aboutsummaryrefslogtreecommitdiffstats
path: root/lib/compiler/src
diff options
context:
space:
mode:
authorBjörn Gustavsson <[email protected]>2018-08-21 08:58:15 +0200
committerBjörn Gustavsson <[email protected]>2018-09-17 06:41:02 +0200
commit70bb14fe3fd27e8c429bf402e364b144a5807d9a (patch)
treec8c86337d5324e900ce483eb49b32227a3a314d2 /lib/compiler/src
parentec1f35c9f52be894ba295b9a48237020855e3c46 (diff)
downloadotp-70bb14fe3fd27e8c429bf402e364b144a5807d9a.tar.gz
otp-70bb14fe3fd27e8c429bf402e364b144a5807d9a.tar.bz2
otp-70bb14fe3fd27e8c429bf402e364b144a5807d9a.zip
Add beam_ssa_dead.erl
Add beam_ssa_dead to perform the main optimizations done by beam_dead: * Shortcut branches that jump to another block with a branch. If it can be seen that the second branch will always branch to a specific block, replace the target of the first branch. * Combined nested sequences of '=:=' tests and switch instructions operating on the same variable to a single switch. Diffing the compiler output, it seems that beam_ssa_dead finds many more opportunities for optimizations than beam_dead, although it does not find all opportunities that beam_dead does. In total, beam_ssa_dead is such improvement over beam_dead that there is no reason to keep beam_dead as well as beam_ssa_dead. Note that beam_ssa_dead does not attempt to optimize away redundant bs_context_binary instructions, because that instruction will be superseded by new instructions in the near future.
Diffstat (limited to 'lib/compiler/src')
-rw-r--r--lib/compiler/src/Makefile2
-rw-r--r--lib/compiler/src/beam_ssa_dead.erl1001
-rw-r--r--lib/compiler/src/beam_ssa_opt.erl20
-rw-r--r--lib/compiler/src/compile.erl1
-rw-r--r--lib/compiler/src/compiler.app.src1
5 files changed, 1022 insertions, 3 deletions
diff --git a/lib/compiler/src/Makefile b/lib/compiler/src/Makefile
index bd35f20442..71fcfc2d2a 100644
--- a/lib/compiler/src/Makefile
+++ b/lib/compiler/src/Makefile
@@ -64,6 +64,7 @@ MODULES = \
beam_split \
beam_ssa \
beam_ssa_codegen \
+ beam_ssa_dead \
beam_ssa_lint \
beam_ssa_opt \
beam_ssa_pp \
@@ -194,6 +195,7 @@ $(EBIN)/beam_listing.beam: core_parse.hrl v3_kernel.hrl beam_ssa.hrl
$(EBIN)/beam_kernel_to_ssa.beam: v3_kernel.hrl beam_ssa.hrl
$(EBIN)/beam_ssa.beam: beam_ssa.hrl
$(EBIN)/beam_ssa_codegen.beam: beam_ssa.hrl
+$(EBIN)/beam_ssa_dead.beam: beam_ssa.hrl
$(EBIN)/beam_ssa_lint.beam: beam_ssa.hrl
$(EBIN)/beam_ssa_opt.beam: beam_ssa.hrl
$(EBIN)/beam_ssa_pp.beam: beam_ssa.hrl
diff --git a/lib/compiler/src/beam_ssa_dead.erl b/lib/compiler/src/beam_ssa_dead.erl
new file mode 100644
index 0000000000..7cdb4315fe
--- /dev/null
+++ b/lib/compiler/src/beam_ssa_dead.erl
@@ -0,0 +1,1001 @@
+%%
+%% %CopyrightBegin%
+%%
+%% Copyright Ericsson AB 2018. All Rights Reserved.
+%%
+%% Licensed under the Apache License, Version 2.0 (the "License");
+%% you may not use this file except in compliance with the License.
+%% You may obtain a copy of the License at
+%%
+%% http://www.apache.org/licenses/LICENSE-2.0
+%%
+%% Unless required by applicable law or agreed to in writing, software
+%% distributed under the License is distributed on an "AS IS" BASIS,
+%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+%% See the License for the specific language governing permissions and
+%% limitations under the License.
+%%
+%% %CopyrightEnd%
+%%
+%% Dead code is code that is executed but has no effect. This
+%% optimization pass either removes dead code or jumps around it,
+%% potentially making it unreachable so that it can be dropped
+%% the next time beam_ssa:linearize/1 is called.
+%%
+
+-module(beam_ssa_dead).
+-export([opt/1]).
+
+-include("beam_ssa.hrl").
+-import(lists, [append/1,last/1,member/2,takewhile/2,reverse/1]).
+
+-type used_vars() :: #{beam_ssa:label():=ordsets:ordset(beam_ssa:var_name())}.
+
+-type basic_type_test() :: atom() | {'is_tagged_tuple',pos_integer(),atom()}.
+-type type_test() :: basic_type_test() | {'not',basic_type_test()}.
+-type op_name() :: atom().
+-type basic_rel_op() :: {op_name(),beam_ssa:b_var(),beam_ssa:value()} |
+ {basic_type_test(),beam_ssa:value()}.
+-type rel_op() :: {op_name(),beam_ssa:b_var(),beam_ssa:value()} |
+ {type_test(),beam_ssa:value()}.
+
+-record(st,
+ {bs :: beam_ssa:block_map(),
+ us :: used_vars(),
+ skippable :: #{beam_ssa:label():='true'},
+ rel_op=none :: 'none' | rel_op(),
+ target=any :: 'any' | 'one_way' | beam_ssa:label()
+ }).
+
+-spec opt([{Label0,Block0}]) -> [{Label,Block}] when
+ Label0 :: beam_ssa:label(),
+ Block0 :: beam_ssa:b_blk(),
+ Label :: beam_ssa:label(),
+ Block :: beam_ssa:b_blk().
+
+opt(Linear) ->
+ {Used,Skippable} = used_vars(Linear),
+ Blocks0 = maps:from_list(Linear),
+ St0 = #st{bs=Blocks0,us=Used,skippable=Skippable},
+ St = shortcut_opt(St0),
+ #st{bs=Blocks} = combine_eqs(St),
+ beam_ssa:linearize(Blocks).
+
+%%%
+%%% Shortcut br/switch targets.
+%%%
+%%% A br/switch may branch to another br/switch that in turn always
+%%% branches to another target. Rewrite br/switch to refer to the
+%%% ultimate targets directly. That will save execution time, but
+%%% could also reduce the size of the code if some of the original
+%%% targets become unreachable and be deleted.
+%%%
+%%% When rewriting branches, we must be careful not to skip instructions
+%%% that have side effects or that bind variables that will be used
+%%% at the new target.
+%%%
+%%% We must also avoid branching to phi nodes. The reason is
+%%% twofold. First, we might create a critical edge which is strictly
+%%% forbidden. Second, there will be a branch from a block that is not
+%%% listed in the list of predecessors in the phi node. Those
+%%% limitations could probably be overcome, but it is not clear how
+%%% much that would improve the code.
+%%%
+
+shortcut_opt(#st{bs=Blocks}=St) ->
+ %% Processing the blocks in reverse post order seems to give more
+ %% opportunities for optimizations compared to post order. (Based on
+ %% running scripts/diffable with both PO and RPO and looking at
+ %% the diff.)
+ Ls = beam_ssa:rpo(Blocks),
+ shortcut_opt(Ls, #{from=>0}, St).
+
+shortcut_opt([L|Ls], Bs0, #st{bs=Blocks0}=St) ->
+ #b_blk{is=Is,last=Last0} = Blk0 = get_block(L, St),
+ Bs = Bs0#{from:=L},
+ case shortcut_terminator(Last0, Is, Bs, St) of
+ Last0 ->
+ %% No change. No need to update the block.
+ shortcut_opt(Ls, Bs, St);
+ Last ->
+ %% The terminator was simplified in some way.
+ %% Update the block.
+ Blk = Blk0#b_blk{last=Last},
+ Blocks = Blocks0#{L=>Blk},
+ shortcut_opt(Ls, Bs, St#st{bs=Blocks})
+ end;
+shortcut_opt([], _, St) -> St.
+
+shortcut_terminator(#b_br{bool=#b_literal{val=true},succ=Succ0},
+ _Is, Bs, St0) ->
+ St = St0#st{rel_op=none},
+ shortcut(Succ0, Bs, St);
+shortcut_terminator(#b_br{bool=#b_var{}=Bool,succ=Succ0,fail=Fail0}=Br,
+ Is, Bs, St0) ->
+ St = St0#st{target=one_way},
+ RelOp = get_rel_op(Bool, Is),
+ SuccBs = bind_var(Bool, #b_literal{val=true}, Bs),
+ BrSucc = shortcut(Succ0, SuccBs, St#st{rel_op=RelOp}),
+ FailBs = bind_var(Bool, #b_literal{val=false}, Bs),
+ BrFail = shortcut(Fail0, FailBs, St#st{rel_op=invert_op(RelOp)}),
+ case {BrSucc,BrFail} of
+ {#b_br{bool=#b_literal{val=true},succ=Succ},
+ #b_br{bool=#b_literal{val=true},succ=Fail}}
+ when Succ =/= Succ0; Fail =/= Fail0 ->
+ %% One or both of the targets were cut short.
+ beam_ssa:normalize(Br#b_br{succ=Succ,fail=Fail});
+ {_,_} ->
+ %% No change.
+ Br
+ end;
+shortcut_terminator(#b_switch{arg=Bool,list=List0}=Sw, _Is, Bs, St) ->
+ List = shortcut_switch(List0, Bool, Bs, St),
+ beam_ssa:normalize(Sw#b_switch{list=List});
+shortcut_terminator(Last, _Is, _Bs, _St) ->
+ Last.
+
+shortcut_switch([{Lit,L0}|T], Bool, Bs, St0) ->
+ St = St0#st{rel_op=normalize_op({bif,'=:='}, [Bool,Lit])},
+ #b_br{bool=#b_literal{val=true},succ=L} =
+ shortcut(L0, bind_var(Bool, Lit, Bs), St#st{target=one_way}),
+ [{Lit,L}|shortcut_switch(T, Bool, Bs, St0)];
+shortcut_switch([], _, _, _) -> [].
+
+shortcut(L, Bs, St) ->
+ shortcut_1(L, Bs, ordsets:new(), St).
+
+shortcut_1(L, Bs0, UnsetVars0, St) ->
+ case shortcut_2(L, Bs0, UnsetVars0, St) of
+ none ->
+ %% No more shortcuts found. Package up the previous
+ %% label in an unconditional branch.
+ #b_br{bool=#b_literal{val=true},succ=L,fail=L};
+ {#b_br{bool=#b_var{}}=Br,_,_} ->
+ %% This is a two-way branch. We can't do any better.
+ Br;
+ {#b_br{bool=#b_literal{val=true},succ=Succ},Bs,UnsetVars} ->
+ %% This is a safe `br`, but try to find a better one.
+ shortcut_1(Succ, Bs#{from:=L}, UnsetVars, St)
+ end.
+
+%% Try to shortcut this block, branching to a successor.
+shortcut_2(L, Bs0, UnsetVars0, St) ->
+ #b_blk{is=Is,last=Last} = get_block(L, St),
+ case eval_is(Is, Bs0, St) of
+ none ->
+ %% It is not safe to avoid this block because it
+ %% has instructions with potential side effects.
+ none;
+ Bs ->
+ %% The instructions in the block (if any) don't
+ %% have any side effects and can be skipped.
+ %% Evaluate the terminator.
+ case eval_terminator(Last, Bs, St) of
+ none ->
+ %% The terminator is not suitable (could be
+ %% because it is a switch that can't be simplified
+ %% or it is a ret instruction).
+ none;
+ #b_br{}=Br ->
+ %% We have a potentially suitable br.
+ %% Now update the set of variables that will never
+ %% be set if this block will be skipped.
+ UnsetVars1 = [V || #b_set{dst=#b_var{name=V}} <- Is],
+ UnsetVars = ordsets:union(UnsetVars0,
+ ordsets:from_list(UnsetVars1)),
+
+ %% Continue checking whether this br is suitable.
+ shortcut_3(Br, Bs#{from:=L}, UnsetVars, St)
+ end
+ end.
+
+shortcut_3(Br, Bs, UnsetVars, #st{target=Target}=St) ->
+ case is_br_safe(UnsetVars, Br, St) of
+ false ->
+ %% Branching using this `br` is unsafe, either because it
+ %% is an unconditional branch to a phi node, or because
+ %% one or more of the variables that are not set will be
+ %% used. Try to follow branches of this `br`, to find a
+ %% safe `br`.
+ case Br of
+ #b_br{bool=#b_literal{val=true},succ=L} ->
+ case Target of
+ L ->
+ %% We have reached the forced target, and it
+ %% is unsafe. Give up.
+ none;
+ _ ->
+ %% Try following this branch to see whether it
+ %% leads to a safe `br`.
+ shortcut_2(L, Bs, UnsetVars, St)
+ end;
+ #b_br{bool=#b_var{},succ=Succ,fail=Fail} ->
+ case {Succ,Fail} of
+ {L,Target} ->
+ %% The failure label is the forced target.
+ %% Try following the success label to see
+ %% whether it also ultimately ends up at the
+ %% forced target.
+ shortcut_2(L, Bs, UnsetVars, St);
+ {Target,L} ->
+ %% The success label is the forced target.
+ %% Try following the failure label to see
+ %% whether it also ultimately ends up at the
+ %% forced target.
+ shortcut_2(L, Bs, UnsetVars, St);
+ {_,_} ->
+ case Target of
+ any ->
+ %% This two-way branch is unsafe. Try reducing
+ %% it to a one-way branch.
+ shortcut_two_way(Br, Bs, UnsetVars, St);
+ one_way ->
+ %% This two-way branch is unsafe. Try reducing
+ %% it to a one-way branch.
+ shortcut_two_way(Br, Bs, UnsetVars, St);
+ _ when is_integer(Target) ->
+ %% This two-way branch is unsafe, and
+ %% there already is a forced target.
+ %% Give up.
+ none
+ end
+ end
+ end;
+ true ->
+ %% This `br` instruction is safe. It does not
+ %% branch to a phi node, and all variables that
+ %% will be used are guaranteed to be defined.
+ case Br of
+ #b_br{bool=#b_literal{val=true},succ=L} ->
+ %% This is a one-way branch.
+ case Target of
+ any ->
+ %% No forced target. Success!
+ {Br,Bs,UnsetVars};
+ one_way ->
+ %% The target must be a one-way branch, which this
+ %% `br` is. Success!
+ {Br,Bs,UnsetVars};
+ L when is_integer(Target) ->
+ %% The forced target is L. Success!
+ {Br,Bs,UnsetVars};
+ _ when is_integer(Target) ->
+ %% Wrong forced target. Try following this branch
+ %% to see if it ultimately ends up at the forced
+ %% target.
+ shortcut_2(L, Bs, UnsetVars, St)
+ end;
+ #b_br{bool=#b_var{}} ->
+ %% This is a two-way branch.
+ if
+ Target =:= any; Target =:= one_way ->
+ %% No specific forced target. Try to reduce the
+ %% two-way branch to an one-way branch.
+ case shortcut_two_way(Br, Bs, UnsetVars, St) of
+ none when Target =:= any ->
+ %% This `br` can't be reduced to a one-way
+ %% branch. Return the `br` as-is.
+ {Br,Bs,UnsetVars};
+ none when Target =:= one_way ->
+ %% This `br` can't be reduced to a one-way
+ %% branch. The caller wants a one-way branch.
+ %% Give up.
+ none;
+ {_,_,_}=Res ->
+ %% This `br` was successfully reduced to a
+ %% one-way branch.
+ Res
+ end;
+ is_integer(Target) ->
+ %% There is a forced target, which can't
+ %% be reached because this `br` is a two-way
+ %% branch. Give up.
+ none
+ end
+ end
+ end.
+
+shortcut_two_way(#b_br{succ=Succ,fail=Fail}, Bs0, UnsetVars0, St) ->
+ case shortcut_2(Succ, Bs0, UnsetVars0, St#st{target=Fail}) of
+ {#b_br{bool=#b_literal{},succ=Fail},_,_}=Res ->
+ Res;
+ none ->
+ case shortcut_2(Fail, Bs0, UnsetVars0, St#st{target=Succ}) of
+ {#b_br{bool=#b_literal{},succ=Succ},_,_}=Res ->
+ Res;
+ none ->
+ none
+ end
+ end.
+
+get_block(L, St) ->
+ #st{bs=#{L:=Blk}} = St,
+ Blk.
+
+is_br_safe(UnsetVars, Br, #st{us=Us}=St) ->
+ %% Check that none of the unset variables will be used.
+ case Br of
+ #b_br{bool=#b_var{name=V},succ=Succ,fail=Fail} ->
+ #{Succ:=Used0,Fail:=Used1} = Us,
+
+ %% A two-way branch never branches to a phi node, so there
+ %% is no need to check for phi nodes here.
+ not member(V, UnsetVars) andalso
+ ordsets:is_disjoint(Used0, UnsetVars) andalso
+ ordsets:is_disjoint(Used1, UnsetVars);
+ #b_br{succ=Same,fail=Same} ->
+ %% An unconditional branch must not jump to
+ %% a phi node.
+ not is_forbidden(Same, St) andalso
+ ordsets:is_disjoint(map_get(Same, Us), UnsetVars)
+ end.
+
+is_forbidden(L, St) ->
+ case get_block(L, St) of
+ #b_blk{is=[#b_set{op=phi}|_]} -> true;
+ #b_blk{is=[#b_set{op=peek_message}|_]} -> true;
+ #b_blk{} -> false
+ end.
+
+
+%% Evaluate the instructions in the block.
+%% Return the updated bindings, or 'none' if there is
+%% any instruction with potential side effects.
+
+eval_is([#b_set{op=phi,dst=Dst,args=Args}|Is], Bs0, St) ->
+ From = maps:get(from, Bs0),
+ [Val] = [Val || {Val,Pred} <- Args, Pred =:= From],
+ Bs = bind_var(Dst, Val, Bs0),
+ eval_is(Is, Bs, St);
+eval_is([#b_set{op={bif,_},dst=Dst}=I0|Is], Bs, St) ->
+ I = sub(I0, Bs),
+ case eval_bif(I, St) of
+ #b_literal{}=Val ->
+ eval_is(Is, bind_var(Dst, Val, Bs), St);
+ none ->
+ eval_is(Is, Bs, St)
+ end;
+eval_is([#b_set{op=Op,dst=Dst}=I|Is], Bs, St)
+ when Op =:= is_tagged_tuple; Op =:= is_nonempty_list ->
+ #b_set{args=Args} = sub(I, Bs),
+ case eval_rel_op(Op, Args, St) of
+ #b_literal{}=Val ->
+ eval_is(Is, bind_var(Dst, Val, Bs), St);
+ none ->
+ eval_is(Is, Bs, St)
+ end;
+eval_is([#b_set{}=I|Is], Bs, St) ->
+ case beam_ssa:no_side_effect(I) of
+ true ->
+ %% This instruction has no side effects. It can
+ %% safely be omitted.
+ eval_is(Is, Bs, St);
+ false ->
+ %% This instruction may have some side effect.
+ %% It is not safe to avoid this instruction.
+ none
+ end;
+eval_is([], Bs, _St) -> Bs.
+
+eval_terminator(#b_br{bool=#b_var{}=Bool}=Br, Bs, _St) ->
+ Val = get_value(Bool, Bs),
+ beam_ssa:normalize(Br#b_br{bool=Val});
+eval_terminator(#b_br{bool=#b_literal{}}=Br, _Bs, _St) ->
+ beam_ssa:normalize(Br);
+eval_terminator(#b_switch{arg=Arg,fail=Fail,list=List}=Sw, Bs, St) ->
+ case get_value(Arg, Bs) of
+ #b_literal{}=Val ->
+ %% Literal argument. Simplify to a `br`.
+ beam_ssa:normalize(Sw#b_switch{arg=Val});
+ #b_var{} ->
+ case St of
+ #st{rel_op=none} ->
+ %% No previous relational operator is stored.
+ %% Give up.
+ none;
+ #st{} ->
+ %% There is a previous relational operator stored.
+ %% Try optimizing the switch.
+ case eval_switch(List, Arg, St, Fail) of
+ none ->
+ none;
+ To when is_integer(To) ->
+ %% Either one of the values in the switch
+ %% matched a previous value in a '=:=' test, or
+ %% none of the values matched a previous test.
+ #b_br{bool=#b_literal{val=true},succ=To,fail=To}
+ end
+ end
+ end;
+eval_terminator(#b_ret{}, _Bs, _St) ->
+ none.
+
+eval_switch([{Lit,Lbl}|T], Arg, St, Fail) ->
+ case eval_rel_op({bif,'=:='}, [Arg,Lit], St) of
+ none ->
+ %% This label could be reached.
+ eval_switch(T, Arg, St, none);
+ #b_literal{val=false} ->
+ %% This branch will never be taken.
+ eval_switch(T, Arg, St, Fail);
+ #b_literal{val=true} ->
+ %% Success. This branch will always be taken.
+ Lbl
+ end;
+eval_switch([], _Arg, _St, Fail) ->
+ %% Fail is now either the failure label or 'none'.
+ Fail.
+
+bind_var(Var, Val0, Bs) ->
+ Val = get_value(Val0, Bs),
+ Bs#{Var=>Val}.
+
+get_value(#b_var{}=Var, Bs) ->
+ case Bs of
+ #{Var:=Val} -> get_value(Val, Bs);
+ #{} -> Var
+ end;
+get_value(#b_literal{}=Lit, _Bs) -> Lit.
+
+eval_bif(#b_set{op={bif,Bif},args=Args}, St) ->
+ Arity = length(Args),
+ case erl_bifs:is_pure(erlang, Bif, Arity) of
+ false ->
+ none;
+ true ->
+ case [Lit || #b_literal{val=Lit} <- Args] of
+ LitArgs when length(LitArgs) =:= Arity ->
+ try apply(erlang, Bif, LitArgs) of
+ Val -> #b_literal{val=Val}
+ catch
+ error:_ -> none
+ end;
+ _ ->
+ %% Not literal arguments. Try to evaluate
+ %% it based on a previous relational operator.
+ eval_rel_op({bif,Bif}, Args, St)
+ end
+ end.
+
+%%%
+%%% Handling of relational operators.
+%%%
+
+get_rel_op(Bool, [_|_]=Is) ->
+ case last(Is) of
+ #b_set{op=Op,dst=Bool,args=Args} ->
+ normalize_op(Op, Args);
+ #b_set{} ->
+ none
+ end;
+get_rel_op(_, []) -> none.
+
+%% normalize_op(Instruction) -> {Normalized,FailLabel} | error
+%% Normalized = {Operator,Variable,Variable|Literal} |
+%% {TypeTest,Variable}
+%% Operation = '<' | '=<' | '=:=' | '=/=' | '>=' | '>'
+%% TypeTest = is_atom | is_integer ...
+%% Variable = #b_var{}
+%% Literal = #b_literal{}
+%%
+%% Normalize a relational operator to facilitate further
+%% comparisons between operators. Always make the register
+%% operand the first operand. If there are two registers,
+%% order the registers in lexical order.
+%%
+%% For example, this instruction:
+%%
+%% #b_set{op={bif,=<},args=[#b_literal{}, #b_var{}}
+%%
+%% will be normalized to:
+%%
+%% {'=<',#b_var{},#b_literal{}}
+
+-spec normalize_op(Op, Args) -> NormalizedOp | 'none' when
+ Op :: beam_ssa:op(),
+ Args :: [beam_ssa:value()],
+ NormalizedOp :: basic_rel_op().
+
+normalize_op(is_tagged_tuple, [Arg,#b_literal{val=Size},#b_literal{val=Tag}])
+ when is_integer(Size), is_atom(Tag) ->
+ {{is_tagged_tuple,Size,Tag},Arg};
+normalize_op(is_nonempty_list, [Arg]) ->
+ {is_nonempty_list,Arg};
+normalize_op({bif,Bif}, [Arg]) ->
+ case erl_internal:new_type_test(Bif, 1) of
+ true -> {Bif,Arg};
+ false -> none
+ end;
+normalize_op({bif,Bif}, [_,_]=Args) ->
+ case erl_internal:comp_op(Bif, 2) of
+ true ->
+ normalize_op_1(Bif, Args);
+ false ->
+ none
+ end;
+normalize_op(_, _) -> none.
+
+normalize_op_1(Bif, Args) ->
+ case Args of
+ [#b_literal{}=Arg1,#b_var{}=Arg2] ->
+ {turn_op(Bif),Arg2,Arg1};
+ [#b_var{}=Arg1,#b_literal{}=Arg2] ->
+ {Bif,Arg1,Arg2};
+ [#b_var{}=A,#b_var{}=B] ->
+ if A < B -> {Bif,A,B};
+ true -> {turn_op(Bif),B,A}
+ end;
+ [#b_literal{},#b_literal{}] ->
+ none
+ end.
+
+-spec invert_op(basic_rel_op() | 'none') -> rel_op() | 'none'.
+
+invert_op({Op,Arg1,Arg2}) ->
+ {invert_op_1(Op),Arg1,Arg2};
+invert_op({TypeTest,Arg}) ->
+ {{'not',TypeTest},Arg};
+invert_op(none) -> none.
+
+invert_op_1('>=') -> '<';
+invert_op_1('<') -> '>=';
+invert_op_1('=<') -> '>';
+invert_op_1('>') -> '=<';
+invert_op_1('=:=') -> '=/=';
+invert_op_1('=/=') -> '=:=';
+invert_op_1('==') -> '/=';
+invert_op_1('/=') -> '=='.
+
+turn_op('<') -> '>';
+turn_op('=<') -> '>=';
+turn_op('>') -> '<';
+turn_op('>=') -> '=<';
+turn_op('=:='=Op) -> Op;
+turn_op('=/='=Op) -> Op;
+turn_op('=='=Op) -> Op;
+turn_op('/='=Op) -> Op.
+
+eval_rel_op(_Bif, _Args, #st{rel_op=none}) ->
+ none;
+eval_rel_op(Bif, Args, #st{rel_op=Prev}) ->
+ case normalize_op(Bif, Args) of
+ none ->
+ none;
+ RelOp ->
+ case will_succeed(Prev, RelOp) of
+ yes -> #b_literal{val=true};
+ no -> #b_literal{val=false};
+ maybe -> none
+ end
+ end.
+
+%% will_succeed(PrevCondition, Condition) -> yes | no | maybe
+%% PrevCondition is a condition known to be true. This function
+%% will tell whether Condition will succeed.
+
+will_succeed({_Op,_Var,_Value}=Same, {_Op,_Var,_Value}=Same) ->
+ %% Repeated test.
+ yes;
+will_succeed({Op1,Var,#b_literal{val=A}}, {Op2,Var,#b_literal{val=B}}) ->
+ will_succeed_1(Op1, A, Op2, B);
+will_succeed({Op1,Var,#b_var{}=A}, {Op2,Var,#b_var{}=B}) ->
+ will_succeed_vars(Op1, A, Op2, B);
+will_succeed({'=:=',Var,#b_literal{val=A}}, {TypeTest,Var}) ->
+ eval_type_test(TypeTest, A);
+will_succeed({_,_}=Same, {_,_}=Same) ->
+ %% Repeated type test.
+ yes;
+will_succeed({Test1,Var}, {Test2,Var}) ->
+ will_succeed_test(Test1, Test2);
+will_succeed({_,_}, {_,_}) ->
+ maybe;
+will_succeed({_,_}, {_,_,_}) ->
+ maybe;
+will_succeed({_,_,_}, {_,_}) ->
+ maybe;
+will_succeed({_,_,_}, {_,_,_}) ->
+ maybe.
+
+will_succeed_test({'not',Test1}, Test2) ->
+ case Test1 =:= Test2 of
+ true -> no;
+ false -> maybe
+ end;
+will_succeed_test(is_tuple, {is_tagged_tuple,_,_}) ->
+ maybe;
+will_succeed_test({is_tagged_tuple,_,_}, is_tuple) ->
+ yes;
+will_succeed_test(is_list, is_nonempty_list) ->
+ maybe;
+will_succeed_test(is_nonempty_list, is_list) ->
+ yes;
+will_succeed_test(T1, T2) ->
+ case is_numeric_test(T1) andalso is_numeric_test(T2) of
+ true -> maybe;
+ false -> no
+ end.
+
+will_succeed_1('=:=', A, '<', B) ->
+ if
+ B =< A -> no;
+ true -> yes
+ end;
+will_succeed_1('=:=', A, '=<', B) ->
+ if
+ B < A -> no;
+ true -> yes
+ end;
+will_succeed_1('=:=', A, '=:=', B) when A =/= B ->
+ no;
+will_succeed_1('=:=', A, '=/=', B) ->
+ if
+ A =:= B -> no;
+ true -> yes
+ end;
+will_succeed_1('=:=', A, '>=', B) ->
+ if
+ B > A -> no;
+ true -> yes
+ end;
+will_succeed_1('=:=', A, '>', B) ->
+ if
+ B >= A -> no;
+ true -> yes
+ end;
+
+will_succeed_1('=/=', A, '=:=', B) when A =:= B -> no;
+
+will_succeed_1('<', A, '=:=', B) when B >= A -> no;
+will_succeed_1('<', A, '=/=', B) when B >= A -> yes;
+will_succeed_1('<', A, '<', B) when B >= A -> yes;
+will_succeed_1('<', A, '=<', B) when B > A -> yes;
+will_succeed_1('<', A, '>=', B) when B > A -> no;
+will_succeed_1('<', A, '>', B) when B >= A -> no;
+
+will_succeed_1('=<', A, '=:=', B) when B > A -> no;
+will_succeed_1('=<', A, '=/=', B) when B > A -> yes;
+will_succeed_1('=<', A, '<', B) when B > A -> yes;
+will_succeed_1('=<', A, '=<', B) when B >= A -> yes;
+will_succeed_1('=<', A, '>=', B) when B > A -> no;
+will_succeed_1('=<', A, '>', B) when B >= A -> no;
+
+will_succeed_1('>=', A, '=:=', B) when B < A -> no;
+will_succeed_1('>=', A, '=/=', B) when B < A -> yes;
+will_succeed_1('>=', A, '<', B) when B =< A -> no;
+will_succeed_1('>=', A, '=<', B) when B < A -> no;
+will_succeed_1('>=', A, '>=', B) when B =< A -> yes;
+will_succeed_1('>=', A, '>', B) when B < A -> yes;
+
+will_succeed_1('>', A, '=:=', B) when B =< A -> no;
+will_succeed_1('>', A, '=/=', B) when B =< A -> yes;
+will_succeed_1('>', A, '<', B) when B =< A -> no;
+will_succeed_1('>', A, '=<', B) when B < A -> no;
+will_succeed_1('>', A, '>=', B) when B =< A -> yes;
+will_succeed_1('>', A, '>', B) when B < A -> yes;
+
+will_succeed_1('==', A, '==', B) ->
+ if
+ A == B -> yes;
+ true -> no
+ end;
+will_succeed_1('==', A, '/=', B) ->
+ if
+ A == B -> no;
+ true -> yes
+ end;
+will_succeed_1('/=', A, '/=', B) when A == B -> yes;
+will_succeed_1('/=', A, '==', B) when A == B -> no;
+
+will_succeed_1(_, _, _, _) -> maybe.
+
+will_succeed_vars('=/=', Val, '=:=', Val) -> no;
+will_succeed_vars('=:=', Val, '=/=', Val) -> no;
+will_succeed_vars('=:=', Val, '>=', Val) -> yes;
+will_succeed_vars('=:=', Val, '=<', Val) -> yes;
+
+will_succeed_vars('/=', Val1, '==', Val2) when Val1 == Val2 -> no;
+will_succeed_vars('==', Val1, '/=', Val2) when Val1 == Val2 -> no;
+
+will_succeed_vars(_, _, _, _) -> maybe.
+
+is_numeric_test(is_float) -> true;
+is_numeric_test(is_integer) -> true;
+is_numeric_test(is_number) -> true;
+is_numeric_test(_) -> false.
+
+eval_type_test(Test, Arg) ->
+ case eval_type_test_1(Test, Arg) of
+ true -> yes;
+ false -> no
+ end.
+
+eval_type_test_1(is_nonempty_list, Arg) ->
+ case Arg of
+ [_|_] -> true;
+ _ -> false
+ end;
+eval_type_test_1({is_tagged_tuple,Sz,Tag}, Arg) ->
+ if
+ tuple_size(Arg) =:= Sz, element(1, Arg) =:= Tag ->
+ true;
+ true ->
+ false
+ end;
+eval_type_test_1(Test, Arg) ->
+ erlang:Test(Arg).
+
+%%%
+%%% Combine bif:'=:=' and switch instructions
+%%% to switch instructions.
+%%%
+%%% Consider this code:
+%%%
+%%% 0:
+%%% @ssa_bool = bif:'=:=' Var, literal 1
+%%% br @ssa_bool, label 2, label 3
+%%%
+%%% 2:
+%%% ret literal a
+%%%
+%%% 3:
+%%% @ssa_bool:7 = bif:'=:=' Var, literal 2
+%%% br @ssa_bool:7, label 4, label 999
+%%%
+%%% 4:
+%%% ret literal b
+%%%
+%%% 999:
+%%% .
+%%% .
+%%% .
+%%%
+%%% The two bif:'=:=' instructions can be combined
+%%% to a switch:
+%%%
+%%% 0:
+%%% switch Var, label 999, [ { literal 1, label 2 },
+%%% { literal 2, label 3 } ]
+%%%
+%%% 2:
+%%% ret literal a
+%%%
+%%% 4:
+%%% ret literal b
+%%%
+%%% 999:
+%%% .
+%%% .
+%%% .
+%%%
+
+combine_eqs(#st{bs=Blocks}=St) ->
+ Ls = reverse(beam_ssa:rpo(Blocks)),
+ combine_eqs_1(Ls, St).
+
+combine_eqs_1([L|Ls], #st{bs=Blocks0}=St0) ->
+ case comb_get_sw(L, St0) of
+ none ->
+ combine_eqs_1(Ls, St0);
+ {_,Arg,_,Fail0,List0} ->
+ case comb_get_sw(Fail0, St0) of
+ {true,Arg,Fail1,Fail,List1} ->
+ %% Another switch/br with the same arguments was
+ %% found. Try combining them.
+ case combine_lists(Fail1, List0, List1, Blocks0) of
+ none ->
+ %% Different types of literals in the lists,
+ %% or the success cases in the first switch
+ %% could branch to the second switch
+ %% (increasing code size and repeating tests).
+ combine_eqs_1(Ls, St0);
+ List ->
+ %% Everything OK! Combine the lists.
+ Sw0 = #b_switch{arg=Arg,fail=Fail,list=List},
+ Sw = beam_ssa:normalize(Sw0),
+ Blk0 = maps:get(L, Blocks0),
+ Blk = Blk0#b_blk{last=Sw},
+ Blocks = Blocks0#{L:=Blk},
+ St = St0#st{bs=Blocks},
+ combine_eqs_1(Ls, St)
+ end;
+ {true,_OtherArg,_,_,_} ->
+ %% The other switch/br uses a different Arg.
+ combine_eqs_1(Ls, St0);
+ {false,_,_,_,_} ->
+ %% Not safe: Bindings of variables that will be used
+ %% or execution of instructions with potential
+ %% side effects will be skipped.
+ combine_eqs_1(Ls, St0);
+ none ->
+ %% No switch/br at this label.
+ combine_eqs_1(Ls, St0)
+ end
+ end;
+combine_eqs_1([], St) -> St.
+
+comb_get_sw(L, Blocks) ->
+ comb_get_sw(L, true, Blocks).
+
+comb_get_sw(L, Safe0, #st{bs=Blocks,skippable=Skippable}=St) ->
+ #b_blk{is=Is,last=Last} = maps:get(L, Blocks),
+ Safe1 = Safe0 andalso is_map_key(L, Skippable),
+ case Last of
+ #b_ret{} ->
+ none;
+ #b_br{bool=#b_var{}=Bool,succ=Succ,fail=Fail} ->
+ case comb_is(Is, Bool, Safe1) of
+ {none,_} ->
+ none;
+ {#b_set{op={bif,'=:='},args=[#b_var{}=Arg,#b_literal{}=Lit]},Safe} ->
+ {Safe,Arg,L,Fail,[{Lit,Succ}]};
+ {#b_set{},_} ->
+ none
+ end;
+ #b_br{bool=#b_literal{val=true},succ=Succ} ->
+ comb_get_sw(Succ, Safe1, St);
+ #b_switch{arg=#b_var{}=Arg,fail=Fail,list=List} ->
+ {none,Safe} = comb_is(Is, none, Safe1),
+ {Safe,Arg,L,Fail,List}
+ end.
+
+comb_is([#b_set{dst=#b_var{}=Bool}=I], Bool, Safe) ->
+ {I,Safe};
+comb_is([#b_set{}=I|Is], Bool, Safe0) ->
+ Safe = Safe0 andalso beam_ssa:no_side_effect(I),
+ comb_is(Is, Bool, Safe);
+comb_is([], _Bool, Safe) ->
+ {none,Safe}.
+
+%% combine_list(Fail, List1, List2, Blocks) -> List|none.
+%% Try to combine two switch lists, returning the combined
+%% list or 'none' if not possible.
+%%
+%% The values in the two lists must be all of the same type.
+%%
+%% The code reached from the labels in the first list must
+%% not reach the failure label (if they do, tests could
+%% be repeated).
+%%
+
+combine_lists(Fail, L1, L2, Blocks) ->
+ Ls = beam_ssa:rpo([Lbl || {_,Lbl} <- L1], Blocks),
+ case member(Fail, Ls) of
+ true ->
+ %% One or more of labels in the first list
+ %% could reach the failure label. That
+ %% means that the second switch/br instruction
+ %% will be retained, increasing code size and
+ %% potentially also execution time.
+ none;
+ false ->
+ %% The combined switch will replace both original
+ %% br/switch instructions, leading to a reduction in code
+ %% size and potentially also in execution time.
+ combine_lists_1(L1, L2)
+ end.
+
+combine_lists_1(List0, List1) ->
+ case are_lists_compatible(List0, List1) of
+ true ->
+ First = maps:from_list(List0),
+ List0 ++ [{Val,Lbl} || {Val,Lbl} <- List1,
+ not is_map_key(Val, First)];
+ false ->
+ none
+ end.
+
+are_lists_compatible([{#b_literal{val=Val1},_}|_],
+ [{#b_literal{val=Val2},_}|_]) ->
+ case lit_type(Val1) of
+ none -> false;
+ Type -> Type =:= lit_type(Val2)
+ end.
+
+lit_type(Val) ->
+ if
+ is_atom(Val) -> atom;
+ is_float(Val) -> float;
+ is_integer(Val) -> integer;
+ true -> none
+ end.
+
+%%%
+%%% Calculate used variables for each block.
+%%%
+
+used_vars(Linear) ->
+ used_vars(reverse(Linear), #{}, #{}).
+
+used_vars([{L,#b_blk{is=Is}=Blk}|Bs], UsedVars0, Skip0) ->
+ %% Calculate the variables used by each block and its
+ %% successors. This information is used by
+ %% shortcut_opt/1.
+
+ Successors = beam_ssa:successors(Blk),
+ Used0 = used_vars_succ(Successors, L, UsedVars0),
+ Used = used_vars_blk(Blk, Used0),
+ UsedVars = used_vars_phis(Is, L, Used, UsedVars0),
+
+ %% combine_eqs/1 needs different variable usage
+ %% information than shortcut_opt/1. The Skip
+ %% map will have an entry for each block that
+ %% can be skipped (does not bind any variable used
+ %% in successor).
+
+ Defined0 = [Def || #b_set{dst=#b_var{name=Def}} <- Is],
+ Defined = ordsets:from_list(Defined0),
+ MaySkip = ordsets:is_disjoint(Defined, Used0),
+ case MaySkip of
+ true ->
+ Skip = Skip0#{L=>true},
+ used_vars(Bs, UsedVars, Skip);
+ false ->
+ used_vars(Bs, UsedVars, Skip0)
+ end;
+used_vars([], UsedVars, Skip) ->
+ {UsedVars,Skip}.
+
+used_vars_succ([S|Ss], L, UsedVars) ->
+ Live0 = used_vars_succ(Ss, L, UsedVars),
+ Key = {S,L},
+ case UsedVars of
+ #{Key:=Live} ->
+ ordsets:union(Live, Live0);
+ #{S:=Live} ->
+ ordsets:union(Live, Live0);
+ #{} ->
+ Live0
+ end;
+used_vars_succ([], _, _) ->
+ ordsets:new().
+
+used_vars_phis(Is, L, Live0, UsedVars0) ->
+ UsedVars = UsedVars0#{L=>Live0},
+ Phis = takewhile(fun(#b_set{op=Op}) -> Op =:= phi end, Is),
+ case Phis of
+ [] ->
+ UsedVars;
+ [_|_] ->
+ PhiArgs = append([Args || #b_set{args=Args} <- Phis]),
+ case [{P,V} || {#b_var{name=V},P} <- PhiArgs] of
+ [_|_]=PhiVars ->
+ PhiLive0 = rel2fam(PhiVars),
+ PhiLive = [{{L,P},ordsets:union(ordsets:from_list(Vs), Live0)} ||
+ {P,Vs} <- PhiLive0],
+ maps:merge(UsedVars, maps:from_list(PhiLive));
+ [] ->
+ %% There were only literals in the phi node(s).
+ UsedVars
+ end
+ end.
+
+used_vars_blk(#b_blk{is=Is,last=Last}, Used0) ->
+ Used = ordsets:union(Used0, beam_ssa:used(Last)),
+ used_vars_is(reverse(Is), Used).
+
+used_vars_is([#b_set{op=phi}|Is], Used) ->
+ used_vars_is(Is, Used);
+used_vars_is([#b_set{dst=#b_var{name=Dst}}=I|Is], Used0) ->
+ Used1 = ordsets:union(Used0, beam_ssa:used(I)),
+ Used = ordsets:del_element(Dst, Used1),
+ used_vars_is(Is, Used);
+used_vars_is([], Used) ->
+ Used.
+
+%%%
+%%% Common utilities.
+%%%
+
+sub(#b_set{args=Args}=I, Sub) ->
+ I#b_set{args=[sub_arg(A, Sub) || A <- Args]}.
+
+sub_arg(Old, Sub) ->
+ case Sub of
+ #{Old:=New} -> New;
+ #{} -> Old
+ end.
+
+rel2fam(S0) ->
+ S1 = sofs:relation(S0),
+ S = sofs:rel2fam(S1),
+ sofs:to_external(S).
diff --git a/lib/compiler/src/beam_ssa_opt.erl b/lib/compiler/src/beam_ssa_opt.erl
index 2174c6aa31..83ee918b25 100644
--- a/lib/compiler/src/beam_ssa_opt.erl
+++ b/lib/compiler/src/beam_ssa_opt.erl
@@ -52,10 +52,19 @@ passes(Opts0) ->
?PASS(ssa_opt_element),
?PASS(ssa_opt_linearize),
?PASS(ssa_opt_record),
+
+ %% Run ssa_opt_cse twice, because it will help ssa_opt_dead,
+ %% and ssa_opt_dead will help ssa_opt_cse. Run ssa_opt_live
+ %% twice, because it will help ssa_opt_dead and ssa_opt_dead
+ %% will help ssa_opt_live.
?PASS(ssa_opt_cse),
?PASS(ssa_opt_type),
- ?PASS(ssa_opt_float),
?PASS(ssa_opt_live),
+ ?PASS(ssa_opt_dead),
+ ?PASS(ssa_opt_cse), %Second time.
+ ?PASS(ssa_opt_float),
+ ?PASS(ssa_opt_live), %Second time.
+
?PASS(ssa_opt_bsm),
?PASS(ssa_opt_bsm_shortcut),
?PASS(ssa_opt_misc),
@@ -91,6 +100,9 @@ function(#b_function{anno=Anno,bs=Blocks0,args=Args,cnt=Count0}=F, Ps) ->
%%% Trivial sub passes.
%%%
+ssa_opt_dead(#st{ssa=Linear}=St) ->
+ St#st{ssa=beam_ssa_dead:opt(Linear)}.
+
ssa_opt_linearize(#st{ssa=Blocks}=St) ->
St#st{ssa=beam_ssa:linearize(Blocks)}.
@@ -493,11 +505,13 @@ cse_suitable(#b_set{op={bif,tuple_size}}) ->
%% and beam_validator could fail to understand
%% that tuple operations that follow are safe.
false;
-cse_suitable(#b_set{op={bif,Name},args=Args}) ->
+cse_suitable(#b_set{anno=Anno,op={bif,Name},args=Args}) ->
+ %% Doing CSE for floating point operators is unsafe.
%% Doing CSE for comparison operators would prevent
%% creation of 'test' instructions.
Arity = length(Args),
- not (erl_internal:new_type_test(Name, Arity) orelse
+ not (is_map_key(float_op, Anno) orelse
+ erl_internal:new_type_test(Name, Arity) orelse
erl_internal:comp_op(Name, Arity) orelse
erl_internal:bool_op(Name, Arity));
cse_suitable(#b_set{}) -> false.
diff --git a/lib/compiler/src/compile.erl b/lib/compiler/src/compile.erl
index 0a7f723c64..a0ebfff5f7 100644
--- a/lib/compiler/src/compile.erl
+++ b/lib/compiler/src/compile.erl
@@ -2047,6 +2047,7 @@ pre_load() ->
beam_peep,
beam_ssa,
beam_ssa_codegen,
+ beam_ssa_dead,
beam_ssa_opt,
beam_ssa_pre_codegen,
beam_ssa_recv,
diff --git a/lib/compiler/src/compiler.app.src b/lib/compiler/src/compiler.app.src
index 74529f7fef..f5e2b31b87 100644
--- a/lib/compiler/src/compiler.app.src
+++ b/lib/compiler/src/compiler.app.src
@@ -39,6 +39,7 @@
beam_peep,
beam_ssa,
beam_ssa_codegen,
+ beam_ssa_dead,
beam_ssa_lint,
beam_ssa_opt,
beam_ssa_pp,