{arity=Arity}, make_fun, [], {x,0}, Vst); +%% Other BIFs 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,Op,{f,Fail},[Src]=Ss,Dst}, Vst) when Op =:= hd; Op =:= tl -> - assert_term(Src, Vst), - branch(Fail, Vst, - fun(FailVst) -> - update_type(fun subtract/2, cons, Src, FailVst) - end, - fun(SuccVst0) -> - SuccVst = update_type(fun meet/2, cons, Src, SuccVst0), - extract_term(term, {bif,Op}, Ss, Dst, SuccVst) - end); valfun_4({bif,Op,{f,Fail},Ss,Dst}, Vst) -> validate_src(Ss, Vst), - branch(Fail, Vst, - fun(SuccVst0) -> - %% Infer argument types. Note that we can't subtract - %% types as the BIF could fail for reasons other than - %% bad argument types. - ArgTypes = bif_arg_types(Op, Ss), - SuccVst = foldl(fun({Arg, T}, V) -> - update_type(fun meet/2, T, Arg, V) - end, SuccVst0, zip(Ss, ArgTypes)), - Type = bif_return_type(Op, Ss, SuccVst), - extract_term(Type, {bif,Op}, Ss, Dst, SuccVst) - end); -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), - - %% Heap allocations and X registers are killed regardless of whether we - %% fail or not, as we may fail after GC. - St = kill_heap_allocation(St0), - Vst = prune_x_regs(Live, Vst0#vst{current=St}), - - branch(Fail, Vst, - fun(SuccVst0) -> - ArgTypes = bif_arg_types(Op, Ss), - SuccVst = foldl(fun({Arg, T}, V) -> - update_type(fun meet/2, T, Arg, V) - end, SuccVst0, zip(Ss, ArgTypes)), - - Type = bif_return_type(Op, Ss, SuccVst), - - %% We're passing Vst0 as the original because the - %% registers were pruned before the branch. - extract_term(Type, {gc_bif,Op}, Ss, Dst, SuccVst, Vst0) - end); + validate_bif(Op, Fail, Ss, Dst, Vst); +valfun_4({gc_bif,Op,{f,Fail},Live,Ss,Dst}, Vst) -> + validate_gc_bif(Op, Fail, Ss, Dst, Live, Vst); valfun_4(return, #vst{current=#st{numy=none}}=Vst) -> assert_durable_term({x,0}, Vst), kill_state(Vst); @@ -773,7 +768,7 @@ valfun_4({loop_rec,{f,Fail},Dst}, Vst) -> %% part of this term must be stored in a Y register. branch(Fail, Vst, fun(SuccVst0) -> - {Ref, SuccVst} = new_value(term, loop_rec, [], SuccVst0), + {Ref, SuccVst} = new_value(any, loop_rec, [], SuccVst0), mark_fragile(Dst, set_reg_vref(Ref, Dst, SuccVst)) end); valfun_4({wait,_}, Vst) -> @@ -790,24 +785,13 @@ valfun_4(timeout, Vst) -> prune_x_regs(0, Vst); valfun_4(send, Vst) -> call(send, 2, Vst); -valfun_4({set_tuple_element,Src,Tuple,N}, Vst) -> - I = N + 1, - assert_term(Src, Vst), - assert_type({tuple_element,I}, Tuple, Vst), - %% Manually update the tuple type; we can't rely on the ordinary update - %% helpers as we must support overwriting (rather than just widening or - %% narrowing) known elements, and we can't use extract_term either since - %% the source tuple may be aliased. - {tuple, Sz, Es0} = get_term_type(Tuple, Vst), - Es = set_element_type({integer,I}, get_term_type(Src, Vst), Es0), - override_type({tuple, Sz, Es}, Tuple, Vst); %% Match instructions. valfun_4({select_val,Src,{f,Fail},{list,Choices}}, Vst) -> assert_term(Src, Vst), assert_choices(Choices), validate_select_val(Fail, Choices, Src, Vst); valfun_4({select_tuple_arity,Tuple,{f,Fail},{list,Choices}}, Vst) -> - assert_type(tuple, Tuple, Vst), + assert_type(#t_tuple{}, Tuple, Vst), assert_arities(Choices), validate_select_tuple_arity(Fail, Choices, Tuple, Vst); @@ -817,17 +801,17 @@ valfun_4({test,bs_start_match3,{f,Fail},Live,[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), + assert_type(#t_bs_context{}, Ctx, Vst), branch(Fail, Vst, fun(V) -> V end); valfun_4({test,bs_skip_bits2,{f,Fail},[Ctx,Src,_,_]}, Vst) -> - bsm_validate_context(Ctx, Vst), + assert_type(#t_bs_context{}, Ctx, Vst), assert_term(Src, Vst), branch(Fail, Vst, fun(V) -> V end); valfun_4({test,bs_test_tail2,{f,Fail},[Ctx,_]}, Vst) -> - bsm_validate_context(Ctx, Vst), + assert_type(#t_bs_context{}, Ctx, Vst), branch(Fail, Vst, fun(V) -> V end); valfun_4({test,bs_test_unit,{f,Fail},[Ctx,_]}, Vst) -> - bsm_validate_context(Ctx, Vst), + assert_type(#t_bs_context{}, Ctx, Vst), branch(Fail, Vst, fun(V) -> V end); valfun_4({test,bs_skip_utf8,{f,Fail},[Ctx,Live,_]}, Vst) -> validate_bs_skip_utf(Fail, Ctx, Live, Vst); @@ -835,52 +819,69 @@ valfun_4({test,bs_skip_utf16,{f,Fail},[Ctx,Live,_]}, Vst) -> validate_bs_skip_utf(Fail, Ctx, Live, Vst); valfun_4({test,bs_skip_utf32,{f,Fail},[Ctx,Live,_]}, Vst) -> validate_bs_skip_utf(Fail, Ctx, Live, Vst); -valfun_4({test,bs_get_integer2=Op,{f,Fail},Live,[Ctx,_,_,_],Dst}, Vst) -> - validate_bs_get(Op, Fail, Ctx, Live, {integer, []}, Dst, Vst); +valfun_4({test,bs_get_integer2=Op,{f,Fail},Live, + [Ctx,{integer,Size},Unit,{field_flags,Flags}],Dst},Vst) + when Size * Unit =< 64 -> + Type = case member(unsigned, Flags) of + true -> + NumBits = Size * Unit, + beam_types:make_integer(0, (1 bsl NumBits)-1); + false -> + %% Signed integer or way too large, don't bother. + #t_integer{} + end, + validate_bs_get(Op, Fail, Ctx, Live, Type, Dst, Vst); +valfun_4({test,bs_get_integer2=Op,{f,Fail},Live, + [Ctx,_Size,_Unit,_Flags],Dst},Vst) -> + validate_bs_get(Op, Fail, Ctx, Live, #t_integer{}, Dst, Vst); valfun_4({test,bs_get_float2=Op,{f,Fail},Live,[Ctx,_,_,_],Dst}, Vst) -> - validate_bs_get(Op, Fail, Ctx, Live, {float, []}, Dst, Vst); -valfun_4({test,bs_get_binary2=Op,{f,Fail},Live,[Ctx,_,_,_],Dst}, Vst) -> - validate_bs_get(Op, Fail, Ctx, Live, binary, Dst, Vst); + validate_bs_get(Op, Fail, Ctx, Live, float, Dst, Vst); +valfun_4({test,bs_get_binary2=Op,{f,Fail},Live,[Ctx,_,Unit,_],Dst}, Vst) -> + validate_bs_get(Op, Fail, Ctx, Live, #t_bitstring{unit=Unit}, Dst, Vst); valfun_4({test,bs_get_utf8=Op,{f,Fail},Live,[Ctx,_],Dst}, Vst) -> - validate_bs_get(Op, Fail, Ctx, Live, {integer, []}, Dst, Vst); + Type = beam_types:make_integer(0, ?UNICODE_MAX), + validate_bs_get(Op, Fail, Ctx, Live, Type, Dst, Vst); valfun_4({test,bs_get_utf16=Op,{f,Fail},Live,[Ctx,_],Dst}, Vst) -> - validate_bs_get(Op, Fail, Ctx, Live, {integer, []}, Dst, Vst); + Type = beam_types:make_integer(0, ?UNICODE_MAX), + validate_bs_get(Op, Fail, Ctx, Live, Type, Dst, Vst); valfun_4({test,bs_get_utf32=Op,{f,Fail},Live,[Ctx,_],Dst}, Vst) -> - validate_bs_get(Op, Fail, Ctx, Live, {integer, []}, Dst, Vst); + Type = beam_types:make_integer(0, ?UNICODE_MAX), + validate_bs_get(Op, Fail, Ctx, Live, Type, Dst, Vst); 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), + assert_type(#t_bs_context{}, Ctx, Vst0), verify_live(Live, Vst0), verify_y_init(Vst0), Vst = prune_x_regs(Live, Vst0), - create_term(ms_position, bs_get_position, [Ctx], Dst, Vst, Vst0); + create_term(#t_abstract{kind=ms_position}, bs_get_position, [Ctx], + Dst, Vst, Vst0); valfun_4({bs_set_position, Ctx, Pos}, Vst) -> - bsm_validate_context(Ctx, Vst), - assert_type(ms_position, Pos, Vst), + assert_type(#t_bs_context{}, Ctx, Vst), + assert_type(#t_abstract{kind=ms_position}, Pos, Vst), Vst; %% Other test instructions. valfun_4({test,has_map_fields,{f,Lbl},Src,{list,List}}, Vst) -> - assert_type(map, Src, Vst), + assert_type(#t_map{}, Src, Vst), assert_unique_map_keys(List), branch(Lbl, Vst, fun(V) -> V end); valfun_4({test,is_atom,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, {atom,[]}, Src, Vst); + type_test(Lbl, #t_atom{}, Src, Vst); valfun_4({test,is_binary,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, binary, Src, Vst); + type_test(Lbl, #t_bitstring{unit=8}, Src, Vst); valfun_4({test,is_bitstr,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, binary, Src, Vst); + type_test(Lbl, #t_bitstring{}, Src, Vst); valfun_4({test,is_boolean,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, bool, Src, Vst); + type_test(Lbl, beam_types:make_boolean(), Src, Vst); valfun_4({test,is_float,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, {float,[]}, Src, Vst); + type_test(Lbl, float, Src, Vst); valfun_4({test,is_tuple,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, {tuple,[0],#{}}, Src, Vst); + type_test(Lbl, #t_tuple{}, Src, Vst); valfun_4({test,is_integer,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, {integer,[]}, Src, Vst); + type_test(Lbl, #t_integer{}, Src, Vst); valfun_4({test,is_nonempty_list,{f,Lbl},[Src]}, Vst) -> type_test(Lbl, cons, Src, Vst); valfun_4({test,is_number,{f,Lbl},[Src]}, Vst) -> @@ -888,7 +889,7 @@ valfun_4({test,is_number,{f,Lbl},[Src]}, Vst) -> valfun_4({test,is_list,{f,Lbl},[Src]}, Vst) -> type_test(Lbl, list, Src, Vst); valfun_4({test,is_map,{f,Lbl},[Src]}, Vst) -> - type_test(Lbl, map, Src, Vst); + type_test(Lbl, #t_map{}, Src, Vst); valfun_4({test,is_nil,{f,Lbl},[Src]}, Vst) -> %% is_nil is an exact check against the 'nil' value, and should not be %% treated as a simple type test. @@ -901,12 +902,13 @@ valfun_4({test,is_nil,{f,Lbl},[Src]}, Vst) -> update_eq_types(Src, nil, SuccVst) end); valfun_4({test,test_arity,{f,Lbl},[Tuple,Sz]}, Vst) when is_integer(Sz) -> - assert_type(tuple, Tuple, Vst), - Type = {tuple, Sz, #{}}, + assert_type(#t_tuple{}, Tuple, Vst), + Type = #t_tuple{exact=true,size=Sz}, type_test(Lbl, Type, Tuple, Vst); valfun_4({test,is_tagged_tuple,{f,Lbl},[Src,Sz,Atom]}, Vst) -> assert_term(Src, Vst), - Type = {tuple, Sz, #{ {integer,1} => Atom }}, + Es = #{ 1 => get_literal_type(Atom) }, + Type = #t_tuple{exact=true,size=Sz,elements=Es}, type_test(Lbl, Type, Src, Vst); valfun_4({test,is_eq_exact,{f,Lbl},[Src,Val]=Ss}, Vst) -> validate_src(Ss, Vst), @@ -935,19 +937,19 @@ valfun_4({bs_add,{f,Fail},[A,B,_],Dst}, Vst) -> assert_term(B, Vst), branch(Fail, Vst, fun(SuccVst) -> - create_term({integer,[]}, bs_add, [A, B], Dst, SuccVst) + create_term(#t_integer{}, bs_add, [A, B], Dst, SuccVst) end); valfun_4({bs_utf8_size,{f,Fail},A,Dst}, Vst) -> assert_term(A, Vst), branch(Fail, Vst, fun(SuccVst) -> - create_term({integer,[]}, bs_utf8_size, [A], Dst, SuccVst) + create_term(#t_integer{}, bs_utf8_size, [A], Dst, SuccVst) end); valfun_4({bs_utf16_size,{f,Fail},A,Dst}, Vst) -> assert_term(A, Vst), branch(Fail, Vst, fun(SuccVst) -> - create_term({integer,[]}, bs_utf16_size, [A], Dst, SuccVst) + create_term(#t_integer{}, bs_utf16_size, [A], Dst, SuccVst) end); valfun_4({bs_init2,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> verify_live(Live, Vst0), @@ -962,7 +964,8 @@ valfun_4({bs_init2,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> branch(Fail, Vst, fun(SuccVst0) -> SuccVst = prune_x_regs(Live, SuccVst0), - create_term(binary, bs_init2, [], Dst, SuccVst, SuccVst0) + create_term(#t_bitstring{unit=8}, bs_init2, [], Dst, + SuccVst, SuccVst0) end); valfun_4({bs_init_bits,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> verify_live(Live, Vst0), @@ -977,9 +980,9 @@ valfun_4({bs_init_bits,{f,Fail},Sz,Heap,Live,_,Dst}, Vst0) -> branch(Fail, Vst, fun(SuccVst0) -> SuccVst = prune_x_regs(Live, SuccVst0), - create_term(binary, bs_init_bits, [], Dst, SuccVst) + create_term(#t_bitstring{}, bs_init_bits, [], Dst, SuccVst) end); -valfun_4({bs_append,{f,Fail},Bits,Heap,Live,_Unit,Bin,_Flags,Dst}, Vst0) -> +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), @@ -988,14 +991,16 @@ valfun_4({bs_append,{f,Fail},Bits,Heap,Live,_Unit,Bin,_Flags,Dst}, Vst0) -> branch(Fail, Vst, fun(SuccVst0) -> SuccVst = prune_x_regs(Live, SuccVst0), - create_term(binary, bs_append, [Bin], Dst, SuccVst, SuccVst0) + create_term(#t_bitstring{unit=Unit}, bs_append, + [Bin], Dst, SuccVst, SuccVst0) end); -valfun_4({bs_private_append,{f,Fail},Bits,_Unit,Bin,_Flags,Dst}, Vst) -> +valfun_4({bs_private_append,{f,Fail},Bits,Unit,Bin,_Flags,Dst}, Vst) -> assert_term(Bits, Vst), assert_term(Bin, Vst), branch(Fail, Vst, fun(SuccVst) -> - create_term(binary, bs_private_append, [Bin], Dst, SuccVst) + create_term(#t_bitstring{unit=Unit}, bs_private_append, + [Bin], Dst, SuccVst) end); valfun_4({bs_put_string,Sz,_}, Vst) when is_integer(Sz) -> Vst; @@ -1004,39 +1009,39 @@ valfun_4({bs_put_binary,{f,Fail},Sz,_,_,Src}, Vst) -> assert_term(Src, Vst), branch(Fail, Vst, fun(SuccVst) -> - update_type(fun meet/2, binary, Src, SuccVst) + update_type(fun meet/2, #t_bitstring{}, Src, SuccVst) end); valfun_4({bs_put_float,{f,Fail},Sz,_,_,Src}, Vst) -> assert_term(Sz, Vst), assert_term(Src, Vst), branch(Fail, Vst, fun(SuccVst) -> - update_type(fun meet/2, {float,[]}, Src, SuccVst) + update_type(fun meet/2, float, Src, SuccVst) end); valfun_4({bs_put_integer,{f,Fail},Sz,_,_,Src}, Vst) -> assert_term(Sz, Vst), assert_term(Src, Vst), branch(Fail, Vst, fun(SuccVst) -> - update_type(fun meet/2, {integer,[]}, Src, SuccVst) + update_type(fun meet/2, #t_integer{}, Src, SuccVst) end); valfun_4({bs_put_utf8,{f,Fail},_,Src}, Vst) -> assert_term(Src, Vst), branch(Fail, Vst, fun(SuccVst) -> - update_type(fun meet/2, {integer,[]}, Src, SuccVst) + update_type(fun meet/2, #t_integer{}, Src, SuccVst) end); valfun_4({bs_put_utf16,{f,Fail},_,Src}, Vst) -> assert_term(Src, Vst), branch(Fail, Vst, fun(SuccVst) -> - update_type(fun meet/2, {integer,[]}, Src, SuccVst) + update_type(fun meet/2, #t_integer{}, Src, SuccVst) end); valfun_4({bs_put_utf32,{f,Fail},_,Src}, Vst) -> assert_term(Src, Vst), branch(Fail, Vst, fun(SuccVst) -> - update_type(fun meet/2, {integer,[]}, Src, SuccVst) + update_type(fun meet/2, #t_integer{}, Src, SuccVst) end); %% Map instructions. valfun_4({put_map_assoc=Op,{f,Fail},Src,Dst,Live,{list,List}}, Vst) -> @@ -1050,7 +1055,7 @@ valfun_4(_, _) -> verify_get_map(Fail, Src, List, Vst0) -> assert_not_literal(Src), %OTP 22. - assert_type(map, Src, Vst0), + assert_type(#t_map{}, Src, Vst0), branch(Fail, Vst0, fun(FailVst) -> @@ -1074,7 +1079,7 @@ verify_get_map(Fail, Src, List, Vst0) -> clobber_map_vals([Key,Dst|T], Map, Vst0) -> case is_reg_initialized(Dst, Vst0) of true -> - Vst = extract_term(term, {bif,map_get}, [Key, Map], Dst, Vst0), + Vst = extract_term(any, {bif,map_get}, [Key, Map], Dst, Vst0), clobber_map_vals(T, Map, Vst); false -> clobber_map_vals(T, Map, Vst0) @@ -1099,13 +1104,13 @@ extract_map_keys([]) -> []. extract_map_vals([Key,Dst|Vs], Map, Vst0, Vsti0) -> assert_term(Key, Vst0), - Vsti = extract_term(term, {bif,map_get}, [Key, Map], Dst, Vsti0), + Vsti = extract_term(any, {bif,map_get}, [Key, Map], Dst, Vsti0), extract_map_vals(Vs, Map, Vst0, Vsti); extract_map_vals([], _Map, _Vst0, Vst) -> Vst. verify_put_map(Op, Fail, Src, Dst, Live, List, Vst0) -> - assert_type(map, Src, Vst0), + assert_type(#t_map{}, Src, Vst0), verify_live(Live, Vst0), verify_y_init(Vst0), _ = [assert_term(Term, Vst0) || Term <- List], @@ -1116,10 +1121,73 @@ verify_put_map(Op, Fail, Src, Dst, Live, List, Vst0) -> SuccVst = prune_x_regs(Live, SuccVst0), Keys = extract_map_keys(List), assert_unique_map_keys(Keys), - create_term(map, Op, [Src], Dst, SuccVst, SuccVst0) + create_term(#t_map{}, Op, [Src], Dst, SuccVst, SuccVst0) end). %% +%% Common code for validating BIFs. +%% +%% OrigVst is the state we entered the instruction with, which is needed for +%% gc_bifs as X registers are pruned prior to calling this function, which may +%% have clobbered the sources. +%% + +validate_bif(Op, Fail, Ss, Dst, Vst) -> + validate_src(Ss, Vst), + validate_bif_1(bif, Op, Fail, Ss, Dst, Vst, Vst). + +validate_gc_bif(Op, Fail, Ss, Dst, Live, #vst{current=St0}=Vst0) -> + validate_src(Ss, Vst0), + verify_live(Live, Vst0), + verify_y_init(Vst0), + + %% Heap allocations and X registers are killed regardless of whether we + %% fail or not, as we may fail after GC. + St = kill_heap_allocation(St0), + Vst = prune_x_regs(Live, Vst0#vst{current=St}), + validate_src(Ss, Vst), + + validate_bif_1(gc_bif, Op, Fail, Ss, Dst, Vst, Vst). + +validate_bif_1(Kind, Op, cannot_fail, Ss, Dst, OrigVst, Vst0) -> + %% This BIF explicitly cannot fail; it will not jump to a guard nor throw + %% an exception. Validation will fail if it returns 'none' or has a type + %% conflict on one of its arguments. + + {Type, ArgTypes, _CanSubtract} = bif_types(Op, Ss, Vst0), + ZippedArgs = zip(Ss, ArgTypes), + + Vst = foldl(fun({A, T}, V) -> + update_type(fun meet/2, T, A, V) + end, Vst0, ZippedArgs), + + true = Type =/= none, %Assertion. + + extract_term(Type, {Kind, Op}, Ss, Dst, Vst, OrigVst); +validate_bif_1(Kind, Op, Fail, Ss, Dst, OrigVst, Vst) -> + {Type, ArgTypes, CanSubtract} = bif_types(Op, Ss, Vst), + ZippedArgs = zip(Ss, ArgTypes), + + FailFun = case CanSubtract of + true -> + fun(FailVst0) -> + foldl(fun({A, T}, V) -> + update_type(fun subtract/2, T, A, V) + end, FailVst0, ZippedArgs) + end; + false -> + fun(S) -> S end + end, + SuccFun = fun(SuccVst0) -> + SuccVst = foldl(fun({A, T}, V) -> + update_type(fun meet/2, T, A, V) + end, SuccVst0, ZippedArgs), + extract_term(Type, {Kind, Op}, Ss, Dst, SuccVst, OrigVst) + end, + + branch(Fail, Vst, FailFun, SuccFun). + +%% %% Common code for validating bs_start_match* instructions. %% @@ -1127,18 +1195,18 @@ validate_bs_start_match(Fail, Live, Type, Src, Dst, Vst) -> verify_live(Live, Vst), verify_y_init(Vst), - %% #ms{} can represent either a match context or a term, so we have to mark - %% the source as a term if it fails with a match context as an input. This - %% hack is only needed until we get proper union types. + %% #t_bs_context{} can represent either a match context or a term, so we + %% have to mark the source as a term if it fails with a match context as an + %% input. This hack is only needed until we get proper union types. branch(Fail, Vst, fun(FailVst) -> case get_movable_term_type(Src, FailVst) of - #ms{} -> override_type(term, Src, FailVst); + #t_bs_context{} -> override_type(any, Src, FailVst); _ -> FailVst end end, fun(SuccVst0) -> - SuccVst1 = update_type(fun meet/2, binary, + SuccVst1 = update_type(fun meet/2, #t_bitstring{}, Src, SuccVst0), SuccVst = prune_x_regs(Live, SuccVst1), extract_term(Type, bs_start_match, [Src], Dst, @@ -1149,7 +1217,7 @@ validate_bs_start_match(Fail, Live, Type, Src, Dst, Vst) -> %% Common code for validating bs_get* instructions. %% validate_bs_get(Op, Fail, Ctx, Live, Type, Dst, Vst) -> - bsm_validate_context(Ctx, Vst), + assert_type(#t_bs_context{}, Ctx, Vst), verify_live(Live, Vst), verify_y_init(Vst), @@ -1163,7 +1231,7 @@ validate_bs_get(Op, Fail, Ctx, Live, Type, Dst, Vst) -> %% Common code for validating bs_skip_utf* instructions. %% validate_bs_skip_utf(Fail, Ctx, Live, Vst) -> - bsm_validate_context(Ctx, Vst), + assert_type(#t_bs_context{}, Ctx, Vst), verify_y_init(Vst), verify_live(Live, Vst), @@ -1217,12 +1285,13 @@ kill_state(Vst) -> call(Name, Live, #vst{current=St0}=Vst0) -> verify_call_args(Name, Live, Vst0), verify_y_init(Vst0), - case call_return_type(Name, Vst0) of - Type when Type =/= exception -> - %% Type is never 'exception' because it has been handled earlier. + case call_types(Name, Live, Vst0) of + {none, _, _} -> + kill_state(Vst0); + {RetType, _, _} -> St = St0#st{f=init_fregs()}, Vst = prune_x_regs(0, Vst0#vst{current=St}), - create_term(Type, call, [], {x,0}, Vst) + create_term(RetType, call, [], {x,0}, Vst) end. %% Tail call. @@ -1237,8 +1306,15 @@ tail_call(Name, Live, Vst0) -> verify_call_args(_, 0, #vst{}) -> ok; -verify_call_args({f,Lbl}, Live, Vst) when is_integer(Live)-> - verify_local_args(Live - 1, Lbl, #{}, Vst); +verify_call_args({f,Lbl}, Live, #vst{ft=Ft}=Vst) when is_integer(Live) -> + case gb_trees:lookup(Lbl, Ft) of + {value, FuncInfo} -> + #{ arity := Live, + parameter_types := ParamTypes } = FuncInfo, + verify_local_args(Live - 1, ParamTypes, #{}, Vst); + none -> + error(local_call_to_unknown_function) + end; verify_call_args(_, Live, Vst) when is_integer(Live)-> verify_remote_args_1(Live - 1, Vst); verify_call_args(_, Live, _) -> @@ -1250,87 +1326,50 @@ verify_remote_args_1(X, Vst) -> assert_durable_term({x, X}, Vst), verify_remote_args_1(X - 1, Vst). -verify_local_args(-1, _Lbl, _CtxIds, _Vst) -> +verify_local_args(-1, _ParamTypes, _CtxIds, _Vst) -> ok; -verify_local_args(X, Lbl, CtxIds, Vst) -> +verify_local_args(X, ParamTypes, CtxRefs, Vst) -> Reg = {x, X}, assert_not_fragile(Reg, Vst), case get_movable_term_type(Reg, Vst) of - #ms{id=Id}=Type -> - case CtxIds of - #{ Id := Other } -> + #t_bs_context{}=Type -> + VRef = get_reg_vref(Reg, Vst), + case CtxRefs of + #{ VRef := Other } -> error({multiple_match_contexts, [Reg, Other]}); #{} -> - verify_arg_type(Lbl, Reg, Type, Vst), - verify_local_args(X - 1, Lbl, CtxIds#{ Id => Reg }, Vst) + verify_arg_type(Reg, Type, ParamTypes), + verify_local_args(X - 1, ParamTypes, + CtxRefs#{ VRef => Reg }, Vst) end; Type -> - verify_arg_type(Lbl, Reg, Type, Vst), - verify_local_args(X - 1, Lbl, CtxIds, Vst) + verify_arg_type(Reg, Type, ParamTypes), + verify_local_args(X - 1, ParamTypes, CtxRefs, Vst) end. %% Verifies that the given argument narrows to what the function expects. -verify_arg_type(Lbl, Reg, #ms{}, #vst{ft=Ft}) -> +verify_arg_type(Reg, #t_bs_context{}, ParamTypes) -> %% 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) + case ParamTypes of + #{ Reg := #t_bs_context{}} -> 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, #ms{}} -> +verify_arg_type(Reg, GivenType, ParamTypes) -> + case ParamTypes of + #{ Reg := #t_bs_context{}} -> %% Functions that accept match contexts also accept all other %% terms. This will change once we support union types. ok; - {value, RequiredType} -> - case vat_1(GivenType, RequiredType) of - true -> ok; - false -> error({bad_arg_type, Reg, GivenType, RequiredType}) + #{ Reg := RequiredType } -> + case meet(GivenType, RequiredType) of + GivenType -> ok; + _ -> error({bad_arg_type, Reg, GivenType, RequiredType}) end; - none -> + #{} -> ok end. -%% Checks whether the Given argument is compatible with the Required one. This -%% is essentially a relaxed version of 'meet(Given, Req) =:= Given', where we -%% accept that the Given value has the right type but not necessarily the exact -%% same value; if {atom,gurka} is required, we'll consider {atom,[]} valid. -%% -%% This will catch all problems that could crash the emulator, like passing a -%% 1-tuple when the callee expects a 3-tuple, but some value errors might slip -%% through. -vat_1(Same, Same) -> true; -vat_1({atom,A}, {atom,B}) -> A =:= B orelse is_list(A) orelse is_list(B); -vat_1({atom,A}, bool) -> is_boolean(A) orelse is_list(A); -vat_1(bool, {atom,B}) -> is_boolean(B) orelse is_list(B); -vat_1(cons, list) -> true; -vat_1({float,A}, {float,B}) -> A =:= B orelse is_list(A) orelse is_list(B); -vat_1({float,_}, number) -> true; -vat_1({integer,A}, {integer,B}) -> A =:= B orelse is_list(A) orelse is_list(B); -vat_1({integer,_}, number) -> true; -vat_1(_, {literal,_}) -> false; -vat_1({literal,_}=Lit, Required) -> vat_1(get_literal_type(Lit), Required); -vat_1(nil, list) -> true; -vat_1({tuple,SzA,EsA}, {tuple,SzB,EsB}) -> - if - is_list(SzB) -> - tuple_sz(SzA) >= tuple_sz(SzB) andalso vat_elements(EsA, EsB); - SzA =:= SzB -> - vat_elements(EsA, EsB); - SzA =/= SzB -> - false - end; -vat_1(_, _) -> false. - -vat_elements(EsA, EsB) -> - maps:fold(fun(Key, Req, Acc) -> - case EsA of - #{ Key := Given } -> Acc andalso vat_1(Given, Req); - #{} -> false - end - end, true, EsB). - allocate(Tag, Stk, Heap, Live, #vst{current=#st{numy=none}=St}=Vst0) -> verify_live(Live, Vst0), Vst1 = Vst0#vst{current=St#st{numy=Stk}}, @@ -1512,48 +1551,39 @@ assert_unique_map_keys([_,_|_]=Ls) -> %%% bsm_match_state() -> - #ms{}. + #t_bs_context{}. bsm_match_state(Slots) -> - #ms{slots=Slots}. - -bsm_validate_context(Reg, Vst) -> - _ = bsm_get_context(Reg, Vst), - ok. - -bsm_get_context({Kind,_}=Reg, Vst) when Kind =:= x; Kind =:= y-> - case get_movable_term_type(Reg, Vst) of - #ms{}=Ctx -> Ctx; - _ -> error({no_bsm_context,Reg}) - end; -bsm_get_context(Reg, _) -> - error({bad_source,Reg}). + #t_bs_context{slots=Slots}. bsm_save(Reg, {atom,start}, Vst) -> %% Save point refering to where the match started. %% It is always valid. But don't forget to validate the context register. - bsm_validate_context(Reg, Vst), + assert_type(#t_bs_context{}, Reg, Vst), Vst; bsm_save(Reg, SavePoint, Vst) -> - case bsm_get_context(Reg, Vst) of - #ms{valid=Bits,slots=Slots}=Ctxt0 when SavePoint < Slots -> - Ctx = Ctxt0#ms{valid=Bits bor (1 bsl SavePoint),slots=Slots}, - override_type(Ctx, Reg, Vst); - _ -> error({illegal_save,SavePoint}) + case get_movable_term_type(Reg, Vst) of + #t_bs_context{valid=Bits,slots=Slots}=Ctxt0 when SavePoint < Slots -> + Ctx = Ctxt0#t_bs_context{valid=Bits bor (1 bsl SavePoint), + slots=Slots}, + override_type(Ctx, Reg, Vst); + _ -> + error({illegal_save, SavePoint}) end. bsm_restore(Reg, {atom,start}, Vst) -> %% (Mostly) automatic save point refering to where the match started. %% It is always valid. But don't forget to validate the context register. - bsm_validate_context(Reg, Vst), + assert_type(#t_bs_context{}, Reg, Vst), Vst; bsm_restore(Reg, SavePoint, Vst) -> - case bsm_get_context(Reg, Vst) of - #ms{valid=Bits,slots=Slots} when SavePoint < Slots -> - case Bits band (1 bsl SavePoint) of - 0 -> error({illegal_restore,SavePoint,not_set}); - _ -> Vst - end; - _ -> error({illegal_restore,SavePoint,range}) + case get_movable_term_type(Reg, Vst) of + #t_bs_context{valid=Bits,slots=Slots} when SavePoint < Slots -> + case Bits band (1 bsl SavePoint) of + 0 -> error({illegal_restore, SavePoint, not_set}); + _ -> Vst + end; + _ -> + error({illegal_restore, SavePoint, range}) end. validate_select_val(_Fail, _Choices, _Src, #vst{current=none}=Vst) -> @@ -1569,7 +1599,7 @@ validate_select_val(Fail, [Val,{f,L}|T], Src, Vst0) -> update_ne_types(Src, Val, FailVst) end), validate_select_val(Fail, T, Src, Vst); -validate_select_val(Fail, [], _, Vst) -> +validate_select_val(Fail, [], _Src, Vst) -> branch(Fail, Vst, fun(SuccVst) -> %% The next instruction is never executed. @@ -1581,7 +1611,7 @@ validate_select_tuple_arity(_Fail, _Choices, _Src, #vst{current=none}=Vst) -> %% can't reach the fail label or any of the remaining choices. Vst; validate_select_tuple_arity(Fail, [Arity,{f,L}|T], Tuple, Vst0) -> - Type = {tuple, Arity, #{}}, + Type = #t_tuple{exact=true,size=Arity}, Vst = branch(L, Vst0, fun(BranchVst) -> update_type(fun meet/2, Type, Tuple, BranchVst) @@ -1597,63 +1627,94 @@ validate_select_tuple_arity(Fail, [], _, #vst{}=Vst) -> kill_state(SuccVst) end). -infer_types({Kind,_}=Reg, Vst) when Kind =:= x; Kind =:= y -> - infer_types(get_reg_vref(Reg, Vst), Vst); -infer_types(#value_ref{}=Ref, #vst{current=#st{vs=Vs}}) -> +%% +%% Infers types from comparisons, looking at the expressions that produced the +%% compared values and updates their types if we've learned something new from +%% the comparison. +%% + +infer_types(CompareOp, {Kind,_}=LHS, RHS, Vst) when Kind =:= x; Kind =:= y -> + infer_types(CompareOp, get_reg_vref(LHS, Vst), RHS, Vst); +infer_types(CompareOp, LHS, {Kind,_}=RHS, Vst) when Kind =:= x; Kind =:= y -> + infer_types(CompareOp, LHS, get_reg_vref(RHS, Vst), Vst); +infer_types(CompareOp, LHS, RHS, #vst{current=#st{vs=Vs}}=Vst0) -> case Vs of - #{ Ref := Entry } -> infer_types_1(Entry); - #{} -> fun(_, S) -> S end + #{ LHS := LEntry, RHS := REntry } -> + Vst = infer_types_1(LEntry, RHS, CompareOp, Vst0), + infer_types_1(REntry, LHS, CompareOp, Vst); + #{ LHS := LEntry } -> + infer_types_1(LEntry, RHS, CompareOp, Vst0); + #{ RHS := REntry } -> + infer_types_1(REntry, LHS, CompareOp, Vst0); + #{} -> + Vst0 + end. + +infer_types_1(#value{op={bif,'=:='},args=[LHS,RHS]}, Val, Op, Vst) -> + case Val of + {atom, Bool} when Op =:= eq_exact, Bool; Op =:= ne_exact, not Bool -> + update_eq_types(LHS, RHS, Vst); + {atom, Bool} when Op =:= ne_exact, Bool; Op =:= eq_exact, not Bool -> + update_ne_types(LHS, RHS, Vst); + _ -> + Vst end; -infer_types(_, #vst{}) -> - fun(_, S) -> S end. - -infer_types_1(#value{op={bif,'=:='},args=[LHS,RHS]}) -> - fun({atom,true}, S) -> - %% Either side might contain something worth inferring, so we need - %% to check them both. - Infer_L = infer_types(RHS, S), - Infer_R = infer_types(LHS, S), - Infer_R(RHS, Infer_L(LHS, S)); - (_, S) -> S +infer_types_1(#value{op={bif,'=/='},args=[LHS,RHS]}, Val, Op, Vst) -> + case Val of + {atom, Bool} when Op =:= ne_exact, Bool; Op =:= eq_exact, not Bool -> + update_ne_types(LHS, RHS, Vst); + {atom, Bool} when Op =:= eq_exact, Bool; Op =:= ne_exact, not Bool -> + update_eq_types(LHS, RHS, Vst); + _ -> + Vst end; -infer_types_1(#value{op={bif,element},args=[{integer,Index}=Key,Tuple]}) -> - fun(Val, S) -> - Type = {tuple,[Index], #{ Key => get_term_type(Val, S) }}, - update_type(fun meet/2, Type, Tuple, S) +infer_types_1(#value{op={bif,element},args=[{integer,Index},Tuple]}, + Val, Op, Vst) when Index >= 1 -> + ElementType = get_term_type(Val, Vst), + Es = beam_types:set_element_type(Index, ElementType, #{}), + Type = #t_tuple{size=Index,elements=Es}, + case Op of + eq_exact -> update_type(fun meet/2, Type, Tuple, Vst); + ne_exact -> update_type(fun subtract/2, Type, Tuple, Vst) end; -infer_types_1(#value{op={bif,is_atom},args=[Src]}) -> - infer_type_test_bif({atom,[]}, Src); -infer_types_1(#value{op={bif,is_boolean},args=[Src]}) -> - infer_type_test_bif(bool, Src); -infer_types_1(#value{op={bif,is_binary},args=[Src]}) -> - infer_type_test_bif(binary, Src); -infer_types_1(#value{op={bif,is_bitstring},args=[Src]}) -> - infer_type_test_bif(binary, Src); -infer_types_1(#value{op={bif,is_float},args=[Src]}) -> - infer_type_test_bif(float, Src); -infer_types_1(#value{op={bif,is_integer},args=[Src]}) -> - infer_type_test_bif({integer,{}}, Src); -infer_types_1(#value{op={bif,is_list},args=[Src]}) -> - infer_type_test_bif(list, Src); -infer_types_1(#value{op={bif,is_map},args=[Src]}) -> - infer_type_test_bif(map, Src); -infer_types_1(#value{op={bif,is_number},args=[Src]}) -> - infer_type_test_bif(number, Src); -infer_types_1(#value{op={bif,is_tuple},args=[Src]}) -> - infer_type_test_bif({tuple,[0],#{}}, Src); -infer_types_1(#value{op={bif,tuple_size}, args=[Tuple]}) -> - fun({integer,Arity}, S) -> - update_type(fun meet/2, {tuple,Arity,#{}}, Tuple, S); - (_, S) -> S +infer_types_1(#value{op={bif,is_atom},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(#t_atom{}, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_boolean},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(beam_types:make_boolean(), Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_binary},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(#t_bitstring{unit=8}, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_bitstring},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(#t_bitstring{}, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_float},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(float, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_integer},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(#t_integer{}, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_list},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(list, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_map},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(#t_map{}, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_number},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(number, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,is_tuple},args=[Src]}, Val, Op, Vst) -> + infer_type_test_bif(#t_tuple{}, Src, Val, Op, Vst); +infer_types_1(#value{op={bif,tuple_size}, args=[Tuple]}, + {integer,Arity}, Op, Vst) -> + Type = #t_tuple{exact=true,size=Arity}, + case Op of + eq_exact -> update_type(fun meet/2, Type, Tuple, Vst); + ne_exact -> update_type(fun subtract/2, Type, Tuple, Vst) end; -infer_types_1(_) -> - fun(_, S) -> S end. - -infer_type_test_bif(Type, Src) -> - fun({atom,true}, S) -> - update_type(fun meet/2, Type, Src, S); - (_, S) -> - S +infer_types_1(_, _, _, Vst) -> + Vst. + +infer_type_test_bif(Type, Src, Val, Op, Vst) -> + case Val of + {atom, Bool} when Op =:= eq_exact, Bool; Op =:= ne_exact, not Bool -> + update_type(fun meet/2, Type, Src, Vst); + {atom, Bool} when Op =:= ne_exact, Bool; Op =:= eq_exact, not Bool -> + update_type(fun subtract/2, Type, Src, Vst); + _ -> + Vst end. %%% @@ -1764,43 +1825,58 @@ update_type(Merge, With, #value_ref{}=Ref, Vst) -> update_type(Merge, With, {Kind,_}=Reg, Vst) when Kind =:= x; Kind =:= y -> update_type(Merge, With, get_reg_vref(Reg, Vst), Vst); update_type(Merge, With, Literal, Vst) -> - assert_literal(Literal), %% Literals always retain their type, but we still need to bail on type %% conflicts. - case Merge(Literal, With) of - none -> throw({type_conflict, Literal, With}); + Type = get_literal_type(Literal), + case Merge(Type, With) of + none -> throw({type_conflict, Type, With}); _Type -> Vst end. -update_ne_types(LHS, RHS, Vst) -> +update_eq_types(LHS, RHS, Vst0) -> + LType = get_term_type(LHS, Vst0), + RType = get_term_type(RHS, Vst0), + + Vst1 = update_type(fun meet/2, RType, LHS, Vst0), + Vst = update_type(fun meet/2, LType, RHS, Vst1), + + infer_types(eq_exact, LHS, RHS, Vst). + +update_ne_types(LHS, RHS, Vst0) -> + Vst1 = update_ne_types_1(LHS, RHS, Vst0), + Vst = update_ne_types_1(RHS, LHS, Vst1), + + infer_types(ne_exact, LHS, RHS, Vst). + +update_ne_types_1(LHS, RHS, Vst0) -> %% While updating types on equality is fairly straightforward, inequality %% is a bit trickier since all we know is that the *value* of LHS differs %% from RHS, so we can't blindly subtract their types. %% - %% Consider `number =/= {integer,[]}`; all we know is that LHS isn't equal + %% Consider `number =/= #t_integer{}`; all we know is that LHS isn't equal %% to some *specific integer* of unknown value, and if we were to subtract - %% {integer,[]} we would erroneously infer that the new type is {float,[]}. + %% #t_integer{} we would erroneously infer that the new type is float. %% %% Therefore, we only subtract when we know that RHS has a specific value. - RType = get_term_type(RHS, Vst), - case is_literal(RType) of - true -> update_type(fun subtract/2, RType, LHS, Vst); - false -> Vst + RType = get_term_type(RHS, Vst0), + case beam_types:is_singleton_type(RType) of + true -> + Vst = update_type(fun subtract/2, RType, LHS, Vst0), + + %% If LHS has a specific value after subtraction we can infer types + %% as if we've made an exact match, which is much stronger than + %% ne_exact. + LType = get_term_type(LHS, Vst), + case beam_types:get_singleton_value(LType) of + {ok, Value} -> + infer_types(eq_exact, LHS, value_to_literal(Value), Vst); + error -> + Vst + end; + false -> + Vst0 end. -update_eq_types(LHS, RHS, Vst0) -> - %% Either side might contain something worth inferring, so we need - %% to check them both. - Infer_L = infer_types(RHS, Vst0), - Infer_R = infer_types(LHS, Vst0), - Vst1 = Infer_R(RHS, Infer_L(LHS, Vst0)), - - T1 = get_term_type(LHS, Vst1), - T2 = get_term_type(RHS, Vst1), - - Vst = update_type(fun meet/2, T2, LHS, Vst1), - update_type(fun meet/2, T1, RHS, Vst). - %% Helper functions for the above. assign_1(Src, Dst, Vst0) -> @@ -1851,16 +1927,9 @@ get_reg_vref({y,_}=Src, #vst{current=#st{ys=Ys}}) -> end. set_type(Type, #value_ref{}=Ref, #vst{current=#st{vs=Vs0}=St}=Vst) -> - case Vs0 of - #{ Ref := #value{}=Entry } -> - Vs = Vs0#{ Ref => Entry#value{type=Type} }, - Vst#vst{current=St#st{vs=Vs}}; - #{} -> - %% Dead references may happen during type inference and are not an - %% error in and of themselves. If a problem were to arise from this - %% it'll explode elsewhere. - Vst - end. + #{ Ref := #value{}=Entry } = Vs0, + Vs = Vs0#{ Ref => Entry#value{type=Type} }, + Vst#vst{current=St#st{vs=Vs}}. new_value(Type, Op, Ss, #vst{current=#st{vs=Vs0}=St,ref_ctr=Counter}=Vst) -> Ref = #value_ref{id=Counter}, @@ -1868,9 +1937,9 @@ new_value(Type, Op, Ss, #vst{current=#st{vs=Vs0}=St,ref_ctr=Counter}=Vst) -> {Ref, Vst#vst{current=St#st{vs=Vs},ref_ctr=Counter+1}}. -kill_catch_tag(Reg, #vst{current=#st{ct=[Fail|Fails]}=St}=Vst0) -> - Vst = Vst0#vst{current=St#st{ct=Fails,fls=undefined}}, - {_, Fail} = get_tag_type(Reg, Vst), %Assertion. +kill_catch_tag(Reg, #vst{current=#st{ct=[Tag|Tags]}=St}=Vst0) -> + Vst = Vst0#vst{current=St#st{ct=Tags,fls=undefined}}, + Tag = get_tag_type(Reg, Vst), %Assertion. kill_tag(Reg, Vst). check_try_catch_tags(Type, {y,N}=Reg, Vst) -> @@ -1915,308 +1984,45 @@ is_literal({integer,I}) when is_integer(I) -> true; is_literal({literal,_L}) -> true; is_literal(_) -> false. -%% The possible types. -%% -%% First non-term types: -%% -%% initialized Only for Y registers. Means that the Y register -%% has been initialized with some valid term so that -%% it is safe to pass to the garbage collector. -%% NOT safe to use in any other way (will not crash the -%% emulator, but clearly points to a bug in the compiler). -%% -%% {catchtag,[Lbl]} A special term used within a catch. Must only be used -%% by the catch instructions; NOT safe to use in other -%% instructions. -%% -%% {trytag,[Lbl]} A special term used within a try block. Must only be -%% used by the catch instructions; NOT safe to use in other -%% instructions. -%% -%% exception Can only be used as a type returned by -%% call_return_type/2 (which gives the type of the value -%% returned by a call). Thus 'exception' is never stored -%% as type descriptor for a register. -%% -%% #ms{} A match context for bit syntax matching. We do allow -%% it to moved/to from stack, but otherwise it must only -%% be accessed by bit syntax matching instructions. -%% -%% -%% Normal terms: -%% -%% 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],Es} 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. Es is a map -%% containing known types by tuple index. -%% -%% {tuple,Sz,Es} Tuple. A test_arity instruction has been seen -%% so that it is known that the size is exactly Sz. -%% -%% {atom,[]} Atom. -%% {atom,Atom} -%% -%% {integer,[]} Integer. -%% {integer,Integer} -%% -%% {float,[]} Float. -%% {float,Float} -%% -%% number Integer or Float of unknown value -%% -%% map Map. -%% -%% none A conflict in types. There will be an exception at runtime. -%% - -%% join(Type1, Type2) -> Type -%% Return the most specific type possible. -join(Same, Same) -> - Same; -join(none, Other) -> - Other; -join(Other, none) -> - Other; -join({literal,_}=T1, T2) -> - join_literal(T1, T2); -join(T1, {literal,_}=T2) -> - join_literal(T2, T1); -join({tuple,Size,EsA}, {tuple,Size,EsB}) -> - Es = join_tuple_elements(tuple_sz(Size), EsA, EsB), - {tuple, Size, Es}; -join({tuple,A,EsA}, {tuple,B,EsB}) -> - Size = min(tuple_sz(A), tuple_sz(B)), - Es = join_tuple_elements(Size, EsA, EsB), - {tuple, [Size], Es}; -join({Type,A}, {Type,B}) - when Type =:= atom; Type =:= integer; Type =:= float -> - if A =:= B -> {Type,A}; - true -> {Type,[]} - end; -join({Type,_}, number) - when Type =:= integer; Type =:= float -> - number; -join(number, {Type,_}) - when Type =:= integer; Type =:= float -> - number; -join({integer,_}, {float,_}) -> - number; -join({float,_}, {integer,_}) -> - number; -join(bool, {atom,A}) -> - join_bool(A); -join({atom,A}, bool) -> - join_bool(A); -join({atom,A}, {atom,B}) when is_boolean(A), is_boolean(B) -> - bool; -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)}; -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). - -join_tuple_elements(Limit, EsA, EsB) -> - Es0 = join_elements(EsA, EsB), - maps:filter(fun({integer,Index}, _Type) -> Index =< Limit end, Es0). - -join_elements(Es1, Es2) -> - Keys = if - map_size(Es1) =< map_size(Es2) -> maps:keys(Es1); - map_size(Es1) > map_size(Es2) -> maps:keys(Es2) - end, - join_elements_1(Keys, Es1, Es2, #{}). - -join_elements_1([Key | Keys], Es1, Es2, Acc0) -> - Type = case {Es1, Es2} of - {#{ Key := Same }, #{ Key := Same }} -> Same; - {#{ Key := Type1 }, #{ Key := Type2 }} -> join(Type1, Type2); - {#{}, #{}} -> term - end, - Acc = set_element_type(Key, Type, Acc0), - join_elements_1(Keys, Es1, Es2, Acc); -join_elements_1([], _Es1, _Es2, Acc) -> - Acc. +%% `dialyzer` complains about the float and general literal cases never being +%% matched and I don't like suppressing warnings. Should they become possible +%% I'm sure `dialyzer` will warn about it. +value_to_literal([]) -> nil; +value_to_literal(A) when is_atom(A) -> {atom,A}; +value_to_literal(I) when is_integer(I) -> {integer,I}. -%% Joins 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,_}=Lit, T) -> - join_literal(T, get_literal_type(Lit)); -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. - -join_bool([]) -> {atom,[]}; -join_bool(true) -> bool; -join_bool(false) -> bool; -join_bool(_) -> {atom,[]}. - -%% 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(#ms{}, binary) -> - #ms{}; -meet(binary, #ms{}) -> - #ms{}; -meet({literal,_}, {literal,_}) -> - none; -meet(T1, {literal,_}=T2) -> - meet(T2, T1); -meet({literal,_}=T1, T2) -> - case meet(get_literal_type(T1), T2) of - none -> none; - _ -> T1 - end; -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,Es1},{tuple,Size2,Es2}} -> - Es = meet_elements(Es1, Es2), - case {Size1,Size2,Es} of - {_, _, none} -> - none; - {[Sz1],[Sz2],_} -> - Sz = erlang:max(Sz1, Sz2), - assert_tuple_elements(Sz, Es), - {tuple,[Sz],Es}; - {Sz1,[Sz2],_} when Sz2 =< Sz1 -> - assert_tuple_elements(Sz1, Es), - {tuple,Sz1,Es}; - {Sz,Sz,_} -> - assert_tuple_elements(Sz, Es), - {tuple,Sz,Es}; - {_,_,_} -> - none - end; - {_,_} -> none +%% These are just wrappers around their equivalents in beam_types, which +%% handle the validator-specific #t_abstract{} type. +%% +%% The funny-looking abstract types produced here are intended to provoke +%% errors on actual use; they do no harm just lying around. + +normalize(#t_abstract{}=A) -> error({abstract_type, A}); +normalize(T) -> beam_types:normalize(T). + +join(Same, Same) -> Same; +join(#t_abstract{}=A, B) -> #t_abstract{kind={join, A, B}}; +join(A, #t_abstract{}=B) -> #t_abstract{kind={join, A, B}}; +join(A, B) -> beam_types:join(A, B). + +meet(Same, Same) -> Same; +meet(#t_abstract{}=A, B) -> #t_abstract{kind={meet, A, B}}; +meet(A, #t_abstract{}=B) -> #t_abstract{kind={meet, A, B}}; +meet(A, B) -> beam_types:meet(A, B). + +subtract(#t_abstract{}=A, B) -> #t_abstract{kind={subtract, A, B}}; +subtract(A, #t_abstract{}=B) -> #t_abstract{kind={subtract, A, B}}; +subtract(A, B) -> beam_types:subtract(A, B). + +assert_type(RequiredType, Term, Vst) -> + GivenType = get_movable_term_type(Term, Vst), + case meet(RequiredType, GivenType) of + GivenType -> + ok; + _RequiredType -> + error({bad_type,{needed,RequiredType},{actual,GivenType}}) end. -meet_elements(Es1, Es2) -> - Keys = maps:keys(Es1) ++ maps:keys(Es2), - meet_elements_1(Keys, Es1, Es2, #{}). - -meet_elements_1([Key | Keys], Es1, Es2, Acc) -> - case {Es1, Es2} of - {#{ Key := Type1 }, #{ Key := Type2 }} -> - case meet(Type1, Type2) of - none -> none; - Type -> meet_elements_1(Keys, Es1, Es2, Acc#{ Key => Type }) - end; - {#{ Key := Type1 }, _} -> - meet_elements_1(Keys, Es1, Es2, Acc#{ Key => Type1 }); - {_, #{ Key := Type2 }} -> - meet_elements_1(Keys, Es1, Es2, Acc#{ Key => Type2 }) - end; -meet_elements_1([], _Es1, _Es2, Acc) -> - Acc. - -%% No tuple elements may have an index above the known size. -assert_tuple_elements(Limit, Es) -> - true = maps:fold(fun({integer,Index}, _T, true) -> - Index =< Limit - end, true, Es). %Assertion. - -%% subtract(Type1, Type2) -> Type -%% Subtract Type2 from Type2. Example: -%% subtract(list, nil) -> cons - -subtract(Same, Same) -> none; -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_term_type(Term, Vst), - assert_type(WantedType, Type). - -assert_type(Correct, Correct) -> ok; -assert_type(float, {float,_}) -> ok; -assert_type(tuple, {tuple,_,_}) -> ok; -assert_type(tuple, {literal,Tuple}) when is_tuple(Tuple) -> ok; -assert_type({tuple_element,I}, {tuple,[Sz],_}) - when 1 =< I, I =< Sz -> - ok; -assert_type({tuple_element,I}, {tuple,Sz,_}) - when is_integer(Sz), 1 =< I, I =< Sz -> - ok; -assert_type({tuple_element,I}, {literal,Lit}) when I =< tuple_size(Lit) -> - ok; -assert_type(cons, {literal,[_|_]}) -> - ok; -assert_type(Needed, Actual) -> - error({bad_type,{needed,Needed},{actual,Actual}}). - -get_element_type(Key, Src, Vst) -> - get_element_type_1(Key, get_term_type(Src, Vst)). - -get_element_type_1({integer,_}=Key, {tuple,_Sz,Es}) -> - case Es of - #{ Key := Type } -> Type; - #{} -> term - end; -get_element_type_1(_Index, _Type) -> - term. - -set_element_type(_Key, none, Es) -> - Es; -set_element_type(Key, term, Es) -> - maps:remove(Key, Es); -set_element_type(Key, Type, Es) -> - Es#{ Key => Type }. - -get_tuple_size({integer,[]}) -> 0; -get_tuple_size({integer,Sz}) -> Sz; -get_tuple_size(_) -> 0. - validate_src(Ss, Vst) when is_list(Ss) -> _ = [assert_term(S, Vst) || S <- Ss], ok. @@ -2227,7 +2033,8 @@ validate_src(Ss, Vst) when is_list(Ss) -> get_term_type(Src, Vst) -> case get_movable_term_type(Src, Vst) of - #ms{} -> error({match_context,Src}); + #t_bs_context{} -> error({match_context,Src}); + #t_abstract{} -> error({abstract_term,Src}); Type -> Type end. @@ -2237,12 +2044,11 @@ get_term_type(Src, Vst) -> get_movable_term_type(Src, Vst) -> case get_raw_type(Src, Vst) of + #t_abstract{kind=unfinished_tuple=Kind} -> error({Kind,Src}); initialized -> error({unassigned,Src}); uninitialized -> error({uninitialized_reg,Src}); {catchtag,_} -> error({catchtag,Src}); {trytag,_} -> error({trytag,Src}); - tuple_in_progress -> error({tuple_in_progress,Src}); - {literal,_}=Lit -> get_literal_type(Lit); Type -> Type end. @@ -2284,30 +2090,18 @@ get_raw_type(#value_ref{}=Ref, #vst{current=#st{vs=Vs}}) -> get_raw_type(Src, #vst{}) -> get_literal_type(Src). -get_literal_type(nil=T) -> T; -get_literal_type({atom,A}=T) when is_atom(A) -> T; -get_literal_type({float,F}=T) when is_float(F) -> T; -get_literal_type({integer,I}=T) when is_integer(I) -> T; -get_literal_type({literal,[_|_]}) -> cons; -get_literal_type({literal,Bitstring}) when is_bitstring(Bitstring) -> binary; -get_literal_type({literal,Map}) when is_map(Map) -> map; -get_literal_type({literal,Tuple}) when is_tuple(Tuple) -> glt_1(Tuple); -get_literal_type({literal,_}) -> term; -get_literal_type(T) -> error({not_literal,T}). - -glt_1([]) -> nil; -glt_1(A) when is_atom(A) -> {atom, A}; -glt_1(F) when is_float(F) -> {float, F}; -glt_1(I) when is_integer(I) -> {integer, I}; -glt_1(T) when is_tuple(T) -> - {Es,_} = foldl(fun(Val, {Es0, Index}) -> - Type = glt_1(Val), - Es = set_element_type({integer,Index}, Type, Es0), - {Es, Index + 1} - end, {#{}, 1}, tuple_to_list(T)), - {tuple, tuple_size(T), Es}; -glt_1(L) -> - {literal, L}. +get_literal_type(nil) -> + beam_types:make_type_from_value([]); +get_literal_type({atom,A}) when is_atom(A) -> + beam_types:make_type_from_value(A); +get_literal_type({float,F}) when is_float(F) -> + beam_types:make_type_from_value(F); +get_literal_type({integer,I}) when is_integer(I) -> + beam_types:make_type_from_value(I); +get_literal_type({literal,L}) -> + beam_types:make_type_from_value(L); +get_literal_type(T) -> + error({not_literal,T}). %%% %%% Branch tracking @@ -2325,10 +2119,12 @@ glt_1(L) -> SuccFun :: BranchFun) -> #vst{} when BranchFun :: fun((#vst{}) -> #vst{}). branch(Lbl, Vst0, FailFun, SuccFun) -> + validate_branch(Lbl, Vst0), #vst{current=St0} = Vst0, + try FailFun(Vst0) of Vst1 -> - Vst2 = branch_state(Lbl, Vst1), + Vst2 = fork_state(Lbl, Vst1), Vst = Vst2#vst{current=St0}, try SuccFun(Vst) of V -> V @@ -2346,6 +2142,24 @@ branch(Lbl, Vst0, FailFun, SuccFun) -> SuccFun(Vst0) end. +validate_branch(Lbl, #vst{current=#st{ct=Tags}}) -> + validate_branch_1(Lbl, Tags). + +validate_branch_1(Lbl, [{trytag, FailLbls} | Tags]) -> + %% 'try_case' assumes that an exception has been thrown, so a direct branch + %% will crash the emulator. + %% + %% (Jumping to a 'catch_end' is fine however as it will simply nop in the + %% absence of an exception.) + case ordsets:is_element(Lbl, FailLbls) of + true -> error({illegal_branch, try_handler, Lbl}); + false -> validate_branch_1(Lbl, Tags) + end; +validate_branch_1(Lbl, [_ | Tags]) -> + validate_branch_1(Lbl, Tags); +validate_branch_1(_Lbl, []) -> + ok. + %% A shorthand version of branch/4 for when the state is only altered on %% success. branch(Fail, Vst, SuccFun) -> @@ -2353,12 +2167,12 @@ branch(Fail, Vst, SuccFun) -> %% Directly branches off the state. This is an "internal" operation that should %% be used sparingly. -branch_state(0, #vst{}=Vst) -> +fork_state(0, #vst{}=Vst) -> %% If the instruction fails, the stack may be scanned looking for a catch %% tag. Therefore the Y registers must be initialized at this point. verify_y_init(Vst), Vst; -branch_state(L, #vst{current=St,branched=B,ref_ctr=Counter0}=Vst) -> +fork_state(L, #vst{current=St,branched=B,ref_ctr=Counter0}=Vst) -> case gb_trees:is_defined(L, B) of true -> {MergedSt, Counter} = merge_states(L, St, B, Counter0), @@ -2438,10 +2252,10 @@ merge_tags(uninitialized, _) -> uninitialized; merge_tags(_, uninitialized) -> uninitialized; -merge_tags({catchtag,T0}, {catchtag,T1}) -> - {catchtag, ordsets:from_list(T0 ++ T1)}; -merge_tags({trytag,T0}, {trytag,T1}) -> - {trytag, ordsets:from_list(T0 ++ T1)}; +merge_tags({trytag, LblsA}, {trytag, LblsB}) -> + {trytag, ordsets:union(LblsA, LblsB)}; +merge_tags({catchtag, LblsA}, {catchtag, LblsB}) -> + {catchtag, ordsets:union(LblsA, LblsB)}; merge_tags(_A, _B) -> %% All other combinations leave the register initialized. Errors arising %% from this will be caught later on. @@ -2512,13 +2326,14 @@ merge_stk(_, _) -> undecided. merge_ct(S, S) -> S; merge_ct(Ct0, Ct1) -> merge_ct_1(Ct0, Ct1). -merge_ct_1([C0|Ct0], [C1|Ct1]) -> - [ordsets:from_list(C0++C1)|merge_ct_1(Ct0, Ct1)]; -merge_ct_1([], []) -> []; -merge_ct_1(_, _) -> undecided. - -tuple_sz([Sz]) -> Sz; -tuple_sz(Sz) -> Sz. +merge_ct_1([], []) -> + []; +merge_ct_1([{trytag, LblsA} | CtA], [{trytag, LblsB} | CtB]) -> + [{trytag, ordsets:union(LblsA, LblsB)} | merge_ct_1(CtA, CtB)]; +merge_ct_1([{catchtag, LblsA} | CtA], [{catchtag, LblsB} | CtB]) -> + [{catchtag, ordsets:union(LblsA, LblsB)} | merge_ct_1(CtA, CtB)]; +merge_ct_1(_, _) -> + undecided. verify_y_init(#vst{current=#st{numy=NumY,ys=Ys}}=Vst) when is_integer(NumY) -> HighestY = maps:fold(fun({y,Y}, _, Acc) -> max(Y, Acc) end, -1, Ys), @@ -2661,319 +2476,46 @@ assert_not_fragile(Lit, #vst{}) -> ok. %%% -%%% Return/argument types of BIFs +%%% Return/argument types of calls and BIFs %%% -bif_return_type('-', Src, Vst) -> - arith_return_type(Src, Vst); -bif_return_type('+', Src, Vst) -> - arith_return_type(Src, Vst); -bif_return_type('*', Src, Vst) -> - arith_return_type(Src, Vst); -bif_return_type(abs, [Num], Vst) -> - case get_term_type(Num, Vst) of - {float,_}=T -> T; - {integer,_}=T -> T; - _ -> number - end; -bif_return_type(float, _, _) -> {float,[]}; -bif_return_type('/', _, _) -> {float,[]}; -%% Binary operations -bif_return_type('binary_part', [_,_], _) -> binary; -bif_return_type('binary_part', [_,_,_], _) -> binary; -bif_return_type('bit_size', [_], _) -> {integer,[]}; -bif_return_type('byte_size', [_], _) -> {integer,[]}; -%% Integer operations. -bif_return_type(ceil, [_], _) -> {integer,[]}; -bif_return_type('div', [_,_], _) -> {integer,[]}; -bif_return_type(floor, [_], _) -> {integer,[]}; -bif_return_type('rem', [_,_], _) -> {integer,[]}; -bif_return_type(length, [_], _) -> {integer,[]}; -bif_return_type(size, [_], _) -> {integer,[]}; -bif_return_type(trunc, [_], _) -> {integer,[]}; -bif_return_type(round, [_], _) -> {integer,[]}; -bif_return_type('band', [_,_], _) -> {integer,[]}; -bif_return_type('bor', [_,_], _) -> {integer,[]}; -bif_return_type('bxor', [_,_], _) -> {integer,[]}; -bif_return_type('bnot', [_], _) -> {integer,[]}; -bif_return_type('bsl', [_,_], _) -> {integer,[]}; -bif_return_type('bsr', [_,_], _) -> {integer,[]}; -%% Booleans. -bif_return_type('==', [_,_], _) -> bool; -bif_return_type('/=', [_,_], _) -> bool; -bif_return_type('=<', [_,_], _) -> bool; -bif_return_type('<', [_,_], _) -> bool; -bif_return_type('>=', [_,_], _) -> bool; -bif_return_type('>', [_,_], _) -> bool; -bif_return_type('=:=', [_,_], _) -> bool; -bif_return_type('=/=', [_,_], _) -> bool; -bif_return_type('not', [_], _) -> bool; -bif_return_type('and', [_,_], _) -> bool; -bif_return_type('or', [_,_], _) -> bool; -bif_return_type('xor', [_,_], _) -> bool; -bif_return_type(is_atom, [_], _) -> bool; -bif_return_type(is_boolean, [_], _) -> bool; -bif_return_type(is_binary, [_], _) -> bool; -bif_return_type(is_float, [_], _) -> bool; -bif_return_type(is_function, [_], _) -> bool; -bif_return_type(is_function, [_,_], _) -> bool; -bif_return_type(is_integer, [_], _) -> bool; -bif_return_type(is_list, [_], _) -> bool; -bif_return_type(is_map, [_], _) -> bool; -bif_return_type(is_number, [_], _) -> bool; -bif_return_type(is_pid, [_], _) -> bool; -bif_return_type(is_port, [_], _) -> bool; -bif_return_type(is_reference, [_], _) -> bool; -bif_return_type(is_tuple, [_], _) -> bool; -%% Misc. -bif_return_type(tuple_size, [_], _) -> {integer,[]}; -bif_return_type(map_size, [_], _) -> {integer,[]}; -bif_return_type(node, [], _) -> {atom,[]}; -bif_return_type(node, [_], _) -> {atom,[]}; -bif_return_type(hd, [_], _) -> term; -bif_return_type(tl, [_], _) -> term; -bif_return_type(get, [_], _) -> term; -bif_return_type(Bif, _, _) when is_atom(Bif) -> term. - -%% Generic -bif_arg_types(tuple_size, [_]) -> [{tuple,[0],#{}}]; -bif_arg_types(map_size, [_]) -> [map]; -bif_arg_types(is_map_key, [_,_]) -> [term, map]; -bif_arg_types(map_get, [_,_]) -> [term, map]; -bif_arg_types(length, [_]) -> [list]; -bif_arg_types(hd, [_]) -> [cons]; -bif_arg_types(tl, [_]) -> [cons]; -%% Boolean -bif_arg_types('not', [_]) -> [bool]; -bif_arg_types('and', [_,_]) -> [bool, bool]; -bif_arg_types('or', [_,_]) -> [bool, bool]; -bif_arg_types('xor', [_,_]) -> [bool, bool]; -%% Binary -bif_arg_types('binary_part', [_,_]) -> - PosLen = {tuple, 2, #{ {integer,1} => {integer,[]}, - {integer,2} => {integer,[]} }}, - [binary, PosLen]; -bif_arg_types('binary_part', [_,_,_]) -> - [binary, {integer,[]}, {integer,[]}]; -bif_arg_types('bit_size', [_]) -> [binary]; -bif_arg_types('byte_size', [_]) -> [binary]; -%% Numerical -bif_arg_types('-', [_]) -> [number]; -bif_arg_types('-', [_,_]) -> [number,number]; -bif_arg_types('+', [_]) -> [number]; -bif_arg_types('+', [_,_]) -> [number,number]; -bif_arg_types('*', [_,_]) -> [number, number]; -bif_arg_types('/', [_,_]) -> [number, number]; -bif_arg_types(abs, [_]) -> [number]; -bif_arg_types(ceil, [_]) -> [number]; -bif_arg_types(float, [_]) -> [number]; -bif_arg_types(floor, [_]) -> [number]; -bif_arg_types(trunc, [_]) -> [number]; -bif_arg_types(round, [_]) -> [number]; -%% Integer-specific -bif_arg_types('div', [_,_]) -> [{integer,[]}, {integer,[]}]; -bif_arg_types('rem', [_,_]) -> [{integer,[]}, {integer,[]}]; -bif_arg_types('band', [_,_]) -> [{integer,[]}, {integer,[]}]; -bif_arg_types('bor', [_,_]) -> [{integer,[]}, {integer,[]}]; -bif_arg_types('bxor', [_,_]) -> [{integer,[]}, {integer,[]}]; -bif_arg_types('bnot', [_]) -> [{integer,[]}]; -bif_arg_types('bsl', [_,_]) -> [{integer,[]}, {integer,[]}]; -bif_arg_types('bsr', [_,_]) -> [{integer,[]}, {integer,[]}]; -%% Unsafe type tests that may fail if an argument doesn't have the right type. -bif_arg_types(is_function, [_,_]) -> [term, {integer,[]}]; -bif_arg_types(_, Args) -> [term || _Arg <- Args]. - -is_bif_safe('/=', 2) -> true; -is_bif_safe('<', 2) -> true; -is_bif_safe('=/=', 2) -> true; -is_bif_safe('=:=', 2) -> true; -is_bif_safe('=<', 2) -> true; -is_bif_safe('==', 2) -> true; -is_bif_safe('>', 2) -> true; -is_bif_safe('>=', 2) -> true; -is_bif_safe(is_atom, 1) -> true; -is_bif_safe(is_boolean, 1) -> true; -is_bif_safe(is_binary, 1) -> true; -is_bif_safe(is_bitstring, 1) -> true; -is_bif_safe(is_float, 1) -> true; -is_bif_safe(is_function, 1) -> true; -is_bif_safe(is_integer, 1) -> true; -is_bif_safe(is_list, 1) -> true; -is_bif_safe(is_map, 1) -> true; -is_bif_safe(is_number, 1) -> true; -is_bif_safe(is_pid, 1) -> true; -is_bif_safe(is_port, 1) -> true; -is_bif_safe(is_reference, 1) -> true; -is_bif_safe(is_tuple, 1) -> true; -is_bif_safe(get, 1) -> true; -is_bif_safe(self, 0) -> true; -is_bif_safe(node, 0) -> true; -is_bif_safe(_, _) -> false. - -arith_return_type([A], Vst) -> - %% Unary '+' or '-'. - case get_term_type(A, Vst) of - {integer,_} -> {integer,[]}; - {float,_} -> {float,[]}; - _ -> number - end; -arith_return_type([A,B], Vst) -> - TypeA = get_term_type(A, Vst), - TypeB = get_term_type(B, Vst), - case {TypeA, TypeB} of - {{integer,_},{integer,_}} -> {integer,[]}; - {{float,_},_} -> {float,[]}; - {_,{float,_}} -> {float,[]}; - {_,_} -> number - end; -arith_return_type(_, _) -> number. - -%%% -%%% Return/argument types of calls -%%% - -call_return_type({extfunc,M,F,A}, Vst) -> call_return_type_1(M, F, A, Vst); -call_return_type(_, _) -> term. - -call_return_type_1(erlang, setelement, 3, Vst) -> - IndexType = get_term_type({x,0}, Vst), - TupleType = - case get_term_type({x,1}, Vst) of - {literal,Tuple}=Lit when is_tuple(Tuple) -> get_literal_type(Lit); - {tuple,_,_}=TT -> TT; - _ -> {tuple,[0],#{}} - end, - case IndexType of - {integer,I} when is_integer(I) -> - case meet({tuple,[I],#{}}, TupleType) of - {tuple, Sz, Es0} -> - ValueType = get_term_type({x,2}, Vst), - Es = set_element_type({integer,I}, ValueType, Es0), - {tuple, Sz, Es}; - none -> - TupleType - end; - _ -> - %% The index could point anywhere, so we must discard all element - %% information. - setelement(3, TupleType, #{}) - end; -call_return_type_1(erlang, '++', 2, Vst) -> - LType = get_term_type({x,0}, Vst), - RType = get_term_type({x,1}, Vst), - case LType =:= cons orelse RType =:= cons of - true -> - cons; - false -> - %% `[] ++ RHS` yields RHS, even if RHS is not a list - join(list, RType) - end; -call_return_type_1(erlang, '--', 2, _Vst) -> - list; -call_return_type_1(erlang, F, A, _) -> - erlang_mod_return_type(F, A); -call_return_type_1(lists, F, A, Vst) -> - lists_mod_return_type(F, A, Vst); -call_return_type_1(math, F, A, _) -> - math_mod_return_type(F, A); -call_return_type_1(M, F, A, _) when is_atom(M), is_atom(F), is_integer(A), A >= 0 -> - term. - -erlang_mod_return_type(exit, 1) -> exception; -erlang_mod_return_type(throw, 1) -> exception; -erlang_mod_return_type(error, 1) -> exception; -erlang_mod_return_type(error, 2) -> exception; -erlang_mod_return_type(F, A) when is_atom(F), is_integer(A), A >= 0 -> term. - -math_mod_return_type(cos, 1) -> {float,[]}; -math_mod_return_type(cosh, 1) -> {float,[]}; -math_mod_return_type(sin, 1) -> {float,[]}; -math_mod_return_type(sinh, 1) -> {float,[]}; -math_mod_return_type(tan, 1) -> {float,[]}; -math_mod_return_type(tanh, 1) -> {float,[]}; -math_mod_return_type(acos, 1) -> {float,[]}; -math_mod_return_type(acosh, 1) -> {float,[]}; -math_mod_return_type(asin, 1) -> {float,[]}; -math_mod_return_type(asinh, 1) -> {float,[]}; -math_mod_return_type(atan, 1) -> {float,[]}; -math_mod_return_type(atanh, 1) -> {float,[]}; -math_mod_return_type(erf, 1) -> {float,[]}; -math_mod_return_type(erfc, 1) -> {float,[]}; -math_mod_return_type(exp, 1) -> {float,[]}; -math_mod_return_type(log, 1) -> {float,[]}; -math_mod_return_type(log2, 1) -> {float,[]}; -math_mod_return_type(log10, 1) -> {float,[]}; -math_mod_return_type(sqrt, 1) -> {float,[]}; -math_mod_return_type(atan2, 2) -> {float,[]}; -math_mod_return_type(pow, 2) -> {float,[]}; -math_mod_return_type(ceil, 1) -> {float,[]}; -math_mod_return_type(floor, 1) -> {float,[]}; -math_mod_return_type(fmod, 2) -> {float,[]}; -math_mod_return_type(pi, 0) -> {float,[]}; -math_mod_return_type(F, A) when is_atom(F), is_integer(A), A >= 0 -> term. - -lists_mod_return_type(all, 2, _Vst) -> - bool; -lists_mod_return_type(any, 2, _Vst) -> - bool; -lists_mod_return_type(keymember, 3, _Vst) -> - bool; -lists_mod_return_type(member, 2, _Vst) -> - bool; -lists_mod_return_type(prefix, 2, _Vst) -> - bool; -lists_mod_return_type(suffix, 2, _Vst) -> - bool; -lists_mod_return_type(dropwhile, 2, _Vst) -> - list; -lists_mod_return_type(duplicate, 2, _Vst) -> - list; -lists_mod_return_type(filter, 2, _Vst) -> - list; -lists_mod_return_type(flatten, 1, _Vst) -> - list; -lists_mod_return_type(map, 2, Vst) -> - same_length_type({x,1}, Vst); -lists_mod_return_type(MF, 3, Vst) when MF =:= mapfoldl; MF =:= mapfoldr -> - ListType = same_length_type({x,2}, Vst), - {tuple,2,#{ {integer,1} => ListType} }; -lists_mod_return_type(partition, 2, _Vst) -> - two_tuple(list, list); -lists_mod_return_type(reverse, 1, Vst) -> - same_length_type({x,0}, Vst); -lists_mod_return_type(seq, 2, _Vst) -> - list; -lists_mod_return_type(sort, 1, Vst) -> - same_length_type({x,0}, Vst); -lists_mod_return_type(sort, 2, Vst) -> - same_length_type({x,1}, Vst); -lists_mod_return_type(splitwith, 2, _Vst) -> - two_tuple(list, list); -lists_mod_return_type(takewhile, 2, _Vst) -> - list; -lists_mod_return_type(unzip, 1, Vst) -> - ListType = same_length_type({x,0}, Vst), - two_tuple(ListType, ListType); -lists_mod_return_type(usort, 1, Vst) -> - same_length_type({x,0}, Vst); -lists_mod_return_type(zip, 2, _Vst) -> - list; -lists_mod_return_type(zipwith, 3, _Vst) -> - list; -lists_mod_return_type(_, _, _) -> - term. - -two_tuple(Type1, Type2) -> - {tuple,2,#{ {integer,1} => Type1, - {integer,2} => Type2 }}. - -same_length_type(Reg, Vst) -> - case get_term_type(Reg, Vst) of - {literal,[_|_]} -> cons; - cons -> cons; - nil -> nil; - _ -> list - end. +bif_types(Op, Ss, Vst) -> + Args = [normalize(get_term_type(Arg, Vst)) || Arg <- Ss], + beam_call_types:types(erlang, Op, Args). + +call_types({extfunc,M,F,A}, A, Vst) -> + Args = get_call_args(A, Vst), + beam_call_types:types(M, F, Args); +call_types(_, A, Vst) -> + {any, get_call_args(A, Vst), false}. + +will_bif_succeed(fadd, [_,_], _Vst) -> + maybe; +will_bif_succeed(fdiv, [_,_], _Vst) -> + maybe; +will_bif_succeed(fmul, [_,_], _Vst) -> + maybe; +will_bif_succeed(fnegate, [_], _Vst) -> + maybe; +will_bif_succeed(fsub, [_,_], _Vst) -> + maybe; +will_bif_succeed(Op, Ss, Vst) -> + Args = [normalize(get_term_type(Arg, Vst)) || Arg <- Ss], + beam_call_types:will_succeed(erlang, Op, Args). + +will_call_succeed({extfunc,M,F,A}, Vst) -> + beam_call_types:will_succeed(M, F, get_call_args(A, Vst)); +will_call_succeed(_Call, _Vst) -> + maybe. + +get_call_args(Arity, Vst) -> + get_call_args_1(0, Arity, Vst). + +get_call_args_1(Arity, Arity, _) -> + []; +get_call_args_1(N, Arity, Vst) when N < Arity -> + ArgType = normalize(get_movable_term_type({x,N}, Vst)), + [ArgType | get_call_args_1(N + 1, Arity, Vst)]. check_limit({x,X}=Src) when is_integer(X) -> if diff --git a/lib/compiler/src/cerl_sets.erl b/lib/compiler/src/cerl_sets.erl index f489baf238..84e488fc55 100644 --- a/lib/compiler/src/cerl_sets.erl +++ b/lib/compiler/src/cerl_sets.erl @@ -153,14 +153,21 @@ intersection1(S1, []) -> S1. Set1 :: set(Element), Set2 :: set(Element). -is_disjoint(S1, S2) when map_size(S1) < map_size(S2) -> - fold(fun (_, false) -> false; - (E, true) -> not is_element(E, S2) - end, true, S1); +is_disjoint(S1, S2) when map_size(S1) > map_size(S2) -> + is_disjoint_1(S1, maps:iterator(S2)); is_disjoint(S1, S2) -> - fold(fun (_, false) -> false; - (E, true) -> not is_element(E, S1) - end, true, S2). + is_disjoint_1(S2, maps:iterator(S1)). + +is_disjoint_1(Set, Iter) -> + case maps:next(Iter) of + {K, _, NextIter} -> + case Set of + #{K := _} -> false; + #{} -> is_disjoint_1(Set, NextIter) + end; + none -> + true + end. %% subtract(Set1, Set2) -> Set. %% Return all and only the elements of Set1 which are not also in @@ -180,8 +187,21 @@ subtract(S1, S2) -> Set1 :: set(Element), Set2 :: set(Element). +is_subset(S1, S2) when map_size(S1) > map_size(S2) -> + false; is_subset(S1, S2) -> - fold(fun (E, Sub) -> Sub andalso is_element(E, S2) end, true, S1). + is_subset_1(S2, maps:iterator(S1)). + +is_subset_1(Set, Iter) -> + case maps:next(Iter) of + {K, _, NextIter} -> + case Set of + #{K := _} -> is_subset_1(Set, NextIter); + #{} -> false + end; + none -> + true + end. %% fold(Fun, Accumulator, Set) -> Accumulator. %% Fold function Fun over all elements in Set and return Accumulator. @@ -193,8 +213,16 @@ is_subset(S1, S2) -> AccIn :: Acc, AccOut :: Acc. -fold(F, Init, D) -> - lists:foldl(fun(E,Acc) -> F(E,Acc) end,Init,maps:keys(D)). +fold(Fun, Init, Set) -> + fold_1(Fun, Init, maps:iterator(Set)). + +fold_1(Fun, Acc, Iter) -> + case maps:next(Iter) of + {K, _, NextIter} -> + fold_1(Fun, Fun(K,Acc), NextIter); + none -> + Acc + end. %% filter(Fun, Set) -> Set. %% Filter Set with Fun. @@ -203,5 +231,18 @@ fold(F, Init, D) -> Set1 :: set(Element), Set2 :: set(Element). -filter(F, D) -> - maps:filter(fun(K,_) -> F(K) end, D). +filter(Fun, Set) -> + maps:from_list(filter_1(Fun, maps:iterator(Set))). + +filter_1(Fun, Iter) -> + case maps:next(Iter) of + {K, _, NextIter} -> + case Fun(K) of + true -> + [{K,ok} | filter_1(Fun, NextIter)]; + false -> + filter_1(Fun, NextIter) + end; + none -> + [] + end. diff --git a/lib/compiler/src/compile.erl b/lib/compiler/src/compile.erl index 0325c714d0..21d67f5649 100644 --- a/lib/compiler/src/compile.erl +++ b/lib/compiler/src/compile.erl @@ -265,7 +265,10 @@ expand_opt(r19, Os) -> expand_opt(r20, Os) -> expand_opt_before_21(Os); expand_opt(r21, Os) -> - [no_put_tuple2 | expand_opt(no_bsm3, Os)]; + [no_shared_fun_wrappers, + no_swap, no_put_tuple2 | expand_opt(no_bsm3, Os)]; +expand_opt(r22, Os) -> + [no_shared_fun_wrappers, no_swap | Os]; expand_opt({debug_info_key,_}=O, Os) -> [encrypt_debug_info,O|Os]; expand_opt(no_type_opt=O, Os) -> @@ -278,7 +281,8 @@ expand_opt(no_type_opt=O, Os) -> expand_opt(O, Os) -> [O|Os]. expand_opt_before_21(Os) -> - [no_put_tuple2, no_get_hd_tl, no_ssa_opt_record, + [no_shared_fun_wrappers, no_swap, + no_put_tuple2, no_get_hd_tl, no_ssa_opt_record, no_utf8_atoms | expand_opt(no_bsm3, Os)]. %% format_error(ErrorDescriptor) -> string() @@ -863,8 +867,6 @@ asm_passes() -> {unless,no_postopt, [{pass,beam_block}, {iff,dblk,{listing,"block"}}, - {unless,no_except,{pass,beam_except}}, - {iff,dexcept,{listing,"except"}}, {unless,no_jopt,{pass,beam_jump}}, {iff,djmp,{listing,"jump"}}, {unless,no_peep_opt,{pass,beam_peep}}, @@ -2096,9 +2098,9 @@ pre_load() -> L = [beam_a, beam_asm, beam_block, + beam_call_types, beam_clean, beam_dict, - beam_except, beam_flatten, beam_jump, beam_kernel_to_ssa, @@ -2115,6 +2117,7 @@ pre_load() -> beam_ssa_share, beam_ssa_type, beam_trim, + beam_types, beam_utils, beam_validator, beam_z, diff --git a/lib/compiler/src/compiler.app.src b/lib/compiler/src/compiler.app.src index a086a3a8d3..092757ae65 100644 --- a/lib/compiler/src/compiler.app.src +++ b/lib/compiler/src/compiler.app.src @@ -24,10 +24,10 @@ beam_a, beam_asm, beam_block, + beam_call_types, beam_clean, beam_dict, beam_disasm, - beam_except, beam_flatten, beam_jump, beam_kernel_to_ssa, @@ -47,6 +47,7 @@ beam_ssa_share, beam_ssa_type, beam_trim, + beam_types, beam_utils, beam_validator, beam_z, diff --git a/lib/compiler/src/genop.tab b/lib/compiler/src/genop.tab index 86590fad87..0a38d17857 100755 --- a/lib/compiler/src/genop.tab +++ b/lib/compiler/src/genop.tab @@ -596,3 +596,9 @@ BEAM_FORMAT_NUMBER=0 ## @spec bs_set_positon Ctx Pos ## @doc Sets the current position of Ctx to Pos 168: bs_set_position/2 + +# OTP 23 + +## @spec swap Register1 Register2 +## @doc Swaps the contents of two registers. +169: swap/2 diff --git a/lib/compiler/src/sys_core_fold.erl b/lib/compiler/src/sys_core_fold.erl index 4939a94a92..63c67639d4 100644 --- a/lib/compiler/src/sys_core_fold.erl +++ b/lib/compiler/src/sys_core_fold.erl @@ -99,10 +99,6 @@ t=#{} :: map(), %Types in_guard=false}). %In guard or not. --type type_info() :: cerl:cerl() | 'bool' | 'integer' | {'fun', pos_integer()}. --type yes_no_maybe() :: 'yes' | 'no' | 'maybe'. --type sub() :: #sub{}. - -spec module(cerl:c_module(), [compile:option()]) -> {'ok', cerl:c_module(), [_]}. @@ -315,10 +311,10 @@ expr(#c_seq{arg=Arg0,body=B0}=Seq0, Ctxt, Sub) -> false -> %% Arg cannot be "values" here - only a single value %% make sense here. - case {Ctxt,is_safe_simple(Arg, Sub)} of + case {Ctxt,is_safe_simple(Arg)} of {effect,true} -> B1; {effect,false} -> - case is_safe_simple(B1, Sub) of + case is_safe_simple(B1) of true -> Arg; false -> Seq0#c_seq{arg=Arg,body=B1} end; @@ -442,7 +438,7 @@ expr(#c_catch{anno=Anno,body=B}, effect, Sub) -> expr(#c_catch{body=B0}=Catch, _, Sub) -> %% We can remove catch if the value is simple B1 = body(B0, value, Sub), - case is_safe_simple(B1, Sub) of + case is_safe_simple(B1) of true -> B1; false -> Catch#c_catch{body=B1} end; @@ -458,7 +454,7 @@ expr(#c_try{arg=E0,vars=[#c_var{name=X}],body=#c_var{name=X}, %% We can remove try/catch if the expression is an %% expression that cannot fail. - case is_safe_bool_expr(E2, Sub) orelse is_safe_simple(E2, Sub) of + case is_safe_bool_expr(E2) orelse is_safe_simple(E2) of true -> E2; false -> Try#c_try{arg=E2} end; @@ -472,7 +468,7 @@ expr(#c_try{anno=A,arg=E0,vars=Vs0,body=B0,evars=Evs0,handler=H0}=Try, _, Sub0) E1 = body(E0, value, Sub0), {Vs1,Sub1} = var_list(Vs0, Sub0), B1 = body(B0, value, Sub1), - case is_safe_simple(E1, Sub0) of + case is_safe_simple(E1) of true -> expr(#c_let{anno=A,vars=Vs1,arg=E1,body=B1}, value, Sub0); false -> @@ -602,20 +598,20 @@ is_literal_fun(_) -> false. %% Currently, we don't attempt to check binaries because they %% are difficult to check. -is_safe_simple(#c_var{}=Var, _) -> +is_safe_simple(#c_var{}=Var) -> not cerl:is_c_fname(Var); -is_safe_simple(#c_cons{hd=H,tl=T}, Sub) -> - is_safe_simple(H, Sub) andalso is_safe_simple(T, Sub); -is_safe_simple(#c_tuple{es=Es}, Sub) -> is_safe_simple_list(Es, Sub); -is_safe_simple(#c_literal{}, _) -> true; +is_safe_simple(#c_cons{hd=H,tl=T}) -> + is_safe_simple(H) andalso is_safe_simple(T); +is_safe_simple(#c_tuple{es=Es}) -> is_safe_simple_list(Es); +is_safe_simple(#c_literal{}) -> true; is_safe_simple(#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=Name}, - args=Args}, Sub) when is_atom(Name) -> + args=Args}) when is_atom(Name) -> NumArgs = length(Args), case erl_internal:bool_op(Name, NumArgs) of true -> %% Boolean operators are safe if the arguments are boolean. - all(fun(C) -> is_boolean_type(C, Sub) =:= yes end, Args); + all(fun is_bool_expr/1, Args); false -> %% We need a rather complicated test to ensure that %% we only allow safe calls that are allowed in a guard. @@ -624,9 +620,9 @@ is_safe_simple(#c_call{module=#c_literal{val=erlang}, (erl_internal:comp_op(Name, NumArgs) orelse erl_internal:new_type_test(Name, NumArgs)) end; -is_safe_simple(_, _) -> false. +is_safe_simple(_) -> false. -is_safe_simple_list(Es, Sub) -> all(fun(E) -> is_safe_simple(E, Sub) end, Es). +is_safe_simple_list(Es) -> all(fun(E) -> is_safe_simple(E) end, Es). %% will_fail(Expr) -> true|false. %% Determine whether the expression will fail with an exception. @@ -853,7 +849,7 @@ useless_call(_, _) -> no. %% Anything that will not have any effect will be thrown away. make_effect_seq([H|T], Sub) -> - case is_safe_simple(H, Sub) of + case is_safe_simple(H) of true -> make_effect_seq(T, Sub); false -> #c_seq{arg=H,body=make_effect_seq(T, Sub)} end; @@ -959,138 +955,14 @@ fold_lit_args(Call, Module, Name, Args0) -> %% Attempt to evaluate some pure BIF calls with one or more %% non-literals arguments. %% -fold_non_lit_args(Call, erlang, is_boolean, [Arg], Sub) -> - eval_is_boolean(Call, Arg, Sub); fold_non_lit_args(Call, erlang, length, [Arg], _) -> eval_length(Call, Arg); fold_non_lit_args(Call, erlang, '++', [Arg1,Arg2], _) -> eval_append(Call, Arg1, Arg2); fold_non_lit_args(Call, lists, append, [Arg1,Arg2], _) -> eval_append(Call, Arg1, Arg2); -fold_non_lit_args(Call, erlang, is_function, [Arg1], Sub) -> - eval_is_function_1(Call, Arg1, Sub); -fold_non_lit_args(Call, erlang, is_function, [Arg1,Arg2], Sub) -> - eval_is_function_2(Call, Arg1, Arg2, Sub); -fold_non_lit_args(Call, erlang, N, Args, Sub) -> - NumArgs = length(Args), - case erl_internal:comp_op(N, NumArgs) of - true -> - eval_rel_op(Call, N, Args, Sub); - false -> - case erl_internal:bool_op(N, NumArgs) of - true -> - eval_bool_op(Call, N, Args, Sub); - false -> - Call - end - end; fold_non_lit_args(Call, _, _, _, _) -> Call. -eval_is_function_1(Call, Arg1, Sub) -> - case get_type(Arg1, Sub) of - none -> Call; - {'fun',_} -> #c_literal{anno=cerl:get_ann(Call),val=true}; - _ -> #c_literal{anno=cerl:get_ann(Call),val=false} - end. - -eval_is_function_2(Call, Arg1, #c_literal{val=Arity}, Sub) - when is_integer(Arity), Arity > 0 -> - case get_type(Arg1, Sub) of - none -> Call; - {'fun',Arity} -> #c_literal{anno=cerl:get_ann(Call),val=true}; - _ -> #c_literal{anno=cerl:get_ann(Call),val=false} - end; -eval_is_function_2(Call, _Arg1, _Arg2, _Sub) -> Call. - -%% Evaluate a relational operation using type information. -eval_rel_op(Call, Op, [#c_var{name=V},#c_var{name=V}], _) -> - Bool = erlang:Op(same, same), - #c_literal{anno=cerl:get_ann(Call),val=Bool}; -eval_rel_op(Call, '=:=', [Term,#c_literal{val=true}], Sub) -> - %% BoolVar =:= true ==> BoolVar - case is_boolean_type(Term, Sub) of - yes -> Term; - maybe -> Call; - no -> #c_literal{val=false} - end; -eval_rel_op(Call, '==', Ops, Sub) -> - case is_exact_eq_ok(Ops, Sub) of - true -> - Name = #c_literal{anno=cerl:get_ann(Call),val='=:='}, - Call#c_call{name=Name}; - false -> - Call - end; -eval_rel_op(Call, '/=', Ops, Sub) -> - case is_exact_eq_ok(Ops, Sub) of - true -> - Name = #c_literal{anno=cerl:get_ann(Call),val='=/='}, - Call#c_call{name=Name}; - false -> - Call - end; -eval_rel_op(Call, _, _, _) -> Call. - -is_exact_eq_ok([A,B]=L, Sub) -> - case is_int_type(A, Sub) =:= yes andalso is_int_type(B, Sub) =:= yes of - true -> true; - false -> is_exact_eq_ok_1(L) - end. - -is_exact_eq_ok_1([#c_literal{val=Lit}|_]) -> - is_non_numeric(Lit); -is_exact_eq_ok_1([_|T]) -> - is_exact_eq_ok_1(T); -is_exact_eq_ok_1([]) -> false. - -is_non_numeric([H|T]) -> - is_non_numeric(H) andalso is_non_numeric(T); -is_non_numeric(Tuple) when is_tuple(Tuple) -> - is_non_numeric_tuple(Tuple, tuple_size(Tuple)); -is_non_numeric(Map) when is_map(Map) -> - %% Note that 17.x and 18.x compare keys in different ways. - %% Be very conservative -- require that both keys and values - %% are non-numeric. - is_non_numeric(maps:to_list(Map)); -is_non_numeric(Num) when is_number(Num) -> - false; -is_non_numeric(_) -> true. - -is_non_numeric_tuple(Tuple, El) when El >= 1 -> - is_non_numeric(element(El, Tuple)) andalso - is_non_numeric_tuple(Tuple, El-1); -is_non_numeric_tuple(_Tuple, 0) -> true. - -%% Evaluate a bool op using type information. We KNOW that -%% there must be at least one non-literal argument (i.e. -%% there is no need to handle the case that all argments -%% are literal). - -eval_bool_op(Call, 'and', [#c_literal{val=true},Term], Sub) -> - eval_bool_op_1(Call, Term, Term, Sub); -eval_bool_op(Call, 'and', [Term,#c_literal{val=true}], Sub) -> - eval_bool_op_1(Call, Term, Term, Sub); -eval_bool_op(Call, 'and', [#c_literal{val=false}=Res,Term], Sub) -> - eval_bool_op_1(Call, Res, Term, Sub); -eval_bool_op(Call, 'and', [Term,#c_literal{val=false}=Res], Sub) -> - eval_bool_op_1(Call, Res, Term, Sub); -eval_bool_op(Call, _, _, _) -> Call. - -eval_bool_op_1(Call, Res, Term, Sub) -> - case is_boolean_type(Term, Sub) of - yes -> Res; - no -> eval_failure(Call, badarg); - maybe -> Call - end. - -%% Evaluate is_boolean/1 using type information. -eval_is_boolean(Call, Term, Sub) -> - case is_boolean_type(Term, Sub) of - no -> #c_literal{val=false}; - yes -> #c_literal{val=true}; - maybe -> Call - end. - %% eval_length(Call, List) -> Val. %% Evaluates the length for the prefix of List which has a known %% shape. @@ -1804,7 +1676,7 @@ opt_bool_case_guard(#c_case{arg=#c_literal{}}=Case) -> %% Case; opt_bool_case_guard(#c_case{arg=Arg,clauses=Cs0}=Case) -> - case is_safe_bool_expr(Arg, sub_new()) of + case is_safe_bool_expr(Arg) of false -> Case; true -> @@ -1945,7 +1817,7 @@ case_opt_arg(E0, Sub, Cs, LitExpr) -> {error,Cs}; false -> %% If possible, expand this variable to a previously - %% matched term. + %% constructed tuple E = case_expand_var(E0, Sub), case_opt_arg_1(E, Cs, LitExpr) end @@ -2004,13 +1876,8 @@ case_opt_compiler_generated(Core) -> case_expand_var(E, #sub{t=Tdb}) -> Key = cerl:var_name(E), case Tdb of - #{Key:=T} -> - case cerl:is_c_tuple(T) of - false -> E; - true -> T - end; - _ -> - E + #{Key:=T} -> T; + _ -> E end. %% case_opt_nomatch(E, Clauses, LitExpr) -> Clauses' @@ -2302,43 +2169,30 @@ is_simple_case_arg(_) -> false. %% Check whether the Core expression is guaranteed to return %% a boolean IF IT RETURNS AT ALL. %% -is_bool_expr(Core) -> - is_bool_expr(Core, sub_new()). -%% is_bool_expr(Core, Sub) -> true|false -%% Check whether the Core expression is guaranteed to return -%% a boolean IF IT RETURNS AT ALL. Uses type information -%% to be able to identify more expressions as booleans. -%% is_bool_expr(#c_call{module=#c_literal{val=erlang}, - name=#c_literal{val=Name},args=Args}=Call, _) -> + name=#c_literal{val=Name},args=Args}=Call) -> NumArgs = length(Args), erl_internal:comp_op(Name, NumArgs) orelse erl_internal:new_type_test(Name, NumArgs) orelse erl_internal:bool_op(Name, NumArgs) orelse will_fail(Call); is_bool_expr(#c_try{arg=E,vars=[#c_var{name=X}],body=#c_var{name=X}, - handler=#c_literal{val=false}}, Sub) -> - is_bool_expr(E, Sub); -is_bool_expr(#c_case{clauses=Cs}, Sub) -> - is_bool_expr_list(Cs, Sub); -is_bool_expr(#c_clause{body=B}, Sub) -> - is_bool_expr(B, Sub); -is_bool_expr(#c_let{vars=[V],arg=Arg,body=B}, Sub0) -> - Sub = case is_bool_expr(Arg, Sub0) of - true -> update_types(V, [bool], Sub0); - false -> Sub0 - end, - is_bool_expr(B, Sub); -is_bool_expr(#c_let{body=B}, Sub) -> - %% Binding of multiple variables. - is_bool_expr(B, Sub); -is_bool_expr(C, Sub) -> - is_boolean_type(C, Sub) =:= yes. - -is_bool_expr_list([C|Cs], Sub) -> - is_bool_expr(C, Sub) andalso is_bool_expr_list(Cs, Sub); -is_bool_expr_list([], _) -> true. + handler=#c_literal{val=false}}) -> + is_bool_expr(E); +is_bool_expr(#c_case{clauses=Cs}) -> + is_bool_expr_list(Cs); +is_bool_expr(#c_clause{body=B}) -> + is_bool_expr(B); +is_bool_expr(#c_let{body=B}) -> + is_bool_expr(B); +is_bool_expr(#c_literal{val=Val}) -> + is_boolean(Val); +is_bool_expr(_) -> false. + +is_bool_expr_list([C|Cs]) -> + is_bool_expr(C) andalso is_bool_expr_list(Cs); +is_bool_expr_list([]) -> true. %% is_safe_bool_expr(Core) -> true|false %% Check whether the Core expression ALWAYS returns a boolean @@ -2346,17 +2200,17 @@ is_bool_expr_list([], _) -> true. %% is suitable for a guard (no calls to non-guard BIFs, local %% functions, or is_record/2). %% -is_safe_bool_expr(Core, Sub) -> - is_safe_bool_expr_1(Core, Sub, cerl_sets:new()). +is_safe_bool_expr(Core) -> + is_safe_bool_expr_1(Core, cerl_sets:new()). is_safe_bool_expr_1(#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_record}, args=[A,#c_literal{val=Tag},#c_literal{val=Size}]}, - Sub, _BoolVars) when is_atom(Tag), is_integer(Size) -> - is_safe_simple(A, Sub); + _BoolVars) when is_atom(Tag), is_integer(Size) -> + is_safe_simple(A); is_safe_bool_expr_1(#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_record}}, - _Sub, _BoolVars) -> + _BoolVars) -> %% The is_record/2 BIF is NOT allowed in guards. %% The is_record/3 BIF where its second argument is not an atom or its third %% is not an integer is NOT allowed in guards. @@ -2368,49 +2222,49 @@ is_safe_bool_expr_1(#c_call{module=#c_literal{val=erlang}, is_safe_bool_expr_1(#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}, args=[A,#c_literal{val=Arity}]}, - Sub, _BoolVars) when is_integer(Arity), Arity >= 0 -> - is_safe_simple(A, Sub); + _BoolVars) when is_integer(Arity), Arity >= 0 -> + is_safe_simple(A); is_safe_bool_expr_1(#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=is_function}}, - _Sub, _BoolVars) -> + _BoolVars) -> false; is_safe_bool_expr_1(#c_call{module=#c_literal{val=erlang}, name=#c_literal{val=Name},args=Args}, - Sub, BoolVars) -> + BoolVars) -> NumArgs = length(Args), case (erl_internal:comp_op(Name, NumArgs) orelse erl_internal:new_type_test(Name, NumArgs)) andalso - is_safe_simple_list(Args, Sub) of + is_safe_simple_list(Args) of true -> true; false -> %% Boolean operators are safe if all arguments are boolean. erl_internal:bool_op(Name, NumArgs) andalso - is_safe_bool_expr_list(Args, Sub, BoolVars) + is_safe_bool_expr_list(Args, BoolVars) end; -is_safe_bool_expr_1(#c_let{vars=Vars,arg=Arg,body=B}, Sub, BoolVars) -> - case is_safe_simple(Arg, Sub) of +is_safe_bool_expr_1(#c_let{vars=Vars,arg=Arg,body=B}, BoolVars) -> + case is_safe_simple(Arg) of true -> - case {is_safe_bool_expr_1(Arg, Sub, BoolVars),Vars} of + case {is_safe_bool_expr_1(Arg, BoolVars),Vars} of {true,[#c_var{name=V}]} -> - is_safe_bool_expr_1(B, Sub, cerl_sets:add_element(V, BoolVars)); + is_safe_bool_expr_1(B, cerl_sets:add_element(V, BoolVars)); {false,_} -> - is_safe_bool_expr_1(B, Sub, BoolVars) + is_safe_bool_expr_1(B, BoolVars) end; false -> false end; -is_safe_bool_expr_1(#c_literal{val=Val}, _Sub, _) -> +is_safe_bool_expr_1(#c_literal{val=Val}, _BoolVars) -> is_boolean(Val); -is_safe_bool_expr_1(#c_var{name=V}, _Sub, BoolVars) -> +is_safe_bool_expr_1(#c_var{name=V}, BoolVars) -> cerl_sets:is_element(V, BoolVars); -is_safe_bool_expr_1(_, _, _) -> false. +is_safe_bool_expr_1(_, _) -> false. -is_safe_bool_expr_list([C|Cs], Sub, BoolVars) -> - case is_safe_bool_expr_1(C, Sub, BoolVars) of - true -> is_safe_bool_expr_list(Cs, Sub, BoolVars); +is_safe_bool_expr_list([C|Cs], BoolVars) -> + case is_safe_bool_expr_1(C, BoolVars) of + true -> is_safe_bool_expr_list(Cs, BoolVars); false -> false end; -is_safe_bool_expr_list([], _, _) -> true. +is_safe_bool_expr_list([], _) -> true. %% simplify_let(Let, Sub) -> Expr | impossible %% If the argument part of an let contains a complex expression, such @@ -2785,7 +2639,7 @@ opt_simple_let_2(Let0, Vs0, Arg0, Body, PrevBody, Sub) -> %% with exported variables, but the return value is %% ignored). We can remove the first variable and the %% the first value returned from the 'let' argument. - Arg2 = remove_first_value(Arg1, Sub), + Arg2 = remove_first_value(Arg1), Let1 = Let0#c_let{vars=Vars,arg=Arg2,body=Body}, post_opt_let(Let1, Sub); true -> @@ -2805,36 +2659,36 @@ post_opt_let(Let0, Sub) -> opt_build_stacktrace(Let1). -%% remove_first_value(Core0, Sub) -> Core. +%% remove_first_value(Core0) -> Core. %% Core0 is an expression that returns at least two values. %% Remove the first value returned from Core0. -remove_first_value(#c_values{es=[V|Vs]}, Sub) -> +remove_first_value(#c_values{es=[V|Vs]}) -> Values = core_lib:make_values(Vs), - case is_safe_simple(V, Sub) of + case is_safe_simple(V) of false -> #c_seq{arg=V,body=Values}; true -> Values end; -remove_first_value(#c_case{clauses=Cs0}=Core, Sub) -> - Cs = remove_first_value_cs(Cs0, Sub), +remove_first_value(#c_case{clauses=Cs0}=Core) -> + Cs = remove_first_value_cs(Cs0), Core#c_case{clauses=Cs}; -remove_first_value(#c_receive{clauses=Cs0,action=Act0}=Core, Sub) -> - Cs = remove_first_value_cs(Cs0, Sub), - Act = remove_first_value(Act0, Sub), +remove_first_value(#c_receive{clauses=Cs0,action=Act0}=Core) -> + Cs = remove_first_value_cs(Cs0), + Act = remove_first_value(Act0), Core#c_receive{clauses=Cs,action=Act}; -remove_first_value(#c_let{body=B}=Core, Sub) -> - Core#c_let{body=remove_first_value(B, Sub)}; -remove_first_value(#c_seq{body=B}=Core, Sub) -> - Core#c_seq{body=remove_first_value(B, Sub)}; -remove_first_value(#c_primop{}=Core, _Sub) -> +remove_first_value(#c_let{body=B}=Core) -> + Core#c_let{body=remove_first_value(B)}; +remove_first_value(#c_seq{body=B}=Core) -> + Core#c_seq{body=remove_first_value(B)}; +remove_first_value(#c_primop{}=Core) -> Core; -remove_first_value(#c_call{}=Core, _Sub) -> +remove_first_value(#c_call{}=Core) -> Core. -remove_first_value_cs(Cs, Sub) -> - [C#c_clause{body=remove_first_value(B, Sub)} || +remove_first_value_cs(Cs) -> + [C#c_clause{body=remove_first_value(B)} || #c_clause{body=B}=C <- Cs]. %% maybe_suppress_warnings(Arg, #c_var{}, PreviousBody) -> Arg' @@ -2962,54 +2816,6 @@ move_case_into_arg(Expr, _) -> Expr. %%% -%%% Retrieving information about types. -%%% - --spec get_type(cerl:cerl(), #sub{}) -> type_info() | 'none'. - -get_type(#c_var{name=V}, #sub{t=Tdb}) -> - case Tdb of - #{V:=Type} -> Type; - _ -> none - end; -get_type(C, _) -> - case cerl:type(C) of - binary -> C; - map -> C; - _ -> - case cerl:is_data(C) of - true -> C; - false -> none - end - end. - --spec is_boolean_type(cerl:cerl(), sub()) -> yes_no_maybe(). - -is_boolean_type(Var, Sub) -> - case get_type(Var, Sub) of - none -> - maybe; - bool -> - yes; - C -> - B = cerl:is_c_atom(C) andalso - is_boolean(cerl:atom_val(C)), - yes_no(B) - end. - --spec is_int_type(cerl:cerl(), sub()) -> yes_no_maybe(). - -is_int_type(Var, Sub) -> - case get_type(Var, Sub) of - none -> maybe; - integer -> yes; - C -> yes_no(cerl:is_c_int(C)) - end. - -yes_no(true) -> yes; -yes_no(false) -> no. - -%%% %%% Update type information. %%% @@ -3020,70 +2826,14 @@ update_let_types(_Vs, _Arg, Sub) -> %% that returns multiple values. Sub. -update_let_types_1([#c_var{}=V|Vs], [A|As], Sub0) -> - Sub = update_types_from_expr(V, A, Sub0), +update_let_types_1([#c_var{name=V}|Vs], [A|As], Sub0) -> + Sub = update_types(V, A, Sub0), update_let_types_1(Vs, As, Sub); update_let_types_1([], [], Sub) -> Sub. -update_types_from_expr(V, Expr, Sub) -> - Type = extract_type(Expr, Sub), - update_types(V, [Type], Sub). - -extract_type(#c_call{module=#c_literal{val=erlang}, - name=#c_literal{val=Name}, - args=Args}=Call, Sub) -> - case returns_integer(Name, Args) of - true -> integer; - false -> extract_type_1(Call, Sub) - end; -extract_type(Expr, Sub) -> - extract_type_1(Expr, Sub). - -extract_type_1(Expr, Sub) -> - case is_bool_expr(Expr, Sub) of - false -> Expr; - true -> bool - end. - -returns_integer('band', [_,_]) -> true; -returns_integer('bnot', [_]) -> true; -returns_integer('bor', [_,_]) -> true; -returns_integer('bxor', [_,_]) -> true; -returns_integer(bit_size, [_]) -> true; -returns_integer('bsl', [_,_]) -> true; -returns_integer('bsr', [_,_]) -> true; -returns_integer(byte_size, [_]) -> true; -returns_integer(ceil, [_]) -> true; -returns_integer('div', [_,_]) -> true; -returns_integer(floor, [_]) -> true; -returns_integer(length, [_]) -> true; -returns_integer('rem', [_,_]) -> true; -returns_integer('round', [_]) -> true; -returns_integer(size, [_]) -> true; -returns_integer(tuple_size, [_]) -> true; -returns_integer(trunc, [_]) -> true; -returns_integer(_, _) -> false. - -%% update_types(Expr, Pattern, Sub) -> Sub' -%% Update the type database. - --spec update_types(cerl:c_var(), [type_info()], sub()) -> sub(). - -update_types(#c_var{name=V}, Pat, #sub{t=Tdb0}=Sub) -> - Tdb = update_types_1(V, Pat, Tdb0), - Sub#sub{t=Tdb}. - -update_types_1(V, [#c_tuple{}=P], Types) -> - Types#{V=>P}; -update_types_1(V, [#c_literal{val=Bool}], Types) when is_boolean(Bool) -> - Types#{V=>bool}; -update_types_1(V, [#c_fun{vars=Vars}], Types) -> - Types#{V=>{'fun',length(Vars)}}; -update_types_1(V, [#c_var{name={_,Arity}}], Types) -> - Types#{V=>{'fun',Arity}}; -update_types_1(V, [Type], Types) when is_atom(Type) -> - Types#{V=>Type}; -update_types_1(_, _, Types) -> Types. +update_types(V, #c_tuple{}=P, #sub{t=Tdb}=Sub) -> + Sub#sub{t=Tdb#{V=>P}}; +update_types(_, _, Sub) -> Sub. %% kill_types(V, Tdb) -> Tdb' %% Kill any entries that references the variable, @@ -3099,10 +2849,6 @@ kill_types2(V, [{_,#c_tuple{}=Tuple}=Entry|Tdb]) -> false -> [Entry|kill_types2(V, Tdb)]; true -> kill_types2(V, Tdb) end; -kill_types2(V, [{_, {'fun',_}}=Entry|Tdb]) -> - [Entry|kill_types2(V, Tdb)]; -kill_types2(V, [{_,Atom}=Entry|Tdb]) when is_atom(Atom) -> - [Entry|kill_types2(V, Tdb)]; kill_types2(_, []) -> []. %% copy_type(DestVar, SrcVar, Tdb) -> Tdb' diff --git a/lib/compiler/src/v3_kernel.erl b/lib/compiler/src/v3_kernel.erl index e2b8787224..49fb66126f 100644 --- a/lib/compiler/src/v3_kernel.erl +++ b/lib/compiler/src/v3_kernel.erl @@ -81,13 +81,17 @@ -export([module/2,format_error/1]). --import(lists, [map/2,foldl/3,foldr/3,mapfoldl/3,splitwith/2,member/2, - keyfind/3,partition/2,droplast/1,last/1,sort/1,reverse/1]). +-import(lists, [droplast/1,flatten/1,foldl/3,foldr/3, + map/2,mapfoldl/3,member/2, + keyfind/3,keyreplace/4, + last/1,partition/2,reverse/1, + splitwith/2,sort/1]). -import(ordsets, [add_element/2,del_element/2,union/2,union/1,subtract/2]). -import(cerl, [c_tuple/1]). -include("core_parse.hrl"). -include("v3_kernel.hrl"). +-define(EXPAND_MAX_SIZE_SEGMENT, 1024). %% These are not defined in v3_kernel.hrl. get_kanno(Kthing) -> element(2, Kthing). @@ -120,15 +124,19 @@ copy_anno(Kdst, Ksrc) -> funs=[], %Fun functions free=#{}, %Free variables ws=[] :: [warning()], %Warnings. - guard_refc=0}). %> 0 means in guard + guard_refc=0, %> 0 means in guard + no_shared_fun_wrappers=false :: boolean() + }). -spec module(cerl:c_module(), [compile:option()]) -> {'ok', #k_mdef{}, [warning()]}. -module(#c_module{anno=A,name=M,exports=Es,attrs=As,defs=Fs}, _Options) -> +module(#c_module{anno=A,name=M,exports=Es,attrs=As,defs=Fs}, Options) -> Kas = attributes(As), Kes = map(fun (#c_var{name={_,_}=Fname}) -> Fname end, Es), - St0 = #kern{}, + NoSharedFunWrappers = proplists:get_bool(no_shared_fun_wrappers, + Options), + St0 = #kern{no_shared_fun_wrappers=NoSharedFunWrappers}, {Kfs,St} = mapfoldl(fun function/2, St0, Fs), {ok,#k_mdef{anno=A,name=M#c_literal.val,exports=Kes,attributes=Kas, body=Kfs ++ St#kern.funs},lists:sort(St#kern.ws)}. @@ -716,16 +724,27 @@ gexpr_test_add(Ke, St0) -> %% expr(Cexpr, Sub, State) -> {Kexpr,[PreKexpr],State}. %% Convert a Core expression, flattening it at the same time. -expr(#c_var{anno=A,name={_Name,Arity}}=Fname, Sub, St) -> - %% A local in an expression. - %% For now, these are wrapped into a fun by reverse - %% eta-conversion, but really, there should be exactly one - %% such "lambda function" for each escaping local name, - %% instead of one for each occurrence as done now. +expr(#c_var{anno=A0,name={Name,Arity}}=Fname, Sub, St) -> Vs = [#c_var{name=list_to_atom("V" ++ integer_to_list(V))} || - V <- integers(1, Arity)], - Fun = #c_fun{anno=A,vars=Vs,body=#c_apply{anno=A,op=Fname,args=Vs}}, - expr(Fun, Sub, St); + V <- integers(1, Arity)], + case St#kern.no_shared_fun_wrappers of + false -> + %% Generate a (possibly shared) wrapper function for calling + %% this function. + Wrapper0 = ["-fun.",atom_to_list(Name),"/",integer_to_list(Arity),"-"], + Wrapper = list_to_atom(flatten(Wrapper0)), + Id = {id,{0,0,Wrapper}}, + A = keyreplace(id, 1, A0, Id), + Fun = #c_fun{anno=A,vars=Vs,body=#c_apply{anno=A,op=Fname,args=Vs}}, + expr(Fun, Sub, St); + true -> + %% For backward compatibility with OTP 22 and earlier, + %% use the pre-generated name for the fun wrapper. + %% There will be one wrapper function for each occurrence + %% of `fun F/A`. + Fun = #c_fun{anno=A0,vars=Vs,body=#c_apply{anno=A0,op=Fname,args=Vs}}, + expr(Fun, Sub, St) + end; expr(#c_var{anno=A,name=V}, Sub, St) -> {#k_var{anno=A,name=get_vsub(V, Sub)},[],St}; expr(#c_literal{anno=A,val=V}, _Sub, St) -> @@ -1152,7 +1171,7 @@ validate_bin_element_size(#k_int{val=V}) when V >= 0 -> ok; validate_bin_element_size(#k_atom{val=all}) -> ok; validate_bin_element_size(#k_atom{val=undefined}) -> ok; validate_bin_element_size(_) -> throw(bad_element_size). - + %% atomic_list([Cexpr], Sub, State) -> {[Kexpr],[PreKexpr],State}. atomic_list(Ces, Sub, St) -> @@ -1278,14 +1297,63 @@ pattern_bin_1([#c_bitstr{anno=A,val=E0,size=S0,unit=U,type=T,flags=Fs}|Es0], _ -> Isub0 end, {Es,{Isub,Osub},St3} = pattern_bin_1(Es0, Isub1, Osub1, St2), - {#k_bin_seg{anno=A,size=S, - unit=U0, - type=cerl:concrete(T), - flags=Fs0, - seg=E,next=Es}, - {Isub,Osub},St3}; + {build_bin_seg(A, S, U0, cerl:concrete(T), Fs0, E, Es),{Isub,Osub},St3}; pattern_bin_1([], Isub, Osub, St) -> {#k_bin_end{},{Isub,Osub},St}. +%% build_bin_seg(Anno, Size, Unit, Type, Flags, Seg, Next) -> #k_bin_seg{}. +%% This function normalizes literal integers with size > 8 and literal +%% utf8 segments into integers with size = 8 (and potentially an integer +%% with size less than 8 at the end). This is so further optimizations +%% have a normalized view of literal integers, allowing us to generate +%% more literals and group more clauses. Those integers may be "squeezed" +%% later into the largest integer possible. +%% +build_bin_seg(A, #k_int{val=Bits} = Sz, U, integer=Type, [unsigned,big]=Flags, #k_literal{val=Int}=Seg, Next) -> + Size = Bits * U, + case integer_fits_and_is_expandable(Int, Size) of + true -> build_bin_seg_integer_recur(A, Size, Int, Next); + false -> #k_bin_seg{anno=A,size=Sz,unit=U,type=Type,flags=Flags,seg=Seg,next=Next} + end; +build_bin_seg(A, Sz, U, utf8=Type, [unsigned,big]=Flags, #k_literal{val=Utf8} = Seg, Next) -> + case utf8_fits(Utf8) of + {Int, Bits} -> build_bin_seg_integer_recur(A, Bits, Int, Next); + error -> #k_bin_seg{anno=A,size=Sz,unit=U,type=Type,flags=Flags,seg=Seg,next=Next} + end; +build_bin_seg(A, Sz, U, Type, Flags, Seg, Next) -> + #k_bin_seg{anno=A,size=Sz,unit=U,type=Type,flags=Flags,seg=Seg,next=Next}. + +build_bin_seg_integer_recur(A, Bits, Val, Next) when Bits > 8 -> + NextBits = Bits - 8, + NextVal = Val band ((1 bsl NextBits) - 1), + Last = build_bin_seg_integer_recur(A, NextBits, NextVal, Next), + build_bin_seg_integer(A, 8, Val bsr NextBits, Last); + +build_bin_seg_integer_recur(A, Bits, Val, Next) -> + build_bin_seg_integer(A, Bits, Val, Next). + +build_bin_seg_integer(A, Bits, Val, Next) -> + Sz = #k_int{anno=A,val=Bits}, + Seg = #k_literal{anno=A,val=Val}, + #k_bin_seg{anno=A,size=Sz,unit=1,type=integer,flags=[unsigned,big],seg=Seg,next=Next}. + +integer_fits_and_is_expandable(Int, Size) when 0 < Size, Size =< ?EXPAND_MAX_SIZE_SEGMENT -> + case <<Int:Size>> of + <<Int:Size>> -> true; + _ -> false + end; +integer_fits_and_is_expandable(_Int, _Size) -> + false. + +utf8_fits(Utf8) -> + try + Bin = <<Utf8/utf8>>, + Bits = bit_size(Bin), + <<Int:Bits>> = Bin, + {Int, Bits} + catch + _:_ -> error + end. + %% pattern_list([Cexpr], Sub, State) -> {[Kexpr],Sub,State}. pattern_list(Ces, Sub, St) -> @@ -1535,7 +1603,7 @@ maybe_add_warning(Ke, MatchAnno, St) -> 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 @@ -1743,27 +1811,10 @@ do_combine_lit_pat(#k_tuple{anno=A,es=Es0}) -> do_combine_lit_pat(_) -> throw(not_possible). -combine_bin_segs(#k_bin_seg{size=Size0,unit=Unit,type=integer, - flags=[unsigned,big],seg=Seg,next=Next}) -> - #k_literal{val=Size1} = do_combine_lit_pat(Size0), - #k_literal{val=Int} = do_combine_lit_pat(Seg), - Size = Size1 * Unit, - if - 0 < Size, Size < 64 -> - Bin = <<Int:Size>>, - case Bin of - <<Int:Size>> -> - NextBin = combine_bin_segs(Next), - <<Bin/bits,NextBin/bits>>; - _ -> - %% The integer Int does not fit in the segment, - %% thus it will not match. - throw(not_possible) - end; - true -> - %% Avoid creating huge binary literals. - throw(not_possible) - end; +combine_bin_segs(#k_bin_seg{size=#k_int{val=8},unit=1,type=integer, + flags=[unsigned,big],seg=#k_literal{val=Int},next=Next}) + when is_integer(Int), 0 =< Int, Int =< 255 -> + <<Int,(combine_bin_segs(Next))/bits>>; combine_bin_segs(#k_bin_end{}) -> <<>>; combine_bin_segs(_) -> @@ -1833,11 +1884,10 @@ handle_bin_con_not_possible([]) -> []. select_bin_int([#iclause{pats=[#k_bin_seg{anno=A,type=integer, size=#k_int{val=Bits0}=Sz,unit=U, flags=Fl,seg=#k_literal{val=Val}, - next=N}|Ps]}=C|Cs0]) - when is_integer(Val) -> + next=N}|Ps]}=C|Cs0]) -> Bits = U * Bits0, if - Bits > 1024 -> throw(not_possible); %Expands the code too much. + Bits > ?EXPAND_MAX_SIZE_SEGMENT -> throw(not_possible); %Expands the code too much. true -> ok end, select_assert_match_possible(Bits, Val, Fl), @@ -1848,16 +1898,6 @@ select_bin_int([#iclause{pats=[#k_bin_seg{anno=A,type=integer, end, Cs = select_bin_int_1(Cs0, Bits, Fl, Val), [{k_bin_int,[C#iclause{pats=[P|Ps]}|Cs]}]; -select_bin_int([#iclause{pats=[#k_bin_seg{anno=A,type=utf8, - flags=[unsigned,big]=Fl, - seg=#k_literal{val=Val0}, - next=N}|Ps]}=C|Cs0]) - when is_integer(Val0) -> - {Val,Bits} = select_utf8(Val0), - P = #k_bin_int{anno=A,size=#k_int{val=Bits},unit=1, - flags=Fl,val=Val,next=N}, - Cs = select_bin_int_1(Cs0, Bits, Fl, Val), - [{k_bin_int,[C#iclause{pats=[P|Ps]}|Cs]}]; select_bin_int(_) -> throw(not_possible). select_bin_int_1([#iclause{pats=[#k_bin_seg{anno=A,type=integer, @@ -1872,18 +1912,6 @@ select_bin_int_1([#iclause{pats=[#k_bin_seg{anno=A,type=integer, end, P = #k_bin_int{anno=A,size=Sz,unit=U,flags=Fl,val=Val,next=N}, [C#iclause{pats=[P|Ps]}|select_bin_int_1(Cs, Bits, Fl, Val)]; -select_bin_int_1([#iclause{pats=[#k_bin_seg{anno=A,type=utf8, - flags=Fl, - seg=#k_literal{val=Val0}, - next=N}|Ps]}=C|Cs], - Bits, Fl, Val) when is_integer(Val0) -> - case select_utf8(Val0) of - {Val,Bits} -> ok; - {_,_} -> throw(not_possible) - end, - P = #k_bin_int{anno=A,size=#k_int{val=Bits},unit=1, - flags=[unsigned,big],val=Val,next=N}, - [C#iclause{pats=[P|Ps]}|select_bin_int_1(Cs, Bits, Fl, Val)]; select_bin_int_1([], _, _, _) -> []; select_bin_int_1(_, _, _, _) -> throw(not_possible). @@ -1909,17 +1937,6 @@ match_fun(Val) -> {match,Bs} end. -select_utf8(Val0) -> - try - Bin = <<Val0/utf8>>, - Size = bit_size(Bin), - <<Val:Size>> = Bin, - {Val,Size} - catch - error:_ -> - throw(not_possible) - end. - %% match_value([Var], Con, [Clause], Default, State) -> {SelectExpr,State}. %% At this point all the clauses have the same constructor, we must %% now separate them according to value. @@ -2039,7 +2056,8 @@ match_clause([U|Us], [C|_]=Cs0, Def, St0) -> {Match0,Vs,St1} = get_match(get_con(Cs0), St0), Match = sub_size_var(Match0, Cs0), {Cs1,St2} = new_clauses(Cs0, U, St1), - {B,St3} = match(Vs ++ Us, Cs1, Def, St2), + Cs2 = squeeze_clauses_by_bin_integer_count(Cs1, []), + {B,St3} = match(Vs ++ Us, Cs2, Def, St2), {#k_val_clause{anno=Anno,val=Match,body=B},St3}. sub_size_var(#k_bin_seg{size=#k_var{name=Name}=Kvar}=BinSeg, [#iclause{isub=Sub}|_]) -> @@ -2109,6 +2127,102 @@ new_clauses(Cs0, U, St) -> end, Cs0), {Cs1,St}. +%% group and squeeze +%% The goal of those functions is to group subsequent integer k_bin_seg +%% literals by count so we can leverage bs_get_integer_16 whenever possible. +%% +%% The priority is to create large groups. So if we have three clauses matching +%% on 16-bits/16-bits/8-bits, we will first have a single 8-bits match for all +%% three clauses instead of clauses (one with 16 and another with 8). But note +%% the algorithm is recursive, so the remaining 8-bits for the first two clauses +%% will be grouped next. +%% +%% We also try to not create too large groups. If we have too many clauses, +%% it is preferrable to match on 8-bits, select a branch, then match on the +%% next 8-bits, rather than match on 16-bits which would force us to have +%% to select to many values at the same time, which would not be efficient. +%% +%% Another restriction is that we create groups only if the end of the +%% group is a variadic clause or the end of the binary. That's because +%% if we have 16-bits/16-bits/catch-all, breaking it into a 16-bits lookup +%% will make the catch-all more expensive. +%% +%% Clauses are grouped in reverse when squeezing and then flattened and +%% re-reversed at the end. +squeeze_clauses_by_bin_integer_count([Clause | Clauses], Acc) -> + case clause_count_bin_integer_segments(Clause) of + {literal, N} -> squeeze_clauses_by_bin_integer_count(Clauses, N, 1, [Clause], Acc); + _ -> squeeze_clauses_by_bin_integer_count(Clauses, [[Clause] | Acc]) + end; +squeeze_clauses_by_bin_integer_count(_, Acc) -> + flat_reverse(Acc, []). + +squeeze_clauses_by_bin_integer_count([], N, Count, GroupAcc, Acc) -> + Squeezed = squeeze_clauses(GroupAcc, fix_count_without_variadic_segment(N), Count), + flat_reverse([Squeezed | Acc], []); +squeeze_clauses_by_bin_integer_count([#iclause{pats=[#k_bin_end{} | _]} = Clause], N, Count, GroupAcc, Acc) -> + Squeezed = squeeze_clauses(GroupAcc, fix_count_without_variadic_segment(N), Count), + flat_reverse([[Clause | Squeezed] | Acc], []); +squeeze_clauses_by_bin_integer_count([Clause | Clauses], N, Count, GroupAcc, Acc) -> + case clause_count_bin_integer_segments(Clause) of + {literal, NewN} -> + squeeze_clauses_by_bin_integer_count(Clauses, min(N, NewN), Count + 1, [Clause | GroupAcc], Acc); + + {variadic, NewN} when NewN =< N -> + Squeezed = squeeze_clauses(GroupAcc, NewN, Count), + squeeze_clauses_by_bin_integer_count(Clauses, [[Clause | Squeezed] | Acc]); + + _ -> + squeeze_clauses_by_bin_integer_count(Clauses, [[Clause | GroupAcc] | Acc]) + end. + +clause_count_bin_integer_segments(#iclause{pats=[#k_bin_seg{seg=#k_literal{}} = BinSeg | _]}) -> + count_bin_integer_segments(BinSeg, 0); +clause_count_bin_integer_segments(#iclause{pats=[#k_bin_seg{size=#k_int{val=Size},unit=Unit, + type=integer,flags=[unsigned,big], seg=#k_var{}} | _]}) + when ((Size * Unit) rem 8) =:= 0 -> + {variadic, (Size * Unit) div 8}; +clause_count_bin_integer_segments(_) -> + error. + +count_bin_integer_segments(#k_bin_seg{size=#k_int{val=8},unit=1,type=integer,flags=[unsigned,big], + seg=#k_literal{val=Int},next=Next}, Count) when is_integer(Int), 0 =< Int, Int =< 255 -> + count_bin_integer_segments(Next, Count + 1); +count_bin_integer_segments(_, Count) when Count > 0 -> + {literal, Count}; +count_bin_integer_segments(_, _Count) -> + error. + +%% Since 4 bytes in on 32-bits systems are bignums, we convert +%% anything more than 3 into 2 bytes lookup. The goal is to convert +%% any multi-clause segment into 2-byte lookups with a potential +%% 3 byte lookup at the end. +fix_count_without_variadic_segment(N) when N > 3 -> 2; +fix_count_without_variadic_segment(N) -> N. + +%% If we have more than 16 clauses, then it is better +%% to branch multiple times than getting a large integer. +%% We also abort if we have nothing to squeeze. +squeeze_clauses(Clauses, Size, Count) when Count >= 16; Size == 1 -> Clauses; +squeeze_clauses(Clauses, Size, _Count) -> squeeze_clauses(Clauses, Size). + +squeeze_clauses([#iclause{pats=[#k_bin_seg{seg=#k_literal{}} = BinSeg | Pats]} = Clause | Clauses], Size) -> + [Clause#iclause{pats=[squeeze_segments(BinSeg, 0, 0, Size) | Pats]} | + squeeze_clauses(Clauses, Size)]; +squeeze_clauses([], _Size) -> + []. + +squeeze_segments(#k_bin_seg{size=Sz, seg=#k_literal{val=Val}=Lit} = BinSeg, Acc, Size, 1) -> + BinSeg#k_bin_seg{size=Sz#k_int{val=Size + 8}, seg=Lit#k_literal{val=(Acc bsl 8) bor Val}}; +squeeze_segments(#k_bin_seg{seg=#k_literal{val=Val},next=Next}, Acc, Size, Count) -> + squeeze_segments(Next, (Acc bsl 8) bor Val, Size + 8, Count - 1). + +flat_reverse([Head | Tail], Acc) -> flat_reverse(Tail, flat_reverse_1(Head, Acc)); +flat_reverse([], Acc) -> Acc. + +flat_reverse_1([Head | Tail], Acc) -> flat_reverse_1(Tail, [Head | Acc]); +flat_reverse_1([], Acc) -> Acc. + %% build_guard([GuardClause]) -> GuardExpr. build_guard([]) -> fail; @@ -2446,8 +2560,21 @@ uexpr(Lit, {break,Rs0}, St0) -> {#k_put{anno=#k{us=Used,ns=lit_list_vars(Rs),a=get_kanno(Lit)}, arg=Lit,ret=Rs},Used,St1}. -add_local_function(_, #kern{funs=ignore}=St) -> St; -add_local_function(F, #kern{funs=Funs}=St) -> St#kern{funs=[F|Funs]}. +add_local_function(_, #kern{funs=ignore}=St) -> + St; +add_local_function(#k_fdef{func=Name,arity=Arity}=F, #kern{funs=Funs}=St) -> + case is_defined(Name, Arity, Funs) of + false -> + St#kern{funs=[F|Funs]}; + true -> + St + end. + +is_defined(Name, Arity, [#k_fdef{func=Name,arity=Arity}|_]) -> + true; +is_defined(Name, Arity, [#k_fdef{}|T]) -> + is_defined(Name, Arity, T); +is_defined(_, _, []) -> false. %% Make a #k_fdef{}, making sure that the body is always a #k_match{}. make_fdef(Anno, Name, Arity, Vs, #k_match{}=Body) -> |