
Hello Masahiro,
On 09/30/2013 10:01 AM, Masahiro Yamada wrote:
Hello Jeroen,
diff --git a/config.mk b/config.mk index 48913f6..177f685 100644 --- a/config.mk +++ b/config.mk @@ -117,7 +117,7 @@ CC_TEST_OFILE := $(OBJTREE)/include/generated/cc_test_file.o -include $(CC_OPTIONS_CACHE_FILE)
cc-option-sys = $(shell mkdir -p $(dir $(CC_TEST_OFILE)); \
if $(CC) $(CFLAGS) $(1) -S -xc /dev/null -o
$(CC_TEST_OFILE) \
if $(CC) -Werror $(CFLAGS) $(1) -S -xc /dev/null -o
$(CC_TEST_OFILE) \ > /dev/null 2>&1; then \ echo 'CC_OPTIONS += $(strip $1)' >> $(CC_OPTIONS_CACHE_FILE); \ echo "$(1)"; fi)
It looks like this was already suggested by Tom too. See http://patchwork.ozlabs.org/patch/183174/
I tested this patch but unfortunatelly it did not work.
I downloaded bfin-linux-gcc 4.6.3 from ftp://ftp.kernel.org/pub/tools/crosstool/index.html
I added -Werror to config.mk and I tried:
CROSS_COMPILE=bfin-linux- ./MAKEALL -a blackfin
The log message was still sprinkled with lots of warnings like warning: -fstack-usage not supported for this target [enabled by default]
So, I looked into it more closely and I found gcc can compile the input file /dev/null successfully even if -fstack-usage is not supported.
$ bfin-uclinux-gcc -fstack-usage -S -xc /dev/null $ echo $? 0 $ bfin-uclinux-gcc -Werror -fstack-usage -S -xc /dev/null $ echo $? 0
Instead of /dev/null, I prepared a very simple C source code as an input.
$ cat test.c int test(void) { return 0; } $ bfin-uclinux-gcc -fstack-usage -S -xc test.c test.c: In function 'test': test.c:4:1: warning: -fstack-usage not supported for this target [enabled by default] $ echo $? 0 $ bfin-uclinux-gcc -fstack-usage -Werror -S -xc test.c test.c: In function 'test': test.c:4:1: error: -fstack-usage not supported for this target [-Werror] cc1: all warnings being treated as errors $ echo $? 1
This time we can detect unsupported -fstack-usage by the exit status.
So, along with -Werror, I think /dev/null should also be replaced with some _REAL_ C source file. But my concern about this treak is we might go far from Kbuild. In Linux Kernel, cc-option is impilemented in scripts/Kbuild.include. It uses /dev/null as input. I wish U-Boot would someday adopt (not nessarily the same but very similar) Kbuild.
If you see Linux Kernel top Makefile, many of optional CFLAGS are provided thru cc-option and also can be turned on/off by CONFIG_ switch.
That is one reason I suggested to introduce CONFIG_CC_STACKUSAGE option. But this is just my opinion and I would not necessarily stick to this idea. Please feel free to suggest your thought.
I tried to google this a bit, one thing I found was http://sourceware.org/bugzilla/show_bug.cgi?id=5210, but I fail to (easily) find a reliable gcc reference. Perhaps we should file a bug? -Werror doesn't work with /dev/null when -fstack-usage is not supported...
I cannot comment on your concern regarding Kbuild, since I am unaware of its internals. I just suggested above since it would be a cleaner approach (if it worked...).
Regards, Jeroen