I certainly wouldn't object to such a patch, but I'm not sure who it would actually help. ;-)
As you may already know, I 'make sparse >sp-out 2>&1' on the master branch, use vim to check for warnings/errors, and then diff the corresponding files for 'next' and 'seen' branches (nsp-out, ssp-out). Similarly, I 'make -k hdr-check >hcout 2>&1' on the master branch, use vim ... etc. Note, however, the use of '-k' on the make invocation to make sure I catch all warnings/errors, since '-Werror' is given with the DEVELOPER variable set. So, I would have to do the same with the 'make sparse' invocation after this patch [1].
Yes, anybody who just does 'make sparse' will notice the failure, so that would be a definite improvement. (How many people run 'make sparse' though?). It probably would help anybody using 'cgcc' to develop, but this doesn't quite work (I just tried this for the first time in ages and was surprised to find it almost works for me, YMMV):
$ git branch -v | grep '^\*'
* seen 4141ae2199 Merge branch 'dd/upload-pack-stateless-eof' into seen
$ make clean >/dev/null
$
$ make CC=cgcc >sout2 2>&1
$ ./git version
git version 2.29.2.389.g4141ae2199
$ git describe
v2.29.2-389-g4141ae2199
$
$ grep 'warn' sout2
pack-revindex.c:66:23: warning: memset with byte count of 262144
upload-pack.c:1114:86: warning: Using plain integer as NULL pointer
$
So, this would have failed, because some extra flags ('SP_EXTRA_FLAGS' to be precise) are not being passed to sparse. We can see that:
$ grep SP_EXTRA_FLAGS Makefile
SP_EXTRA_FLAGS = -Wno-universal-initializer
http-push.sp http.sp http-walker.sp remote-curl.sp imap-send.sp: SP_EXTRA_FLAGS += \
pack-revindex.sp: SP_EXTRA_FLAGS += -Wno-memcpy-max-count
compat/nedmalloc/nedmalloc.sp: SP_EXTRA_FLAGS += -Wno-non-pointer-null
$(SPARSE_FLAGS) $(SP_EXTRA_FLAGS) $<
$
(Note that you can't see -DCURL_DISABLE_TYPECHECK being set for all the libcurl using files). So, either sparse has improved to the point that the 'typecheck' shenanigans used in the libcurl headers don't cause it problems anymore, or the headers have changed sufficiently. (Hmm, that header doesn't seem to have changed much).
I don't build with nedmalloc, but last time I did it was littered with uses of '0' as a NULL pointer (hence the -Wno-non-pointer-null). Given that we will probably not re-import that code, I suppose all of those sparse warnings could be fixed.
Which leaves 'pack-revindex.c'. Unfortunately, I don't see an easy fix for that (-Wno-universal-initializer is now the default, so we don't need that anymore). Well, I suppose anybody could add 'CFLAGS+=-Wno-\ memcpy-max-count' to their config.mak file (this '-W' argument is filtered out by cgcc and not passed to gcc). [EDIT: I just tried this and, yes, this does work! I'm still surprised about curl headers ;-) ]
Note also that my build slowed down by 14% using CC=cgcc, since it is running both sparse and gcc on each source file.
I guess it would be most useful on a CI build, but I don't know what would be involved in setting that up.
ATB, Ramsay Jones
[1] It took about one month for me to get used to the 'pu'->'seen' change,
but that was mostly training my fingers to type all of the output
files that are keyed by the branch name: pout->sout, psp-out->ssp-out,
psc->ssc, phcout->shcout and ptest-out->stest-out. :D
Ramsay Jones
