Re: [dpdk-dev] [PATCH v3 7/8] mk: sort object files when building deps lists

[Thomas Monjalon <thomas@monjalon.net>]

28/06/2017 16:07, Luca Boccassi:
> On Tue, 2017-06-27 at 18:14 +0200, Thomas Monjalon wrote:
> > 27/06/2017 16:51, Luca Boccassi:
> > > On Tue, 2017-06-27 at 15:52 +0200, Thomas Monjalon wrote:
> > > > 27/06/2017 12:43, Luca Boccassi:
> > > > > On Tue, 2017-06-27 at 01:20 +0200, Thomas Monjalon wrote:
> > > > > > 23/06/2017 20:41, lboccass@brocade.com:
> > > > > > > From: Luca Boccassi <luca.boccassi@gmail.com>
> > > > > > > 
> > > > > > > In order to achieve reproducible builds, always use the
> > > > > > > same
> > > > > > > order when listing object files to build dependencies
> > > > > > > lists.
> > > > > > > 
> > > > > > > Signed-off-by: Luca Boccassi <luca.boccassi@gmail.com>
> > > > > > > ---
> > > > > > >  mk/rte.app.mk     | 4 ++--
> > > > > > >  mk/rte.hostapp.mk | 4 ++--
> > > > > > >  mk/rte.shared.mk  | 4 ++--
> > > > > > >  3 files changed, 6 insertions(+), 6 deletions(-)
> > > > > > > 
> > > > > > > --- a/mk/rte.app.mk
> > > > > > > +++ b/mk/rte.app.mk
> > > > > > > @@ -263,8 +263,8 @@ LDLIBS_NAMES += $(patsubst -Wl$(comma)-
> > > > > > > l%,lib%.a,$(filter -Wl$(comma)-l%,$(LDLIB
> > > > > > >  
> > > > > > >  # list of found libraries files (useful for deps). If not
> > > > > > > found,
> > > > > > > the
> > > > > > >  # library is silently ignored and dep won't be checked
> > > > > > > -LDLIBS_FILES := $(wildcard $(foreach dir,$(LDLIBS_PATH),\
> > > > > > > -	$(addprefix $(dir)/,$(LDLIBS_NAMES))))
> > > > > > > +LDLIBS_FILES := $(sort $(wildcard $(foreach
> > > > > > > dir,$(LDLIBS_PATH),\
> > > > > > > +	$(addprefix $(dir)/,$(LDLIBS_NAMES)))))
> > > > > > 
> > > > > > You cannot sort libraries.
> > > > > > Check - for instance - this comment above in this file:
> > > > > > 	# Eliminate duplicates without sorting, only keep the
> > > > > > last
> > > > > > occurrence
> > > > > > 	filter-libs = \
> > > > > 
> > > > > Not sure I follow - what's the reason for avoiding to sort the
> > > > > list
> > > > > of
> > > > > libs to link against?
> > > > 
> > > > Sorry, the ordering issue is with LDLIBS not LDLIBS_FILES.
> > > > 
> > > > > > Why sorting them?
> > > > > > What is random in libraries list?
> > > > > 
> > > > > The issue is that the output of wildcard is not fully
> > > > > deterministic. It
> > > > > can depend on the filesystem, and even on the locale settings
> > > > > [1].
> > > > > Before GNU Make 3.82 (2009) it used to automatically sort the
> > > > > output,
> > > > > but that was removed (to make it faster... sigh). [2]
> > > > 
> > > > It is not a true wildcard here. It is just filtering files which
> > > > do not exist.
> > > > I think you do not need this patch for deterministic build.
> > > 
> > > But then those lists are passed down in the .SECONDEXPANSION rule
> > > right?
> > 
> > I do not follow you.
> > Please explain what is the benefit of the patch in the next version.
> 
> I thought that these lists are used to determine which files to
> recompile - and incidentally, in which order as they are passed in this
> snippet in rte.compile-pre.mk:
> 
> .SECONDEXPANSION:
> %.o: %.c $$(wildcard $$(dep_$$@)) $$(DEP_$$(@)) FORCE
> 	@[ -d $(dir $@) ] || mkdir -p $(dir $@)
> 	$(if $(D),\
> 		@echo -n "$< -> $@ " ; \
> 		echo -n "file_missing=$(call boolean,$(file_missing)) " ; \
> 		echo -n "cmdline_changed=$(call boolean,$(call cmdline_changed,$(C_TO_O))) " ; \
> 		echo -n "depfile_missing=$(call boolean,$(depfile_missing)) " ; \
> 		echo "depfile_newer=$(call boolean,$(depfile_newer))")
> 	$(if $(or \
> 		$(file_missing),\
> 		$(call cmdline_changed,$(C_TO_O)),\
> 		$(depfile_missing),\
> 		$(depfile_newer)),\
> 		$(C_TO_O_DO))
> 
> Did I get that wrong? (It is a bit convoluted :-P )

LDLIBS_FILES are just used to link the executable AFAIK.
The libraries are built in separate recursive make calls.

> But nevertheless, I've finally found the root cause for the "wandering
> header" - when building the libraries, CFLAGS lists
> -I$(RTE_OUTDIR)/include first and then -I$(SRCDIR) last. There is a
> race, and sometimes when GCC is called the public header has already
> been installed in RTE_OUTDIR and sometimes it has not. This causes the
> instability, and causes the expansion of __FILE__ and the directory
> listing in the DWARF objects to randomly list the full path to either
> RTE_OUTDIR/include or SRCDIR for each library.
> 
> By always passing -I$(SRCDIR) first this instability is solved.

Good news.