diff options
author | Stavros Aronis <[email protected]> | 2016-06-08 13:13:04 +0200 |
---|---|---|
committer | Stavros Aronis <[email protected]> | 2016-06-08 13:13:04 +0200 |
commit | b13412d2e4331fa426a6722a04310cbcbf1731d6 (patch) | |
tree | 17fba1d7099d652eeab812b4b1a72e44cc9277bd /lib | |
parent | d26c15e07229c90ba8353bd78d5406ada0f13271 (diff) | |
download | otp-b13412d2e4331fa426a6722a04310cbcbf1731d6.tar.gz otp-b13412d2e4331fa426a6722a04310cbcbf1731d6.tar.bz2 otp-b13412d2e4331fa426a6722a04310cbcbf1731d6.zip |
Fix a bug in Dialyzer related to call-site analysis
Dialyzer's "dataflow" module is using information from the callgraph to
determine which functions may be called at a particular call-site. Unfortunately
this information can include functions that are certainly not among the possible
choices. We don't want to emit warnings in such cases, so a "reasonable"
compromise is to stay silent if there are many possible funs and at least one of
them can succeed.
Bug reported by Dan Gudmundsson, test shrunk down by Magnus Lång.
Diffstat (limited to 'lib')
5 files changed, 82 insertions, 11 deletions
diff --git a/lib/dialyzer/src/dialyzer_dataflow.erl b/lib/dialyzer/src/dialyzer_dataflow.erl index 5ab0c39c04..3349b12932 100644 --- a/lib/dialyzer/src/dialyzer_dataflow.erl +++ b/lib/dialyzer/src/dialyzer_dataflow.erl @@ -430,17 +430,35 @@ handle_apply(Tree, Map, State) -> handle_apply_or_call(FunInfoList, Args, ArgTypes, Map, Tree, State) -> None = t_none(), + %% Call-site analysis may be inaccurate and consider more funs than those that + %% are actually possible. If all of them are incorrect, then warnings can be + %% emitted. If at least one fun is ok, however, then no warning is emitted, + %% just in case the bad ones are not really possible. The last argument is + %% used for this, with the following encoding: + %% Initial value: {none, []} + %% First fun checked: {one, <List of warns>} + %% More funs checked: {many, <List of warns>} + %% A '{one, []}' can only become '{many, []}'. + %% If at any point an fun does not add warnings, then the list is also + %% replaced with an empty list. handle_apply_or_call(FunInfoList, Args, ArgTypes, Map, Tree, State, - [None || _ <- ArgTypes], None, false). + [None || _ <- ArgTypes], None, false, {none, []}). handle_apply_or_call([{local, external}|Left], Args, ArgTypes, Map, Tree, State, - _AccArgTypes, _AccRet, _HadExternal) -> + _AccArgTypes, _AccRet, _HadExternal, Warns) -> + {HowMany, _} = Warns, + NewHowMany = + case HowMany of + none -> one; + _ -> many + end, + NewWarns = {NewHowMany, []}, handle_apply_or_call(Left, Args, ArgTypes, Map, Tree, State, - ArgTypes, t_any(), true); + ArgTypes, t_any(), true, NewWarns); handle_apply_or_call([{TypeOfApply, {Fun, Sig, Contr, LocalRet}}|Left], Args, ArgTypes, Map, Tree, #state{opaques = Opaques} = State, - AccArgTypes, AccRet, HadExternal) -> + AccArgTypes, AccRet, HadExternal, Warns) -> Any = t_any(), AnyArgs = [Any || _ <- Args], GenSig = {AnyArgs, fun(_) -> t_any() end}, @@ -586,16 +604,32 @@ handle_apply_or_call([{TypeOfApply, {Fun, Sig, Contr, LocalRet}}|Left], end, NewAccRet = t_sup(AccRet, TotalRet), ?debug("NewAccRet: ~s\n", [t_to_string(NewAccRet)]), + {NewWarnings, State4} = state__remove_added_warnings(State, State3), + {HowMany, OldWarnings} = Warns, + NewWarns = + case HowMany of + none -> {one, NewWarnings}; + _ -> + case OldWarnings =:= [] of + true -> {many, []}; + false -> + case NewWarnings =:= [] of + true -> {many, []}; + false -> {many, NewWarnings ++ OldWarnings} + end + end + end, handle_apply_or_call(Left, Args, ArgTypes, Map, Tree, - State3, NewAccArgTypes, NewAccRet, HadExternal); + State4, NewAccArgTypes, NewAccRet, HadExternal, NewWarns); handle_apply_or_call([], Args, _ArgTypes, Map, _Tree, State, - AccArgTypes, AccRet, HadExternal) -> + AccArgTypes, AccRet, HadExternal, {_, Warnings}) -> + State1 = state__add_warnings(Warnings, State), case HadExternal of false -> NewMap = enter_type_lists(Args, AccArgTypes, Map), - {State, NewMap, AccRet}; + {State1, NewMap, AccRet}; true -> - {had_external, State} + {had_external, State1} end. apply_fail_reason(FailedSig, FailedBif, FailedContract) -> @@ -3038,6 +3072,14 @@ state__add_warning(#state{warnings = Warnings, warning_mode = true} = State, end end. +state__remove_added_warnings(OldState, NewState) -> + #state{warnings = OldWarnings} = OldState, + #state{warnings = NewWarnings} = NewState, + {NewWarnings -- OldWarnings, NewState#state{warnings = OldWarnings}}. + +state__add_warnings(Warns, #state{warnings = Warnings} = State) -> + State#state{warnings = Warns ++ Warnings}. + -spec state__set_curr_fun(curr_fun(), state()) -> state(). state__set_curr_fun(undefined, State) -> diff --git a/lib/dialyzer/test/small_SUITE_data/results/higher_order_discrepancy b/lib/dialyzer/test/small_SUITE_data/results/higher_order_discrepancy index 7ce440a60d..11b9ecade6 100644 --- a/lib/dialyzer/test/small_SUITE_data/results/higher_order_discrepancy +++ b/lib/dialyzer/test/small_SUITE_data/results/higher_order_discrepancy @@ -1,4 +1,3 @@ -higher_order_discrepancy.erl:11: The call higher_order_discrepancy:g('foo') will never return since it differs in the 1st argument from the success typing arguments: ('bar') -higher_order_discrepancy.erl:14: Function g/1 has no local return -higher_order_discrepancy.erl:14: The pattern 'bar' can never match the type 'foo' +higher_order_discrepancy.erl:19: Function g/1 has no local return +higher_order_discrepancy.erl:19: The pattern 'bar' can never match the type 'foo' diff --git a/lib/dialyzer/test/small_SUITE_data/src/higher_order_discrepancy.erl b/lib/dialyzer/test/small_SUITE_data/src/higher_order_discrepancy.erl index ff5ee6bac4..f9547d4929 100644 --- a/lib/dialyzer/test/small_SUITE_data/src/higher_order_discrepancy.erl +++ b/lib/dialyzer/test/small_SUITE_data/src/higher_order_discrepancy.erl @@ -1,3 +1,8 @@ +%% With the patch introduced to avoid false warnings in +%% user_SUITE_data/src/wpc_hlines.erl we can unfortunately no longer precisely +%% catch problems like this one... The refinement procedure is still enough to +%% keep some of the details, nevertheless. + -module(higher_order_discrepancy). -export([test/1]). diff --git a/lib/dialyzer/test/user_SUITE_data/results/wpc_hlines b/lib/dialyzer/test/user_SUITE_data/results/wpc_hlines new file mode 100644 index 0000000000..d6e3f29ab9 --- /dev/null +++ b/lib/dialyzer/test/user_SUITE_data/results/wpc_hlines @@ -0,0 +1,3 @@ + +wpc_hlines.erl:22: Function bad/1 has no local return +wpc_hlines.erl:22: The pattern 'false' can never match the type 'true' diff --git a/lib/dialyzer/test/user_SUITE_data/src/wpc_hlines.erl b/lib/dialyzer/test/user_SUITE_data/src/wpc_hlines.erl new file mode 100644 index 0000000000..8c205a8247 --- /dev/null +++ b/lib/dialyzer/test/user_SUITE_data/src/wpc_hlines.erl @@ -0,0 +1,22 @@ +%% Bug reported by Dan Gudmundsson, test shrunk down by Magnus Lång. + +%% The problem is that dialyzer_dep generates edges from the fun +%% application to both of the functions, and then during the warning pass +%% dialyzer_dataflow:handle_apply_or_call generates warnings for any such +%% edge that won't return. + +%% Since dialyzer_dep is currently supposed to overapproximate rather than +%% underapproximate, the fix was to modify handle_apply_or_call to not generate +%% warnings if some of the possible funs can succeed. + +-module(wpc_hlines). + +-export([do_export/0]). + +do_export() -> + {Proj, _} = % The culprit seems to be putting the funs in a tuple + {fun good/1, fun bad/1}, + Proj(true). + +good(_) -> ok. +bad(false) -> ok. |