TODO: Remove a couple of build system issues that I think have been fixed.
This commit is contained in:
Gregory Nutt
parent
f81ef55834
commit
425f769f0f
105
TODO
105
TODO
@ -25,7 +25,7 @@ nuttx/:
|
||||
(12) Libraries (libc/, libm/)
|
||||
(10) File system/Generic drivers (fs/, drivers/)
|
||||
(9) Graphics Subsystem (graphics/)
|
||||
(3) Build system / Toolchains
|
||||
(1) Build system / Toolchains
|
||||
(3) Linux/Cywgin simulation (arch/sim)
|
||||
(4) ARM (arch/arm/)
|
||||
|
||||
@ -2027,109 +2027,6 @@ o Build system
|
||||
Status: Open
|
||||
Priority: Low.
|
||||
|
||||
Title: CONTROL-C CAN BREAK DEPENDENCIES
|
||||
Description: If you control C out of a make, then there are things that can go
|
||||
wrong. For one, you can break the dependencies in this scenario:
|
||||
|
||||
- The build in a given directory begins with all of the compilations.
|
||||
On terminal, this the long phase with CC: on each line. As each
|
||||
.o file is created, it is timestamped with the current time.
|
||||
|
||||
- The dependencies on each .o are such that the C file will be re-
|
||||
compile if the .o file is OLDER that the corresponding .a archive
|
||||
file.
|
||||
|
||||
- The compilation phase is followed by a single, relatively short
|
||||
AR: phase that adds each of the file to the .a archive file. As
|
||||
each file is added to archive, the timestamp of the of archive is
|
||||
updated to the current time. After the first .o file has been
|
||||
added, then archive file will have a newer timestamp than any of
|
||||
the newly compiled .o file.
|
||||
|
||||
- If the user aborts with control-C during this AR: phase, then we
|
||||
are left with: (1) not all of the files have bee added to the
|
||||
archive, and (2) the archive file has a newer timestamp than any
|
||||
of the .o file.
|
||||
|
||||
So when the make is restarted after a control, the dependencies will
|
||||
see that the .a archive file has the newer time stamp and those .o
|
||||
file will never be added to the archive until the directory is cleaned
|
||||
or some other dependency changes.
|
||||
|
||||
NOTE: This may not really be an issue because the the timestamp on
|
||||
libapps.a is not really used but rather the timestamp on an empty
|
||||
file:
|
||||
|
||||
.built: $(OBJS)
|
||||
$(call ARCHIVE, $(BIN), $(OBJS))
|
||||
$(Q) touch $@
|
||||
|
||||
UPDATE: But there is another way that Control-C can break dependencies:
|
||||
If you control-c out of the make during the apps/ part of the build,
|
||||
the archive at apps/libapps.a is deleted (but all of the .built files
|
||||
remain in place). You can see this in the make outout, for example:
|
||||
|
||||
CC: ieee802154_getsaddr.c
|
||||
make[2]: *** [Makefile:104: ieee802154_getsaddr.o] Interrupt
|
||||
make: *** Deleting file '../apps/libapps.a'
|
||||
|
||||
When you rebuild the system, the first file archived will recreate
|
||||
libapps.a and set the timestamp to the current time. Then, none of
|
||||
the other object files will be added to the archive because they are
|
||||
all older.. or, more correctly, none of the other object files will
|
||||
be addred because .built files remained and say that there is no
|
||||
need to update the libapps.a file.
|
||||
|
||||
The typical symptom of such an issue is a link time error like:
|
||||
|
||||
LD: nuttx libsched.a(os_bringup.o): In function `os_bringup':
|
||||
os_bringup.c:(.text+0x34): undefined reference to `nsh_main'
|
||||
|
||||
This is becuase the libapps.a file was deleted and an new empty
|
||||
libapps.a file was created (which the object containing nsh_main()).
|
||||
The object containing nsh_main() will not be added because the
|
||||
.built file exists and says that there is not need to add the
|
||||
nsh_main() object to libapps.a.
|
||||
|
||||
The work-around for now is:
|
||||
|
||||
$ make apps_distclean
|
||||
|
||||
One solution to this might be to making the special target
|
||||
.PRECIOUS depend on apps/libapps.a. Then if make receives a
|
||||
signal, it will not delete apps/libapps.a. This would have to
|
||||
be done in all Makefiles.
|
||||
|
||||
Status Open
|
||||
Priority: Medium-High. It is a rare event that control-C happens at just the
|
||||
point in time. However, when it does occur the resulting code may
|
||||
have binary incompatiblies in the code taken from the out-of-sync
|
||||
archives and cost a lot of debug time before you realize the issue.
|
||||
|
||||
The first stated problem is not really an issue: There is already
|
||||
the spurious .built file that should handle the described case:
|
||||
If you control-C out of the build then the timestamp on the .built
|
||||
file will not be updated and the archiving should be okay on the
|
||||
next build.
|
||||
|
||||
A work-around for the second stated problem is to do 'make clean'
|
||||
if you ever decide to control-C out of a make and see that the
|
||||
libapps.a file was deleted.
|
||||
|
||||
UPDATE: This is a potential fix for the second problem in place
|
||||
in in all Makefiles under apps/. This fix adds
|
||||
|
||||
.PRECIOUS: $(BIN)
|
||||
|
||||
to all Makefiles. It has not yet been confirmed that this fix
|
||||
eliminates the dependency issue or not.
|
||||
|
||||
Title: DEPENDENCIES OBJECT SUB-DIRECTORIES
|
||||
Descripton: Dependencies do not work in directories that keep binaries in
|
||||
a sub-directory like bin, ubin, kbin.
|
||||
Status: Open
|
||||
Priority: Medium-Low. Definitely a build issue once in awhile.
|
||||
|
||||
o Other drivers (drivers/)
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user