aboutsummaryrefslogtreecommitdiffstats
path: root/lib/compiler/src
diff options
context:
space:
mode:
Diffstat (limited to 'lib/compiler/src')
-rw-r--r--lib/compiler/src/beam_bsm.erl13
-rw-r--r--lib/compiler/src/beam_jump.erl15
-rw-r--r--lib/compiler/src/beam_utils.erl9
-rw-r--r--lib/compiler/src/compile.erl14
-rw-r--r--lib/compiler/src/sys_core_bsm.erl62
5 files changed, 91 insertions, 22 deletions
diff --git a/lib/compiler/src/beam_bsm.erl b/lib/compiler/src/beam_bsm.erl
index abc6e96c85..1c8e0e9854 100644
--- a/lib/compiler/src/beam_bsm.erl
+++ b/lib/compiler/src/beam_bsm.erl
@@ -310,18 +310,7 @@ btb_reaches_match_2([{test,bs_start_match2,{f,F},Live,[Bin,_],Ctx}|Is],
end;
btb_reaches_match_2([{test,_,{f,F},Ss}=I|Is], Regs, D0) ->
btb_ensure_not_used(Ss, I, Regs),
- D1 = btb_follow_branch(F, Regs, D0),
- D = case Is of
- [{bs_context_to_binary,_}|_] ->
- %% bs_context_to_binary following a test instruction
- %% probably needs the current position to be saved as
- %% the new start position, but we can't be sure.
- %% Therefore, conservatively disable the optimization
- %% (instead of forcing a saving of the position).
- D1#btb{must_save=true,must_not_save=true};
- _ ->
- D1
- end,
+ D = btb_follow_branch(F, Regs, D0),
btb_reaches_match_1(Is, Regs, D);
btb_reaches_match_2([{test,_,{f,F},_,Ss,_}=I|Is], Regs, D0) ->
btb_ensure_not_used(Ss, I, Regs),
diff --git a/lib/compiler/src/beam_jump.erl b/lib/compiler/src/beam_jump.erl
index 9eee56d604..22974da398 100644
--- a/lib/compiler/src/beam_jump.erl
+++ b/lib/compiler/src/beam_jump.erl
@@ -128,7 +128,7 @@
%%% on the program state.
%%%
--import(lists, [reverse/1,reverse/2,foldl/3]).
+-import(lists, [dropwhile/2,reverse/1,reverse/2,foldl/3]).
-type instruction() :: beam_utils:instruction().
@@ -411,14 +411,19 @@ opt_useless_loads([{test,_,{f,L},_}=I|Is], L, St) ->
opt_useless_loads(Is, _L, St) ->
{Is,St}.
-opt_useless_block_loads([{set,[Dst],_,_}=I|Is], L, Index) ->
- BlockJump = [{block,Is},{jump,{f,L}}],
+opt_useless_block_loads([{set,[Dst],_,_}=I|Is0], L, Index) ->
+ BlockJump = [{block,Is0},{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
+ %% The register is killed and not used, we can remove the load.
+ %% Remove any `put` instructions in case we just
+ %% removed a `put_tuple` instruction.
+ Is = dropwhile(fun({set,_,_,put}) -> true;
+ (_) -> false
+ end, Is0),
opt_useless_block_loads(Is, L, Index);
false ->
- [I|opt_useless_block_loads(Is, L, Index)]
+ [I|opt_useless_block_loads(Is0, L, Index)]
end;
opt_useless_block_loads([I|Is], L, Index) ->
[I|opt_useless_block_loads(Is, L, Index)];
diff --git a/lib/compiler/src/beam_utils.erl b/lib/compiler/src/beam_utils.erl
index 5580d2f123..6b2ab5a2a4 100644
--- a/lib/compiler/src/beam_utils.erl
+++ b/lib/compiler/src/beam_utils.erl
@@ -745,8 +745,11 @@ 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),
+
+ %% Note that is_function/2 is a type test but is not safe.
case erl_internal:comp_op(Op, Arity) orelse
- erl_internal:new_type_test(Op, Arity) of
+ (erl_internal:new_type_test(Op, Arity) andalso
+ erl_bifs:is_safe(erlang, Op, Arity)) of
true ->
{killed,St};
false ->
@@ -1115,6 +1118,10 @@ defs([{bs_init,{f,L},_,Live,_,Dst}=I|Is], Regs0, D) ->
end,
Regs = def_regs([Dst], Regs1),
[I|defs(Is, Regs, update_regs(L, Regs, D))];
+defs([{test,bs_start_match2,{f,L},Live,_,Dst}=I|Is], _Regs, D) ->
+ Regs0 = init_def_regs(Live),
+ Regs = def_regs([Dst], Regs0),
+ [I|defs(Is, Regs, update_regs(L, Regs0, D))];
defs([{bs_put,{f,L},_,_}=I|Is], Regs, D) ->
[I|defs(Is, Regs, update_regs(L, Regs, D))];
defs([build_stacktrace=I|Is], _Regs, D) ->
diff --git a/lib/compiler/src/compile.erl b/lib/compiler/src/compile.erl
index 562d57a6ef..6510571441 100644
--- a/lib/compiler/src/compile.erl
+++ b/lib/compiler/src/compile.erl
@@ -931,11 +931,17 @@ 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),
diff --git a/lib/compiler/src/sys_core_bsm.erl b/lib/compiler/src/sys_core_bsm.erl
index d7b26c3a56..62657933ee 100644
--- a/lib/compiler/src/sys_core_bsm.erl
+++ b/lib/compiler/src/sys_core_bsm.erl
@@ -44,6 +44,14 @@ function([{#c_var{name={F,Arity}}=Name,B0}|Fs], FsAcc, Ws0) ->
{B,Ws} ->
function(Fs, [{Name,B}|FsAcc], Ws)
catch
+ throw:unsafe_bs_context_to_binary ->
+ %% Unsafe bs_context_to_binary (in the sense that the
+ %% contents of the binary will probably be wrong).
+ %% Disable binary optimizations for the entire function.
+ %% We don't generate an INFO message, because this happens
+ %% very infrequently and it would be hard to explain in
+ %% a comprehensible way in an INFO message.
+ function(Fs, [{Name,B0}|FsAcc], Ws0);
Class:Error:Stack ->
io:fwrite("Function: ~w/~w\n", [F,Arity]),
erlang:raise(Class, Error, Stack)
@@ -116,12 +124,66 @@ move_from_col(Pos, L) ->
[Col|First] ++ Rest.
bsm_do_an([#c_var{name=Vname}=V0|Vs0], Cs0, Case) ->
+ bsm_inner_context_to_binary(Cs0),
Cs = bsm_do_an_var(Vname, Cs0),
V = bsm_annotate_for_reuse(V0),
Vs = core_lib:make_values([V|Vs0]),
Case#c_case{arg=Vs,clauses=Cs};
bsm_do_an(_Vs, _Cs, Case) -> Case.
+bsm_inner_context_to_binary([#c_clause{body=B}|Cs]) ->
+ %% Consider:
+ %%
+ %% foo(<<Length, Data/binary>>) -> %Line 1
+ %% case {Data, Length} of %Line 2
+ %% {_, 0} -> Data; %Line 3
+ %% {<<...>>, 4} -> ... %Line 4
+ %% end.
+ %%
+ %% No sub binary will be created for Data in line 1. The match
+ %% context will be passed on to the `case` in line 2. In line 3,
+ %% this pass inserts a `bs_context_to_binary` instruction to
+ %% convert the match context representing Data to a binary before
+ %% returning it. The problem is that the binary created will be
+ %% the original binary (including the matched out Length field),
+ %% not the tail of the binary as it is supposed to be.
+ %%
+ %% Here follows a heuristic to disable the binary optimizations
+ %% for the entire function if this code kind of code is found.
+
+ case cerl_trees:free_variables(B) of
+ [] ->
+ %% Since there are no free variables in the body of
+ %% this clause, there can't be any troublesome
+ %% bs_context_to_binary instructions.
+ bsm_inner_context_to_binary(Cs);
+ [_|_]=Free ->
+ %% One of the free variables could refer to a match context
+ %% created by the outer binary match.
+ F = fun(#c_primop{name=#c_literal{val=bs_context_to_binary},
+ args=[#c_var{name=V}]}, _) ->
+ case member(V, Free) of
+ true ->
+ %% This bs_context_to_binary instruction will
+ %% make a binary of the match context from an
+ %% outer binary match. It is very likely that
+ %% the contents of the binary will be wrong
+ %% (the original binary as opposed to only
+ %% the tail binary).
+ throw(unsafe_bs_context_to_binary);
+ false ->
+ %% Safe. This bs_context_to_binary instruction
+ %% will make a binary from a match context
+ %% defined in the body of the clause.
+ ok
+ end;
+ (_, _) ->
+ ok
+ end,
+ cerl_trees:fold(F, ok, B)
+ end;
+bsm_inner_context_to_binary([]) -> ok.
+
bsm_do_an_var(V, [#c_clause{pats=[P|_],guard=G,body=B0}=C0|Cs]) ->
case P of
#c_var{name=VarName} ->
generated by cgit v1.2.3 (git 2.25.1) at 2025-05-05 06:59:22 +0000