Parallel build failure due to automakes BUILT_SOURCES and subdir-objects

[ncopa]

I get build failure on my 12 core (with HT so 24 "cpus") build server with -j24:

...
  GEN      t/test-full.pb.cc
  GEN      t/test-full.pb.h
  CXX      protoc-c/protoc_c_protoc_c-c_bytes_field.o
  CXX      protoc-c/protoc_c_protoc_c-c_enum.o
  CXX      protoc-c/protoc_c_protoc_c-c_enum_field.o
  CXX      protoc-c/protoc_c_protoc_c-c_extension.o
  CXX      protoc-c/protoc_c_protoc_c-c_field.o
  CXX      protoc-c/protoc_c_protoc_c-c_file.o
  CXX      protoc-c/protoc_c_protoc_c-c_generator.o
  CXX      protoc-c/protoc_c_protoc_c-c_helpers.o
  CXX      protoc-c/protoc_c_protoc_c-c_message.o
  CXX      protoc-c/protoc_c_protoc_c-c_message_field.o
  CXX      protoc-c/protoc_c_protoc_c-c_primitive_field.o
  CXX      protoc-c/protoc_c_protoc_c-c_service.o
  CXX      protoc-c/protoc_c_protoc_c-c_string_field.o
  CXX      protoc-c/protoc_c_protoc_c-main.o
  CXX      t/generated-code2/t_generated_code2_cxx_generate_packed_data-cxx-generate-packed-data.o
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_B" uses the same enum value as "foo.VALUE_A". If this is intended, set 'option allow_alias = true;' to the enum definition.
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_C" uses the same enum value as "foo.VALUE_A". If this is intended, set 'option allow_alias = true;' to the enum definition.
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_E" uses the same enum value as "foo.VALUE_D". If this is intended, set 'option allow_alias = true;' to the enum definition.
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_AA" uses the same enum value as "foo.VALUE_F". If this is intended, set 'option allow_alias = true;' to the enum definition.
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_B" uses the same enum value as "foo.VALUE_A". If this is intended, set 'option allow_alias = true;' to the enum definition.
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_C" uses the same enum value as "foo.VALUE_A". If this is intended, set 'option allow_alias = true;' to the enum definition.
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_E" uses the same enum value as "foo.VALUE_D". If this is intended, set 'option allow_alias = true;' to the enum definition.
[libprotobuf ERROR google/protobuf/descriptor.cc:4153] "foo.VALUE_AA" uses the same enum value as "foo.VALUE_F". If this is intended, set 'option allow_alias = true;' to the enum definition.
t/generated-code2/cxx-generate-packed-data.cc:5:28: fatal error: t/test-full.pb.h: No such file or directory
 #include "t/test-full.pb.h"
                            ^
compilation terminated.
  CXX      t/t_generated_code2_cxx_generate_packed_data-test-full.pb.o
Preprocessed source stored into /tmp/ccpjigCB.out file, please attach this to your bugreport.
Makefile:1351: recipe for target 't/generated-code2/t_generated_code2_cxx_generate_packed_data-cxx-generate-packed-data.o' failed
make: *** [t/generated-code2/t_generated_code2_cxx_generate_packed_data-cxx-generate-packed-data.o] Error 1
make: *** Waiting for unfinished jobs....

I dont know if this is an automake bug or "feature". It seems like this row in Makefile.am is supposed to make sure that the generated t/test-full.pb.h is built first:

t/generated-code2/cxx-generate-packed-data.$(OBJEXT): t/test-full.pb.h

but it seems like automake adds some extra prefix due to building things in subdirs (and enabling 'subdir-objects' in with AM_INIT_AUTOMAKE) so there should probably be another line with:

--- ./Makefile.am.orig
+++ ./Makefile.am
@@ -142,6 +142,7 @@
        t/generated-code2/cxx-generate-packed-data.cc \
        t/test-full.pb.cc
 t/generated-code2/cxx-generate-packed-data.$(OBJEXT): t/test-full.pb.h
+t/generated-code2/t_generated_code2_cxx_generate_packed_data-cxx-generate-packed-data.$(OBJEXT): t/test-full.pb.h
 t_generated_code2_cxx_generate_packed_data_CXXFLAGS = \
        $(AM_CXXFLAGS) \
        $(protobuf_CFLAGS)

[lipnitsk]

@ncopa, can you submit a Pull Request so we can review and merge in your fix? Do you think it would be possible to test this issue in the Travis CI environment?

[edmonds]

This is a bug that got overlooked in the conversion to subdir-objects. I'm guessing something like this is the right way to express this:

diff --git a/Makefile.am b/Makefile.am
index 3ec4b6e..c5dd657 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -141,7 +141,7 @@ noinst_PROGRAMS += \
 t_generated_code2_cxx_generate_packed_data_SOURCES = \
        t/generated-code2/cxx-generate-packed-data.cc \
        t/test-full.pb.cc
-t/generated-code2/cxx-generate-packed-data.$(OBJEXT): t/test-full.pb.h
+EXTRA_t_generated_code_test_generated_code_DEPENDENCIES = t/test-full.pb.h
 t_generated_code2_cxx_generate_packed_data_CXXFLAGS = \
        $(AM_CXXFLAGS) \
        $(protobuf_CFLAGS)

I've never run into the reporter's problem on my desktop system, but I only have four CPU cores, so...

I'm not sure we can test for this in travis-ci, though.

[edmonds]

Actually, wait, listing t/test-full.pb.h in BUILT_SOURCES is supposed to already take care of this, but "BUILT_SOURCES is honored only by ‘make all’, ‘make check’ and ‘make install’" (per http://www.gnu.org/software/automake/manual/html_node/Sources.html).

@ncopa Can you confirm what make command you used to build the source? You trimmed this information from your initial report.

[ncopa]

I doubt this will be reproduceable in travis. looks like they give you 1.5 virutal cpus...

the make command I used is something like export MAKEFLAGS=-j25; make

It also happens with make all. I think that is becuase, it does a pre-build to generate all the BUILT_SOURCES before it starts the 'real' built. But this prebuild is also run in parallel, so the deps of the prebuild needs be exact.

-t/generated-code2/cxx-generate-packed-data.$(OBJEXT): t/test-full.pb.h
+EXTRA_t_generated_code_test_generated_code_DEPENDENCIES = t/test-full.pb.h

This did not solve it. I think you maybe meant:

t_generated_code2_cxx_generate_packed_data_DEPENDENCIES = t/test-full.pb.h

But that is also wrong because it adds the needed header as a link time dependency of the final executable. What we need is adding t/test-full.pb.h as a dependency of the object generated from t/generated-code2/cxx-generate-packed-data.cc, not a dep fo the final generated executable.

I don't think automake gives any proper way to specify dependencies of the _SOURCES.

But what we could do is add this line:

$(t_generated_code2_cxx_generate_packed_data_OBJECTS): t/test-full.pb.h

That will the t/test-full.pb.h as a dependency for all the code2_cxx_generate_packed_data objects, including the cxx-generate-packed-data.$(OBJEXT) and the test-full.pb.h should always get generated first.

I tested it, it works - and I think that is better than my originally proposed patch since this is not as dependant on the automake internals.

[edmonds]

Fix commited to next.

[juhao5208]

hi，but it's not slove my problem,it is not work,i don't konw why.