aboutsummaryrefslogtreecommitdiffstats
path: root/lib/compiler
diff options
context:
space:
mode:
authorBjörn Gustavsson <[email protected]>2016-08-05 12:01:56 +0200
committerBjörn Gustavsson <[email protected]>2016-08-05 12:01:56 +0200
commit08a5b907960aba230a3790cc8502c4ca5444ad96 (patch)
tree406ea2cf76a4aabf5efeb75e9f20c3ad6c96ef01 /lib/compiler
parente487bb1b14e8abfa5fa4dfd310d7fae87c941bc7 (diff)
parent65266b2db99ea2651606acd3a796738f1a59dc40 (diff)
downloadotp-08a5b907960aba230a3790cc8502c4ca5444ad96.tar.gz
otp-08a5b907960aba230a3790cc8502c4ca5444ad96.tar.bz2
otp-08a5b907960aba230a3790cc8502c4ca5444ad96.zip
Merge branch 'maint'
* maint: Update primary bootstrap beam_block: Fix potentially unsafe optimization in move_allocates/1
Diffstat (limited to 'lib/compiler')
-rw-r--r--lib/compiler/src/beam_block.erl98
-rw-r--r--lib/compiler/test/beam_block_SUITE.erl27
2 files changed, 88 insertions, 37 deletions
diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl
index 85d332c56e..ec41925beb 100644
--- a/lib/compiler/src/beam_block.erl
+++ b/lib/compiler/src/beam_block.erl
@@ -159,14 +159,43 @@ find_fixpoint(OptFun, Is0) ->
end.
%% move_allocates(Is0) -> Is
-%% Move allocate instructions upwards in the instruction stream, in the
-%% hope of getting more possibilities for optimizing away moves later.
+%% Move allocate instructions upwards in the instruction stream
+%% (within the same block), in the hope of getting more possibilities
+%% for optimizing away moves later.
%%
-%% NOTE: Moving allocation instructions is only safe because it is done
-%% immediately after code generation so that we KNOW that if {x,X} is
-%% initialized, all x registers with lower numbers are also initialized.
-%% That assumption may not be true after other optimizations, such as
-%% the beam_utils:live_opt/1 optimization.
+%% For example, we can transform the following instructions:
+%%
+%% get_tuple_element x(1) Element => x(2)
+%% allocate_zero StackSize 3 %% x(0), x(1), x(2) are live
+%%
+%% to the following instructions:
+%%
+%% allocate_zero StackSize 2 %% x(0) and x(1) are live
+%% get_tuple_element x(1) Element => x(2)
+%%
+%% NOTE: Since the beam_reorder pass has been run, it is no longer
+%% safe to assume that if x(N) is initialized, then all lower-numbered
+%% x registers are also initialized.
+%%
+%% For example, in general it is not safe to transform the following
+%% instructions:
+%%
+%% get_tuple_element x(0) Element => x(1)
+%% allocate_zero StackSize 3 %x(0), x(1), x(2) are live
+%%
+%% to the following instructions:
+%%
+%% allocate_zero StackSize 3
+%% get_tuple_element x(0) Element => x(1)
+%%
+%% The transformation is safe if and only if x(1) has been
+%% initialized previously. Unfortunately, beam_reorder may have moved
+%% a get_tuple_element instruction so that x(1) is not always
+%% initialized when this code is reached. To find whether or not x(1)
+%% is initialized, we would need to analyze all code preceding these
+%% two instructions (across branches). Since we currently don't have
+%% any practical mechanism for doing that, we will have to
+%% conservatively assume that the transformation is unsafe.
move_allocates([{block,Bl0}|Is]) ->
Bl = move_allocates_1(reverse(Bl0), []),
@@ -175,27 +204,19 @@ move_allocates([I|Is]) ->
[I|move_allocates(Is)];
move_allocates([]) -> [].
-move_allocates_1([{set,[],[],{alloc,_,_}=Alloc}|Is0], Acc0) ->
- {Is,Acc} = move_allocates_2(Alloc, Is0, Acc0),
- move_allocates_1(Is, Acc);
+move_allocates_1([I|Is], [{set,[],[],{alloc,Live0,Info}}|Acc]=Acc0) ->
+ case {alloc_may_pass(I),alloc_live_regs(I, Live0)} of
+ {false,_} ->
+ move_allocates_1(Is, [I|Acc0]);
+ {true,not_possible} ->
+ move_allocates_1(Is, [I|Acc0]);
+ {true,Live} when is_integer(Live) ->
+ A = {set,[],[],{alloc,Live,Info}},
+ move_allocates_1(Is, [A,I|Acc])
+ end;
move_allocates_1([I|Is], Acc) ->
move_allocates_1(Is, [I|Acc]);
-move_allocates_1([], Is) -> Is.
-
-move_allocates_2({alloc,Live,Info}, [{set,[],[],{alloc,Live0,Info0}}|Is], Acc) ->
- Live = Live0, % Assertion.
- Alloc = {alloc,Live,combine_alloc(Info0, Info)},
- move_allocates_2(Alloc, Is, Acc);
-move_allocates_2({alloc,Live,Info}=Alloc0, [I|Is]=Is0, Acc) ->
- case alloc_may_pass(I) of
- false ->
- {Is0,[{set,[],[],Alloc0}|Acc]};
- true ->
- Alloc = {alloc,alloc_live_regs(I, Live),Info},
- move_allocates_2(Alloc, Is, [I|Acc])
- end;
-move_allocates_2(Alloc, [], Acc) ->
- {[],[{set,[],[],Alloc}|Acc]}.
+move_allocates_1([], Acc) -> Acc.
alloc_may_pass({set,_,_,{alloc,_,_}}) -> false;
alloc_may_pass({set,_,_,{set_tuple_element,_}}) -> false;
@@ -204,9 +225,6 @@ alloc_may_pass({set,_,_,put_list}) -> false;
alloc_may_pass({set,_,_,put}) -> false;
alloc_may_pass({set,_,_,_}) -> true.
-combine_alloc({_,Ns,Nh1,Init}, {_,nostack,Nh2,[]}) ->
- {zero,Ns,beam_utils:combine_heap_needs(Nh1, Nh2),Init}.
-
%% opt([Instruction]) -> [Instruction]
%% Optimize the instruction stream inside a basic block.
@@ -393,10 +411,19 @@ eliminate_use_of_from_reg([I]=Is, From, _To, Acc) ->
%% opt_alloc(Instructions) -> Instructions'
%% Optimises all allocate instructions.
+opt_alloc([{set,[],[],{alloc,Live0,Info0}},
+ {set,[],[],{alloc,Live,Info}}|Is]) ->
+ Live = Live0, %Assertion.
+ Alloc = combine_alloc(Info0, Info),
+ I = {set,[],[],{alloc,Live,Alloc}},
+ opt_alloc([I|Is]);
opt_alloc([{set,[],[],{alloc,R,{_,Ns,Nh,[]}}}|Is]) ->
[{set,[],[],opt_alloc(Is, Ns, Nh, R)}|Is];
opt_alloc([I|Is]) -> [I|opt_alloc(Is)];
opt_alloc([]) -> [].
+
+combine_alloc({_,Ns,Nh1,Init}, {_,nostack,Nh2,[]}) ->
+ {zero,Ns,beam_utils:combine_heap_needs(Nh1, Nh2),Init}.
%% opt_alloc(Instructions, FrameSize, HeapNeed, LivingRegs) -> [Instr]
%% Generates the optimal sequence of instructions for
@@ -445,13 +472,14 @@ count_ones(Bits, Acc) ->
alloc_live_regs({set,Ds,Ss,_}, Regs0) ->
Rset = x_live(Ss, x_dead(Ds, (1 bsl Regs0)-1)),
- live_regs(Rset).
+ live_regs(0, Rset).
-live_regs(Regs) ->
- live_regs_1(0, Regs).
-
-live_regs_1(N, 0) -> N;
-live_regs_1(N, Regs) -> live_regs_1(N+1, Regs bsr 1).
+live_regs(N, 0) ->
+ N;
+live_regs(N, Regs) when Regs band 1 =:= 1 ->
+ live_regs(N+1, Regs bsr 1);
+live_regs(_, _) ->
+ not_possible.
x_dead([{x,N}|Rs], Regs) -> x_dead(Rs, Regs band (bnot (1 bsl N)));
x_dead([_|Rs], Regs) -> x_dead(Rs, Regs);
diff --git a/lib/compiler/test/beam_block_SUITE.erl b/lib/compiler/test/beam_block_SUITE.erl
index 4bcb252833..9fcb6e497d 100644
--- a/lib/compiler/test/beam_block_SUITE.erl
+++ b/lib/compiler/test/beam_block_SUITE.erl
@@ -21,7 +21,8 @@
-export([all/0,suite/0,groups/0,init_per_suite/1,end_per_suite/1,
init_per_group/2,end_per_group/2,
- get_map_elements/1,otp_7345/1,move_opt_across_gc_bif/1]).
+ get_map_elements/1,otp_7345/1,move_opt_across_gc_bif/1,
+ erl_202/1]).
%% The only test for the following functions is that
%% the code compiles and is accepted by beam_validator.
@@ -37,7 +38,8 @@ groups() ->
[{p,[parallel],
[get_map_elements,
otp_7345,
- move_opt_across_gc_bif
+ move_opt_across_gc_bif,
+ erl_202
]}].
init_per_suite(Config) ->
@@ -135,6 +137,27 @@ positive(speaking) ->
paris([], P) -> P + 1.
+
+%% See https://bugs.erlang.org/browse/ERL-202.
+%% Test that move_allocates/1 in beam_block doesn't move allocate
+%% when it would not be safe.
+
+-record(erl_202_r1, {y}).
+-record(erl_202_r2, {x}).
+
+erl_202(_Config) ->
+ Ref = make_ref(),
+ Ref = erl_202({{1,2},Ref}, 42),
+
+ {Ref} = erl_202({7,8}, #erl_202_r1{y=#erl_202_r2{x=Ref}}),
+
+ ok.
+
+erl_202({{_, _},X}, _) ->
+ X;
+erl_202({_, _}, #erl_202_r1{y=R2}) ->
+ {R2#erl_202_r2.x}.
+
%%%
%%% The only test of the following code is that it compiles.
%%%