aboutsummaryrefslogtreecommitdiffstats
path: root/lib/compiler/src
diff options
context:
space:
mode:
Diffstat (limited to 'lib/compiler/src')
-rw-r--r--lib/compiler/src/Makefile1
-rw-r--r--lib/compiler/src/beam_a.erl3
-rw-r--r--lib/compiler/src/beam_block.erl3
-rw-r--r--lib/compiler/src/beam_clean.erl18
-rw-r--r--lib/compiler/src/beam_except.erl96
-rw-r--r--lib/compiler/src/beam_flatten.erl72
-rw-r--r--lib/compiler/src/beam_jump.erl84
-rw-r--r--lib/compiler/src/beam_listing.erl2
-rw-r--r--lib/compiler/src/beam_ssa.erl19
-rw-r--r--lib/compiler/src/beam_ssa_bsm.erl22
-rw-r--r--lib/compiler/src/beam_ssa_codegen.erl175
-rw-r--r--lib/compiler/src/beam_ssa_dead.erl59
-rw-r--r--lib/compiler/src/beam_ssa_funs.erl149
-rw-r--r--lib/compiler/src/beam_ssa_pre_codegen.erl116
-rw-r--r--lib/compiler/src/beam_ssa_type.erl103
-rw-r--r--lib/compiler/src/beam_trim.erl316
-rw-r--r--lib/compiler/src/beam_utils.erl610
-rw-r--r--lib/compiler/src/beam_validator.erl9
-rw-r--r--lib/compiler/src/compile.erl60
-rw-r--r--lib/compiler/src/compiler.app.src1
-rw-r--r--lib/compiler/src/erl_bifs.erl1
21 files changed, 969 insertions, 950 deletions
diff --git a/lib/compiler/src/Makefile b/lib/compiler/src/Makefile
index 26ae6566e6..d475e5a19a 100644
--- a/lib/compiler/src/Makefile
+++ b/lib/compiler/src/Makefile
@@ -63,6 +63,7 @@ MODULES = \
beam_ssa_bsm \
beam_ssa_codegen \
beam_ssa_dead \
+ beam_ssa_funs \
beam_ssa_lint \
beam_ssa_opt \
beam_ssa_pp \
diff --git a/lib/compiler/src/beam_a.erl b/lib/compiler/src/beam_a.erl
index 0abc845310..dd2537a699 100644
--- a/lib/compiler/src/beam_a.erl
+++ b/lib/compiler/src/beam_a.erl
@@ -59,6 +59,9 @@ rename_instrs([{test,is_eq_exact,_,[Dst,Src]}=Test,
rename_instrs([{test,is_eq_exact,_,[Same,Same]}|Is]) ->
%% Same literal or same register. Will always succeed.
rename_instrs(Is);
+rename_instrs([{loop_rec,{f,Fail},{x,0}},{loop_rec_end,_},{label,Fail}|Is]) ->
+ %% This instruction sequence does nothing.
+ rename_instrs(Is);
rename_instrs([{apply_last,A,N}|Is]) ->
[{apply,A},{deallocate,N},return|rename_instrs(Is)];
rename_instrs([{call_last,A,F,N}|Is]) ->
diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl
index d28c0fd9e4..9d8d5b2b0c 100644
--- a/lib/compiler/src/beam_block.erl
+++ b/lib/compiler/src/beam_block.erl
@@ -49,9 +49,6 @@ function({function,Name,Arity,CLabel,Is0}) ->
blockify(Is) ->
blockify(Is, []).
-blockify([{loop_rec,{f,Fail},{x,0}},{loop_rec_end,_Lbl},{label,Fail}|Is], Acc) ->
- %% Useless instruction sequence.
- blockify(Is, Acc);
blockify([I|Is0]=IsAll, Acc) ->
case collect(I) of
error -> blockify(Is0, [I|Acc]);
diff --git a/lib/compiler/src/beam_clean.erl b/lib/compiler/src/beam_clean.erl
index f5f0ac2218..7299654476 100644
--- a/lib/compiler/src/beam_clean.erl
+++ b/lib/compiler/src/beam_clean.erl
@@ -23,17 +23,15 @@
-export([module/2]).
-export([clean_labels/1]).
--import(lists, [foldl/3]).
-spec module(beam_utils:module_code(), [compile:option()]) ->
{'ok',beam_utils:module_code()}.
module({Mod,Exp,Attr,Fs0,_}, Opts) ->
Order = [Lbl || {function,_,_,Lbl,_} <- Fs0],
- All = foldl(fun({function,_,_,Lbl,_}=Func,D) -> dict:store(Lbl, Func, D) end,
- dict:new(), Fs0),
+ All = maps:from_list([{Lbl,Func} || {function,_,_,Lbl,_}=Func <- Fs0]),
WorkList = rootset(Fs0, Exp, Attr),
- Used = find_all_used(WorkList, All, sets:from_list(WorkList)),
+ Used = find_all_used(WorkList, All, cerl_sets:from_list(WorkList)),
Fs1 = remove_unused(Order, Used, All),
{Fs2,Lc} = clean_labels(Fs1),
Fs = maybe_remove_lines(Fs2, Opts),
@@ -55,16 +53,16 @@ rootset(Fs, Root0, Attr) ->
%% Remove the unused functions.
remove_unused([F|Fs], Used, All) ->
- case sets:is_element(F, Used) of
+ case cerl_sets:is_element(F, Used) of
false -> remove_unused(Fs, Used, All);
- true -> [dict:fetch(F, All)|remove_unused(Fs, Used, All)]
+ true -> [map_get(F, All)|remove_unused(Fs, Used, All)]
end;
remove_unused([], _, _) -> [].
-
+
%% Find all used functions.
find_all_used([F|Fs0], All, Used0) ->
- {function,_,_,_,Code} = dict:fetch(F, All),
+ {function,_,_,_,Code} = map_get(F, All),
{Fs,Used} = update_work_list(Code, {Fs0,Used0}),
find_all_used(Fs, All, Used);
find_all_used([], _All, Used) -> Used.
@@ -78,9 +76,9 @@ update_work_list([_|Is], Sets) ->
update_work_list([], Sets) -> Sets.
add_to_work_list(F, {Fs,Used}=Sets) ->
- case sets:is_element(F, Used) of
+ case cerl_sets:is_element(F, Used) of
true -> Sets;
- false -> {[F|Fs],sets:add_element(F, Used)}
+ false -> {[F|Fs],cerl_sets:add_element(F, Used)}
end.
diff --git a/lib/compiler/src/beam_except.erl b/lib/compiler/src/beam_except.erl
index 98831d87a7..49bfb5606f 100644
--- a/lib/compiler/src/beam_except.erl
+++ b/lib/compiler/src/beam_except.erl
@@ -31,7 +31,7 @@
%%% erlang:error(function_clause, Args) => jump FuncInfoLabel
%%%
--import(lists, [reverse/1,seq/2]).
+-import(lists, [reverse/1,seq/2,splitwith/2]).
-spec module(beam_utils:module_code(), [compile:option()]) ->
{'ok',beam_utils:module_code()}.
@@ -74,13 +74,13 @@ translate([I|Is], St, Acc) ->
translate([], _, Acc) ->
reverse(Acc).
-translate_1(Ar, I, Is, St, [{line,_}=Line|Acc1]=Acc0) ->
- case dig_out(Ar, Acc1) of
+translate_1(Ar, I, Is, #st{arity=Arity}=St, [{line,_}=Line|Acc1]=Acc0) ->
+ case dig_out(Ar, Arity, Acc1) of
no ->
translate(Is, St, [I|Acc0]);
- {yes,{function_clause,Arity},Acc2} ->
+ {yes,function_clause,Acc2} ->
case {Line,St} of
- {{line,Loc},#st{lbl=Fi,loc=Loc,arity=Arity}} ->
+ {{line,Loc},#st{lbl=Fi,loc=Loc}} ->
Instr = {jump,{f,Fi}},
translate(Is, St, [Instr|Acc2]);
{_,_} ->
@@ -92,9 +92,13 @@ translate_1(Ar, I, Is, St, [{line,_}=Line|Acc1]=Acc0) ->
translate(Is, St, [Instr,Line|Acc2])
end.
-dig_out(Ar, [{kill,_}|Is]) ->
- dig_out(Ar, Is);
-dig_out(1, [{block,Bl0}|Is]) ->
+dig_out(1, _Arity, Is) ->
+ dig_out(Is);
+dig_out(2, Arity, Is) ->
+ dig_out_fc(Arity, Is);
+dig_out(_, _, _) -> no.
+
+dig_out([{block,Bl0}|Is]) ->
case dig_out_block(reverse(Bl0)) of
no -> no;
{yes,What,[]} ->
@@ -102,12 +106,7 @@ dig_out(1, [{block,Bl0}|Is]) ->
{yes,What,Bl} ->
{yes,What,[{block,Bl}|Is]}
end;
-dig_out(2, [{block,Bl}|Is]) ->
- case dig_out_block_fc(Bl) of
- no -> no;
- {yes,What} -> {yes,What,Is}
- end;
-dig_out(_, _) -> no.
+dig_out(_) -> no.
dig_out_block([{set,[{x,0}],[{atom,if_clause}],move}]) ->
{yes,if_end,[]};
@@ -141,33 +140,64 @@ fix_block_1([{set,[],[],{alloc,Live,{F1,F2,Needed0,F3}}}|Is], Words) ->
fix_block_1([I|Is], Words) ->
[I|fix_block_1(Is, Words)].
-dig_out_block_fc([{set,[],[],{alloc,Live,_}}|Bl]) ->
- Regs = maps:from_list([{{x,X},{arg,X}} || X <- seq(0, Live-1)]),
- dig_out_fc(Bl, Regs);
-dig_out_block_fc(_) -> no.
-dig_out_fc([{set,[Dst],[Hd,Tl],put_list}|Is], Regs0) ->
+dig_out_fc(Arity, Is0) ->
+ Regs0 = maps:from_list([{{x,X},{arg,X}} || X <- seq(0, Arity-1)]),
+ {Is,Acc0} = splitwith(fun({label,_}) -> false;
+ ({test,_,_,_}) -> false;
+ (_) -> true
+ end, Is0),
+ {Regs,Acc} = dig_out_fc_1(reverse(Is), Regs0, Acc0),
+ case is_fc(Arity, Regs) of
+ true ->
+ {yes,function_clause,Acc};
+ false ->
+ no
+ end.
+
+dig_out_fc_1([{block,Bl}|Is], Regs0, Acc) ->
+ Regs = dig_out_fc_block(Bl, Regs0),
+ dig_out_fc_1(Is, Regs, Acc);
+dig_out_fc_1([{bs_set_position,_,_}=I|Is], Regs, Acc) ->
+ dig_out_fc_1(Is, Regs, [I|Acc]);
+dig_out_fc_1([{bs_get_tail,_,_,Live}=I|Is], Regs0, Acc) ->
+ Regs = prune_xregs(Live, Regs0),
+ dig_out_fc_1(Is, Regs, [I|Acc]);
+dig_out_fc_1([_|_], _Regs, _Acc) ->
+ {#{},[]};
+dig_out_fc_1([], Regs, Acc) ->
+ {Regs,Acc}.
+
+dig_out_fc_block([{set,[],[],{alloc,Live,_}}|Is], Regs0) ->
+ Regs = prune_xregs(Live, Regs0),
+ dig_out_fc_block(Is, Regs);
+dig_out_fc_block([{set,[Dst],[Hd,Tl],put_list}|Is], Regs0) ->
Regs = Regs0#{Dst=>{cons,get_reg(Hd, Regs0),get_reg(Tl, Regs0)}},
- dig_out_fc(Is, Regs);
-dig_out_fc([{set,[Dst],[Src],move}|Is], 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(Is, Regs);
-dig_out_fc([{set,_,_,_}|_], _Regs) ->
- %% Unknown instruction. It is not a function_clause error.
- no;
-dig_out_fc([], Regs) ->
+ dig_out_fc_block(Is, Regs);
+dig_out_fc_block([{set,_,_,_}|_], _Regs) ->
+ %% Unknown instruction. Fail.
+ #{};
+dig_out_fc_block([], Regs) -> Regs.
+
+prune_xregs(Live, Regs) ->
+ maps:filter(fun({x,X}, _) -> X < Live end, Regs).
+
+is_fc(Arity, Regs) ->
case Regs of
#{{x,0}:={atom,function_clause},{x,1}:=Args} ->
- dig_out_fc_1(Args, 0);
+ is_fc_1(Args, 0) =:= Arity;
#{} ->
- no
+ false
end.
-dig_out_fc_1({cons,{arg,I},T}, I) ->
- dig_out_fc_1(T, I+1);
-dig_out_fc_1(nil, I) ->
- {yes,{function_clause,I}};
-dig_out_fc_1(_, _) -> no.
+is_fc_1({cons,{arg,I},T}, I) ->
+ is_fc_1(T, I+1);
+is_fc_1(nil, I) ->
+ I;
+is_fc_1(_, _) -> -1.
get_reg(R, Regs) ->
case Regs of
diff --git a/lib/compiler/src/beam_flatten.erl b/lib/compiler/src/beam_flatten.erl
index 973d16a1bc..3e6bc1b1ed 100644
--- a/lib/compiler/src/beam_flatten.erl
+++ b/lib/compiler/src/beam_flatten.erl
@@ -32,8 +32,7 @@ module({Mod,Exp,Attr,Fs,Lc}, _Opt) ->
{ok,{Mod,Exp,Attr,[function(F) || F <- Fs],Lc}}.
function({function,Name,Arity,CLabel,Is0}) ->
- Is1 = block(Is0),
- Is = opt(Is1),
+ Is = block(Is0),
{function,Name,Arity,CLabel,Is}.
block(Is) ->
@@ -43,21 +42,12 @@ block([{block,Is0}|Is1], Acc) -> block(Is1, norm_block(Is0, Acc));
block([I|Is], Acc) -> block(Is, [I|Acc]);
block([], Acc) -> reverse(Acc).
-norm_block([{set,[],[],{alloc,R,{_,nostack,_,_}=Alloc}}|Is], Acc0) ->
- case insert_alloc_in_bs_init(Acc0, Alloc) of
- impossible ->
- norm_block(Is, reverse(norm_allocate(Alloc, R), Acc0));
- Acc ->
- norm_block(Is, Acc)
- end;
norm_block([{set,[],[],{alloc,R,Alloc}}|Is], Acc0) ->
norm_block(Is, reverse(norm_allocate(Alloc, R), Acc0));
-norm_block([{set,[D1],[S],get_hd},{set,[D2],[S],get_tl}|Is], Acc) ->
- I = {get_list,S,D1,D2},
- norm_block(Is, [I|Acc]);
-norm_block([I|Is], Acc) -> norm_block(Is, [norm(I)|Acc]);
+norm_block([I|Is], Acc) ->
+ norm_block(Is, [norm(I)|Acc]);
norm_block([], Acc) -> Acc.
-
+
norm({set,[D],As,{bif,N,F}}) -> {bif,N,F,As,D};
norm({set,[D],As,{alloc,R,{gc_bif,N,F}}}) -> {gc_bif,N,F,R,As,D};
norm({set,[D],[],init}) -> {init,D};
@@ -91,57 +81,3 @@ norm_allocate({nozero,Ns,0,Inits}, Regs) ->
[{allocate,Ns,Regs}|Inits];
norm_allocate({nozero,Ns,Nh,Inits}, Regs) ->
[{allocate_heap,Ns,Nh,Regs}|Inits].
-
-%% insert_alloc_in_bs_init(ReverseInstructionStream, AllocationInfo) ->
-%% impossible | ReverseInstructionStream'
-%% A bs_init/6 instruction should not be followed by a test heap instruction.
-%% Given the AllocationInfo from a test heap instruction, merge the
-%% allocation amounts into the previous bs_init/6 instruction (if any).
-%%
-insert_alloc_in_bs_init([{bs_put,_,_,_}=I|Is], Alloc) ->
- %% The instruction sequence ends with an bs_put/4 instruction.
- %% We'll need to search backwards for the bs_init/6 instruction.
- insert_alloc_1(Is, Alloc, [I]);
-insert_alloc_in_bs_init(_, _) -> impossible.
-
-insert_alloc_1([{bs_init=Op,Fail,Info0,Live,Ss,Dst}|Is],
- {_,nostack,Ws2,[]}, Acc) when is_integer(Live) ->
- %% The number of extra heap words is always in the second position
- %% in the Info tuple.
- Ws1 = element(2, Info0),
- Al = beam_utils:combine_heap_needs(Ws1, Ws2),
- Info = setelement(2, Info0, Al),
- I = {Op,Fail,Info,Live,Ss,Dst},
- reverse(Acc, [I|Is]);
-insert_alloc_1([{bs_put,_,_,_}=I|Is], Alloc, Acc) ->
- insert_alloc_1(Is, Alloc, [I|Acc]).
-
-%% opt(Is0) -> Is
-%% Simple peep-hole optimization to move a {move,Any,{x,0}} past
-%% any kill up to the next call instruction. (To give the loader
-%% an opportunity to combine the 'move' and the 'call' instructions.)
-%%
-opt(Is) ->
- opt_1(Is, []).
-
-opt_1([{move,_,{x,0}}=I|Is0], Acc0) ->
- case move_past_kill(Is0, I, Acc0) of
- impossible -> opt_1(Is0, [I|Acc0]);
- {Is,Acc} -> opt_1(Is, Acc)
- end;
-opt_1([I|Is], Acc) ->
- opt_1(Is, [I|Acc]);
-opt_1([], Acc) -> reverse(Acc).
-
-move_past_kill([{kill,Src}|_], {move,Src,_}, _) ->
- impossible;
-move_past_kill([{kill,_}=I|Is], Move, Acc) ->
- move_past_kill(Is, Move, [I|Acc]);
-move_past_kill([{trim,N,_}=I|Is], {move,Src,Dst}=Move, Acc) ->
- case Src of
- {y,Y} when Y < N-> impossible;
- {y,Y} -> {Is,[{move,{y,Y-N},Dst},I|Acc]};
- _ -> {Is,[Move,I|Acc]}
- end;
-move_past_kill(Is, Move, Acc) ->
- {Is,[Move|Acc]}.
diff --git a/lib/compiler/src/beam_jump.erl b/lib/compiler/src/beam_jump.erl
index fbff4cfd79..d3a618d211 100644
--- a/lib/compiler/src/beam_jump.erl
+++ b/lib/compiler/src/beam_jump.erl
@@ -144,13 +144,19 @@ module({Mod,Exp,Attr,Fs0,Lc0}, _Opt) ->
%%
%% NOTE: This function assumes that there are no labels inside blocks.
function({function,Name,Arity,CLabel,Asm0}, Lc0) ->
- Asm1 = eliminate_moves(Asm0),
- {Asm2,Lc} = insert_labels(Asm1, Lc0, []),
- Asm3 = share(Asm2),
- Asm4 = move(Asm3),
- Asm5 = opt(Asm4, CLabel),
- Asm = remove_unused_labels(Asm5),
- {{function,Name,Arity,CLabel,Asm},Lc}.
+ try
+ Asm1 = eliminate_moves(Asm0),
+ {Asm2,Lc} = insert_labels(Asm1, Lc0, []),
+ Asm3 = share(Asm2),
+ Asm4 = move(Asm3),
+ Asm5 = opt(Asm4, CLabel),
+ Asm = remove_unused_labels(Asm5),
+ {{function,Name,Arity,CLabel,Asm},Lc}
+ catch
+ Class:Error:Stack ->
+ io:fwrite("Function: ~w/~w\n", [Name,Arity]),
+ erlang:raise(Class, Error, Stack)
+ end.
%%%
%%% Scan instructions in execution order and remove redundant 'move'
@@ -196,22 +202,19 @@ no_fallthrough([I|_]) ->
is_unreachable_after(I).
already_has_value(Lit, Lbl, Reg, D) ->
- Key = {Lbl,Reg},
case D of
- #{Lbl:=unsafe} ->
- false;
- #{Key:=Lit} ->
+ #{Lbl:={Reg,Lit}} ->
true;
#{} ->
false
end.
update_value_dict([Lit,{f,Lbl}|T], Reg, D0) ->
- Key = {Lbl,Reg},
D = case D0 of
- #{Key := inconsistent} -> D0;
- #{Key := _} -> D0#{Key := inconsistent};
- _ -> D0#{Key => Lit}
+ #{Lbl:=unsafe} -> D0;
+ #{Lbl:={Reg,Lit}} -> D0;
+ #{Lbl:=_} -> D0#{Lbl:=unsafe};
+ #{} -> D0#{Lbl=>{Reg,Lit}}
end,
update_value_dict(T, Reg, D);
update_value_dict([], _, D) -> D.
@@ -407,7 +410,7 @@ find_fixpoint(OptFun, Is0) ->
Is -> find_fixpoint(OptFun, Is)
end.
-opt([{test,_,{f,L}=Lbl,_}=I|[{jump,{f,L}}|_]=Is], Acc0, St0) ->
+opt([{test,_,{f,L}=Lbl,_}=I|[{jump,{f,L}}|_]=Is], Acc, St) ->
%% We have
%% Test Label Ops
%% jump Label
@@ -416,23 +419,20 @@ opt([{test,_,{f,L}=Lbl,_}=I|[{jump,{f,L}}|_]=Is], Acc0, St0) ->
case beam_utils:is_pure_test(I) of
false ->
%% Test is not pure; we must keep it.
- opt(Is, [I|Acc0], label_used(Lbl, St0));
+ opt(Is, [I|Acc], label_used(Lbl, St));
true ->
%% The test is pure and its failure label is the same
%% as in the jump that follows -- thus it is not needed.
- %% Check if any of the previous instructions could also be eliminated.
- {Acc,St} = opt_useless_loads(Acc0, L, St0),
opt(Is, Acc, St)
end;
-opt([{test,_,{f,L}=Lbl,_}=I|[{label,L}|_]=Is], Acc0, St0) ->
+opt([{test,_,{f,L}=Lbl,_}=I|[{label,L}|_]=Is], Acc, St) ->
%% Similar to the above, except we have a fall-through rather than jump
%% Test Label Ops
%% label Label
case beam_utils:is_pure_test(I) of
false ->
- opt(Is, [I|Acc0], label_used(Lbl, St0));
+ opt(Is, [I|Acc], label_used(Lbl, St));
true ->
- {Acc,St} = opt_useless_loads(Acc0, L, St0),
opt(Is, Acc, St)
end;
opt([{test,Test0,{f,L}=Lbl,Ops}=I|[{jump,To}|Is]=Is0], Acc, St) ->
@@ -499,46 +499,6 @@ normalize_replace([{From,To0}|Rest], Replace, Acc) ->
normalize_replace([], _Replace, Acc) ->
maps:from_list(Acc).
-%% After eliminating a test, it might happen, that a register was only used
-%% in this test. Let's check if that was the case and if it was so, we can
-%% eliminate the load into the register completely.
-opt_useless_loads([{block,_}|_]=Is, L, #st{index={lazy,FIs}}=St) ->
- opt_useless_loads(Is, L, St#st{index=beam_utils:index_labels(FIs)});
-opt_useless_loads([{block,Block0}|Is], L, #st{index=Index}=St) ->
- case opt_useless_block_loads(Block0, L, Index) of
- [] ->
- opt_useless_loads(Is, L, St);
- [_|_]=Block ->
- {[{block,Block}|Is],St}
- end;
-%% After eliminating the test and useless blocks, it might happen,
-%% that the previous test could also be eliminated.
-%% It might be that the label was already marked as used, even if ultimately,
-%% it never will be - we can't do much about it at that point, though
-opt_useless_loads([{test,_,{f,L},_}=I|Is], L, St) ->
- case beam_utils:is_pure_test(I) of
- false ->
- {[I|Is],St};
- true ->
- opt_useless_loads(Is, L, St)
- end;
-opt_useless_loads(Is, _L, St) ->
- {Is,St}.
-
-opt_useless_block_loads([{set,[Dst],_,_}=I|Is], L, Index) ->
- BlockJump = [{block,Is},{jump,{f,L}}],
- case beam_utils:is_killed(Dst, BlockJump, Index) of
- true ->
- %% The register is killed and not used, we can remove the load
- opt_useless_block_loads(Is, L, Index);
- false ->
- [I|opt_useless_block_loads(Is, L, Index)]
- end;
-opt_useless_block_loads([I|Is], L, Index) ->
- [I|opt_useless_block_loads(Is, L, Index)];
-opt_useless_block_loads([], _L, _Index) ->
- [].
-
collect_labels(Is, Label, #st{entry=Entry,replace=Replace} = St) ->
collect_labels_1(Is, Label, Entry, Replace, St).
diff --git a/lib/compiler/src/beam_listing.erl b/lib/compiler/src/beam_listing.erl
index 8a0ce5b50a..6121593b11 100644
--- a/lib/compiler/src/beam_listing.erl
+++ b/lib/compiler/src/beam_listing.erl
@@ -66,7 +66,7 @@ module(Stream, [_|_]=Fs) ->
foreach(fun (F) -> io:format(Stream, "~p.\n", [F]) end, Fs).
format_asm([{label,L}|Is]) ->
- [" {label,",integer_to_list(L),"}.\n"|format_asm(Is)];
+ [io_lib:format(" {label,~p}.\n", [L])|format_asm(Is)];
format_asm([I|Is]) ->
[io_lib:format(" ~p", [I]),".\n"|format_asm(Is)];
format_asm([]) -> [].
diff --git a/lib/compiler/src/beam_ssa.erl b/lib/compiler/src/beam_ssa.erl
index 1a2e759965..c5e23d2ae0 100644
--- a/lib/compiler/src/beam_ssa.erl
+++ b/lib/compiler/src/beam_ssa.erl
@@ -198,6 +198,7 @@ no_side_effect(#b_set{op=Op}) ->
has_map_field -> true;
is_nonempty_list -> true;
is_tagged_tuple -> true;
+ make_fun -> true;
put_map -> true;
put_list -> true;
put_tuple -> true;
@@ -578,11 +579,11 @@ used(_) -> [].
-spec definitions(Blocks :: block_map()) -> definition_map().
definitions(Blocks) ->
- beam_ssa:fold_instrs_rpo(fun(#b_set{ dst = Var }=I, Acc) ->
- maps:put(Var, I, Acc);
- (_Terminator, Acc) ->
- Acc
- end, [0], #{}, Blocks).
+ fold_instrs_rpo(fun(#b_set{ dst = Var }=I, Acc) ->
+ maps:put(Var, I, Acc);
+ (_Terminator, Acc) ->
+ Acc
+ end, [0], #{}, Blocks).
-spec uses(Blocks :: block_map()) -> usage_map().
uses(Blocks) ->
@@ -592,7 +593,7 @@ uses(Blocks) ->
From :: [label()],
Blocks :: block_map().
uses(From, Blocks) ->
- beam_ssa:fold_rpo(fun fold_uses_block/3, From, #{}, Blocks).
+ fold_rpo(fun fold_uses_block/3, From, #{}, Blocks).
fold_uses_block(Lbl, #b_blk{is=Is,last=Last}, UseMap0) ->
F = fun(I, UseMap) ->
@@ -600,7 +601,7 @@ fold_uses_block(Lbl, #b_blk{is=Is,last=Last}, UseMap0) ->
Uses0 = maps:get(Var, Acc, []),
Uses = [{Lbl, I} | Uses0],
maps:put(Var, Uses, Acc)
- end, UseMap, beam_ssa:used(I))
+ end, UseMap, used(I))
end,
F(Last, foldl(F, UseMap0, Is)).
@@ -797,8 +798,8 @@ split_blocks_1([L|Ls], P, Blocks0, Count0) ->
BefBlk = Blk#b_blk{is=Bef,last=Br},
NewBlk = Blk#b_blk{is=Aft},
Blocks1 = Blocks0#{L:=BefBlk,NewLbl=>NewBlk},
- Successors = beam_ssa:successors(NewBlk),
- Blocks = beam_ssa:update_phi_labels(Successors, L, NewLbl, Blocks1),
+ Successors = successors(NewBlk),
+ Blocks = update_phi_labels(Successors, L, NewLbl, Blocks1),
split_blocks_1([NewLbl|Ls], P, Blocks, Count);
no ->
split_blocks_1(Ls, P, Blocks0, Count0)
diff --git a/lib/compiler/src/beam_ssa_bsm.erl b/lib/compiler/src/beam_ssa_bsm.erl
index 2efeb6b5b6..9631bf3334 100644
--- a/lib/compiler/src/beam_ssa_bsm.erl
+++ b/lib/compiler/src/beam_ssa_bsm.erl
@@ -354,6 +354,11 @@ amb_get_alias(#b_var{}=Arg, Lbl, State) ->
error ->
{Arg, State}
end;
+amb_get_alias(#b_remote{mod=Mod0,name=Name0}=Arg0, Lbl, State0) ->
+ {Mod, State1} = amb_get_alias(Mod0, Lbl, State0),
+ {Name, State} = amb_get_alias(Name0, Lbl, State1),
+ Arg = Arg0#b_remote{mod=Mod,name=Name},
+ {Arg, State};
amb_get_alias(Arg, _Lbl, State) ->
{Arg, State}.
@@ -391,15 +396,22 @@ amb_insert_promotions(Blocks0, State) ->
Block = maps:get(Lbl, Blocks),
Alias = Promotion#b_set.dst,
- {Before, After} = splitwith(fun(#b_set{args=Args}) ->
- not member(Alias, Args)
- end, Block#b_blk.is),
+ {Before, After} = splitwith(
+ fun(#b_set{args=Args}) ->
+ not is_var_in_args(Alias, Args)
+ end, Block#b_blk.is),
Is = Before ++ [Promotion | After],
maps:put(Lbl, Block#b_blk{is=Is}, Blocks)
end,
maps:fold(F, Blocks0, State#amb.promotions).
+is_var_in_args(Var, [Var | _]) -> true;
+is_var_in_args(Var, [#b_remote{name=Var} | _]) -> true;
+is_var_in_args(Var, [#b_remote{mod=Var} | _]) -> true;
+is_var_in_args(Var, [_ | Args]) -> is_var_in_args(Var, Args);
+is_var_in_args(_Var, []) -> false.
+
%%%
%%% Subpasses
%%%
@@ -732,6 +744,10 @@ aca_cs_args([Arg | Args], VRs) ->
aca_cs_args([], _VRs) ->
[].
+aca_cs_arg(#b_remote{mod=Mod0,name=Name0}=Rem, VRs) ->
+ Mod = aca_cs_arg(Mod0, VRs),
+ Name = aca_cs_arg(Name0, VRs),
+ Rem#b_remote{mod=Mod,name=Name};
aca_cs_arg(Arg, VRs) ->
case VRs of
#{ Arg := New } -> New;
diff --git a/lib/compiler/src/beam_ssa_codegen.erl b/lib/compiler/src/beam_ssa_codegen.erl
index 1c7563faa0..3c14062d0b 100644
--- a/lib/compiler/src/beam_ssa_codegen.erl
+++ b/lib/compiler/src/beam_ssa_codegen.erl
@@ -231,7 +231,7 @@ need_heap_never(_) -> false.
need_heap_blks([{L,#cg_blk{is=Is0}=Blk0}|Bs], H0, Acc) ->
{Is1,H1} = need_heap_is(reverse(Is0), H0, []),
- {Ns,H} = need_heap_terminator(Bs, H1),
+ {Ns,H} = need_heap_terminator(Bs, L, H1),
Is = Ns ++ Is1,
Blk = Blk0#cg_blk{is=Is},
need_heap_blks(Bs, H, [{L,Blk}|Acc]);
@@ -241,6 +241,13 @@ need_heap_blks([], H, Acc) ->
need_heap_is([#cg_alloc{words=Words}=Alloc0|Is], N, Acc) ->
Alloc = Alloc0#cg_alloc{words=add_heap_words(N, Words)},
need_heap_is(Is, #need{}, [Alloc|Acc]);
+need_heap_is([#cg_set{anno=Anno,op=bs_init}=I0|Is], N, Acc) ->
+ Alloc = case need_heap_need(N) of
+ [#cg_alloc{words=Need}] -> alloc(Need);
+ [] -> 0
+ end,
+ I = I0#cg_set{anno=Anno#{alloc=>Alloc}},
+ need_heap_is(Is, #need{}, [I|Acc]);
need_heap_is([#cg_set{op=Op,args=Args}=I|Is], N, Acc) ->
case classify_heap_need(Op, Args) of
{put,Words} ->
@@ -256,11 +263,31 @@ need_heap_is([#cg_set{op=Op,args=Args}=I|Is], N, Acc) ->
need_heap_is([], N, Acc) ->
{Acc,N}.
-need_heap_terminator([{_,#cg_blk{last=#cg_br{succ=Same,fail=Same}}}|_], N) ->
+need_heap_terminator([{_,#cg_blk{last=#cg_br{succ=L,fail=L}}}|_], L, N) ->
+ %% Fallthrough.
{[],N};
-need_heap_terminator([{_,#cg_blk{}}|_], N) ->
+need_heap_terminator([{_,#cg_blk{is=Is,last=#cg_br{succ=L}}}|_], L, N) ->
+ case need_heap_need(N) of
+ [] ->
+ {[],#need{}};
+ [_|_]=Alloc ->
+ %% If the preceding instructions are a binary construction,
+ %% hoist the allocation and incorporate into the bs_init
+ %% instruction.
+ case reverse(Is) of
+ [#cg_set{op=succeeded},#cg_set{op=bs_init}|_] ->
+ {[],N};
+ [#cg_set{op=bs_put}|_] ->
+ {[],N};
+ _ ->
+ %% Not binary construction. Must emit an allocation
+ %% instruction in this block.
+ {Alloc,#need{}}
+ end
+ end;
+need_heap_terminator([{_,#cg_blk{}}|_], _, N) ->
{need_heap_need(N),#need{}};
-need_heap_terminator([], H) ->
+need_heap_terminator([], _, H) ->
{need_heap_need(H),#need{}}.
need_heap_need(#need{h=0,f=0}) -> [];
@@ -1041,12 +1068,13 @@ cg_block([#cg_set{op=bs_init,dst=Dst0,args=Args0,anno=Anno}=I,
#cg_set{op=succeeded,dst=Bool}], {Bool,Fail0}, St) ->
Fail = bif_fail(Fail0),
Line = line(Anno),
+ Alloc = map_get(alloc, Anno),
[#b_literal{val=Kind}|Args1] = Args0,
case Kind of
new ->
[Dst,Size,{integer,Unit}] = beam_args([Dst0|Args1], St),
Live = get_live(I),
- {[Line|cg_bs_init(Dst, Size, Unit, Live, Fail)],St};
+ {[Line|cg_bs_init(Dst, Size, Alloc, Unit, Live, Fail)],St};
private_append ->
[Dst,Src,Bits,{integer,Unit}] = beam_args([Dst0|Args1], St),
Flags = {field_flags,[]},
@@ -1056,7 +1084,7 @@ cg_block([#cg_set{op=bs_init,dst=Dst0,args=Args0,anno=Anno}=I,
[Dst,Src,Bits,{integer,Unit}] = beam_args([Dst0|Args1], St),
Flags = {field_flags,[]},
Live = get_live(I),
- Is = [Line,{bs_append,Fail,Bits,0,Live,Unit,Src,Flags,Dst}],
+ Is = [Line,{bs_append,Fail,Bits,Alloc,Live,Unit,Src,Flags,Dst}],
{Is,St}
end;
cg_block([#cg_set{anno=Anno,op=bs_start_match,dst=Ctx0,args=[Bin0]}=I,
@@ -1204,6 +1232,12 @@ cg_copy_1([#cg_set{dst=Dst0,args=Args}|T], St) ->
end;
cg_copy_1([], _St) -> [].
+-define(IS_LITERAL(Val), (Val =:= nil orelse
+ element(1, Val) =:= integer orelse
+ element(1, Val) =:= float orelse
+ element(1, Val) =:= atom orelse
+ element(1, Val) =:= literal)).
+
bif_to_test('and', [V1,V2], Fail) ->
[{test,is_eq_exact,Fail,[V1,{atom,true}]},
{test,is_eq_exact,Fail,[V2,{atom,true}]}];
@@ -1217,15 +1251,99 @@ bif_to_test('or', [V1,V2], {f,Lbl}=Fail) when Lbl =/= 0 ->
bif_to_test('not', [Var], Fail) ->
[{test,is_eq_exact,Fail,[Var,{atom,false}]}];
bif_to_test(Name, Args, Fail) ->
- [beam_utils:bif_to_test(Name, Args, Fail)].
+ [bif_to_test_1(Name, Args, Fail)].
+
+bif_to_test_1(is_atom, [_]=Ops, Fail) ->
+ {test,is_atom,Fail,Ops};
+bif_to_test_1(is_boolean, [_]=Ops, Fail) ->
+ {test,is_boolean,Fail,Ops};
+bif_to_test_1(is_binary, [_]=Ops, Fail) ->
+ {test,is_binary,Fail,Ops};
+bif_to_test_1(is_bitstring,[_]=Ops, Fail) ->
+ {test,is_bitstr,Fail,Ops};
+bif_to_test_1(is_float, [_]=Ops, Fail) ->
+ {test,is_float,Fail,Ops};
+bif_to_test_1(is_function, [_]=Ops, Fail) ->
+ {test,is_function,Fail,Ops};
+bif_to_test_1(is_function, [_,_]=Ops, Fail) ->
+ {test,is_function2,Fail,Ops};
+bif_to_test_1(is_integer, [_]=Ops, Fail) ->
+ {test,is_integer,Fail,Ops};
+bif_to_test_1(is_list, [_]=Ops, Fail) ->
+ {test,is_list,Fail,Ops};
+bif_to_test_1(is_map, [_]=Ops, Fail) ->
+ {test,is_map,Fail,Ops};
+bif_to_test_1(is_number, [_]=Ops, Fail) ->
+ {test,is_number,Fail,Ops};
+bif_to_test_1(is_pid, [_]=Ops, Fail) ->
+ {test,is_pid,Fail,Ops};
+bif_to_test_1(is_port, [_]=Ops, Fail) ->
+ {test,is_port,Fail,Ops};
+bif_to_test_1(is_reference, [_]=Ops, Fail) ->
+ {test,is_reference,Fail,Ops};
+bif_to_test_1(is_tuple, [_]=Ops, Fail) ->
+ {test,is_tuple,Fail,Ops};
+bif_to_test_1('=<', [A,B], Fail) ->
+ {test,is_ge,Fail,[B,A]};
+bif_to_test_1('>', [A,B], Fail) ->
+ {test,is_lt,Fail,[B,A]};
+bif_to_test_1('<', [_,_]=Ops, Fail) ->
+ {test,is_lt,Fail,Ops};
+bif_to_test_1('>=', [_,_]=Ops, Fail) ->
+ {test,is_ge,Fail,Ops};
+bif_to_test_1('==', [C,A], Fail) when ?IS_LITERAL(C) ->
+ {test,is_eq,Fail,[A,C]};
+bif_to_test_1('==', [_,_]=Ops, Fail) ->
+ {test,is_eq,Fail,Ops};
+bif_to_test_1('/=', [C,A], Fail) when ?IS_LITERAL(C) ->
+ {test,is_ne,Fail,[A,C]};
+bif_to_test_1('/=', [_,_]=Ops, Fail) ->
+ {test,is_ne,Fail,Ops};
+bif_to_test_1('=:=', [C,A], Fail) when ?IS_LITERAL(C) ->
+ {test,is_eq_exact,Fail,[A,C]};
+bif_to_test_1('=:=', [_,_]=Ops, Fail) ->
+ {test,is_eq_exact,Fail,Ops};
+bif_to_test_1('=/=', [C,A], Fail) when ?IS_LITERAL(C) ->
+ {test,is_ne_exact,Fail,[A,C]};
+bif_to_test_1('=/=', [_,_]=Ops, Fail) ->
+ {test,is_ne_exact,Fail,Ops}.
opt_call_moves(Is0, Arity) ->
{Moves0,Is} = splitwith(fun({move,_,_}) -> true;
+ ({kill,_}) -> true;
(_) -> false
end, Is0),
Moves = opt_call_moves_1(Moves0, Arity),
Moves ++ Is.
+opt_call_moves_1([{move,Src,{x,_}=Tmp}=M1|[{kill,_}|_]=Is], Arity) ->
+ %% There could be a {move,Tmp,{x,0}} instruction after the
+ %% kill/1 instructions (moved to there by opt_move_to_x0/1).
+ case splitwith(fun({kill,_}) -> true;
+ (_) -> false
+ end, Is) of
+ {Kills,[{move,{x,_}=Tmp,{x,0}}=M2]} ->
+ %% The two move/2 instructions (M1 and M2) can be combined
+ %% to one. The question is, though, is it safe to place
+ %% them after the kill/1 instructions?
+ case is_killed(Src, Kills, Arity) of
+ true ->
+ %% Src (a Y register) is killed by one of the
+ %% kill/1 instructions. Thus M1 and M2
+ %% must be placed before the kill/1 instructions
+ %% (essentially undoing what opt_move_to_x0/1
+ %% did, which turned out to be a pessimization
+ %% in this case).
+ opt_call_moves_1([M1,M2|Kills], Arity);
+ false ->
+ %% Src is not killed by any of the kill/1
+ %% instructions. Thus it is safe to place
+ %% M1 and M2 after the kill/1 instructions.
+ opt_call_moves_1(Kills++[M1,M2], Arity)
+ end;
+ {_,_} ->
+ [M1|Is]
+ end;
opt_call_moves_1([{move,Src,{x,_}=Tmp}=M1,{move,Tmp,Dst}=M2|Is], Arity) ->
case is_killed(Tmp, Is, Arity) of
true ->
@@ -1239,6 +1357,10 @@ opt_call_moves_1([M|Ms], Arity) ->
[M|opt_call_moves_1(Ms, Arity)];
opt_call_moves_1([], _Arity) -> [].
+is_killed(Y, [{kill,Y}|_], _) ->
+ true;
+is_killed(R, [{kill,_}|Is], Arity) ->
+ is_killed(R, Is, Arity);
is_killed(R, [{move,R,_}|_], _) ->
false;
is_killed(R, [{move,_,R}|_], _) ->
@@ -1246,7 +1368,9 @@ is_killed(R, [{move,_,R}|_], _) ->
is_killed(R, [{move,_,_}|Is], Arity) ->
is_killed(R, Is, Arity);
is_killed({x,X}, [], Arity) ->
- X >= Arity.
+ X >= Arity;
+is_killed({y,_}, [], _) ->
+ false.
cg_alloc(#cg_alloc{stack=none,words=#need{h=0,f=0}}, _St) ->
[];
@@ -1527,13 +1651,13 @@ cg_bs_put(Fail, [{atom,Type},{literal,Flags}|Args]) ->
[{Op,Fail,{field_flags,Flags},Src}]
end.
-cg_bs_init(Dst, Size0, Unit, Live, Fail) ->
+cg_bs_init(Dst, Size0, Alloc, Unit, Live, Fail) ->
Op = case Unit of
1 -> bs_init_bits;
8 -> bs_init2
end,
Size = cg_bs_init_size(Size0),
- [{Op,Fail,Size,0,Live,{field_flags,[]},Dst}].
+ [{Op,Fail,Size,Alloc,Live,{field_flags,[]},Dst}].
cg_bs_init_size({x,_}=R) -> R;
cg_bs_init_size({y,_}=R) -> R;
@@ -1652,12 +1776,41 @@ phi_copies([#b_set{dst=Dst,args=PhiArgs}|Sets], L) ->
[#cg_set{op=copy,dst=Dst,args=CopyArgs}|phi_copies(Sets, L)];
phi_copies([], _) -> [].
+%% opt_move_to_x0([Instruction]) -> [Instruction].
+%% Simple peep-hole optimization to move a {move,Any,{x,0}} past
+%% any kill up to the next call instruction. (To give the loader
+%% an opportunity to combine the 'move' and the 'call' instructions.)
+
+opt_move_to_x0(Moves) ->
+ opt_move_to_x0(Moves, []).
+
+opt_move_to_x0([{move,_,{x,0}}=I|Is0], Acc0) ->
+ case move_past_kill(Is0, I, Acc0) of
+ impossible -> opt_move_to_x0(Is0, [I|Acc0]);
+ {Is,Acc} -> opt_move_to_x0(Is, Acc)
+ end;
+opt_move_to_x0([I|Is], Acc) ->
+ opt_move_to_x0(Is, [I|Acc]);
+opt_move_to_x0([], Acc) -> reverse(Acc).
+
+move_past_kill([{kill,Src}|_], {move,Src,_}, _) ->
+ impossible;
+move_past_kill([{kill,_}=I|Is], Move, Acc) ->
+ move_past_kill(Is, Move, [I|Acc]);
+move_past_kill(Is, Move, Acc) ->
+ {Is,[Move|Acc]}.
+
%% setup_args(Args, Anno, Context) -> [Instruction].
%% setup_args(Args) -> [Instruction].
%% Set up X registers for a call.
setup_args(Args, Anno, none, St) ->
- setup_args(Args) ++ kill_yregs(Anno, St);
+ case {setup_args(Args),kill_yregs(Anno, St)} of
+ {Moves,[]} ->
+ Moves;
+ {Moves,Kills} ->
+ opt_move_to_x0(Moves ++ Kills)
+ end;
setup_args(Args, _, _, _) ->
setup_args(Args).
diff --git a/lib/compiler/src/beam_ssa_dead.erl b/lib/compiler/src/beam_ssa_dead.erl
index c20652580d..067d9a6741 100644
--- a/lib/compiler/src/beam_ssa_dead.erl
+++ b/lib/compiler/src/beam_ssa_dead.erl
@@ -135,7 +135,8 @@ 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])},
+ RelOp = {'=:=',Bool,Lit},
+ St = St0#st{rel_op=RelOp},
#b_br{bool=#b_literal{val=true},succ=L} =
shortcut(L0, bind_var(Bool, Lit, Bs), St#st{target=one_way}),
[{Lit,L}|shortcut_switch(T, Bool, Bs, St0)];
@@ -388,41 +389,43 @@ eval_terminator(#b_switch{arg=Arg,fail=Fail,list=List}=Sw, Bs, St) ->
%% 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.
+ %% Try optimizing the switch.
+ case eval_switch(List, Arg, St, Fail) of
+ none ->
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
+ To when is_integer(To) ->
+ %% Either one of the values in the switch
+ %% matched a previous value in a '=:=' test, or
+ %% none of the values matched a previous test.
+ #b_br{bool=#b_literal{val=true},succ=To,fail=To}
end
end;
eval_terminator(#b_ret{}, _Bs, _St) ->
none.
-eval_switch([{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} ->
+eval_switch(List, Arg, #st{rel_op={_,Arg,_}=PrevOp}, Fail) ->
+ %% There is a previous relational operator testing the same variable.
+ %% Optimization may be possible.
+ eval_switch_1(List, Arg, PrevOp, Fail);
+eval_switch(_, _, _, _) ->
+ %% There is either no previous relational operator, or it tests
+ %% a different variable. Nothing to optimize.
+ none.
+
+eval_switch_1([{Lit,Lbl}|T], Arg, PrevOp, Fail) ->
+ RelOp = {'=:=',Arg,Lit},
+ case will_succeed(PrevOp, RelOp) of
+ yes ->
%% Success. This branch will always be taken.
- Lbl
+ Lbl;
+ no ->
+ %% This branch will never be taken.
+ eval_switch_1(T, Arg, PrevOp, Fail);
+ maybe ->
+ %% This label could be reached.
+ eval_switch_1(T, Arg, PrevOp, none)
end;
-eval_switch([], _Arg, _St, Fail) ->
+eval_switch_1([], _Arg, _PrevOp, Fail) ->
%% Fail is now either the failure label or 'none'.
Fail.
diff --git a/lib/compiler/src/beam_ssa_funs.erl b/lib/compiler/src/beam_ssa_funs.erl
new file mode 100644
index 0000000000..38df50fd74
--- /dev/null
+++ b/lib/compiler/src/beam_ssa_funs.erl
@@ -0,0 +1,149 @@
+%%
+%% %CopyrightBegin%
+%%
+%% Copyright Ericsson AB 2018. All Rights Reserved.
+%%
+%% Licensed under the Apache License, Version 2.0 (the "License");
+%% you may not use this file except in compliance with the License.
+%% You may obtain a copy of the License at
+%%
+%% http://www.apache.org/licenses/LICENSE-2.0
+%%
+%% Unless required by applicable law or agreed to in writing, software
+%% distributed under the License is distributed on an "AS IS" BASIS,
+%% WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+%% See the License for the specific language governing permissions and
+%% limitations under the License.
+%%
+%% %CopyrightEnd%
+%%
+
+%%%
+%%% If a fun is defined locally and only used for calls, it can be replaced
+%%% with direct calls to the relevant function. This greatly speeds up "named
+%%% functions" (which rely on make_fun to recreate themselves) and macros that
+%%% wrap their body in a fun.
+%%%
+
+-module(beam_ssa_funs).
+
+-export([module/2]).
+
+-include("beam_ssa.hrl").
+
+-import(lists, [foldl/3]).
+
+-spec module(Module, Options) -> Result when
+ Module :: beam_ssa:b_module(),
+ Options :: [compile:option()],
+ Result :: {ok, beam_ssa:b_module()}.
+
+module(#b_module{body=Fs0}=Module, _Opts) ->
+ Trampolines = foldl(fun find_trampolines/2, #{}, Fs0),
+ Fs = [lfo(F, Trampolines) || F <- Fs0],
+ {ok, Module#b_module{body=Fs}}.
+
+%% If a function does absolutely nothing beyond calling another function with
+%% the same arguments in the same order, we can shave off a call by short-
+%% circuiting it.
+find_trampolines(#b_function{args=Args,bs=Blocks}=F, Trampolines) ->
+ case maps:get(0, Blocks) of
+ #b_blk{is=[#b_set{op=call,
+ args=[#b_local{}=Actual | Args],
+ dst=Dst}],
+ last=#b_ret{arg=Dst}} ->
+ {_, Name, Arity} = beam_ssa:get_anno(func_info, F),
+ Trampoline = #b_local{name=#b_literal{val=Name},arity=Arity},
+ maps:put(Trampoline, Actual, Trampolines);
+ _ ->
+ Trampolines
+ end.
+
+lfo(#b_function{bs=Blocks0}=F, Trampolines) ->
+ Linear0 = beam_ssa:linearize(Blocks0),
+ Linear = lfo_optimize(Linear0, lfo_analyze(Linear0, #{}), Trampolines),
+ F#b_function{bs=maps:from_list(Linear)}.
+
+%% Gather a map of the locally defined funs that are only used for calls.
+lfo_analyze([{_L,#b_blk{is=Is,last=Last}}|Bs], LFuns0) ->
+ LFuns = lfo_analyze_last(Last, lfo_analyze_is(Is, LFuns0)),
+ lfo_analyze(Bs, LFuns);
+lfo_analyze([], LFuns) ->
+ LFuns.
+
+lfo_analyze_is([#b_set{op=make_fun,
+ dst=Dst,
+ args=[#b_local{} | FreeVars]}=Def | Is],
+ LFuns0) ->
+ LFuns = maps:put(Dst, Def, maps:without(FreeVars, LFuns0)),
+ lfo_analyze_is(Is, LFuns);
+lfo_analyze_is([#b_set{op=call,
+ args=[Fun | CallArgs]} | Is],
+ LFuns) when is_map_key(Fun, LFuns) ->
+ #b_set{args=[#b_local{arity=Arity} | FreeVars]} = maps:get(Fun, LFuns),
+ case length(CallArgs) + length(FreeVars) of
+ Arity ->
+ lfo_analyze_is(Is, maps:without(CallArgs, LFuns));
+ _ ->
+ %% This will `badarity` at runtime, and it's easier to disable the
+ %% optimization than to simulate it.
+ lfo_analyze_is(Is, maps:without([Fun | CallArgs], LFuns))
+ end;
+lfo_analyze_is([#b_set{args=Args} | Is], LFuns) when map_size(LFuns) =/= 0 ->
+ %% We disqualify funs that are used outside calls because this forces them
+ %% to be created anyway, and the slight performance gain from direct calls
+ %% is not enough to offset the potential increase in stack frame size (the
+ %% free variables need to be kept alive until the call).
+ %%
+ %% This is also a kludge to make HiPE work, as the latter will generate
+ %% code with the assumption that the functions referenced in a make_fun
+ %% will only be used by funs, which will not be the case if we mix it with
+ %% direct calls. See cerl_cconv.erl for details.
+ %%
+ %% Future optimizations like delaying fun creation until use may require us
+ %% to copy affected functions so that HiPE gets its own to play with (until
+ %% HiPE is fixed anyway).
+ lfo_analyze_is(Is, maps:without(Args, LFuns));
+lfo_analyze_is([_ | Is], LFuns) ->
+ lfo_analyze_is(Is, LFuns);
+lfo_analyze_is([], LFuns) ->
+ LFuns.
+
+lfo_analyze_last(#b_switch{arg=Arg}, LFuns) ->
+ maps:remove(Arg, LFuns);
+lfo_analyze_last(#b_ret{arg=Arg}, LFuns) ->
+ maps:remove(Arg, LFuns);
+lfo_analyze_last(_, LFuns) ->
+ LFuns.
+
+%% Replace all calls of suitable funs with a direct call to their
+%% implementation. Liveness optimization will get rid of the make_fun
+%% instruction.
+lfo_optimize(Linear, LFuns, _Trampolines) when map_size(LFuns) =:= 0 ->
+ Linear;
+lfo_optimize(Linear, LFuns, Trampolines) ->
+ lfo_optimize_1(Linear, LFuns, Trampolines).
+
+lfo_optimize_1([{L,#b_blk{is=Is0}=Blk}|Bs], LFuns, Trampolines) ->
+ Is = lfo_optimize_is(Is0, LFuns, Trampolines),
+ [{L,Blk#b_blk{is=Is}} | lfo_optimize_1(Bs, LFuns, Trampolines)];
+lfo_optimize_1([], _LFuns, _Trampolines) ->
+ [].
+
+lfo_optimize_is([#b_set{op=call,
+ args=[Fun | CallArgs]}=Call0 | Is],
+ LFuns, Trampolines) when is_map_key(Fun, LFuns) ->
+ #b_set{args=[Local | FreeVars]} = maps:get(Fun, LFuns),
+ Args = [lfo_short_circuit(Local, Trampolines) | CallArgs ++ FreeVars],
+ Call = beam_ssa:add_anno(local_fun_opt, Fun, Call0#b_set{args=Args}),
+ [Call | lfo_optimize_is(Is, LFuns, Trampolines)];
+lfo_optimize_is([I | Is], LFuns, Trampolines) ->
+ [I | lfo_optimize_is(Is, LFuns, Trampolines)];
+lfo_optimize_is([], _LFuns, _Trampolines) ->
+ [].
+
+lfo_short_circuit(Call, Trampolines) ->
+ case maps:find(Call, Trampolines) of
+ {ok, Other} -> lfo_short_circuit(Other, Trampolines);
+ error -> Call
+ end.
diff --git a/lib/compiler/src/beam_ssa_pre_codegen.erl b/lib/compiler/src/beam_ssa_pre_codegen.erl
index 36137ef046..c60c6da9ea 100644
--- a/lib/compiler/src/beam_ssa_pre_codegen.erl
+++ b/lib/compiler/src/beam_ssa_pre_codegen.erl
@@ -131,6 +131,10 @@ passes(Opts) ->
?PASS(find_yregs),
?PASS(reserve_yregs),
+ %% Handle legacy binary match instruction that don't
+ %% accept a Y register as destination.
+ ?PASS(legacy_bs),
+
%% Improve reuse of Y registers to potentially
%% reduce the size of the stack frame.
?PASS(copy_retval),
@@ -612,6 +616,59 @@ bs_subst_ctx(#b_var{}=Var, CtxChain) ->
bs_subst_ctx(Other, _CtxChain) ->
Other.
+%% legacy_bs(St0) -> St.
+%% Binary matching instructions in OTP 21 and earlier don't support
+%% a Y register as destination. If St#st.use_bsm3 is false,
+%% we will need to rewrite those instructions so that the result
+%% is first put in an X register and then moved to a Y register
+%% if the operation succeeded.
+
+legacy_bs(#st{use_bsm3=false,ssa=Blocks0,cnt=Count0,res=Res}=St) ->
+ IsYreg = maps:from_list([{V,true} || {V,{y,_}} <- Res]),
+ Linear0 = beam_ssa:linearize(Blocks0),
+ {Linear,Count} = legacy_bs(Linear0, IsYreg, Count0, #{}, []),
+ Blocks = maps:from_list(Linear),
+ St#st{ssa=Blocks,cnt=Count};
+legacy_bs(#st{use_bsm3=true}=St) -> St.
+
+legacy_bs([{L,Blk}|Bs], IsYreg, Count0, Copies0, Acc) ->
+ #b_blk{is=Is0,last=Last} = Blk,
+ Is1 = case Copies0 of
+ #{L:=Copy} -> [Copy|Is0];
+ #{} -> Is0
+ end,
+ {Is,Count,Copies} = legacy_bs_is(Is1, Last, IsYreg, Count0, Copies0, []),
+ legacy_bs(Bs, IsYreg, Count, Copies, [{L,Blk#b_blk{is=Is}}|Acc]);
+legacy_bs([], _IsYreg, Count, _Copies, Acc) ->
+ {Acc,Count}.
+
+legacy_bs_is([#b_set{op=Op,dst=Dst}=I0,
+ #b_set{op=succeeded,dst=SuccDst,args=[Dst]}=SuccI0],
+ Last, IsYreg, Count0, Copies0, Acc) ->
+ NeedsFix = is_map_key(Dst, IsYreg) andalso
+ case Op of
+ bs_get -> true;
+ bs_init -> true;
+ _ -> false
+ end,
+ case NeedsFix of
+ true ->
+ TempDst = #b_var{name={'@bs_temp_dst',Count0}},
+ Count = Count0 + 1,
+ I = I0#b_set{dst=TempDst},
+ SuccI = SuccI0#b_set{args=[TempDst]},
+ Copy = #b_set{op=copy,dst=Dst,args=[TempDst]},
+ #b_br{bool=SuccDst,succ=SuccL} = Last,
+ Copies = Copies0#{SuccL=>Copy},
+ legacy_bs_is([], Last, IsYreg, Count, Copies, [SuccI,I|Acc]);
+ false ->
+ legacy_bs_is([], Last, IsYreg, Count0, Copies0, [SuccI0,I0|Acc])
+ end;
+legacy_bs_is([I|Is], Last, IsYreg, Count, Copies, Acc) ->
+ legacy_bs_is(Is, Last, IsYreg, Count, Copies, [I|Acc]);
+legacy_bs_is([], _Last, _IsYreg, Count, Copies, Acc) ->
+ {reverse(Acc),Count,Copies}.
+
%% sanitize(St0) -> St.
%% Remove constructs that can cause problems later:
%%
@@ -652,21 +709,24 @@ sanitize([], Count, Blocks0, Values) ->
false -> remove_unreachable(Ls, Blocks, Reachable, [])
end,Count}.
-sanitize_is([#b_set{op=get_map_element,
- args=[#b_literal{}=Map,Key]}=I0|Is],
- Count0, Values, _Changed, Acc) ->
- {MapVar,Count} = new_var('@ssa_map', Count0),
- I = I0#b_set{args=[MapVar,Key]},
- Copy = #b_set{op=copy,dst=MapVar,args=[Map]},
- sanitize_is(Is, Count, Values, true, [I,Copy|Acc]);
+sanitize_is([#b_set{op=get_map_element,args=Args0}=I0|Is],
+ Count0, Values, Changed, Acc) ->
+ case sanitize_args(Args0, Values) of
+ [#b_literal{}=Map,Key] ->
+ %% Bind the literal map to a variable.
+ {MapVar,Count} = new_var('@ssa_map', Count0),
+ I = I0#b_set{args=[MapVar,Key]},
+ Copy = #b_set{op=copy,dst=MapVar,args=[Map]},
+ sanitize_is(Is, Count, Values, true, [I,Copy|Acc]);
+ [_,_]=Args0 ->
+ sanitize_is(Is, Count0, Values, Changed, [I0|Acc]);
+ [_,_]=Args ->
+ I = I0#b_set{args=Args},
+ sanitize_is(Is, Count0, Values, Changed, [I|Acc])
+ end;
sanitize_is([#b_set{op=Op,dst=Dst,args=Args0}=I0|Is0],
- Count, Values, Changed, Acc) ->
- Args = map(fun(Var) ->
- case Values of
- #{Var:=New} -> New;
- #{} -> Var
- end
- end, Args0),
+ Count, Values, Changed0, Acc) ->
+ Args = sanitize_args(Args0, Values),
case sanitize_instr(Op, Args, I0) of
{value,Value0} ->
Value = #b_literal{val=Value0},
@@ -674,7 +734,9 @@ sanitize_is([#b_set{op=Op,dst=Dst,args=Args0}=I0|Is0],
{ok,I} ->
sanitize_is(Is0, Count, Values, true, [I|Acc]);
ok ->
- sanitize_is(Is0, Count, Values, Changed, [I0|Acc])
+ I = I0#b_set{args=Args},
+ Changed = Changed0 orelse Args =/= Args0,
+ sanitize_is(Is0, Count, Values, Changed, [I|Acc])
end;
sanitize_is([], Count, Values, Changed, Acc) ->
case Changed of
@@ -684,6 +746,14 @@ sanitize_is([], Count, Values, Changed, Acc) ->
no_change
end.
+sanitize_args(Args, Values) ->
+ map(fun(Var) ->
+ case Values of
+ #{Var:=New} -> New;
+ #{} -> Var
+ end
+ end, Args).
+
sanitize_instr({bif,Bif}, [#b_literal{val=Lit}], _I) ->
case erl_bifs:is_pure(erlang, Bif, 1) of
false ->
@@ -973,11 +1043,11 @@ need_frame_1([#b_set{op=call,args=[Func|_]}|Is], Context) ->
#b_remote{} ->
%% This is an apply(), which always needs a frame.
true;
- #b_var{} ->
- %% A fun call always needs a frame.
- true;
+ #b_local{} ->
+ Context =:= body orelse Is =/= [];
_ ->
- Context =:= body orelse Is =/= []
+ %% A fun call always needs a frame.
+ true
end;
need_frame_1([I|Is], Context) ->
beam_ssa:clobbers_xregs(I) orelse need_frame_1(Is, Context);
@@ -2141,7 +2211,13 @@ linear_scan(#st{intervals=Intervals0,res=Res}=St0) ->
Free = init_free(maps:to_list(Res)),
Intervals1 = [init_interval(Int, Res) || Int <- Intervals0],
Intervals = sort(Intervals1),
- IsReserved = fun (#i{reg=Reg}) -> Reg =/= none end,
+ IsReserved = fun(#i{reg=Reg}) ->
+ case Reg of
+ none -> false;
+ {prefer,{_,_}} -> false;
+ {_,_} -> true
+ end
+ end,
{UnhandledRes,Unhandled} = partition(IsReserved, Intervals),
L = #l{unhandled_res=UnhandledRes,
unhandled_any=Unhandled,free=Free},
diff --git a/lib/compiler/src/beam_ssa_type.erl b/lib/compiler/src/beam_ssa_type.erl
index 18e6e73a46..95fc3bb0e9 100644
--- a/lib/compiler/src/beam_ssa_type.erl
+++ b/lib/compiler/src/beam_ssa_type.erl
@@ -27,9 +27,10 @@
-define(UNICODE_INT, #t_integer{elements={0,16#10FFFF}}).
--record(d, {ds :: #{beam_ssa:var_name():=beam_ssa:b_set()},
+-record(d, {ds :: #{beam_ssa:b_var():=beam_ssa:b_set()},
ls :: #{beam_ssa:label():=type_db()},
- sub :: #{beam_ssa:var_name():=beam_ssa:value()}
+ once :: cerl_sets:set(beam_ssa:b_var()),
+ sub :: #{beam_ssa:b_var():=beam_ssa:value()}
}).
-define(ATOM_SET_SIZE, 5).
@@ -56,13 +57,15 @@
Block :: beam_ssa:b_blk().
opt(Linear, Args) ->
+ UsedOnce = used_once(Linear, Args),
Ts = maps:from_list([{V,any} || #b_var{}=V <- Args]),
FakeCall = #b_set{op=call,args=[#b_remote{mod=#b_literal{val=unknown},
name=#b_literal{val=unknown},
arity=0}]},
Defs = maps:from_list([{Var,FakeCall#b_set{dst=Var}} ||
#b_var{}=Var <- Args]),
- D = #d{ds=Defs,ls=#{0=>Ts},sub=#{}},
+ D = #d{ds=Defs,ls=#{0=>Ts,?BADARG_BLOCK=>#{}},
+ once=UsedOnce,sub=#{}},
opt_1(Linear, D).
opt_1([{L,Blk}|Bs], #d{ls=Ls}=D) ->
@@ -425,16 +428,43 @@ opt_terminator(#b_ret{}=Ret, _Ts, _Ds) -> Ret.
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}, Ts, D0) ->
- D = update_successor_bool(Bool, false, Fail, Ts, D0),
- SuccTs = infer_types(Bool, Ts, D0),
- update_successor_bool(Bool, true, Succ, SuccTs, D);
-update_successors(#b_switch{arg=#b_var{}=V,fail=Fail,list=List}, Ts, D0) ->
- D = update_successor(Fail, Ts, D0),
- foldl(fun({Val,S}, A) ->
- T = get_type(Val, Ts),
- update_successor(S, Ts#{V=>T}, A)
- end, D, List);
+update_successors(#b_br{bool=#b_var{}=Bool,succ=Succ,fail=Fail}, Ts0, D0) ->
+ case cerl_sets:is_element(Bool, D0#d.once) of
+ true ->
+ %% This variable is defined in this block and is only
+ %% referenced by this br terminator. Therefore, there is
+ %% no need to include the type database passed on to the
+ %% successors of this block.
+ Ts = maps:remove(Bool, Ts0),
+ D = update_successor(Fail, Ts, D0),
+ SuccTs = infer_types(Bool, Ts, D0),
+ update_successor(Succ, SuccTs, D);
+ false ->
+ D = update_successor_bool(Bool, false, Fail, Ts0, D0),
+ SuccTs = infer_types(Bool, Ts0, D0),
+ update_successor_bool(Bool, true, Succ, SuccTs, D)
+ end;
+update_successors(#b_switch{arg=#b_var{}=V,fail=Fail,list=List}, Ts0, D0) ->
+ case cerl_sets:is_element(V, D0#d.once) of
+ true ->
+ %% This variable is defined in this block and is only
+ %% referenced by this switch terminator. Therefore, there is
+ %% no need to include the type database passed on to the
+ %% successors of this block.
+ Ts = maps:remove(V, Ts0),
+ D = update_successor(Fail, Ts, D0),
+ F = fun({_Val,S}, A) ->
+ update_successor(S, Ts, A)
+ end,
+ foldl(F, D, List);
+ false ->
+ D = update_successor(Fail, Ts0, D0),
+ F = fun({Val,S}, A) ->
+ T = get_type(Val, Ts0),
+ update_successor(S, Ts0#{V=>T}, A)
+ end,
+ foldl(F, D, List)
+ end;
update_successors(#b_ret{}, _Ts, D) -> D.
update_successor_bool(#b_var{}=Var, BoolValue, S, Ts, D) ->
@@ -447,6 +477,11 @@ update_successor_bool(#b_var{}=Var, BoolValue, S, Ts, D) ->
update_successor(S, Ts, D)
end.
+update_successor(?BADARG_BLOCK, _Ts, #d{}=D) ->
+ %% We KNOW that no variables are used in the ?BADARG_BLOCK,
+ %% so there is no need to update the type information. That
+ %% can be a huge timesaver for huge functions.
+ D;
update_successor(S, Ts0, #d{ls=Ls}=D) ->
case Ls of
#{S:=Ts1} ->
@@ -766,6 +801,48 @@ simplify_not(#b_br{bool=#b_var{}=V,succ=Succ,fail=Fail}=Br0, Ts, Ds) ->
Br0
end.
+%%%
+%%% Calculate the set of variables that are only used once in the
+%%% block that they are defined in. That will allow us to discard type
+%%% information for variables that will never be referenced by the
+%%% successor blocks, potentially improving compilation times.
+%%%
+
+used_once(Linear, Args) ->
+ Map0 = used_once_1(reverse(Linear), #{}),
+ Map = maps:without(Args, Map0),
+ cerl_sets:from_list(maps:keys(Map)).
+
+used_once_1([{L,#b_blk{is=Is,last=Last}}|Bs], Uses0) ->
+ Uses = used_once_2([Last|reverse(Is)], L, Uses0),
+ used_once_1(Bs, Uses);
+used_once_1([], Uses) -> Uses.
+
+used_once_2([I|Is], L, Uses0) ->
+ Uses = used_once_uses(beam_ssa:used(I), L, Uses0),
+ case I of
+ #b_set{dst=Dst} ->
+ case Uses of
+ #{Dst:=[L]} ->
+ used_once_2(Is, L, Uses);
+ #{} ->
+ used_once_2(Is, L, maps:remove(Dst, Uses))
+ end;
+ _ ->
+ used_once_2(Is, L, Uses)
+ end;
+used_once_2([], _, Uses) -> Uses.
+
+used_once_uses([V|Vs], L, Uses) ->
+ case Uses of
+ #{V:=Us} ->
+ used_once_uses(Vs, L, Uses#{V:=[L|Us]});
+ #{} ->
+ used_once_uses(Vs, L, Uses#{V=>[L]})
+ end;
+used_once_uses([], _, Uses) -> Uses.
+
+
get_types(Values, Ts) ->
[get_type(Val, Ts) || Val <- Values].
-spec get_type(beam_ssa:value(), type_db()) -> type().
diff --git a/lib/compiler/src/beam_trim.erl b/lib/compiler/src/beam_trim.erl
index 1acbedd45b..51ff580a7a 100644
--- a/lib/compiler/src/beam_trim.erl
+++ b/lib/compiler/src/beam_trim.erl
@@ -21,12 +21,11 @@
-module(beam_trim).
-export([module/2]).
--import(lists, [reverse/1,reverse/2,splitwith/2,sort/1]).
+-import(lists, [any/2,member/2,reverse/1,reverse/2,splitwith/2,sort/1]).
-record(st,
- {safe :: gb_sets:set(beam_asm:label()), %Safe labels.
- lbl :: beam_utils:code_index() %Code at each label.
- }).
+ {safe :: cerl_sets:set(beam_asm:label()) %Safe labels.
+ }).
-spec module(beam_utils:module_code(), [compile:option()]) ->
{'ok',beam_utils:module_code()}.
@@ -36,10 +35,15 @@ module({Mod,Exp,Attr,Fs0,Lc}, _Opts) ->
{ok,{Mod,Exp,Attr,Fs,Lc}}.
function({function,Name,Arity,CLabel,Is0}) ->
- %%ok = io:fwrite("~w: ~p\n", [?LINE,{Name,Arity}]),
- St = #st{safe=safe_labels(Is0, []),lbl=beam_utils:index_labels(Is0)},
- Is = trim(Is0, St, []),
- {function,Name,Arity,CLabel,Is}.
+ try
+ St = #st{safe=safe_labels(Is0, [])},
+ Is = trim(Is0, St, []),
+ {function,Name,Arity,CLabel,Is}
+ catch
+ Class:Error:Stack ->
+ io:fwrite("Function: ~w/~w\n", [Name,Arity]),
+ erlang:raise(Class, Error, Stack)
+ end.
trim([{kill,_}|_]=Is0, St, Acc) ->
{Kills0,Is1} = splitwith(fun({kill,_}) -> true;
@@ -47,14 +51,33 @@ trim([{kill,_}|_]=Is0, St, Acc) ->
end, Is0),
Kills = sort(Kills0),
try
- {FrameSize,Layout} = frame_layout(Is1, Kills, St),
- Configs = trim_instructions(Layout),
- try_remap(Configs, Is1, FrameSize)
- of
+ %% Find out the size and layout of the stack frame.
+ %% Example of a layout:
+ %%
+ %% [{kill,{y,0}},{dead,{y,1},{live,{y,2}},{kill,{y,3}}]
+ %%
+ %% That means that y0 and y3 are to be killed, that y1
+ %% has been killed previously, and that y2 is live.
+ {FrameSize,Layout} = frame_layout(Is1, Kills, St),
+
+ %% Calculate all recipes that are not worse in terms
+ %% of estimated execution time. The recipes are ordered
+ %% in descending order from how much they trim.
+ Recipes = trim_recipes(Layout),
+
+ %% Try the recipes in order. A recipe may not work out because
+ %% a register that was previously killed may be
+ %% resurrected. If that happens, the next recipe, which trims
+ %% less, will be tried.
+ try_remap(Recipes, Is1, FrameSize)
+ of
{Is,TrimInstr} ->
+ %% One of the recipes was applied.
trim(Is, St, reverse(TrimInstr)++Acc)
catch
not_possible ->
+ %% No recipe worked out. Use the original kill
+ %% instructions.
trim(Is1, St, reverse(Kills, Acc))
end;
trim([I|Is], St, Acc) ->
@@ -62,34 +85,42 @@ trim([I|Is], St, Acc) ->
trim([], _, Acc) ->
reverse(Acc).
-%% trim_instructions([{kill,R}|{live,R}|{dead,R}]) -> {[Instruction],MapFun}
-%% Figure out the sequence of moves and trim to use.
+%% trim_recipes([{kill,R}|{live,R}|{dead,R}]) -> [Recipe].
+%% Recipe = {Kills,NumberToTrim,Moves}
+%% Kills = [{kill,Y}]
+%% Moves = [{move,SrcY,DstY}]
+%%
+%% Calculate how to best trim the stack and kill the correct
+%% Y registers. Return a list of possible recipes. The best
+%% recipe (the one that trims the most) is first in the list.
+%% All of the recipes are no worse in estimated execution time
+%% than the original sequences of kill instructions.
-trim_instructions(Layout) ->
+trim_recipes(Layout) ->
Cost = length([I || {kill,_}=I <- Layout]),
- trim_instructions_1(Layout, 0, [], {Cost,[]}).
+ trim_recipes_1(Layout, 0, [], {Cost,[]}).
-trim_instructions_1([{kill,{y,Trim0}}|Ks], Trim0, Moves, Config0) ->
+trim_recipes_1([{kill,{y,Trim0}}|Ks], Trim0, Moves, Recipes0) ->
Trim = Trim0 + 1,
- Config = save_config(Ks, Trim, Moves, Config0),
- trim_instructions_1(Ks, Trim, Moves, Config);
-trim_instructions_1([{dead,{y,Trim0}}|Ks], Trim0, Moves, Config0) ->
+ Recipes = save_recipe(Ks, Trim, Moves, Recipes0),
+ trim_recipes_1(Ks, Trim, Moves, Recipes);
+trim_recipes_1([{dead,{y,Trim0}}|Ks], Trim0, Moves, Recipes0) ->
Trim = Trim0 + 1,
- Config = save_config(Ks, Trim, Moves, Config0),
- trim_instructions_1(Ks, Trim, Moves, Config);
-trim_instructions_1([{live,{y,Trim0}=Src}|Ks0], Trim0, Moves0, Config0) ->
+ Recipes = save_recipe(Ks, Trim, Moves, Recipes0),
+ trim_recipes_1(Ks, Trim, Moves, Recipes);
+trim_recipes_1([{live,{y,Trim0}=Src}|Ks0], Trim0, Moves0, Recipes0) ->
case take_last_dead(Ks0) of
none ->
- {_,ConfigList} = Config0,
- ConfigList;
+ {_,RecipesList} = Recipes0,
+ RecipesList;
{Dst,Ks} ->
Trim = Trim0 + 1,
Moves = [{move,Src,Dst}|Moves0],
- Config = save_config(Ks, Trim, Moves, Config0),
- trim_instructions_1(Ks, Trim, Moves, Config)
+ Recipes = save_recipe(Ks, Trim, Moves, Recipes0),
+ trim_recipes_1(Ks, Trim, Moves, Recipes)
end;
-trim_instructions_1([], _, _, {_,ConfigList}) ->
- ConfigList.
+trim_recipes_1([], _, _, {_,RecipesList}) ->
+ RecipesList.
take_last_dead(L) ->
take_last_dead_1(reverse(L)).
@@ -100,28 +131,48 @@ take_last_dead_1([{dead,Reg}|Is]) ->
{Reg,reverse(Is)};
take_last_dead_1(_) -> none.
-save_config(Ks, Trim, Moves, {MaxCost,Acc}=Config) ->
- case config_cost(Ks, Moves) of
- Cost when Cost =< MaxCost ->
- {MaxCost,[{Ks,Trim,Moves}|Acc]};
+save_recipe(Ks, Trim, Moves, {MaxCost,Acc}=Recipes) ->
+ case recipe_cost(Ks, Moves) of
+ Cost when Cost =< MaxCost ->
+ %% The price is right.
+ {MaxCost,[{Ks,Trim,Moves}|Acc]};
_Cost ->
- Config
+ %% Too expensive.
+ Recipes
end.
-config_cost(Ks, Moves) ->
+recipe_cost(Ks, Moves) ->
%% We estimate that a {move,{y,_},{y,_}} instruction is roughly twice as
%% expensive as a {kill,{y,_}} instruction. A {trim,_} instruction is
%% roughly as expensive as a {kill,{y,_}} instruction.
- config_cost_1(Ks, 1+2*length(Moves)).
+ recipe_cost_1(Ks, 1+2*length(Moves)).
-config_cost_1([{kill,_}|Ks], Cost) ->
- config_cost_1(Ks, Cost+1);
-config_cost_1([_|Ks], Cost) ->
- config_cost_1(Ks, Cost);
-config_cost_1([], Cost) -> Cost.
+recipe_cost_1([{kill,_}|Ks], Cost) ->
+ recipe_cost_1(Ks, Cost+1);
+recipe_cost_1([_|Ks], Cost) ->
+ recipe_cost_1(Ks, Cost);
+recipe_cost_1([], Cost) -> Cost.
+
+%% try_remap([Recipe], [Instruction], FrameSize) ->
+%% {[Instruction],[TrimInstruction]}.
+%% Try to renumber Y registers in the instruction stream. The
+%% first rececipe that works will be used.
+%%
+%% This function will issue a `not_possible` exception if none
+%% of the recipes were possible to apply.
+
+try_remap([R|Rs], Is, FrameSize) ->
+ {TrimInstr,Map} = expand_recipe(R, FrameSize),
+ try
+ {remap(Is, Map, []),TrimInstr}
+ catch
+ throw:not_possible ->
+ try_remap(Rs, Is, FrameSize)
+ end;
+try_remap([], _, _) -> throw(not_possible).
-expand_config({Layout,Trim,Moves}, FrameSize) ->
+expand_recipe({Layout,Trim,Moves}, FrameSize) ->
Kills = [Kill || {kill,_}=Kill <- Layout],
{Kills++reverse(Moves, [{trim,Trim,FrameSize-Trim}]),create_map(Trim, Moves)}.
@@ -132,16 +183,16 @@ create_map(Trim, []) ->
(Any) -> Any
end;
create_map(Trim, Moves) ->
- GbTree0 = [{Src,Dst-Trim} || {move,{y,Src},{y,Dst}} <- Moves],
- GbTree = gb_trees:from_orddict(sort(GbTree0)),
- IllegalTargets = gb_sets:from_list([Dst || {move,_,{y,Dst}} <- Moves]),
+ Map0 = [{Src,Dst-Trim} || {move,{y,Src},{y,Dst}} <- Moves],
+ Map = maps:from_list(Map0),
+ IllegalTargets = cerl_sets:from_list([Dst || {move,_,{y,Dst}} <- Moves]),
fun({y,Y0}) when Y0 < Trim ->
- case gb_trees:lookup(Y0, GbTree) of
- {value,Y} -> {y,Y};
- none -> throw(not_possible)
- end;
+ case Map of
+ #{Y0:=Y} -> {y,Y};
+ #{} -> throw(not_possible)
+ end;
({y,Y}) ->
- case gb_sets:is_element(Y, IllegalTargets) of
+ case cerl_sets:is_element(Y, IllegalTargets) of
true -> throw(not_possible);
false -> {y,Y-Trim}
end;
@@ -149,19 +200,15 @@ create_map(Trim, Moves) ->
(Any) -> Any
end.
-try_remap([C|Cs], Is, FrameSize) ->
- {TrimInstr,Map} = expand_config(C, FrameSize),
- try
- {remap(Is, Map, []),TrimInstr}
- catch
- throw:not_possible ->
- try_remap(Cs, Is, FrameSize)
- end;
-try_remap([], _, _) -> throw(not_possible).
-
remap([{block,Bl0}|Is], Map, Acc) ->
Bl = remap_block(Bl0, Map, []),
remap(Is, Map, [{block,Bl}|Acc]);
+remap([{bs_get_tail,Src,Dst,Live}|Is], Map, Acc) ->
+ I = {bs_get_tail,Map(Src),Map(Dst),Live},
+ remap(Is, Map, [I|Acc]);
+remap([{bs_set_position,Src1,Src2}|Is], Map, Acc) ->
+ I = {bs_set_position,Map(Src1),Map(Src2)},
+ remap(Is, Map, [I|Acc]);
remap([{call_fun,_}=I|Is], Map, Acc) ->
remap(Is, Map, [I|Acc]);
remap([{call,_,_}=I|Is], Map, Acc) ->
@@ -205,35 +252,66 @@ remap([return|_]=Is, _, Acc) ->
reverse(Acc, Is);
remap([{line,_}=I|Is], Map, Acc) ->
remap(Is, Map, [I|Acc]).
-
+
remap_block([{set,Ds0,Ss0,Info}|Is], Map, Acc) ->
Ds = [Map(D) || D <- Ds0],
Ss = [Map(S) || S <- Ss0],
remap_block(Is, Map, [{set,Ds,Ss,Info}|Acc]);
remap_block([], _, Acc) -> reverse(Acc).
-
-safe_labels([{label,L},{line,_},{badmatch,{Tag,_}}|Is], Acc) when Tag =/= y ->
- safe_labels(Is, [L|Acc]);
-safe_labels([{label,L},{line,_},{case_end,{Tag,_}}|Is], Acc) when Tag =/= y ->
- safe_labels(Is, [L|Acc]);
-safe_labels([{label,L},{line,_},if_end|Is], Acc) ->
- safe_labels(Is, [L|Acc]);
-safe_labels([{label,L},
- {block,[{set,[{x,0}],[{Tag,_}],move}]},
- {line,_},
- {call_ext,1,{extfunc,erlang,error,1}}|Is], Acc) when Tag =/= y ->
- safe_labels(Is, [L|Acc]);
+
+%% safe_labels([Instruction], Accumulator) -> gb_set()
+%% Build a gb_set of safe labels. The code at a safe
+%% label does not depend on the values in a specific
+%% Y register, only that all Y registers are initialized
+%% so that it safe to scan the stack when an exception
+%% is generated.
+%%
+%% In other words, code at a safe label will continue
+%% to work if Y registers have been renumbered and
+%% the size of the stack frame has changed.
+
+safe_labels([{label,L}|Is], Acc) ->
+ case is_safe_label(Is) of
+ true -> safe_labels(Is, [L|Acc]);
+ false -> safe_labels(Is, Acc)
+ end;
safe_labels([_|Is], Acc) ->
safe_labels(Is, Acc);
-safe_labels([], Acc) -> gb_sets:from_list(Acc).
+safe_labels([], Acc) -> cerl_sets:from_list(Acc).
+
+is_safe_label([{line,_}|Is]) ->
+ is_safe_label(Is);
+is_safe_label([{badmatch,{Tag,_}}|_]) ->
+ Tag =/= y;
+is_safe_label([{case_end,{Tag,_}}|_]) ->
+ Tag =/= y;
+is_safe_label([{try_case_end,{Tag,_}}|_]) ->
+ Tag =/= y;
+is_safe_label([if_end|_]) ->
+ true;
+is_safe_label([{block,Bl}|Is]) ->
+ is_safe_label_block(Bl) andalso is_safe_label(Is);
+is_safe_label([{call_ext,_,{extfunc,M,F,A}}|_]) ->
+ erl_bifs:is_exit_bif(M, F, A);
+is_safe_label(_) -> false.
+
+is_safe_label_block([{set,Ds,Ss,_}|Is]) ->
+ IsYreg = fun({y,_}) -> true;
+ (_) -> false
+ end,
+ %% This instruction is safe if the instruction
+ %% neither reads or writes Y registers.
+ not (any(IsYreg, Ss) orelse any(IsYreg, Ds)) andalso
+ is_safe_label_block(Is);
+is_safe_label_block([]) -> true.
%% frame_layout([Instruction], [{kill,_}], St) ->
%% [{kill,Reg} | {live,Reg} | {dead,Reg}]
%% Figure out the layout of the stack frame.
-frame_layout(Is, Kills, #st{safe=Safe,lbl=D}) ->
+frame_layout(Is, Kills, #st{safe=Safe}) ->
N = frame_size(Is, Safe),
- IsKilled = fun(R) -> beam_utils:is_not_used(R, Is, D) end,
+ IsKilled = fun(R) -> is_not_used(R, Is) end,
{N,frame_layout_1(Kills, 0, N, IsKilled, [])}.
frame_layout_1([{kill,{y,Y}}=I|Ks], Y, N, IsKilled, Acc) ->
@@ -253,6 +331,11 @@ frame_layout_2(Is) -> reverse(Is).
%% frame_size([Instruction], SafeLabels) -> FrameSize
%% Find out the frame size by looking at the code that follows.
+%%
+%% Implicitly, also check that the instructions are a straight
+%% sequence of code that ends in a return. Any branches are
+%% to safe labels (i.e., the code at those labels don't depend
+%% on the contents of any Y register).
frame_size([{block,_}|Is], Safe) ->
frame_size(Is, Safe);
@@ -285,15 +368,92 @@ frame_size([{make_fun2,_,_,_,_}|Is], Safe) ->
frame_size(Is, Safe);
frame_size([{get_map_elements,{f,L},_,_}|Is], Safe) ->
frame_size_branch(L, Is, Safe);
-frame_size([{deallocate,N}|_], _) -> N;
+frame_size([{deallocate,N}|_], _) ->
+ N;
frame_size([{line,_}|Is], Safe) ->
frame_size(Is, Safe);
+frame_size([{bs_set_position,_,_}|Is], Safe) ->
+ frame_size(Is, Safe);
+frame_size([{bs_get_tail,_,_,_}|Is], Safe) ->
+ frame_size(Is, Safe);
frame_size(_, _) -> throw(not_possible).
frame_size_branch(0, Is, Safe) ->
frame_size(Is, Safe);
frame_size_branch(L, Is, Safe) ->
- case gb_sets:is_member(L, Safe) of
+ case cerl_sets:is_element(L, Safe) of
false -> throw(not_possible);
true -> frame_size(Is, Safe)
end.
+
+%% is_not_used(Y, [Instruction]) -> true|false.
+%% Test whether the value of Y is unused in the instruction sequence.
+%% Return true if the value of Y is not used, and false if it is used.
+%%
+%% This function handles the same instructions as frame_size/2. It
+%% assumes that any labels in the instructions are safe labels.
+
+is_not_used(Y, [{apply,_}|Is]) ->
+ is_not_used(Y, Is);
+is_not_used(Y, [{bif,_,{f,_},Ss,Dst}|Is]) ->
+ is_not_used_ss_dst(Y, Ss, Dst, Is);
+is_not_used(Y, [{block,Bl}|Is]) ->
+ case is_not_used_block(Y, Bl) of
+ used -> false;
+ killed -> true;
+ transparent -> is_not_used(Y, Is)
+ end;
+is_not_used(Y, [{bs_get_tail,Src,Dst,_}|Is]) ->
+ is_not_used_ss_dst(Y, [Src], Dst, Is);
+is_not_used(Y, [{bs_init,_,_,_,Ss,Dst}|Is]) ->
+ is_not_used_ss_dst(Y, Ss, Dst, Is);
+is_not_used(Y, [{bs_put,{f,_},_,Ss}|Is]) ->
+ not member(Y, Ss) andalso is_not_used(Y, Is);
+is_not_used(Y, [{bs_set_position,Src1,Src2}|Is]) ->
+ Y =/= Src1 andalso Y =/= Src2 andalso
+ is_not_used(Y, Is);
+is_not_used(Y, [{call,_,_}|Is]) ->
+ is_not_used(Y, Is);
+is_not_used(Y, [{call_ext,_,_}=I|Is]) ->
+ beam_jump:is_exit_instruction(I) orelse is_not_used(Y, Is);
+is_not_used(Y, [{call_fun,_}|Is]) ->
+ is_not_used(Y, Is);
+is_not_used(_Y, [{deallocate,_}|_]) ->
+ true;
+is_not_used(Y, [{gc_bif,_,{f,_},_Live,Ss,Dst}|Is]) ->
+ is_not_used_ss_dst(Y, Ss, Dst, Is);
+is_not_used(Y, [{get_map_elements,{f,_},S,{list,List}}|Is]) ->
+ {Ss,Ds} = beam_utils:split_even(List),
+ case member(Y, [S|Ss]) of
+ true ->
+ false;
+ false ->
+ member(Y, Ds) orelse is_not_used(Y, Is)
+ end;
+is_not_used(Y, [{kill,Yreg}|Is]) ->
+ Y =:= Yreg orelse is_not_used(Y, Is);
+is_not_used(Y, [{line,_}|Is]) ->
+ is_not_used(Y, Is);
+is_not_used(Y, [{make_fun2,_,_,_,_}|Is]) ->
+ is_not_used(Y, Is);
+is_not_used(Y, [{test,_,_,Ss}|Is]) ->
+ not member(Y, Ss) andalso is_not_used(Y, Is);
+is_not_used(Y, [{test,_Op,{f,_},_Live,Ss,Dst}|Is]) ->
+ is_not_used_ss_dst(Y, Ss, Dst, Is).
+
+is_not_used_block(Y, [{set,Ds,Ss,_}|Is]) ->
+ case member(Y, Ss) of
+ true ->
+ used;
+ false ->
+ case member(Y, Ds) of
+ true ->
+ killed;
+ false ->
+ is_not_used_block(Y, Is)
+ end
+ end;
+is_not_used_block(_Y, []) -> transparent.
+
+is_not_used_ss_dst(Y, Ss, Dst, Is) ->
+ not member(Y, Ss) andalso (Y =:= Dst orelse is_not_used(Y, Is)).
diff --git a/lib/compiler/src/beam_utils.erl b/lib/compiler/src/beam_utils.erl
index 626e041ea0..6e6574c0b3 100644
--- a/lib/compiler/src/beam_utils.erl
+++ b/lib/compiler/src/beam_utils.erl
@@ -18,27 +18,16 @@
%% %CopyrightEnd%
%%
%% Purpose : Common utilities used by several optimization passes.
-%%
+%%
-module(beam_utils).
--export([is_killed/3,is_killed_at/3,is_not_used/3,
- empty_label_index/0,index_label/3,index_labels/1,replace_labels/4,
- code_at/2,bif_to_test/3,is_pure_test/1,
- combine_heap_needs/2,
- split_even/1
- ]).
+-export([replace_labels/4,is_pure_test/1,split_even/1]).
-export_type([code_index/0,module_code/0,instruction/0]).
--import(lists, [flatmap/2,map/2,member/2,sort/1,reverse/1]).
-
--define(is_const(Val), (Val =:= nil orelse
- element(1, Val) =:= integer orelse
- element(1, Val) =:= float orelse
- element(1, Val) =:= atom orelse
- element(1, Val) =:= literal)).
+-import(lists, [map/2,reverse/1]).
-%% instruction() describes all instructions that are used during optimzation
+%% instruction() describes all instructions that are used during optimization
%% (from beam_a to beam_z).
-type instruction() :: atom() | tuple().
@@ -54,97 +43,6 @@
-type fail() :: beam_asm:fail() | 'fail'.
-type test() :: {'test',atom(),fail(),[beam_asm:src()]} |
{'test',atom(),fail(),integer(),list(),beam_asm:reg()}.
--type result_cache() :: gb_trees:tree(beam_asm:label(), 'killed' | 'used').
-
--record(live,
- {lbl :: code_index(), %Label to code index.
- res :: result_cache()}). %Result cache for each label.
-
-%% is_killed(Register, [Instruction], State) -> true|false
-%% Determine whether a register is killed by the instruction sequence.
-%% If true is returned, it means that the register will not be
-%% referenced in ANY way (not even indirectly by an allocate instruction);
-%% i.e. it is OK to enter the instruction sequence with Register
-%% containing garbage.
-%%
-%% The state (constructed by index_instructions/1) is used to allow us
-%% to determine the kill state across branches.
-
--spec is_killed(beam_asm:reg(), [instruction()], code_index()) -> boolean().
-
-is_killed(R, Is, D) ->
- St = #live{lbl=D,res=gb_trees:empty()},
- case check_liveness(R, Is, St) of
- {killed,_} -> true;
- {exit_not_used,_} -> false;
- {_,_} -> false
- end.
-
-%% is_killed_at(Reg, Lbl, State) -> true|false
-%% Determine whether Reg is killed at label Lbl.
-
--spec is_killed_at(beam_asm:reg(), beam_asm:label(), code_index()) -> boolean().
-
-is_killed_at(R, Lbl, D) when is_integer(Lbl) ->
- St0 = #live{lbl=D,res=gb_trees:empty()},
- case check_liveness_at(R, Lbl, St0) of
- {killed,_} -> true;
- {exit_not_used,_} -> false;
- {_,_} -> false
- end.
-
-%% is_not_used(Register, [Instruction], State) -> true|false
-%% Determine whether a register is never used in the instruction sequence
-%% (it could still be referenced by an allocate instruction, meaning that
-%% it MUST be initialized, but that its value does not matter).
-%% The state is used to allow us to determine the usage state
-%% across branches.
-
--spec is_not_used(beam_asm:reg(), [instruction()], code_index()) -> boolean().
-
-is_not_used(R, Is, D) ->
- St = #live{lbl=D,res=gb_trees:empty()},
- case check_liveness(R, Is, St) of
- {used,_} -> false;
- {exit_not_used,_} -> true;
- {_,_} -> true
- end.
-
-%% index_labels(FunctionIs) -> State
-%% Index the instruction sequence so that we can quickly
-%% look up the instruction following a specific label.
-
--spec index_labels([instruction()]) -> code_index().
-
-index_labels(Is) ->
- index_labels_1(Is, []).
-
-%% empty_label_index() -> State
-%% Create an empty label index.
-
--spec empty_label_index() -> code_index().
-
-empty_label_index() ->
- gb_trees:empty().
-
-%% index_label(Label, [Instruction], State) -> State
-%% Add an index for a label.
-
--spec index_label(beam_asm:label(), [instruction()], code_index()) ->
- code_index().
-
-index_label(Lbl, Is0, Acc) ->
- Is = drop_labels(Is0),
- gb_trees:enter(Lbl, Is, Acc).
-
-
-%% code_at(Label, State) -> [I].
-%% Retrieve the code at the given label.
-
--spec code_at(beam_asm:label(), code_index()) -> [instruction()].
-
-code_at(L, Ll) ->
- gb_trees:get(L, Ll).
%% replace_labels(FunctionIs, Tail, ReplaceDb, Fallback) -> FunctionIs.
%% Replace all labels in instructions according to the ReplaceDb.
@@ -158,44 +56,6 @@ code_at(L, Ll) ->
replace_labels(Is, Acc, D, Fb) ->
replace_labels_1(Is, Acc, D, Fb).
-%% bif_to_test(Bif, [Op], Fail) -> {test,Test,Fail,[Op]}
-%% Convert a BIF to a test. Fail if not possible.
-
--spec bif_to_test(atom(), list(), fail()) -> test().
-
-bif_to_test(is_atom, [_]=Ops, Fail) -> {test,is_atom,Fail,Ops};
-bif_to_test(is_boolean, [_]=Ops, Fail) -> {test,is_boolean,Fail,Ops};
-bif_to_test(is_binary, [_]=Ops, Fail) -> {test,is_binary,Fail,Ops};
-bif_to_test(is_bitstring,[_]=Ops, Fail) -> {test,is_bitstr,Fail,Ops};
-bif_to_test(is_float, [_]=Ops, Fail) -> {test,is_float,Fail,Ops};
-bif_to_test(is_function, [_]=Ops, Fail) -> {test,is_function,Fail,Ops};
-bif_to_test(is_function, [_,_]=Ops, Fail) -> {test,is_function2,Fail,Ops};
-bif_to_test(is_integer, [_]=Ops, Fail) -> {test,is_integer,Fail,Ops};
-bif_to_test(is_list, [_]=Ops, Fail) -> {test,is_list,Fail,Ops};
-bif_to_test(is_map, [_]=Ops, Fail) -> {test,is_map,Fail,Ops};
-bif_to_test(is_number, [_]=Ops, Fail) -> {test,is_number,Fail,Ops};
-bif_to_test(is_pid, [_]=Ops, Fail) -> {test,is_pid,Fail,Ops};
-bif_to_test(is_port, [_]=Ops, Fail) -> {test,is_port,Fail,Ops};
-bif_to_test(is_reference, [_]=Ops, Fail) -> {test,is_reference,Fail,Ops};
-bif_to_test(is_tuple, [_]=Ops, Fail) -> {test,is_tuple,Fail,Ops};
-bif_to_test('=<', [A,B], Fail) -> {test,is_ge,Fail,[B,A]};
-bif_to_test('>', [A,B], Fail) -> {test,is_lt,Fail,[B,A]};
-bif_to_test('<', [_,_]=Ops, Fail) -> {test,is_lt,Fail,Ops};
-bif_to_test('>=', [_,_]=Ops, Fail) -> {test,is_ge,Fail,Ops};
-bif_to_test('==', [C,A], Fail) when ?is_const(C) ->
- {test,is_eq,Fail,[A,C]};
-bif_to_test('==', [_,_]=Ops, Fail) -> {test,is_eq,Fail,Ops};
-bif_to_test('/=', [C,A], Fail) when ?is_const(C) ->
- {test,is_ne,Fail,[A,C]};
-bif_to_test('/=', [_,_]=Ops, Fail) -> {test,is_ne,Fail,Ops};
-bif_to_test('=:=', [C,A], Fail) when ?is_const(C) ->
- {test,is_eq_exact,Fail,[A,C]};
-bif_to_test('=:=', [_,_]=Ops, Fail) -> {test,is_eq_exact,Fail,Ops};
-bif_to_test('=/=', [C,A], Fail) when ?is_const(C) ->
- {test,is_ne_exact,Fail,[A,C]};
-bif_to_test('=/=', [_,_]=Ops, Fail) -> {test,is_ne_exact,Fail,Ops}.
-
-
%% is_pure_test({test,Op,Fail,Ops}) -> true|false.
%% Return 'true' if the test instruction does not modify any
%% registers and/or bit syntax matching state.
@@ -215,22 +75,9 @@ is_pure_test({test,test_arity,_,[_,_]}) -> true;
is_pure_test({test,has_map_fields,_,[_|_]}) -> true;
is_pure_test({test,is_bitstr,_,[_]}) -> true;
is_pure_test({test,is_function2,_,[_,_]}) -> true;
-is_pure_test({test,Op,_,Ops}) ->
+is_pure_test({test,Op,_,Ops}) ->
erl_internal:new_type_test(Op, length(Ops)).
-%% combine_heap_needs(HeapNeed1, HeapNeed2) -> HeapNeed
-%% Combine the heap need for two allocation instructions.
-
--type heap_need_tag() :: 'floats' | 'words'.
--type heap_need() :: non_neg_integer() |
- {'alloc',[{heap_need_tag(),non_neg_integer()}]}.
--spec combine_heap_needs(heap_need(), heap_need()) -> heap_need().
-
-combine_heap_needs(H1, H2) when is_integer(H1), is_integer(H2) ->
- H1 + H2;
-combine_heap_needs(H1, H2) ->
- {alloc,combine_alloc_lists([H1,H2])}.
-
%% split_even/1
%% [1,2,3,4,5,6] -> {[1,3,5],[2,4,6]}
@@ -242,438 +89,6 @@ split_even(Rs) -> split_even(Rs, [], []).
%%% Local functions.
%%%
-
-%% check_liveness(Reg, [Instruction], #live{}) ->
-%% {killed | not_used | used, #live{}}
-%% Find out whether Reg is used or killed in instruction sequence.
-%%
-%% killed - Reg is assigned or killed by an allocation instruction.
-%% not_used - the value of Reg is not used, but Reg must not be garbage
-%% exit_not_used - the value of Reg is not used, but must not be garbage
-%% because the stack will be scanned because an
-%% exit BIF will raise an exception
-%% used - Reg is used
-
-check_liveness({fr,_}, _, St) ->
- %% Conservatively always consider the floating point register used.
- {used,St};
-check_liveness(R, [{block,Blk}|Is], St0) ->
- case check_liveness_block(R, Blk, St0) of
- {transparent,St1} ->
- check_liveness(R, Is, St1);
- {alloc_used,St1} ->
- %% Used by an allocating instruction, but value not referenced.
- %% Must check the rest of the instructions.
- not_used(check_liveness(R, Is, St1));
- {Other,_}=Res when is_atom(Other) ->
- Res
- end;
-check_liveness(R, [{label,_}|Is], St) ->
- check_liveness(R, Is, St);
-check_liveness(R, [{test,_,{f,Fail},As}|Is], St0) ->
- case member(R, As) of
- true ->
- {used,St0};
- false ->
- case check_liveness_at(R, Fail, St0) of
- {killed,St1} ->
- check_liveness(R, Is, St1);
- {exit_not_used,St1} ->
- not_used(check_liveness(R, Is, St1));
- {not_used,St1} ->
- not_used(check_liveness(R, Is, St1));
- {used,_}=Used ->
- Used
- end
- end;
-check_liveness(R, [{test,Op,Fail,Live,Ss,Dst}|Is], St) ->
- %% Check this instruction as a block to get a less conservative
- %% result if the caller is is_not_used/3.
- Block = [{set,[Dst],Ss,{alloc,Live,{bif,Op,Fail}}}],
- check_liveness(R, [{block,Block}|Is], St);
-check_liveness(R, [{select,_,R,_,_}|_], St) ->
- {used,St};
-check_liveness(R, [{select,_,_,Fail,Branches}|_], St) ->
- check_liveness_everywhere(R, [Fail|Branches], St);
-check_liveness(R, [{jump,{f,F}}|_], St) ->
- check_liveness_at(R, F, St);
-check_liveness(R, [{case_end,Used}|_], St) ->
- check_liveness_exit(R, Used, St);
-check_liveness(R, [{try_case_end,Used}|_], St) ->
- check_liveness_exit(R, Used, St);
-check_liveness(R, [{badmatch,Used}|_], St) ->
- check_liveness_exit(R, Used, St);
-check_liveness(R, [if_end|_], St) ->
- check_liveness_exit(R, ignore, St);
-check_liveness(R, [{func_info,_,_,Ar}|_], St) ->
- case R of
- {x,X} when X < Ar -> {used,St};
- _ -> {killed,St}
- end;
-check_liveness(R, [{kill,R}|_], St) ->
- {killed,St};
-check_liveness(R, [{kill,_}|Is], St) ->
- check_liveness(R, Is, St);
-check_liveness(R, [{bs_init,_,_,none,Ss,Dst}|Is], St) ->
- case member(R, Ss) of
- true ->
- {used,St};
- false ->
- if
- R =:= Dst -> {killed,St};
- true -> check_liveness(R, Is, St)
- end
- end;
-check_liveness(R, [{bs_init,_,_,Live,Ss,Dst}|Is], St) ->
- case R of
- {x,X} ->
- case member(R, Ss) of
- true ->
- {used,St};
- false ->
- if
- X < Live ->
- not_used(check_liveness(R, Is, St));
- true ->
- {killed,St}
- end
- end;
- {y,_} ->
- case member(R, Ss) of
- true -> {used,St};
- false ->
- %% If the exception is taken, the stack may
- %% be scanned. Therefore the register is not
- %% guaranteed to be killed.
- if
- R =:= Dst -> {not_used,St};
- true -> not_used(check_liveness(R, Is, St))
- end
- end
- end;
-check_liveness(R, [{deallocate,_}|Is], St) ->
- case R of
- {y,_} -> {killed,St};
- _ -> check_liveness(R, Is, St)
- end;
-check_liveness({x,_}=R, [return|_], St) ->
- case R of
- {x,0} -> {used,St};
- {x,_} -> {killed,St}
- end;
-check_liveness(R, [{call,Live,_}|Is], St) ->
- case R of
- {x,X} when X < Live -> {used,St};
- {x,_} -> {killed,St};
- {y,_} -> not_used(check_liveness(R, Is, St))
- end;
-check_liveness(R, [{call_ext,Live,_}=I|Is], St) ->
- case R of
- {x,X} when X < Live ->
- {used,St};
- {x,_} ->
- {killed,St};
- {y,_} ->
- case beam_jump:is_exit_instruction(I) of
- false ->
- not_used(check_liveness(R, Is, St));
- true ->
- %% We must make sure we don't check beyond this
- %% instruction or we will fall through into random
- %% unrelated code and get stuck in a loop.
- {exit_not_used,St}
- end
- end;
-check_liveness(R, [{call_fun,Live}|Is], St) ->
- case R of
- {x,X} when X =< Live -> {used,St};
- {x,_} -> {killed,St};
- {y,_} -> not_used(check_liveness(R, Is, St))
- end;
-check_liveness(R, [{apply,Args}|Is], St) ->
- case R of
- {x,X} when X < Args+2 -> {used,St};
- {x,_} -> {killed,St};
- {y,_} -> not_used(check_liveness(R, Is, St))
- end;
-check_liveness(R, [{bif,Op,Fail,Ss,D}|Is], St) ->
- Set = {set,[D],Ss,{bif,Op,Fail}},
- check_liveness(R, [{block,[Set]}|Is], St);
-check_liveness(R, [{gc_bif,Op,{f,Fail},Live,Ss,D}|Is], St) ->
- Set = {set,[D],Ss,{alloc,Live,{gc_bif,Op,Fail}}},
- check_liveness(R, [{block,[Set]}|Is], St);
-check_liveness(R, [{bs_put,{f,0},_,Ss}|Is], St) ->
- case member(R, Ss) of
- true -> {used,St};
- false -> check_liveness(R, Is, St)
- end;
-check_liveness(R, [{bs_restore2,S,_}|Is], St) ->
- case R of
- S -> {used,St};
- _ -> check_liveness(R, Is, St)
- end;
-check_liveness(R, [{bs_save2,S,_}|Is], St) ->
- case R of
- S -> {used,St};
- _ -> check_liveness(R, Is, St)
- end;
-check_liveness(R, [{move,S,D}|Is], St) ->
- case R of
- S -> {used,St};
- D -> {killed,St};
- _ -> check_liveness(R, Is, St)
- end;
-check_liveness(R, [{make_fun2,_,_,_,NumFree}|Is], St) ->
- case R of
- {x,X} when X < NumFree -> {used,St};
- {x,_} -> {killed,St};
- {y,_} -> not_used(check_liveness(R, Is, St))
- end;
-check_liveness(R, [{'catch'=Op,Y,Fail}|Is], St) ->
- Set = {set,[Y],[],{try_catch,Op,Fail}},
- check_liveness(R, [{block,[Set]}|Is], St);
-check_liveness(R, [{'try'=Op,Y,Fail}|Is], St) ->
- Set = {set,[Y],[],{try_catch,Op,Fail}},
- check_liveness(R, [{block,[Set]}|Is], St);
-check_liveness(R, [{try_end,Y}|Is], St) ->
- case R of
- Y ->
- {killed,St};
- {y,_} ->
- %% y registers will be used if an exception occurs and
- %% control transfers to the label given in the previous
- %% try/2 instruction.
- {used,St};
- _ ->
- check_liveness(R, Is, St)
- end;
-check_liveness(R, [{catch_end,Y}|Is], St) ->
- case R of
- Y -> {killed,St};
- _ -> check_liveness(R, Is, St)
- end;
-check_liveness(R, [{get_tuple_element,S,_,D}|Is], St) ->
- case R of
- S -> {used,St};
- D -> {killed,St};
- _ -> check_liveness(R, Is, St)
- end;
-check_liveness(R, [{loop_rec,{f,_},{x,0}}|_], St) ->
- case R of
- {x,_} ->
- {killed,St};
- _ ->
- %% y register. Rarely happens. Be very conversative and
- %% assume it's used.
- {used,St}
- end;
-check_liveness(R, [{loop_rec_end,{f,Fail}}|_], St) ->
- check_liveness_at(R, Fail, St);
-check_liveness(R, [{line,_}|Is], St) ->
- check_liveness(R, Is, St);
-check_liveness(R, [{get_map_elements,{f,Fail},S,{list,L}}|Is], St0) ->
- {Ss,Ds} = split_even(L),
- case member(R, [S|Ss]) of
- true ->
- {used,St0};
- false ->
- case check_liveness_at(R, Fail, St0) of
- {killed,St}=Killed ->
- case member(R, Ds) of
- true -> Killed;
- false -> check_liveness(R, Is, St)
- end;
- Other ->
- Other
- end
- end;
-check_liveness(R, [{put_map,F,Op,S,D,Live,{list,Puts}}|Is], St) ->
- Set = {set,[D],[S|Puts],{alloc,Live,{put_map,Op,F}}},
- check_liveness(R, [{block,[Set]}||Is], St);
-check_liveness(R, [{put_tuple,Ar,D}|Is], St) ->
- Set = {set,[D],[],{put_tuple,Ar}},
- check_liveness(R, [{block,[Set]}||Is], St);
-check_liveness(R, [{put_list,S1,S2,D}|Is], St) ->
- Set = {set,[D],[S1,S2],put_list},
- check_liveness(R, [{block,[Set]}||Is], St);
-check_liveness(R, [{test_heap,N,Live}|Is], St) ->
- I = {block,[{set,[],[],{alloc,Live,{nozero,nostack,N,[]}}}]},
- check_liveness(R, [I|Is], St);
-check_liveness(R, [{allocate_zero,N,Live}|Is], St) ->
- I = {block,[{set,[],[],{alloc,Live,{zero,N,0,[]}}}]},
- check_liveness(R, [I|Is], St);
-check_liveness(R, [{get_hd,S,D}|Is], St) ->
- I = {block,[{set,[D],[S],get_hd}]},
- check_liveness(R, [I|Is], St);
-check_liveness(R, [{get_tl,S,D}|Is], St) ->
- I = {block,[{set,[D],[S],get_tl}]},
- check_liveness(R, [I|Is], St);
-check_liveness(R, [remove_message|Is], St) ->
- check_liveness(R, Is, St);
-check_liveness({x,X}, [build_stacktrace|_], St) when X > 0 ->
- {killed,St};
-check_liveness(R, [{recv_mark,_}|Is], St) ->
- check_liveness(R, Is, St);
-check_liveness(R, [{recv_set,_}|Is], St) ->
- check_liveness(R, Is, St);
-check_liveness(R, [{'%',_}|Is], St) ->
- check_liveness(R, Is, St);
-check_liveness(_R, Is, St) when is_list(Is) ->
- %% Not implemented. Conservatively assume that the register is used.
- {used,St}.
-
-check_liveness_everywhere(R, Lbls, St0) ->
- check_liveness_everywhere_1(R, Lbls, killed, St0).
-
-check_liveness_everywhere_1(R, [{f,Lbl}|T], Res0, St0) ->
- {Res1,St} = check_liveness_at(R, Lbl, St0),
- Res = case Res1 of
- killed -> Res0;
- _ -> Res1
- end,
- case Res of
- used -> {used,St};
- _ -> check_liveness_everywhere_1(R, T, Res, St)
- end;
-check_liveness_everywhere_1(R, [_|T], Res, St) ->
- check_liveness_everywhere_1(R, T, Res, St);
-check_liveness_everywhere_1(_, [], Res, St) ->
- {Res,St}.
-
-check_liveness_at(R, Lbl, #live{lbl=Ll,res=ResMemorized}=St0) ->
- case gb_trees:lookup(Lbl, ResMemorized) of
- {value,Res} ->
- {Res,St0};
- none ->
- {Res,St} = case gb_trees:lookup(Lbl, Ll) of
- {value,Is} -> check_liveness(R, Is, St0);
- none -> {used,St0}
- end,
- {Res,St#live{res=gb_trees:insert(Lbl, Res, St#live.res)}}
- end.
-
-not_used({used,_}=Res) -> Res;
-not_used({_,St}) -> {not_used,St}.
-
-check_liveness_exit(R, R, St) -> {used,St};
-check_liveness_exit({x,_}, _, St) -> {killed,St};
-check_liveness_exit({y,_}, _, St) -> {exit_not_used,St}.
-
-%% check_liveness_block(Reg, [Instruction], State) ->
-%% {killed | not_used | used | alloc_used | transparent,State'}
-%% Finds out how Reg is used in the instruction sequence inside a block.
-%% Returns one of:
-%% killed - Reg is assigned a new value or killed by an
-%% allocation instruction
-%% not_used - The value is not used, but the register is referenced
-%% e.g. by an allocation instruction
-%% transparent - Reg is neither used nor killed
-%% alloc_used - Used only in an allocate instruction
-%% used - Reg is explicitly used by an instruction
-%%
-%% Annotations are not allowed.
-%%
-%% (Unknown instructions will cause an exception.)
-
-check_liveness_block({x,X}=R, [{set,Ds,Ss,{alloc,Live,Op}}|Is], St0) ->
- if
- X >= Live ->
- {killed,St0};
- true ->
- case check_liveness_block_1(R, Ss, Ds, Op, Is, St0) of
- {transparent,St} -> {alloc_used,St};
- {_,_}=Res -> not_used(Res)
- end
- end;
-check_liveness_block({y,_}=R, [{set,Ds,Ss,{alloc,_Live,Op}}|Is], St0) ->
- case check_liveness_block_1(R, Ss, Ds, Op, Is, St0) of
- {transparent,St} -> {alloc_used,St};
- {_,_}=Res -> not_used(Res)
- end;
-check_liveness_block({y,_}=R, [{set,Ds,Ss,{try_catch,_,Op}}|Is], St0) ->
- case Ds of
- [R] ->
- {killed,St0};
- _ ->
- case check_liveness_block_1(R, Ss, Ds, Op, Is, St0) of
- {exit_not_used,St} ->
- {used,St};
- {transparent,St} ->
- %% Conservatively assumed that it is used.
- {used,St};
- {_,_}=Res ->
- Res
- end
- end;
-check_liveness_block(R, [{set,Ds,Ss,Op}|Is], St) ->
- check_liveness_block_1(R, Ss, Ds, Op, Is, St);
-check_liveness_block(_, [], St) -> {transparent,St}.
-
-check_liveness_block_1(R, Ss, Ds, Op, Is, St0) ->
- case member(R, Ss) of
- true ->
- {used,St0};
- false ->
- case check_liveness_block_2(R, Op, Ss, St0) of
- {killed,St} ->
- case member(R, Ds) of
- true -> {killed,St};
- false -> check_liveness_block(R, Is, St)
- end;
- {exit_not_used,St} ->
- case member(R, Ds) of
- true -> {exit_not_used,St};
- false -> check_liveness_block(R, Is, St)
- end;
- {not_used,St} ->
- not_used(case member(R, Ds) of
- true -> {killed,St};
- false -> check_liveness_block(R, Is, St)
- end);
- {used,St} ->
- {used,St}
- end
- end.
-
-check_liveness_block_2(R, {gc_bif,Op,{f,Lbl}}, Ss, St) ->
- check_liveness_block_3(R, Lbl, {Op,length(Ss)}, St);
-check_liveness_block_2(R, {bif,Op,{f,Lbl}}, Ss, St) ->
- Arity = length(Ss),
- case erl_internal:comp_op(Op, Arity) orelse
- erl_internal:new_type_test(Op, Arity) of
- true ->
- {killed,St};
- false ->
- check_liveness_block_3(R, Lbl, {Op,length(Ss)}, St)
- end;
-check_liveness_block_2(R, {put_map,_Op,{f,Lbl}}, _Ss, St) ->
- check_liveness_block_3(R, Lbl, {unsafe,0}, St);
-check_liveness_block_2(_, _, _, St) ->
- {killed,St}.
-
-check_liveness_block_3({x,_}, 0, _FA, St) ->
- {killed,St};
-check_liveness_block_3({y,_}, 0, {F,A}, St) ->
- %% If the exception is thrown, the stack may be scanned,
- %% thus implicitly using the y register.
- case erl_bifs:is_safe(erlang, F, A) of
- true -> {killed,St};
- false -> {used,St}
- end;
-check_liveness_block_3(R, Lbl, _FA, St0) ->
- check_liveness_at(R, Lbl, St0).
-
-index_labels_1([{label,Lbl}|Is0], Acc) ->
- Is = drop_labels(Is0),
- index_labels_1(Is0, [{Lbl,Is}|Acc]);
-index_labels_1([_|Is], Acc) ->
- index_labels_1(Is, Acc);
-index_labels_1([], Acc) -> gb_trees:from_orddict(sort(Acc)).
-
-drop_labels([{label,_}|Is]) -> drop_labels(Is);
-drop_labels(Is) -> Is.
-
-
replace_labels_1([{test,Test,{f,Lbl},Ops}|Is], Acc, D, Fb) ->
replace_labels_1(Is, [{test,Test,{f,label(Lbl, D, Fb)},Ops}|Acc], D, Fb);
replace_labels_1([{test,Test,{f,Lbl},Live,Ops,Dst}|Is], Acc, D, Fb) ->
@@ -729,21 +144,6 @@ label(Old, D, Fb) ->
_ -> Fb(Old)
end.
-%% Help function for combine_heap_needs.
-
-combine_alloc_lists(Al0) ->
- Al1 = flatmap(fun(Words) when is_integer(Words) ->
- [{words,Words}];
- ({alloc,List}) ->
- List
- end, Al0),
- Al2 = sofs:relation(Al1),
- Al3 = sofs:relation_to_family(Al2),
- Al4 = sofs:to_external(Al3),
- [{Tag,lists:sum(L)} || {Tag,L} <- Al4].
-
-%% live_opt/4.
-
split_even([], Ss, Ds) ->
{reverse(Ss),reverse(Ds)};
split_even([S,D|Rs], Ss, Ds) ->
diff --git a/lib/compiler/src/beam_validator.erl b/lib/compiler/src/beam_validator.erl
index ca065295d6..7d908df3bf 100644
--- a/lib/compiler/src/beam_validator.erl
+++ b/lib/compiler/src/beam_validator.erl
@@ -479,16 +479,20 @@ valfun_1({try_case,Reg}, #vst{current=#st{ct=[Fail|Fails]}}=Vst0) ->
error({bad_type,Type})
end;
valfun_1({get_list,Src,D1,D2}, Vst0) ->
+ assert_not_literal(Src),
assert_type(cons, Src, Vst0),
Vst = set_type_reg(term, Src, D1, Vst0),
set_type_reg(term, Src, D2, Vst);
valfun_1({get_hd,Src,Dst}, Vst) ->
+ assert_not_literal(Src),
assert_type(cons, Src, Vst),
set_type_reg(term, Src, Dst, Vst);
valfun_1({get_tl,Src,Dst}, Vst) ->
+ assert_not_literal(Src),
assert_type(cons, Src, Vst),
set_type_reg(term, Src, Dst, Vst);
valfun_1({get_tuple_element,Src,I,Dst}, Vst) ->
+ assert_not_literal(Src),
assert_type({tuple_element,I+1}, Src, Vst),
set_type_reg(term, Src, Dst, Vst);
valfun_1({jump,{f,Lbl}}, Vst) ->
@@ -917,6 +921,7 @@ valfun_4(_, _) ->
error(unknown_instruction).
verify_get_map(Fail, Src, List, Vst0) ->
+ assert_not_literal(Src), %OTP 22.
assert_type(map, Src, Vst0),
Vst1 = foldl(fun(D, Vsti) ->
case is_reg_defined(D,Vsti) of
@@ -1466,6 +1471,10 @@ assert_term(Src, Vst) ->
get_term_type(Src, Vst),
ok.
+assert_not_literal({x,_}) -> ok;
+assert_not_literal({y,_}) -> ok;
+assert_not_literal(Literal) -> error({literal_not_allowed,Literal}).
+
%% The possible types.
%%
%% First non-term types:
diff --git a/lib/compiler/src/compile.erl b/lib/compiler/src/compile.erl
index d894694c79..65c4f140c9 100644
--- a/lib/compiler/src/compile.erl
+++ b/lib/compiler/src/compile.erl
@@ -210,8 +210,11 @@ do_compile(Input, Opts0) ->
{error,Reason}
end
end,
- %% Dialyzer has already spawned workers.
- case lists:member(dialyzer, Opts) of
+ %% Some tools, like Dialyzer, has already spawned workers
+ %% and spawning extra workers actually slow the compilation
+ %% down instead of speeding it up, so we provide a mechanism
+ %% to bypass the compiler process.
+ case lists:member(no_spawn_compiler_process, Opts) of
true ->
IntFun();
false ->
@@ -823,6 +826,9 @@ kernel_passes() ->
{unless,no_bsm_opt,{pass,beam_ssa_bsm}},
{iff,dssabsm,{listing,"ssabsm"}},
{iff,ssalint,{pass,beam_ssa_lint}},
+ {unless,no_fun_opt,{pass,beam_ssa_funs}},
+ {iff,dssafuns,{listing,"ssafuns"}},
+ {iff,ssalint,{pass,beam_ssa_lint}},
{unless,no_ssa_opt,{pass,beam_ssa_opt}},
{iff,dssaopt,{listing,"ssaopt"}},
{iff,ssalint,{pass,beam_ssa_lint}},
@@ -862,7 +868,9 @@ asm_passes() ->
%% need to do a few clean-ups to code.
{iff,no_postopt,[{pass,beam_clean}]},
+ {iff,diffable,?pass(diffable)},
{pass,beam_z},
+ {iff,diffable,{listing,"S"}},
{iff,dz,{listing,"z"}},
{iff,dopt,{listing,"optimize"}},
{iff,'S',{listing,"S"}},
@@ -1006,11 +1014,17 @@ parse_module(_Code, St0) ->
end.
do_parse_module(DefEncoding, #compile{ifile=File,options=Opts,dir=Dir}=St) ->
+ SourceName0 = proplists:get_value(source, Opts, File),
+ SourceName = case member(deterministic, Opts) of
+ true -> filename:basename(SourceName0);
+ false -> SourceName0
+ end,
R = epp:parse_file(File,
- [{includes,[".",Dir|inc_paths(Opts)]},
- {macros,pre_defs(Opts)},
- {default_encoding,DefEncoding},
- extra]),
+ [{includes,[".",Dir|inc_paths(Opts)]},
+ {source_name, SourceName},
+ {macros,pre_defs(Opts)},
+ {default_encoding,DefEncoding},
+ extra]),
case R of
{ok,Forms,Extra} ->
Encoding = proplists:get_value(encoding, Extra),
@@ -1914,6 +1928,39 @@ restore_expand_module([F|Fs]) ->
[F|restore_expand_module(Fs)];
restore_expand_module([]) -> [].
+%%%
+%%% Transform the BEAM code to make it more friendly for
+%%% diffing: using function names instead of labels for
+%%% local calls and number labels relative to each function.
+%%%
+
+diffable(Code0, St) ->
+ {Mod,Exp,Attr,Fs0,NumLabels} = Code0,
+ EntryLabels0 = [{Entry,{Name,Arity}} ||
+ {function,Name,Arity,Entry,_} <- Fs0],
+ EntryLabels = maps:from_list(EntryLabels0),
+ Fs = [diffable_fix_function(F, EntryLabels) || F <- Fs0],
+ Code = {Mod,Exp,Attr,Fs,NumLabels},
+ {ok,Code,St}.
+
+diffable_fix_function({function,Name,Arity,Entry0,Is0}, LabelMap0) ->
+ Entry = maps:get(Entry0, LabelMap0),
+ {Is1,LabelMap} = diffable_label_map(Is0, 1, LabelMap0, []),
+ Fb = fun(Old) -> error({no_fb,Old}) end,
+ Is = beam_utils:replace_labels(Is1, [], LabelMap, Fb),
+ {function,Name,Arity,Entry,Is}.
+
+diffable_label_map([{label,Old}|Is], New, Map, Acc) ->
+ case Map of
+ #{Old:=NewLabel} ->
+ diffable_label_map(Is, New, Map, [{label,NewLabel}|Acc]);
+ #{} ->
+ diffable_label_map(Is, New+1, Map#{Old=>New}, [{label,New}|Acc])
+ end;
+diffable_label_map([I|Is], New, Map, Acc) ->
+ diffable_label_map(Is, New, Map, [I|Acc]);
+diffable_label_map([], _New, Map, Acc) ->
+ {Acc,Map}.
-spec options() -> 'ok'.
@@ -2047,6 +2094,7 @@ pre_load() ->
beam_ssa_bsm,
beam_ssa_codegen,
beam_ssa_dead,
+ beam_ssa_funs,
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 7b802fdd62..86259270bd 100644
--- a/lib/compiler/src/compiler.app.src
+++ b/lib/compiler/src/compiler.app.src
@@ -39,6 +39,7 @@
beam_ssa_bsm,
beam_ssa_codegen,
beam_ssa_dead,
+ beam_ssa_funs,
beam_ssa_lint,
beam_ssa_opt,
beam_ssa_pp,
diff --git a/lib/compiler/src/erl_bifs.erl b/lib/compiler/src/erl_bifs.erl
index 71ab0e872a..ce9762899e 100644
--- a/lib/compiler/src/erl_bifs.erl
+++ b/lib/compiler/src/erl_bifs.erl
@@ -108,6 +108,7 @@ is_pure(erlang, list_to_atom, 1) -> true;
is_pure(erlang, list_to_binary, 1) -> true;
is_pure(erlang, list_to_float, 1) -> true;
is_pure(erlang, list_to_integer, 1) -> true;
+is_pure(erlang, list_to_integer, 2) -> true;
is_pure(erlang, list_to_pid, 1) -> true;
is_pure(erlang, list_to_tuple, 1) -> true;
is_pure(erlang, max, 2) -> true;
generated by cgit v1.2.3 (git 2.25.1) at 2025-05-07 04:43:25 +0000