liboil softfloat fix questions

liboil softfloat fix questions

[Koen Kooi]

Hi,

I noticed liboil got patched:

+if test x"${ENABLE_SOFTFLOAT}" = "xyes"; then
+      VFP_CFLAGS="$VFP_CFLAGS -mfloat-abi=softfp"
+else
+      VFP_CFLAGS="$VFP_CFLAGS -mfpu=vfp"

That looks a bit wrong to me, since for EABI you'd want 
'-mfloat-abi=softfp -mfpu=vfp' to make it use the VFP coprocessor using 
softfloat calling conventions[1] and plain '-mfloat-abi=softfp' to do 
softfloat, and with OABI you don't want to pass '-mfloat-abi=softfp'
To make it even more complicated, on armv7a[2] you want to pass 
-mfpu=neon -mfloat-abi=softfp. If you look at the output of gcc -S 
you'll see that EABI won't use an FPU unless you pass in 
-mfloat-abi=softfp *and* -mfpu=<FPU>.

AFAICT OE passes all the needed magic already in TARGET_CCFLAGS, so it 
seems that this entire construct can be removed.

Robert, could you beat the cacao people with a cluebat so they get their 
stuff right?

regards,

Koen

[1] There are more permutations, but let's focus on the one OE is 
currently using
[2] On cortex-a8 the VFP isn't pipelined, so double precision math is 
twice as slow compared to ARM11 cores, the NEON unit is pipelined, but 
can only do single-precision. So don't be surprised if your new omap3 
nokia tablet performs FP math slower than your n810.

Re: liboil softfloat fix questions

[Khem Raj]

On Thu, Sep 25, 2008 at 12:31 AM, Koen Kooi <k.kooi@student.utwente.nl> wrote:
> Hi,
>
> I noticed liboil got patched:
>
> +if test x"${ENABLE_SOFTFLOAT}" = "xyes"; then
> +      VFP_CFLAGS="$VFP_CFLAGS -mfloat-abi=softfp"
> +else
> +      VFP_CFLAGS="$VFP_CFLAGS -mfpu=vfp"
>
> That looks a bit wrong to me, since for EABI you'd want '-mfloat-abi=softfp
> -mfpu=vfp' to make it use the VFP coprocessor using softfloat calling
> conventions[1] and plain '-mfloat-abi=softfp' to do softfloat, and with OABI
> you don't want to pass '-mfloat-abi=softfp'
> To make it even more complicated, on armv7a[2] you want to pass -mfpu=neon
> -mfloat-abi=softfp. If you look at the output of gcc -S you'll see that EABI
> won't use an FPU unless you pass in -mfloat-abi=softfp *and* -mfpu=<FPU>.

yes thats right

>
> AFAICT OE passes all the needed magic already in TARGET_CCFLAGS, so it seems
> that this entire construct can be removed.
>
> Robert, could you beat the cacao people with a cluebat so they get their
> stuff right?
>
> regards,
>
> Koen
>
> [1] There are more permutations, but let's focus on the one OE is currently
> using
> [2] On cortex-a8 the VFP isn't pipelined, so double precision math is twice
> as slow compared to ARM11 cores, the NEON unit is pipelined, but can only do
> single-precision. '

should get better with cortex-a9

So don't be surprised if your new omap3 nokia tablet
> performs FP math slower than your n810.

>
>
> _______________________________________________
> Openembedded-devel mailing list
> Openembedded-devel@lists.openembedded.org
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-devel
>

Re: liboil softfloat fix questions

[Koen Kooi]

Khem Raj wrote:
> On Thu, Sep 25, 2008 at 12:31 AM, Koen Kooi<k.kooi@student.utwente.nl>  wrote:

>> [2] On cortex-a8 the VFP isn't pipelined, so double precision math is twice
>> as slow compared to ARM11 cores, the NEON unit is pipelined, but can only do
>> single-precision. '
>
> should get better with cortex-a9

Yes, cortex-a9 dumps the vpf-lite and has a proper pipelined version, 
but sadly there isn't any cortex-a9 silicon on the market yet.

regards,

Koen

Re: liboil softfloat fix questions

[Henning Heinold]

On Thu, Sep 25, 2008 at 09:31:43AM +0200, Koen Kooi wrote:
> Hi,
>
> I noticed liboil got patched:
>
> +if test x"${ENABLE_SOFTFLOAT}" = "xyes"; then
> +      VFP_CFLAGS="$VFP_CFLAGS -mfloat-abi=softfp"
> +else
> +      VFP_CFLAGS="$VFP_CFLAGS -mfpu=vfp"
>
> That looks a bit wrong to me, since for EABI you'd want  
> '-mfloat-abi=softfp -mfpu=vfp' to make it use the VFP coprocessor using  
> softfloat calling conventions[1] and plain '-mfloat-abi=softfp' to do  
> softfloat, and with OABI you don't want to pass '-mfloat-abi=softfp'
> To make it even more complicated, on armv7a[2] you want to pass  
> -mfpu=neon -mfloat-abi=softfp. If you look at the output of gcc -S  
> you'll see that EABI won't use an FPU unless you pass in  
> -mfloat-abi=softfp *and* -mfpu=<FPU>.
>
> AFAICT OE passes all the needed magic already in TARGET_CCFLAGS, so it  
> seems that this entire construct can be removed.
>
> Robert, could you beat the cacao people with a cluebat so they get their  
> stuff right?

Hi,

okay seems I have to learn more about fpu stuff.

From cacao I only took the configure option. The Cflags comes from
liboil itself.

I will remove this cruft from the configure.ac and will try again.

Bye Henning