Re: Problems with parallel make on seq-2-1-7

[Benjamin Franksen <benjamin.franksen@helmholtz-berlin.de>]

Hi Mark

I add tech-talk to the list since this is a problem in the EPICS base rules.

On Friday, July 06, 2012, you wrote:
> With the old version it failed about 1 in 10 times at most.
> 
> With the new version there is a more serious problem.  It fails even
> without a parallel make!
> 
> I did a parallel make on the clean and uninstall, but then a non-parallel
> make on the build:
> 
> corvette:~/devel/seq>make -sj uninstall
> 
> corvette:~/devel/seq>make -s
> mkdir ../../../include
> mkdir ../../../lib/linux-x86
> mkdir ../../../lib/vxWorks-68040
> mkdir ../../../bin/vxWorks-68040
> mkdir ../../../lib/vxWorks-ppc603_long
> mkdir ../../../bin/vxWorks-ppc603_long
> mkdir ../../../lib/vxWorks-ppc604_long
> mkdir ../../../bin/vxWorks-ppc604_long
> mkdir ../../../bin/linux-x86
> mkdir ../../../dbd
> ../snl.re:16:17: fatal error: snl.h: No such file or directory
> compilation terminated.
> make[3]: *** [lexer.o] Error 1
> make[2]: *** [install.linux-x86] Error 2
> make[1]: *** [snc.install] Error 2
> make: *** [src.install] Error 2
> corvette:~/devel/seq>
> 
> If I do a parallel make on the build it actually works most if not all of
> the time, but a non-parallel make fails.  This seems strange.

I tried this on another comupter that runs a fedora linux today and I could 
reproduce this. The make version is 3.82 (on my system it is 3.81) and this 
is probably the same version you have.

After the failed make (without -j) search for lexer.d; you won't find it. 
That means there is no dependency of lexer.o on snl.h and thus it does not 
get created before trying to compile lexer.c. If you add snl.h explicitly to 
the prerequisites of lexer.o

lexer.o: snl.h

Then the build works fine. However, this dependency should have been 
generated automatically, there should be a file lexer.d that lists all the 
dependencies of lexer.o on all the header files lexer.c includes.

The parallel build works because the rules that generates snl.c and snl.h is 
executed in parallel and therefore in this case the snl.h (accidentally) 
exists. This is a excerpt of the relevanmt rules in the sequencer's Makefile 
(the fixed version, not the one in 2.1.7):

snl.c: snl.lem snl.lt $(LEMON)
	$(RM) snl.c snl.h
	$(LEMON) o=. $< # this command creates snl.c and snl.h

snl.h: snl.c
	# empty recipy needed to work around make bug #35455

vpath %.re ..

lexer.c: snl.re
	$(RE2C) -s -b -o $@ $<

These rules are fine. What's wrong is teh rules in base that build the 
dependency files (.d). The following is from RULES_BUILD:

%$(OBJ): %.c
	@$(HDEPENDS_CMD)
	@$(RM) $@
	$(COMPILE.c) $(call PATH_FILTER,$<) $(COMPILE_FILTER.c)

%$(OBJ): %.cc
	@$(HDEPENDS_CMD)
	@$(RM) $@
	$(COMPILE.cpp) $(call PATH_FILTER,$<) $(COMPILE_FILTER.cpp)

%$(OBJ): %.cpp
	@$(HDEPENDS_CMD)
	@$(RM) $@
	$(COMPILE.cpp) $(call PATH_FILTER,$<) $(COMPILE_FILTER.cpp)

This means .d files are not generated before the rules to compile a .c file 
fires. Instead they are generated together with the object files. This is 
bad, since it means that the header files must all exist or be explicitly 
added as dependencies.

Instead, there should explicit rules that generate .d from .c:

%$(DEP): %.c
	@$(HDEPENDS_CMD)

and similar for .cc and .cpp. When I add such a rule to the Makefile, 
everything build just fine on the fedora system with make 3.82. The 
$(HDEPENDS_CMD) can then be removed from the object generation rules.

There remains the question what do about this problem? Should I add the 
extra dependency (lexer.o: snl.h) to the Makefile? Or should I add the 
missing rules for .d files?

Cheers
Ben

Attachment: signature.asc
Description: This is a digitally signed message part.