Help with hard versus soft float on PPC

[Till Straumann]

At least for classic PPC (I'm not familiar with the SPE ABI) the
hard- and soft-float ABIs are *not* binary compatible.

Luckily later gcc toolchains set and check object file
attributes to detect such situations.

E.g., when using -hard-float the classic PPC ABI requires
the compiler to set a CRC bit-6 when floating-point
arguments are (or clear the bit if this is not the case) passed in
registers to functions with variable arguments.

Code compiled with -soft-float never uses FP registers and thus
does not maintain the bit in question.

Hence, if a function compiled with -msoft-float calls a
vararg function compiled with -mhard-float then the
state of CRC[6] is undefined and may be wrong.
This then may crash a FP-only task because the
-mhard-float compiled function may find CRC[6]==1
and try to save some FP registers on the stack...

Another incompatibility is that -hard-float compiled
code passes FP args in registers whereas -soft-float
compiled code obviously does not. Hence, passing of
FP args is not possible between code with this ABI
incompatibility.

Again, that all was about classic PPC but I strongly
recommend to take these warnings seriously and
not try to mix incompatible compiler flags.

HTH
-- Till

Peter Dufault wrote:
> I've been building my application for my new MPC5554 BSP using the MPC5554 hard-float (single precision SPE hard-float, I added the context switch hooks) while the BSP has been built using soft float.  This has been working OK for about a month.  I tried to rebuild the BSP with the same set of flags as my application, but now I get warnings, though I didn't before even though there was a mismatch then too.
>
> /opt/rtems-4.10/lib/gcc/powerpc-rtems4.10/4.4.2/../../../../powerpc-rtems4.10/bin/ld: Warning: /home/dufault/work/scratch/Andrea_Doria/powerpcmpc5566evbrtems4.10/lib/libmotobank.a(servo.o) uses hard float, o-optimize/motobank.exe uses soft float
>
> All the application and bsp .o files are now built using:
>     -meabi \
>     -mfloat-gprs=single \
>     -mspe
>
> and the final link looks like this:
>
> powerpc-rtems4.10-g++ \
>     -B/opt/rtems-4.10/powerpc-rtems4.10/mpc5566evb/lib/ \
>     -specs bsp_specs \
>     -qrtems \
>     -g \
>     -Wall \
>     -mcpu=8540 \
>     -meabi \
>     -msdata \
>     -fno-common \
>     -mfloat-gprs=single \
>     -mspe \
>     -D__ppc_generic \
>     -mstrict-align \
>     -o o-optimize/motobank.exe \
>     o-optimize/_init.o \
>     o-optimize/flash.o \
>     -L/home/dufault/work/scratch/Andrea_Doria/powerpcmpc5566evbrtems4.10/lib \
>     -lmotobank \
>     -letpu \
>     -ltelnetd \
>     -lnfs
>
> I haven't done anything about rebuilding newlib.
>
> What's the right way to go about properly building for -float-gprs=single -mspe?  
>
> Peter
> _______________________________________________
> rtems-users mailing list
> rtems-users at rtems.org
> http://www.rtems.org/mailman/listinfo/rtems-users
>