aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorBjörn Gustavsson <[email protected]>2010-10-12 10:08:18 +0200
committerBjörn Gustavsson <[email protected]>2010-10-12 14:15:36 +0200
commitf6a2fb81fd87e671ceed85261a1c28a6082f010c (patch)
treefaf8abfab935ac33adcb3fece584fd6d1017a341
parent499082d25170aa7edf922a6aee0465a7ffb144d4 (diff)
downloadotp-f6a2fb81fd87e671ceed85261a1c28a6082f010c.tar.gz
otp-f6a2fb81fd87e671ceed85261a1c28a6082f010c.tar.bz2
otp-f6a2fb81fd87e671ceed85261a1c28a6082f010c.zip
beam_block: Do optimizations in the safe order
Moving of allocation instructions upwards in the instruction stream (in order to enable further optimizations) in beam_block, is implemented with the assumption that if a register {x,X} contains a valid term, then all other x register with lower numbers than X also contain valid terms. That assumption is true after code generation. The beam_utils:live_opt/1 optimization, however, may invalidate that assumption. For instance, if a receive statement exports a variable that is used, but the return value of the receive statement is not used, then {x,1} but not {x,0} contains a valid term at the end of the receive statement. If the receive statement is followed by {bif,self,{f,0},[],{x,0}}. {test_heap,NumberOfWords,2}. moving the allocation upwards will produce {test_heap,NumberOfWords,2}. {bif,self,{f,0},[],{x,0}}. which will cause the beam_validator pass to scream loudly that {x,0} is not live at the test_heap instruction. Fix the problem by doing the optimizations in reverse order. Reported-by: Jim Engquist
-rw-r--r--lib/compiler/src/beam_block.erl35
-rw-r--r--lib/compiler/test/receive_SUITE.erl25
2 files changed, 44 insertions, 16 deletions
diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl
index 9c6f835ab0..c45874597a 100644
--- a/lib/compiler/src/beam_block.erl
+++ b/lib/compiler/src/beam_block.erl
@@ -36,12 +36,13 @@ function({function,Name,Arity,CLabel,Is0}, Lc0) ->
%% Collect basic blocks and optimize them.
Is2 = blockify(Is1),
- Is3 = beam_utils:live_opt(Is2),
- Is4 = opt_blocks(Is3),
- Is5 = beam_utils:delete_live_annos(Is4),
+ Is3 = move_allocates(Is2),
+ Is4 = beam_utils:live_opt(Is3),
+ Is5 = opt_blocks(Is4),
+ Is6 = beam_utils:delete_live_annos(Is5),
%% Optimize bit syntax.
- {Is,Lc} = bsm_opt(Is5, Lc0),
+ {Is,Lc} = bsm_opt(Is6, Lc0),
%% Done.
{{function,Name,Arity,CLabel,Is},Lc}
@@ -156,11 +157,7 @@ opt_blocks([I|Is]) ->
opt_blocks([]) -> [].
opt_block(Is0) ->
- %% We explicitly move any allocate instruction upwards before optimising
- %% moves, to avoid any potential problems with the calculation of live
- %% registers.
- Is1 = move_allocates(Is0),
- Is = find_fixpoint(fun opt/1, Is1),
+ Is = find_fixpoint(fun opt/1, Is0),
opt_alloc(Is).
find_fixpoint(OptFun, Is0) ->
@@ -170,11 +167,21 @@ find_fixpoint(OptFun, Is0) ->
end.
%% move_allocates(Is0) -> Is
-%% Move allocates upwards in the instruction stream, in the hope of
-%% getting more possibilities for optimizing away moves later.
-
-move_allocates(Is) ->
- move_allocates_1(reverse(Is), []).
+%% Move allocate instructions upwards in the instruction stream, 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.
+
+move_allocates([{block,Bl0}|Is]) ->
+ Bl = move_allocates_1(reverse(Bl0), []),
+ [{block,Bl}|move_allocates(Is)];
+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),
diff --git a/lib/compiler/test/receive_SUITE.erl b/lib/compiler/test/receive_SUITE.erl
index fca3f0387b..2a592dd669 100644
--- a/lib/compiler/test/receive_SUITE.erl
+++ b/lib/compiler/test/receive_SUITE.erl
@@ -21,7 +21,7 @@
-module(receive_SUITE).
-export([all/1,init_per_testcase/2,fin_per_testcase/2,
- recv/1,coverage/1,otp_7980/1,ref_opt/1]).
+ recv/1,coverage/1,otp_7980/1,ref_opt/1,export/1]).
-include("test_server.hrl").
@@ -36,7 +36,7 @@ fin_per_testcase(_Case, Config) ->
all(suite) ->
test_lib:recompile(?MODULE),
- [recv,coverage,otp_7980,ref_opt].
+ [recv,coverage,otp_7980,ref_opt,export].
-record(state, {ena = true}).
@@ -205,4 +205,25 @@ collect_recv_opt_instrs(Code) ->
end] || {function,_,_,_,Is} <- Code],
lists:append(L).
+export(Config) when is_list(Config) ->
+ Ref = make_ref(),
+ ?line self() ! {result,Ref,42},
+ ?line 42 = export_1(Ref),
+ ?line {error,timeout} = export_1(Ref),
+ ok.
+
+export_1(Reference) ->
+ id(Reference),
+ receive
+ {result,Reference,Result} ->
+ Result
+ after 1 ->
+ Result = {error,timeout}
+ end,
+ %% Result ({x,1}) is used, but not the return value ({x,0})
+ %% of the receive. Used to be incorrectly optimized
+ %% by beam_block.
+ id({build,self()}),
+ Result.
+
id(I) -> I.