diff options
Diffstat (limited to 'lib/compiler')
23 files changed, 614 insertions, 151 deletions
diff --git a/lib/compiler/doc/src/notes.xml b/lib/compiler/doc/src/notes.xml index f0d869381b..f11444137d 100644 --- a/lib/compiler/doc/src/notes.xml +++ b/lib/compiler/doc/src/notes.xml @@ -32,6 +32,50 @@ <p>This document describes the changes made to the Compiler application.</p> +<section><title>Compiler 7.4.4</title> + + <section><title>Fixed Bugs and Malfunctions</title> + <list> + <item> + <p>Fixed a compiler crash introduced in <c>22.0.6</c> + (OTP-15952).</p> + <p> + Own Id: OTP-15953 Aux Id: ERL-999 </p> + </item> + </list> + </section> + +</section> + +<section><title>Compiler 7.4.3</title> + + <section><title>Fixed Bugs and Malfunctions</title> + <list> + <item> + <p>Fixed an unsafe optimization when matching + <c>tuple_size/1</c> outside of guards, which could crash + the emulator if the argument was not a tuple.</p> + <p> + Own Id: OTP-15945</p> + </item> + <item> + <p>Fixed a rare bug that could cause the wrong kind of + exception to be thrown when a BIF failed in a function + that matched bitstrings.</p> + <p> + Own Id: OTP-15946</p> + </item> + <item> + <p>Fixed a bug where receive statements inside try/catch + blocks could return incorrect results.</p> + <p> + Own Id: OTP-15952</p> + </item> + </list> + </section> + +</section> + <section><title>Compiler 7.4.2</title> <section><title>Fixed Bugs and Malfunctions</title> diff --git a/lib/compiler/src/beam_except.erl b/lib/compiler/src/beam_except.erl index 2305502800..2b9c1b0cf5 100644 --- a/lib/compiler/src/beam_except.erl +++ b/lib/compiler/src/beam_except.erl @@ -152,7 +152,7 @@ dig_out_fc(Arity, Is0) -> ({test,_,_,_}) -> false; (_) -> true end, Is0), - {Regs,Acc} = dig_out_fc_1(reverse(Is), Regs0, Acc0), + {Regs,Acc} = dig_out_fc_1(reverse(Is), Arity, Regs0, Acc0), case Regs of #{{x,0}:={atom,function_clause},{x,1}:=Args} -> case moves_from_stack(Args, 0, []) of @@ -165,19 +165,27 @@ dig_out_fc(Arity, Is0) -> no end. -dig_out_fc_1([{block,Bl}|Is], Regs0, Acc) -> +dig_out_fc_1([{block,Bl}|Is], Arity, 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,Src,Dst,Live0}|Is], Regs0, Acc) -> - Regs = prune_xregs(Live0, Regs0), - Live = dig_out_stack_live(Regs, Live0), - I = {bs_get_tail,Src,Dst,Live}, - dig_out_fc_1(Is, Regs, [I|Acc]); -dig_out_fc_1([_|_], _Regs, _Acc) -> + dig_out_fc_1(Is, Arity, Regs, Acc); +dig_out_fc_1([{bs_set_position,_,_}=I|Is], Arity, Regs, Acc) -> + dig_out_fc_1(Is, Arity, Regs, [I|Acc]); +dig_out_fc_1([{bs_get_tail,Src,Dst,Live0}|Is], Arity, Regs0, Acc) -> + case Src of + {x,X} when X < Arity -> + %% The heuristic for determining the number of live + %% registers is likely to give an incorrect result. + %% Give up. + {#{},[]}; + _ -> + Regs = prune_xregs(Live0, Regs0), + Live = dig_out_stack_live(Regs, Live0), + I = {bs_get_tail,Src,Dst,Live}, + dig_out_fc_1(Is, Arity, Regs, [I|Acc]) + end; +dig_out_fc_1([_|_], _Arity, _Regs, _Acc) -> {#{},[]}; -dig_out_fc_1([], Regs, Acc) -> +dig_out_fc_1([], _Arity, Regs, Acc) -> {Regs,Acc}. dig_out_fc_block([{set,[],[],{alloc,Live,_}}|Is], Regs0) -> diff --git a/lib/compiler/src/beam_ssa.erl b/lib/compiler/src/beam_ssa.erl index a9977b0b1d..6492d1e1bf 100644 --- a/lib/compiler/src/beam_ssa.erl +++ b/lib/compiler/src/beam_ssa.erl @@ -96,7 +96,8 @@ %% To avoid the collapsing, change the value of SET_LIMIT to 50 in the %% file erl_types.erl in the hipe application. --type prim_op() :: 'bs_add' | 'bs_extract' | 'bs_init' | 'bs_init_writable' | +-type prim_op() :: 'bs_add' | 'bs_extract' | 'bs_get_tail' | + 'bs_init' | 'bs_init_writable' | 'bs_match' | 'bs_put' | 'bs_start_match' | 'bs_test_tail' | 'bs_utf16_size' | 'bs_utf8_size' | 'build_stacktrace' | 'call' | 'catch_end' | @@ -117,11 +118,12 @@ '+' | '-' | '*' | '/'. %% Primops only used internally during code generation. --type cg_prim_op() :: 'bs_get' | 'bs_match_string' | 'bs_restore' | 'bs_skip' | +-type cg_prim_op() :: 'bs_get' | 'bs_get_position' | 'bs_match_string' | + 'bs_restore' | 'bs_save' | 'bs_set_position' | 'bs_skip' | 'copy' | 'put_tuple_arity' | 'put_tuple_element' | - 'set_tuple_element'. + 'put_tuple_elements' | 'set_tuple_element'. --import(lists, [foldl/3,keyfind/3,mapfoldl/3,member/2,reverse/1]). +-import(lists, [foldl/3,keyfind/3,mapfoldl/3,member/2,reverse/1,umerge/1]). -spec add_anno(Key, Value, Construct) -> Construct when Key :: atom(), @@ -649,12 +651,18 @@ is_commutative('=/=') -> true; is_commutative('/=') -> true; is_commutative(_) -> false. -def_used_1([#b_blk{is=Is,last=Last}|Bs], Preds, Def0, Used0) -> - {Def,Used1} = def_used_is(Is, Preds, Def0, Used0), - Used = ordsets:union(used(Last), Used1), - def_used_1(Bs, Preds, Def, Used); -def_used_1([], _Preds, Def, Used) -> - {ordsets:from_list(Def),Used}. +def_used_1([#b_blk{is=Is,last=Last}|Bs], Preds, Def0, UsedAcc) -> + {Def,Used} = def_used_is(Is, Preds, Def0, used(Last)), + case Used of + [] -> + def_used_1(Bs, Preds, Def, UsedAcc); + [_|_] -> + def_used_1(Bs, Preds, Def, [Used|UsedAcc]) + end; +def_used_1([], _Preds, Def0, UsedAcc) -> + Def = ordsets:from_list(Def0), + Used = umerge(UsedAcc), + {Def,Used}. def_used_is([#b_set{op=phi,dst=Dst,args=Args}|Is], Preds, Def0, Used0) -> diff --git a/lib/compiler/src/beam_ssa_bsm.erl b/lib/compiler/src/beam_ssa_bsm.erl index 382e6f635e..1ac9e6a3bb 100644 --- a/lib/compiler/src/beam_ssa_bsm.erl +++ b/lib/compiler/src/beam_ssa_bsm.erl @@ -683,8 +683,12 @@ aca_copy_successors(Lbl0, Blocks0, Counter0) -> Lbl = maps:get(Lbl0, BRs), {Lbl, Blocks, Counter}. +aca_cs_build_brs([?BADARG_BLOCK=Lbl | Path], Counter, Acc) -> + %% ?BADARG_BLOCK is a marker and not an actual block, so renaming it will + %% break exception handling. + aca_cs_build_brs(Path, Counter, Acc#{ Lbl => Lbl }); aca_cs_build_brs([Lbl | Path], Counter0, Acc) -> - aca_cs_build_brs(Path, Counter0 + 1, maps:put(Lbl, Counter0, Acc)); + aca_cs_build_brs(Path, Counter0 + 1, Acc#{ Lbl => Counter0 }); aca_cs_build_brs([], Counter, Acc) -> {Acc, Counter}. diff --git a/lib/compiler/src/beam_ssa_codegen.erl b/lib/compiler/src/beam_ssa_codegen.erl index 07f4c8b461..08641e2abc 100644 --- a/lib/compiler/src/beam_ssa_codegen.erl +++ b/lib/compiler/src/beam_ssa_codegen.erl @@ -764,9 +764,8 @@ defined(Linear, #cg{regs=Regs}) -> def([{L,#cg_blk{is=Is0,last=Last}=Blk0}|Bs], DefMap0, Regs) -> Def0 = def_get(L, DefMap0), - {Is,Def} = def_is(Is0, Regs, Def0, []), - Successors = successors(Last), - DefMap = def_successors(Successors, Def, DefMap0), + {Is,Def,MaybeDef} = def_is(Is0, Regs, Def0, []), + DefMap = def_successors(Last, Def, MaybeDef, DefMap0), Blk = Blk0#cg_blk{is=Is}, [{L,Blk}|def(Bs, DefMap, Regs)]; def([], _, _) -> []. @@ -780,6 +779,11 @@ def_get(L, DefMap) -> def_is([#cg_alloc{anno=Anno0}=I0|Is], Regs, Def, Acc) -> I = I0#cg_alloc{anno=Anno0#{def_yregs=>Def}}, def_is(Is, Regs, Def, [I|Acc]); +def_is([#cg_set{op=succeeded,args=[Var]}=I], Regs, Def, Acc) -> + %% Var will only be defined on the success branch of the `br` + %% for this block. + MaybeDef = def_add_yreg(Var, [], Regs), + {reverse(Acc, [I]),Def,MaybeDef}; def_is([#cg_set{op=kill_try_tag,args=[#b_var{}=Tag]}=I|Is], Regs, Def0, Acc) -> Def = ordsets:del_element(Tag, Def0), def_is(Is, Regs, Def, [I|Acc]); @@ -822,7 +826,7 @@ def_is([#cg_set{anno=Anno0,dst=Dst}=I0|Is], Regs, Def0, Acc) -> Def = def_add_yreg(Dst, Def0, Regs), def_is(Is, Regs, Def, [I|Acc]); def_is([], _, Def, Acc) -> - {reverse(Acc),Def}. + {reverse(Acc),Def,[]}. def_add_yreg(Dst, Def, Regs) -> case is_yreg(Dst, Regs) of @@ -830,6 +834,12 @@ def_add_yreg(Dst, Def, Regs) -> false -> Def end. +def_successors(#cg_br{bool=#b_var{},succ=Succ,fail=Fail}, Def, MaybeDef, DefMap0) -> + DefMap = def_successors([Fail], ordsets:subtract(Def, MaybeDef), DefMap0), + def_successors([Succ], Def, DefMap); +def_successors(Last, Def, [], DefMap) -> + def_successors(successors(Last), Def, DefMap). + def_successors([S|Ss], Def0, DefMap) -> case DefMap of #{S:=Def1} -> diff --git a/lib/compiler/src/beam_ssa_dead.erl b/lib/compiler/src/beam_ssa_dead.erl index 64b9b3e222..e78e4647a8 100644 --- a/lib/compiler/src/beam_ssa_dead.erl +++ b/lib/compiler/src/beam_ssa_dead.erl @@ -30,7 +30,7 @@ -import(lists, [append/1,keymember/3,last/1,member/2, takewhile/2,reverse/1]). --type used_vars() :: #{beam_ssa:label():=ordsets:ordset(beam_ssa:var_name())}. +-type used_vars() :: #{beam_ssa:label():=cerl_sets:set(beam_ssa:var_name())}. -type basic_type_test() :: atom() | {'is_tagged_tuple',pos_integer(),atom()}. -type type_test() :: basic_type_test() | {'not',basic_type_test()}. @@ -90,13 +90,11 @@ shortcut_opt(#st{bs=Blocks}=St) -> %% the diff.) %% %% Unfortunately, processing the blocks in reverse post order - %% potentially makes the time complexity quadratic or even cubic if - %% the ordset of unset variables grows large, instead of - %% linear for post order processing. We try to still get reasonable - %% compilation times by optimizations that will keep the constant - %% factor as low as possible, and we try to avoid the cubic time - %% complexity by trying to keep the set of unset variables as small - %% as possible. + %% potentially makes the time complexity quadratic, instead of + %% linear for post order processing. We avoid drastic slowdowns by + %% limiting how far we search forward to a common block that + %% both the success and failure label will reach (see the comment + %% in the first clause of shortcut_2/5). Ls = beam_ssa:rpo(Blocks), shortcut_opt(Ls, #{}, St). @@ -124,10 +122,15 @@ shortcut_terminator(#b_br{bool=#b_var{}=Bool,succ=Succ0,fail=Fail0}=Br, Is, From, Bs, St0) -> St = St0#st{target=one_way}, RelOp = get_rel_op(Bool, Is), - SuccBs = bind_var(Bool, #b_literal{val=true}, Bs), + + %% The boolean in a `br` is seldom used by the successors. By + %% not binding its value unless it is actually used we might be able + %% to skip some work in shortcut/4 and sub/2. + SuccBs = bind_var_if_used(Succ0, Bool, #b_literal{val=true}, Bs, St), BrSucc = shortcut(Succ0, From, SuccBs, St#st{rel_op=RelOp}), - FailBs = bind_var(Bool, #b_literal{val=false}, Bs), + FailBs = bind_var_if_used(Fail0, Bool, #b_literal{val=false}, Bs, St), BrFail = shortcut(Fail0, From, FailBs, St#st{rel_op=invert_op(RelOp)}), + case {BrSucc,BrFail} of {#b_br{bool=#b_literal{val=true},succ=Succ}, #b_br{bool=#b_literal{val=true},succ=Fail}} @@ -152,8 +155,14 @@ shortcut_switch([{Lit,L0}|T], Bool, From, Bs, St0) -> [{Lit,L}|shortcut_switch(T, Bool, From, Bs, St0)]; shortcut_switch([], _, _, _, _) -> []. +shortcut(L, _From, Bs, #st{rel_op=none,target=one_way}) when map_size(Bs) =:= 0 -> + %% There is no way that we can find a suitable branch, because there is no + %% relational operator stored, there are no bindings, and the block L can't + %% have any phi nodes from which we could pick bindings because when the target + %% is `one_way`, it implies the From block has a two-way `br` terminator. + #b_br{bool=#b_literal{val=true},succ=L,fail=L}; shortcut(L, From, Bs, St) -> - shortcut_1(L, From, Bs, ordsets:new(), St). + shortcut_1(L, From, Bs, cerl_sets:new(), St). shortcut_1(L, From, Bs0, UnsetVars0, St) -> case shortcut_2(L, From, Bs0, UnsetVars0, St) of @@ -170,7 +179,19 @@ shortcut_1(L, From, Bs0, UnsetVars0, St) -> end. %% Try to shortcut this block, branching to a successor. -shortcut_2(L, From, Bs0, UnsetVars0, St) -> +shortcut_2(L, From, Bs, UnsetVars, St) -> + case cerl_sets:size(UnsetVars) of + SetSize when SetSize > 128 -> + %% This is an heuristic to limit the search for a forced label + %% before it drastically slows down the compiler. Experiments + %% with scripts/diffable showed that limits larger than 31 did not + %% find any more opportunities for optimization. + none; + _SetSize -> + shortcut_3(L, From, Bs, UnsetVars, St) + end. + +shortcut_3(L, From, Bs0, UnsetVars0, St) -> #b_blk{is=Is,last=Last} = get_block(L, St), case eval_is(Is, From, Bs0, St) of none -> @@ -347,7 +368,7 @@ update_unset_vars(L, Is, Br, UnsetVars, #st{skippable=Skippable}) -> %% Some variables defined in this block are used by %% successors. We must update the set of unset variables. SetInThisBlock = [V || #b_set{dst=V} <- Is], - ordsets:union(UnsetVars, ordsets:from_list(SetInThisBlock)) + cerl_sets:union(UnsetVars, cerl_sets:from_list(SetInThisBlock)) end. shortcut_two_way(#b_br{succ=Succ,fail=Fail}, From, Bs0, UnsetVars0, St0) -> @@ -376,14 +397,14 @@ is_br_safe(UnsetVars, Br, #st{us=Us}=St) -> %% A two-way branch never branches to a phi node, so there %% is no need to check for phi nodes here. - not member(V, UnsetVars) andalso - ordsets:is_disjoint(Used0, UnsetVars) andalso - ordsets:is_disjoint(Used1, UnsetVars); + not cerl_sets:is_element(V, UnsetVars) andalso + cerl_sets:is_disjoint(Used0, UnsetVars) andalso + cerl_sets:is_disjoint(Used1, UnsetVars); #b_br{succ=Same,fail=Same} -> %% An unconditional branch must not jump to %% a phi node. not is_forbidden(Same, St) andalso - ordsets:is_disjoint(map_get(Same, Us), UnsetVars) + cerl_sets:is_disjoint(map_get(Same, Us), UnsetVars) end. is_forbidden(L, St) -> @@ -500,6 +521,15 @@ eval_switch_1([], _Arg, _PrevOp, Fail) -> %% Fail is now either the failure label or 'none'. Fail. +bind_var_if_used(L, Var, Val0, Bs, #st{us=Us}) -> + case cerl_sets:is_element(Var, map_get(L, Us)) of + true -> + Val = get_value(Val0, Bs), + Bs#{Var=>Val}; + false -> + Bs + end. + bind_var(Var, Val0, Bs) -> Val = get_value(Val0, Bs), Bs#{Var=>Val}. @@ -989,7 +1019,7 @@ used_vars([{L,#b_blk{is=Is}=Blk}|Bs], UsedVars0, Skip0) -> %% shortcut_opt/1. Successors = beam_ssa:successors(Blk), - Used0 = used_vars_succ(Successors, L, UsedVars0, []), + Used0 = used_vars_succ(Successors, L, UsedVars0, cerl_sets:new()), Used = used_vars_blk(Blk, Used0), UsedVars = used_vars_phis(Is, L, Used, UsedVars0), @@ -1000,8 +1030,8 @@ used_vars([{L,#b_blk{is=Is}=Blk}|Bs], UsedVars0, Skip0) -> %% shortcut_opt/1. Defined0 = [Def || #b_set{dst=Def} <- Is], - Defined = ordsets:from_list(Defined0), - MaySkip = ordsets:is_disjoint(Defined, Used0), + Defined = cerl_sets:from_list(Defined0), + MaySkip = cerl_sets:is_disjoint(Defined, Used0), case MaySkip of true -> Skip = Skip0#{L=>true}, @@ -1018,11 +1048,11 @@ used_vars_succ([S|Ss], L, LiveMap, Live0) -> #{Key:=Live} -> %% The successor has a phi node, and the value for %% this block in the phi node is a variable. - used_vars_succ(Ss, L, LiveMap, ordsets:union(Live, Live0)); + used_vars_succ(Ss, L, LiveMap, cerl_sets:union(Live, Live0)); #{S:=Live} -> %% No phi node in the successor, or the value for %% this block in the phi node is a literal. - used_vars_succ(Ss, L, LiveMap, ordsets:union(Live, Live0)); + used_vars_succ(Ss, L, LiveMap, cerl_sets:union(Live, Live0)); #{} -> %% A peek_message block which has not been processed yet. used_vars_succ(Ss, L, LiveMap, Live0) @@ -1040,7 +1070,7 @@ used_vars_phis(Is, L, Live0, UsedVars0) -> case [{P,V} || {#b_var{}=V,P} <- PhiArgs] of [_|_]=PhiVars -> PhiLive0 = rel2fam(PhiVars), - PhiLive = [{{L,P},ordsets:union(ordsets:from_list(Vs), Live0)} || + PhiLive = [{{L,P},cerl_sets:union(cerl_sets:from_list(Vs), Live0)} || {P,Vs} <- PhiLive0], maps:merge(UsedVars, maps:from_list(PhiLive)); [] -> @@ -1050,14 +1080,14 @@ used_vars_phis(Is, L, Live0, UsedVars0) -> end. used_vars_blk(#b_blk{is=Is,last=Last}, Used0) -> - Used = ordsets:union(Used0, beam_ssa:used(Last)), + Used = cerl_sets:union(Used0, cerl_sets:from_list(beam_ssa:used(Last))), used_vars_is(reverse(Is), Used). used_vars_is([#b_set{op=phi}|Is], Used) -> used_vars_is(Is, Used); used_vars_is([#b_set{dst=Dst}=I|Is], Used0) -> - Used1 = ordsets:union(Used0, beam_ssa:used(I)), - Used = ordsets:del_element(Dst, Used1), + Used1 = cerl_sets:union(Used0, cerl_sets:from_list(beam_ssa:used(I))), + Used = cerl_sets:del_element(Dst, Used1), used_vars_is(Is, Used); used_vars_is([], Used) -> Used. @@ -1066,8 +1096,9 @@ used_vars_is([], Used) -> %%% Common utilities. %%% -sub(#b_set{args=Args}=I, Sub) -> - I#b_set{args=[sub_arg(A, Sub) || A <- Args]}. +sub(#b_set{args=Args}=I, Sub) when map_size(Sub) =/= 0 -> + I#b_set{args=[sub_arg(A, Sub) || A <- Args]}; +sub(I, _Sub) -> I. sub_arg(#b_var{}=Old, Sub) -> case Sub of diff --git a/lib/compiler/src/beam_ssa_opt.erl b/lib/compiler/src/beam_ssa_opt.erl index 90c0d3cf16..d87c66c272 100644 --- a/lib/compiler/src/beam_ssa_opt.erl +++ b/lib/compiler/src/beam_ssa_opt.erl @@ -904,8 +904,7 @@ cse_suitable(#b_set{}) -> false. }). ssa_opt_float({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) -> - NonGuards0 = float_non_guards(Linear0), - NonGuards = gb_sets:from_list(NonGuards0), + NonGuards = non_guards(Linear0), Blocks = maps:from_list(Linear0), Fs = #fs{non_guards=NonGuards,bs=Blocks}, {Linear,Count} = float_opt(Linear0, Count0, Fs), @@ -916,15 +915,6 @@ float_blk_is_in_guard(#b_blk{last=#b_br{fail=F}}, #fs{non_guards=NonGuards}) -> float_blk_is_in_guard(#b_blk{}, #fs{}) -> false. -float_non_guards([{L,#b_blk{is=Is}}|Bs]) -> - case Is of - [#b_set{op=landingpad}|_] -> - [L|float_non_guards(Bs)]; - _ -> - float_non_guards(Bs) - end; -float_non_guards([]) -> [?BADARG_BLOCK]. - float_opt([{L,Blk}|Bs0], Count0, Fs) -> case float_blk_is_in_guard(Blk, Fs) of true -> @@ -1774,35 +1764,44 @@ opt_bs_put_split_int_1(Int, L, R) -> %%% ssa_opt_tuple_size({#st{ssa=Linear0,cnt=Count0}=St, FuncDb}) -> - {Linear,Count} = opt_tup_size(Linear0, Count0, []), + %% This optimization is only safe in guards, as prefixing tuple_size with + %% an is_tuple check prevents it from throwing an exception. + NonGuards = non_guards(Linear0), + {Linear,Count} = opt_tup_size(Linear0, NonGuards, Count0, []), {St#st{ssa=Linear,cnt=Count}, FuncDb}. -opt_tup_size([{L,#b_blk{is=Is,last=Last}=Blk}|Bs], Count0, Acc0) -> +opt_tup_size([{L,#b_blk{is=Is,last=Last}=Blk}|Bs], NonGuards, Count0, Acc0) -> case {Is,Last} of {[#b_set{op={bif,'=:='},dst=Bool,args=[#b_var{}=Tup,#b_literal{val=Arity}]}], #b_br{bool=Bool}} when is_integer(Arity), Arity >= 0 -> - {Acc,Count} = opt_tup_size_1(Tup, L, Count0, Acc0), - opt_tup_size(Bs, Count, [{L,Blk}|Acc]); + {Acc,Count} = opt_tup_size_1(Tup, L, NonGuards, Count0, Acc0), + opt_tup_size(Bs, NonGuards, Count, [{L,Blk}|Acc]); {_,_} -> - opt_tup_size(Bs, Count0, [{L,Blk}|Acc0]) + opt_tup_size(Bs, NonGuards, Count0, [{L,Blk}|Acc0]) end; -opt_tup_size([], Count, Acc) -> +opt_tup_size([], _NonGuards, Count, Acc) -> {reverse(Acc),Count}. -opt_tup_size_1(Size, EqL, Count0, [{L,Blk0}|Acc]) -> - case Blk0 of - #b_blk{is=Is0,last=#b_br{bool=Bool,succ=EqL,fail=Fail}} -> - case opt_tup_size_is(Is0, Bool, Size, []) of - none -> +opt_tup_size_1(Size, EqL, NonGuards, Count0, [{L,Blk0}|Acc]) -> + #b_blk{is=Is0,last=Last} = Blk0, + case Last of + #b_br{bool=Bool,succ=EqL,fail=Fail} -> + case gb_sets:is_member(Fail, NonGuards) of + true -> {[{L,Blk0}|Acc],Count0}; - {PreIs,TupleSizeIs,Tuple} -> - opt_tup_size_2(PreIs, TupleSizeIs, L, EqL, - Tuple, Fail, Count0, Acc) + false -> + case opt_tup_size_is(Is0, Bool, Size, []) of + none -> + {[{L,Blk0}|Acc],Count0}; + {PreIs,TupleSizeIs,Tuple} -> + opt_tup_size_2(PreIs, TupleSizeIs, L, EqL, + Tuple, Fail, Count0, Acc) + end end; - #b_blk{} -> + _ -> {[{L,Blk0}|Acc],Count0} end; -opt_tup_size_1(_, _, Count, Acc) -> +opt_tup_size_1(_, _, _, Count, Acc) -> {Acc,Count}. opt_tup_size_2(PreIs, TupleSizeIs, PreL, EqL, Tuple, Fail, Count0, Acc) -> @@ -1940,12 +1939,24 @@ verify_merge_is(_) -> is_merge_allowed(_, #b_blk{}, #b_blk{is=[#b_set{op=peek_message}|_]}) -> false; -is_merge_allowed(L, #b_blk{last=#b_br{}}=Blk, #b_blk{}) -> +is_merge_allowed(L, #b_blk{last=#b_br{}}=Blk, #b_blk{is=Is}) -> %% The predecessor block must have exactly one successor (L) for %% the merge to be safe. case beam_ssa:successors(Blk) of - [L] -> true; - [_|_] -> false + [L] -> + case Is of + [#b_set{op=phi,args=[_]}|_] -> + %% The type optimizer pass must have been + %% turned off, since it would have removed this + %% redundant phi node. Refuse to merge the blocks + %% to ensure that this phi node remains at the + %% beginning of a block. + false; + _ -> + true + end; + [_|_] -> + false end; is_merge_allowed(_, #b_blk{last=#b_switch{}}, #b_blk{}) -> false. @@ -2241,6 +2252,19 @@ gcd(A, B) -> X -> gcd(B, X) end. +non_guards(Linear) -> + gb_sets:from_list(non_guards_1(Linear)). + +non_guards_1([{L,#b_blk{is=Is}}|Bs]) -> + case Is of + [#b_set{op=landingpad}|_] -> + [L | non_guards_1(Bs)]; + _ -> + non_guards_1(Bs) + end; +non_guards_1([]) -> + [?BADARG_BLOCK]. + rel2fam(S0) -> S1 = sofs:relation(S0), S = sofs:rel2fam(S1), diff --git a/lib/compiler/src/beam_ssa_pre_codegen.erl b/lib/compiler/src/beam_ssa_pre_codegen.erl index 9af72afca7..7f816b9802 100644 --- a/lib/compiler/src/beam_ssa_pre_codegen.erl +++ b/lib/compiler/src/beam_ssa_pre_codegen.erl @@ -156,7 +156,9 @@ passes(Opts) -> %% Allocate registers. ?PASS(linear_scan), ?PASS(frame_size), - ?PASS(turn_yregs)], + ?PASS(turn_yregs), + + ?PASS(assert_no_critical_edges)], [P || P <- Ps, P =/= ignore]. function(#b_function{anno=Anno,args=Args,bs=Blocks0,cnt=Count0}=F0, @@ -1321,10 +1323,11 @@ fix_receives_1([{L,Blk}|Ls], Blocks0, Count0) -> LoopExit = find_loop_exit(Rm, Blocks0), Defs0 = beam_ssa:def([L], Blocks0), CommonUsed = recv_common(Defs0, LoopExit, Blocks0), - {Blocks1,Count1} = recv_fix_common(CommonUsed, LoopExit, Rm, - Blocks0, Count0), + {Blocks1,Count1} = recv_crit_edges(Rm, LoopExit, Blocks0, Count0), + {Blocks2,Count2} = recv_fix_common(CommonUsed, LoopExit, Rm, + Blocks1, Count1), Defs = ordsets:subtract(Defs0, CommonUsed), - {Blocks,Count} = fix_receive(Rm, Defs, Blocks1, Count1), + {Blocks,Count} = fix_receive(Rm, Defs, Blocks2, Count2), fix_receives_1(Ls, Blocks, Count); #b_blk{} -> fix_receives_1(Ls, Blocks0, Count0) @@ -1341,6 +1344,57 @@ recv_common(Defs, Exit, Blocks) -> Def = ordsets:subtract(Defs, ExitDefs), ordsets:intersection(Def, ExitUsed). +%% recv_crit_edges([RemoveMessageLabel], LoopExit, +%% Blocks0, Count0) -> {Blocks,Count}. +%% +%% Adds dummy blocks on all conditional jumps to the exit block so that +%% recv_fix_common/5 can insert phi nodes without having to worry about +%% critical edges. + +recv_crit_edges(_Rms, none, Blocks0, Count0) -> + {Blocks0, Count0}; +recv_crit_edges(Rms, Exit, Blocks0, Count0) -> + Ls = beam_ssa:rpo(Rms, Blocks0), + rce_insert_edges(Ls, Exit, Count0, Blocks0). + +rce_insert_edges([L | Ls], Exit, Count0, Blocks0) -> + Successors = beam_ssa:successors(map_get(L, Blocks0)), + case member(Exit, Successors) of + true when Successors =/= [Exit] -> + {Blocks, Count} = rce_insert_edge(L, Exit, Count0, Blocks0), + rce_insert_edges(Ls, Exit, Count, Blocks); + _ -> + rce_insert_edges(Ls, Exit, Count0, Blocks0) + end; +rce_insert_edges([], _Exit, Count, Blocks) -> + {Blocks, Count}. + +rce_insert_edge(L, Exit, Count, Blocks0) -> + #b_blk{last=Last0} = FromBlk0 = map_get(L, Blocks0), + + ToExit = #b_br{bool=#b_literal{val=true},succ=Exit,fail=Exit}, + + FromBlk = FromBlk0#b_blk{last=rce_reroute_terminator(Last0, Exit, Count)}, + EdgeBlk = #b_blk{anno=#{},is=[],last=ToExit}, + + Blocks = Blocks0#{ Count => EdgeBlk, L => FromBlk }, + {Blocks, Count + 1}. + +rce_reroute_terminator(#b_br{succ=Exit}=Last, Exit, New) -> + rce_reroute_terminator(Last#b_br{succ=New}, Exit, New); +rce_reroute_terminator(#b_br{fail=Exit}=Last, Exit, New) -> + rce_reroute_terminator(Last#b_br{fail=New}, Exit, New); +rce_reroute_terminator(#b_br{}=Last, _Exit, _New) -> + Last; +rce_reroute_terminator(#b_switch{fail=Exit}=Last, Exit, New) -> + rce_reroute_terminator(Last#b_switch{fail=New}, Exit, New); +rce_reroute_terminator(#b_switch{list=List0}=Last, Exit, New) -> + List = [if + Lbl =:= Exit -> {Arg, New}; + Lbl =/= Exit -> {Arg, Lbl} + end || {Arg, Lbl} <- List0], + Last#b_switch{list=List}. + %% recv_fix_common([CommonVar], LoopExit, [RemoveMessageLabel], %% Blocks0, Count0) -> {Blocks,Count}. %% Handle variables alwys defined in a receive and used @@ -1409,21 +1463,51 @@ fix_receive([], _Defs, Blocks, Count) -> {Blocks,Count}. %% find_loop_exit([Label], Blocks) -> Label | none. -%% Find the block to which control is transferred when the -%% the receive loop is exited. - -find_loop_exit([L1,L2|_Ls], Blocks) -> - Path1 = beam_ssa:rpo([L1], Blocks), - Path2 = beam_ssa:rpo([L2], Blocks), - find_loop_exit_1(Path1, cerl_sets:from_list(Path2)); -find_loop_exit(_, _) -> none. - -find_loop_exit_1([H|T], OtherPath) -> - case cerl_sets:is_element(H, OtherPath) of - true -> H; - false -> find_loop_exit_1(T, OtherPath) +%% Given the list of all blocks with the remove_message instructions +%% for this receive, find the block to which control is transferred +%% when the receive loop is exited (if any). + +find_loop_exit([_,_|_]=RmBlocks, Blocks) -> + %% We used to only analyze the path from two of the remove_message + %% blocks. That would fail to find a common block if one or both + %% of the blocks happened to raise an exception. To be sure that + %% we always find a common block if there is one (shared by at + %% least two clauses), we must analyze the path from all + %% remove_message blocks. + {Dominators,_} = beam_ssa:dominators(Blocks), + RmSet = cerl_sets:from_list(RmBlocks), + Rpo = beam_ssa:rpo(RmBlocks, Blocks), + find_loop_exit_1(Rpo, RmSet, Dominators); +find_loop_exit(_, _) -> + %% There is (at most) a single clause. There is no common + %% loop exit block. + none. + +find_loop_exit_1([?BADARG_BLOCK|Ls], RmSet, Dominators) -> + %% ?BADARG_BLOCK is a marker and not an actual block, so it is not + %% the block we are looking for. + find_loop_exit_1(Ls, RmSet, Dominators); +find_loop_exit_1([L|Ls], RmSet, Dominators) -> + DomBy = map_get(L, Dominators), + case any(fun(E) -> cerl_sets:is_element(E, RmSet) end, DomBy) of + true -> + %% This block is dominated by one of the remove_message blocks, + %% which means that the block is part of only one clause. + %% It is not the block we are looking for. + find_loop_exit_1(Ls, RmSet, Dominators); + false -> + %% This block is the first block that is not dominated by + %% any of the blocks with remove_message instructions, + %% which means that at least two of the receive clauses + %% will ultimately transfer control to it. It is the block + %% we are looking for. + L end; -find_loop_exit_1([], _) -> none. +find_loop_exit_1([], _, _) -> + %% None of clauses transfers control to a common block after the receive + %% statement. That means that the receive statement is a the end of a + %% function (or that all clauses raise exceptions). + none. %% find_rm_blocks(StartLabel, Blocks) -> [Label]. %% Find all blocks that start with remove_message within the receive diff --git a/lib/compiler/src/beam_ssa_share.erl b/lib/compiler/src/beam_ssa_share.erl index 426efa2cc9..73983bd34a 100644 --- a/lib/compiler/src/beam_ssa_share.erl +++ b/lib/compiler/src/beam_ssa_share.erl @@ -303,8 +303,12 @@ canonical_is([#b_ret{arg=Arg}], VarMap, Acc0) -> Acc0 end, {{ret,canonical_arg(Arg, VarMap),Acc1},VarMap}; -canonical_is([#b_br{bool=#b_var{},fail=Fail}], VarMap, Acc) -> - {{br,succ,Fail,Acc},VarMap}; +canonical_is([#b_br{bool=#b_var{}=Arg,fail=Fail}], VarMap, Acc) -> + %% A previous buggy version of this code omitted the canonicalized + %% argument in the return value. Unfortunately, that worked most + %% of the time, except when `br` terminator referenced a variable + %% defined in a previous block instead of in the same block. + {{br,canonical_arg(Arg, VarMap),succ,Fail,Acc},VarMap}; canonical_is([#b_br{succ=Succ}], VarMap, Acc) -> {{br,Succ,Acc},VarMap}; canonical_is([], VarMap, Acc) -> diff --git a/lib/compiler/src/beam_ssa_type.erl b/lib/compiler/src/beam_ssa_type.erl index 68920e7dd3..3c06c83e2e 100644 --- a/lib/compiler/src/beam_ssa_type.erl +++ b/lib/compiler/src/beam_ssa_type.erl @@ -160,6 +160,10 @@ opt_finish_1([Arg | Args], [TypeMap | TypeMaps], ParamInfo0) -> case join(maps:values(TypeMap)) of any -> opt_finish_1(Args, TypeMaps, ParamInfo0); + none -> + %% This function will never be called. Pretend that we don't + %% know the type for this argument. + opt_finish_1(Args, TypeMaps, ParamInfo0); JoinedType -> JoinedType = verified_type(JoinedType), ParamInfo = ParamInfo0#{ Arg => validator_anno(JoinedType) }, diff --git a/lib/compiler/src/beam_validator.erl b/lib/compiler/src/beam_validator.erl index ebe9631e09..349d74eb58 100644 --- a/lib/compiler/src/beam_validator.erl +++ b/lib/compiler/src/beam_validator.erl @@ -1068,8 +1068,11 @@ verify_get_map(Fail, Src, List, Vst0) -> %% {get_map_elements,{f,7},{x,1},{list,[{atom,a},{x,1},{atom,b},{x,2}]}}. %% %% If 'a' exists but not 'b', {x,1} is overwritten when we jump to {f,7}. +%% +%% We must be careful to preserve the uninitialized status for Y registers +%% that have been allocated but not yet defined. clobber_map_vals([Key,Dst|T], Map, Vst0) -> - case is_reg_defined(Dst, Vst0) of + case is_reg_initialized(Dst, Vst0) of true -> Vst = extract_term(term, {bif,map_get}, [Key, Map], Dst, Vst0), clobber_map_vals(T, Map, Vst); @@ -1079,6 +1082,17 @@ clobber_map_vals([Key,Dst|T], Map, Vst0) -> clobber_map_vals([], _Map, Vst) -> Vst. +is_reg_initialized({x,_}=Reg, #vst{current=#st{xs=Xs}}) -> + is_map_key(Reg, Xs); +is_reg_initialized({y,_}=Reg, #vst{current=#st{ys=Ys}}) -> + case Ys of + #{Reg:=Val} -> + Val =/= uninitialized; + #{} -> + false + end; +is_reg_initialized(V, #vst{}) -> error({not_a_register, V}). + extract_map_keys([Key,_Val|T]) -> [Key|extract_map_keys(T)]; extract_map_keys([]) -> []. @@ -1604,13 +1618,8 @@ infer_types_1(#value{op={bif,'=:='},args=[LHS,RHS]}) -> end; infer_types_1(#value{op={bif,element},args=[{integer,Index}=Key,Tuple]}) -> fun(Val, S) -> - case is_value_alive(Tuple, S) of - true -> - Type = {tuple,[Index], #{ Key => get_term_type(Val, S) }}, - update_type(fun meet/2, Type, Tuple, S); - false -> - S - end + Type = {tuple,[Index], #{ Key => get_term_type(Val, S) }}, + update_type(fun meet/2, Type, Tuple, S) end; infer_types_1(#value{op={bif,is_atom},args=[Src]}) -> infer_type_test_bif({atom,[]}, Src); @@ -1634,10 +1643,7 @@ infer_types_1(#value{op={bif,is_tuple},args=[Src]}) -> infer_type_test_bif({tuple,[0],#{}}, Src); infer_types_1(#value{op={bif,tuple_size}, args=[Tuple]}) -> fun({integer,Arity}, S) -> - case is_value_alive(Tuple, S) of - true -> update_type(fun meet/2, {tuple,Arity,#{}}, Tuple, S); - false -> S - end; + update_type(fun meet/2, {tuple,Arity,#{}}, Tuple, S); (_, S) -> S end; infer_types_1(_) -> @@ -1645,10 +1651,7 @@ infer_types_1(_) -> infer_type_test_bif(Type, Src) -> fun({atom,true}, S) -> - case is_value_alive(Src, S) of - true -> update_type(fun meet/2, Type, Src, S); - false -> S - end; + update_type(fun meet/2, Type, Src, S); (_, S) -> S end. @@ -1885,10 +1888,6 @@ check_try_catch_tags(Type, {y,N}=Reg, Vst) -> ok end. -is_reg_defined({x,_}=Reg, #vst{current=#st{xs=Xs}}) -> is_map_key(Reg, Xs); -is_reg_defined({y,_}=Reg, #vst{current=#st{ys=Ys}}) -> is_map_key(Reg, Ys); -is_reg_defined(V, #vst{}) -> error({not_a_register, V}). - assert_term(Src, Vst) -> _ = get_term_type(Src, Vst), ok. @@ -2285,9 +2284,6 @@ get_raw_type(#value_ref{}=Ref, #vst{current=#st{vs=Vs}}) -> get_raw_type(Src, #vst{}) -> get_literal_type(Src). -is_value_alive(#value_ref{}=Ref, #vst{current=#st{vs=Vs}}) -> - is_map_key(Ref, Vs). - get_literal_type(nil=T) -> T; get_literal_type({atom,A}=T) when is_atom(A) -> T; get_literal_type({float,F}=T) when is_float(F) -> T; @@ -2469,25 +2465,44 @@ merge_vrefs(RefA, RefB, Merge, Counter) -> merge_values(Merge, VsA, VsB) -> maps:fold(fun(Spec, New, Acc) -> - merge_values_1(Spec, New, VsA, VsB, Acc) + mv_1(Spec, New, VsA, VsB, Acc) end, #{}, Merge). -merge_values_1(Same, Same, VsA, VsB, Acc) -> +mv_1(Same, Same, VsA, VsB, Acc0) -> %% We're merging different versions of the same value, so it's safe to %% reuse old entries if the type's unchanged. - #value{type=TypeA}=EntryA = map_get(Same, VsA), - #value{type=TypeB}=EntryB = map_get(Same, VsB), + #value{type=TypeA,args=Args}=EntryA = map_get(Same, VsA), + #value{type=TypeB,args=Args}=EntryB = map_get(Same, VsB), + Entry = case join(TypeA, TypeB) of TypeA -> EntryA; TypeB -> EntryB; JoinedType -> EntryA#value{type=JoinedType} end, - Acc#{ Same => Entry }; -merge_values_1({RefA, RefB}, New, VsA, VsB, Acc) -> + + Acc = Acc0#{ Same => Entry }, + + %% Type inference may depend on values that are no longer reachable from a + %% register, so all arguments must be merged into the new state. + mv_args(Args, VsA, VsB, Acc); +mv_1({RefA, RefB}, New, VsA, VsB, Acc) -> #value{type=TypeA} = map_get(RefA, VsA), #value{type=TypeB} = map_get(RefB, VsB), Acc#{ New => #value{op=join,args=[],type=join(TypeA, TypeB)} }. +mv_args([#value_ref{}=Arg | Args], VsA, VsB, Acc0) -> + case Acc0 of + #{ Arg := _ } -> + mv_args(Args, VsA, VsB, Acc0); + #{} -> + Acc = mv_1(Arg, Arg, VsA, VsB, Acc0), + mv_args(Args, VsA, VsB, Acc) + end; +mv_args([_ | Args], VsA, VsB, Acc) -> + mv_args(Args, VsA, VsB, Acc); +mv_args([], _VsA, _VsB, Acc) -> + Acc. + merge_fragility(FragileA, FragileB) -> cerl_sets:union(FragileA, FragileB). diff --git a/lib/compiler/src/cerl.erl b/lib/compiler/src/cerl.erl index 62cd5b5120..bc28f58712 100644 --- a/lib/compiler/src/cerl.erl +++ b/lib/compiler/src/cerl.erl @@ -263,7 +263,7 @@ %% @see subtrees/1 %% @see meta/1 --type ctype() :: 'alias' | 'apply' | 'binary' | 'bitrst' | 'call' | 'case' +-type ctype() :: 'alias' | 'apply' | 'binary' | 'bitstr' | 'call' | 'case' | 'catch' | 'clause' | 'cons' | 'fun' | 'let' | 'letrec' | 'literal' | 'map' | 'map_pair' | 'module' | 'primop' | 'receive' | 'seq' | 'try' | 'tuple' | 'values' | 'var'. diff --git a/lib/compiler/src/compile.erl b/lib/compiler/src/compile.erl index 28db8986ff..0325c714d0 100644 --- a/lib/compiler/src/compile.erl +++ b/lib/compiler/src/compile.erl @@ -268,8 +268,11 @@ expand_opt(r21, Os) -> [no_put_tuple2 | expand_opt(no_bsm3, Os)]; expand_opt({debug_info_key,_}=O, Os) -> [encrypt_debug_info,O|Os]; -expand_opt(no_type_opt, Os) -> - [no_ssa_opt_type_start, +expand_opt(no_type_opt=O, Os) -> + %% Be sure to keep the no_type_opt option so that it will + %% be recorded in the BEAM file, allowing the test suites + %% to recompile the file with this option. + [O,no_ssa_opt_type_start, no_ssa_opt_type_continue, no_ssa_opt_type_finish | Os]; expand_opt(O, Os) -> [O|Os]. diff --git a/lib/compiler/test/Makefile b/lib/compiler/test/Makefile index db8eb7e2e1..7be23fbb93 100644 --- a/lib/compiler/test/Makefile +++ b/lib/compiler/test/Makefile @@ -109,6 +109,8 @@ NO_MOD_OPT = $(NO_OPT) NO_SSA_OPT = $(NO_OPT) +NO_TYPE_OPT = $(NO_OPT) + NO_OPT_MODULES= $(NO_OPT:%=%_no_opt_SUITE) NO_OPT_ERL_FILES= $(NO_OPT_MODULES:%=%.erl) POST_OPT_MODULES= $(NO_OPT:%=%_post_opt_SUITE) @@ -121,6 +123,8 @@ NO_MOD_OPT_MODULES= $(NO_MOD_OPT:%=%_no_module_opt_SUITE) NO_MOD_OPT_ERL_FILES= $(NO_MOD_OPT_MODULES:%=%.erl) NO_SSA_OPT_MODULES= $(NO_SSA_OPT:%=%_no_ssa_opt_SUITE) NO_SSA_OPT_ERL_FILES= $(NO_SSA_OPT_MODULES:%=%.erl) +NO_TYPE_OPT_MODULES= $(NO_TYPE_OPT:%=%_no_type_opt_SUITE) +NO_TYPE_OPT_ERL_FILES= $(NO_TYPE_OPT_MODULES:%=%.erl) ERL_FILES= $(MODULES:%=%.erl) CORE_FILES= $(CORE_MODULES:%=%.core) @@ -150,7 +154,7 @@ EBIN = . # ---------------------------------------------------- make_emakefile: $(NO_OPT_ERL_FILES) $(POST_OPT_ERL_FILES) $(NO_SSA_OPT_ERL_FILES) \ - $(INLINE_ERL_FILES) $(R21_ERL_FILES) $(NO_MOD_OPT_ERL_FILES) + $(INLINE_ERL_FILES) $(R21_ERL_FILES) $(NO_MOD_OPT_ERL_FILES) $(NO_TYPE_OPT_ERL_FILES) $(ERL_TOP)/make/make_emakefile $(ERL_COMPILE_FLAGS) -o$(EBIN) $(MODULES) \ > $(EMAKEFILE) $(ERL_TOP)/make/make_emakefile +no_copt +no_postopt \ @@ -169,6 +173,8 @@ make_emakefile: $(NO_OPT_ERL_FILES) $(POST_OPT_ERL_FILES) $(NO_SSA_OPT_ERL_FILES -o$(EBIN) $(NO_MOD_OPT_MODULES) >> $(EMAKEFILE) $(ERL_TOP)/make/make_emakefile +from_core $(ERL_COMPILE_FLAGS) \ -o$(EBIN) $(CORE_MODULES) >> $(EMAKEFILE) + $(ERL_TOP)/make/make_emakefile +no_type_opt $(ERL_COMPILE_FLAGS) \ + -o$(EBIN) $(NO_TYPE_OPT_MODULES) >> $(EMAKEFILE) tests debug opt: make_emakefile erl $(ERL_MAKE_FLAGS) -make @@ -202,6 +208,10 @@ docs: %_no_module_opt_SUITE.erl: %_SUITE.erl sed -e 's;-module($(basename $<));-module($(basename $@));' $< > $@ +%_no_type_opt_SUITE.erl: %_SUITE.erl + sed -e 's;-module($(basename $<));-module($(basename $@));' $< > $@ + + # ---------------------------------------------------- # Release Target # ---------------------------------------------------- @@ -216,7 +226,8 @@ release_tests_spec: make_emakefile $(INSTALL_DATA) $(NO_OPT_ERL_FILES) $(POST_OPT_ERL_FILES) \ $(INLINE_ERL_FILES) $(R21_ERL_FILES) \ $(NO_MOD_OPT_ERL_FILES) \ - $(NO_SSA_OPT_ERL_FILES) "$(RELSYSDIR)" + $(NO_SSA_OPT_ERL_FILES) \ + $(NO_TYPE_OPT_ERL_FILES) "$(RELSYSDIR)" $(INSTALL_DATA) $(CORE_FILES) "$(RELSYSDIR)" for file in $(ERL_DUMMY_FILES); do \ module=`basename $$file .erl`; \ diff --git a/lib/compiler/test/beam_except_SUITE.erl b/lib/compiler/test/beam_except_SUITE.erl index 67947dc292..f52239f2a8 100644 --- a/lib/compiler/test/beam_except_SUITE.erl +++ b/lib/compiler/test/beam_except_SUITE.erl @@ -72,11 +72,25 @@ bs_get_tail(Config) -> {function_clause, [{?MODULE,bs_get_tail_1,[<<>>,0,0,Config],_}|_]}} = (catch bs_get_tail_1(id(<<>>), 0, 0, Config)), + + ok = bs_get_tail_2(<<"W">>, <<"X">>, <<"Z">>), + ok = bs_get_tail_2(<<"M">>, <<"X">>, <<"Z">>), + {'EXIT', + {function_clause, + [{?MODULE,do_get_bs_tail_2,[<<"A">>,<<"B">>,[],<<"C">>],_}|_]}} = + (catch bs_get_tail_2(<<"A">>, <<"B">>, <<"C">>)), + ok. bs_get_tail_1(<<_:32, Rest/binary>>, Z1, Z2, F1) -> {Rest,Z1,Z2,F1}. +bs_get_tail_2(A, B, C) -> + do_get_bs_tail_2(A, B, [], C). + +do_get_bs_tail_2(<<"W">>, <<"X">>, _, <<"Z">>) -> ok; +do_get_bs_tail_2(<<"M">>, <<"X">>, _, <<"Z">>) -> ok. + coverage(_) -> File = {file,"fake.erl"}, ok = fc(a), diff --git a/lib/compiler/test/beam_ssa_SUITE.erl b/lib/compiler/test/beam_ssa_SUITE.erl index a741ebbdf9..3b510f3528 100644 --- a/lib/compiler/test/beam_ssa_SUITE.erl +++ b/lib/compiler/test/beam_ssa_SUITE.erl @@ -23,7 +23,7 @@ init_per_group/2,end_per_group/2, calls/1,tuple_matching/1,recv/1,maps/1, cover_ssa_dead/1,combine_sw/1,share_opt/1, - beam_ssa_dead_crash/1]). + beam_ssa_dead_crash/1,stack_init/1]). suite() -> [{ct_hooks,[ts_install_cth]}]. @@ -39,7 +39,8 @@ groups() -> cover_ssa_dead, combine_sw, share_opt, - beam_ssa_dead_crash + beam_ssa_dead_crash, + stack_init ]}]. init_per_suite(Config) -> @@ -190,6 +191,15 @@ recv(_Config) -> self() ! {[self(),r1],{2,99,<<"data">>}}, {Parent,r1,<<1:32,2:8,99:8,"data">>} = tricky_recv_4(), + %% Test tricky_recv_5/0. + self() ! 1, + a = tricky_recv_5(), + self() ! 2, + b = tricky_recv_5(), + + %% tricky_recv_6/0 is a compile-time error. + tricky_recv_6(), + ok. sync_wait_mon({Pid, Ref}, Timeout) -> @@ -295,6 +305,38 @@ tricky_recv_4() -> end, id({Pid,R,Request}). +%% beam_ssa_pre_codegen would accidentally create phi nodes on critical edges +%% when fixing up receives; the call to id/2 can either succeed or land in the +%% catch block, and we added a phi node to its immediate successor. +tricky_recv_5() -> + try + receive + X=1 -> + id(42), + a; + X=2 -> + b + end, + case X of + 1 -> a; + 2 -> b + end + catch + _:_ -> c + end. + +%% When fixing tricky_recv_5, we introduced a compiler crash when the common +%% exit block was ?BADARG_BLOCK and floats were in the picture. +tricky_recv_6() -> + RefA = make_ref(), + RefB = make_ref(), + receive + {RefA, Number} -> Number + 1.0; + {RefB, Number} -> Number + 2.0 + after 0 -> + ok + end. + maps(_Config) -> {'EXIT',{{badmatch,#{}},_}} = (catch maps_1(any)), ok. @@ -483,9 +525,11 @@ do_comb_sw_2(X) -> erase(?MODULE). share_opt(_Config) -> - ok = do_share_opt(0). + ok = do_share_opt_1(0), + ok = do_share_opt_2(), + ok. -do_share_opt(A) -> +do_share_opt_1(A) -> %% The compiler would be stuck in an infinite loop in beam_ssa_share. case A of 0 -> a; @@ -494,6 +538,26 @@ do_share_opt(A) -> end, receive after 1 -> ok end. +do_share_opt_2() -> + ok = sopt_2({[pointtopoint], [{dstaddr,any}]}, ok), + ok = sopt_2({[broadcast], [{broadaddr,any}]}, ok), + ok = sopt_2({[], []}, ok), + ok. + +sopt_2({Flags, Opts}, ok) -> + Broadcast = lists:member(broadcast, Flags), + P2P = lists:member(pointtopoint, Flags), + case Opts of + %% The following two clauses would be combined to one, silently + %% discarding the guard test of the P2P variable. + [{broadaddr,_}|Os] when Broadcast -> + sopt_2({Flags, Os}, ok); + [{dstaddr,_}|Os] when P2P -> + sopt_2({Flags, Os}, ok); + [] -> + ok + end. + beam_ssa_dead_crash(_Config) -> not_A_B = do_beam_ssa_dead_crash(id(false), id(true)), not_A_not_B = do_beam_ssa_dead_crash(false, false), @@ -548,6 +612,30 @@ do_beam_ssa_dead_crash(A, B) -> end end. +stack_init(_Config) -> + 6 = stack_init(a, #{a => [1,2,3]}), + 0 = stack_init(missing, #{}), + ok. + +stack_init(Key, Map) -> + %% beam_ssa_codegen would wrongly assume that y(0) would always be + %% initialized by the `get_map_elements` instruction that follows, and + %% would set up the stack frame using an `allocate` instruction and + %% would not generate an `init` instruction to initialize y(0). + Res = case Map of + #{Key := Elements} -> + %% Elements will be assigned to y(0) if the key Key exists. + lists:foldl(fun(El, Acc) -> + Acc + El + end, 0, Elements); + #{} -> + %% y(0) will be left uninitialized when the key is not + %% present in the map. + 0 + end, + %% y(0) would be uninitialized here if the key was not present in the map + %% (if the second clause was executed). + id(Res). %% The identity function. id(I) -> I. diff --git a/lib/compiler/test/beam_type_SUITE.erl b/lib/compiler/test/beam_type_SUITE.erl index 076a604aa4..a99dee48aa 100644 --- a/lib/compiler/test/beam_type_SUITE.erl +++ b/lib/compiler/test/beam_type_SUITE.erl @@ -24,7 +24,8 @@ integers/1,numbers/1,coverage/1,booleans/1,setelement/1, cons/1,tuple/1,record_float/1,binary_float/1,float_compare/1, arity_checks/1,elixir_binaries/1,find_best/1, - test_size/1,cover_lists_functions/1,list_append/1,bad_binary_unit/1]). + test_size/1,cover_lists_functions/1,list_append/1,bad_binary_unit/1, + none_argument/1]). suite() -> [{ct_hooks,[ts_install_cth]}]. @@ -49,7 +50,8 @@ groups() -> test_size, cover_lists_functions, list_append, - bad_binary_unit + bad_binary_unit, + none_argument ]}]. init_per_suite(Config) -> @@ -518,5 +520,24 @@ bad_binary_unit(_Config) -> false = is_binary(Bitstring), ok. +%% ERL-1013: The compiler would crash during the type optimization pass. +none_argument(_Config) -> + Binary = id(<<3:16, 42>>), + error = id(case Binary of + <<Len:16, Body/binary>> when length(Body) == Len - 2 -> + %% The type for Body will be none. It means + %% that this clause will never match and that + %% uncompress/1 will never be called. + uncompress(Body); + _ -> + error + end), + ok. + +uncompress(CompressedBinary) -> + %% The type for CompressedBinary is none, which beam_ssa_type + %% did not handle properly. + zlib:uncompress(CompressedBinary). + id(I) -> I. diff --git a/lib/compiler/test/beam_validator_SUITE.erl b/lib/compiler/test/beam_validator_SUITE.erl index 6b1438abdd..20f6cb2691 100644 --- a/lib/compiler/test/beam_validator_SUITE.erl +++ b/lib/compiler/test/beam_validator_SUITE.erl @@ -681,11 +681,16 @@ infer_on_eq_4(T) -> %% ERIERL-348; types were inferred for dead values, causing validation to fail. +-record(idv, {key}). + infer_dead_value(Config) when is_list(Config) -> a = idv_1({a, b, c, d, e, f, g}, {0, 0, 0, 0, 0, 0, 0}), b = idv_1({a, b, c, d, 0, 0, 0}, {a, b, c, d, 0, 0, 0}), c = idv_1({0, 0, 0, 0, 0, f, g}, {0, 0, 0, 0, 0, f, g}), error = idv_1(gurka, gaffel), + + ok = idv_2(id(#idv{})), + ok. idv_1({_A, _B, _C, _D, _E, _F, _G}, @@ -700,6 +705,23 @@ idv_1({_A, _B, _C, _D, _E, F, G}, idv_1(_A, _B) -> error. +%% ERL-995: The first solution to ERIERL-348 was incomplete and caused +%% validation to fail when living values depended on delayed type inference on +%% "dead" values. + +idv_2(State) -> + Flag = (State#idv.key == undefined), + case id(gurka) of + {_} -> id([Flag]); + _ -> ok + end, + if + Flag -> idv_called_once(State); + true -> ok + end. + +idv_called_once(_State) -> ok. + %%%------------------------------------------------------------------------- transform_remove(Remove, Module) -> diff --git a/lib/compiler/test/bs_match_SUITE.erl b/lib/compiler/test/bs_match_SUITE.erl index d97f49c56e..145a50f4ad 100644 --- a/lib/compiler/test/bs_match_SUITE.erl +++ b/lib/compiler/test/bs_match_SUITE.erl @@ -44,7 +44,8 @@ beam_bsm/1,guard/1,is_ascii/1,non_opt_eq/1, expression_before_match/1,erl_689/1,restore_on_call/1, restore_after_catch/1,matches_on_parameter/1,big_positions/1, - matching_meets_apply/1,bs_start_match2_defs/1]). + matching_meets_apply/1,bs_start_match2_defs/1, + exceptions_after_match_failure/1]). -export([coverage_id/1,coverage_external_ignore/2]). @@ -80,7 +81,8 @@ groups() -> beam_bsm,guard,is_ascii,non_opt_eq, expression_before_match,erl_689,restore_on_call, matches_on_parameter,big_positions, - matching_meets_apply,bs_start_match2_defs]}]. + matching_meets_apply,bs_start_match2_defs, + exceptions_after_match_failure]}]. init_per_suite(Config) -> @@ -2005,4 +2007,17 @@ do_matching_meets_apply(_Bin, {Handler, State}) -> %% Another case of the above. Handler:abs(State). +%% Exception handling was broken on the failure path of bs_start_match as +%% beam_ssa_bsm accidentally cloned and renamed the ?BADARG_BLOCK. +exceptions_after_match_failure(_Config) -> + {'EXIT', {badarith, _}} = (catch do_exceptions_after_match_failure(atom)), + ok = do_exceptions_after_match_failure(<<0, 1, "gurka">>), + ok = do_exceptions_after_match_failure(2.0). + +do_exceptions_after_match_failure(<<_A, _B, "gurka">>) -> + ok; +do_exceptions_after_match_failure(Other) -> + Other / 2.0, + ok. + id(I) -> I. diff --git a/lib/compiler/test/match_SUITE.erl b/lib/compiler/test/match_SUITE.erl index aac9de278d..bc74ec4984 100644 --- a/lib/compiler/test/match_SUITE.erl +++ b/lib/compiler/test/match_SUITE.erl @@ -25,7 +25,8 @@ match_in_call/1,untuplify/1,shortcut_boolean/1,letify_guard/1, selectify/1,deselectify/1,underscore/1,match_map/1,map_vars_used/1, coverage/1,grab_bag/1,literal_binary/1, - unary_op/1,eq_types/1,match_after_return/1,match_right_tuple/1]). + unary_op/1,eq_types/1,match_after_return/1,match_right_tuple/1, + tuple_size_in_try/1]). -include_lib("common_test/include/ct.hrl"). @@ -41,7 +42,8 @@ groups() -> shortcut_boolean,letify_guard,selectify,deselectify, underscore,match_map,map_vars_used,coverage, grab_bag,literal_binary,unary_op,eq_types, - match_after_return,match_right_tuple]}]. + match_after_return,match_right_tuple, + tuple_size_in_try]}]. init_per_suite(Config) -> @@ -922,4 +924,19 @@ match_right_tuple_1(T) -> force_succ_regs(_A, B) -> B. +tuple_size_in_try(Config) when is_list(Config) -> + %% The tuple_size optimization was applied outside of guards, causing + %% either the emulator or compiler to crash. + ok = tsit(gurka), + ok = tsit(gaffel). + +tsit(A) -> + try + id(ignored), + 1 = tuple_size(A), + error + catch + _:_ -> ok + end. + id(I) -> I. diff --git a/lib/compiler/test/receive_SUITE.erl b/lib/compiler/test/receive_SUITE.erl index 752491f0f8..8cd864c59e 100644 --- a/lib/compiler/test/receive_SUITE.erl +++ b/lib/compiler/test/receive_SUITE.erl @@ -431,6 +431,20 @@ elusive_common_exit(_Config) -> self() ! {1, a}, self() ! {2, b}, {[z], [{2,b},{1,a}]} = elusive_loop([x,y,z], 2, []), + + CodeServer = whereis(code_server), + Self = self(), + Self ! {Self, abc}, + Self ! {CodeServer, []}, + Self ! {Self, other}, + try elusive2([]) of + Unexpected -> + ct:fail("Expected an exception; got ~p\n", [Unexpected]) + catch + throw:[other, CodeServer, Self] -> + ok + end, + ok. elusive_loop(List, 0, Results) -> @@ -449,4 +463,25 @@ elusive_loop(List, ToReceive, Results) -> %% that it would not insert all necessary copy instructions. elusive_loop(RemList, ToReceive-1, [Result | Results]). + +elusive2(Acc) -> + receive + {Pid, abc} -> + ok; + {Pid, []} -> + ok; + {Pid, Res} -> + %% beam_ssa_pre_codegen:find_loop_exit/2 attempts to find + %% the first block of the common code after the receive + %% statement. It used to only look at the two last clauses + %% of the receive. In this function, the last two clauses + %% don't have any common block, so it would be assumed + %% that there was no common block for any of the + %% clauses. That would mean that copy instructions would + %% not be inserted as needed. + throw([Res | Acc]) + end, + %% Common code. + elusive2([Pid | Acc]). + id(I) -> I. diff --git a/lib/compiler/test/test_lib.erl b/lib/compiler/test/test_lib.erl index 3348c6e9ea..34410e4b2a 100644 --- a/lib/compiler/test/test_lib.erl +++ b/lib/compiler/test/test_lib.erl @@ -97,7 +97,8 @@ get_data_dir(Config) -> Data2 = re:replace(Data1, "_post_opt_SUITE", "_SUITE", Opts), Data3 = re:replace(Data2, "_inline_SUITE", "_SUITE", Opts), Data4 = re:replace(Data3, "_r21_SUITE", "_SUITE", Opts), - Data = re:replace(Data4, "_no_module_opt_SUITE", "_SUITE", Opts), + Data5 = re:replace(Data4, "_no_module_opt_SUITE", "_SUITE", Opts), + Data = re:replace(Data5, "_no_type_opt_SUITE", "_SUITE", Opts), re:replace(Data, "_no_ssa_opt_SUITE", "_SUITE", Opts). is_cloned_mod(Mod) -> diff --git a/lib/compiler/vsn.mk b/lib/compiler/vsn.mk index 508bbc902c..7192ddca15 100644 --- a/lib/compiler/vsn.mk +++ b/lib/compiler/vsn.mk @@ -1 +1 @@ -COMPILER_VSN = 7.4.2 +COMPILER_VSN = 7.4.4 |