Floating point problems when building for ARM v5

Brian Clarke brian.clarke@oxsemi.com
Wed Oct 13 07:28:00 GMT 2004


I take it nobody has encountered either of the problems I mentioned in 
my original post.

Have have since found solutions that have allowed me to build a working 
toolchain for ARMv5, but I have not discovered a way to make a common 
toolchain for v4 and v5.

Brian Clarke wrote:
> Hello all,
> 
> I am attempting to use a GCC 3.4.2 based ARM toolchain (binutils-2.15, 
> newlib-1.12.0) to build for both v4 and v5 instruction targets.
> 
> When building for a v4 target all is well, the toolchain's libs (libc, 
> libgcc) clain to be built for FPA floating point as do my object files 
> and the code runs fine on the target.
> 
> When I try to build for a v5 target (actually ARM926EJ-S, which has no 
> hardware FP support) the linker complains about floating point 
> conflicts. The toolchain libs are again FPA, but my object files are VFP 
> format.
> 
> I've tried convincing the toolchain to build my libraries for FPA, via 
> experimenting with -msoft-float, -mhard-float and -fp=2/3 but each time 
> I still get VFP libraries. I tried specifying the -mfpu=fpa switch, but 
> this is rejected by the compiler.
> 
> Can anyone suggest a way to get the toolchain to build FPA libraries?
> 
> I've also tried to understand how to build the toolchain so it will 
> build VFP libraries (i.e. for libc, libgcc). I have specified 
> --with-arch=armv5te to the configure scripts of gcc (pt1 and pt2), 
> newlib and binutils. This has resulted in libgcc being built for VFP, 
> but libc is still built for FPA, suggesting I need to provide some other 
> option while configuring newlib. The options I'm currently giving to 
> newlib are:
> 
> configure --target=arm-elf --with-arch=armv5te  --enable-multilib 
> --enable-interwork --prefix=$(PREFIX-PATH) --exec-prefix=$(EXEC-PATH)
> 
> I expected to see multiple versions of libc etc. built due to the 
> --enable-multilib option, but only one appears, with FPA as the floating 
> point type. Can anyone suggest an option to specify to the newlib 
> configure to produce VFP libraries? The newlib config-ml.in file doesn't 
> seem to allow for selecting VFP.
> 
> Thanks,
> Brian.
> 
> 



More information about the Gcc-help mailing list