Damaged dependency files

Damaged dependency files

[Alex Peshkoff]

 Hello, all!

When running "make" I get an error:
make[3]: *** No rule to make target
`/usr/home/firebird/firebird/trunk/src/yvalve/gds_proto.h)', needed by
`/usr/home/firebird/firebird/trunk/temp/Release/gpre/boot/gpre_meta_boot.o'. 
Stop.

Let me explain, why do I ask in this list. Error happens due to broken
structure of dependency files, created using -MMD switch of g++.
Normally that file has the following layout:

/usr/home/firebird/firebird/trunk/temp/Release/gpre/c_cxx.o: $(wildcard  \
 /usr/home/firebird/firebird/trunk/src/gpre/c_cxx.cpp \
  /usr/home/firebird/firebird/trunk/src/include/firebird.h \
  /usr/home/firebird/firebird/trunk/src/include/gen/autoconfig.h \
....................
  /usr/home/firebird/firebird/trunk/src/include/ProviderInterface.h \
  /usr/home/firebird/firebird/trunk/src/common/classes/Interface.h)

But for some files I get:

 \
 /usr/home/firebird/firebird/trunk/temp/Release/gpre/boot/gpre_meta_boot.o: 
\
 /usr/home/firebird/firebird/trunk/src/gpre/boot/gpre_meta_boot.cpp \
  /usr/home/firebird/firebird/trunk/src/include/firebird.h \
........................
  /usr/home/firebird/firebird/trunk/src/yvalve/gds_proto.h)

I.e. extra line with single backslash in the beginning and missing
'$(wildcard' make's function after object file name.

My problem happens only for x86 builds with optimization, for x64 or
without optimization I've never seen it. I can reliably reproduce it for
the following versions:
gcc (Debian 4.3.2-1.1) 4.3.2
gcc (GCC) 4.3.3
gcc (GCC) 3.3.6
First one (where bug was noted initially) is Debian binary, two last
were built from clean sources, taken from official gcc repository (base
& g++ tar files were used). I've specially tried rather old gcc version
to check, does it help or not, but bug happens for all compilers.
What is specially surprising - for different gcc versions I get damaged
dependency files for different source files. On the other hand, for
particular compiler it happens always with particular source file and
does not depend upon 'make -j1' or 'make -j4' was used.

All this taken together makes me think, that this is gcc problem. But
may be I miss something obvious?

I can provide any additional information required, but do not want to
overload first letter with probably unneeded details. May be command
line options only:
g++ -ggdb -DFB_SEND_FLAGS=MSG_NOSIGNAL -DLINUX -pipe -MMD -fPIC
-fmessage-length=0 -O3 -march=i586 -mcpu=i686 -fno-omit-frame-pointer
-I/usr/home/firebird/firebird/trunk/src/include/gen
-I/usr/home/firebird/firebird/trunk/src/include -pthread  -pthread -c
/usr/home/firebird/firebird/trunk/src/gpre/boot/gpre_meta_boot.cpp -o
/usr/home/firebird/firebird/trunk/temp/Release/gpre/boot/gpre_meta_boot.o

And one thought - typically this happens for relatively long paths. May
be compiler tries to build better looking ddependency file but has some
old bug in this less tested branch of code?

Alex.

Re: Damaged dependency files

[Segher Boessenkool]

> I.e. extra line with single backslash in the beginning and missing
> '$(wildcard' make's function after object file name.

This isn't caused by GCC, or not directly anyway.  Firebird's
make.rules has:

	$(OBJ)/%.o: $(SRC_ROOT)/%.c
	$(CC) $(WCFLAGS) -c $(firstword $<) -o $@
	@sed -i -e "1s/:/: \$$(wildcard/" -e "\$$s/\(.*\)/\\1)/" $(patsubst
%.o,%.d,$@)

Something in that sed command does not work as expected; you should
look at the .d file before it is changed to see what's up (you can give
the -i flag an argument which it will use to make a backup of the original
file before it edits it; e.g.  -i.orig  will make a bla.d.orig).


Segher

Re: Damaged dependency files

[Alex Peshkoff]

 On 11/29/10 21:48, Segher Boessenkool wrote:
>> I.e. extra line with single backslash in the beginning and missing
>> '$(wildcard' make's function after object file name.
> This isn't caused by GCC, or not directly anyway.  Firebird's
> make.rules has:
>
> 	$(OBJ)/%.o: $(SRC_ROOT)/%.c
> 	$(CC) $(WCFLAGS) -c $(firstword $<) -o $@
> 	@sed -i -e "1s/:/: \$$(wildcard/" -e "\$$s/\(.*\)/\\1)/" $(patsubst
> %.o,%.d,$@)
>
> Something in that sed command does not work as expected; you should
> look at the .d file before it is changed to see what's up (you can give
> the -i flag an argument which it will use to make a backup of the original
> file before it edits it; e.g.  -i.orig  will make a bla.d.orig).
>
Thank you, Segher!