Hi Masahiro,
On 22 December 2014 at 03:30, Masahiro Yamada yamada.m@jp.panasonic.com wrote:
Hi Simon,
On Tue, 16 Dec 2014 21:44:00 -0700 Simon Glass sjg@chromium.org wrote:
Hi Masahiro,
On 15 December 2014 at 18:47, Masahiro YAMADA yamada.m@jp.panasonic.com wrote:
Simon,
2014-12-02 5:06 GMT+09:00 Simon Glass sjg@chromium.org:
Hi Masahiro,
On 26 November 2014 at 00:45, Masahiro Yamada yamada.m@jp.panasonic.com wrote:
Hi Tom, Simon, other developers,
Since commit 0d296cc2d3b8 (Provide option to avoid defining a custom version of uintptr_t) and commit 4166ecb247a1 (Add some standard headers external code might need), I have been wondering if they were right decisions.
As arch/arm/include/asm/types.h of Linux Kernel says, using 'stdint.h' is not feasible.
-------------------------------------->8------------------------------------- /*
- The C99 types uintXX_t that are usually defined in 'stdint.h' are not as
- unambiguous on ARM as you would expect. For the types below, there is a
- difference on ARM between GCC built for bare metal ARM, GCC built for glibc
- and the kernel itself, which results in build errors if you try to build with
- -ffreestanding and include 'stdint.h' (such as when you include 'arm_neon.h'
- in order to use NEON intrinsics)
- As the typedefs for these types in 'stdint.h' are based on builtin defines
- supplied by GCC, we can tweak these to align with the kernel's idea of those
- types, so 'linux/types.h' and 'stdint.h' can be safely included from the same
- source file (provided that -ffreestanding is used).
int32_t uint32_t uintptr_t- bare metal GCC long unsigned long unsigned int
- glibc GCC int unsigned int unsigned int
- kernel int unsigned int unsigned long
*/ --------------------------------------8<----------------------------------------
To me this doesn't matter. I feel that int32_t should probably be int on 32-bit ARM, but actually so long as it is consistent then it is fine if it is long. In fact so long as these types are defined consistently, everything works.
Gabe Black and you broke the consistency. See my explanation below.
Actually, the kernel never includes <stdint.h> except host programs.
Commit 0d296cc2d3b8 introduced "USE_STDINT", but it causes type conflicts depending on which GCC is used.
With ARM bare metal GCC,
yamada@beagle:~/workspace/u-boot$ make omap3_beagle_defconfig all CROSS_COMPILE=arm-none-eabi- USE_STDINT=1 # # configuration written to .config # # # configuration written to spl/.config # scripts/kconfig/conf --silentoldconfig Kconfig scripts/kconfig/conf --silentoldconfig Kconfig CHK include/config.h GEN include/autoconf.mk GEN include/autoconf.mk.dep GEN spl/include/autoconf.mk CHK include/config/uboot.release CHK include/generated/version_autogenerated.h CHK include/generated/timestamp_autogenerated.h UPD include/generated/timestamp_autogenerated.h CC lib/asm-offsets.s In file included from /opt/arm-2011.03/bin/../lib/gcc/arm-none-eabi/4.5.2/include/stdint.h:5:0, from include/compiler.h:117, from include/image.h:19, from include/common.h:85, from lib/asm-offsets.c:15: /opt/arm-2011.03/bin/../lib/gcc/arm-none-eabi/4.5.2/include/stdint-gcc.h:40:24: error: conflicting types for 'int32_t' include/linux/types.h:99:17: note: previous declaration of 'int32_t' was here /opt/arm-2011.03/bin/../lib/gcc/arm-none-eabi/4.5.2/include/stdint-gcc.h:52:25: error: conflicting types for 'uint32_t' include/linux/types.h:105:17: note: previous declaration of 'uint32_t' was here make[2]: *** [lib/asm-offsets.s] Error 1 make[1]: *** [prepare0] Error 2 make: *** [__build_one_by_one] Error 2
While toolchain is this? I don't see this problem - it seems broken.
Not broken at all. I downloaded it from Mentor Graphics http://www.mentor.com/embedded-software/codesourcery
This is a bare-metal compiler. If you do not understand, you should read arch/arm/include/asm/types.h of Linux Kernel once again.
OK, so perhaps for that compiler you cannot use USE_STDINT=1? What does it define for the conflicting type?
It does define "uint32_t" as "unsigned long".
OK
Until commit 0d296cc2d, U-Boot has used the hard-coded defines like Linux.
Well it still does, only that it also now has the *option* of using stdint.h.
No. It is not "Do not worry, it is optional" things.
You have changed code here ard threr for the optional feature.
Commit aac618a32 (ext4: Use inttypes for printf() string) Commit 19ea4678c (Use int64_t for time types) Commit 6bf672592 (Use uint64_t instead of u64 in put_dec()) Commit c6da9ae8a (Tidy up data sizes and function comment in display_options)
etc.
These are all very small commits and only deal with 64-bit printf()s. It seems a small price to pay for the compatibility benefits of supporting stdint.h. The 32-bit madness is a red herring I think.
But anyway it seems like we can fix this problem so that stdint.h can be included without changing the types, as the kernel apparently does.
In my understaing, we should only use ILP32 and LP64 compilers.
short int long longlong pointerILP32 (32bit system) 16 32 32 64 32 LP64 (64bit Unix-like system) 16 32 64 64 64
Whether it is 32bit or 64bit system, we can hard-code
u32/uint32_t as unsigned int u64/uint64_t as unsigned long long uintptr_t as unsigned long
We do not need to refer to compiler-provided headers.
Moreover, we __should not__ refer to compiler-provided <stdint.h>.
Including <stdint.h> means that we use uint32_t defined by the compiler. It is "unsigned int" on some compilers, and "unsigned long" on some compilers.
I don't seem to have that problem, or at least I have not noticed it with printf().
You are not trying to see the problem. You should try. Go to Linaro page and download the bare-metal toolchain. http://www.linaro.org/downloads/
Fair enough, I do use Linaro gcc-linaro-arm-linux-gnueabihf-4.8-2013 which doesn't seem to be broken in this way. Obviously I have just not tried enough compilers, although I did see the problem you describe on m68k.
We still have the hard-coded uint32_t define in <linux/types.h> This causes the type conflict error I showed above.
OK, we can fix it, but horrible problems still remain.
If we depends on <stdint.h>, we do not know if uint32_t is defined as "unsigned int" or "unsigned long".
To print out 32bit variables, we would always have to use PRId32. Do you want to modify all the printf() printing 32bit variables just to make them unreadable?
Ick, I have not seen this. Can you given an example of where this happens?
Again, you are not trying to see that.
Fine. Check my series: http://patchwork.ozlabs.org/patch/423341/
Checked :-)
OK, we can fix "int32_t" and "uint32_t", but it still seems strange to see that "uint32_t" is defined as "unsigned long", whereas "u32" is defined as "unsigned int".
If so, must we fix "u32", "s32", ... all the fixed-width typedefs ?
I notice including <linux/types.h> in the U-Boot source tree and <stdint.h> provided by your compiler at the same time is a nightmare.
If we lean toward <stdint.h>, we must ban <linux/types.h>, but <stdint.h> is not available all the time. For example, kernel.org tool-chains do not provide <stdint.h>.
Maybe we can drastically re-write <linux/types.h> and friends to resolve the type conflicts, but I do not think we should not drift apart from the kernel because we have borrowed many source files from Linux.
So far I don't see a big problem.
Horrible problem!
I feel that, were stdint.h available earlier, then types.h might not have been written and we would just use stdint.h. Presumably stdint.h has been created to fix these sorts of problems, and in fact for new projects, they would not define their own types.
What is missing is how to pass the variable width nicely to printf()/scanf().
printf("foo =" PRId32 "\n", foo_32); printf("bar =" PRId64 "\n", bar_64);
This is ridiculous. Extremely unreadable.
I want something like this: printf("foo = %32x \n", foo_32); printf("bar =" %64x \n", bar_64);
If this had been provided, Linux and U-Boot might have adopted <stdint.h>, but in fact, PRI* is an awful workaround.
Agreed it's not very readable but so long as it only applies to 64-bit values and only to printf() it doesn't see bad to me. Provided we resolve what you have brought up above.
Check my series.
IMO in time the kernel and U-Boot might move to stdint.h, and having it as an option at the moment helps us understand the issues. It does not break any builds.
I double it. Using <stdint.h> is a misjudge. It will mess up printf() everywhere.
Linux Kernel has guideline how to print each variable type. Documentation/printk-formats.txt
I could be wrong though, time will tell. We should keep an eye on it.
If we keep wrong code in the code base, someone might continue development base on it, which is also wrong.
In order not to lose our time, wrong code should be immediately removed.
How do you explain stdint.h? So many projects use it - it is now part of the C99 standard. Has the world gone mad?
It is trade-off and consistency things.
If you make a decision to use <stdint.h>, it must be consistent everywhere.
<stdint.h> gives int{8,16,32,64}_t and uint{8,16,32,64}_t, uintptr_t etc.
We should always use them for fixed-width variable types and should never use hard-coded ones.
That means we should not use include/linux/types.h and friends to hard-code u8/u16/u32/u64 etc.
We always have to use PRIxN, PRIdN etc. to avoid printf-related warnings. For that, compiler-provided <inttypes.h> must be included.
When you start a new project, you can include <stdint.h> and <inttypes.h> and follow the that rule from the beginning.
You will see horrible things if you try to apply that rule on U-Boot.
I would like to find a solution to this. It does not seem like rocket science. If we accept that stdint.h is useful (I believe it is) then we can make it work perhaps as the kernel does. In other words, we can redefine the types (__INT32_TYPE__ etc.) instead of using the stdint.h ones. The effect is the same because it is not the bit widths that are different - it is just the type names.
I'm not sure if that is what you are suggesting or not. But it seems like it should work, and avoid all the 64-bit PRI defines that you seem very upset about. There are only 26 uses in U-Boot as of now!
Regards, Simon