Big-endian in combined one-pass build

[Jeff Johnston]

On 01/24/2011 06:55 PM, Drasko DRASKOVIC wrote:
> Hi Jeff,
> Thank you very much for this clarification.
>
> I have some additional question at this point :
>
> On Mon, Jan 24, 2011 at 11:10 PM, Jeff Johnston<jjohnstn@redhat.com>  wrote:
>> On 01/24/2011 09:52 AM, Drasko DRASKOVIC wrote:
>>> Questions are :
>>> 1) Why is this happening and how to force newlib to compile in BE mode ?
>>
>> Normally, newlib is built with a target and using multilib.  Gcc then
>> reports the permutations to build newlib with.  When you don't use multilib,
>> then newlib gets built with the default options and whatever options you
>> override.  If the default for arm-elf is to be LE, then that's what you end
>> up with.
> I have not found how to pass GCC options to be BE. So I used armeb-elf
> target, which seems only way to force GCC to BE.
> However, I noticed that newlib has no support for armeb-elf atrget. is
> this correct ?
>
> There is some patch I found to force armeb-elf target in newlib :
> http://sourceware.org/ml/newlib/2008/msg00698.html.
> However, toolchain produced in that way (multilib disabled), although
> capable to link well my software, it is creating binaries that do not
> function well.
> Is this patch incorrect ?
>

I stand corrected, there was a request.  I didn't comment on it so it 
got missed and the sender doesn't appear to have followed up on it.  It
appears to be missing some other changes in the file which look for 
arm-* and won't trigger on armeb-*.  You might have fixed that on your 
own, but if not, be aware.

In theory, the armeb compiler should compile with BE by default.  It 
should set compiler flags that can be used to figure out the float 
endian-ness.  See libc/include/machine/ieeefp.h for arm logic.  If those 
flags aren't being set or the ieeefp.h isn't interpreting correctly 
under such a scenario, it might get the floating-point format wrong 
which may be the cause of errors you are seeing.

If the fixes are made and the armeb-elf configuration isn't creating 
valid executables, you're going to have to do some debug investigation. 
  Do you get through crt0 ok?  Can you run a simple empty program?  If 
not, where are you failing?  If yes, try other simple programs like 
HelloWorld.  Check constants and make sure they are in BE format.

>>
>> No one has ever requested or added armeb-elf as a newlib target.
> That is strange, considering that Crosstool disables multilib and
> forces armeb-elf target when BE is selected...
> So, basically this target does not exist and mentioned patch is risky ?
>
>>
>>> 2) Can --with-endian=big flag work even if I disabled multilib (i.e.
>>> is multilib necessary for BE mode build ?)
>>
>> By your test results, the with-endian flag is not being used to set
>> compilation flags passed to newlib.  Newlib doesn't interpret or do anything
>> with the --with-endian setting.
>
> Yes, it seems like that. With multilib enabled I forced
> --enable-biendian, which produced BE libs. They compile binaries that
> seems to be fine. However patched newlib with armeb-elf target
> produced bad binaries.
>
>>
>>> 3) Do I always have to use armeb-elf instead of arm-elf target when
>>> multilib is disabled
>>
>> If you don't override the compilation flags manually, then yes.
>
> OK, that would mean calling "make" with TARGET_CFLAGS variable set
> correctly, right ? Would that be sufficient, or something else should
> be changed in the code before ?
>

Depends on the questions above regarding debugging.

>>
>>> 4) Do I have to use make CFLAGS_FOR_TARGET="-mbig-endian -msoft-float"
>>> CCASFLAGS="-mbig-endian -msoft-float" when calling GCC (with newlib
>>> combined, one-pass) build to force Newlib to compile in big-endian
>>> mode ?
>>>
>>
>> Actually, for newlib, use TARGET_CFLAGS="-mbig-endian -msoft-float". You
>> might require the CFLAGS_FOR_TARGET to ensure that other target libs are
>> built with the same options (e.g. libgcc).
>
> My question here is : would TARGET_CFLAGS="-mbig-endian -msoft-float"
> to affect C++ part also (i.e. would this produce BE C++ libs) ?
>
> Can you please clarify these variables in a little more detail, as it
> seems to be crucial :
> 1) What would be difference between TARGET_CFLAGS and
> CFLAGS_FOR_TARGET, and what is the role of the CCASFLAGS ?

TARGET_CFLAGS is a flag used by newlib and is used both for compiling C 
files and assembly files.  CCASFLAGS are flags used to compile .S files 
which are assembly files with C preprocessor statements (e.g. #ifdef 
HAVE_xxxx).  Newlib uses .S files for stuff such as setjmp/longjmp, 
crt0.  This allows a platform to change any C file over to assembly if 
desired and to react to various newlib_cflags used in configure.host in 
a reasonable manner (e.g. HAVE_xxxx tells the code it can make a call to 
xxxx).

The TARGET_CFLAGS flag is used both for compiling newlib C files and is 
used to set CCASFLAGS.  You don't have to set CCASFLAGS in addition.

CFLAGS_FOR_TARGET is set from the top-level and passed to components 
that are for the target (e.g. target libraries).  It works for newlib 
since newlib is a target library and it will be the CFLAGS variable 
passed to newlib at its top-level.  You might want to add in the initial 
default CFLAGS_FOR_TARGET settings if you override (namely "-g -O2") 
because you will stomp over these.

So, if you want to pass flags just to newlib, use TARGET_CFLAGS.  If you 
want to pass flags to all target libraries, use CFLAGS_FOR_TARGET.

If the armeb-elf compiler should by default be setting BE and soft-float 
options, then you shouldn't be passing any flags - it should just work. 
  You may have to do save the build log and do some assembly debugging 
to figure out what is going wrong.

> 2) Both GCC with C++ supptor and newlib will be compiled in combined
> build calling "make" one time, and re-defining these variables. So who
> will use these, GCC or Newlib, or both ? I am asking thisbecause I had
> compilation breaking when issuing :
> make CFLAGS_FOR_TARGET="-mbig-endian -msoft-float"
> CCASFLAGS="-mbig-endian -msoft-float"
>

See above.  You don't need to set CCASFLAGS.

> What would be correct "make" invocation then, having in mind that I'd
> like to disable multilib and use toolchain to compile and link code
> for big-endian ARM946ES ?
>
>
> Thank you very much again for your time.
>
> Kind regards,
> Drasko