Re: SoftFloat Benchmarks

[Eric BENARD / Free <ebenard@xxxxxxx>]

Le Lundi 8 Mars 2004 23:14, Leon Woestenberg a écrit :
> I am however, also interested in learning how you built your toolchain
> with Nico's patches. Are you using a cross compilation tool or was this
> hand-crafted?
>
Hand-crafted in order to have a good control on the build process (I had to 
give several try for some stages). I'm using patches from crosstool's patches 
directory and patches from mailing lists.
I can post some details if necessary as I wrote the process on a paper.

> Did you also make sure the "printf("%f", 1.0)" issue was solved in your
> toolchain as discussed here?
>
> http://sources.redhat.com/ml/crossgcc/2004-02/msg00106.html
>
Yes. In fact I'm using these patches :
http://lists.arm.linux.org.uk/pipermail/linux-arm/2003-November/006590.html

Which are giving good results with several test programs (I'll post tests 
results later today as these results are also interesting).

Patches I'm using :
- binutils-2.15.90.0.1.1 :
	arm-be.patch
- GCC-3.3.3 :
	gcc-3.3.2-arm-softfloat.diff  gcc-3.3.2-arm-targetbug.diff  t-linux
- Glibc-2.3.2 :
	arm-asm-clobber.patch     sscanf.patch
	string2-typedef.patch
	errlist-awk.patch         sysdep-cancel-arm-1.2-1.6.patch
	patch-glibc-ieee754-arm-ieee754.h  patch-glibc-ieee754-dbl64-mpn2dbl
	patch-glibc-ieee754-dbl64-dbl2mpn  patch-glibc-ieee754-flt32-mpn2flt
- NWFPE : 
	nwfpe-double_cpdo.c  nwfpe-fpa11_cpdt.c

> Also, there seem to be outstanding issues for non-XScale targets, as the
> assembler is forced to target XScales (but that is good for IXP425 :).
>
> http://sources.redhat.com/ml/crossgcc/2004-02/msg00129.html
>
I'm not using this patch, nor this one : 
http://gcc.gnu.org/ml/gcc-patches/2001-12/msg01370.html

The toolchain is optimized and dedicated to big endian Xscale, but that's not 
a real problem as it's easy to have an optimized toolchain for each target 
and switch the PATH depending on the wanted target.

If you want a generic toolchain optimized for each configuration you will have 
to compile several glibc with each option and to handle several directories 
and options for gcc to get and link to the correct configuration (big-endian+ 
soft-float for example) and this is not very clear to me, so I compiled 
several toolchains using the same process but not the same options/patches :
	- standard (let the kernel handle FP)
	- softfloat, with Nicolas' patch
	- -msoft-float : let have an external libfloat library.

Eric



-------------------------------------------------------------------
Subscription options: http://lists.arm.linux.org.uk/mailman/listinfo/linux-arm
FAQ/Etiquette:       http://www.arm.linux.org.uk/armlinux/mailinglists.php